diff --git a/agent/orchestrator/SKILL.md b/agent/orchestrator/SKILL.md
index 1b68bdb22..7e371352d 100644
--- a/agent/orchestrator/SKILL.md
+++ b/agent/orchestrator/SKILL.md
@@ -10,8 +10,8 @@ description: Create and coordinate native agent sessions (Codex, Claude, Qwen, K
 A session started from CanvasTTY's launcher with the **Orchestrator** role has the `canvastty_agents` MCP tools. Use them in this order:
 
 1. `list_providers`: which agents CanvasTTY can launch here. Each entry's `id` is the exact `spawn_agent.provider` value; prefer entries that are `available` with `signIn: "ok"` (`unknown` only means CanvasTTY has not checked yet). When plugins offer launch options (for example an accounts plugin's `list_routes` tool), the answer names that tool.
-2. `spawn_agent` once per part of the task: a known `provider` id, an absolute `cwd` and a self-contained prompt. An unknown provider is refused with a pointer back to `list_providers`. **If the person names a model, pass it as `model`** in the format `list_providers` gives for that provider (`model.format`; OpenCode takes `provider/model`, for example `zai-coding-plan/glm-5.3-flash`, and `model.known` lists what `opencode models` reported); without `model` the CLI uses its own default. `effort` sets the reasoning effort where the CLI has one (`efforts`). Without `profile` the subagent gets your own launch profile, or the next lower one its CLI has (never more than yours, never YOLO); `list_providers` shows each provider's `profiles`. `cwd` must be this project's folder or a folder inside it. A refusal says why (a folder outside the project, a profile above yours, a limit the person set): change the request, do not retry it.
-3. `wait_for_agent` for each subagent (`timeoutSeconds` up to 600; call it again after `timeout`) instead of polling. `needs_approval` means the person must answer a prompt in that card: never answer it yourself.
+2. `spawn_agent` once per part of the task: a known `provider` id, an absolute `cwd` and a self-contained prompt. An unknown provider is refused with a pointer back to `list_providers`. **If the person names a model, pass it as `model`** in the format `list_providers` gives for that provider (`model.format`; OpenCode takes `provider/model`, for example `zai-coding-plan/glm-5.3-flash`, and `model.known` lists what `opencode models` reported); without `model` a subagent of your own CLI keeps your plugin launch options (for example your model account), and any other CLI uses its own default. A plugin you do not name in `launchOptions` keeps your own values; passing `model` next to inherited options that already set the model is refused, and a subagent of another CLI is refused while your model comes from a plugin's options (spawn one of your own CLI instead). The answer's `servedBy` names the CLI, the model asked for, and whether each plugin's options were `inherited` or `explicit`. `effort` sets the reasoning effort where the CLI has one (`efforts`). Without `profile` the subagent gets your own launch profile, or the next lower one its CLI has (never more than yours, never YOLO); `list_providers` shows each provider's `profiles`. `cwd` must be this project's folder or a folder inside it. A refusal says why (a folder outside the project, a profile above yours, a limit the person set): change the request, do not retry it.
+3. `wait_for_agent` for each subagent (`timeoutSeconds` up to 100; a call returns within 50 seconds unless you are Claude Code; `timeout` means it is still running, so read `output` as progress and call it again) instead of polling. `needs_approval` means the person must answer a prompt in that card (`output` shows it, for example a shell command): never answer it yourself, and `send_to_agent` is refused meanwhile. Tell the person what it asks, then call `wait_for_agent` again: it waits for their answer.
 4. `get_agent_result` to read what the subagent produced; check the actual files before reporting success.
 
 Do not explore the filesystem, `PATH` or config folders (for example `~/.local/bin` or `~/.config/<cli>`) looking for agents or their configuration: `list_providers` is what CanvasTTY can launch, and delegation goes through `spawn_agent`.
@@ -32,10 +32,10 @@ Use the bundled `canvastty-control.mjs` CLI with a running CanvasTTY instance wh
 
    Run `providers` first to see which providers are installed and signed in. If the person names a model, add `--model <id>` (OpenCode: `provider/model`); `--effort <level>` works for Codex, Claude and Grok.
 
-   `--provider` accepts codex, claude, qwen, kimi, opencode, hermes, grok, omp and pi; each worker uses that provider's own normal/YOLO launch flags. The create response includes `capabilities`: only Codex workers have `result: true` (captured final answer) and `menus: true` (startup/approval menus that `choose`/`dismiss` can act on). For other providers, `result` ends as `no_result` and menus must be resolved by the user; treat their `screen` text as the only evidence. From the person's own terminal CLI creation defaults to YOLO, using the provider's full-access, no-approval launch flag; inside an Orchestrator session it defaults to that session's profile. It does not change global provider settings. `--profile normal` is available when the user requests their ordinary configured profile; do not silently replace their selected permissions.
+   `--provider` accepts codex, claude, qwen, kimi, opencode, hermes, grok, omp and pi; each worker uses that provider's own normal/YOLO launch flags. The create response includes `capabilities`: only Codex workers have `result: true` (captured final answer) and `menus: true` (startup menus that `choose`/`dismiss` can act on). For other providers, `result` ends as `no_result` and menus must be resolved by the user; treat their `screen` text as the only evidence. From the person's own terminal CLI creation defaults to YOLO, using the provider's full-access, no-approval launch flag; inside an Orchestrator session it defaults to that session's profile. It does not change global provider settings. `--profile normal` is available when the user requests their ordinary configured profile; do not silently replace their selected permissions.
 4. Save the returned session ID and request ID. Verify provider, profile, cwd and launch status. A returned ID or an idle status alone does not prove a working agent.
 5. Inspect `screen <id>` until the worker is at an empty composer (verified automatically for Codex; for other providers judge from the screen text). Trust, authentication, permission menus and startup failures are not task prompts. Do not submit a task to dismiss them.
-6. Resolve only reviewed, authorized menu actions: `choose <id> --choice <observed-number> --revision <interaction.revision>`. For an idle menu offering Escape, `dismiss <id> --revision <screen.revision>` closes it. A stale revision requires another inspection. In particular, inspect new or changed lifecycle hooks before trusting them; YOLO does not imply approval of unknown hooks. Verify the next screen after each choice.
+6. An approval or permission prompt (`status` is `needs_approval`, or Codex asks to run a command, apply edits or grant access) is the person's to answer, never yours: inside an Orchestrator session `choose`, `dismiss` and `send` are refused with `NEEDS_PERSON` meanwhile. Tell the person what it asks, then check `status` again. Resolve only reviewed, authorized startup menu actions: `choose <id> --choice <observed-number> --revision <interaction.revision>`. For an idle menu offering Escape, `dismiss <id> --revision <screen.revision>` closes it. A stale revision requires another inspection. In particular, inspect new or changed lifecycle hooks before trusting them; YOLO does not imply approval of unknown hooks. Verify the next screen after each choice.
 
 ## Run and verify
 
diff --git a/docs/ARCHITECTURE.md b/docs/ARCHITECTURE.md
index aa004a203..218c35143 100644
--- a/docs/ARCHITECTURE.md
+++ b/docs/ARCHITECTURE.md
@@ -70,7 +70,7 @@ Electron main process
 - `src/main/services/BrowserService.ts` is the only owner of the built-in browser's `WebContentsView` tabs and shared persistent partition. Remote pages have no preload or Node access, keep context isolation and sandbox enabled, and cannot request hardware, location, notification, clipboard-read, certificate-bypass, or external-protocol capabilities. HTTP(S) popups are adopted as internal tabs; other schemes are rejected.
 - `src/main/services/browser/` contains the browser kernel. `BrowserStore` atomically persists only tab order, active tab, and safe restore URLs. `BrowserPolicyService` centralizes URL, permission, download, and upload rules; validated uploads are copied through an already-open no-follow file descriptor into private staging before Chromium sees them. `BrowserAutomationService` attaches Electron's internal debugger to the existing live tab without a remote-debugging port. `BrowserCommandDispatcher` adds revisions, revision-bound refs, mutation request deduplication, per-tab FIFO mutation lanes, bounded concurrency, typed errors, and redacted fail-closed audit for agent mutations.
 - `src/main/services/agent-browser/` exposes the kernel only through an authenticated user-local Unix socket (`0600`) or Windows named pipe. The Windows pipe is created by the bundled native host with a protected DACL containing only the exact current-user SID and rejects remote clients. Each agent PTY receives a one-use bootstrap capability through its child environment. A successful authentication rotates it to a session-scoped reconnect capability held only in helper memory; duplicate bootstrap authentication is accepted only while the same `connectionId` is already live, and every capability is revoked when the PTY ends. The bundled stdio MCP helper is the only protocol adapter; no TCP listener, cookie/storage endpoint, arbitrary evaluation tool, or raw CDP surface exists.
-- `src/main/services/agent-runtime/` is a separate lifecycle boundary and is not controlled by the Browser access switch. When CanvasTTY status hooks are enabled, every agent PTY receives a distinct capability for a protected user-local socket/pipe. Provider command hooks and the OpenCode event plugin may report only the fixed status enum, bounded event name, and optional opaque turn/prompt ID; prompt text, responses, tool input, and arbitrary telemetry are rejected by the exact gateway schema. On macOS and Linux the helpers are the native `canvastty-helper` (`native/canvastty-helper`, Go, standard library only; subcommands `mcp-browser`, `mcp-orchestration`, `permission-gate`, `hook`), which speaks exactly the wire protocols of the `.mjs` helpers (`tests/native-helper-*-parity.test.mjs` run both) without an Electron-as-Node process per agent and per tool call; `src/main/services/agentHelpers.ts` falls back to the `.mjs` helpers when the binary is absent, on Windows, or with `CANVASTTY_HELPERS=node`. Electron helper commands carry `ELECTRON_RUN_AS_NODE=1` inside the exact hook command only; the provider PTY never inherits that process-mode flag, so a provider cannot accidentally launch a second CanvasTTY GUI instance. The user can revoke this capability from Agents settings, immediately returning live agent status to `unavailable`; re-enabling requires a new/restarted PTY. Explicitly trusted plugin hooks use a separate process runner which re-checks the private PluginManager registry on every invocation and strips CanvasTTY internal capabilities before passing the provider payload to third-party code. Provider-native review remains an independent gate; CanvasTTY does not bypass Codex hook trust globally.
+- `src/main/services/agent-runtime/` is a separate lifecycle boundary and is not controlled by the Browser access switch. When CanvasTTY status hooks are enabled, every agent PTY receives a distinct capability for a protected user-local socket/pipe. Provider command hooks and the OpenCode event plugin may report only the fixed status enum, bounded event name, and optional opaque turn/prompt ID; prompt text, responses, tool input, and arbitrary telemetry are rejected by the exact gateway schema. On macOS and Linux the helpers are the native `canvastty-helper` (`native/canvastty-helper`, Go, standard library only; subcommands `mcp-browser`, `mcp-orchestration`, `permission-gate`, `hook`), which speaks exactly the wire protocols of the `.mjs` helpers (`tests/native-helper-*-parity.test.mjs` run both) without an Electron-as-Node process per agent and per tool call; `src/main/services/agentHelpers.ts` falls back to the `.mjs` helpers when the binary is absent, on Windows, or with `CANVASTTY_HELPERS=node`. Electron helper commands carry `ELECTRON_RUN_AS_NODE=1` inside the exact hook command only; the provider PTY never inherits that process-mode flag, so a provider cannot accidentally launch a second CanvasTTY GUI instance. The user can revoke this capability from Agents settings, immediately returning live agent status to `unavailable`; re-enabling requires a new/restarted PTY. Explicitly trusted plugin hooks use a separate process runner which re-checks the private PluginManager registry on every invocation and strips CanvasTTY internal capabilities before passing the provider payload to third-party code. Provider-native review remains an independent gate; CanvasTTY does not bypass Codex hook trust globally. A Codex or OpenCode subagent's final answer (Codex `Stop` hook, the OpenCode plugin at `session.idle`) travels on the same hooks, so it exists only while status hooks are on for that launch: whether a launch has them is decided when it is planned (a deferred plugin launch included), and a launch planned with them off has no lease and no lifecycle hooks at all. `wait_for_agent` and `get_agent_result` then return `answerUnavailable: { reason: "status_hooks_off", message }` and the orchestrator reads the output instead. An answer is kept only for its own turn (`src/main/services/agent-runtime/applyRuntimeSignal.ts`, the glue main runs for each signal): OpenCode names each turn, a submitted input other than an approval answer or an empty Enter in an idle card starts a new one, and a signal of the current turn that predates that input still ends the wait but brings no answer.
 - Lifecycle adapters use launch-only settings for Claude, Codex, Qwen, and OpenCode. Kimi, Hermes, and Grok, whose hook discovery is home-config based, receive ownership-checked temporary entries shared across live CanvasTTY sessions. Kimi and Hermes keep recovery journals and exact backups; Grok uses a dedicated owned hook file. Cleanup restores exact original bytes when no concurrent edit occurred and otherwise removes only CanvasTTY-owned entries.
 - `TerminalManager` injects the MCP helper per launch without leaving permanent provider configuration. Claude Code, Codex, and Qwen Code receive CLI arguments; Qwen gets one inline `--mcp-config` entry that overrides only the CanvasTTY server name and leaves unrelated user servers available. OpenCode receives a merged launch-only `OPENCODE_CONFIG_CONTENT` entry plus one scoped browser-tool permission; Kimi uses its per-run MCP configuration when supported. Older Kimi versions receive a compare-and-swap temporary CanvasTTY entry and one exact permission rule with an atomic recovery journal. Hermes receives a temporary `mcp_servers.canvastty_browser` entry in `HERMES_HOME/config.yaml` (defaulting to `~/.hermes/config.yaml` on POSIX or `%LOCALAPPDATA%\hermes\config.yaml` on Windows); sensitive capability values stay as child-environment placeholders. Temporary Kimi and Hermes configuration remains until the final owning PTY session ends, then exact original bytes are restored when safe. A journal repairs an interrupted Hermes launch at the next CanvasTTY startup, while compare-and-swap checks preserve concurrent user edits. Unrelated MCP entries, credentials, and file/shell permissions are preserved. Qwen, OpenCode, and Hermes YOLO remain launch-only and do not change persistent permission settings.
 - `src/main/services/providerCliRegistry.ts` is the single owner of provider CLI discovery. During main-process startup it creates a shared snapshot for every provider in `PROVIDER_CLI_DEFINITIONS` — each definition declares the executable command names it may install as (which may differ from the provider ID, e.g. a provider shipping as `mcode`) and optional known home-relative or Windows LOCALAPPDATA-relative directories — by checking smoke-only overrides, the inherited `PATH`, platform defaults, and those known directories in that order. Available entries retain an absolute executable, launcher kind, and supplemented child `PATH`; POSIX entries must be executable files and Windows entries must be supported native or batch launchers. `TerminalManager`, `LimitsService`, agent-browser probes, and provider smoke tests consume that same snapshot and never repeat command lookup. Missing entries produce a failed session with copyable checked-path diagnostics before PTY or temporary browser configuration creation, while the limit adapter reports `cli-not-found`; its HOME row is hidden until explicitly selected after CLI detection. Launchers keep a provider without a local CLI visible when an account bound to a remote computer or a saved container profile provides another route. CanvasTTY never reads shell startup scripts. Agents settings can recheck candidate paths, atomically replace the registry snapshot, reconcile saved launcher and limit selections, and refresh CLI-bound adapters without restarting the app. Existing sessions keep running; a newly found CLI remains disabled until selected.
diff --git a/docs/ARCHITECTURE.ru.md b/docs/ARCHITECTURE.ru.md
index ae66a88d0..d7e0054a4 100644
--- a/docs/ARCHITECTURE.ru.md
+++ b/docs/ARCHITECTURE.ru.md
@@ -48,7 +48,7 @@ Electron main process
 - `src/main/services/PluginSecretsService.ts` сериализует запись секретов каждого плагина, шифрует весь ограниченный payload через Electron `safeStorage`, отклоняет plaintext-only backend и удаляет зашифрованный файл при uninstall. `ProviderSecretsService.ts` применяет ту же архитектуру к API-ключам провайдеров для CLI с BYOK: значения остаются в main-процессе, а renderer-контракт раскрывает только флаги `configured` и операции set/clear. Записи настроек `ApiProfile` именуют model-бэкенды (протокол, HTTPS base URL, ссылка на секрет) для тех же BYOK-рантаймов; это не agent providers, а normalizer настроек отбрасывает невалидные профили вместо «ремонта».
 - `src/main/services/PluginMediaService.ts` сохраняет разрешения только после нативного выбора папки, скрывает абсолютные пути, пропускает symlinks и отдаёт аудио с HTTP Range. Чтение плейлистов остаётся внутри разрешённых библиотек; ограниченная атомарная запись разрешена только в `Playlists/`.
 - `src/main/services/BrowserService.ts` владеет вкладками встроенного браузера в `WebContentsView`. Удалённые страницы используют отдельный persistent partition с выключенным Node, включёнными context isolation/sandbox и отклонением website permissions по умолчанию. Это core service, а не возможность runtime-плагина.
-- `src/main/services/agent-runtime/` — отдельная всегда включённая lifecycle-граница, не зависящая от переключателя Browser access. Каждый agent PTY получает собственный capability для защищённого user-local socket/pipe. Provider command hooks и OpenCode event plugin могут передать только фиксированный status enum, ограниченное имя события и необязательный opaque turn/prompt ID; точная schema Gateway отклоняет prompt text, ответы, tool input и произвольную telemetry. При завершении PTY capability и временные файлы отзываются.
+- `src/main/services/agent-runtime/` — отдельная всегда включённая lifecycle-граница, не зависящая от переключателя Browser access. Каждый agent PTY получает собственный capability для защищённого user-local socket/pipe. Provider command hooks и OpenCode event plugin могут передать только фиксированный status enum, ограниченное имя события и необязательный opaque turn/prompt ID; точная schema Gateway отклоняет prompt text, ответы, tool input и произвольную telemetry. При завершении PTY capability и временные файлы отзываются. Финальный ответ субагента Codex или OpenCode (хук Codex `Stop`, плагин OpenCode на `session.idle`) идёт через те же хуки, поэтому он есть только пока status hooks включены для этого запуска: это решается при планировании запуска (включая отложенный запуск с плагинами), и запуск, спланированный с выключенными хуками, не получает ни lease, ни lifecycle-хуков. Тогда `wait_for_agent` и `get_agent_result` возвращают `answerUnavailable: { reason: "status_hooks_off", message }`, и оркестратор читает вывод. Ответ хранится только для своего хода (`applyRuntimeSignal.ts`): OpenCode именует каждый ход, отправленный ввод (кроме ответа на запрос подтверждения и пустого Enter в idle-карточке) начинает новый, а сигнал текущего хода, начатого до этого ввода, завершает ожидание, но без ответа.
 - Claude, Codex, Qwen и OpenCode получают lifecycle hooks только на текущий запуск. Для Kimi, Hermes и Grok, которые ищут hooks в home-конфигурации, используются ownership-checked временные записи с совместным владением живых сессий. Kimi и Hermes используют recovery journals и точные backups, Grok — отдельный owned hook file; cleanup восстанавливает исходные байты или удаляет только записи CanvasTTY при конкурентных изменениях.
 - `TerminalManager` подмешивает MCP helper, не оставляя постоянных изменений в provider-конфигах. Claude Code, Codex и Qwen Code получают CLI arguments; Qwen получает одну inline-запись `--mcp-config`, которая переопределяет только имя сервера CanvasTTY и не скрывает сторонние user servers. OpenCode — объединённый launch-only `OPENCODE_CONFIG_CONTENT` с одной scoped browser-tool permission, Kimi — per-run MCP config или временную запись с compare-and-swap и recovery journal для старых версий. Hermes получает временную запись `mcp_servers.canvastty_browser` в `HERMES_HOME/config.yaml` (по умолчанию `~/.hermes/config.yaml` в POSIX или `%LOCALAPPDATA%\hermes\config.yaml` в Windows); чувствительные capability-значения остаются ссылками на окружение дочернего процесса. Временная конфигурация Kimi и Hermes живёт до завершения последней владеющей PTY-сессии, после чего исходные байты точно восстанавливаются, если файл не менялся параллельно. Journal восстанавливает Hermes после прерванного запуска при следующем старте CanvasTTY, а compare-and-swap сохраняет одновременные пользовательские изменения. Сторонние MCP-записи, credentials и file/shell permissions не затрагиваются. Qwen, OpenCode и Hermes YOLO остаются launch-only и не меняют постоянные permission-настройки.
 - `src/main/services/providerCliRegistry.ts` — единственный владелец обнаружения provider CLI. При запуске main-процесса он создаёт общий snapshot для каждого провайдера из `PROVIDER_CLI_DEFINITIONS` — каждое определение объявляет имена executable-команд, под которыми провайдер может устанавливаться (они могут отличаться от ID провайдера, например провайдер с командой `mcode`), и опциональные известные каталоги (относительно home или Windows LOCALAPPDATA), — последовательно проверяя smoke-only overrides, унаследованный `PATH`, системные каталоги платформы и эти известные каталоги. Доступная запись хранит абсолютный executable, тип launcher-а и дополненный дочерний `PATH`; POSIX-кандидат обязан быть исполняемым файлом, а Windows-кандидат — поддерживаемым native или batch launcher-ом. `TerminalManager`, `LimitsService`, agent-browser probes и provider smoke используют один и тот же snapshot и не повторяют поиск команды. Недоступный CLI создаёт failed-сессию с копируемой диагностикой проверенных путей до создания PTY или временной browser-конфигурации, а адаптер лимитов сообщает `cli-not-found`; строка HOME скрыта до ручного выбора после обнаружения CLI. Провайдер без локального CLI остаётся в панелях запуска, если для него есть аккаунт на удалённом компьютере или сохранённый контейнерный профиль. CanvasTTY не читает shell startup scripts. Настройки агентов позволяют повторно проверить пути, атомарно заменить snapshot registry, согласовать сохранённые списки запуска и лимитов и обновить зависящие от CLI адаптеры без перезапуска. Работающие сессии продолжаются; найденный позже CLI остаётся выключенным до ручного выбора.
diff --git a/docs/ARCHITECTURE.zh-CN.md b/docs/ARCHITECTURE.zh-CN.md
index 34dcc04e0..1f7d0b66e 100644
--- a/docs/ARCHITECTURE.zh-CN.md
+++ b/docs/ARCHITECTURE.zh-CN.md
@@ -49,7 +49,7 @@ Electron main process
 - `src/main/services/PluginSecretsService.ts` 串行化每个插件的机密写入，通过 Electron `safeStorage` 加密完整的有界 payload，拒绝 plaintext-only backend，并在卸载时删除加密文件。`ProviderSecretsService.ts` 将同一架构应用于面向 BYOK CLI 的服务商 API key：值只留在 main 进程，renderer 契约只暴露每个 key 的 `configured` 标志与 set/clear 操作。`ApiProfile` 设置项为同一批 BYOK 运行时命名 model 后端（协议、HTTPS base URL、secret 引用）；它们不是 agent provider，且 settings normalizer 会丢弃而非修复无效 profile。
 - `src/main/services/PluginMediaService.ts` 仅在原生目录选择后保存授权，隐藏绝对路径，跳过 symlink，并以 HTTP Range 提供音频。Playlist 读取限制在授权媒体库内；写入受大小限制，并且只能原子写入 `Playlists/`。
 - `src/main/services/BrowserService.ts` 管理内置浏览器的 `WebContentsView` tab。远程页面使用独立 persistent partition，禁用 Node，启用 context isolation/sandbox，并默认拒绝网站权限。这是 core service，不是 runtime 插件能力。
-- `src/main/services/agent-runtime/` 是独立且始终启用的 lifecycle 边界，不受 Browser access 开关控制。每个 agent PTY 都为受保护的 user-local socket/pipe 获得独立 capability。Provider command hook 与 OpenCode event plugin 只能提交固定 status enum、受限 event 名称和可选 opaque turn/prompt ID；Gateway 的精确 schema 会拒绝 prompt text、回复、tool input 与任意 telemetry。PTY 退出时 capability 与临时文件都会被撤销。
+- `src/main/services/agent-runtime/` 是独立且始终启用的 lifecycle 边界，不受 Browser access 开关控制。每个 agent PTY 都为受保护的 user-local socket/pipe 获得独立 capability。Provider command hook 与 OpenCode event plugin 只能提交固定 status enum、受限 event 名称和可选 opaque turn/prompt ID；Gateway 的精确 schema 会拒绝 prompt text、回复、tool input 与任意 telemetry。PTY 退出时 capability 与临时文件都会被撤销。 Codex 或 OpenCode 子 agent 的最终回答（Codex `Stop` hook、OpenCode 插件在 `session.idle` 时）经由同一组 hook 传递，因此只有该次启动开启了 status hooks 时才存在：是否开启在规划启动时决定（包括延迟的插件启动），在关闭状态下规划的启动既没有 lease 也没有任何 lifecycle hook。此时 `wait_for_agent` 与 `get_agent_result` 返回 `answerUnavailable: { reason: "status_hooks_off", message }`，编排者改为读取输出。回答只归属于它自己的 turn（`applyRuntimeSignal.ts`）：OpenCode 为每个 turn 命名；除审批回答和空闲卡片中的空 Enter 外，提交的输入会开始新 turn；在该输入之前开始的当前 turn 的信号仍会结束等待，但不带回答。
 - Claude、Codex、Qwen 与 OpenCode 使用仅本次启动有效的 lifecycle hook。Kimi、Hermes 与 Grok 只能从 home 配置发现 hook，因此使用由实时 CanvasTTY 会话共享、带 ownership 检查的临时条目。Kimi 与 Hermes 使用 recovery journal 和精确 backup；Grok 使用独立 owned hook 文件。Cleanup 在无并发编辑时逐字恢复原文件，否则只移除 CanvasTTY 自己的条目。
 - `TerminalManager` 注入 MCP helper 时不会留下永久的服务商配置变更。Claude Code、Codex 与 Qwen Code 使用 CLI 参数；Qwen 使用一个 inline `--mcp-config`，只覆盖 CanvasTTY 服务名，不隐藏无关用户服务。OpenCode 使用合并后的、仅本次启动有效的 `OPENCODE_CONFIG_CONTENT` 和一条 scoped browser-tool 权限；Kimi 使用 per-run MCP 配置，旧版本则使用带 compare-and-swap 与 recovery journal 的临时配置。Hermes 会在 `HERMES_HOME/config.yaml` 中获得临时 `mcp_servers.canvastty_browser` 配置项（POSIX 默认路径为 `~/.hermes/config.yaml`，Windows 默认路径为 `%LOCALAPPDATA%\hermes\config.yaml`），敏感 capability 值仍以子进程环境变量占位符保存。Kimi 与 Hermes 的临时配置会保留到最后一个所属 PTY 会话结束；若文件未被并发修改，则精确恢复原始字节。若 Hermes 启动意外中断，journal 会在 CanvasTTY 下次启动时修复配置，compare-and-swap 则保留用户的并发修改。其他 MCP 配置项、凭据和文件/shell 权限不会受影响。Qwen、OpenCode 与 Hermes 的 YOLO 都不修改持久权限设置。
 - `src/main/services/providerCliRegistry.ts` 是服务商 CLI 发现的唯一职责边界。main 进程启动时，它按 smoke-only override、继承的 `PATH`、平台默认目录、已知用户/服务商目录的顺序，为 `PROVIDER_CLI_DEFINITIONS` 中的每个服务商创建一个共享快照——每个定义声明该服务商可能安装的 executable 命令名（可以与服务商 ID 不同，例如命令为 `mcode` 的服务商），以及可选的已知目录（相对 home 或 Windows LOCALAPPDATA）。可用条目保存绝对 executable、launcher 类型以及补充后的子进程 `PATH`；POSIX 候选必须是可执行文件，Windows 候选必须是受支持的 native 或 batch launcher。`TerminalManager`、`LimitsService`、agent-browser probe 与 provider smoke 共用该快照，不再各自查找命令。CLI 不可用时，系统会在创建 PTY 或临时 browser 配置之前生成 failed session，并提供可复制的已检查路径诊断；限额适配器报告 `cli-not-found`；HOME 行保持隐藏，直到检测到 CLI 后由用户手动选择。若某服务商绑定了远程计算机上的账户或已保存的容器配置，即使本地没有 CLI，启动器仍会显示它。CanvasTTY 不读取 shell startup script。Agents 设置可重新检查候选路径、原子替换 registry 快照、调整已保存的启动器和限额选择，并在无需重启的情况下刷新依赖 CLI 的适配器。运行中的 session 保持不变；新检测到的 CLI 需手动启用。
diff --git a/docs/UI_CONTRACT.md b/docs/UI_CONTRACT.md
index 7acaf490d..609d8cb14 100644
--- a/docs/UI_CONTRACT.md
+++ b/docs/UI_CONTRACT.md
@@ -65,7 +65,7 @@ This contract preserves the approved MVP concept and prevents feature ownership
 - An authenticated connection or heartbeat alone never shows agent presence. A branded badge appears only after that agent actually issues a browser command, and its cursor appears only after a real pointer position exists. Claude uses `#D97757`, Codex `#10A37F`, Qwen `#6D44E8`, Kimi `#7C5CFC`, OpenCode `#5A5858`, Hermes `#D6A700`, and an unknown provider `#7A8291`. Browser Settings separately controls indicator visibility without revoking browser access; the agent-access kill switch revokes the connection itself.
 - With window snapping enabled, drag and resize use a hidden `10px` grid and a `10px` magnetic threshold for neighboring edges, centers, and a consistent `20px` gap.
 - One trusted context dispatcher owns right-clicks on the canvas. Truly empty canvas offers region creation, a sticky note, the Settings-configured launcher submenu, Browser, and Settings. Holding RMB on empty canvas also opens the configurable radial quick launcher: moving toward an action and releasing activates it, while a stationary right-click continues to open the standard empty-canvas menu. Region and note right-clicks remain context-specific. A region offers rename, color, a note at the pressed point, and region-only deletion. A note frame offers edit, bring-to-front, and deletion. Terminal, Browser, plugin, textarea/input, and contenteditable surfaces keep their native context menus; middle-button panning is unchanged. `Cmd/Ctrl+K` opens a searchable palette over live sessions and the same global creation/launch actions. Regions persist independently of terminal restore, render behind HOME and windows, resize from every edge/corner, and move every terminal, plugin, Browser, or note whose complete bounds were inside the region at drag start. Attached contents follow the region continuously during the gesture and persist once at pointer release; deleting a region never deletes its contents. Sticky notes persist bounded text and geometry, snap like other windows, and resize from every edge/corner. General Settings exposes independent **Save colored regions after exit** and **Save notes after exit** controls. Both default on to preserve existing profiles. Turning either off removes only that collection from disk persistence without deleting its live canvas objects; turning it back on saves the current collection, while leaving it off starts the next launch with that collection empty.
-- Context menus, region editors, the radial launcher, and the command palette use official Lucide/provider assets and shared application color tokens. Their dimensions follow root `--ui-scale`. Appearance owns the `0.85–1.25` scale in `0.05` steps and the sage/lilac/night system palette; the scale changes UI chrome rather than canvas zoom or terminal font size. Agents owns the independent context-menu and radial-launcher subsets and their safe reset defaults.
+- Context menus, region editors, the radial launcher, and the command palette use official Lucide/provider assets and shared application color tokens. Their dimensions follow root `--ui-scale`, as do the terminal card options menu and the canvas shortcut-hints panel. Appearance owns the `0.85–1.25` scale in `0.05` steps and the sage/lilac/night system palette; the scale changes UI chrome rather than canvas zoom or terminal font size. Agents owns the independent context-menu and radial-launcher subsets and their safe reset defaults.
 - Settings uses the approved two-pane modal grammar: a persistent icon-labelled sidebar for all seven sections, a separate current-section header, and a scrollable content pane. Settings must not fall back to the legacy light slide-over with one horizontal segmented tab strip. Simple settings use compact copy/control rows; all Settings chrome reads the active system tokens, official Lucide assets, and root `--ui-scale`.
 - General Settings places the macOS/Windows/Linux/Custom keyboard preset after language, followed by an expandable Keyboard shortcuts section for canvas, terminal, Codex editor and canvas navigation bindings. Recording a shortcut selects Custom; choosing a named preset replaces bindings. Optional commands can be disabled, while submit, newline and whole-draft selection remain bound. Canvas and terminal bindings update immediately; native Codex editor bindings apply to new or restarted cards. Focused terminal and text inputs retain ownership of canvas keyboard shortcuts on every platform.
 - Empty-canvas drag and middle-button drag pan the canvas. On fresh profiles ordinary scroll over empty canvas pans on both axes; pinch and `Cmd/Ctrl + scroll` zoom around the pointer. **Use scroll wheel to zoom** restores historical ordinary-wheel zoom and is enabled by migration for existing profiles. Only a focused input widget interrupts canvas wheel/pinch, except the HOME session list keeps ordinary local scrolling without a focus precondition; unfocused or non-focusable widgets otherwise leave those events to the canvas. **Wheel/pinch capture over widgets** has Off, On, and Key modes: Off preserves local ownership, On always transfers only wheel/pinch to the canvas, and Key transfers only wheel/pinch while its binding is held. Key and full-navigation bindings accept modifier chords or Mouse3/Mouse4/Mouse5. A focused live Browser page therefore scrolls normally in Off or released-Key mode; unfocused Browser content, On, an active Key binding, the full override, pinch, and `Cmd/Ctrl + scroll` are canvas-owned. The owner selected for a Browser wheel sequence is held until 250 ms of wheel inactivity. Browser summaries and placeholders are always canvas-owned. Fresh profiles use Key with `Command` on macOS or `Ctrl` elsewhere. A separate configurable **Canvas navigation override** (default `Option` on macOS and `Alt` elsewhere) transfers wheel/pinch and drag while held. While full override is held all canvas surfaces show `grab`; an owned drag shows `grabbing` until pointer up/cancel even after the binding is released.
diff --git a/docs/agent-orchestration.md b/docs/agent-orchestration.md
index d53c4e212..cfff9ecdd 100644
--- a/docs/agent-orchestration.md
+++ b/docs/agent-orchestration.md
@@ -9,10 +9,10 @@ A session launched from the desktop with the **Orchestrator** role gets the `can
 | Tool | What it does |
 | --- | --- |
 | `list_providers` | The agent providers CanvasTTY can launch as subagents: `id` (the exact `spawn_agent.provider` value), `name`, `installed`/`available` (from the provider CLI registry resolved at startup or on a recheck), `signIn` (`ok`, `signed_out`, `expired` or `unknown`, from the last usage-limits read; it never starts a read and never reads credentials), `subagent` and `orchestrator` support, `model` (whether the CLI takes one, its format, and for OpenCode the models a cached `opencode models` listing reported: run in the background with a 5 s timeout, refreshed after 10 minutes, never awaited) and `efforts`, and plugin launch options (`launchOptions`, plus `launchOptionTools` such as an accounts plugin's `list_routes`). Fast: cached state only, no network calls. |
-| `spawn_agent` | Launches a subagent (profile: see below). `provider` must be a `list_providers` id (the schema lists them); an unknown one is refused with `INVALID_REQUEST` naming `list_providers`. Optional `model` goes to the CLI's own `--model` for that run (OpenCode `provider/model`; Codex, Claude, Qwen, Kimi alias, Grok, OMP, Pi, Cursor; Hermes, MiniMax, Devin and Antigravity take none) and optional `effort` to its reasoning effort (Codex `-c model_reasoning_effort`, Claude `--effort`, Grok `--reasoning-effort`). Both are checked per provider, kept on restart and restore, and never written to the CLI's configuration. If the person names a model, the orchestrator passes it as `model`. An OpenCode model its own `opencode models` list does not contain is refused before launch with up to five closest listed ids (OpenCode would otherwise stop with only "Unexpected server error"); when that list cannot be read, the model is passed as given. A subagent that exits anyway reports the last lines of its screen as `exitLines` in `wait_for_agent`, `observe_agent` and `get_agent_result`. |
-| `wait_for_agent` | `{ sessionId, timeoutSeconds ≤ 600 }` (default 55). Returns when the subagent is `idle` (after a prompt: only once a turn that started after the latest delivered prompt has ended, so the CLI's startup idle does not count), `needs_approval`, `done`/`failed` (exited), `quiet` (reports no status and its screen stopped changing), `closed`, or on `timeout`, with `status`, `exitCode`, `waitedMs` and the masked terminal tail (masked before it is cut). Only this orchestrator's own subagents; it stops at once when the call is canceled or the orchestrator disconnects. |
-| `get_agent_result`, `observe_agent` | `get_agent_result`: `answer`, the final reply of the subagent's last turn as the agent itself reported it (Codex through its Stop hook, OpenCode through CanvasTTY's plugin reading the session's last assistant message at `session.idle`; at most 4,096 characters with the end kept, masked, in memory only and cleared when the next turn starts), plus `status`, the exit state and the masked terminal tail. Other providers have no `answer`: their tail is raw screen output. `observe_agent`: the current status and tail. |
-| `send_to_agent`, `cancel_agent`, `list_agents` | Follow-up prompts, disposing a subagent, listing this session's subagents. |
+| `spawn_agent` | Launches a subagent (profile: see below). `provider` must be a `list_providers` id (the schema lists them); an unknown one is refused with `INVALID_REQUEST` naming `list_providers`. Optional `model` goes to the CLI's own `--model` for that run (OpenCode `provider/model`; Codex, Claude, Qwen, Kimi alias, Grok, OMP, Pi, Cursor; Hermes, MiniMax, Devin and Antigravity take none) and optional `effort` to its reasoning effort (Codex `-c model_reasoning_effort`, Claude `--effort`, Grok `--reasoning-effort`). Both are checked per provider, kept on restart and restore, and never written to the CLI's configuration. If the person names a model, the orchestrator passes it as `model`. Plugin launch options the call does not name are inherited from the orchestrator for a subagent of the same CLI (see Plugin launch options below); `model` next to inherited options that set the model is refused, and so is a subagent of another CLI while the orchestrator's model comes from a plugin's options. The answer's `servedBy` is `{ provider, model, launchOptions: { <pluginId>: "inherited" | "explicit" } }`. An OpenCode model its own `opencode models` list does not contain is refused before launch with up to five closest listed ids (OpenCode would otherwise stop with only "Unexpected server error"); when that list cannot be read, the model is passed as given. A subagent that exits anyway reports the last lines of its screen as `exitLines` in `wait_for_agent`, `observe_agent` and `get_agent_result`. |
+| `wait_for_agent` | `{ sessionId, timeoutSeconds ≤ 100 }` (default 50). For an orchestrator other than Claude Code a call returns after at most 50 seconds, because its MCP client (Codex, OpenCode and others) ends a tool call after about 60 seconds; `timeout` then means the subagent is still running, with its progress in `output` and a `message` to call again. Returns when the subagent is `idle` (after a prompt: only once a turn that started after the latest delivered prompt has ended, so the CLI's startup idle does not count), `needs_approval` (once per prompt: a later wait on the same prompt waits for the person's answer instead of returning it again), `done`/`failed` (exited), `quiet` (reports no status and its screen stopped changing), `closed`, or on `timeout`, with `status`, `exitCode`, `waitedMs`, the end of the subagent's screen as masked plain text (as `get_agent_result`) and, for Codex and OpenCode subagents, `answer` (see `get_agent_result`). A turn that the person interrupted by submitting their own input in the subagent's card still ends the wait, without an `answer`: that answer is no longer for the orchestrator's prompt. With CanvasTTY status hooks off (Settings → Agents) a subagent has no hooks at all, so no `answer` can be captured: the result carries `answerUnavailable: { reason: "status_hooks_off", message }` instead and the wait ends `quiet`; read `output` (or `observe_agent`). Turning status hooks on applies to subagents started or restarted afterwards. Only this orchestrator's own subagents; it stops at once when the call is canceled or the orchestrator disconnects. |
+| `get_agent_result`, `observe_agent` | `get_agent_result`: `answer`, the final reply of the subagent's last turn as the agent itself reported it (Codex through its Stop hook, OpenCode through CanvasTTY's plugin reading the session's last assistant message at `session.idle`; at most 4,096 characters with the end kept, masked, in memory only, kept only for the turn it ended and cleared when the next input is submitted or the next turn starts; an Enter with nothing typed in an idle card keeps it; an answer reported as the subagent exits still counts; `answerUnavailable` instead when status hooks are off), plus `status`, the exit state and the end of the subagent's screen as masked plain text (a headless terminal renders the redraws of a full-screen TUI such as OpenCode, so no control sequences remain; `wait_for_agent` returns the same). Other providers have no `answer`: read that text. `observe_agent`: the current status and the end of the screen as masked plain text, rendered like `wait_for_agent`'s `output` (at most `maxChars`). |
+| `send_to_agent`, `cancel_agent`, `list_agents` | Follow-up prompts, disposing a subagent, listing this session's subagents. Prompts to one subagent arrive in the order they were sent. OpenCode drops keys typed while it starts, so a prompt for it (also `spawn_agent`'s first one) is typed only once its input field is up: the terminal control sequence that shows its text cursor on its full screen, or a lifecycle hook only 2 s after its first one (OpenCode's plugin can report shortly before the field takes keys). This wait is used on macOS and Linux; on Windows, whose pseudo console may not pass those sequences, OpenCode gets input as soon as its PTY exists, as before. `spawn_agent` (its checks, the launch and this wait together) and `send_to_agent` wait at most 45 s, under the MCP client's tool deadline; then the call fails with the reason and the session id, and nothing is typed. A send cancelled while it waits returns at once. `send_to_agent` is refused while the subagent is `needs_approval` (any CLI's own approval or permission prompt: Claude Code, Codex, Grok, Kimi, Qwen, Hermes, OpenCode's permission and question prompts): what it typed (or its Enter) would answer the person's prompt. |
 
 ### Subagent profiles
 
@@ -28,7 +28,11 @@ OpenCode has no auto flag, so its auto profile is a per-run `OPENCODE_CONFIG_CON
 - `external_directory` is untouched: anything outside the project folder asks as before (each tool checks it first).
 - The person's own deny and ask rules for these tools (global and project `opencode.json(c)`, `OPENCODE_CONFIG`, the inline config) are appended after the auto rules, so they still win: auto never allows what the person denied or wanted to be asked about.
 
-A launch contributor that puts OpenCode on another model keeps `bash` asking, like accept-edits for the other CLIs.
+A launch contributor that puts OpenCode on another model keeps `bash` asking, like accept-edits for the other CLIs. In a subagent's card that ask goes to the person as every subagent's prompt does:
+
+- CanvasTTY's guard sees the command first. A decision service's deny fails the call with its reason before OpenCode asks; an allow from a plugin the person let allow answers OpenCode's prompt once. Neither leaves a prompt waiting.
+- With no verdict, OpenCode's own prompt stays. The card turns `needs_approval`, which puts it in Needs attention (and the notification, when it is on). Its row opens the card, where the person answers the prompt OpenCode shows for that exact call.
+- The orchestrator's `wait_for_agent` returns `needs_approval` with the prompt in `output`. It cannot answer it: `send_to_agent` is refused while the subagent waits, and a second `wait_for_agent` waits for the person's answer, then for the subagent's turn to end.
 
 The control CLI's `create --profile auto|normal|acceptEdits|plan|yolo` is the equivalent for its workers (from the person's own terminal it defaults to YOLO, which needs the person's acknowledgement for that CLI).
 
@@ -45,16 +49,17 @@ The core applies these to every way an agent can start another one: `spawn_agent
 | Profile ceiling | A subagent's profile is at most its orchestrator's, never YOLO. A saved subagent whose record says more is restored with its orchestrator's ceiling. |
 | Project folder | A subagent's `cwd` must be the folder the person chose for the agent it descends from, or a folder inside it, compared as real paths in both Unicode spellings (NFC and NFD name the same folder on macOS). `/`, the home folder or another project is refused: "A subagent works only inside this project's folder …". A relative `cwd` is taken from the orchestrator's folder. |
 | Limits | At most `orchestrationMaxDepth` levels of subagents below the agent the person started (2 by default) and `orchestrationMaxSubagents` live subagents in one orchestration, all levels together (8 by default). Only the person sets them, in Settings → Agents. Exited subagents do not count. |
-| Plugin launch options | `launchOptions` from an orchestrator reach only plugins that declared `launch.delegable: true`; everything a plugin contributes still passes the core-owned flag and configuration checks (no permission, sandbox, hook or bypass flags, no OpenCode/Kimi configuration that sets approvals). |
+| Plugin launch options | `launchOptions` from an orchestrator reach only plugins that declared `launch.delegable: true` (a plugin the spawn does not name keeps the orchestrator's own values for a subagent of the same CLI; repeating them needs no declaration; a plugin turned off since is left out, or refused when it set the model; changed plugin fields refuse with that reason); everything a plugin contributes still passes the core-owned flag and configuration checks (no permission, sandbox, hook or bypass flags, no OpenCode/Kimi configuration that sets approvals). |
 | YOLO | Enforced in the main process: only for a CLI the person acknowledged YOLO for in the launcher, never for a subagent, never from an orchestrator's control connection; a plugin with `sessions:launch` needs the same acknowledgement. |
 | Isolation | Every subagent and every agent a plugin starts runs inside agent isolation (below) when it is on; where there is no layer, a subagent runs in normal (it asks) and its card says why. |
+| Provider keys | A subagent and every card a plugin starts (a plain terminal card included) inherit only an allowlist of the person's environment (path, home, locale, terminal, temporary/XDG folders, proxy and certificate variables, the CLIs' home folders): no provider API key (`OPENAI_API_KEY`, `ANTHROPIC_API_KEY`, …), token (`GH_TOKEN`, `CLAUDE_CODE_OAUTH_TOKEN`, …), cloud credential, admin key, CanvasTTY credential (`CANVASTTY_GITHUB_TOKEN`) or OpenCode inline config, also not through the configuration CanvasTTY adds for the launch. A key it needs comes from a launch contribution's `secretEnv`, or the person lists the variable in `CANVASTTY_AGENT_PASS_ENV`. Orchestrators cannot start plain terminal cards. See [installing-and-security.md](installing-and-security.md#agent-protection-layers). |
 | No settings | The `canvastty_agents` tools and an orchestrator's control connection cannot change settings, base protection, profiles, plugin trust or isolation. |
 
 ## Agent isolation
 
 Settings → Agents → Agent isolation (on by default) wraps an agent's whole process tree (the CLI, its commands, its MCP servers and hooks) in an operating-system layer. It applies to every subagent, every agent a plugin starts, and every agent in a profile other than normal (manual), whoever launched it. Inside it:
 
-- files can be written only in the project folder (both spellings), the launch's own temporary folder (`TMPDIR`), the CLI's own state, configuration and cache folders, and the npm/bun caches; the project's git hooks and an existing repository's git config, and the CLI's own permission settings files (`~/.codex/config.toml`, `~/.claude/settings.json`, OpenCode's config) are not writable;
+- files can be written only in the project folder (both spellings), the launch's own temporary folder (`TMPDIR`), the CLI's own state, configuration and cache folders, and the npm/bun caches; the project's git hooks and an existing repository's git config, and the CLI's own permission settings files (`~/.codex/config.toml`, `~/.claude/settings.json`, Kimi's `config.toml` in `~/.kimi-code` or `~/.kimi`, OpenCode's config) are not writable;
 - SSH, cloud and package-registry credentials, other CLIs' folders (their credentials) and CanvasTTY's own tokens, secret stores, account homes and launch files are unreadable, except what this launch was handed (its own control connection, its account home);
 - no other process can be signalled, no app opened through Launch Services, no Apple events sent (osascript driving Terminal), no preferences written through cfprefsd, no launchd job submitted, and Unix sockets reach only DNS, syslog, the launch's own folder and CanvasTTY's token-checked gateways (not Docker, tmux or an SSH agent);
 - network and the keychain's services work as before: the CLI reaches its provider and reads its own sign-in.
@@ -127,7 +132,7 @@ Automated Codex delivery waits for the current launch's empty composer, includin
 multiline draft checks, and sends task text with explicit paste boundaries before
 Enter. It does not treat startup output or a previous launch's screen as readiness.
 
-For a visible numbered menu, `screen` returns `interaction.options`, the selected option and a screen revision. After inspecting the choices and confirming the action is within the task's authorization, use `choose SESSION_ID --choice N --revision HASH`. To close an idle menu that explicitly offers Escape, use `dismiss SESSION_ID --revision HASH`. Both commands reject a changed screen, act only on the owned PTY and return delivery receipts; inspect the next screen to confirm the outcome. They do not automatically approve anything.
+For a visible numbered menu, `screen` returns `interaction.options`, the selected option and a screen revision. After inspecting the choices and confirming the action is within the task's authorization, use `choose SESSION_ID --choice N --revision HASH`. To close an idle menu that explicitly offers Escape, use `dismiss SESSION_ID --revision HASH`. Both commands reject a changed screen, act only on the owned PTY and return delivery receipts; inspect the next screen to confirm the outcome. They do not automatically approve anything. On an orchestrator's own connection they never answer an approval or permission prompt: while the worker is `needs_approval`, or the menu is Codex's own approval prompt (run a command, apply edits, grant permissions or network access, an MCP server's request), `choose`, `dismiss` and `send` are refused with `NEEDS_PERSON` and nothing is written. The person answers it in the card (or a decision service they allowed does, before the CLI asks); the orchestrator tells them what it asks and checks `status` again. Other menus (hooks review, model and update notices) stay the orchestrator's to resolve. The person's own connection (the app-wide descriptor) is unchanged. `interrupt` still stops the orchestrator's own submitted turn: stopping is not an answer, and the asked call does not run.
 
 Codex may show **Hooks need review** on first launch or after the bundled lifecycle helpers change. Review the listed commands before trusting them through the observed menu. This is a separate startup step from selecting YOLO and is never dismissed by submitting a task. Controlled launches disable Codex's terminal animations for stable composer detection; ordinary sessions keep their configuration.
 
diff --git a/docs/installing-and-security.md b/docs/installing-and-security.md
index c81480251..35562e338 100644
--- a/docs/installing-and-security.md
+++ b/docs/installing-and-security.md
@@ -93,6 +93,8 @@ Agents work in their own mode; CanvasTTY's layers sit outside it as a quiet safe
 
 Isolation keeps approval configuration read-only, including OpenCode's project and custom config files and its `agent`/`agents` directories. CLI state and caches remain writable. A moved CLI home is refused if its real path overlaps host credentials or private CanvasTTY data; a selected account folder below `account-homes` remains available. Credential exclusions use the host environment captured before launch contributions, even if the launch changes `HOME`.
 
+**Provider keys in the environment.** A plain terminal card you start keeps your whole environment, except CanvasTTY's own credentials (`CANVASTTY_GITHUB_TOKEN`, any `CANVASTTY_*TOKEN`/`*KEY`/`*SECRET`/`*PASSWORD`), which only the app uses. An agent card never inherits those nor organization-level admin credentials (`OPENAI_ADMIN_KEY`, `ANTHROPIC_ADMIN_KEY`, any `*_ADMIN_*KEY` or `*_ADMIN_*TOKEN`), whoever starts it. An agent you start yourself otherwise keeps your environment, ordinary provider keys such as `OPENAI_API_KEY` or `ANTHROPIC_API_KEY` included, in every profile, so a CLI signed in by API key keeps working (an exported `ANTHROPIC_API_KEY` still takes precedence over Claude Code's subscription sign-in). Subagents and every card a plugin starts (a plain terminal card included, since the plugin drives it) inherit only what a process needs: `PATH`, `HOME`, user, shell, locale and terminal variables, temporary and XDG folders, the Windows system folders, proxy variables (`HTTP(S)_PROXY`, `ALL_PROXY`, `NO_PROXY`), certificate settings (`SSL_CERT_FILE`, `SSL_CERT_DIR`, `NODE_EXTRA_CA_CERTS`, `REQUESTS_CA_BUNDLE`, `CURL_CA_BUNDLE`), `EDITOR`/`VISUAL`/`PAGER` and the CLIs' own home folders (`CODEX_HOME`, `CLAUDE_CONFIG_DIR`, …). No provider key, token, cloud credential, `SSH_AUTH_SOCK` or OpenCode inline config (`OPENCODE_CONFIG_CONTENT`, which can hold keys) reaches them, also not through the configuration CanvasTTY adds for the launch; a key they need comes explicitly, through a launch contribution's `secretEnv`. To pass a variable anyway, list it in `CANVASTTY_AGENT_PASS_ENV` (comma-separated, e.g. `CANVASTTY_AGENT_PASS_ENV=OPENAI_ADMIN_KEY`) in the environment CanvasTTY starts with. This is about inherited variables only: without the isolation layer an agent runs as you and can still read a key from your shell profile file, and a plain terminal card runs your login shell, which reads that file.
+
 OpenCode Auto inspects only regular local config files of at most 1 MiB. An existing source that cannot be inspected adds no Auto allowances. A launch contributor's uninspectable config is refused; only absent optional directory candidates are skipped. JSONC comments and trailing commas are parsed without changing quoted strings.
 
 **Git audit.** When an isolated agent session ends, is closed or is restored after a quit, CanvasTTY checks the repositories under its folder whose git folder changed. If git would now run something outside isolation (a `core.hooksPath`, a filter or diff driver, `fsmonitor`, a hook file, `info/attributes`), a notice lists exactly what changed; **Neutralize** removes those keys and disables those files, **Keep as is** leaves them. It does not undo other file changes inside the project.
diff --git a/docs/installing-and-security.ru.md b/docs/installing-and-security.ru.md
index 0532080dd..c18a7cf80 100644
--- a/docs/installing-and-security.ru.md
+++ b/docs/installing-and-security.ru.md
@@ -89,6 +89,8 @@ gh variable set SPARKLE_PUBLIC_ED_KEY --repo howdeploy/CanvasTTY \
 
 **Аудит git.** Когда изолированная сессия агента заканчивается, закрывается или восстанавливается после выхода, CanvasTTY проверяет репозитории в её папке, у которых изменилась папка git. Если git теперь запустит что-то вне изоляции (`core.hooksPath`, фильтр или diff-драйвер, `fsmonitor`, файл хука, `info/attributes`), появляется уведомление со списком изменений; **Обезвредить** удаляет эти ключи и отключает файлы, **Оставить как есть** ничего не трогает. Другие изменения файлов в проекте аудит не откатывает.
 
+**Ключи провайдеров в окружении.** Обычный терминал, запущенный вами, получает всё ваше окружение, кроме собственных учётных данных CanvasTTY (`CANVASTTY_GITHUB_TOKEN`, любые `CANVASTTY_*TOKEN`/`*KEY`/`*SECRET`/`*PASSWORD`), которые нужны только приложению. Карточка агента никогда не наследует ни их, ни админ-ключи и админ-токены организации (`OPENAI_ADMIN_KEY`, `ANTHROPIC_ADMIN_KEY`, любые `*_ADMIN_*KEY` и `*_ADMIN_*TOKEN`). В остальном агент, запущенный вами, получает ваше окружение, включая обычные ключи вроде `OPENAI_API_KEY` и `ANTHROPIC_API_KEY`, в любом режиме, так что CLI со входом по API-ключу работает как раньше. Субагенты и любые карточки, запущенные плагинами (включая обычный терминал, ведь им управляет плагин), наследуют только то, что нужно процессу: `PATH`, `HOME`, переменные пользователя, оболочки, локали и терминала, временные и XDG-папки, системные папки Windows, прокси (`HTTP(S)_PROXY`, `ALL_PROXY`, `NO_PROXY`), настройки сертификатов (`SSL_CERT_FILE`, `SSL_CERT_DIR`, `NODE_EXTRA_CA_CERTS`, `REQUESTS_CA_BUNDLE`, `CURL_CA_BUNDLE`), `EDITOR`/`VISUAL`/`PAGER` и домашние папки CLI (`CODEX_HOME`, `CLAUDE_CONFIG_DIR`, …). Ни ключей провайдеров, ни токенов, ни облачных учётных данных, ни `SSH_AUTH_SOCK`, ни встроенной конфигурации OpenCode (`OPENCODE_CONFIG_CONTENT`, в ней бывают ключи) они не получают, в том числе через конфигурацию, которую CanvasTTY добавляет к запуску; нужный ключ передаётся явно, через `secretEnv` вклада в запуск. Чтобы всё же передать переменную, перечислите её в `CANVASTTY_AGENT_PASS_ENV` (через запятую) в окружении, с которым запускается CanvasTTY. Речь только о наследуемых переменных: без слоя изоляции агент работает от вашего имени и может прочитать ключ из файла профиля оболочки, а обычный терминал запускает вашу login-оболочку, которая этот файл читает.
+
 ### Linux: если bubblewrap не запускается
 
 Ubuntu 24.04 и новее (и другие дистрибутивы с `kernel.apparmor_restrict_unprivileged_userns=1`) разрешают создавать непривилегированные пространства имён пользователя только программам с профилем AppArmor, а bubblewrap без них не работает. CanvasTTY проверяет это один раз (после неудачи — снова через минуту) и, если `bwrap` установлен, но не запускается, ведёт себя как на компьютере без слоя изоляции: субагенты и агенты от плагинов работают вручную, а карточка объясняет почему. Чтобы разрешить, дайте bubblewrap собственный профиль (рекомендуется, касается только `bwrap`):
diff --git a/docs/installing-and-security.zh-CN.md b/docs/installing-and-security.zh-CN.md
index 605deda0b..b9d2d080a 100644
--- a/docs/installing-and-security.zh-CN.md
+++ b/docs/installing-and-security.zh-CN.md
@@ -89,6 +89,8 @@ gh variable set SPARKLE_PUBLIC_ED_KEY --repo howdeploy/CanvasTTY \
 
 **Git 审计。** 隔离的智能体会话结束、被关闭或在退出后恢复时，CanvasTTY 会检查其文件夹下 git 目录发生变化的仓库。如果 git 现在会在隔离外运行某些东西（`core.hooksPath`、filter 或 diff 驱动、`fsmonitor`、hook 文件、`info/attributes`），会出现一条列出具体变化的通知；**Neutralize** 删除这些键并停用这些文件，**Keep as is** 保持不变。它不会撤销项目内的其他文件改动。
 
+**环境中的服务商密钥。** 你自己启动的普通终端卡片保留你的全部环境，但不包括 CanvasTTY 自己的凭据（`CANVASTTY_GITHUB_TOKEN`，以及任何 `CANVASTTY_*TOKEN`/`*KEY`/`*SECRET`/`*PASSWORD`），它们只供应用本身使用。智能体卡片永远不会继承这些凭据，也不会继承组织级管理员凭据（`OPENAI_ADMIN_KEY`、`ANTHROPIC_ADMIN_KEY`，以及任何 `*_ADMIN_*KEY` 或 `*_ADMIN_*TOKEN`）。除此之外，你自己启动的智能体在任何模式下都保留你的环境，包括 `OPENAI_API_KEY`、`ANTHROPIC_API_KEY` 等普通密钥，因此用 API 密钥登录的 CLI 照常工作。子智能体以及插件启动的任何卡片（包括普通终端卡片，因为由插件操控）只继承进程所需的变量：`PATH`、`HOME`、用户、shell、语言区域和终端变量、临时目录和 XDG 目录、Windows 系统目录、代理变量（`HTTP(S)_PROXY`、`ALL_PROXY`、`NO_PROXY`）、证书设置（`SSL_CERT_FILE`、`SSL_CERT_DIR`、`NODE_EXTRA_CA_CERTS`、`REQUESTS_CA_BUNDLE`、`CURL_CA_BUNDLE`）、`EDITOR`/`VISUAL`/`PAGER` 以及各 CLI 自己的主目录（`CODEX_HOME`、`CLAUDE_CONFIG_DIR` 等）。它们拿不到任何服务商密钥、令牌、云凭据、`SSH_AUTH_SOCK` 或 OpenCode 内联配置（`OPENCODE_CONFIG_CONTENT`，其中可能含有密钥），通过 CanvasTTY 为启动添加的配置也不会；需要的密钥通过启动贡献的 `secretEnv` 显式传入。如仍要传递某个变量，请在 CanvasTTY 启动时的环境中把它列入 `CANVASTTY_AGENT_PASS_ENV`（逗号分隔）。这里只涉及继承的变量：没有隔离层时，智能体以你的身份运行，仍可从 shell 配置文件中读取密钥；普通终端卡片运行你的登录 shell，也会读取该文件。
+
 ### Linux：bubblewrap 无法启动时
 
 Ubuntu 24.04 及更新版本（以及其他设置了 `kernel.apparmor_restrict_unprivileged_userns=1` 的发行版）只允许带 AppArmor 配置文件的程序创建非特权用户命名空间，而 bubblewrap 需要它。CanvasTTY 会检查一次（失败后一分钟再查），如果 `bwrap` 已安装却无法启动，就按没有隔离层的电脑处理：子智能体和插件启动的智能体以手动模式运行，卡片会说明原因。要允许它，可以为 bubblewrap 单独添加配置文件（推荐，只影响 `bwrap`）：
diff --git a/docs/plugin-api.d.ts b/docs/plugin-api.d.ts
index 8ab1b19a3..65c14bd51 100644
--- a/docs/plugin-api.d.ts
+++ b/docs/plugin-api.d.ts
@@ -287,7 +287,11 @@ export interface CanvasTTYDecisionRequest {
   /** The agent's current folder, when its CLI reports it. */
   agentCwd: string | null;
   tool: { name: string; kind: "shell" | "edit" | "other"; command: string | null; paths: string[] };
-  /** The tool input as the agent sent it; null when it was over 40 KB (then `truncated`). */
+  /**
+   * The tool input as the agent sent it; null when it was over 40 KB (then `truncated`). OpenCode's tools come in Claude
+   * Code's names and shapes: bash {command, workdir}, Write {file_path, content}, Edit {file_path, old_string, new_string,
+   * replace_all}, MultiEdit {file_path, edits: [{old_string, new_string, replace_all}]}, apply_patch {patch}.
+   */
   input: unknown;
   truncated: boolean;
   /** How long the host waits for this answer (the service's `timeoutMs`). */
diff --git a/docs/plugins.md b/docs/plugins.md
index 0047ad17b..54aace18a 100644
--- a/docs/plugins.md
+++ b/docs/plugins.md
@@ -196,7 +196,7 @@ Up to 8 fields; `kind` is `boolean`, `select` (1–16 options) or `text` (at mos
 
 A `select` with `"optionsFrom": "service"` also lists choices the service offers, such as its own accounts. When the launcher opens, CanvasTTY asks the service `canvastty.launch.options` `{ provider, fields: [keys] }` and waits at most 3 s; the answer `{ "<key>": [{ value, label }] }` adds up to 64 choices per field after the declared ones (which stay required and are all the launcher shows when the service does not answer). Because such a list can change after a card was saved, its value is accepted as any text up to 200 characters without control characters, and `canvastty.launch.prepare` must check it and refuse a value it no longer knows.
 
-Orchestrators pass the same values to `spawn_agent` as `launchOptions` (`{ "<pluginId>": { "<key>": value } }`), checked exactly like the launcher's, but only to a plugin whose `launch` declares `"delegable": true` (its options are safe for an agent to choose; without it only the person chooses them, and `list_providers` does not list them); a plugin tool can hand them out (for example the account it picked). Whatever a contribution adds is checked the same way for every launch: no permission, sandbox, hook or bypass flags, and no configuration the CLI reads that sets approvals (an OpenCode `OPENCODE_CONFIG` file or `OPENCODE_CONFIG_DIR`, `OPENCODE_PERMISSION`, Kimi's `--config`/`--config-file`); a file CanvasTTY cannot read and check is refused. While a child's launch waits for its plugins (launch options, a launch policy, an environment), `spawn_agent` answers only after its `prompt` reached the started agent, and `send_to_agent` waits the same way. A refused, failed or cancelled launch fails the call with the reason and the session id (the card stays); the text is dropped, never kept for a later restart. The control CLI answers `NOT_READY` for such a card.
+Orchestrators pass the same values to `spawn_agent` as `launchOptions` (`{ "<pluginId>": { "<key>": value } }`), checked exactly like the launcher's, but only to a plugin whose `launch` declares `"delegable": true` (its options are safe for an agent to choose; without it only the person chooses them, and `list_providers` does not list them); a plugin tool can hand them out (for example the account it picked). A plugin the spawn does not name keeps the orchestrator's own values for a subagent of the same CLI (the person chose them; this needs no `delegable`), so a subagent never falls back silently to its CLI's default model; a subagent of another CLI is refused while a plugin's contribution sets the orchestrator's model (its CLI model argument or `thirdPartyModel`) unless the spawn names that plugin, and `spawn_agent` answers with `servedBy` (CLI, requested model, and whether each plugin's options were inherited or explicit). An inherited plugin that was turned off since the orchestrator started is left out, unless it set the model (then the spawn is refused); one whose fields changed refuses the spawn with that reason. For a subagent's launch, a contribution that chooses the model (`--model`/`--model=`, `-m` for OpenCode, Codex, Qwen, Kimi and Grok, and for Codex also `--profile`/`-p` and `-c`/`--config` setting `model`, `model_provider` or `profile`) refuses a launch that also asks for a model, and two contributions that both set the model are refused, instead of starting the CLI with two. A person's own launches and restores start as before. Whatever a contribution adds is checked the same way for every launch: no permission, sandbox, hook or bypass flags, and no configuration the CLI reads that sets approvals (an OpenCode `OPENCODE_CONFIG` file or `OPENCODE_CONFIG_DIR`, `OPENCODE_PERMISSION`, Kimi's `--config`/`--config-file`); a file CanvasTTY cannot read and check is refused. While a child's launch waits for its plugins (launch options, a launch policy, an environment), `spawn_agent` answers only after its `prompt` reached the started agent, and `send_to_agent` waits the same way. A refused, failed or cancelled launch fails the call with the reason and the session id (the card stays); the text is dropped, never kept for a later restart. The control CLI answers `NOT_READY` for such a card.
 
 Before the agent starts, the host sends the service a `canvastty.launch.prepare` request, which surfaces cannot send:
 
@@ -286,7 +286,7 @@ Before a local agent's shell command or file write runs, CanvasTTY can ask a plu
 }]
 ```
 
-`pre-tool` is every shell and file-writing tool call, before it runs and in every permission mode, YOLO included: Claude Code, Codex and Qwen Code through their `PreToolUse` hook, OpenCode through CanvasTTY's OpenCode plugin (`tool.execute.before`). `appliesTo` limits the agents; all four when omitted. The host sends `canvastty.decide` (host-only) and waits at most `timeoutMs` (1000 to 60000; 3000 when omitted). The agent's call waits as long, so ask for more only when the answer needs it (for example a local model reading the command); CanvasTTY sizes each card's hook for the longest budget of the services that apply when the card starts, and a service trusted later gets no more than its card allows. The request carries the budget as `budgetMs`:
+`pre-tool` is every shell and file-writing tool call, before it runs and in every permission mode, YOLO included: Claude Code, Codex and Qwen Code through their `PreToolUse` hook, OpenCode through CanvasTTY's OpenCode plugin (`tool.execute.before`). OpenCode's tools arrive under Claude Code's names and input shapes: `bash` `{command, workdir}`, `Write` `{file_path, content}` (the whole new file), `Edit` `{file_path, old_string, new_string, replace_all}`, `MultiEdit` `{file_path, edits: [{old_string, new_string, replace_all}]}` and `apply_patch` `{patch}`. `appliesTo` limits the agents; all four when omitted. The host sends `canvastty.decide` (host-only) and waits at most `timeoutMs` (1000 to 60000; 3000 when omitted). The agent's call waits as long, so ask for more only when the answer needs it (for example a local model reading the command); CanvasTTY sizes each card's hook for the longest budget of the services that apply when the card starts, and a service trusted later gets no more than its card allows. The request carries the budget as `budgetMs`:
 
 ```ts
 interface DecisionRequest {
diff --git a/docs/plugins.ru.md b/docs/plugins.ru.md
index 5d2cfe9d1..ac4965b95 100644
--- a/docs/plugins.ru.md
+++ b/docs/plugins.ru.md
@@ -264,7 +264,7 @@ CanvasTTY владеет окном, PTY, сохранённой записью
 }]
 ```
 
-`pre-tool` — каждый вызов shell-инструмента и записи файла до выполнения, в любом режиме разрешений, включая YOLO: у Claude Code, Codex и Qwen Code через их хук `PreToolUse`, у OpenCode через плагин CanvasTTY для OpenCode (`tool.execute.before`). `appliesTo` ограничивает агентов; без него — все четыре. Хост отправляет `canvastty.decide` (только хост) и ждёт не больше `timeoutMs` (от 1000 до 60000; без него 3000). Вызов агента ждёт столько же, поэтому просите больше, только когда ответу это нужно (например, локальная модель читает команду); CanvasTTY настраивает хук каждого окна под самый долгий бюджет сервисов, которые к нему относятся на момент запуска, а сервис, которому доверились позже, получает не больше, чем позволяет окно. Бюджет приходит в запросе как `budgetMs`:
+`pre-tool` — каждый вызов shell-инструмента и записи файла до выполнения, в любом режиме разрешений, включая YOLO: у Claude Code, Codex и Qwen Code через их хук `PreToolUse`, у OpenCode через плагин CanvasTTY для OpenCode (`tool.execute.before`). Инструменты OpenCode приходят с именами и формой ввода Claude Code: `bash` `{command, workdir}`, `Write` `{file_path, content}` (весь новый файл), `Edit` `{file_path, old_string, new_string, replace_all}`, `MultiEdit` `{file_path, edits: [{old_string, new_string, replace_all}]}` и `apply_patch` `{patch}`. `appliesTo` ограничивает агентов; без него — все четыре. Хост отправляет `canvastty.decide` (только хост) и ждёт не больше `timeoutMs` (от 1000 до 60000; без него 3000). Вызов агента ждёт столько же, поэтому просите больше, только когда ответу это нужно (например, локальная модель читает команду); CanvasTTY настраивает хук каждого окна под самый долгий бюджет сервисов, которые к нему относятся на момент запуска, а сервис, которому доверились позже, получает не больше, чем позволяет окно. Бюджет приходит в запросе как `budgetMs`:
 
 ```ts
 interface DecisionRequest {
diff --git a/docs/plugins.zh-CN.md b/docs/plugins.zh-CN.md
index c91ff2c60..0f2e185cf 100644
--- a/docs/plugins.zh-CN.md
+++ b/docs/plugins.zh-CN.md
@@ -262,7 +262,7 @@ CanvasTTY 负责卡片、PTY、保存的记录和恢复顺序；服务回答五
 }]
 ```
 
-`pre-tool` 指每一次 shell 和写文件的工具调用，在运行之前，并且适用于所有权限模式（包括 YOLO）：Claude Code、Codex 和 Qwen Code 通过它们的 `PreToolUse` hook，OpenCode 通过 CanvasTTY 的 OpenCode 插件（`tool.execute.before`）。`appliesTo` 限定 agent；省略时为全部四个。主机发送 `canvastty.decide`（仅主机可发），最多等待 `timeoutMs`（1000 到 60000；省略时 3000）。agent 的调用会等待同样长的时间，所以只有在回答确实需要时才申请更长时间（例如让本地模型阅读命令）；CanvasTTY 在卡片启动时按适用服务中最长的预算设置该卡片的 hook，之后才被信任的服务所得时间不超过卡片允许的范围。请求以 `budgetMs` 携带该预算：
+`pre-tool` 指每一次 shell 和写文件的工具调用，在运行之前，并且适用于所有权限模式（包括 YOLO）：Claude Code、Codex 和 Qwen Code 通过它们的 `PreToolUse` hook，OpenCode 通过 CanvasTTY 的 OpenCode 插件（`tool.execute.before`）。OpenCode 的工具以 Claude Code 的名称和输入形状到达：`bash` `{command, workdir}`、`Write` `{file_path, content}`（整个新文件）、`Edit` `{file_path, old_string, new_string, replace_all}`、`MultiEdit` `{file_path, edits: [{old_string, new_string, replace_all}]}` 和 `apply_patch` `{patch}`。`appliesTo` 限定 agent；省略时为全部四个。主机发送 `canvastty.decide`（仅主机可发），最多等待 `timeoutMs`（1000 到 60000；省略时 3000）。agent 的调用会等待同样长的时间，所以只有在回答确实需要时才申请更长时间（例如让本地模型阅读命令）；CanvasTTY 在卡片启动时按适用服务中最长的预算设置该卡片的 hook，之后才被信任的服务所得时间不超过卡片允许的范围。请求以 `budgetMs` 携带该预算：
 
 ```ts
 interface DecisionRequest {
diff --git a/native/canvastty-helper/catalog.json b/native/canvastty-helper/catalog.json
index d3ec39a27..a02ac59ed 100644
--- a/native/canvastty-helper/catalog.json
+++ b/native/canvastty-helper/catalog.json
@@ -964,7 +964,7 @@
       },
       {
         "name": "spawn_agent",
-        "description": "Launch another provider's agent as a CanvasTTY subagent of this session and optionally deliver a first prompt. Returns the new session id. provider must be an id from list_providers (known ids: codex, claude, qwen, kimi, opencode, hermes, grok, omp, pi, cursor, minimax, devin, antigravity); call list_providers first to see which are installed and signed in. Give each subagent one self-contained part of the task and an absolute cwd. If the person names a model, pass it as model in the format list_providers gives for that provider (OpenCode: provider/model); effort sets the reasoning effort where that CLI has one (list_providers shows its efforts). An unsupported model or effort is refused with the reason. profile: without it the subagent gets this session's launch profile, or the next lower one its CLI has; a subagent never gets more than this session (plan < normal < acceptEdits < auto) and never YOLO. \"auto\" lets it work without asking the person for each edit or command inside the project; CanvasTTY's outer layers (agent isolation, base protection) still stop writes outside the project, secrets and system changes. The answer says the profile it got. cwd must be this project's folder or a folder inside it. launchOptions passes plugin launch options exactly as a plugin tool gives them (for example the account a plugin picked), for plugins that allow an orchestrator to choose them. Refusals say why (a limit the person set, a folder outside the project, a profile above this session's): change the request instead of retrying it. Then call wait_for_agent and get_agent_result.",
+        "description": "Launch another provider's agent as a CanvasTTY subagent of this session and optionally deliver a first prompt. Returns the new session id. provider must be an id from list_providers (known ids: codex, claude, qwen, kimi, opencode, hermes, grok, omp, pi, cursor, minimax, devin, antigravity); call list_providers first to see which are installed and signed in. Give each subagent one self-contained part of the task and an absolute cwd. If the person names a model, pass it as model in the format list_providers gives for that provider (OpenCode: provider/model); effort sets the reasoning effort where that CLI has one (list_providers shows its efforts). An unsupported model or effort is refused with the reason. profile: without it the subagent gets this session's launch profile, or the next lower one its CLI has; a subagent never gets more than this session (plan < normal < acceptEdits < auto) and never YOLO. \"auto\" lets it work without asking the person for each edit or command inside the project; CanvasTTY's outer layers (agent isolation, base protection) still stop writes outside the project, secrets and system changes. The answer says the profile it got. cwd must be this project's folder or a folder inside it. launchOptions passes plugin launch options exactly as a plugin tool gives them (for example the account a plugin picked), for plugins that allow an orchestrator to choose them. A plugin this call does not name keeps this session's own launch options (for example its model account), so the subagent never falls back silently to its CLI's default model; a subagent of another CLI is refused while this session's model comes from a plugin's options (spawn one of this CLI instead), and model next to inherited options that set the model is refused. The answer's servedBy says which CLI and requested model serve it and where each plugin's options came from. Refusals say why (a limit the person set, a folder outside the project, a profile above this session's): change the request instead of retrying it. Then call wait_for_agent and get_agent_result.",
         "inputSchema": {
           "type": "object",
           "properties": {
@@ -1045,7 +1045,7 @@
       },
       {
         "name": "send_to_agent",
-        "description": "Write a prompt into one of this session's subagents. Plain terminal sessions are not agents.",
+        "description": "Write a prompt into one of this session's subagents. Plain terminal sessions are not agents. Refused while the subagent is needs_approval: that prompt is the person's to answer.",
         "inputSchema": {
           "type": "object",
           "properties": {
@@ -1072,7 +1072,7 @@
       },
       {
         "name": "observe_agent",
-        "description": "Read the capped terminal tail and status of one of this session's subagents.",
+        "description": "Read the status of one of this session's subagents and the end of its screen as masked plain text (once its process exited, also exitLines).",
         "inputSchema": {
           "type": "object",
           "properties": {
@@ -1095,7 +1095,7 @@
       },
       {
         "name": "wait_for_agent",
-        "description": "Wait until one of this session's subagents stops working, instead of polling observe_agent or get_agent_result; nothing is sent to it while it waits. Returns reason \"idle\" (the turn that answers your latest prompt ended and it waits for input; an idle before that turn started does not count), \"needs_approval\" (its card shows a prompt only the person may answer; never answer it yourself), \"done\" or \"failed\" (its process exited), \"quiet\" (it reports no status or no turn start and its screen stopped changing, so judge from output), \"closed\" (its card was closed) or \"timeout\" after timeoutSeconds (default 55, at most 600), with status, exitCode, waitedMs, the masked terminal tail as output, and answer (the final reply of the turn that ended, for Codex and OpenCode subagents); when the subagent's process exited (for example at once, on a model its CLI does not know), exitLines holds the last lines of its screen as plain text, which say why. After a timeout, call it again. Then read get_agent_result.",
+        "description": "Wait until one of this session's subagents stops working, instead of polling observe_agent or get_agent_result; nothing is sent to it while it waits. Returns reason \"idle\" (the turn that answers your latest prompt ended and it waits for input; an idle before that turn started does not count), \"needs_approval\" (its card shows a prompt only the person may answer, such as a shell command it wants to run, shown in output; never answer it yourself: tell the person what it asks, then call wait_for_agent again, which waits for their answer instead of returning the same prompt again), \"done\" or \"failed\" (its process exited), \"quiet\" (it reports no status or no turn start and its screen stopped changing, so judge from output), \"closed\" (its card was closed) or \"timeout\" after timeoutSeconds (default 50, at most 100; unless this session is Claude Code, a call returns after at most 50 seconds, before its MCP client would end the call), with status, exitCode, waitedMs, the end of its screen as masked plain text as output, and answer (the final reply of the turn that ended, for Codex and OpenCode subagents; when CanvasTTY status hooks are off it cannot be captured and answerUnavailable says so instead: then read output); when the subagent's process exited (for example at once, on a model its CLI does not know), exitLines holds the last lines of its screen as plain text, which say why. A timeout means it is still running: read output as its progress and call wait_for_agent again. Then read get_agent_result.",
         "inputSchema": {
           "type": "object",
           "properties": {
@@ -1107,7 +1107,7 @@
             "timeoutSeconds": {
               "type": "integer",
               "minimum": 1,
-              "maximum": 600
+              "maximum": 100
             }
           },
           "required": [
@@ -1118,7 +1118,7 @@
       },
       {
         "name": "get_agent_result",
-        "description": "Get one of this session's subagents' result: answer (its last turn's final reply as the agent reported it, for Codex and OpenCode subagents; truncated:true when only the end was kept), status (idle once its turn ended), the exit state (running | done | failed; an interactive CLI stays running after a task) and the masked terminal tail as output. Prefer answer; the terminal tail is raw screen output.",
+        "description": "Get one of this session's subagents' result: answer (its last turn's final reply as the agent reported it, for Codex and OpenCode subagents; truncated:true when only the end was kept), status (idle once its turn ended), the exit state (running | done | failed; an interactive CLI stays running after a task) and the end of its screen as masked plain text as output (what the terminal shows, without control sequences). Prefer answer. answerUnavailable (reason status_hooks_off) means no answer can be captured because CanvasTTY status hooks are off: read output instead.",
         "inputSchema": {
           "type": "object",
           "properties": {
@@ -1178,6 +1178,6 @@
     ],
     "maxPluginToolNameLength": 64,
     "maxLaunchOptionsBytes": 16384,
-    "defaultAgentWaitSeconds": 55
+    "defaultAgentWaitSeconds": 50
   }
 }
diff --git a/scripts/canvastty-control.mjs b/scripts/canvastty-control.mjs
index 78e3762e9..50050ae92 100644
--- a/scripts/canvastty-control.mjs
+++ b/scripts/canvastty-control.mjs
@@ -155,8 +155,10 @@ create --cwd <directory> [--provider <id>] [--title <name>] [--yolo | --profile
   --model: the CLI's own --model for this worker (OpenCode: provider/model); if
   the person names a model, pass it. --effort: codex, claude and grok only.
   providers shows each provider's model format and effort levels.
-  Only Codex workers report a captured result and expose startup/approval menus
-  to choose/dismiss; the create response lists each worker's capabilities.
+  Only Codex workers report a captured result and expose startup menus to
+  choose/dismiss; the create response lists each worker's capabilities. From an
+  Orchestrator session, choose, dismiss and send are refused (NEEDS_PERSON) while
+  a worker waits at an approval or permission prompt: only the person answers it.
 list
 status <session-id>
 screen <session-id>
diff --git a/src/agent-browser/orchestration-catalog.d.mts b/src/agent-browser/orchestration-catalog.d.mts
index 48685eb78..1af1e815c 100644
--- a/src/agent-browser/orchestration-catalog.d.mts
+++ b/src/agent-browser/orchestration-catalog.d.mts
@@ -3,6 +3,7 @@ export const MAX_ORCHESTRATION_PAYLOAD_BYTES: number;
 export const AGENT_PROVIDER_IDS: readonly string[];
 export const MAX_AGENT_WAIT_SECONDS: number;
 export const DEFAULT_AGENT_WAIT_SECONDS: number;
+export const CLIENT_SAFE_AGENT_WAIT_SECONDS: number;
 export const REASONING_EFFORT_IDS: readonly string[];
 export function unknownProviderMessage(value: unknown): string;
 
diff --git a/src/agent-browser/orchestration-catalog.mjs b/src/agent-browser/orchestration-catalog.mjs
index ad7b53d22..7233b08fc 100644
--- a/src/agent-browser/orchestration-catalog.mjs
+++ b/src/agent-browser/orchestration-catalog.mjs
@@ -28,10 +28,13 @@ const prompt = string({ minLength: 1, maxLength: 65_536 });
 export const AGENT_PROVIDER_IDS = Object.freeze([
   "codex", "claude", "qwen", "kimi", "opencode", "hermes", "grok", "omp", "pi", "cursor", "minimax", "devin", "antigravity"
 ]);
-/** wait_for_agent: the longest wait one call may ask for, and the wait without timeoutSeconds (some MCP clients
- *  end a tool call after 60 seconds). */
-export const MAX_AGENT_WAIT_SECONDS = 600;
-export const DEFAULT_AGENT_WAIT_SECONDS = 55;
+/** wait_for_agent: the longest wait one call may ask for, and the wait without timeoutSeconds. Most MCP clients end a
+ *  tool call after 60 seconds (Codex's default tool_timeout_sec, OpenCode's client), so a call from any orchestrator
+ *  but Claude Code (whose client waits far longer) returns after at most CLIENT_SAFE_AGENT_WAIT_SECONDS: a "timeout"
+ *  that says the subagent still runs, never a call its client cut off. */
+export const MAX_AGENT_WAIT_SECONDS = 100;
+export const DEFAULT_AGENT_WAIT_SECONDS = 50;
+export const CLIENT_SAFE_AGENT_WAIT_SECONDS = 50;
 /** spawn_agent.effort: every level some CLI takes (src/shared/launchModel.ts REASONING_EFFORTS; a test keeps them equal). */
 export const REASONING_EFFORT_IDS = Object.freeze(["minimal", "low", "medium", "high", "xhigh", "max"]);
 const provider = string({ minLength: 1, maxLength: 32, enum: [...AGENT_PROVIDER_IDS] });
@@ -58,7 +61,7 @@ export const ORCHESTRATION_TOOL_DEFINITIONS = Object.freeze([
   ),
   tool(
     "spawn_agent",
-    `Launch another provider's agent as a CanvasTTY subagent of this session and optionally deliver a first prompt. Returns the new session id. provider must be an id from list_providers (known ids: ${AGENT_PROVIDER_IDS.join(", ")}); call list_providers first to see which are installed and signed in. Give each subagent one self-contained part of the task and an absolute cwd. If the person names a model, pass it as model in the format list_providers gives for that provider (OpenCode: provider/model); effort sets the reasoning effort where that CLI has one (list_providers shows its efforts). An unsupported model or effort is refused with the reason. profile: without it the subagent gets this session's launch profile, or the next lower one its CLI has; a subagent never gets more than this session (plan < normal < acceptEdits < auto) and never YOLO. "auto" lets it work without asking the person for each edit or command inside the project; CanvasTTY's outer layers (agent isolation, base protection) still stop writes outside the project, secrets and system changes. The answer says the profile it got. cwd must be this project's folder or a folder inside it. launchOptions passes plugin launch options exactly as a plugin tool gives them (for example the account a plugin picked), for plugins that allow an orchestrator to choose them. Refusals say why (a limit the person set, a folder outside the project, a profile above this session's): change the request instead of retrying it. Then call wait_for_agent and get_agent_result.`,
+    `Launch another provider's agent as a CanvasTTY subagent of this session and optionally deliver a first prompt. Returns the new session id. provider must be an id from list_providers (known ids: ${AGENT_PROVIDER_IDS.join(", ")}); call list_providers first to see which are installed and signed in. Give each subagent one self-contained part of the task and an absolute cwd. If the person names a model, pass it as model in the format list_providers gives for that provider (OpenCode: provider/model); effort sets the reasoning effort where that CLI has one (list_providers shows its efforts). An unsupported model or effort is refused with the reason. profile: without it the subagent gets this session's launch profile, or the next lower one its CLI has; a subagent never gets more than this session (plan < normal < acceptEdits < auto) and never YOLO. "auto" lets it work without asking the person for each edit or command inside the project; CanvasTTY's outer layers (agent isolation, base protection) still stop writes outside the project, secrets and system changes. The answer says the profile it got. cwd must be this project's folder or a folder inside it. launchOptions passes plugin launch options exactly as a plugin tool gives them (for example the account a plugin picked), for plugins that allow an orchestrator to choose them. A plugin this call does not name keeps this session's own launch options (for example its model account), so the subagent never falls back silently to its CLI's default model; a subagent of another CLI is refused while this session's model comes from a plugin's options (spawn one of this CLI instead), and model next to inherited options that set the model is refused. The answer's servedBy says which CLI and requested model serve it and where each plugin's options came from. Refusals say why (a limit the person set, a folder outside the project, a profile above this session's): change the request instead of retrying it. Then call wait_for_agent and get_agent_result.`,
     {
       provider,
       cwd: string({ minLength: 1, maxLength: 4_096 }),
@@ -73,25 +76,25 @@ export const ORCHESTRATION_TOOL_DEFINITIONS = Object.freeze([
   ),
   tool(
     "send_to_agent",
-    "Write a prompt into one of this session's subagents. Plain terminal sessions are not agents.",
+    "Write a prompt into one of this session's subagents. Plain terminal sessions are not agents. Refused while the subagent is needs_approval: that prompt is the person's to answer.",
     { sessionId, prompt, submit: boolean() },
     ["sessionId", "prompt"]
   ),
   tool(
     "observe_agent",
-    "Read the capped terminal tail and status of one of this session's subagents.",
+    "Read the status of one of this session's subagents and the end of its screen as masked plain text (once its process exited, also exitLines).",
     { sessionId, maxChars: integer({ minimum: 256, maximum: 8_192 }) },
     ["sessionId"]
   ),
   tool(
     "wait_for_agent",
-    `Wait until one of this session's subagents stops working, instead of polling observe_agent or get_agent_result; nothing is sent to it while it waits. Returns reason "idle" (the turn that answers your latest prompt ended and it waits for input; an idle before that turn started does not count), "needs_approval" (its card shows a prompt only the person may answer; never answer it yourself), "done" or "failed" (its process exited), "quiet" (it reports no status or no turn start and its screen stopped changing, so judge from output), "closed" (its card was closed) or "timeout" after timeoutSeconds (default ${DEFAULT_AGENT_WAIT_SECONDS}, at most ${MAX_AGENT_WAIT_SECONDS}), with status, exitCode, waitedMs, the masked terminal tail as output, and answer (the final reply of the turn that ended, for Codex and OpenCode subagents); when the subagent's process exited (for example at once, on a model its CLI does not know), exitLines holds the last lines of its screen as plain text, which say why. After a timeout, call it again. Then read get_agent_result.`,
+    `Wait until one of this session's subagents stops working, instead of polling observe_agent or get_agent_result; nothing is sent to it while it waits. Returns reason "idle" (the turn that answers your latest prompt ended and it waits for input; an idle before that turn started does not count), "needs_approval" (its card shows a prompt only the person may answer, such as a shell command it wants to run, shown in output; never answer it yourself: tell the person what it asks, then call wait_for_agent again, which waits for their answer instead of returning the same prompt again), "done" or "failed" (its process exited), "quiet" (it reports no status or no turn start and its screen stopped changing, so judge from output), "closed" (its card was closed) or "timeout" after timeoutSeconds (default ${DEFAULT_AGENT_WAIT_SECONDS}, at most ${MAX_AGENT_WAIT_SECONDS}; unless this session is Claude Code, a call returns after at most ${CLIENT_SAFE_AGENT_WAIT_SECONDS} seconds, before its MCP client would end the call), with status, exitCode, waitedMs, the end of its screen as masked plain text as output, and answer (the final reply of the turn that ended, for Codex and OpenCode subagents; when CanvasTTY status hooks are off it cannot be captured and answerUnavailable says so instead: then read output); when the subagent's process exited (for example at once, on a model its CLI does not know), exitLines holds the last lines of its screen as plain text, which say why. A timeout means it is still running: read output as its progress and call wait_for_agent again. Then read get_agent_result.`,
     { sessionId, timeoutSeconds: integer({ minimum: 1, maximum: MAX_AGENT_WAIT_SECONDS }) },
     ["sessionId"]
   ),
   tool(
     "get_agent_result",
-    "Get one of this session's subagents' result: answer (its last turn's final reply as the agent reported it, for Codex and OpenCode subagents; truncated:true when only the end was kept), status (idle once its turn ended), the exit state (running | done | failed; an interactive CLI stays running after a task) and the masked terminal tail as output. Prefer answer; the terminal tail is raw screen output.",
+    "Get one of this session's subagents' result: answer (its last turn's final reply as the agent reported it, for Codex and OpenCode subagents; truncated:true when only the end was kept), status (idle once its turn ended), the exit state (running | done | failed; an interactive CLI stays running after a task) and the end of its screen as masked plain text as output (what the terminal shows, without control sequences). Prefer answer. answerUnavailable (reason status_hooks_off) means no answer can be captured because CanvasTTY status hooks are off: read output instead.",
     { sessionId },
     ["sessionId"]
   ),
diff --git a/src/agent-runtime/opencode-decisions.mjs b/src/agent-runtime/opencode-decisions.mjs
index d9e9c4ca3..f3ec617e6 100644
--- a/src/agent-runtime/opencode-decisions.mjs
+++ b/src/agent-runtime/opencode-decisions.mjs
@@ -62,9 +62,13 @@ export function createOpenCodeDecisions({ client, env = process.env, send = exch
 }
 
 /**
- * An OpenCode tool call in the shape the decision hook reads (the Claude/Codex hook tool names), or null for a tool
- * that neither runs commands nor writes files. OpenCode's own names: bash {command, workdir}, write {filePath,
- * content}, edit {filePath, oldString, newString}, multiedit {filePath, edits}, apply_patch {patchText}.
+ * An OpenCode tool call in the shape the decision hook reads (Claude Code's hook tool names and input shapes), or null
+ * for a tool that neither runs commands nor writes files. OpenCode's own names: bash {command, workdir}, write
+ * {filePath, content}, edit {filePath, oldString, newString, replaceAll}, multiedit {filePath, edits: [{filePath,
+ * oldString, newString, replaceAll}]}, apply_patch {patchText}. They go out as bash {command, workdir}, Write {file_path,
+ * content} (the whole new file), Edit {file_path, old_string, new_string, replace_all} (one replacement in the file),
+ * MultiEdit {file_path, edits: [{old_string, new_string, replace_all}]} and apply_patch {patch}, so a decision service
+ * can tell a whole-file write from a partial edit and sees the text an edit replaces.
  */
 export function guardedCall(tool, args) {
   if (!tool || !args || typeof args !== "object") return null;
@@ -74,13 +78,15 @@ export function guardedCall(tool, args) {
     return { toolName: "bash", toolInput: { command: args.command, ...(workdir ? { workdir } : {}) } };
   }
   if (tool === "write" || tool === "edit" || tool === "multiedit") {
-    const filePath = typeof args.filePath === "string" && args.filePath.length > 0 ? args.filePath
-      : typeof args.file_path === "string" && args.file_path.length > 0 ? args.file_path : null;
+    const filePath = pathOf(args);
     if (!filePath) return null;
-    const content = typeof args.content === "string" ? args.content
-      : typeof args.newString === "string" ? args.newString
-        : multieditText(args.edits);
-    return { toolName: "edit", toolInput: { file_path: filePath, ...(content !== null ? { content } : {}) } };
+    if (tool === "write") {
+      return { toolName: "Write", toolInput: { file_path: filePath, ...(typeof args.content === "string" ? { content: args.content } : {}) } };
+    }
+    if (tool === "edit") return { toolName: "Edit", toolInput: { file_path: filePath, ...replacement(args) } };
+    const edits = Array.isArray(args.edits) ? args.edits.filter((edit) => edit && typeof edit === "object").map(replacement)
+      .filter((edit) => Object.keys(edit).length > 0) : [];
+    return { toolName: "MultiEdit", toolInput: { file_path: filePath, edits } };
   }
   if (tool === "apply_patch" || tool === "patch") {
     const patch = typeof args.patchText === "string" ? args.patchText : typeof args.patch === "string" ? args.patch : null;
@@ -89,11 +95,22 @@ export function guardedCall(tool, args) {
   return null;
 }
 
-/** Every multiedit edit's new text, one per line (what the batch writes), or null. */
-function multieditText(edits) {
-  if (!Array.isArray(edits)) return null;
-  const texts = edits.flatMap((edit) => edit && typeof edit === "object" && typeof edit.newString === "string" ? [edit.newString] : []);
-  return texts.length > 0 ? texts.join("\n") : null;
+function pathOf(args) {
+  return typeof args.filePath === "string" && args.filePath.length > 0 ? args.filePath
+    : typeof args.file_path === "string" && args.file_path.length > 0 ? args.file_path : null;
+}
+
+/** One OpenCode replacement {oldString, newString, replaceAll} as Claude's Edit input (fields that are not text are left out). */
+function replacement(edit) {
+  const text = (camel, snake) => typeof edit[camel] === "string" ? edit[camel] : typeof edit[snake] === "string" ? edit[snake] : null;
+  const oldString = text("oldString", "old_string");
+  const newString = text("newString", "new_string");
+  const replaceAll = typeof edit.replaceAll === "boolean" ? edit.replaceAll : typeof edit.replace_all === "boolean" ? edit.replace_all : null;
+  return {
+    ...(oldString !== null ? { old_string: oldString } : {}),
+    ...(newString !== null ? { new_string: newString } : {}),
+    ...(replaceAll !== null ? { replace_all: replaceAll } : {})
+  };
 }
 
 /** POST /permission/{requestID}/reply with `once`. False on any failure (OpenCode's prompt stays for the person). */
diff --git a/src/agent-runtime/opencode-plugin.mjs b/src/agent-runtime/opencode-plugin.mjs
index bd8151b86..69ae5f00c 100644
--- a/src/agent-runtime/opencode-plugin.mjs
+++ b/src/agent-runtime/opencode-plugin.mjs
@@ -3,10 +3,13 @@ import { CAPTURE_RESULT_ENV } from "./runtime-protocol.mjs";
 import { finalAnswer } from "./opencode-final-answer.mjs";
 import { createOpenCodeDecisions } from "./opencode-decisions.mjs";
 import { spawn } from "node:child_process";
+import { randomUUID } from "node:crypto";
 import { HOOK_TIMEOUT_MS, preparePluginHook } from "./plugin-hook-dispatch.mjs";
 
 let rootSessionId = null;
 let rootWorking = false;
+// One id per root turn (not the session id), so a late answer of an earlier turn is never taken for the current one.
+let rootTurnId = null;
 const lifecycleEnabled = process.env.CANVASTTY_LIFECYCLE_HOOKS_ENABLED !== "0";
 const pluginHookRegistry = process.env.CANVASTTY_PLUGIN_HOOK_REGISTRY ?? "";
 const pluginHookRunnerCommand = process.env.CANVASTTY_PLUGIN_HOOK_RUNNER_COMMAND ?? "";
@@ -45,6 +48,7 @@ async function lifecycleEvent(event, decisions, client) {
     if (session.parentID || session.parentId) return;
     rootSessionId = stringField(session.id, sessionId);
     rootWorking = false;
+    rootTurnId = null;
     if (!rootSessionId) return;
     if (lifecycleEnabled) {
       await reportLifecycle({ state: "idle", event: event.type, threadId: rootSessionId });
@@ -63,44 +67,55 @@ async function lifecycleEvent(event, decisions, client) {
         ? statusValue.type
         : null;
     if (status === "busy" || status === "retry") {
+      const startsTurn = status === "busy" && !rootWorking;
+      if (startsTurn || !rootTurnId) rootTurnId = randomUUID();
+      rootWorking = true;
       if (lifecycleEnabled) {
-        await reportLifecycle({ state: "working", event: `session.status:${status}`, turnId: rootSessionId });
+        await reportLifecycle({ state: "working", event: `session.status:${status}`, turnId: rootTurnId });
       }
-      if (status === "busy" && !rootWorking) {
+      if (startsTurn) {
         runPluginHooks("prompt-submit", `session.status:${status}`, event);
       }
-      rootWorking = true;
     } else if (status === "idle") {
       rootWorking = false;
-      if (lifecycleEnabled) await reportLifecycle({ state: "idle", event: "session.status:idle", turnId: rootSessionId });
+      // A result-capturing turn ends with session.idle, once its final answer was read: an idle reported here would
+      // let a waiting orchestrator take the turn as finished before its answer arrived.
+      if (lifecycleEnabled && !captureResult) await reportLifecycle({ state: "idle", event: "session.status:idle", turnId: rootTurnId });
     }
     return;
   }
   if (event.type === "session.idle") {
     rootWorking = false;
-    const result = captureResult ? await finalAnswer(client, rootSessionId) : undefined;
-    if (lifecycleEnabled || result) {
-      await reportLifecycle({ state: "idle", event: event.type, turnId: rootSessionId, ...(result ? { result } : {}) });
+    const endingSessionId = rootSessionId, endingTurnId = rootTurnId;
+    // With CanvasTTY status hooks off nothing is reported, the answer included: it is never read.
+    const result = lifecycleEnabled && captureResult && endingTurnId ? await finalAnswer(client, endingSessionId) : undefined;
+    // The read is asynchronous: when the next turn started meanwhile, this answer and this idle are no longer current.
+    const superseded = endingSessionId !== rootSessionId || endingTurnId !== rootTurnId || rootWorking;
+    if (lifecycleEnabled && !superseded) {
+      await reportLifecycle({ state: "idle", event: event.type, turnId: endingTurnId, ...(result ? { result } : {}) });
     }
     runPluginHooks("stop", event.type, event);
   } else if (event.type === "permission.asked") {
-    if (lifecycleEnabled) await reportLifecycle({ state: "needs_approval", event: event.type, turnId: rootSessionId });
+    if (lifecycleEnabled) await reportLifecycle({ state: "needs_approval", event: event.type, turnId: rootTurnId });
     runPluginHooks("permission-request", event.type, event);
   } else if (event.type === "permission.replied") {
-    if (lifecycleEnabled) await reportLifecycle({ state: "working", event: event.type, turnId: rootSessionId });
+    rootWorking = true;
+    if (lifecycleEnabled) await reportLifecycle({ state: "working", event: event.type, turnId: rootTurnId });
     runPluginHooks("permission-result", event.type, event);
   } else if (event.type === "question.asked") {
-    if (lifecycleEnabled) await reportLifecycle({ state: "needs_approval", event: event.type, turnId: rootSessionId });
+    if (lifecycleEnabled) await reportLifecycle({ state: "needs_approval", event: event.type, turnId: rootTurnId });
   } else if (event.type === "question.replied" || event.type === "question.rejected") {
-    if (lifecycleEnabled) await reportLifecycle({ state: "working", event: event.type, turnId: rootSessionId });
+    rootWorking = true;
+    if (lifecycleEnabled) await reportLifecycle({ state: "working", event: event.type, turnId: rootTurnId });
   } else if (event.type === "session.error") {
     rootWorking = false;
-    if (lifecycleEnabled) await reportLifecycle({ state: "idle", event: event.type, turnId: rootSessionId });
+    if (lifecycleEnabled) await reportLifecycle({ state: "idle", event: event.type, turnId: rootTurnId });
     runPluginHooks("stop", event.type, event);
   } else if (event.type === "session.deleted") {
     runPluginHooks("session-end", event.type, event);
     rootWorking = false;
     rootSessionId = null;
+    rootTurnId = null;
   } else if (event.type === "tool.execute.after") {
     runPluginHooks("after-tool", event.type, event);
   }
diff --git a/src/main/index.ts b/src/main/index.ts
index b9832dcee..4f22b438c 100644
--- a/src/main/index.ts
+++ b/src/main/index.ts
@@ -36,6 +36,7 @@ import { AgentChatHistoryService } from "./services/AgentChatHistoryService";
 import { LimitsService, withoutAccountScope } from "./services/LimitsService";
 import { UsageHistoryService } from "./services/UsageHistoryService";
 import { LocalUsageCollector } from "./services/LocalUsageCollector";
+import { runShutdownSteps } from "./services/shutdownSteps";
 import {
   createProviderCliRegistry,
   providerCliAvailability,
@@ -91,6 +92,7 @@ import {
   RuntimeGateway
 } from "./services/agent-runtime";
 import type { RuntimeHookHelperLaunch } from "./services/agent-runtime/ProviderRuntimeLaunch";
+import { applyRuntimeSignal } from "./services/agent-runtime/applyRuntimeSignal";
 import { agentHelperLaunches } from "./services/agentHelpers";
 import {
   recoverHermesConfigurationOnStartup,
@@ -518,7 +520,8 @@ async function initializeServices(ipc: IpcRegistrar): Promise<void> {
       providerClis,
       runtimeDirectory,
       hermesHomeDirectory,
-      kimiHomeDirectory
+      kimiHomeDirectory,
+      accountHomesRoot: join(userDataPath, "plugin-data")
     });
     // Off the startup path: the first Kimi launch then finds the probe answered instead of blocking on it.
     const bridge = agentBrowserBridge;
@@ -529,15 +532,7 @@ async function initializeServices(ipc: IpcRegistrar): Promise<void> {
       runtimeDirectory: lifecycleRuntimeDirectory,
       windowsHostPath,
       onSignal: (terminalSessionId, signal) => {
-        terminalManager?.applyProviderSignal(terminalSessionId, {
-          kind: "lifecycle",
-          state: signal.state,
-          event: signal.event,
-          ...(signal.turnId ? { requestId: signal.turnId } : {}),
-          ...(signal.threadId ? { threadId: signal.threadId } : {})
-        });
-        // A subagent's final answer (Codex Stop hook, OpenCode plugin) for get_agent_result and wait_for_agent.
-        if (signal.result) terminalManager?.recordAnswer(terminalSessionId, signal.result);
+        if (terminalManager) applyRuntimeSignal(terminalManager, terminalSessionId, signal);
         agentControl?.onSignal(terminalSessionId, signal);
         if (signal.lastAssistantMessage !== undefined && signal.answerCaptureGrantExpiresAt !== undefined) {
           evenG2?.answer(
@@ -568,6 +563,8 @@ async function initializeServices(ipc: IpcRegistrar): Promise<void> {
       openCodePluginPath,
       hermesHomeDirectory,
       kimiHomeDirectory,
+      // Account homes a launch contributor starts Grok/Kimi with live in plugin data (ProviderRuntimeLaunch.ownAccountHome).
+      accountHomesRoot: join(userDataPath, "plugin-data"),
       recoverOnStart: true,
       coreHooksEnabled: settings.get().agentLifecycleHooksEnabled,
       permissionGate: helpers.permissionGate,
@@ -1275,30 +1272,36 @@ void IPC.terminalData;
 
 async function shutdownServices(): Promise<void> {
   diagnostics.record("info", "application", "shutdown.started");
-  mascotManager?.dispose();
-  agentChatHistory?.dispose();
-  if (agentControl) await Promise.allSettled([agentControl.close()]);
   if (updateTimer) clearTimeout(updateTimer);
   if (updateInterval) clearInterval(updateInterval);
   for (const request of browserRequests.values()) { clearTimeout(request.timer); request.reject(new Error("App closing")); }
   browserRequests.clear();
-  await evenG2?.close();
-  if (terminalManager) await terminalManager.shutdown();
+  // Each step runs even when an earlier one fails, so the terminals are always hung up and saved.
+  await runShutdownSteps([
+    { name: "mascots", run: () => mascotManager?.dispose() },
+    { name: "agent chat history", run: () => agentChatHistory?.dispose() },
+    { name: "agent control", run: () => agentControl?.close() },
+    { name: "Even G2", run: () => evenG2?.close() },
+    { name: "terminal processes", run: async () => { if (terminalManager) await terminalManager.shutdown(); } }
+  ]);
   // The hung-up PTYs exit while the other services close; quitting waits for them (see waitForProcessExits).
   const ptyExits = terminalManager?.waitForProcessExits().then((left) => {
     if (left > 0) console.warn(`CanvasTTY quit with ${left} terminal process(es) that did not exit after SIGKILL.`);
   });
-  // Stop the timer first, then abort in-flight limit requests so the final save is prompt.
-  const usageHistoryClosed = usageHistory?.dispose();
-  limitsService?.dispose();
-  await usageHistoryClosed;
-  if (agentGateway) await Promise.allSettled([agentGateway.close()]);
-  if (runtimeGateway) await Promise.allSettled([runtimeGateway.close()]);
-  if (browserService) await Promise.allSettled([browserService.dispose()]);
-  if (pluginServices) await Promise.allSettled([pluginServices.dispose()]);
-  if (pluginManager) await Promise.allSettled([pluginManager.dispose()]);
-  await handoffService?.dispose();
-  if (materialService) await Promise.allSettled([materialService.dispose()]);
+  // Stop the usage-history timer first, then abort in-flight limit requests so its final save is prompt.
+  let usageHistoryClosed: Promise<unknown> | undefined;
+  await runShutdownSteps([
+    { name: "usage history", run: () => { usageHistoryClosed = usageHistory?.dispose(); } },
+    { name: "limits", run: () => limitsService?.dispose() },
+    { name: "usage history save", run: () => usageHistoryClosed },
+    { name: "agent gateway", run: () => agentGateway?.close() },
+    { name: "runtime gateway", run: () => runtimeGateway?.close() },
+    { name: "browser", run: () => browserService?.dispose() },
+    { name: "plugin services", run: () => pluginServices?.dispose() },
+    { name: "plugins", run: () => pluginManager?.dispose() },
+    { name: "handoff", run: () => handoffService?.dispose() },
+    { name: "materials", run: () => materialService?.dispose() }
+  ]);
   await ptyExits;
   diagnostics.record("info", "application", "shutdown.completed");
   await diagnostics.flush();
diff --git a/src/main/ipc/registerIpc.ts b/src/main/ipc/registerIpc.ts
index 00d3feb6c..9702686f1 100644
--- a/src/main/ipc/registerIpc.ts
+++ b/src/main/ipc/registerIpc.ts
@@ -919,6 +919,10 @@ export function registerIpc(ipcMain: IpcRegistrar, {
     assertMainRenderer(event, getMainWindow);
     return githubAuth.signOut();
   });
+  ipcMain.handle(IPC.githubAuthCancel, (event) => {
+    assertMainRenderer(event, getMainWindow);
+    githubAuth.cancelDeviceFlow();
+  });
   ipcMain.handle(IPC.githubAuthOpenUrl, (event, value: unknown) => {
     assertMainRenderer(event, getMainWindow);
     if (typeof value !== "string") throw new Error("URL is required.");
diff --git a/src/main/services/AgentControlService.ts b/src/main/services/AgentControlService.ts
index 34a5fda11..bfee70141 100644
--- a/src/main/services/AgentControlService.ts
+++ b/src/main/services/AgentControlService.ts
@@ -15,6 +15,7 @@ import type { TerminalManager } from "./TerminalManager.ts";
 import { onDiskPath, otherSpellings } from "./onDiskPath.ts";
 import { LaunchRefusal } from "./launchRefusal.ts";
 import { terminalFailureDetails } from "./terminalFailureDetails.ts";
+import { renderedTerminalText } from "./renderedTerminalText.ts";
 import { RESULT_CAPTURE_PROVIDERS } from "./resultCapture.ts";
 
 // Cards one parent may have in total, live or exited (the live limit is the person's setting).
@@ -30,8 +31,12 @@ export interface DelegationLimits {
 }
 const MAX_OBSERVE_CHARS = 8_192;
 // An exited agent's reason is in its last screen lines (OpenCode: "Error: Unexpected server error" for an unknown model).
-const MAX_EXIT_WINDOW_CHARS = 16_384;
 const MAX_EXIT_LINES = 20;
+/**
+ * How much of the scrollback (masked first) the plain-text output is rendered from: several full redraws of a
+ * full-screen TUI, and many screens of a line-based CLI, while each call still masks a bounded window.
+ */
+const MAX_SCREEN_SOURCE_CHARS = 65_536;
 const CHILD_POSITION_STEP = { x: 60, y: 60 };
 
 export interface SpawnAgentRequest {
@@ -52,7 +57,7 @@ export interface SpawnAgentRequest {
 export interface AgentObservation {
   sessionId: string;
   status: SessionSnapshot["status"];
-  /** Raw terminal tail, capped; capabilities with result \"none\" see nothing. */
+  /** The screen's text as a terminal shows it (no control sequences), masked, capped to its last characters. */
   output: string;
   /** Once the process exited: its exit code and the last lines of its screen as plain text, masked. */
   exitCode?: number | null;
@@ -70,12 +75,14 @@ export interface AgentWaitResult {
   status?: SessionSnapshot["status"];
   exitCode: number | null;
   waitedMs: number;
-  /** Masked terminal tail (masked before the cut). */
+  /** The screen's text as a terminal shows it (no control sequences), masked, capped to its last characters. */
   output: string;
   /** Once the process exited: the last lines of its screen as plain text, masked (why it stopped, e.g. a bad model). */
   exitLines?: string;
   /** The final answer of the turn that ended (Codex, OpenCode subagents), masked. */
   answer?: AgentAnswer;
+  /** Why a Codex or OpenCode subagent has no answer and never will in this launch (read its output instead). */
+  answerUnavailable?: AgentAnswerUnavailable;
 }
 
 export interface AgentWaitTiming {
@@ -106,13 +113,26 @@ export interface AgentResult {
   /** The session's status (idle once its turn ended; state stays "running" while the CLI is open). */
   status: SessionSnapshot["status"];
   exitCode: number | null;
+  /** The screen's text as a terminal shows it (no control sequences), masked, capped to its last characters. */
   output: string;
   /** Once the process exited: the last lines of its screen as plain text, masked. */
   exitLines?: string;
   /** The last turn's final answer as the agent reported it (Codex, OpenCode subagents), masked; absent otherwise. */
   answer?: AgentAnswer;
+  /** Why a Codex or OpenCode subagent has no answer and never will in this launch (read its output instead). */
+  answerUnavailable?: AgentAnswerUnavailable;
+}
+
+export interface AgentAnswerUnavailable {
+  /** "status_hooks_off": CanvasTTY status hooks were off for its launch, so it has no hooks to report an answer. */
+  reason: "status_hooks_off";
+  message: string;
 }
 
+const STATUS_HOOKS_OFF_MESSAGE = "This subagent's final answer cannot be captured because CanvasTTY status hooks are off "
+  + "(Settings → Agents; turning them on applies to agents started or restarted afterwards). Read its output instead: "
+  + "the output field here, or observe_agent.";
+
 export interface AgentAnswer {
   text: string;
   /** Only the end of a longer answer was kept. */
@@ -133,6 +153,11 @@ export class PromptNotDeliveredError extends Error {
 export class AgentControlService {
   private readonly terminals: TerminalManager;
   private readonly options: AgentControlOptions;
+  /**
+   * The card revision at which wait_for_agent last returned needs_approval for a subagent. While the card is still at
+   * that same prompt, the next wait waits for the person's answer instead of returning it again.
+   */
+  private readonly reportedApprovals = new Map<string, number | undefined>();
 
   constructor(terminals: TerminalManager, options: AgentControlOptions = {}) {
     this.terminals = terminals;
@@ -142,9 +167,10 @@ export class AgentControlService {
   /**
    * Creates the subagent card. With an initial prompt it resolves only once the prompt reached the agent's PTY
    * (after an asynchronous launch has started) and rejects with PromptNotDeliveredError when that launch did not
-   * start; the card stays, so the caller can inspect or cancel it.
+   * start; the card stays, so the caller can inspect or cancel it. `promptWaitMs` bounds that wait (the launch and
+   * the CLI's readiness together; LAUNCH_INPUT_WAIT_MS without it).
    */
-  spawn(request: SpawnAgentRequest, signal?: AbortSignal): Promise<SessionMetadata> {
+  spawn(request: SpawnAgentRequest, signal?: AbortSignal, promptWaitMs?: number): Promise<SessionMetadata> {
     if (!request || typeof request.parentSessionId !== "string") {
       throw new Error("A parent session id is required.");
     }
@@ -152,6 +178,8 @@ export class AgentControlService {
     const capabilities = PROVIDER_CAPABILITIES[request.provider];
     if (!capabilities) throw new Error("Unknown agent provider.");
     if (!capabilities.send) throw new Error(`${request.provider} cannot receive prompts.`);
+    const inheritance = this.subagentLaunchOptions(parent.id, request.provider, request.launchOptions, request.model);
+    if (inheritance.launchOptions !== request.launchOptions) request = { ...request, launchOptions: inheritance.launchOptions };
 
     const children = this.children(parent.id);
     if (children.length >= MAX_CHILDREN_PER_PARENT) {
@@ -194,7 +222,7 @@ export class AgentControlService {
     if (request.initialPrompt === undefined || request.initialPrompt.length === 0) return Promise.resolve(created);
     const input = request.provider === "codex"
       ? codexPromptInput(request.initialPrompt) : `${request.initialPrompt}\r`;
-    return this.deliver(created.id, input, "prompt", signal)
+    return this.deliver(created.id, input, "prompt", signal, promptWaitMs)
       .then(() => this.terminals.getMetadata(created.id) ?? created, (error: unknown) => {
         // Cancelled while it started: nobody receives its id, so the card is closed instead of left running.
         if (!signal?.aborted) throw error;
@@ -203,21 +231,75 @@ export class AgentControlService {
       });
   }
 
+  /**
+   * The plugin launch options a subagent of this parent starts with. A plugin the spawn does not name keeps the
+   * orchestrator's own values, which the person chose: a subagent never falls back silently to its CLI's default model
+   * and sign-in (for OpenCode that is OpenCode Zen, a third-party service) while its orchestrator runs on a model the
+   * person picked through a plugin. Naming a plugin is an explicit choice (checked by the launch: another value needs
+   * the plugin's delegable declaration). Another CLI cannot inherit a plugin choice that set the orchestrator's model,
+   * and is refused unless the spawn names that plugin. An explicit model next to an inherited choice that sets the model is refused.
+   */
+  subagentLaunchOptions(
+    parentSessionId: string,
+    provider: AgentProviderId,
+    requested: SpawnAgentRequest["launchOptions"],
+    model?: string
+  ): { launchOptions: SpawnAgentRequest["launchOptions"]; sources: Record<string, "inherited" | "explicit"> } {
+    const parent = this.requireSession(parentSessionId);
+    const sources: Record<string, "inherited" | "explicit"> = {};
+    for (const pluginId of Object.keys(requested ?? {})) sources[pluginId] = "explicit";
+    const parentOptions = this.terminals.launchOptionsOf(parent.id);
+    const unnamed = Object.keys(parentOptions ?? {}).filter((pluginId) => !Object.hasOwn(sources, pluginId));
+    if (!parentOptions || unnamed.length === 0) return { launchOptions: requested, sources };
+    const modelSetBy = new Set(this.terminals.modelSetByOf(parent.id));
+    const modelPlugins = unnamed.filter((pluginId) => modelSetBy.has(pluginId));
+    if (parent.provider !== provider) {
+      if (modelPlugins.length === 0) return { launchOptions: requested, sources };
+      const named = modelPlugins.map((pluginId) => pluginId.slice(0, 80)).join(", ");
+      // The person chose those options for this CLI; another CLI would run on its own default model and sign-in.
+      throw new DelegationRefusal(`This orchestrator's model is set by its ${named} launch options, which the person chose for ${parent.provider}; `
+        + `a ${provider} subagent would not get that model. Spawn a ${parent.provider} subagent instead.`);
+    }
+    if (model !== undefined && modelPlugins.length > 0) {
+      const named = modelPlugins.map((pluginId) => pluginId.slice(0, 80)).join(", ");
+      throw new DelegationRefusal(`This subagent inherits its orchestrator's ${named} launch options, which set its model; it cannot also take model ${model.slice(0, 80)}. `
+        + `Omit model, or pass launchOptions for ${named} that choose the model you want.`);
+    }
+    // A plugin turned off since the orchestrator started is left out, unless it set the model: then the subagent would
+    // run on its CLI's own default model and sign-in, so it is refused instead.
+    const gone = new Set(this.terminals.unavailableLaunchPlugins(Object.fromEntries(unnamed.map((pluginId) => [pluginId, parentOptions[pluginId]]))));
+    const goneModel = modelPlugins.filter((pluginId) => gone.has(pluginId));
+    if (goneModel.length > 0) {
+      const named = goneModel.map((pluginId) => pluginId.slice(0, 80)).join(", ");
+      throw new DelegationRefusal(`This orchestrator's model is set by its ${named} launch options, and ${named} is now disabled, removed or not trusted; `
+        + "a subagent would run on its CLI's own default model instead. The person can turn the plugin back on, or start the orchestrator again.");
+    }
+    const launchOptions: NonNullable<SpawnAgentRequest["launchOptions"]> = { ...(requested ?? {}) };
+    for (const pluginId of unnamed.filter((id) => !gone.has(id))) {
+      launchOptions[pluginId] = parentOptions[pluginId]!;
+      sources[pluginId] = "inherited";
+    }
+    return { launchOptions, sources };
+  }
+
   /** The profile a subagent of this parent gets for this request (what spawn will use), or why it gets none. */
   profileFor(parentSessionId: string, provider: AgentProviderId, requested?: unknown): { profile: LaunchProfile; inherited: boolean } | { error: string } {
     return subagentProfile(this.requireSession(parentSessionId).profile, provider, requested, this.containment());
   }
 
   /** Validates at once (throws); resolves once the text reached the agent, and rejects when it did not. */
-  send(sessionId: string, text: string, submit = true, signal?: AbortSignal): Promise<void> {
+  send(sessionId: string, text: string, submit = true, signal?: AbortSignal, waitMs?: number): Promise<void> {
     const session = this.requireSession(sessionId);
     if (session.provider === "terminal") throw new Error("Plain terminals are not agents.");
     const capabilities = PROVIDER_CAPABILITIES[session.provider as AgentProviderId];
     if (!capabilities.send) throw new Error(`${session.provider} cannot receive prompts.`);
     if (typeof text !== "string" || text.length === 0) throw new Error("Prompt text is required.");
     if (session.exitCode !== null) throw new Error("Agent session has already exited.");
+    // The card shows a prompt only the person may answer (OpenCode's own shell prompt, for example): text typed into
+    // it now, or the Enter after it, would answer that prompt on the person's behalf.
+    if (session.status === "needs_approval") throw new DelegationRefusal(APPROVAL_REFUSAL);
     const input = session.provider === "codex" ? codexPromptInput(text, submit) : submit ? `${text}\r` : text;
-    return this.deliver(sessionId, input, "text", signal);
+    return this.deliver(sessionId, input, "text", signal, waitMs);
   }
 
   status(sessionId: string): SessionMetadata {
@@ -288,25 +370,28 @@ export class AgentControlService {
     return false;
   }
 
-  observe(sessionId: string, maxChars = MAX_OBSERVE_CHARS): AgentObservation {
-    const session = this.requireSession(sessionId);
-    if (session.provider === "terminal") throw new Error("Plain terminals are not agents.");
-    const capabilities = PROVIDER_CAPABILITIES[session.provider as AgentProviderId];
-    if (!capabilities.observe) throw new Error(`${session.provider} cannot be observed.`);
+  async observe(sessionId: string, maxChars = MAX_OBSERVE_CHARS): Promise<AgentObservation> {
+    const session = this.observableSession(sessionId);
     const buffer = this.terminals.readBuffer(sessionId).buffer;
-    const exitLines = session.exitCode === null ? null : this.exitLines(buffer);
+    const exitLines = session.exitCode === null ? null : await this.screenExitLines(sessionId, buffer);
     return {
       sessionId: session.id,
       status: session.status,
-      // Masked before the cut (a cut inside a secret would leave a tail no pattern recognizes), over a window
-      // wider than any match rather than the whole scrollback.
-      output: this.redactTail(buffer, maxChars),
+      output: await this.screenText(sessionId, buffer, maxChars),
       ...(session.exitCode === null ? {} : { exitCode: session.exitCode }),
       ...(exitLines ? { exitLines } : {})
     };
   }
 
-  result(sessionId: string): AgentResult {
+  private observableSession(sessionId: string): SessionMetadata {
+    const session = this.requireSession(sessionId);
+    if (session.provider === "terminal") throw new Error("Plain terminals are not agents.");
+    const capabilities = PROVIDER_CAPABILITIES[session.provider as AgentProviderId];
+    if (!capabilities.observe) throw new Error(`${session.provider} cannot be observed.`);
+    return session;
+  }
+
+  async result(sessionId: string): Promise<AgentResult> {
     const session = this.requireSession(sessionId);
     if (session.provider === "terminal") throw new Error("Plain terminals are not agents.");
     const capabilities = PROVIDER_CAPABILITIES[session.provider as AgentProviderId];
@@ -317,6 +402,7 @@ export class AgentControlService {
     const buffer = capabilities.result === "terminal"
       ? this.terminals.readBuffer(sessionId).buffer
       : "";
+    const exitLines = session.exitCode === null ? null : await this.screenExitLines(sessionId, buffer);
     return {
       sessionId: session.id,
       state: session.exitCode === null
@@ -324,9 +410,9 @@ export class AgentControlService {
         : session.exitCode === 0 ? "done" : "failed",
       status: session.status,
       exitCode: session.exitCode,
-      output: this.redactTail(buffer, MAX_OBSERVE_CHARS),
-      ...(session.exitCode !== null && this.exitLines(buffer) ? { exitLines: this.exitLines(buffer)! } : {}),
-      ...(answer ? { answer } : {})
+      output: buffer ? await this.screenText(sessionId, buffer, MAX_OBSERVE_CHARS) : "",
+      ...(exitLines ? { exitLines } : {}),
+      ...(answer ? { answer } : this.answerUnavailable(sessionId))
     };
   }
 
@@ -344,14 +430,25 @@ export class AgentControlService {
     const timing = this.options.waitTiming ?? AGENT_WAIT_TIMING;
     const timeoutMs = Math.min(MAX_AGENT_WAIT_MS, Math.max(0, request.timeoutMs));
     const started = Date.now();
-    const answer = (session: SessionMetadata | null, reason: AgentWaitReason): AgentWaitResult => {
+    const answer = async (session: SessionMetadata | null, reason: AgentWaitReason): Promise<AgentWaitResult> => {
       const waitedMs = Date.now() - started;
+      if (reason === "needs_approval") this.reportedApprovals.set(sessionId, session?.revision);
+      else if (session?.status !== "needs_approval") this.reportedApprovals.delete(sessionId);
       if (!session) return { sessionId, reason, exitCode: null, waitedMs, output: "" };
-      let observation: AgentObservation | null = null;
-      try { observation = this.observe(sessionId); } catch { observation = null; }
+      let observable = false;
+      try { observable = Boolean(this.observableSession(sessionId)); } catch { observable = false; }
+      let output = "";
+      let exitLines: string | null = null;
+      if (observable) {
+        try {
+          const buffer = this.terminals.readBuffer(sessionId).buffer;
+          output = await this.screenText(sessionId, buffer, MAX_OBSERVE_CHARS);
+          if (session.exitCode !== null) exitLines = await this.screenExitLines(sessionId, buffer);
+        } catch { output = ""; }
+      }
       const answer = reason === "timeout" || reason === "needs_approval" ? null : this.answer(sessionId);
-      return { sessionId, reason, status: session.status, exitCode: session.exitCode, waitedMs, output: observation?.output ?? "",
-        ...(observation?.exitLines ? { exitLines: observation.exitLines } : {}), ...(answer ? { answer } : {}) };
+      return { sessionId, reason, status: session.status, exitCode: session.exitCode, waitedMs, output,
+        ...(exitLines ? { exitLines } : {}), ...(answer ? { answer } : this.answerUnavailable(sessionId)) };
     };
     let offset = this.outputOffset(sessionId);
     let changedAt = started;
@@ -363,7 +460,10 @@ export class AgentControlService {
       if (current !== offset) { offset = current; changedAt = now; }
       const quietFor = now - changedAt;
       if (session.exitCode !== null) return answer(session, session.exitCode === 0 ? "done" : "failed");
-      if (session.status === "needs_approval") return answer(session, "needs_approval");
+      // A prompt this wait already returned (the card has not changed since) is the person's to answer: wait for it.
+      const reported = this.reportedApprovals.has(sessionId) && this.reportedApprovals.get(sessionId) === session.revision;
+      if (session.status === "needs_approval" && !reported) return answer(session, "needs_approval");
+      if (session.status !== "needs_approval") this.reportedApprovals.delete(sessionId);
       // After a prompt, an idle that no turn followed (the CLI's startup idle, or one reported before the turn began)
       // is not the answer: only an idle after a turn that started since that prompt is. A CLI that never reports its
       // turns still ends as "quiet" once its screen stops changing.
@@ -385,20 +485,53 @@ export class AgentControlService {
   }
 
   /** Through the terminal manager's one delivery rule: exactly once, into the launch that is starting now. */
-  private async deliver(sessionId: string, data: string, what: "prompt" | "text", signal?: AbortSignal): Promise<void> {
-    const delivery = await this.terminals.deliverInput(sessionId, data, undefined, signal, this.requireSession(sessionId).provider === "codex");
+  private async deliver(sessionId: string, data: string, what: "prompt" | "text", signal?: AbortSignal, waitMs?: number): Promise<void> {
+    // waitMs (AGENT_TOOL_DELIVERY_MS for the orchestration tools) also caps the wait for Codex's composer.
+    const codex = this.requireSession(sessionId).provider === "codex";
+    const delivery = await this.terminals.deliverInput(sessionId, data, waitMs, signal, codex);
     if (!delivery.delivered) {
       throw new PromptNotDeliveredError(sessionId, `The ${what} for agent ${sessionId} was not delivered: ${delivery.reason}`);
     }
   }
 
-  /** The last lines of an exited agent's screen as plain text: masked over the whole tail window before it is cut. */
-  private exitLines(buffer: string): string | null {
-    const text = terminalFailureDetails(this.redactTail(buffer, MAX_EXIT_WINDOW_CHARS));
+  /**
+   * The last lines of the screen, rendered like `output` (a TUI that redraws only the cells that changed would
+   * otherwise read as its redraws joined together) from the same window as `output` (a line drawn early and never
+   * redrawn, such as the model's status line, is still on the screen), masked before it is cut and again after.
+   */
+  private async screenExitLines(sessionId: string, buffer: string): Promise<string | null> {
+    const source = fromSequenceBoundary(this.redactTail(buffer, MAX_SCREEN_SOURCE_CHARS), buffer.length > MAX_SCREEN_SOURCE_CHARS);
+    let screen: string;
+    try {
+      screen = this.redactTail(await renderedTerminalText(source, this.geometry(sessionId)), MAX_SCREEN_SOURCE_CHARS);
+    } catch {
+      screen = source;
+    }
+    const text = terminalFailureDetails(screen);
     if (!text) return null;
     return text.split("\n").slice(-MAX_EXIT_LINES).join("\n");
   }
 
+  /**
+   * The text the agent's terminal shows, as plain text: a TUI's redraws (cursor moves, erases, colors) rendered by a
+   * headless terminal instead of returned as escape sequences. Masked twice: the scrollback before it is rendered
+   * (a held value the TUI wrote whole is gone before any redraw can split it), and the rendered text before its
+   * last `maxChars` are cut (a value the screen shows across wrapped rows or box sides). Should the rendering fail,
+   * the masked raw tail is returned as before.
+   */
+  private async screenText(sessionId: string, buffer: string, maxChars: number): Promise<string> {
+    const source = fromSequenceBoundary(this.redactTail(buffer, MAX_SCREEN_SOURCE_CHARS), buffer.length > MAX_SCREEN_SOURCE_CHARS);
+    try {
+      return this.redactTail(await renderedTerminalText(source, this.geometry(sessionId)), maxChars);
+    } catch {
+      return this.redactTail(buffer, maxChars);
+    }
+  }
+
+  private geometry(sessionId: string): { cols: number; rows: number } | null {
+    try { return typeof this.terminals.geometry === "function" ? this.terminals.geometry(sessionId) : null; } catch { return null; }
+  }
+
   /** Plugin launch secrets never reach another agent through observed output: the tail as masking the whole text leaves it. */
   private redactTail(text: string, maxChars: number): string {
     if (typeof this.terminals.redactSecretsTail === "function") return this.terminals.redactSecretsTail(text, maxChars);
@@ -409,6 +542,14 @@ export class AgentControlService {
     try { return typeof this.terminals.turnProgress === "function" ? this.terminals.turnProgress(sessionId) : null; } catch { return null; }
   }
 
+  /** Says so when no answer can be captured for this agent, so an orchestrator reads its output instead of waiting for one. */
+  private answerUnavailable(sessionId: string): { answerUnavailable?: AgentAnswerUnavailable } {
+    try {
+      const reason = typeof this.terminals.answerUnavailable === "function" ? this.terminals.answerUnavailable(sessionId) : null;
+      return reason === "status_hooks_off" ? { answerUnavailable: { reason, message: STATUS_HOOKS_OFF_MESSAGE } } : {};
+    } catch { return {}; }
+  }
+
   private answer(sessionId: string): AgentAnswer | null {
     try {
       const answer = typeof this.terminals.answer === "function" ? this.terminals.answer(sessionId) : null;
@@ -435,6 +576,11 @@ export class AgentControlService {
   }
 }
 
+/** send_to_agent while the subagent's card waits for the person (needs_approval). */
+export const APPROVAL_REFUSAL = "This subagent is waiting for the person to answer a prompt in its card (needs_approval); "
+  + "only the person may answer it, so nothing was sent. Tell the person what it asks (wait_for_agent's output shows it), "
+  + "then call wait_for_agent again: it waits for their answer.";
+
 /** A delegation rule refused the request; the text says why, for the orchestrator to adapt instead of retrying. */
 export class DelegationRefusal extends LaunchRefusal {
   constructor(message: string) {
@@ -525,3 +671,13 @@ function tail(text: string, maxChars: number): string {
 function spawnCanceled(): Error {
   return new DOMException("The spawn was canceled before its agent started.", "AbortError");
 }
+
+/**
+ * A window cut from the end of a terminal stream may begin inside an escape sequence, whose rest (`18;9H`, `m`) a
+ * terminal would print as text. A cut window therefore starts at its first escape or line break instead.
+ */
+function fromSequenceBoundary(window: string, cut: boolean): string {
+  if (!cut) return window;
+  const starts = [window.indexOf("\u001b"), window.indexOf("\n")].filter((index) => index >= 0);
+  return starts.length > 0 ? window.slice(Math.min(...starts)) : window;
+}
diff --git a/src/main/services/GithubAuthService.ts b/src/main/services/GithubAuthService.ts
index defbdb6b7..66055a245 100644
--- a/src/main/services/GithubAuthService.ts
+++ b/src/main/services/GithubAuthService.ts
@@ -2,6 +2,7 @@ import { randomUUID } from "node:crypto";
 import * as electron from "electron";
 import { readFile, writeFile, mkdir, rename, rm } from "node:fs/promises";
 import { join, dirname } from "node:path";
+import type { GithubDeviceFlowState } from "../../shared/contracts";
 
 /** GitHub OAuth Device Flow for the plugin showcase. */
 
@@ -10,6 +11,7 @@ export interface GithubAuthStatus {
   authorized: boolean;
   login: string | null;
   tokenExpiresAt: number | null;
+  deviceFlowState: GithubDeviceFlowState;
 }
 
 interface StoredTokens {
@@ -60,6 +62,7 @@ export class GithubAuthService {
   private readonly pollTimeoutMs: number;
   private generation = 0;
   private deviceFlow: { generation: number; controller: AbortController } | null = null;
+  private deviceFlowState: GithubDeviceFlowState = "idle";
   private storeWrite = Promise.resolve();
 
   constructor(userDataPath: string, clientId?: string, options: GithubAuthServiceOptions = {}) {
@@ -126,13 +129,20 @@ export class GithubAuthService {
 
   async status(): Promise<GithubAuthStatus> {
     if (!this.tokens) {
-      return { configured: this.clientConfigured, authorized: false, login: null, tokenExpiresAt: null };
+      return {
+        configured: this.clientConfigured,
+        authorized: false,
+        login: null,
+        tokenExpiresAt: null,
+        deviceFlowState: this.deviceFlowState
+      };
     }
     return {
       configured: this.clientConfigured,
       authorized: true,
       login: this.tokens.login,
-      tokenExpiresAt: this.tokens.expiresAt
+      tokenExpiresAt: this.tokens.expiresAt,
+      deviceFlowState: this.deviceFlowState
     };
   }
 
@@ -146,10 +156,11 @@ export class GithubAuthService {
       throw new Error("GitHub OAuth is not configured (missing client id).");
     }
 
-    this.cancelDeviceFlow();
+    this.abortDeviceFlow();
     const generation = ++this.generation;
     const controller = new AbortController();
     this.deviceFlow = { generation, controller };
+    this.deviceFlowState = "pending";
 
     try {
       const body = new URLSearchParams({ client_id: this.clientId });
@@ -176,7 +187,11 @@ export class GithubAuthService {
 
       void this.pollDeviceCode(deviceCode, interval, flowLifetimeMs, generation, controller.signal)
         .catch((error) => {
-          if (!isAbortError(error)) console.warn("CanvasTTY GitHub OAuth device poll failed.", error);
+          // Only this flow's own abort (Cancel, sign-out, a new sign-in) is a cancellation; a request
+          // that hit its timeout also throws an AbortError, and that is a failure the person must see.
+          if (controller.signal.aborted) return;
+          this.finishDeviceFlow(generation, "failed");
+          console.warn("CanvasTTY GitHub OAuth device poll failed.", error);
         })
         .finally(() => {
           if (this.deviceFlow?.generation === generation) this.deviceFlow = null;
@@ -184,14 +199,26 @@ export class GithubAuthService {
 
       return { userCode, verificationUri, interval, expiresAt: this.now() + flowLifetimeMs };
     } catch (error) {
-      if (this.deviceFlow?.generation === generation) this.deviceFlow = null;
+      if (this.deviceFlow?.generation === generation) {
+        this.deviceFlow = null;
+        // The caller receives this error directly, so no separate outcome is kept.
+        this.deviceFlowState = "idle";
+      }
       throw error;
     }
   }
 
+  /** Stops a pending device sign-in. A completed sign-in or an existing session is left untouched. */
+  cancelDeviceFlow(): void {
+    if (!this.deviceFlow || this.deviceFlowState !== "pending") return;
+    this.abortDeviceFlow();
+    this.deviceFlowState = "cancelled";
+  }
+
   async signOut(): Promise<void> {
     this.generation += 1;
-    this.cancelDeviceFlow();
+    this.abortDeviceFlow();
+    this.deviceFlowState = "idle";
     this.tokens = null;
     this.refreshPromise = null;
     await this.enqueueStoreWrite(async () => {
@@ -199,9 +226,17 @@ export class GithubAuthService {
     });
   }
 
-  private cancelDeviceFlow(): void {
-    this.deviceFlow?.controller.abort();
+  private abortDeviceFlow(): void {
+    const flow = this.deviceFlow;
+    // Clear the flow before aborting so a fetcher that ignores its AbortSignal
+    // cannot apply a late response: every step re-checks isCurrentFlow().
     this.deviceFlow = null;
+    flow?.controller.abort();
+  }
+
+  private finishDeviceFlow(generation: number, state: GithubDeviceFlowState): void {
+    if (this.deviceFlow?.generation !== generation || this.deviceFlowState !== "pending") return;
+    this.deviceFlowState = state;
   }
 
   private isCurrentFlow(generation: number, signal: AbortSignal): boolean {
@@ -217,51 +252,88 @@ export class GithubAuthService {
   ): Promise<void> {
     const started = this.now();
     let interval = initialInterval;
+    let transientFailures = 0;
+    // A dropped connection, the request timeout, HTTP 429/5xx or an unreadable
+    // body is not the end of the flow: the person may still approve the code.
+    // Back off (RFC 8628 section 3.5) and poll again, but give up after a few
+    // failures in a row instead of polling a broken endpoint until the code expires.
+    const retryOrFail = (): boolean => {
+      transientFailures += 1;
+      if (transientFailures >= MAX_CONSECUTIVE_POLL_FAILURES) {
+        this.finishDeviceFlow(generation, "failed");
+        return false;
+      }
+      interval = backedOff(interval);
+      return true;
+    };
     while (this.now() - started < lifetimeMs) {
       await this.wait(interval * 1000, signal);
       if (!this.isCurrentFlow(generation, signal)) return;
+      if (this.now() - started >= lifetimeMs) break;
 
       const body = new URLSearchParams({
         client_id: this.clientId,
         device_code: deviceCode,
         grant_type: "urn:ietf:params:oauth:grant-type:device_code"
       });
-      // A dropped connection, the 15 s request timeout or a GitHub 5xx is not
-      // the end of the flow: the person may still approve the code. Back off
-      // (RFC 8628 section 3.5) and poll again until the code expires.
-      let payload: unknown;
+      let response: Response;
+      let payload: unknown = null;
       try {
-        const response = await this.request("https://github.com/login/oauth/access_token", {
+        response = await this.request("https://github.com/login/oauth/access_token", {
           method: "POST",
           headers: oauthHeaders(),
           body: body.toString()
         }, signal);
-        if (!response.ok) {
-          interval = backedOff(interval);
-          continue;
-        }
-        payload = await response.json();
+        // OAuth errors may arrive with HTTP 400 (RFC 6749 section 5.2), so read the body either way.
+        payload = await response.json().catch(() => null);
       } catch (error) {
         if (signal.aborted) throw error;
-        interval = backedOff(interval);
-        continue;
+        if (!this.isCurrentFlow(generation, signal)) return;
+        if (retryOrFail()) continue;
+        return;
       }
-      if (!isRecord(payload)) {
-        interval = backedOff(interval);
-        continue;
+      if (!this.isCurrentFlow(generation, signal)) return;
+      const providerError = isRecord(payload) && typeof payload.error === "string" ? payload.error : null;
+      // HTTP 429/5xx is transient whatever its body says (e.g. {"error":"temporarily_unavailable"});
+      // slow_down keeps its own handling below, which also raises the interval.
+      if (!response.ok && isTransientStatus(response.status) && providerError !== "slow_down") {
+        if (retryOrFail()) continue;
+        return;
       }
-      if (payload.error === "authorization_pending" || payload.error === "slow_down") {
-        if (payload.error === "slow_down") {
+      if (providerError === null && (!response.ok || !isRecord(payload))) {
+        if (response.ok) {
+          if (retryOrFail()) continue;
+          return;
+        }
+        this.finishDeviceFlow(generation, "failed");
+        return;
+      }
+      if (providerError === "authorization_pending" || providerError === "slow_down") {
+        transientFailures = 0;
+        if (providerError === "slow_down") {
           // +5 s for this and later polls; GitHub may name a longer interval.
-          const requested = typeof payload.interval === "number" && Number.isFinite(payload.interval) ? payload.interval : 0;
+          const requested = isRecord(payload) && typeof payload.interval === "number" && Number.isFinite(payload.interval)
+            ? payload.interval
+            : 0;
           interval = Math.max(interval + 5, Math.min(requested, MAX_POLL_INTERVAL_SECONDS));
         }
         continue;
       }
-      if (payload.error === "access_denied" || payload.error === "expired_token") return;
-      if (typeof payload.access_token === "string") {
+      if (providerError === "access_denied") {
+        this.finishDeviceFlow(generation, "denied");
+        return;
+      }
+      if (providerError === "expired_token") {
+        this.finishDeviceFlow(generation, "expired");
+        return;
+      }
+      if (providerError === null && isRecord(payload) && typeof payload.access_token === "string") {
         const login = await this.fetchLogin(payload.access_token, signal);
-        if (!login || !this.isCurrentFlow(generation, signal)) return;
+        if (!this.isCurrentFlow(generation, signal)) return;
+        if (!login) {
+          this.finishDeviceFlow(generation, "failed");
+          return;
+        }
         const expiresIn = typeof payload.expires_in === "number" && payload.expires_in > 0
           ? payload.expires_in
           : null;
@@ -272,11 +344,15 @@ export class GithubAuthService {
           login
         };
         this.tokens = next;
+        this.deviceFlowState = "idle";
         await this.persist(next, generation);
         return;
       }
+      // Any other provider error (device_flow_disabled, incorrect_client_credentials, ...) is final.
+      this.finishDeviceFlow(generation, "failed");
       return;
     }
+    this.finishDeviceFlow(generation, "expired");
   }
 
   private async refreshAccessToken(expected: StoredTokens, generation: number): Promise<string | null> {
@@ -326,7 +402,8 @@ export class GithubAuthService {
       const payload: unknown = await response.json();
       return isRecord(payload) && typeof payload.login === "string" ? payload.login : null;
     } catch (error) {
-      if (isAbortError(error)) throw error;
+      // Rethrow only a cancelled flow; a timed-out or failed profile request is a failed sign-in.
+      if (signal.aborted && isAbortError(error)) throw error;
       return null;
     }
   }
@@ -433,6 +510,12 @@ function isMissingFile(error: unknown): boolean {
 }
 
 const MAX_POLL_INTERVAL_SECONDS = 60;
+/** Consecutive network, timeout, HTTP 429/5xx or unreadable answers before the sign-in is reported as failed. */
+const MAX_CONSECUTIVE_POLL_FAILURES = 5;
+
+function isTransientStatus(status: number): boolean {
+  return status === 408 || status === 429 || status >= 500;
+}
 
 function backedOff(interval: number): number {
   return Math.min(Math.max(interval * 2, interval + 1), MAX_POLL_INTERVAL_SECONDS);
diff --git a/src/main/services/LaunchPipeline.ts b/src/main/services/LaunchPipeline.ts
index 439370a56..6cccb067c 100644
--- a/src/main/services/LaunchPipeline.ts
+++ b/src/main/services/LaunchPipeline.ts
@@ -1,5 +1,5 @@
 import { randomUUID } from "node:crypto";
-import { mkdir, rm, writeFile } from "node:fs/promises";
+import { mkdir, rm, rmdir, writeFile } from "node:fs/promises";
 import { dirname, join, resolve } from "node:path";
 import type {
   LaunchProfileId,
@@ -12,6 +12,7 @@ import type {
 import { claudeCoreSettingsKey, coreOwnedLaunchArgument, parseInlineSettings } from "./terminalLaunch.ts";
 import { MAX_PLUGIN_SLOT_BYTES } from "./TerminalSessionStore.ts";
 import { LaunchRefusal } from "./launchRefusal.ts";
+import { contributedHookHome } from "./accountHomes.ts";
 import { MAX_INSPECTED_CONFIG_BYTES, parseJsonc, readInspectedFile } from "./inspectedConfig.ts";
 
 /** A trusted plugin service that declared launch options (PluginManager.launchContributors). */
@@ -22,6 +23,8 @@ export interface LaunchContributor {
   launch: PluginServiceLaunch;
   /** The plugin holds the `secrets` permission, so `secretEnv` may name its secrets. */
   secrets: boolean;
+  /** `<userData>/plugin-data/<pluginId>`, the host's record (never the plugin's answer); account homes live below it. */
+  dataDir?: string;
 }
 
 export interface LaunchPipelineDependencies {
@@ -73,6 +76,15 @@ export type PreparedLaunch =
     envSources: Record<string, string>;
     /** A contributor runs the agent on another model than its vendor's: "auto" becomes accept-edits. */
     thirdPartyModel: boolean;
+    /** Plugin ids whose contribution decides the model: it passes the CLI's model flag or runs a third-party model. */
+    modelSetBy: string[];
+    /** This run's own folder below the card's folder; `cleanup` removes it. */
+    runDirectory: string;
+    /**
+     * Grok/Kimi: the home (`GROK_HOME` / `KIMI_CODE_HOME`) the contributor that set it handed this launch inside its own
+     * data folder (contributedHookHome); CanvasTTY writes that CLI's temporary configuration there.
+     */
+    accountHome?: string;
     cleanup(): Promise<void>;
   }
   | { ok: false; reason: string };
@@ -118,9 +130,25 @@ export class LaunchPipeline {
     return rm(this.dependencies.runsRoot, { recursive: true, force: true });
   }
 
-  /** Removes a closed session's file folders. */
-  forgetSession(sessionId: string): Promise<void> {
-    return rm(join(this.dependencies.runsRoot, safeSegment(sessionId)), { recursive: true, force: true });
+  /**
+   * Removes a closed card's run folders that this pipeline prepared for it, then the card's folder if nothing else
+   * is left in it. A card brought back under the same id may already be writing a new run folder there; it is kept.
+   */
+  async forgetSession(sessionId: string, runDirectories: Iterable<string>): Promise<void> {
+    const parent = join(this.dependencies.runsRoot, safeSegment(sessionId));
+    const errors: unknown[] = [];
+    for (const runDirectory of new Set(runDirectories)) {
+      // Only a run folder directly below this card's folder is ever removed here.
+      if (dirname(resolve(runDirectory)) !== resolve(parent)) continue;
+      try { await removeRunFolder(runDirectory); } catch (error) { errors.push(error); }
+    }
+    try {
+      await rmdir(parent);
+    } catch (error) {
+      // Still in use by a newer run of the same card, or already gone; the startup sweep removes what is left.
+      if (!PARENT_KEPT_CODES.has((error as NodeJS.ErrnoException).code ?? "")) errors.push(error);
+    }
+    if (errors.length) throw new AggregateError(errors, "Launch run folders could not be removed.");
   }
 
   /** Launch policies that apply to this agent: every launch of it asks them, chosen or not. */
@@ -145,45 +173,45 @@ export class LaunchPipeline {
    * Checks launcher values against each plugin's declared fields and fills defaults.
    * Throws with a person-readable reason; returns undefined when nothing was chosen.
    */
-  normalizeOptions(provider: ProviderId, candidate: unknown, context?: { delegated?: boolean }): Record<string, PluginLaunchValues> | undefined {
+  normalizeOptions(provider: ProviderId, candidate: unknown, context?: {
+    delegated?: boolean;
+    /** The orchestrator's own options: repeating them for its subagent is the person's choice, not the orchestrator's. */
+    inherited?: Readonly<Record<string, Readonly<PluginLaunchValues>>>;
+  }): Record<string, PluginLaunchValues> | undefined {
     if (candidate === undefined) return undefined;
     if (!isRecord(candidate) || Object.keys(candidate).length > MAX_OPTION_PLUGINS) throw new Error("Launch options are invalid.");
     if (Object.keys(candidate).length === 0) return undefined;
     if (provider === "terminal") throw new Error("A plain terminal takes no launch options.");
     const contributors = new Map(this.dependencies.contributors().map((contributor) => [contributor.pluginId, contributor]));
     const options: Record<string, PluginLaunchValues> = {};
-    for (const [pluginId, raw] of Object.entries(candidate)) {
+    for (const [pluginId, requested] of Object.entries(candidate)) {
+      let raw = requested;
       const contributor = contributors.get(pluginId);
       if (!contributor) throw new Error(unavailableReason(pluginId));
       // An orchestrator picks options for its subagents only where the plugin said that is safe.
+      const notDelegable = `${contributor.pluginName} has not declared its launch options safe for an orchestrator to choose; only the person chooses them, in the launcher.`;
+      // Without the declaration a subagent may only repeat the orchestrator's own values: the keys it leaves out are
+      // the inherited ones, never the plugin's defaults, and the result is checked against them below.
+      const inherited = context?.delegated && contributor.launch.delegable !== true ? context.inherited?.[pluginId] : undefined;
       if (context?.delegated && contributor.launch.delegable !== true) {
-        throw new LaunchRefusal(`${contributor.pluginName} has not declared its launch options safe for an orchestrator to choose; only the person chooses them, in the launcher.`);
+        if (!inherited || !sameAsInherited(raw, inherited)) throw new LaunchRefusal(notDelegable);
+        raw = { ...inherited, ...(raw as Record<string, unknown>) };
       }
-      const { appliesTo, fields } = contributor.launch;
-      if (appliesTo && !appliesTo.includes(provider as never)) {
-        throw new Error(`${contributor.pluginName} launch options do not apply to ${provider}.`);
+      if (!inherited) {
+        options[pluginId] = launchValues(contributor, provider, raw);
+        continue;
       }
-      if (!isRecord(raw)) throw new Error(`${contributor.pluginName} launch options are invalid.`);
-      const known = new Set(fields.map((field) => field.key));
-      const unknown = Object.keys(raw).find((key) => !known.has(key));
-      if (unknown) throw new Error(`${contributor.pluginName} has no launch option ${unknown.slice(0, 40)}.`);
-      const values: PluginLaunchValues = {};
-      for (const field of fields) {
-        const value = raw[field.key] ?? field.default
-          ?? (field.kind === "boolean" ? false : field.kind === "select" ? field.options?.[0]?.value ?? "" : "");
-        const plainText = (limit: number): boolean => typeof value === "string" && value.length <= limit && !/[\u0000-\u001f\u007f]/.test(value);
-        // A service-provided choice may have changed since the launcher showed it: the service checks it when it prepares.
-        const valid = field.kind === "boolean"
-          ? typeof value === "boolean"
-          : field.kind === "select"
-            ? typeof value === "string" && (Boolean(field.options?.some((option) => option.value === value)) || (field.optionsFrom === "service" && plainText(200)))
-            : plainText(field.maxLength ?? 200);
-        if (!valid) throw new Error(`${contributor.pluginName} launch option ${field.label} is invalid.`);
-        values[field.key] = value as boolean | string;
-      }
-      if (Buffer.byteLength(JSON.stringify(values), "utf8") > MAX_PLUGIN_SLOT_BYTES) {
-        throw new Error(`${contributor.pluginName} launch options exceed 4 KB.`);
+      // The person's earlier choice no longer fits the plugin (it changed its fields or what it applies to): a subagent
+      // never gets other values than its orchestrator's, so it is refused with that reason.
+      const changed = `${contributor.pluginName} launch options changed since this orchestrator started, so its subagents cannot repeat the person's choice. `
+        + "The person can start the orchestrator again with the new options.";
+      let values: PluginLaunchValues;
+      try {
+        values = launchValues(contributor, provider, raw);
+      } catch {
+        throw new LaunchRefusal(changed);
       }
+      if (!sameValues(values, inherited)) throw new LaunchRefusal(changed);
       options[pluginId] = values;
     }
     return options;
@@ -239,8 +267,14 @@ export class LaunchPipeline {
    * refuses the whole launch with a reason; nothing is ever launched without a contribution the person selected,
    * or past a policy that did not answer.
    */
-  async prepare(context: LaunchSessionContext): Promise<PreparedLaunch> {
+  async prepare(context: LaunchSessionContext, launch: {
+    model?: string;
+    /** Plugins whose options a subagent repeats unchanged from its orchestrator of the same CLI: the person chose them. */
+    personChosen?: readonly string[];
+  } = {}): Promise<PreparedLaunch> {
     const contributors = new Map(this.dependencies.contributors().map((contributor) => [contributor.pluginId, contributor]));
+    // Only a subagent's launch is refused for a second model; it was chosen for it, not by the person.
+    const delegated = context.role === "subagent";
     const selected = Object.keys(context.options).sort();
     for (const pluginId of selected) {
       if (!contributors.has(pluginId)) return { ok: false, reason: unavailableReason(pluginId) };
@@ -251,7 +285,7 @@ export class LaunchPipeline {
       ...policies.map((contributor) => this.ask(contributor, context, false))
     ]);
     const runDirectory = join(this.dependencies.runsRoot, safeSegment(context.sessionId), randomUUID());
-    const cleanup = (): Promise<void> => rm(runDirectory, { recursive: true, force: true });
+    const cleanup = (): Promise<void> => removeRunFolder(runDirectory);
     const refuse = async (reason: string): Promise<PreparedLaunch> => {
       await cleanup().catch(() => undefined);
       return { ok: false, reason };
@@ -267,6 +301,9 @@ export class LaunchPipeline {
     const args: string[] = [];
     const secrets: string[] = [];
     let thirdPartyModel = false;
+    const modelSetBy: string[] = [];
+    let modelArgumentBy: string | undefined;
+    let accountHome: string | undefined;
     for (const answer of answers) {
       if ("refuse" in answer) return refuse(answer.refuse);
     }
@@ -285,6 +322,25 @@ export class LaunchPipeline {
       }
       const forbidden = contributedArgs.find((argument) => coreOwnedLaunchArgument(context.provider, argument));
       if (forbidden) return refuse(`${name} added ${forbidden.slice(0, 60)}, which only CanvasTTY may pass.`);
+      const modelFlag = modelSettingArgument(context.provider, contributedArgs);
+      if (modelFlag !== undefined || contribution.thirdPartyModel === true) {
+        // A subagent's CLI gets one model: a second model flag next to the launch's own made OpenCode crash at start.
+        // A person's own launches (and their restores) start as they always did; the person chose both.
+        if (delegated && modelFlag !== undefined && launch.model !== undefined) {
+          return refuse(`${name} sets the model of this launch, and the launch also asks for model ${launch.model.slice(0, 80)}. Omit the model, or choose ${name} options for the model you want.`);
+        }
+        // Two model arguments are the crash, and are always refused. A plugin that only routes the CLI to another service
+        // (thirdPartyModel) next to one that picks the model is allowed only as the person's own combination: both
+        // repeated unchanged from the orchestrator. A pair the orchestrator chose itself is refused as before.
+        const personPair = !(modelFlag !== undefined && modelArgumentBy !== undefined)
+          && [...modelSetBy, contributor.pluginId].every((pluginId) => launch.personChosen?.includes(pluginId) === true);
+        if (delegated && modelSetBy.length > 0 && !personPair) {
+          const first = contributors.get(modelSetBy[0]!)?.pluginName ?? modelSetBy[0]!;
+          return refuse(`${first} and ${name} both set the model of this launch; choose launch options from only one of them.`);
+        }
+        if (modelFlag !== undefined) modelArgumentBy ??= contributor.pluginId;
+        modelSetBy.push(contributor.pluginId);
+      }
       // A configuration a CLI reads from the environment or an argument may not decide its approvals either.
       const widening = permissionConfigProblem(context.provider, contribution, context.cwd);
       if (widening) return refuse(`${name} ${widening}`);
@@ -313,6 +369,7 @@ export class LaunchPipeline {
         const conflict = claim(key);
         if (conflict) return refuse(conflict);
         env[key] = expand(value);
+        accountHome = contributedHookHome(context.provider, contributor.dataDir, key, env[key]) ?? accountHome;
       }
       for (const [key, secretKey] of Object.entries(contribution.secretEnv)) {
         const conflict = claim(key);
@@ -327,7 +384,7 @@ export class LaunchPipeline {
       args.push(...contributedArgs.map(expand));
       thirdPartyModel ||= contribution.thirdPartyModel === true;
     }
-    return { ok: true, env, args, secrets, envSources, thirdPartyModel, cleanup };
+    return { ok: true, env, args, secrets, envSources, thirdPartyModel, modelSetBy, runDirectory, ...(accountHome ? { accountHome } : {}), cleanup };
   }
 
   private async ask(
@@ -562,6 +619,17 @@ function unavailableReason(pluginId: string): string {
   return `Needs plugin ${pluginId} for its launch options; it is disabled, removed, or its native code is not trusted.`;
 }
 
+/** Errors that leave a card's folder in place on purpose: another run still uses it, or it is gone. */
+const PARENT_KEPT_CODES = new Set(["ENOENT", "ENOTEMPTY", "EEXIST", "EBUSY", "EPERM"]);
+
+/**
+ * Removes one run folder. On Windows the agent may still hold a file there for a moment after its PTY exits, so
+ * busy/permission errors are retried briefly before giving up.
+ */
+function removeRunFolder(runDirectory: string): Promise<void> {
+  return rm(runDirectory, { recursive: true, force: true, maxRetries: 5, retryDelay: 100 });
+}
+
 function safeSegment(value: string): string {
   return value.replace(/[^A-Za-z0-9._-]/g, "_").replace(/^\.+/, "_").slice(0, 128) || "_";
 }
@@ -575,3 +643,92 @@ export function errorText(error: unknown): string {
 export function isRecord(value: unknown): value is Record<string, unknown> {
   return Boolean(value && typeof value === "object" && !Array.isArray(value));
 }
+
+/** A plugin's launch values checked against its declared fields, with defaults filled in. Throws with a person-readable reason. */
+function launchValues(contributor: LaunchContributor, provider: ProviderId, raw: unknown): PluginLaunchValues {
+  const { appliesTo, fields } = contributor.launch;
+  if (appliesTo && !appliesTo.includes(provider as never)) {
+    throw new Error(`${contributor.pluginName} launch options do not apply to ${provider}.`);
+  }
+  if (!isRecord(raw)) throw new Error(`${contributor.pluginName} launch options are invalid.`);
+  const known = new Set(fields.map((field) => field.key));
+  const unknown = Object.keys(raw).find((key) => !known.has(key));
+  if (unknown) throw new Error(`${contributor.pluginName} has no launch option ${unknown.slice(0, 40)}.`);
+  const values: PluginLaunchValues = {};
+  for (const field of fields) {
+    const value = raw[field.key] ?? field.default
+      ?? (field.kind === "boolean" ? false : field.kind === "select" ? field.options?.[0]?.value ?? "" : "");
+    const plainText = (limit: number): boolean => typeof value === "string" && value.length <= limit && !/[\u0000-\u001f\u007f]/.test(value);
+    // A service-provided choice may have changed since the launcher showed it: the service checks it when it prepares.
+    const valid = field.kind === "boolean"
+      ? typeof value === "boolean"
+      : field.kind === "select"
+        ? typeof value === "string" && (Boolean(field.options?.some((option) => option.value === value)) || (field.optionsFrom === "service" && plainText(200)))
+        : plainText(field.maxLength ?? 200);
+    if (!valid) throw new Error(`${contributor.pluginName} launch option ${field.label} is invalid.`);
+    values[field.key] = value as boolean | string;
+  }
+  if (Buffer.byteLength(JSON.stringify(values), "utf8") > MAX_PLUGIN_SLOT_BYTES) {
+    throw new Error(`${contributor.pluginName} launch options exceed 4 KB.`);
+  }
+  return values;
+}
+
+/** The orchestrator's own choice repeated for its subagent: only keys it chose, each with the same value. */
+function sameAsInherited(raw: unknown, inherited: Readonly<PluginLaunchValues> | undefined): boolean {
+  if (!inherited || !isRecord(raw)) return false;
+  const keys = Object.keys(raw);
+  // An empty choice repeats only an empty one (a plugin without fields, chosen in the launcher).
+  if (keys.length === 0) return Object.keys(inherited).length === 0;
+  return keys.every((key) => Object.hasOwn(inherited, key) && inherited[key] === raw[key]);
+}
+
+/** Exactly the same keys and values. */
+function sameValues(values: Readonly<PluginLaunchValues>, inherited: Readonly<PluginLaunchValues>): boolean {
+  const keys = Object.keys(values);
+  return keys.length === Object.keys(inherited).length
+    && keys.every((key) => Object.hasOwn(inherited, key) && inherited[key] === values[key]);
+}
+
+/**
+ * How each CLI's arguments choose its model. Every CLI here takes `--model`; `short` are its short model flags,
+ * `profile` flags that select a configuration profile (which may name a model), and `config` flags that set a
+ * configuration key, with the keys that decide the model.
+ */
+const MODEL_ARGUMENTS: Readonly<Partial<Record<ProviderId, { short?: readonly string[]; profile?: readonly string[]; config?: { flags: readonly string[]; keys: readonly string[] } }>>> = {
+  codex: { short: ["-m"], profile: ["--profile", "-p"], config: { flags: ["-c", "--config"], keys: ["model", "model_provider", "profile"] } },
+  opencode: { short: ["-m"] },
+  qwen: { short: ["-m"] },
+  kimi: { short: ["-m"] },
+  grok: { short: ["-m"] }
+};
+
+/**
+ * The first argument that chooses the CLI's model, in any form: `--model x`, `--model=x`, a short `-m x` or `-mx`,
+ * and for Codex `--profile`/`-p` and `-c`/`--config` setting `model`, `model_provider` or `profile`. Undefined when none does.
+ */
+export function modelSettingArgument(provider: ProviderId, args: readonly string[]): string | undefined {
+  const table = MODEL_ARGUMENTS[provider] ?? {};
+  const valueFlags = ["--model", ...(table.short ?? []), ...(table.profile ?? [])];
+  const configKeys = new Set(table.config?.keys ?? []);
+  const configSetsModel = (value: string | undefined): boolean => {
+    if (value === undefined) return false;
+    const key = value.split("=", 1)[0]!.trim().replace(/^["']|["']$/gu, "");
+    return value.includes("=") && configKeys.has(key);
+  };
+  for (let index = 0; index < args.length; index++) {
+    const argument = args[index]!;
+    if (valueFlags.some((flag) => argument === flag || attachedValue(argument, flag) !== undefined)) return argument;
+    for (const flag of table.config?.flags ?? []) {
+      const value = argument === flag ? args[index + 1] : attachedValue(argument, flag);
+      if (configSetsModel(value)) return argument === flag ? `${argument} ${value}` : argument;
+    }
+  }
+  return undefined;
+}
+
+/** The value written into the flag's own argument: `--flag=value`, or `-fvalue` for a short flag. */
+function attachedValue(argument: string, flag: string): string | undefined {
+  if (flag.startsWith("--")) return argument.startsWith(`${flag}=`) ? argument.slice(flag.length + 1) : undefined;
+  return argument.length > flag.length && argument.startsWith(flag) ? argument.slice(flag.length) : undefined;
+}
diff --git a/src/main/services/PluginManager.ts b/src/main/services/PluginManager.ts
index 0e86e7192..f68238a79 100644
--- a/src/main/services/PluginManager.ts
+++ b/src/main/services/PluginManager.ts
@@ -48,6 +48,7 @@ import {
 } from "../../shared/contracts.ts";
 import { isValidSemver } from "../../shared/hostVersion.ts";
 import { PLUGIN_VISIBILITY_BRIDGE_SOURCE } from "../../shared/pluginVisibility.ts";
+import { PLUGIN_NATIVE_CODE_OFF_ERROR, PLUGIN_SERVICE_MODULE_OFF_ERROR } from "../../shared/pluginServiceNotice.ts";
 import type { PluginServiceSpec } from "./PluginServiceSupervisor.ts";
 import type { LaunchContributor } from "./LaunchPipeline.ts";
 import type { EnvironmentProvider } from "./EnvironmentRegistry.ts";
@@ -654,7 +655,8 @@ export class PluginManager {
         pluginName: manifest.name,
         serviceId: service.id,
         launch: structuredClone(service.launch),
-        secrets: manifest.permissions.includes("secrets")
+        secrets: manifest.permissions.includes("secrets"),
+        dataDir: join(this.dataRoot, plugin.manifest.id)
       });
     }
     return contributors;
@@ -1203,11 +1205,21 @@ export class PluginManager {
     return structuredClone(contribution);
   }
 
+  /**
+   * The plugin's own UI may ask `serviceId` now: it is in the active manifest and its native code is trusted. A
+   * service whose module is off or whose native code is not trusted gets a fixed text the renderer turns into a
+   * localized notice with a way to the right settings section (pluginServiceNotice).
+   */
   assertService(pluginId: string, serviceId: string): void {
-    const plugin = activePlugin(this.requireEnabledPlugin(pluginId));
+    const installed = this.requireEnabledPlugin(pluginId);
+    const plugin = activePlugin(installed);
     if (!plugin.manifest.services?.some((service) => service.id === serviceId)) {
+      if (installed.manifest.services?.some((service) => service.id === serviceId)) throw new Error(PLUGIN_SERVICE_MODULE_OFF_ERROR);
       throw new Error("Plugin service does not exist.");
     }
+    if (!installed.nativeCodeTrusted || !this.serviceTrust.get(pluginId)?.[serviceId]) {
+      throw new Error(PLUGIN_NATIVE_CODE_OFF_ERROR);
+    }
   }
 
   hasPermission(pluginId: string, permission: PluginPermission): boolean {
diff --git a/src/main/services/PluginServiceSupervisor.ts b/src/main/services/PluginServiceSupervisor.ts
index 1c1b061dc..1236d301b 100644
--- a/src/main/services/PluginServiceSupervisor.ts
+++ b/src/main/services/PluginServiceSupervisor.ts
@@ -288,7 +288,8 @@ export class PluginServiceSupervisor {
   private send(pluginId: string, serviceId: string, method: string, params: unknown, timeoutMs: number): Promise<unknown> {
     const record = this.services.get(serviceKey(pluginId, serviceId));
     if (!record || !record.child || (record.state !== "running" && record.state !== "starting")) {
-      return Promise.reject(new Error("Plugin service is not running."));
+      // A trusted service that crashed, failed to start or is waiting to restart says why.
+      return Promise.reject(new Error(`Plugin service is not running.${record?.lastError ? ` Last error: ${record.lastError}` : ""}`));
     }
     if (record.pending.size >= MAX_PENDING_REQUESTS) {
       return Promise.reject(new Error("Plugin service is busy."));
diff --git a/src/main/services/TerminalManager.ts b/src/main/services/TerminalManager.ts
index 7688dcc1e..5aafb474c 100644
--- a/src/main/services/TerminalManager.ts
+++ b/src/main/services/TerminalManager.ts
@@ -1,4 +1,5 @@
 import { randomUUID } from "node:crypto";
+import { constants as osConstants } from "node:os";
 import { realpathSync, statSync } from "node:fs";
 import { basename, dirname } from "node:path";
 import { isPathInside } from "../../agent-runtime/path-inside.mjs";
@@ -8,6 +9,7 @@ import type {
   CreateSessionRequest,
   GitRiskReport,
   Point,
+  PluginLaunchValues,
   ProviderId,
   SessionBounds,
   SessionEnvironmentChoice,
@@ -53,6 +55,7 @@ import {
   type ControlConnection
 } from "./agent-control/controlCapabilities.ts";
 import { codexTrustArguments, mergeOpenCodeLaunchEnvironment } from "./agent-runtime/ProviderRuntimeLaunch.ts";
+import { accountHookHome } from "./accountHomes.ts";
 import { openCodeProjectFolderEnvironment } from "./openCodeConfig.ts";
 import { onDiskPath } from "./onDiskPath.ts";
 import { RESULT_CAPTURE_PROVIDERS } from "./resultCapture.ts";
@@ -72,6 +75,7 @@ import { LaunchRefusal } from "./launchRefusal.ts";
 import { configuredMode } from "./configuredMode.ts";
 import { envKey, RESERVED_ENV, type LaunchPipeline, type PreparedLaunch } from "./LaunchPipeline.ts";
 import type { EnvironmentRegistry } from "./EnvironmentRegistry.ts";
+import { agentEnvironment, shellEnvironment } from "./ambientProviderKeys.ts";
 import {
   persistedTerminalSession,
   type PersistedEnvironmentRef,
@@ -128,6 +132,20 @@ interface ManagedSession {
   promptTurnMark?: number;
   /** The last turn's final answer its hook or plugin reported (captureResult only); cleared when a turn starts. */
   answer?: { text: string; truncated: boolean; at: number };
+  /**
+   * Status hooks were off when this launch was planned, or were turned off since: it has no lifecycle hooks, so its
+   * card stays unavailable and no answer is captured until a restart with status hooks on.
+   */
+  statusHooksOff?: boolean;
+  /** Something other than Enter was typed since the last submitted input (an Enter alone then submits nothing). */
+  typedSinceSubmit?: boolean;
+  /** Bumped by every submitted input (except an approval answer) and every launch: the turn an answer must belong to. */
+  inputGeneration?: number;
+  /** inputGeneration when the current turn started. */
+  answerTurnGeneration?: number;
+  /** OpenCode's current turn id and the inputGeneration each recent turn id started in. */
+  providerTurnId?: string;
+  providerTurnGenerations?: Map<string, number>;
   /**
    * Plugin options, environment ref (or, until the plugin has prepared it, the launcher's environment choice)
    * and owning plugin carried into the saved record.
@@ -137,10 +155,21 @@ interface ManagedSession {
   launchToken: number;
   /** Removes the current run's plugin files; called when the process exits. */
   launchCleanup: (() => Promise<void>) | null;
+  /** Cleanup already started by an earlier exit; closing the card must drain it before deleting its parent. */
+  launchCleanupPending?: Promise<void>;
+  /** Whole contributed launches, including asynchronous preparation/wrapping and abandoned-run cleanup. */
+  launchTasks?: Set<Promise<unknown>>;
+  /**
+   * The pipelines that prepared launches for this card, each with the run folders it created for this card object,
+   * retained across host reconfiguration/restart. Closing removes only these, never a later card's under the same id.
+   */
+  launchRunFolders?: Map<Pick<LaunchPipeline, "forgetSession">, Set<string>>;
   /** A restored grok card waits for its grid before launching; plugins still learn it is a restore. */
   restoringLaunch: boolean;
   /** The environment was prepared or resumed in this run of the app, so it can be wrapped now. */
   environmentReady: boolean;
+  /** Plugins whose launch contribution decided this run's model (LaunchPipeline.prepare), once prepared. */
+  modelSetBy?: string[];
   /** Bumped by every launch the person or the app asks for (create, restart, restore); input waits for one. */
   launchEpoch: number;
   /** An automation-owned paste-only draft may be extended in this launch; submit or person input clears it. */
@@ -157,6 +186,17 @@ interface ManagedSession {
   hookSignals?: number;
   /** Set after the person answered a hooked Claude prompt; see settleAnsweredPrompt. */
   answeredPromptTimer?: ReturnType<typeof setTimeout>;
+  /** The last deliverInput of this card: the next one waits for it, so sends arrive in order. */
+  inputQueue?: Promise<InputDelivery>;
+  /** This launch's CLI turned its text cursor on in its full-screen view: its input field is up (watchInputReadiness). */
+  inputReady?: boolean;
+  /** While watching for inputReady: this launch's CLI is on its alternate (full-screen) buffer. */
+  alternateScreen?: boolean;
+  /** While watching for inputReady: an escape sequence cut off at the end of the last output chunk. */
+  readinessCarry?: string;
+  /** When this launch's first lifecycle hook stops waiting for inputReady (INPUT_READY_HOOK_GRACE_MS after it). */
+  hookReadyAt?: number;
+  hookReadyTimer?: ReturnType<typeof setTimeout>;
 }
 
 type EnvironmentService = Pick<EnvironmentRegistry,
@@ -197,6 +237,30 @@ export const PTY_KILL_WAIT_MS = 1_000;
 /** Longer than every plugin step of a launch together (prepare, resume, launch options, wrap). */
 export const LAUNCH_INPUT_WAIT_MS = 60_000;
 
+/**
+ * CLIs that drop what is typed before their input field is up. OpenCode's full-screen view throws away every key that
+ * arrives while it starts (a few seconds), so a first prompt typed when the PTY exists is lost and the agent waits at
+ * an empty prompt. Its readiness is structured: a lifecycle hook, or the terminal control sequence that shows its text
+ * cursor on the alternate screen once its prompt has focus (never the human-readable screen text).
+ */
+const WAITS_FOR_INPUT_READY: ReadonlySet<ProviderId> = new Set<ProviderId>(["opencode"]);
+
+/**
+ * Whether input for this CLI waits for its input field. Only on macOS and Linux: Windows' pseudo console (ConPTY,
+ * winpty) may not pass the alternate-screen and cursor sequences through, so there the gate could never open, and
+ * input is typed as soon as the PTY exists, as before.
+ */
+function waitsForInputReady(provider: ProviderId): boolean {
+  return (process.platform === "darwin" || process.platform === "linux") && WAITS_FOR_INPUT_READY.has(provider);
+}
+
+/**
+ * A lifecycle hook alone is not proof that the input field is up: OpenCode's plugin reports (for a resumed session,
+ * idle on session.updated) about 0.6 s before typed keys reach its prompt. The cursor signal decides; a hook counts
+ * only this long after the launch's first one, for a CLI whose cursor sequence never arrives.
+ */
+export const INPUT_READY_HOOK_GRACE_MS = 2_000;
+
 /** What happened to input handed to deliverInput. */
 export type InputDelivery = { delivered: true } | { delivered: false; reason: string };
 
@@ -660,8 +724,15 @@ export class TerminalManager {
     }
 
     const delegated = origin === "subagent" || control.ownerPluginId !== undefined;
+    // The orchestrator's own options are the person's choice: its subagent may repeat them even where the plugin did
+    // not declare its options delegable. Any other choice needs the plugin's delegable declaration.
+    // Only a subagent of the same CLI: the person chose them for that CLI (as AgentControlService.subagentLaunchOptions).
+    const parent = origin === "subagent" && request.parentSessionId !== undefined ? this.sessions.get(request.parentSessionId) : undefined;
+    const parentOptions = parent && parent.metadata.provider === request.provider
+      ? parent.extras.options as Record<string, PluginLaunchValues> | undefined : undefined;
     const launchOptions = this.launchPipeline
-      ? this.launchPipeline.normalizeOptions(request.provider, request.launchOptions, origin === "subagent" ? { delegated: true } : undefined)
+      ? this.launchPipeline.normalizeOptions(request.provider, request.launchOptions, origin === "subagent"
+        ? { delegated: true, ...(parentOptions ? { inherited: parentOptions } : {}) } : undefined)
       : request.launchOptions === undefined ? undefined : failWith("Plugin launch options are not available.");
     // The isolation layer decides before anything starts: a subagent without it runs in normal, a launch that
     // needs it and cannot have it is refused.
@@ -902,18 +973,42 @@ export class TerminalManager {
    * The one delivery rule for text another agent, a plugin or a controller sends to a card (spawn_agent's first
    * prompt, send_to_agent, plugin sessions.send). A running card gets it at once. A card whose launch plugins are
    * still preparing (launch options, a launch policy, an environment) or that waits for its grid gets it exactly
-   * once, when that launch has started. A launch that is refused, fails, is cancelled or superseded (closed,
-   * restarted), or does not start within LAUNCH_INPUT_WAIT_MS delivers nothing, says why, and drops the text:
-   * it never reaches a later launch of the card. An aborted `signal` (the sender cancelled) delivers nothing either.
-   * Automated Codex prompts additionally wait for an empty task composer in this launch; a PTY alone is not ready.
+   * once, when that launch has started. A CLI that drops keys typed while it starts (WAITS_FOR_INPUT_READY) gets it
+   * only once its input field is up; automated Codex prompts (`waitForCodexComposer`) wait for an empty task composer
+   * in this launch, since a PTY alone is not ready. A launch that is refused, fails, is cancelled or superseded
+   * (closed, restarted), or does not start or become ready within `waitMs` delivers nothing, says why, and drops the
+   * text: it never reaches a later launch of the card. An aborted `signal` (the sender cancelled) delivers nothing
+   * either. Sends to one card are delivered in the order they were made; the wait of each counts from its own call.
    */
-  async deliverInput(id: string, data: string, waitMs = LAUNCH_INPUT_WAIT_MS, signal?: AbortSignal, waitForCodexComposer = false): Promise<InputDelivery> {
-    const canceled: InputDelivery = { delivered: false, reason: "The delivery was cancelled." };
-    if (signal?.aborted) return canceled;
+  deliverInput(id: string, data: string, waitMs = LAUNCH_INPUT_WAIT_MS, signal?: AbortSignal, waitForCodexComposer = false): Promise<InputDelivery> {
     const session = this.sessions.get(id);
-    if (!session) return { delivered: false, reason: "The session does not exist." };
+    if (!session) return Promise.resolve({ delivered: false, reason: "The session does not exist." });
+    // The launch this call was made for: a send queued behind another never reaches a later launch either.
     const epoch = session.launchEpoch;
     const deadline = Date.now() + waitMs;
+    const previous = session.inputQueue;
+    const run = (): Promise<InputDelivery> =>
+      this.deliverQueuedInput(id, session, epoch, data, waitMs, deadline, signal, waitForCodexComposer);
+    const pending = previous ? previous.then(run, run) : run();
+    session.inputQueue = pending;
+    void pending.finally(() => { if (session.inputQueue === pending) delete session.inputQueue; }).catch(() => undefined);
+    if (!signal || !previous) return pending;
+    // A send cancelled while it waits behind another settles at once; its turn in the queue then types nothing.
+    const canceled: InputDelivery = { delivered: false, reason: "The delivery was cancelled." };
+    if (signal.aborted) return Promise.resolve(canceled);
+    return new Promise<InputDelivery>((resolve) => {
+      const abort = (): void => resolve(canceled);
+      signal.addEventListener("abort", abort, { once: true });
+      void pending.then(resolve, () => resolve(canceled)).finally(() => signal.removeEventListener("abort", abort));
+    });
+  }
+
+  private async deliverQueuedInput(
+    id: string, session: ManagedSession, epoch: number, data: string, waitMs: number, deadline: number,
+    signal?: AbortSignal, waitForCodexComposer = false
+  ): Promise<InputDelivery> {
+    const canceled: InputDelivery = { delivered: false, reason: "The delivery was cancelled." };
+    if (signal?.aborted) return canceled;
     const launchOutput = (): TerminalBufferSnapshot => {
       const snapshot = this.readBuffer(id);
       const start = snapshot.outputOffset - snapshot.buffer.length;
@@ -922,31 +1017,44 @@ export class TerminalManager {
     // Build the screen only after this launch has a PTY: a pending restart
     // still has its previous process's scrollback until bindProcess records the new offset.
     let screen: SessionScreen | null = null;
-    let inputReady = !waitForCodexComposer || session.codexDraftEpoch === epoch;
+    // Per call, unlike session.inputReady (OpenCode's cursor signal): checked when this send's turn in the queue
+    // comes, so an automation draft extended by the send before it counts.
+    let codexComposerSeen = !waitForCodexComposer || session.codexDraftEpoch === epoch;
+    // What the text still waits for: the launch to start, or (once its PTY exists) the CLI's input field.
+    const awaiting = (): "launch" | "input" | null => {
+      if (this.sessions.get(id) !== session || session.launchEpoch !== epoch || session.metadata.exitCode !== null) return null;
+      if (!session.process) return "launch";
+      return cliAcceptsInput(session) && codexComposerSeen ? null : "input";
+    };
     try {
-      const waiting = (): boolean => this.sessions.get(id) === session && session.launchEpoch === epoch
-        && session.metadata.exitCode === null && (!session.process || !inputReady);
-      while (waiting() && !signal?.aborted) {
-        if (session.process && !inputReady) {
+      for (let stage = awaiting(); stage && !signal?.aborted; stage = awaiting()) {
+        const composerPending = stage === "input" && !codexComposerSeen;
+        if (composerPending) {
           screen ??= new SessionScreen({ geometry: (id) => this.geometry(id), readBuffer: launchOutput }, id);
           const snapshot = launchOutput();
           screen.feed({ id, data: snapshot.buffer, outputOffset: snapshot.outputOffset });
           if (await screen.inspect((terminal) => codexComposerReady(
             terminalScreenText(terminal, true), codexPlaceholderIsDim(terminal), codexContinuationIsEmpty(terminal)))) {
-            inputReady = true;
-            break;
+            codexComposerSeen = true;
+            continue;
           }
         }
         const remaining = deadline - Date.now();
         if (remaining <= 0) {
-          if (session.process && !inputReady) return { delivered: false, reason: `Codex did not become ready for input within ${Math.round(waitMs / 1000)} s.` };
-          return { delivered: false, reason: `The session did not start within ${Math.round(waitMs / 1000)} s.` };
+          // A caller that does not wait (0) is told the state, not a wait it never made.
+          const within = waitMs <= 0 ? "yet" : waitMs >= 1_000 ? `within ${Math.round(waitMs / 1_000)} s` : `within ${waitMs} ms`;
+          return { delivered: false, reason: stage === "launch"
+            ? `The session did not start ${within}.`
+            : composerPending
+              ? `Codex did not become ready for input ${within}.`
+              : `The agent's CLI did not show its input field ${within}; nothing was typed. Look at the card, then send the prompt again.` };
         }
         await new Promise<void>((resolve) => {
-          const timer = setTimeout(wake, session.process && !inputReady ? Math.min(100, remaining) : remaining);
+          // Codex draws its composer without a structured signal: its screen is polled while the deadline runs.
+          const timer = setTimeout(wake, composerPending ? Math.min(100, remaining) : remaining);
           function wake(): void {
             clearTimeout(timer);
-            session!.launchWaiters.delete(wake);
+            session.launchWaiters.delete(wake);
             signal?.removeEventListener("abort", wake);
             resolve();
           }
@@ -974,6 +1082,23 @@ export class TerminalManager {
     }
   }
 
+  /**
+   * Submitted text starts a new task (an approval answer continues the current turn), so an earlier answer is not its.
+   * An Enter with nothing typed before it in an idle card submits nothing: the stored answer stays.
+   */
+  private noteSubmittedInput(session: ManagedSession, data: string): void {
+    const typed = /[^\r\n]/u.test(data);
+    if (!data.endsWith("\r")) {
+      if (typed) session.typedSinceSubmit = true;
+      return;
+    }
+    const empty = !typed && !session.typedSinceSubmit;
+    session.typedSinceSubmit = false;
+    if (empty && session.metadata.status === "idle") return;
+    if (session.metadata.status !== "needs_approval") session.inputGeneration = (session.inputGeneration ?? 0) + 1;
+    delete session.answer;
+  }
+
   private wakeLaunchWaiters(session: ManagedSession): void {
     for (const wake of [...session.launchWaiters]) wake();
   }
@@ -982,9 +1107,19 @@ export class TerminalManager {
     if (typeof data !== "string" || data.length === 0) return false;
     const session = this.sessions.get(id);
     if (!session || session.metadata.exitCode !== null || !session.process) return false;
-    const process = session.process;
+    return this.writeInput(id, session, session.process, data);
+  }
+
+  /**
+   * The one write of person or agent input to a card's CLI, and what it means for turns and answers. Everything that
+   * types into a card goes through here: input(), deliverInput, and any writer that holds input and replays it later,
+   * so a replayed Enter is a submitted prompt like a typed one. Nothing is written to a process the card no longer runs.
+   */
+  private writeInput(id: string, session: ManagedSession, process: IPty, data: string): boolean {
+    if (this.sessions.get(id) !== session || session.process !== process || session.metadata.exitCode !== null) return false;
     const written = tryPtyOperation(() => process.write(data));
     if (written) delete session.codexDraftEpoch;
+    if (written) this.noteSubmittedInput(session, data);
     if (written && ANSWERS_PROMPT.test(data)) this.settleAnsweredPrompt(id, session);
     return written;
   }
@@ -1055,11 +1190,53 @@ export class TerminalManager {
     return publicSessionMetadata(session);
   }
 
-  /** `source` "hook" is the agent's own lifecycle hook (through the runtime gateway); "title" is its terminal title. */
-  applyProviderSignal(id: string, signal: ProviderLifecycleSignal, source: "hook" | "title" = "hook"): void {
+  /**
+   * `source` "hook" is the agent's own lifecycle hook (through the runtime gateway); "title" is its terminal title.
+   * Returns whether the signal belongs to the current turn and the last submitted input, so the answer it carries may
+   * be kept (applyRuntimeSignal). With status hooks off for this launch nothing applies and no answer is kept.
+   */
+  applyProviderSignal(id: string, signal: ProviderLifecycleSignal, source: "hook" | "title" = "hook"): boolean {
     const session = this.sessions.get(id);
-    if (!this.lifecycleHooksEnabled || !session || session.metadata.status === "done" || session.metadata.status === "failed") return;
-    if (source === "hook") session.hookSignals = (session.hookSignals ?? 0) + 1;
+    if (!this.lifecycleHooksEnabled || !session || session.statusHooksOff) return false;
+    // A result-capturing OpenCode names each turn. A signal of an earlier turn is stale and changes nothing; a signal
+    // of the current turn that started before the last submitted input still moves the status (so a wait ends), but
+    // is not that input's turn: it neither counts as a turn start nor brings an answer.
+    const exited = session.metadata.status === "done" || session.metadata.status === "failed";
+    let current = true;
+    let startsProviderTurn = false;
+    if (source === "hook" && session.captureResult && session.metadata.provider === "opencode" && signal.requestId) {
+      const turns = session.providerTurnGenerations ??= new Map();
+      const generation = session.inputGeneration ?? 0;
+      const known = turns.get(signal.requestId);
+      if (known === undefined) {
+        // A turn id not seen yet opens the next turn, also when its start was lost (as the gateway counts it); after
+        // the exit no turn opens.
+        if (exited) return false;
+        startsProviderTurn = true;
+        turns.set(signal.requestId, generation);
+        session.providerTurnId = signal.requestId;
+        while (turns.size > 64) turns.delete(turns.keys().next().value!);
+      } else if (session.providerTurnId !== signal.requestId) {
+        return false;
+      } else {
+        current = known === generation;
+      }
+    }
+    // The process already exited: its status stays, but the answer its last turn reported as it ended still counts.
+    if (exited) return current;
+    if (source === "hook") {
+      session.hookSignals = (session.hookSignals ?? 0) + 1;
+      // The CLI runs, but its input field may not be up yet: input waiting for it (deliverInput) goes on the cursor
+      // signal, or once the grace after this first hook has passed.
+      if (session.hookSignals === 1 && waitsForInputReady(session.metadata.provider) && !session.inputReady) {
+        session.hookReadyAt = Date.now() + INPUT_READY_HOOK_GRACE_MS;
+        session.hookReadyTimer = setTimeout(() => {
+          session.hookReadyTimer = undefined;
+          this.wakeLaunchWaiters(session);
+        }, INPUT_READY_HOOK_GRACE_MS);
+        session.hookReadyTimer.unref?.();
+      }
+    }
 
     const threadId = normalizeThreadId(session.metadata.provider, signal.threadId);
     const threadChanged = Boolean(threadId && threadId !== session.threadId);
@@ -1070,24 +1247,44 @@ export class TerminalManager {
 
     const nextStatus = signal.state;
     // A turn starts when the agent moves to working; wait_for_agent compares it with the last delivered prompt.
-    if (nextStatus === "working" && session.metadata.status !== "working") session.turnStarts = (session.turnStarts ?? 0) + 1;
-    // A new turn: the previous answer is no longer this turn's.
-    if (nextStatus === "working") session.answer = undefined;
+    if (current && nextStatus === "working") {
+      if (startsProviderTurn || session.metadata.status !== "working") session.turnStarts = (session.turnStarts ?? 0) + 1;
+      // A new turn: the previous answer is no longer this turn's.
+      session.answer = undefined;
+      session.answerTurnGeneration = session.inputGeneration ?? 0;
+    }
     const completed = nextStatus === "idle" && ["Stop", "StopFailure", "StopCancelled"].includes(signal.event ?? "");
     const nextTurnCompleted = nextStatus === "working" ? false : completed || Boolean(session.metadata.turnCompleted);
-    if (!threadChanged && session.metadata.status === nextStatus && Boolean(session.metadata.turnCompleted) === nextTurnCompleted) return;
+    if (!threadChanged && session.metadata.status === nextStatus && Boolean(session.metadata.turnCompleted) === nextTurnCompleted) return current;
     session.metadata.status = nextStatus;
     session.metadata.turnCompleted = nextTurnCompleted;
     this.emitSession(session.metadata);
+    return current;
+  }
+
+  /**
+   * Why a result-capturing session's final answer cannot be captured, or null when it can (or the session does not
+   * capture one). With CanvasTTY status hooks off for its launch it gets no hooks, so nothing reports the answer.
+   */
+  answerUnavailable(id: string): "status_hooks_off" | null {
+    const session = this.sessions.get(id);
+    if (!session?.captureResult) return null;
+    return !this.lifecycleHooksEnabled || session.statusHooksOff ? "status_hooks_off" : null;
   }
 
   /**
    * Keeps the final answer a result-capturing session reported with its turn's end (the Codex Stop hook, the OpenCode
-   * plugin's session.idle); read back masked by answer(). In memory only, never saved.
+   * plugin's session.idle), only while it belongs to the current turn and the last submitted input; read back masked
+   * by answer(). OpenCode passes its turn id. In memory only, never saved.
    */
-  recordAnswer(id: string, result: { text: string; truncated: boolean }): void {
+  recordAnswer(id: string, result: { text: string; truncated: boolean }, correlation?: { turnId?: string | null }): void {
     const session = this.sessions.get(id);
     if (!session?.captureResult || typeof result?.text !== "string") return;
+    const generation = session.inputGeneration ?? 0;
+    if (session.metadata.provider === "opencode" && correlation && "turnId" in correlation) {
+      if (!correlation.turnId || correlation.turnId !== session.providerTurnId
+        || session.providerTurnGenerations?.get(correlation.turnId) !== generation) return;
+    } else if (session.promptTurnMark !== undefined && session.answerTurnGeneration !== generation) return;
     session.answer = { text: result.text, truncated: result.truncated === true, at: Date.now() };
   }
 
@@ -1105,6 +1302,7 @@ export class TerminalManager {
     if (next) return;
     for (const session of this.sessions.values()) {
       session.lifecycle = null;
+      session.statusHooksOff = true;
       if (
         session.metadata.provider === "terminal"
         || session.metadata.status === "done"
@@ -1164,6 +1362,37 @@ export class TerminalManager {
     }
   }
 
+  /** Serializes this card's run cleanup across exit/restart; callbacks retain their specific launch ownership. */
+  private cleanupLaunchFiles(session: ManagedSession): Promise<void> {
+    const cleanup = session.launchCleanup;
+    session.launchCleanup = null;
+    const previous = session.launchCleanupPending;
+    if (!cleanup) return previous ?? Promise.resolve();
+    const pending = (async (): Promise<void> => {
+      if (previous) await previous.catch(() => undefined);
+      await cleanup();
+    })();
+    session.launchCleanupPending = pending;
+    void pending.finally(() => {
+      if (session.launchCleanupPending === pending) delete session.launchCleanupPending;
+    }).catch(() => undefined);
+    return pending;
+  }
+
+  /** Closing a card waits for producers of run files, not just the current process's adopted cleanup. */
+  private async forgetLaunchFiles(session: ManagedSession): Promise<void> {
+    while (session.launchTasks?.size) await Promise.allSettled([...session.launchTasks]);
+    try { await this.cleanupLaunchFiles(session); }
+    finally {
+      const errors: unknown[] = [];
+      // Different pipeline instances may share runsRoot; even their parent deletions must not overlap.
+      for (const [pipeline, runDirectories] of session.launchRunFolders ?? []) {
+        try { await pipeline.forgetSession(session.metadata.id, runDirectories); } catch (error) { errors.push(error); }
+      }
+      if (errors.length) throw new AggregateError(errors, "Launch file folders could not be removed.");
+    }
+  }
+
   /**
    * Closes a card. A card in a plugin environment releases it: `keepEnvironmentData` is the person's
    * answer to "Keep environment data?" (kept unless they said no). Quitting releases nothing.
@@ -1197,9 +1426,11 @@ export class TerminalManager {
     this.redaction.clear(`session:${id}`);
     this.releaseIsolation(id);
     session.launchToken += 1;
-    void session.launchCleanup?.().catch(() => undefined);
-    session.launchCleanup = null;
-    if (session.extras.options) void this.launchPipeline?.forgetSession(id).catch(() => undefined);
+    // launchToken and launch waiters were invalidated above: pending preparation/wrapping can now unwind before
+    // parent removal. Its original pipeline remains the owner even if the host configuration has since changed.
+    void this.forgetLaunchFiles(session).catch((error: unknown) => {
+      console.warn(`CanvasTTY could not remove the launch files of card ${id}.`, error);
+    });
     session.agentBrowser?.cleanup();
     session.agentRuntime?.cleanup();
     session.agentOrchestration?.cleanup();
@@ -1721,19 +1952,32 @@ export class TerminalManager {
   ): PlannedSpawn | { failure: LaunchFailure } {
     const providerCli = provider === "terminal" ? undefined : this.providerClis.get(provider);
     if (providerCli?.state === "unavailable") return { failure: providerCli };
+    // The person's environment: a plain shell the person starts keeps all of it but CanvasTTY's own credentials; their
+    // own agent cards also lose admin keys; a delegated card (a subagent, or any card a plugin started, plain shell
+    // included, since its owner drives it) gets only an allowlist (ambientProviderKeys.ts). Every launch builder below
+    // reads this, never process.env; what a launch sets explicitly is added on top.
+    const delegated = this.isDelegated(id, role);
+    const baseEnvironment = provider === "terminal" && !delegated
+      ? shellEnvironment(terminalEnvironment())
+      : agentEnvironment(terminalEnvironment(), { delegated });
     // What decides whether Claude's lifecycle hooks may go over HTTP (ClaudeHttpHooks.ts): where and how it runs.
     const claudeHttp = provider === "claude" && providerCli?.state === "available" ? {
       executable: providerCli.executable,
       profile,
       environmentWrapped,
-      env: { ...terminalEnvironment(), ...providerCli.environment, ...(contribution?.env ?? {}) },
+      env: { ...baseEnvironment, ...providerCli.environment, ...(contribution?.env ?? {}) },
       args: contribution?.args ?? [],
       cwd
     } : undefined;
+    // Grok and Kimi read their hooks (Kimi also its MCP servers) from their home: the launch's own account home (inside
+    // the data folder of the plugin that set it) gets CanvasTTY's temporary configuration there.
+    const hookHome = accountHookHome(provider, contribution?.env);
+    const accountHome = hookHome !== undefined && hookHome === contribution?.accountHome ? hookHome : undefined;
     const agentRuntime = provider === "terminal"
       ? null
-      : this.agentRuntime?.prepareLaunch({ terminalSessionId: id, provider, cwd,
+      : this.agentRuntime?.prepareLaunch({ terminalSessionId: id, provider, cwd, environment: baseEnvironment,
         ...(captureResult ? { captureResult: true } : {}),
+        ...(accountHome ? { accountHome } : {}),
         ...(claudeHttp ? { claudeHttp } : {}),
         ...(answerCaptureGrantExpiresAt === undefined ? {} : { answerCaptureGrantExpiresAt }) }) ?? null;
     let pluginTools: string[] = [];
@@ -1765,12 +2009,13 @@ export class TerminalManager {
           terminalSessionId: id,
           provider,
           cwd,
+          ...(accountHome ? { accountHome } : {}),
+          environment: baseEnvironment,
           ...(bridged ? {
             includeOrchestration: true,
             orchestrationTools: [...(role === "orchestrator" ? ORCHESTRATION_TOOL_NAMES : []), ...pluginTools]
           } : {})
         }) ?? null;
-      const baseEnvironment = terminalEnvironment();
       const browserEnvironment = agentBrowser?.environment ?? {};
       const runtimeEnvironment = agentRuntime?.environment ?? {};
       const orchestrationEnvironment = agentOrchestration?.environment ?? {};
@@ -1836,6 +2081,37 @@ export class TerminalManager {
     }
   }
 
+  /**
+   * The plugin launch options this card was started with (a copy), or undefined without any. A subagent inherits them
+   * unless its spawn names the plugin (AgentControlService.subagentLaunchOptions).
+   */
+  launchOptionsOf(id: string): Record<string, PluginLaunchValues> | undefined {
+    const options = this.sessions.get(id)?.extras.options;
+    return options ? structuredClone(options) as Record<string, PluginLaunchValues> : undefined;
+  }
+
+  /**
+   * The plugins whose options a subagent repeats exactly from its orchestrator of the same CLI (inherited, or named
+   * with the same values): the person chose them. Empty for any other card.
+   */
+  private personChosenLaunchPlugins(metadata: SessionMetadata, options: Record<string, unknown> | undefined): string[] {
+    if (metadata.role !== "subagent" || metadata.parentSessionId === undefined || !options) return [];
+    const parent = this.sessions.get(metadata.parentSessionId);
+    if (!parent || parent.metadata.provider !== metadata.provider) return [];
+    const parentOptions = (parent.extras.options ?? {}) as Record<string, Record<string, unknown>>;
+    return Object.entries(options as Record<string, Record<string, unknown>>).filter(([pluginId, values]) => {
+      const own = parentOptions[pluginId];
+      if (!own || !values) return false;
+      const keys = Object.keys(values);
+      return keys.length === Object.keys(own).length && keys.every((key) => Object.hasOwn(own, key) && own[key] === values[key]);
+    }).map(([pluginId]) => pluginId);
+  }
+
+  /** Plugins whose launch contribution decided this card's model in its current run; empty until it is prepared. */
+  modelSetByOf(id: string): string[] {
+    return [...(this.sessions.get(id)?.modelSetBy ?? [])];
+  }
+
   /** The model and effort this card's launches ask the CLI for. */
   private launchModelOf(id: string): LaunchModelChoice {
     const starting = this.startingModels.get(id);
@@ -1886,7 +2162,8 @@ export class TerminalManager {
     return this.environments?.available(environment) ?? false;
   }
 
-  private unavailableLaunchPlugins(options: Record<string, unknown> | undefined): string[] {
+  /** The plugins among these launch options that are disabled, removed or not trusted now. */
+  unavailableLaunchPlugins(options: Record<string, unknown> | undefined): string[] {
     if (!options) return [];
     return this.launchPipeline ? this.launchPipeline.unavailable(options) : Object.keys(options).sort();
   }
@@ -1906,7 +2183,11 @@ export class TerminalManager {
   ): void {
     const token = ++session.launchToken;
     const { metadata } = session;
-    void this.runContributedLaunch(id, session, token, resume, restoring, answerCaptureGrantExpiresAt)
+    // Each launch decides its model again; until it is prepared no plugin has.
+    session.modelSetBy = [];
+    const tasks = session.launchTasks ??= new Set();
+    // Register before invoking plugin code, which may synchronously close or restart its card.
+    const task = Promise.resolve().then(() => this.runContributedLaunch(id, session, token, resume, restoring, answerCaptureGrantExpiresAt))
       .catch((error: unknown): LaunchOutcome => {
         metadata.failureDetails = this.redactSecrets(`Launch refused: ${error instanceof Error ? error.message : String(error)}`);
         return "failed";
@@ -1920,7 +2201,9 @@ export class TerminalManager {
         this.emitSession(metadata, outcome === "failed" ? failureOrigin : null);
         this.schedulePersistence();
       })
-      .finally(() => this.wakeLaunchWaiters(session));
+      .finally(() => { tasks.delete(task); this.wakeLaunchWaiters(session); });
+    tasks.add(task);
+    void task.catch(() => undefined);
   }
 
   private async runContributedLaunch(
@@ -1939,6 +2222,8 @@ export class TerminalManager {
       return "failed";
     };
     const environments = this.environments;
+    // The launch starts a microtask after it was registered; a card closed or restarted in that same tick is done.
+    if (!live()) return "superseded";
 
     // 1. Place a new session where the person chose. The choice stays saved with the card until the plugin has
     // prepared it, so a card whose preparation was cut short (quit, crash) or failed never restores locally.
@@ -1980,142 +2265,148 @@ export class TerminalManager {
 
     // 3. Chosen launch contributors, and the launch policies that apply.
     let contribution: LaunchContribution | null = null;
-    const trustedFolder = session.extras.environment ? undefined : this.personTrustedFolder(metadata.parentSessionId, metadata.cwd);
-    if (session.extras.options || this.policyApplies(metadata.provider)) {
-      const pipeline = this.launchPipeline;
-      if (!pipeline) return refuse(missingLaunchPlugins(Object.keys(session.extras.options ?? {})));
-      const placedIn = session.extras.environment;
-      const prepared = await pipeline.prepare({
-        sessionId: id,
-        provider: metadata.provider,
-        profile: metadata.profile,
-        role: metadata.role,
-        cwd: metadata.cwd,
-        ...(metadata.parentSessionId !== undefined ? { parentSessionId: metadata.parentSessionId } : {}),
-        restoring,
-        resume: resume !== null,
-        options: structuredClone(session.extras.options ?? {}) as Record<string, Record<string, boolean | string>>,
-        environment: placedIn ? { pluginId: placedIn.pluginId, kind: placedIn.kind } : null,
-        ...(trustedFolder ? { trustedFolder } : {})
-      });
-      if (!live()) {
-        if (prepared.ok) void prepared.cleanup().catch(() => undefined);
-        return "superseded";
+    let adopted = false;
+    let plannedCleanup: (() => void) | undefined;
+    try {
+      const trustedFolder = session.extras.environment ? undefined : this.personTrustedFolder(metadata.parentSessionId, metadata.cwd);
+      if (session.extras.options || this.policyApplies(metadata.provider)) {
+        const pipeline = this.launchPipeline;
+        if (!pipeline) return refuse(missingLaunchPlugins(Object.keys(session.extras.options ?? {})));
+        // This pipeline may write run files for the card, so closing it later removes its folder there.
+        const runFolders = session.launchRunFolders ??= new Map();
+        const ownFolders = runFolders.get(pipeline) ?? new Set<string>();
+        runFolders.set(pipeline, ownFolders);
+        const placedIn = session.extras.environment;
+        const prepared = await pipeline.prepare({
+          sessionId: id,
+          provider: metadata.provider,
+          profile: metadata.profile,
+          role: metadata.role,
+          cwd: metadata.cwd,
+          ...(metadata.parentSessionId !== undefined ? { parentSessionId: metadata.parentSessionId } : {}),
+          restoring,
+          resume: resume !== null,
+          options: structuredClone(session.extras.options ?? {}) as Record<string, Record<string, boolean | string>>,
+          environment: placedIn ? { pluginId: placedIn.pluginId, kind: placedIn.kind } : null,
+          ...(trustedFolder ? { trustedFolder } : {})
+        }, {
+          ...(this.launchModelOf(id).model !== undefined ? { model: this.launchModelOf(id).model } : {}),
+          personChosen: this.personChosenLaunchPlugins(metadata, session.extras.options)
+        });
+        if (prepared.ok) {
+          contribution = prepared;
+          ownFolders.add(prepared.runDirectory);
+        }
+        if (!live()) return "superseded";
+        if (!prepared.ok) return refuse(prepared.reason);
+        session.modelSetBy = prepared.modelSetBy;
+        this.addLaunchSecrets(session, prepared.secrets);
       }
-      if (!prepared.ok) return refuse(prepared.reason);
-      contribution = prepared;
-      this.addLaunchSecrets(session, prepared.secrets);
-    }
-    const dropContribution = (): void => {
-      void contribution?.cleanup().catch(() => undefined);
-    };
 
-    // 4. What the environment keeps of CanvasTTY's protection, and the isolation layer for this launch.
-    const keeps = environment ? this.environments?.keeps?.(environment) ?? {} : {};
-    if (environment && keeps.launch !== true && metadata.profile !== "normal") {
-      dropContribution();
-      return refuse(`${environment.label} does not pass the launch on unchanged (the plugin does not declare it), so the ${metadata.profile} profile's settings and CanvasTTY's hooks would not reach the agent there. Launch it in normal, or use an environment that keeps them.`);
-    }
-    const decision = this.launchIsolation(id, metadata.provider, metadata.profile, metadata.role,
-      environment ? { isolated: keeps.isolated === true, label: environment.label } : null);
-    if (decision.refuse) {
-      dropContribution();
-      return refuse(decision.refuse);
-    }
-    if (environment && keeps.launch !== true) {
-      metadata.isolation = { ...(metadata.isolation ?? { state: "environment" }),
-        reason: `${metadata.isolation?.reason ? `${metadata.isolation.reason} ` : ""}Base protection and CanvasTTY's hooks do not reach the agent in ${environment.label}.` };
-    }
+      // 4. What the environment keeps of CanvasTTY's protection, and the isolation layer for this launch.
+      const keeps = environment ? this.environments?.keeps?.(environment) ?? {} : {};
+      if (environment && keeps.launch !== true && metadata.profile !== "normal") {
+        return refuse(`${environment.label} does not pass the launch on unchanged (the plugin does not declare it), so the ${metadata.profile} profile's settings and CanvasTTY's hooks would not reach the agent there. Launch it in normal, or use an environment that keeps them.`);
+      }
+      const decision = this.launchIsolation(id, metadata.provider, metadata.profile, metadata.role,
+        environment ? { isolated: keeps.isolated === true, label: environment.label } : null);
+      if (decision.refuse) {
+        return refuse(decision.refuse);
+      }
+      if (environment && keeps.launch !== true) {
+        metadata.isolation = { ...(metadata.isolation ?? { state: "environment" }),
+          reason: `${metadata.isolation?.reason ? `${metadata.isolation.reason} ` : ""}Base protection and CanvasTTY's hooks do not reach the agent in ${environment.label}.` };
+      }
 
-    // 5. The host spawns the PTY; an environment only rewrites what is spawned, and the isolation layer wraps that.
-    let planned: PlannedSpawn | { failure: LaunchFailure };
-    try {
-      planned = this.planSpawn(id, metadata.provider, decision.profile, metadata.cwd, resume,
-        session.captureResult, metadata.role, answerCaptureGrantExpiresAt, contribution, trustedFolder, Boolean(environment), decision.apply);
-    } catch (error) {
-      dropContribution();
-      metadata.failureDetails = this.redactSecrets(error instanceof Error ? error.message : String(error));
-      return "failed";
-    }
-    if ("failure" in planned) {
-      dropContribution();
-      applyLaunchFailure(metadata, planned.failure);
-      return "failed";
-    }
-    const abandon = (): void => {
-      planned.cleanup();
-      dropContribution();
-    };
-    this.noteConfiguredMode(id, metadata.provider, decision.profile, planned.env, planned.cwd);
-    let spawn: { command: string; args: string[] | string; cwd: string; env: Record<string, string> } = planned;
-    if (environment && environments) {
-      if (typeof planned.args === "string") {
-        abandon();
-        return refuse("this provider's Windows batch launcher cannot run in a plugin environment.");
+      // 5. The host spawns the PTY; an environment only rewrites what is spawned, and the isolation layer wraps that.
+      let planned: PlannedSpawn | { failure: LaunchFailure };
+      // Whether this launch gets lifecycle hooks is decided as it is planned; the wrap below may take a while.
+      const statusHooks = this.lifecycleHooksEnabled;
+      try {
+        planned = this.planSpawn(id, metadata.provider, decision.profile, metadata.cwd, resume,
+          session.captureResult, metadata.role, answerCaptureGrantExpiresAt, contribution, trustedFolder, Boolean(environment), decision.apply);
+      } catch (error) {
+        metadata.failureDetails = this.redactSecrets(error instanceof Error ? error.message : String(error));
+        return "failed";
       }
-      const secretValues = new Set(contribution?.secrets ?? []);
-      const secretEnvNames = Object.keys(contribution?.env ?? {}).filter((key) => secretValues.has(contribution!.env[key]!));
-      // The environment sees the launch's own variables, never CanvasTTY's reserved ones or secret values.
-      const visible = Object.fromEntries(Object.entries(planned.launchEnvironment)
-        .filter(([key]) => !RESERVED_ENV.test(key) && !secretEnvNames.includes(key)));
-      const wrapped = await environments.wrap(environment, {
-        sessionId: id,
-        provider: metadata.provider,
-        launch: { command: planned.command, args: planned.args, env: visible, cwd: planned.cwd },
-        secretEnvNames,
-        takenEnv: new Set(Object.keys(planned.launchEnvironment)),
-        path: launchSearchPath(planned.env)
-      });
-      if (!live()) {
-        abandon();
-        return "superseded";
+      if ("failure" in planned) {
+        applyLaunchFailure(metadata, planned.failure);
+        return "failed";
       }
-      if (!wrapped.ok) {
-        abandon();
-        return refuse(wrapped.reason);
+      plannedCleanup = planned.cleanup;
+      this.noteConfiguredMode(id, metadata.provider, decision.profile, planned.env, planned.cwd);
+      let spawn: { command: string; args: string[] | string; cwd: string; env: Record<string, string> } = planned;
+      if (environment && environments) {
+        if (typeof planned.args === "string") {
+          return refuse("this provider's Windows batch launcher cannot run in a plugin environment.");
+        }
+        const secretValues = new Set(contribution?.secrets ?? []);
+        const secretEnvNames = Object.keys(contribution?.env ?? {}).filter((key) => secretValues.has(contribution!.env[key]!));
+        // The environment sees the launch's own variables, never CanvasTTY's reserved ones or secret values.
+        const visible = Object.fromEntries(Object.entries(planned.launchEnvironment)
+          .filter(([key]) => !RESERVED_ENV.test(key) && !secretEnvNames.includes(key)));
+        const wrapped = await environments.wrap(environment, {
+          sessionId: id,
+          provider: metadata.provider,
+          launch: { command: planned.command, args: planned.args, env: visible, cwd: planned.cwd },
+          secretEnvNames,
+          takenEnv: new Set(Object.keys(planned.launchEnvironment)),
+          path: launchSearchPath(planned.env)
+        });
+        if (!live()) {
+          return "superseded";
+        }
+        if (!wrapped.ok) {
+          return refuse(wrapped.reason);
+        }
+        this.addLaunchSecrets(session, wrapped.secrets);
+        spawn = { command: wrapped.command, args: wrapped.args, cwd: wrapped.cwd,
+          env: { ...planned.env, ...(wrapped.cwd !== planned.cwd ? { PWD: wrapped.cwd } : {}), ...wrapped.env } };
       }
-      this.addLaunchSecrets(session, wrapped.secrets);
-      spawn = { command: wrapped.command, args: wrapped.args, cwd: wrapped.cwd,
-        env: { ...planned.env, ...(wrapped.cwd !== planned.cwd ? { PWD: wrapped.cwd } : {}), ...wrapped.env } };
-    }
-    if (decision.apply) {
+      if (decision.apply) {
+        try {
+          spawn = { ...this.wrapIsolated(id, metadata.provider, decision.profile, spawn), cwd: spawn.cwd };
+        } catch (error) {
+          return refuse(error instanceof Error ? error.message : String(error));
+        }
+      }
+      let process: IPty;
       try {
-        spawn = { ...this.wrapIsolated(id, metadata.provider, decision.profile, spawn), cwd: spawn.cwd };
+        process = this.spawnPty(spawn.command, spawn.args, {
+          name: "xterm-256color", cols: session.cols, rows: session.rows, cwd: spawn.cwd, env: spawn.env
+        });
       } catch (error) {
-        abandon();
-        return refuse(error instanceof Error ? error.message : String(error));
+        this.releaseIsolation(id);
+        metadata.failureDetails = this.redactSecrets(error instanceof Error ? error.message : String(error));
+        return "failed";
+      }
+      session.process = process;
+      if (planned.launchEnvironment.CANVASTTY_CODEX_KEYBOARD) {
+        metadata.nativeEditor = JSON.parse(planned.launchEnvironment.CANVASTTY_CODEX_KEYBOARD);
+      } else {
+        delete metadata.nativeEditor;
+      }
+      this.launchContexts.set(id, { cwd: spawn.cwd, configDir: spawn.env.CLAUDE_CONFIG_DIR ?? null });
+      session.agentBrowser = planned.agentBrowser;
+      session.agentRuntime = planned.agentRuntime;
+      session.agentOrchestration = planned.agentOrchestration;
+      session.launchCleanup = contribution?.cleanup ?? null;
+      adopted = true;
+      metadata.status = initialSessionStatus(metadata.provider);
+      metadata.exitCode = null;
+      metadata.failureDetails = null;
+      this.bindProcess(id, session, process, statusHooks);
+      const runtimeStatus = this.agentRuntime?.currentStatus(id);
+      if (runtimeStatus) metadata.status = runtimeStatus;
+      if (environment) this.describeEnvironment(id, session, environment);
+      return "launched";
+    } finally {
+      // Unadopted resources remain owned by this tracked launch, including throws during plugin wrapping.
+      if (!adopted) {
+        try { plannedCleanup?.(); }
+        finally { await contribution?.cleanup().catch(() => undefined); }
       }
     }
-    let process: IPty;
-    try {
-      process = this.spawnPty(spawn.command, spawn.args, {
-        name: "xterm-256color", cols: session.cols, rows: session.rows, cwd: spawn.cwd, env: spawn.env
-      });
-    } catch (error) {
-      abandon();
-      this.releaseIsolation(id);
-      metadata.failureDetails = this.redactSecrets(error instanceof Error ? error.message : String(error));
-      return "failed";
-    }
-    session.process = process;
-    if (planned.launchEnvironment.CANVASTTY_CODEX_KEYBOARD) {
-      metadata.nativeEditor = JSON.parse(planned.launchEnvironment.CANVASTTY_CODEX_KEYBOARD);
-    } else {
-      delete metadata.nativeEditor;
-    }
-    this.launchContexts.set(id, { cwd: spawn.cwd, configDir: spawn.env.CLAUDE_CONFIG_DIR ?? null });
-    session.agentBrowser = planned.agentBrowser;
-    session.agentRuntime = planned.agentRuntime;
-    session.agentOrchestration = planned.agentOrchestration;
-    session.launchCleanup = contribution?.cleanup ?? null;
-    metadata.status = initialSessionStatus(metadata.provider);
-    metadata.exitCode = null;
-    metadata.failureDetails = null;
-    this.bindProcess(id, session, process);
-    const runtimeStatus = this.agentRuntime?.currentStatus(id);
-    if (runtimeStatus) metadata.status = runtimeStatus;
-    if (environment) this.describeEnvironment(id, session, environment);
-    return "launched";
   }
 
   private addLaunchSecrets(session: ManagedSession, secrets: readonly string[]): void {
@@ -2133,13 +2424,22 @@ export class TerminalManager {
     }).catch(() => undefined);
   }
 
-  private bindProcess(id: string, session: ManagedSession, process: IPty): void {
+  /**
+   * `statusHooks`: status hooks were on when this launch was planned (and so got its lifecycle hooks). A launch planned
+   * and spawned in one step passes nothing; a contributed launch passes what held before its asynchronous steps.
+   */
+  private bindProcess(id: string, session: ManagedSession, process: IPty, statusHooks = this.lifecycleHooksEnabled): void {
     session.launchOutputOffset = session.outputOffset;
+    if (statusHooks && this.lifecycleHooksEnabled) delete session.statusHooksOff;
+    else session.statusHooksOff = true;
     session.agentBrowser?.retainUntilExit?.();
     process.onData((data) => {
       const current = this.sessions.get(id);
       if (!current || current !== session || current.process !== process) return;
 
+      if (!current.inputReady && waitsForInputReady(current.metadata.provider) && watchInputReadiness(current, data)) {
+        this.wakeLaunchWaiters(current);
+      }
       const lifecycleState = current.lifecycle?.push(data);
       if (lifecycleState) current.titleState = lifecycleState;
       if (lifecycleState && !titleDefersToHooks(current, lifecycleState)) {
@@ -2153,38 +2453,46 @@ export class TerminalManager {
 
     let exited!: () => void;
     this.liveProcesses.set(process, new Promise<void>((resolve) => { exited = resolve; }));
-    process.onExit(({ exitCode }) => {
+    process.onExit(({ exitCode, signal }) => {
       this.liveProcesses.delete(process);
       exited();
       const current = this.sessions.get(id);
       if (!current || current !== session || current.process !== process) return;
       // node-pty calls this from a native callback that aborts the whole app when JavaScript throws in it.
       try {
-        this.recordExit(id, current, exitCode);
+        this.recordExit(id, current, exitCode, signal);
       } catch (error) {
         console.warn(`PTY ${id} exit could not be recorded.`, error);
       }
     });
   }
 
-  private recordExit(id: string, current: ManagedSession, exitCode: number): void {
+  private recordExit(id: string, current: ManagedSession, reportedExitCode: number, signal?: number): void {
     this.flushOutput(id, current);
     this.releaseIsolation(id);
     if (current.extras.gitAuditSince !== undefined) void this.auditGit(id, current);
+    // node-pty reports a process killed by a signal as exitCode 0 plus the signal. That is a crash, not a clean finish:
+    // record it the way a shell does, as 128+signal, so the card, wait_for_agent and the saved session all say failed.
+    const killedBy = typeof signal === "number" && signal > 0 ? signal : 0;
+    const exitCode = killedBy ? 128 + killedBy : reportedExitCode;
     current.metadata.exitCode = exitCode;
     current.metadata.status = exitCode === 0 ? "done" : "failed";
-    current.metadata.failureDetails = exitCode === 0
+    const details = exitCode === 0
       ? null
       : terminalFailureDetails(this.redactSecrets(current.bufferChunks.slice(current.bufferStart).join("")));
+    current.metadata.failureDetails = killedBy
+      ? `${signalExitText(killedBy)}${details ? `\n${details}` : ""}`.slice(0, 8_000)
+      : details;
     current.agentBrowser?.cleanup();
     current.agentBrowser = null;
     current.agentRuntime?.cleanup();
     current.agentRuntime = null;
     current.agentOrchestration?.cleanup();
     current.agentOrchestration = null;
-    void current.launchCleanup?.().catch(() => undefined);
-    current.launchCleanup = null;
+    void this.cleanupLaunchFiles(current).catch(() => undefined);
     this.emitSession(current.metadata);
+    // Input waiting for this CLI to become ready learns at once that it never will.
+    this.wakeLaunchWaiters(current);
     // Recorded at the moment of exit, so a finished agent is never relaunched.
     this.schedulePersistence();
   }
@@ -2285,12 +2593,64 @@ function resetLaunchSignals(session: ManagedSession): void {
   delete session.answer;
   delete session.turnStarts;
   delete session.promptTurnMark;
+  // A late answer of the ended launch never belongs to the new one.
+  session.inputGeneration = (session.inputGeneration ?? 0) + 1;
+  delete session.statusHooksOff;
+  delete session.typedSinceSubmit;
+  delete session.answerTurnGeneration;
+  delete session.providerTurnId;
+  delete session.providerTurnGenerations;
   delete session.hookSignals;
   delete session.titleState;
+  delete session.inputReady;
+  delete session.alternateScreen;
+  delete session.readinessCarry;
+  delete session.hookReadyAt;
+  if (session.hookReadyTimer) clearTimeout(session.hookReadyTimer);
+  session.hookReadyTimer = undefined;
   if (session.answeredPromptTimer) clearTimeout(session.answeredPromptTimer);
   session.answeredPromptTimer = undefined;
 }
 
+/** Whether typed text reaches this launch's CLI now (see WAITS_FOR_INPUT_READY). */
+function cliAcceptsInput(session: ManagedSession): boolean {
+  return !waitsForInputReady(session.metadata.provider) || session.inputReady === true
+    || (session.hookReadyAt !== undefined && Date.now() >= session.hookReadyAt);
+}
+
+/** DEC private modes set or reset (CSI ? Pm h / l): terminal control sequences, not text the CLI shows. */
+const DEC_PRIVATE_MODES = /\x1b\[\?([\d;]+)([hl])/gu;
+/** The longest DEC private mode sequence kept from the end of one chunk for the next. */
+const READINESS_CARRY_CHARS = 32;
+
+/**
+ * Reads only terminal control sequences of one output chunk: the alternate screen (modes 1049, 1047, 47) and the text
+ * cursor (DECTCEM, mode 25). A full-screen CLI hides its cursor while it draws its start-up and shows it once its
+ * input field has focus; that moment is when typed keys start to reach that field. True once that happened; a
+ * sequence cut at the chunk's end is completed by the next chunk.
+ */
+function watchInputReadiness(session: ManagedSession, data: string): boolean {
+  const text = (session.readinessCarry ?? "") + data;
+  let consumed = 0;
+  for (const match of text.matchAll(DEC_PRIVATE_MODES)) {
+    consumed = match.index + match[0].length;
+    const set = match[2] === "h";
+    for (const mode of match[1]!.split(";")) {
+      if (mode === "1049" || mode === "1047" || mode === "47") session.alternateScreen = set;
+      else if (mode === "25" && set && session.alternateScreen) session.inputReady = true;
+    }
+    if (session.inputReady) break;
+  }
+  if (session.inputReady) {
+    delete session.readinessCarry;
+    delete session.alternateScreen;
+    return true;
+  }
+  const escape = text.lastIndexOf("\x1b");
+  session.readinessCarry = escape >= consumed && text.length - escape < READINESS_CARRY_CHARS ? text.slice(escape) : "";
+  return false;
+}
+
 function titleDefersToHooks(session: ManagedSession, state: "idle" | "working" | "needs_approval"): boolean {
   return state !== "working" && session.metadata.provider === "claude" && session.hookSignals !== undefined;
 }
@@ -2509,3 +2869,9 @@ function appendScrollback(session: ManagedSession, data: string): void {
     session.bufferStart = 0;
   }
 }
+
+/** "The process was killed by signal SIGKILL (9)." from node-pty's signal number. */
+function signalExitText(signal: number): string {
+  const name = Object.entries(osConstants.signals).find(([, value]) => value === signal)?.[0];
+  return `The process was killed by signal ${name ? `${name} (${signal})` : signal}.`;
+}
diff --git a/src/main/services/accountHomes.ts b/src/main/services/accountHomes.ts
new file mode 100644
index 000000000..673639c0b
--- /dev/null
+++ b/src/main/services/accountHomes.ts
@@ -0,0 +1,104 @@
+import { lstatSync, mkdirSync, realpathSync } from "node:fs";
+import { isAbsolute, join, relative, resolve, sep } from "node:path";
+import type { ProviderId } from "../../shared/contracts.ts";
+
+// Account homes: a CLI home a launch contributor starts Grok or Kimi with (an account of an account pool). Grok Build
+// and Kimi Code read their hooks, and Kimi its MCP servers and permission rules, only from that home, so CanvasTTY's
+// temporary overlays go there for such a launch. These are the rules for when a home counts as the launch's own.
+
+/** The variable that moves the home whose configuration Grok Build / Kimi Code read. */
+const ACCOUNT_HOME_VARIABLES: Readonly<Partial<Record<ProviderId, string>>> = { grok: "GROK_HOME", kimi: "KIMI_CODE_HOME" };
+
+/** A folder path without trailing separators (`/a/b/` names the same folder as `/a/b`); a root stays as it is. */
+export function withoutTrailingSeparator(path: string): string {
+  const trimmed = path.replace(/[\\/]+$/u, "");
+  return trimmed === "" || /^[A-Za-z]:$/u.test(trimmed) ? path : trimmed;
+}
+
+/** The path segments of `path` strictly below `root`, or null when it is not strictly inside. */
+function segmentsBelow(root: string, path: string): string[] | null {
+  const rest = relative(root, path);
+  if (rest === "" || rest.startsWith("..") || isAbsolute(rest)) return null;
+  return rest.split(sep);
+}
+
+/**
+ * The home a launch contribution started a Grok or Kimi launch with (its `GROK_HOME` / `KIMI_CODE_HOME`); undefined for
+ * other providers or when the contribution leaves the home alone. `KIMI_SHARE_DIR` (the older kimi-cli) is not it.
+ */
+export function accountHookHome(provider: ProviderId, contributedEnvironment: Readonly<Record<string, string>> | undefined): string | undefined {
+  const name = ACCOUNT_HOME_VARIABLES[provider];
+  const value = name ? contributedEnvironment?.[name] : undefined;
+  return typeof value === "string" && value.length > 0 ? withoutTrailingSeparator(value) : undefined;
+}
+
+/**
+ * The account home one contributor handed a Grok/Kimi launch: `value` of the provider's home variable `key` when it is an
+ * exact absolute path strictly inside that contributor's own data folder. `dataDir` is the host's record of the
+ * contributor (`<userData>/plugin-data/<pluginId>`), never the plugin's answer, so a plugin can only name a folder of
+ * its own data.
+ */
+export function contributedHookHome(provider: ProviderId, dataDir: string | undefined, key: string, value: string): string | undefined {
+  if (!dataDir || !isAbsolute(dataDir) || ACCOUNT_HOME_VARIABLES[provider] !== key) return undefined;
+  const folder = withoutTrailingSeparator(value);
+  if (!isAbsolute(folder) || resolve(folder) !== folder) return undefined;
+  return segmentsBelow(resolve(dataDir), folder) ? folder : undefined;
+}
+
+/**
+ * The account home an overlay may be written into, or null for the default home (the launch is then exactly as before).
+ * Only an exact absolute folder at least `<plugin>/<home>` deep inside `root` (`<userData>/plugin-data`), other than the
+ * default home, counts. Inside the root every folder from the root down must be a real folder (a missing one is created,
+ * private) and none of `written` (names in the home CanvasTTY writes) a symbolic link, or this throws and the launch is
+ * refused: CanvasTTY never writes through a link into another place.
+ */
+export function ownedAccountHome(options: {
+  root: string | undefined;
+  candidate: string | undefined;
+  defaultHome: string;
+  written: readonly string[];
+  label: string;
+}): string | null {
+  const { root, candidate } = options;
+  if (!root || candidate === undefined) return null;
+  const home = withoutTrailingSeparator(candidate);
+  if (!isAbsolute(home) || resolve(home) !== home || home === options.defaultHome) return null;
+  const segments = segmentsBelow(resolve(root), home);
+  if (!segments || segments.length < 2) return null;
+  try {
+    let path = resolve(root);
+    let canonical = realpathSync.native(path);
+    for (const segment of ["", ...segments]) {
+      if (segment) {
+        path = join(path, segment);
+        canonical = join(canonical, segment);
+      }
+      // A new account's home (or the folders above it in the plugin's data) may not exist yet: the CLI would create it
+      // on its first run. One private folder at a time, never recursively, so no link is followed while creating it.
+      if (segment && !exists(path)) mkdirSync(path, { mode: 0o700 });
+      const info = lstatSync(path);
+      if (!info.isDirectory() || info.isSymbolicLink() || realpathSync.native(path) !== canonical) throw new Error("not a real folder");
+    }
+  } catch {
+    throw new Error(`The ${options.label} account home is missing or crosses a symbolic link; CanvasTTY did not write its configuration there.`);
+  }
+  for (const name of options.written) {
+    let link = false;
+    try {
+      link = lstatSync(join(home, name)).isSymbolicLink();
+    } catch {
+      continue;
+    }
+    if (link) throw new Error(`The ${options.label} account home holds a symbolic link where CanvasTTY writes its configuration.`);
+  }
+  return home;
+}
+
+function exists(path: string): boolean {
+  try {
+    lstatSync(path);
+    return true;
+  } catch {
+    return false;
+  }
+}
diff --git a/src/main/services/agent-browser/AgentBrowserBridge.ts b/src/main/services/agent-browser/AgentBrowserBridge.ts
index 06fa4f53b..216663f85 100644
--- a/src/main/services/agent-browser/AgentBrowserBridge.ts
+++ b/src/main/services/agent-browser/AgentBrowserBridge.ts
@@ -14,6 +14,10 @@ export interface PrepareAgentBrowserLaunchInput {
   includeOrchestration?: boolean;
   /** The canvastty_agents tools this session may use (default: the core tools). */
   orchestrationTools?: readonly string[];
+  /** Kimi: the launch's own account home (`KIMI_CODE_HOME` inside the contributing plugin's data), for its config. */
+  accountHome?: string;
+  /** The card's own (already filtered) environment, which the launch builds from instead of the app's. */
+  environment?: Readonly<Record<string, string | undefined>>;
 }
 
 export interface PreparedAgentBrowserPtyLaunch {
@@ -66,12 +70,15 @@ export class AgentBrowserBridge implements AgentBrowserLaunchCoordinator {
     // Browser access off: no canvastty_browser and no browser capability, but a session that gets canvastty_agents
     // (orchestrators, plugin tools) still gets it; that server has its own capability and gateway.
     if (!this.gateway.isEnabled) return input.includeOrchestration ? this.prepareOrchestrationOnly(input) : null;
-    const capability = this.gateway.registerAgent(input);
+    const { environment, ...agent } = input;
+    const capability = this.gateway.registerAgent(agent);
     let providerLaunch;
     try {
       providerLaunch = this.providers.prepare(input.provider, capability.connectionId, {
         ...(input.includeOrchestration ? { orchestration: true } : {}),
-        ...(input.orchestrationTools ? { orchestrationTools: input.orchestrationTools } : {})
+        ...(input.orchestrationTools ? { orchestrationTools: input.orchestrationTools } : {}),
+        ...(input.accountHome ? { accountHome: input.accountHome } : {}),
+        ...(environment ? { environment } : {})
       });
     } catch (error) {
       this.gateway.revokeTerminalSession(input.terminalSessionId);
@@ -93,7 +100,9 @@ export class AgentBrowserBridge implements AgentBrowserLaunchCoordinator {
     const providerLaunch = this.providers.prepare(input.provider, connectionId, {
       orchestration: true,
       browser: false,
-      ...(input.orchestrationTools ? { orchestrationTools: input.orchestrationTools } : {})
+      ...(input.orchestrationTools ? { orchestrationTools: input.orchestrationTools } : {}),
+      ...(input.accountHome ? { accountHome: input.accountHome } : {}),
+      ...(input.environment ? { environment: input.environment } : {})
     });
     return this.launchWith(providerLaunch, "", connectionId, () => undefined, {});
   }
diff --git a/src/main/services/agent-browser/OrchestrationTools.ts b/src/main/services/agent-browser/OrchestrationTools.ts
index ba6afebf3..28566b785 100644
--- a/src/main/services/agent-browser/OrchestrationTools.ts
+++ b/src/main/services/agent-browser/OrchestrationTools.ts
@@ -6,6 +6,7 @@ import { PromptNotDeliveredError, type AgentControlService, type SpawnAgentReque
 import { LaunchRefusal } from "../launchRefusal.ts";
 import type { PluginAgentTools } from "../PluginAgentTools.ts";
 import {
+  CLIENT_SAFE_AGENT_WAIT_SECONDS,
   DEFAULT_AGENT_WAIT_SECONDS,
   MAX_AGENT_WAIT_SECONDS,
   ORCHESTRATION_TOOL_DEFINITIONS,
@@ -22,6 +23,13 @@ import type { McpToolDefinition } from "../../../agent-browser/orchestration-cat
  * a foreign session id is a protocol error, never a filtered result, so an
  * orchestrator cannot probe sessions it does not own.
  */
+/**
+ * How long spawn_agent (its checks, the launch and the CLI's readiness together) and send_to_agent wait for a prompt
+ * to reach the agent: under the about 60 s after which most MCP clients end a tool call, so the orchestrator gets
+ * the reason (and the session id) instead of a client timeout.
+ */
+export const AGENT_TOOL_DELIVERY_MS = 45_000;
+
 export class ScopedOrchestrationHandler implements OrchestrationCommandHandler {
   private readonly control: AgentControlService;
   private readonly plugins: Pick<PluginAgentTools, "list" | "call"> | null;
@@ -65,9 +73,9 @@ export class ScopedOrchestrationHandler implements OrchestrationCommandHandler {
         case "send_to_agent":
           return await this.send(sessionId, request.arguments, signal);
         case "observe_agent":
-          return this.observe(sessionId, request.arguments);
+          return await this.observe(sessionId, request.arguments);
         case "get_agent_result":
-          return this.result(sessionId, request.arguments);
+          return await this.result(sessionId, request.arguments);
         case "cancel_agent":
           return this.cancel(sessionId, request.arguments);
         case "list_agents":
@@ -118,12 +126,15 @@ export class ScopedOrchestrationHandler implements OrchestrationCommandHandler {
     }
     this.requireOwned(orchestratorId, target);
     const seconds = typeof args.timeoutSeconds === "number" ? args.timeoutSeconds : DEFAULT_AGENT_WAIT_SECONDS;
+    // Shorter than the orchestrator's MCP client lets a tool call run, so it gets "timeout" and progress, not an error.
+    const limit = this.control.status(orchestratorId).provider === "claude" ? MAX_AGENT_WAIT_SECONDS : CLIENT_SAFE_AGENT_WAIT_SECONDS;
     try {
       const result = await this.control.waitFor(target, {
-        timeoutMs: Math.min(MAX_AGENT_WAIT_SECONDS, Math.max(1, seconds)) * 1_000,
+        timeoutMs: Math.min(limit, Math.max(1, seconds)) * 1_000,
         ...(signal ? { signal } : {})
       });
-      return { ...result };
+      return { ...result, ...(result.reason === "timeout"
+        ? { message: "Still running. Read output for its progress and call wait_for_agent again." } : {}) };
     } catch (error) {
       if (signal?.aborted || (error instanceof Error && error.name === "AbortError")) throw canceledError();
       throw error;
@@ -131,6 +142,7 @@ export class ScopedOrchestrationHandler implements OrchestrationCommandHandler {
   }
 
   private async spawn(orchestratorId: string, args: Record<string, unknown>, signal?: AbortSignal): Promise<Record<string, unknown>> {
+    const deadline = Date.now() + AGENT_TOOL_DELIVERY_MS;
     if (!(AGENT_PROVIDERS as readonly unknown[]).includes(args.provider)) {
       throw orchestrationBridgeError("INVALID_REQUEST", unknownProviderMessage(args.provider), false);
     }
@@ -146,17 +158,26 @@ export class ScopedOrchestrationHandler implements OrchestrationCommandHandler {
       if (unknown) throw orchestrationBridgeError("INVALID_REQUEST", unknown, false);
     }
     if (signal?.aborted) throw canceledError();
+    // A plugin the spawn does not name keeps this session's own options (never silently the CLI's default model).
+    const requested = args.launchOptions as SpawnAgentRequest["launchOptions"];
+    let inheritance: ReturnType<AgentControlService["subagentLaunchOptions"]>;
+    try {
+      inheritance = this.control.subagentLaunchOptions(orchestratorId, provider, requested, args.model as string | undefined);
+    } catch (error) {
+      throw orchestrationBridgeError("INVALID_REQUEST", error instanceof Error ? error.message : String(error), false);
+    }
+    const launchOptions = inheritance.launchOptions;
     const created = await this.control.spawn({
       parentSessionId: orchestratorId,
       provider: args.provider as never,
       cwd: args.cwd as string,
       ...(args.title !== undefined ? { title: args.title as string } : {}),
       ...(args.prompt !== undefined ? { initialPrompt: args.prompt as string } : {}),
-      ...(args.launchOptions !== undefined ? { launchOptions: args.launchOptions as SpawnAgentRequest["launchOptions"] } : {}),
+      ...(launchOptions !== undefined ? { launchOptions } : {}),
       ...(args.model !== undefined ? { model: args.model as string } : {}),
       ...(args.effort !== undefined ? { effort: args.effort as SpawnAgentRequest["effort"] } : {}),
       profile: profile.profile
-    }, signal).catch((error: unknown) => {
+    }, signal, Math.max(0, deadline - Date.now())).catch((error: unknown) => {
       if (signal?.aborted) throw canceledError();
       throw error;
     });
@@ -177,7 +198,10 @@ export class ScopedOrchestrationHandler implements OrchestrationCommandHandler {
       profile: created.profile,
       ...(profile.inherited ? { profileInherited: true } : {}),
       ...(created.model !== undefined ? { model: created.model } : {}),
-      ...(created.effort !== undefined ? { effort: created.effort } : {})
+      ...(created.effort !== undefined ? { effort: created.effort } : {}),
+      // Which CLI serves it, the model it was asked for (null: its launch options or the CLI's own default decide),
+      // and where each plugin's launch options came from.
+      servedBy: { provider: created.provider, model: created.model ?? null, launchOptions: inheritance.sources }
     };
   }
 
@@ -188,7 +212,8 @@ export class ScopedOrchestrationHandler implements OrchestrationCommandHandler {
         args.sessionId as string,
         args.prompt as string,
         args.submit === undefined ? true : Boolean(args.submit),
-        signal
+        signal,
+        AGENT_TOOL_DELIVERY_MS
       );
     } catch (error) {
       if (signal?.aborted) throw canceledError();
@@ -198,9 +223,9 @@ export class ScopedOrchestrationHandler implements OrchestrationCommandHandler {
     return { sessionId: args.sessionId as string, sent: true };
   }
 
-  private observe(orchestratorId: string, args: Record<string, unknown>): Record<string, unknown> {
+  private async observe(orchestratorId: string, args: Record<string, unknown>): Promise<Record<string, unknown>> {
     this.requireOwned(orchestratorId, args.sessionId as string);
-    const observation = this.control.observe(
+    const observation = await this.control.observe(
       args.sessionId as string,
       args.maxChars as number | undefined
     );
@@ -213,9 +238,9 @@ export class ScopedOrchestrationHandler implements OrchestrationCommandHandler {
     };
   }
 
-  private result(orchestratorId: string, args: Record<string, unknown>): Record<string, unknown> {
+  private async result(orchestratorId: string, args: Record<string, unknown>): Promise<Record<string, unknown>> {
     this.requireOwned(orchestratorId, args.sessionId as string);
-    const result = this.control.result(args.sessionId as string);
+    const result = await this.control.result(args.sessionId as string);
     return {
       sessionId: result.sessionId,
       state: result.state,
@@ -223,7 +248,8 @@ export class ScopedOrchestrationHandler implements OrchestrationCommandHandler {
       exitCode: result.exitCode,
       output: result.output,
       ...(result.exitLines ? { exitLines: result.exitLines } : {}),
-      ...(result.answer ? { answer: result.answer } : {})
+      ...(result.answer ? { answer: result.answer } : {}),
+      ...(result.answerUnavailable ? { answerUnavailable: result.answerUnavailable } : {})
     };
   }
 
diff --git a/src/main/services/agent-browser/ProviderLaunch.ts b/src/main/services/agent-browser/ProviderLaunch.ts
index f59cbe12c..b5e2fdcf1 100644
--- a/src/main/services/agent-browser/ProviderLaunch.ts
+++ b/src/main/services/agent-browser/ProviderLaunch.ts
@@ -22,6 +22,7 @@ import {
   resolveHermesHomeDirectory
 } from "../hermesConfig.ts";
 import { openCodeBrowserEnvironment } from "../openCodeConfig.ts";
+import { ownedAccountHome } from "../accountHomes.ts";
 import {
   acquireConfigurationLock,
   atomicWrite,
@@ -74,6 +75,11 @@ export interface ProviderLaunchOptions {
   /** The same probe off the main thread, for warmKimiProbe; defaults to the sync probe's answer when only that is given. */
   probeKimiPerRunConfigAsync?: (cli: AvailableProviderCli) => Promise<boolean>;
   environment?: Readonly<Record<string, string | undefined>>;
+  /**
+   * `<userData>/plugin-data`: where an account home (the `KIMI_CODE_HOME` a launch contributor started Kimi with) may be
+   * for Kimi's temporary configuration to be written into it (accountHomes.ts). Without it no account home gets it.
+   */
+  accountHomesRoot?: string;
 }
 
 interface KimiLockHooks extends ConfigurationLockHooks {
@@ -95,8 +101,8 @@ export class ProviderLaunchAdapters {
   private kimiWarmed: { executable: string; generation: number; result: boolean } | null = null;
   private kimiWarming: Promise<void> | null = null;
   private kimiGeneration = 0;
-  private kimiConfiguration: KimiTemporaryConfiguration | null = null;
-  private kimiConfigurationUsers = 0;
+  /** The temporary Kimi configuration per Kimi home (the default home and account homes) and its launches. */
+  private readonly kimiConfigurations = new Map<string, { configuration: KimiTemporaryConfiguration; users: number }>();
   private hermesConfiguration: HermesTemporaryConfiguration | null = null;
   private hermesConfigurationUsers = 0;
 
@@ -108,6 +114,9 @@ export class ProviderLaunchAdapters {
     this.kimiHomeDirectory = validateKimiHomeDirectory(
       options.kimiHomeDirectory ?? join(homedir(), ".kimi-code")
     );
+    if (options.accountHomesRoot !== undefined && !isAbsolute(options.accountHomesRoot)) {
+      throw new Error("Account homes root must be absolute.");
+    }
     this.environment = options.environment ?? process.env;
     // The probe runs Kimi with this adapter's environment (a fake HOME, an account's), never the app's own.
     this.probe = options.probeKimiPerRunConfig ?? ((cli) => probeKimiPerRunMcpConfig(cli, undefined, this.environment));
@@ -148,9 +157,18 @@ export class ProviderLaunchAdapters {
 
   /**
    * `orchestrationTools`: the canvastty_agents tools this session may use (default: the core tools). `browser: false`
-   * (browser access off) leaves canvastty_browser out, so only canvastty_agents is attached.
+   * (browser access off) leaves canvastty_browser out, so only canvastty_agents is attached. `accountHome` (Kimi only):
+   * the `KIMI_CODE_HOME` a launch contributor started this launch with; Kimi's temporary configuration goes there when
+   * it is an account home (accountHomes.ts), since that is the only home this Kimi reads.
    */
-  prepare(provider: AgentProvider, connectionId: string, options?: { orchestration?: boolean; orchestrationTools?: readonly string[]; browser?: boolean }): PreparedProviderLaunch {
+  prepare(provider: AgentProvider, connectionId: string, options?: {
+    orchestration?: boolean;
+    orchestrationTools?: readonly string[];
+    browser?: boolean;
+    accountHome?: string;
+    /** The card's own (already filtered) environment; the launch reads the person's OpenCode inline config from it. */
+    environment?: Readonly<Record<string, string | undefined>>;
+  }): PreparedProviderLaunch {
     const providerCli = this.providerClis.get(provider);
     if (providerCli.state === "unavailable") throw new Error(providerCli.diagnostic);
     const orchestrationHelper = orchestrationHelperFor(this.options, options);
@@ -181,7 +199,7 @@ export class ProviderLaunchAdapters {
     if (provider === "opencode") {
       return {
         args: [],
-        environment: openCodeBrowserEnvironment(browserHelper, this.environment, orchestrationHelper),
+        environment: openCodeBrowserEnvironment(browserHelper, options?.environment ?? this.environment, orchestrationHelper),
         releaseConfiguration() {}
       };
     }
@@ -192,7 +210,14 @@ export class ProviderLaunchAdapters {
         releaseConfiguration: this.acquireHermesConfiguration(browserHelper, orchestrationHelper)
       };
     }
-    return this.prepareKimi(connectionId, browserHelper, orchestrationHelper);
+    const home = ownedAccountHome({
+      root: this.options.accountHomesRoot,
+      candidate: options?.accountHome,
+      defaultHome: this.kimiHomeDirectory,
+      written: KIMI_WRITTEN,
+      label: "Kimi"
+    }) ?? this.kimiHomeDirectory;
+    return this.prepareKimi(connectionId, browserHelper, orchestrationHelper, home);
   }
 
   recoverKimiConfiguration(): void {
@@ -234,7 +259,7 @@ export class ProviderLaunchAdapters {
     };
   }
 
-  private prepareKimi(connectionId: string, browserHelper: StdioHelperLaunch | null, orchestrationHelper?: StdioHelperLaunch): PreparedProviderLaunch {
+  private prepareKimi(connectionId: string, browserHelper: StdioHelperLaunch | null, orchestrationHelper: StdioHelperLaunch | undefined, home: string): PreparedProviderLaunch {
     const kimiCli = this.providerClis.get("kimi");
     if (kimiCli.state === "unavailable") throw new Error(kimiCli.diagnostic);
     if (this.kimiProbedExecutable !== kimiCli.executable) {
@@ -252,6 +277,7 @@ export class ProviderLaunchAdapters {
     // The per-run document carries orchestration per launch, so the shared
     // fallback configuration only tracks it when mcp.json is actually mutated.
     const releaseShared = this.acquireKimiConfiguration(
+      home,
       !supportsPerRun,
       browserHelper,
       supportsPerRun ? undefined : orchestrationHelper
@@ -286,6 +312,7 @@ export class ProviderLaunchAdapters {
   }
 
   private acquireKimiConfiguration(
+    home: string,
     includeMcpEntry: boolean,
     browserHelper: StdioHelperLaunch | null,
     orchestrationHelper?: StdioHelperLaunch
@@ -293,32 +320,34 @@ export class ProviderLaunchAdapters {
     const requiresOrchestration = includeMcpEntry && orchestrationHelper !== undefined;
     // The shared mcp.json carries the browser entry only while browser access is on.
     const includeBrowser = includeMcpEntry && browserHelper !== null;
-    if (!this.kimiConfiguration) {
-      this.kimiConfiguration = KimiTemporaryConfiguration.begin({
-        homeDirectory: this.kimiHomeDirectory,
+    let shared = this.kimiConfigurations.get(home);
+    if (!shared) {
+      shared = { configuration: KimiTemporaryConfiguration.begin({
+        homeDirectory: home,
         helper: includeBrowser ? browserHelper : null,
         includeMcpEntry,
         ...(requiresOrchestration && orchestrationHelper ? { orchestrationHelper } : {})
-      });
-    } else if (this.kimiConfiguration.includeMcpEntry !== includeMcpEntry) {
+      }), users: 0 };
+      this.kimiConfigurations.set(home, shared);
+    } else if (shared.configuration.includeMcpEntry !== includeMcpEntry) {
       throw new Error("Kimi MCP capability changed while temporary configuration is active.");
-    } else if (includeMcpEntry && this.kimiConfiguration.hasBrowserEntry !== includeBrowser) {
+    } else if (includeMcpEntry && shared.configuration.hasBrowserEntry !== includeBrowser) {
       throw new Error("Kimi browser access cannot change while a temporary Kimi configuration is active.");
-    } else if (requiresOrchestration && !this.kimiConfiguration.hasOrchestrationEntry) {
+    } else if (requiresOrchestration && !shared.configuration.hasOrchestrationEntry) {
       // Mirrors the Hermes guard: never silently drop the canvastty_agents
       // entry an orchestrator needs because a plain launch owns the config.
       throw new Error("Kimi MCP orchestration cannot be enabled while a temporary Kimi configuration is active.");
     }
-    this.kimiConfigurationUsers += 1;
+    const owned = shared;
+    owned.users += 1;
     let released = false;
     return () => {
       if (released) return;
       released = true;
-      this.kimiConfigurationUsers -= 1;
-      if (this.kimiConfigurationUsers !== 0) return;
-      const configuration = this.kimiConfiguration;
-      this.kimiConfiguration = null;
-      configuration?.cleanup();
+      owned.users -= 1;
+      if (owned.users !== 0) return;
+      if (this.kimiConfigurations.get(home) === owned) this.kimiConfigurations.delete(home);
+      owned.configuration.cleanup();
     };
   }
 }
@@ -867,6 +896,9 @@ function parseJournal(raw: string, paths: ReturnType<typeof kimiPaths>): Recover
   return value as unknown as RecoveryJournal;
 }
 
+/** What KimiTemporaryConfiguration writes in a Kimi home (kimiPaths); none may be a symbolic link in an account home. */
+const KIMI_WRITTEN = ["mcp.json", "config.toml", ".canvastty-browser.lock", ".canvastty-browser-recovery.json", ".canvastty-browser-backups"];
+
 function kimiPaths(homeDirectory: string) {
   return {
     mcp: join(homeDirectory, "mcp.json"),
diff --git a/src/main/services/agent-control/AgentControlGateway.ts b/src/main/services/agent-control/AgentControlGateway.ts
index 0d3d140d9..3a47c9df8 100644
--- a/src/main/services/agent-control/AgentControlGateway.ts
+++ b/src/main/services/agent-control/AgentControlGateway.ts
@@ -612,7 +612,16 @@ export class AgentControlGateway {
       this.redact(terminalScreenText(owned.terminal, true)), codexPlaceholderIsDim(owned.terminal), codexContinuationIsEmpty(owned.terminal));
     if (request.method === "screen") return { sessionId: id, text: screen, revision: hash(screen), outputOffset: owned.outputOffset,
       interaction: capabilities.menus ? codexChoices(screen) : null, composerReady };
-    if ((request.method === "choose" || request.method === "dismiss") && !capabilities.menus) {
+    // A prompt only the person may answer: the card waits for them (needs_approval, from the CLI's own permission
+    // hook), or Codex shows its approval prompt (before that hook's report arrived; for text, only while no composer is
+    // ready, so earlier output on screen does not refuse a task). An orchestrator's connection never answers it, with a
+    // choice, Escape or text. The person's own connection and the card itself still can.
+    const menuAction = request.method === "choose" || request.method === "dismiss";
+    const approvalOnScreen = capabilities.menus && (menuAction || !composerReady) && codexApprovalPrompt(screen);
+    if (agent && (menuAction || request.method === "send") && (fresh.status === "needs_approval" || approvalOnScreen)) {
+      throw new ControlError("NEEDS_PERSON", PERSON_ANSWERS_REFUSAL);
+    }
+    if (menuAction && !capabilities.menus) {
       throw new ControlError("NOT_SUPPORTED", `Menus are parsed for Codex only; resolve ${fresh.provider} prompts from the desktop and treat screen as the only evidence.`);
     }
     // Its launch plugins are still preparing it: nothing can be written yet, and nothing is queued.
@@ -777,6 +786,21 @@ export class AgentControlGateway {
   }
 }
 
+/** What an orchestrator's connection gets for choose, dismiss or send while its worker waits for the person. */
+export const PERSON_ANSWERS_REFUSAL = "This worker is waiting for the person to answer an approval or permission prompt in its card "
+  + "(needs_approval); only the person may answer it, so nothing was written. Do not retry: tell the person what it asks "
+  + "(screen shows it), then check status again. Once they have answered it is no longer needs_approval.";
+
+/**
+ * Codex's own approval prompts (running a command, applying edits, granting permissions or network access, an MCP
+ * server's request): the person's to answer, whatever the card's status says yet. Its other menus (hooks review,
+ * model and update notices) are not approvals.
+ */
+export function codexApprovalPrompt(screen: string): boolean {
+  return /Would you like to (?:run the following command|make the following edits|grant these permissions)|\bneeds your approval\b|Do you want to approve\b/i.test(screen)
+    || /^\s*(?:›\s*)?\d+\.\s+(?:Yes, proceed\b|Yes, and don['’]t ask again\b|No, and tell Codex what to do differently\b)/m.test(screen);
+}
+
 function codexChoices(screen: string): { revision: string; selected: number; options: Array<{ number: number; label: string }> } | null {
   const matches = screen.split("\n").map((line) => line.match(/^\s*(›\s*)?(\d+)\.\s+(.+)$/)).filter((m) => m !== null);
   if (matches.length < 2 || matches.length > 12 || matches.filter((m) => m[1]).length !== 1) return null;
diff --git a/src/main/services/agent-control/controlCapabilities.ts b/src/main/services/agent-control/controlCapabilities.ts
index e1b872f5a..2d4d37dc5 100644
--- a/src/main/services/agent-control/controlCapabilities.ts
+++ b/src/main/services/agent-control/controlCapabilities.ts
@@ -17,7 +17,10 @@ export const CONTROL_PROVIDERS: readonly AgentProviderId[] = AGENT_PROVIDERS;
 export interface ControlCapabilities {
   /** The worker reports a captured final answer through `result`. */
   result: boolean;
-  /** `screen` parses startup/approval menus that `choose`/`dismiss` can act on. */
+  /**
+   * `screen` parses numbered menus that `choose`/`dismiss` can act on. An approval or permission prompt is the person's:
+   * an orchestrator's own connection is refused it (NEEDS_PERSON).
+   */
   menus: boolean;
 }
 
diff --git a/src/main/services/agent-runtime/AgentRuntimeBridge.ts b/src/main/services/agent-runtime/AgentRuntimeBridge.ts
index 18b6ef2d2..72aa1e65f 100644
--- a/src/main/services/agent-runtime/AgentRuntimeBridge.ts
+++ b/src/main/services/agent-runtime/AgentRuntimeBridge.ts
@@ -23,6 +23,10 @@ export interface PrepareAgentRuntimeLaunchInput {
   decisions?: boolean;
   /** Claude Code: what decides whether its lifecycle hooks may go over HTTP (see ClaudeHttpHooks.ts). */
   claudeHttp?: ClaudeHttpLaunchFacts;
+  /** Grok/Kimi: the CLI home a launch contributor started the launch with (`accountHookHome`), for its status hooks. */
+  accountHome?: string;
+  /** The card's own (already filtered) environment, which the launch builds from instead of the app's. */
+  environment?: Readonly<Record<string, string | undefined>>;
 }
 
 export interface PreparedAgentRuntimePtyLaunch {
@@ -92,7 +96,7 @@ export class AgentRuntimeBridge implements AgentRuntimeLaunchCoordinator {
     let prepared;
     try {
       prepared = this.providers.prepare(input.provider, input.terminalSessionId, this.coreHooksEnabled, decisions, budgetMs,
-        httpHookBase ?? undefined);
+        httpHookBase ?? undefined, input.accountHome, input.environment);
     } catch (error) {
       if (capability) this.gateway.revokeTerminalSession(input.terminalSessionId, capability.capabilityToken);
       throw error;
diff --git a/src/main/services/agent-runtime/ProviderRuntimeLaunch.ts b/src/main/services/agent-runtime/ProviderRuntimeLaunch.ts
index a945496fe..11b757776 100644
--- a/src/main/services/agent-runtime/ProviderRuntimeLaunch.ts
+++ b/src/main/services/agent-runtime/ProviderRuntimeLaunch.ts
@@ -11,6 +11,7 @@ import { dirname, isAbsolute, join, win32 } from "node:path";
 import { pathToFileURL } from "node:url";
 import { lazyRequire } from "../../lazyRequire.ts";
 import { isOwnLifecycleHookText } from "../agentHelpers.ts";
+import { ownedAccountHome } from "../accountHomes.ts";
 import {
   atomicWrite,
   ensurePrivateDirectory,
@@ -93,6 +94,11 @@ export interface ProviderRuntimeLaunchOptions {
   pluginHooks?: RuntimePluginHookSource;
   /** The decision hook (permission-gate.mjs); without it decision hooks are never installed. */
   permissionGate?: RuntimeHookHelperLaunch;
+  /**
+   * `<userData>/plugin-data`: the only place an account home (a CLI home a launch contributor started the launch with)
+   * may be for the Grok/Kimi lifecycle overlays to be written into it. Without it no account home gets them.
+   */
+  accountHomesRoot?: string;
 }
 
 export interface PreparedProviderRuntimeLaunch {
@@ -101,6 +107,10 @@ export interface PreparedProviderRuntimeLaunch {
   releaseConfiguration(): void;
 }
 
+const KIMI_JOURNAL = ".canvastty-runtime-kimi.json";
+const RUNTIME_BACKUPS = ".canvastty-runtime-backups";
+const GROK_HOOKS_FILE = "canvastty-runtime-hooks.json";
+
 const HOOK_PROVIDERS: ReadonlySet<string> = new Set(["claude", "codex", "qwen", "opencode", "kimi", "hermes", "grok", "omp", "pi"]);
 
 export class ProviderRuntimeLaunchAdapters {
@@ -120,6 +130,8 @@ export class ProviderRuntimeLaunchAdapters {
   private grokOverlay: GrokRuntimeHooks | null = null;
   private grokOverlaySignature: string | null = null;
   private grokUsers = 0;
+  /** Grok/Kimi overlays in account homes, by provider and home; each is removed when its last launch ends. */
+  private readonly accountOverlays = new Map<string, { overlay: { cleanup(): void }; signature: string; users: number }>();
 
   constructor(options: ProviderRuntimeLaunchOptions) {
     validateHelper(options.helper);
@@ -129,6 +141,9 @@ export class ProviderRuntimeLaunchAdapters {
     if (!isAbsolute(options.openCodePluginPath)) {
       throw new Error("OpenCode lifecycle plugin path must be absolute.");
     }
+    if (options.accountHomesRoot !== undefined && !isAbsolute(options.accountHomesRoot)) {
+      throw new Error("Account homes root must be absolute.");
+    }
     if (options.permissionGate) {
       validateHelper(options.permissionGate);
       // Either Electron running one absolute script, or the native helper's permission-gate subcommand.
@@ -175,6 +190,10 @@ export class ProviderRuntimeLaunchAdapters {
    * `decisionBudgetMs` (a decision service's `decide.timeoutMs`) lengthens the hook's deadlines to fit it.
    * `claudeHttpHookBase` (Claude Code only) sends its lifecycle events, except SessionStart, as HTTP hooks to the
    * gateway's loopback listener instead of running the helper; the decision hook stays a command.
+   * `accountHome` (Grok and Kimi only, see accountHomes.ts): the CLI home a launch contributor started this launch
+   * with. When it is an account home inside `accountHomesRoot`, the Grok/Kimi overlay goes there instead of into the
+   * default home, since that is the only home the CLI reads. Claude Code, Codex, Qwen Code and OpenCode take their hooks
+   * from the launch's arguments or environment, so they work in any home already.
    */
   prepare(
     provider: AgentProvider,
@@ -182,7 +201,10 @@ export class ProviderRuntimeLaunchAdapters {
     coreHooksEnabled = true,
     decisions = false,
     decisionBudgetMs?: number,
-    claudeHttpHookBase?: string
+    claudeHttpHookBase?: string,
+    accountHome?: string,
+    /** The card's own (already filtered) environment; the launch reads the person's OpenCode inline config from it. */
+    environment: Readonly<Record<string, string | undefined>> = this.environment
   ): PreparedProviderRuntimeLaunch {
     const pluginRegistrations = this.options.pluginHooks?.list(provider) ?? [];
     const gate = decisions && this.decisionsSupported(provider);
@@ -195,7 +217,7 @@ export class ProviderRuntimeLaunchAdapters {
     // MiniMax, Devin and Antigravity must never reach Grok's shared hook overlay.
     const hasHooks = HOOK_PROVIDERS.has(provider)
       && (coreHooksEnabled || pluginCommands.length > 0 || (provider === "opencode" && (pluginRegistrations.length > 0 || openCodeDecisions)));
-    const environment = hasHooks
+    const hookEnvironment = hasHooks
       ? {
         ...(pluginRegistrations.length > 0 ? {
           [PLUGIN_HOOK_TERMINAL_SESSION_ENV]: terminalSessionId
@@ -204,13 +226,13 @@ export class ProviderRuntimeLaunchAdapters {
       : {};
     if (!hasHooks) {
       this.clearSharedOverlay(provider);
-      return prepared([], environment);
+      return prepared([], hookEnvironment);
     }
     if (provider === "claude") {
-      return prepared(claudeHookArgs(this.options.helper, this.platform, coreHooksEnabled, pluginCommands, claudeHttpHookBase), environment);
+      return prepared(claudeHookArgs(this.options.helper, this.platform, coreHooksEnabled, pluginCommands, claudeHttpHookBase), hookEnvironment);
     }
     if (provider === "codex") {
-      return prepared(codexHookArgs(this.options.helper, this.platform, coreHooksEnabled, pluginCommands), environment);
+      return prepared(codexHookArgs(this.options.helper, this.platform, coreHooksEnabled, pluginCommands), hookEnvironment);
     }
     if (provider === "qwen") {
       const path = createQwenHookSettings({
@@ -222,7 +244,7 @@ export class ProviderRuntimeLaunchAdapters {
         coreHooksEnabled,
         pluginCommands
       });
-      return prepared([], { ...environment, [QWEN_SYSTEM_SETTINGS]: path }, () => unlinkIfExists(path));
+      return prepared([], { ...hookEnvironment, [QWEN_SYSTEM_SETTINGS]: path }, () => unlinkIfExists(path));
     }
     if (provider === "opencode") {
       const pluginEnvironment = this.openCodePluginEnvironment(
@@ -230,11 +252,11 @@ export class ProviderRuntimeLaunchAdapters {
         coreHooksEnabled
       );
       return prepared([], {
-        ...environment,
+        ...hookEnvironment,
         ...pluginEnvironment,
         ...(openCodeDecisions ? { [OPENCODE_DECISIONS_ENV]: "1", ...budgetEnvironment(decisionBudgetMs) } : {}),
         [OPENCODE_CONFIG_CONTENT]: openCodeLifecycleConfig(
-          this.environment[OPENCODE_CONFIG_CONTENT],
+          environment[OPENCODE_CONFIG_CONTENT],
           this.options.openCodePluginPath
         )
       });
@@ -242,16 +264,20 @@ export class ProviderRuntimeLaunchAdapters {
     if (provider === "omp" || provider === "pi") {
       return prepared(coreHooksEnabled
         ? ["--extension", join(dirname(this.options.openCodePluginPath), "omp-extension.mjs")]
-        : [], environment);
+        : [], hookEnvironment);
+    }
+    if (provider === "kimi" || provider === "grok") {
+      const ownHome = this.ownAccountHome(provider, accountHome);
+      if (ownHome) return prepared([], hookEnvironment, this.acquireAccountHome(provider, ownHome, coreHooksEnabled, pluginCommands));
     }
     if (provider === "kimi") {
-      return prepared([], environment, this.acquireKimi(coreHooksEnabled, pluginCommands));
+      return prepared([], hookEnvironment, this.acquireKimi(coreHooksEnabled, pluginCommands));
     }
     if (provider === "hermes") {
-      return prepared([], environment, this.acquireHermes(coreHooksEnabled, pluginCommands));
+      return prepared([], hookEnvironment, this.acquireHermes(coreHooksEnabled, pluginCommands));
     }
-    if (provider === "grok") return prepared([], environment, this.acquireGrok(coreHooksEnabled, pluginCommands));
-    return prepared([], environment);
+    if (provider === "grok") return prepared([], hookEnvironment, this.acquireGrok(coreHooksEnabled, pluginCommands));
+    return prepared([], hookEnvironment);
   }
 
   recoverConfigurations(): void {
@@ -355,6 +381,47 @@ export class ProviderRuntimeLaunchAdapters {
     });
   }
 
+  /** The account home this Grok/Kimi launch's overlay goes into (accountHomes.ts), or null for the default home. */
+  private ownAccountHome(provider: "kimi" | "grok", candidate: string | undefined): string | null {
+    return ownedAccountHome({
+      root: this.options.accountHomesRoot,
+      candidate,
+      defaultHome: provider === "kimi" ? this.kimiHomeDirectory : this.grokHomeDirectory,
+      written: provider === "kimi" ? ["config.toml", KIMI_JOURNAL, RUNTIME_BACKUPS] : ["hooks", join("hooks", GROK_HOOKS_FILE)],
+      label: provider === "kimi" ? "Kimi" : "Grok"
+    });
+  }
+
+  private acquireAccountHome(
+    provider: "kimi" | "grok",
+    home: string,
+    coreHooksEnabled: boolean,
+    pluginCommands: readonly ProviderHookCommand[]
+  ): () => void {
+    const key = `${provider}\0${home}`;
+    const signature = overlaySignature(coreHooksEnabled, pluginCommands);
+    let entry = this.accountOverlays.get(key);
+    if (!entry || entry.signature !== signature) {
+      if (entry) {
+        this.accountOverlays.delete(key);
+        entry.overlay.cleanup();
+      }
+      const overlay = provider === "kimi"
+        ? KimiRuntimeHooks.begin(home, this.options.helper, this.platform, coreHooksEnabled, pluginCommands)
+        : GrokRuntimeHooks.begin(home, this.options.helper, this.platform, coreHooksEnabled, pluginCommands);
+      entry = { overlay, signature, users: 0 };
+      this.accountOverlays.set(key, entry);
+    }
+    const owned = entry;
+    owned.users += 1;
+    return once(() => {
+      owned.users -= 1;
+      if (owned.users !== 0 || this.accountOverlays.get(key) !== owned) return;
+      this.accountOverlays.delete(key);
+      owned.overlay.cleanup();
+    });
+  }
+
   /** Whether this provider can take the decision hook: Claude Code, Codex, Qwen Code (PreToolUse) and OpenCode. */
   decisionsSupported(provider: AgentProvider): boolean {
     return Boolean(this.options.permissionGate) && (provider === "opencode" || Object.hasOwn(DECISION_TOOL_MATCHERS, provider));
@@ -406,6 +473,8 @@ const CLAUDE_HOOKS: readonly HookMapping[] = [
   { event: "UserPromptSubmit", state: "working" },
   { event: "PermissionRequest", state: "needs_approval" },
   { event: "PostToolUse", state: "working" },
+  // An approved tool that fails runs PostToolUseFailure, not PostToolUse: the turn goes on after the approval.
+  { event: "PostToolUseFailure", state: "working" },
   { event: "Stop", state: "idle" },
   { event: "StopFailure", state: "idle" },
   { event: "SessionEnd", state: "idle" },
@@ -794,7 +863,7 @@ class KimiRuntimeHooks {
   ): KimiRuntimeHooks {
     ensurePrivateDirectory(home);
     const path = join(home, "config.toml");
-    const journalPath = join(home, ".canvastty-runtime-kimi.json");
+    const journalPath = join(home, KIMI_JOURNAL);
     this.recover(home);
     const original = readOptional(path);
     const block = kimiHookBlock(helper, platform, coreHooksEnabled, pluginCommands);
@@ -809,7 +878,7 @@ class KimiRuntimeHooks {
   }
 
   static recover(home: string): void {
-    const journalPath = join(home, ".canvastty-runtime-kimi.json");
+    const journalPath = join(home, KIMI_JOURNAL);
     const journal = readJournal(journalPath);
     if (!journal) return;
     const path = join(home, "config.toml");
@@ -895,7 +964,7 @@ class GrokRuntimeHooks {
   ): GrokRuntimeHooks {
     const hooksDirectory = join(home, "hooks");
     ensurePrivateDirectory(hooksDirectory);
-    const path = join(hooksDirectory, "canvastty-runtime-hooks.json");
+    const path = join(hooksDirectory, GROK_HOOKS_FILE);
     this.recover(home);
     if (existsSync(path)) throw new Error("Grok CanvasTTY lifecycle hook path is already occupied.");
     const expected = `${JSON.stringify({
@@ -909,7 +978,7 @@ class GrokRuntimeHooks {
   }
 
   static recover(home: string): void {
-    const path = join(home, "hooks", "canvastty-runtime-hooks.json");
+    const path = join(home, "hooks", GROK_HOOKS_FILE);
     const current = readOptional(path);
     if (current === null) return;
     if (
@@ -948,7 +1017,7 @@ function createTextJournal(
   mutated: string,
   extra: Record<string, unknown>
 ): TextOverlayJournal {
-  const backupDirectory = join(home, ".canvastty-runtime-backups");
+  const backupDirectory = join(home, RUNTIME_BACKUPS);
   ensurePrivateDirectory(backupDirectory);
   const backupPath = join(backupDirectory, `${provider}-${randomUUID()}.bak`);
   if (original !== null) atomicWrite(backupPath, original, modeOf(join(home, provider === "kimi" ? "config.toml" : "config.yaml")));
diff --git a/src/main/services/agent-runtime/RuntimeGateway.ts b/src/main/services/agent-runtime/RuntimeGateway.ts
index 7ad4f9d48..42a27aad8 100644
--- a/src/main/services/agent-runtime/RuntimeGateway.ts
+++ b/src/main/services/agent-runtime/RuntimeGateway.ts
@@ -79,6 +79,8 @@ interface RuntimeLease {
   provider: Exclude<ProviderId, "terminal">;
   tokenDigest: Buffer;
   activeTurnId: string | null;
+  /** A turn started and has not reported idle yet; between turns an unseen turn id may open the next one. */
+  turnOpen: boolean;
   /** Turn ids this session already reported (bounded, oldest dropped): a late start of one of them is stale. */
   seenTurnIds: Set<string>;
   latest: RuntimeLifecycleSignal | null;
@@ -276,6 +278,7 @@ export class RuntimeGateway {
       provider,
       tokenDigest: tokenDigest(capabilityToken),
       activeTurnId: null,
+      turnOpen: false,
       seenTurnIds: new Set(),
       latest: null,
       captureResult: captureResultOrGrantExpiresAt === true,
@@ -469,21 +472,27 @@ export class RuntimeGateway {
       throw new Error("Answer capture is not authorized for this session.");
     }
 
-    if (message.turnId && isTurnStart(message.event)) {
+    const startsTurn = isTurnStart(message.event);
+    if (message.turnId && startsTurn) {
       // Each hook runs on its own connection, so a start of an earlier turn can arrive after the next turn began;
       // it must not make that newer turn's events look stale.
       if (message.turnId !== lease.activeTurnId && lease.seenTurnIds.has(message.turnId)) return null;
       lease.activeTurnId = message.turnId;
+      lease.turnOpen = true;
+      rememberTurn(lease, message.turnId);
+    } else if (message.turnId && message.turnId !== lease.activeTurnId) {
+      // A turn whose start hook was lost (it timed out, or the CLI skipped it) still counts: after an idle, an
+      // unseen id opens the next turn. During an open turn, or for an id already seen, the event is out of order.
+      if (lease.activeTurnId && (lease.turnOpen || lease.seenTurnIds.has(message.turnId))) return null;
+      lease.activeTurnId = message.turnId;
+      lease.turnOpen = true;
       rememberTurn(lease, message.turnId);
-    } else if (
-      message.turnId
-      && lease.activeTurnId
-      && message.turnId !== lease.activeTurnId
-    ) {
-      return null;
     } else if (message.turnId) {
       rememberTurn(lease, message.turnId);
+    } else if (startsTurn) {
+      lease.turnOpen = true;
     }
+    if (message.state === "idle") lease.turnOpen = false;
     const signal: RuntimeLifecycleSignal = {
       state: message.state,
       event: message.event,
@@ -877,7 +886,8 @@ function isTurnStart(event: string): boolean {
   return event === "UserPromptSubmit"
     || event === "TurnStarted"
     || event === "pre_llm_call"
-    || event === "session.status:busy";
+    || event === "session.status:busy"
+    || event === "session.status:retry";
 }
 
 function isRecord(value: unknown): value is Record<string, unknown> {
diff --git a/src/main/services/agent-runtime/applyRuntimeSignal.ts b/src/main/services/agent-runtime/applyRuntimeSignal.ts
new file mode 100644
index 000000000..47b800df6
--- /dev/null
+++ b/src/main/services/agent-runtime/applyRuntimeSignal.ts
@@ -0,0 +1,22 @@
+import type { RuntimeLifecycleSignal } from "./RuntimeGateway.ts";
+import type { TerminalManager } from "../TerminalManager.ts";
+
+/**
+ * What main does with one runtime gateway signal: the card's status first, then the final answer it carries (Codex
+ * Stop hook, OpenCode plugin) for get_agent_result and wait_for_agent, kept only when the signal belongs to the
+ * current turn and the last submitted input.
+ */
+export function applyRuntimeSignal(
+  terminals: Pick<TerminalManager, "applyProviderSignal" | "recordAnswer">,
+  terminalSessionId: string,
+  signal: RuntimeLifecycleSignal
+): void {
+  const current = terminals.applyProviderSignal(terminalSessionId, {
+    kind: "lifecycle",
+    state: signal.state,
+    event: signal.event,
+    ...(signal.turnId ? { requestId: signal.turnId } : {}),
+    ...(signal.threadId ? { threadId: signal.threadId } : {})
+  });
+  if (current && signal.result) terminals.recordAnswer(terminalSessionId, signal.result, { turnId: signal.turnId });
+}
diff --git a/src/main/services/ambientProviderKeys.ts b/src/main/services/ambientProviderKeys.ts
new file mode 100644
index 000000000..3657702f6
--- /dev/null
+++ b/src/main/services/ambientProviderKeys.ts
@@ -0,0 +1,87 @@
+/**
+ * The person's opt-in: a comma- or space-separated list of variable names that agent cards keep from CanvasTTY's
+ * own environment although this module would leave them out (e.g. `OPENAI_ADMIN_KEY`). Read from the app's
+ * environment only, so an agent cannot widen it for itself.
+ */
+export const AGENT_PASS_ENV = "CANVASTTY_AGENT_PASS_ENV";
+
+/**
+ * Organization-level credentials (OPENAI_ADMIN_KEY, ANTHROPIC_ADMIN_KEY, XAI_ADMIN_API_KEY, GITLAB_ADMIN_TOKEN, …):
+ * projects, members, usage, billing. No agent CLI signs in with them.
+ */
+const ADMIN_CREDENTIAL = /(?:^|_)ADMIN_[A-Z0-9_]*(?:KEY|TOKEN)$/u;
+
+/**
+ * CanvasTTY's own credentials in its environment (CANVASTTY_GITHUB_TOKEN, which the plugin catalog uses, and any
+ * CANVASTTY_*TOKEN / *KEY / *SECRET / *PASSWORD): the app's, read by the main process, never needed by a card.
+ * "KEY" counts only as a whole word part (CANVASTTY_CODEX_KEYBOARD is not one).
+ */
+const APP_CREDENTIAL = /^CANVASTTY_(?:[A-Z0-9]+_)*[A-Z0-9]*(?:TOKEN|KEY|SECRET|PASSWORD|PASSWD)(?:_|$)/u;
+
+/**
+ * What a delegated card inherits from the person's environment: what a process needs to run, find its tools, reach
+ * the network and find its CLI's own state, and nothing that signs in anywhere. Proxy variables are kept on purpose
+ * (without them nothing reaches the network behind a corporate proxy), although a proxy URL may carry its own
+ * user:password. SSH_AUTH_SOCK, GH_TOKEN, AWS_*, cloud and provider credentials are left out; the person lists one in
+ * AGENT_PASS_ENV, or a launch contribution passes it explicitly (secretEnv).
+ */
+const DELEGATED_NAMES = new Set<string>([
+  // Process basics
+  "PATH", "HOME", "USER", "LOGNAME", "SHELL", "PWD", "SHLVL", "TZ", "LANG", "LANGUAGE",
+  "TERM", "COLORTERM", "TERM_PROGRAM", "TERM_PROGRAM_VERSION", "COLORFGBG", "NO_COLOR", "FORCE_COLOR",
+  "TMPDIR", "TEMP", "TMP", "EDITOR", "VISUAL", "PAGER", "__CF_USER_TEXT_ENCODING",
+  // Linux desktop session (addresses only)
+  "DISPLAY", "WAYLAND_DISPLAY", "XAUTHORITY", "DBUS_SESSION_BUS_ADDRESS",
+  // Windows
+  "SYSTEMROOT", "SYSTEMDRIVE", "WINDIR", "COMSPEC", "PATHEXT", "APPDATA", "LOCALAPPDATA", "USERPROFILE", "USERNAME",
+  "HOMEDRIVE", "HOMEPATH", "PROGRAMDATA", "PROGRAMFILES", "PROGRAMFILES(X86)", "PROGRAMW6432", "COMMONPROGRAMFILES",
+  "COMMONPROGRAMFILES(X86)", "PUBLIC", "ALLUSERSPROFILE", "OS", "PROCESSOR_ARCHITECTURE", "NUMBER_OF_PROCESSORS",
+  // Network: proxies and the person's own certificate authorities
+  "HTTP_PROXY", "HTTPS_PROXY", "ALL_PROXY", "NO_PROXY",
+  "SSL_CERT_FILE", "SSL_CERT_DIR", "NODE_EXTRA_CA_CERTS", "REQUESTS_CA_BUNDLE", "CURL_CA_BUNDLE",
+  // Where each CLI keeps its own state (a moved home is where the CLI's sign-in lives, not a credential)
+  "CLAUDE_CONFIG_DIR", "CLAUDE_SECURESTORAGE_CONFIG_DIR", "CODEX_HOME", "GROK_HOME", "KIMI_CODE_HOME", "KIMI_SHARE_DIR",
+  "KIMI_HOME", "QWEN_HOME", "QWEN_RUNTIME_DIR", "HERMES_HOME", "PI_CODING_AGENT_DIR", "PI_CODING_AGENT_SESSION_DIR",
+  "PI_CONFIG_DIR", "OPENCODE_CONFIG", "OPENCODE_CONFIG_DIR", "MINIMAX_DATA_DIR", "MAVIS_DATA_DIR", "CURSOR_CONFIG_DIR"
+]);
+/**
+ * Locale (LC_ALL, LC_CTYPE, …) and XDG folders. No CANVASTTY_* variable is inherited: what a card needs of CanvasTTY's
+ * own (its runtime, browser and orchestration addresses and capabilities, the Codex keyboard) the launch sets itself,
+ * and the rest (CANVASTTY_HELPERS, CANVASTTY_USER_DATA_DIR, CANVASTTY_GITHUB_TOKEN, …) is read by the main process.
+ */
+const DELEGATED_PREFIX = /^(?:LC_|XDG_)/u;
+
+export interface AgentEnvironmentInput {
+  /** Not started by the person: a subagent, or any card (agent or plain shell) a plugin started. */
+  delegated: boolean;
+}
+
+/**
+ * The environment a card inherits from CanvasTTY's own (terminalEnvironment), before what the core and launch
+ * contributions set for the launch. A plain shell card the person starts goes through shellEnvironment instead.
+ * Every launch builder (agent runtime, browser bridge, OpenCode inline config) reads this result, not process.env.
+ * - The person's own top-level agent cards keep everything but admin and CanvasTTY credentials: a CLI signed in by
+ *   API key keeps working, in any profile.
+ * - A delegated card gets only DELEGATED_NAMES / DELEGATED_PREFIX (an allowlist, so no provider, cloud or git
+ *   credential, whatever its name); what it needs beyond that it gets explicitly, through a launch contribution's
+ *   secretEnv.
+ * Names in AGENT_PASS_ENV are kept. Names compare case-insensitively (Windows).
+ */
+/** A plain shell card the person starts: their whole environment, without CanvasTTY's own credentials. */
+export function shellEnvironment(source: Readonly<Record<string, string>>): Record<string, string> {
+  return Object.fromEntries(Object.entries(source).filter(([key]) => !APP_CREDENTIAL.test(key.toUpperCase())));
+}
+
+export function agentEnvironment(
+  source: Readonly<Record<string, string>>,
+  input: AgentEnvironmentInput
+): Record<string, string> {
+  const passKey = Object.keys(source).find((key) => key.toUpperCase() === AGENT_PASS_ENV);
+  const pass = new Set((passKey ? source[passKey] : "").split(/[\s,]+/u).filter(Boolean).map((name) => name.toUpperCase()));
+  return Object.fromEntries(Object.entries(source).filter(([key]) => {
+    const name = key.toUpperCase();
+    if (pass.has(name)) return true;
+    if (ADMIN_CREDENTIAL.test(name) || APP_CREDENTIAL.test(name)) return false;
+    return !input.delegated || DELEGATED_NAMES.has(name) || DELEGATED_PREFIX.test(name);
+  }));
+}
diff --git a/src/main/services/isolation/isolationPaths.ts b/src/main/services/isolation/isolationPaths.ts
index 5cdcb9a5e..e2ef706d7 100644
--- a/src/main/services/isolation/isolationPaths.ts
+++ b/src/main/services/isolation/isolationPaths.ts
@@ -87,8 +87,13 @@ function providerFolders(provider: ProviderId, env: IsolationPathInput["env"], h
       return { folders: named("opencode"), files: [] };
     case "hermes":
       return { folders: [env.HERMES_HOME && isAbsolute(env.HERMES_HOME) ? env.HERMES_HOME : join(home, ".hermes"), ...named("hermes")], files: [] };
-    case "kimi":
-      return { folders: [env.KIMI_HOME && isAbsolute(env.KIMI_HOME) ? env.KIMI_HOME : join(home, ".kimi"), ...named("kimi")], files: [] };
+    case "kimi": {
+      // Kimi Code CLI keeps its sign-in in KIMI_CODE_HOME (default ~/.kimi-code); the older kimi-cli in
+      // KIMI_SHARE_DIR (default ~/.kimi). KIMI_HOME is kept as it was read before.
+      const moved = (name: string): string | null => { const value = env[name]; return value && isAbsolute(value) ? value : null; };
+      return { folders: [moved("KIMI_CODE_HOME") ?? join(home, ".kimi-code"), moved("KIMI_SHARE_DIR") ?? join(home, ".kimi"),
+        ...(moved("KIMI_HOME") ? [moved("KIMI_HOME")!] : []), ...named("kimi")], files: [] };
+    }
     case "cursor":
       return { folders: [...named("cursor"), ...named("cursor-agent")], files: [] };
     case "antigravity":
@@ -106,7 +111,7 @@ const HOME_VARIABLES: Partial<Record<ProviderId, readonly string[]>> = {
   claude: ["CLAUDE_CONFIG_DIR"],
   grok: ["GROK_HOME"],
   hermes: ["HERMES_HOME"],
-  kimi: ["KIMI_HOME"],
+  kimi: ["KIMI_CODE_HOME", "KIMI_SHARE_DIR", "KIMI_HOME"],
   opencode: ["OPENCODE_CONFIG_DIR"],
   qwen: ["QWEN_HOME"]
 };
@@ -267,7 +272,8 @@ export function isolationPaths(input: IsolationPathInput): IsolationPaths {
       // A repository that exists keeps its config (hooksPath, fsmonitor, filters run code when the person uses git
       // later, outside the layer); a new one may be created, which writes its config.
       ...(existsSync(join(project, ".git", "config")) ? [join(project, ".git", "config"), join(project, ".git", "config.lock")] : []),
-      ...(input.provider === "codex" ? ownFolders.map((folder) => join(folder, "config.toml")) : []),
+      // Codex's and Kimi's config.toml hold their permission rules and hooks.
+      ...(input.provider === "codex" || input.provider === "kimi" ? ownFolders.map((folder) => join(folder, "config.toml")) : []),
       ...(input.provider === "claude" ? ownFolders.flatMap((folder) => [join(folder, "settings.json"), join(folder, "settings.local.json")]) : []),
       ...configSources.jsonFiles
     ]),
diff --git a/src/main/services/materials/HandoffService.ts b/src/main/services/materials/HandoffService.ts
index 043e00877..6f0dba0ef 100644
--- a/src/main/services/materials/HandoffService.ts
+++ b/src/main/services/materials/HandoffService.ts
@@ -63,6 +63,8 @@ export interface HandoffImageOps {
 }
 
 interface HandoffTiming {
+  /** How long a paste waits for the agent's CLI to take input (OpenCode drops keys until its input field is up). */
+  readyMs: number;
   attachMs: number;
   pasteMs: number;
   pollMs: number;
@@ -114,7 +116,7 @@ type Checked = { ok: true; plan: HandoffPlan } | { ok: false; reason: HandoffBlo
 
 type Reserved = { ok: true; plan: HandoffPlan; screen: SessionScreen | null } | { ok: false; reason: HandoffBlock | MaterialFailure };
 
-const DEFAULT_TIMING: HandoffTiming = { attachMs: 8_000, pasteMs: 6_000, pollMs: 120 };
+const DEFAULT_TIMING: HandoffTiming = { readyMs: 15_000, attachMs: 8_000, pasteMs: 6_000, pollMs: 120 };
 
 export class HandoffService {
   private readonly options: HandoffServiceOptions;
@@ -509,7 +511,7 @@ export class HandoffService {
     const paste = async (data: string): Promise<string | null> => {
       const error = this.deliveryError(plan, screen);
       if (error) return error;
-      const written = await this.options.terminals.deliverInput(id, `${PASTE_START}${data}${PASTE_END}`, 0);
+      const written = await this.options.terminals.deliverInput(id, `${PASTE_START}${data}${PASTE_END}`, this.timing.readyMs);
       return written.delivered ? null : written.reason;
     };
     if (!screen) {
@@ -554,7 +556,7 @@ export class HandoffService {
       ? failed(error)
       : { state: "pasted", note: "enter-failed", sentAt: this.now(), imagesAttached: attached, error };
     this.submitting.set(plan.id, { sessionId: id, turnStartedAt: null });
-    const submitted = await this.options.terminals.deliverInput(id, "\r", 0);
+    const submitted = await this.options.terminals.deliverInput(id, "\r", this.timing.readyMs);
     return submitted.delivered
       ? { state: "submitted", sentAt: this.now(), imagesAttached: attached }
       : { state: "pasted", note: "enter-failed", sentAt: this.now(), imagesAttached: attached, error: submitted.reason };
diff --git a/src/main/services/materials/sessionScreen.ts b/src/main/services/materials/sessionScreen.ts
index e39b33f9f..b9f78f8f1 100644
--- a/src/main/services/materials/sessionScreen.ts
+++ b/src/main/services/materials/sessionScreen.ts
@@ -115,9 +115,18 @@ export function pasteMarkers(screen: string): number {
 
 /** Visible screen rows, optionally reconstructing xterm's soft-wrapped logical lines. */
 export function terminalScreenText(terminal: import("@xterm/headless").Terminal, unwrap = false): string {
+  const buffer = terminal.buffer.active;
+  return terminalRowsText(terminal, buffer.baseY, buffer.baseY + terminal.rows, unwrap);
+}
+
+/**
+ * Rows `from` to `to` (exclusive) of the active buffer as text. `unwrap` joins a row xterm wrapped onto the line it
+ * continues (untrimmed, so a space the wrap fell on stays between the words) and trims each line once it is whole.
+ */
+export function terminalRowsText(terminal: import("@xterm/headless").Terminal, from: number, to: number, unwrap = false): string {
   const buffer = terminal.buffer.active;
   const lines: string[] = [];
-  for (let row = buffer.baseY; row < buffer.baseY + terminal.rows; row++) {
+  for (let row = from; row < to; row++) {
     const line = buffer.getLine(row);
     const text = line?.translateToString(!unwrap) ?? "";
     if (unwrap && line?.isWrapped && lines.length) lines[lines.length - 1] += text;
diff --git a/src/main/services/openCodeConfig.ts b/src/main/services/openCodeConfig.ts
index d9e33e91a..b1f911dbf 100644
--- a/src/main/services/openCodeConfig.ts
+++ b/src/main/services/openCodeConfig.ts
@@ -102,7 +102,8 @@ export function openCodeYoloEnvironment(
  * Auto's rules: read, glob, grep, list allowed, except `.env` files (they ask, like OpenCode's default); edit (edit,
  * write, apply_patch) allowed; bash allowed only when `shellGuarded` (CanvasTTY's base protection is on and its guard
  * runs in this OpenCode), otherwise it asks. external_directory is not touched: a path outside the project still asks.
- * `thirdPartyModel` (a launch contributor put OpenCode on another model) keeps bash asking, like accept-edits.
+ * `thirdPartyModel` (a launch contributor put OpenCode on another model) keeps bash asking, like accept-edits, in a
+ * subagent's card too: that ask reaches the person through needs_approval, and the orchestrator cannot answer it.
  */
 export function openCodeAutoEnvironment(
   environment: Readonly<Record<string, string | undefined>>,
diff --git a/src/main/services/renderedTerminalText.ts b/src/main/services/renderedTerminalText.ts
new file mode 100644
index 000000000..bff948199
--- /dev/null
+++ b/src/main/services/renderedTerminalText.ts
@@ -0,0 +1,39 @@
+import { lazyRequire } from "../lazyRequire.ts";
+import { terminalRowsText } from "./materials/sessionScreen.ts";
+
+const xterm = lazyRequire<typeof import("@xterm/headless")>("@xterm/headless");
+
+/** Rows of scrollback the rendering keeps above the screen: the tail an orchestrator reads is far shorter. */
+const SCROLLBACK_ROWS = 500;
+const DEFAULT_GEOMETRY = { cols: 120, rows: 40 };
+
+/**
+ * What a terminal shows for `data` as plain text: the scrollback rows and the screen (the alternate screen of a
+ * full-screen TUI such as OpenCode, which keeps no scrollback), as a headless xterm renders them, so cursor moves,
+ * redraws, erases and colors leave only the text a person sees. A row the terminal wrapped continues its line;
+ * trailing blanks and empty rows at the end are dropped.
+ */
+export async function renderedTerminalText(data: string, geometry: { cols: number; rows: number } | null): Promise<string> {
+  const { cols, rows } = geometry && geometry.cols > 0 && geometry.rows > 0 ? geometry : DEFAULT_GEOMETRY;
+  const terminal = new (xterm().Terminal)({ cols, rows, scrollback: SCROLLBACK_ROWS, allowProposedApi: true });
+  const rowsText = (): string => terminalRowsText(terminal, 0, terminal.buffer.active.length, true).replace(/\n+$/u, "");
+  // A full-screen TUI leaves its alternate screen as it exits, which discards that screen: its last frame (often why it
+  // stopped) is kept from the moment it leaves, before what it printed afterwards.
+  let leftScreen = "";
+  const leaving = terminal.parser.registerCsiHandler({ prefix: "?", final: "l" }, (params) => {
+    if (terminal.buffer.active.type === "alternate" && params.some((mode) => mode === 1049 || mode === 1047 || mode === 47)) {
+      leftScreen = rowsText();
+    }
+    return false;
+  });
+  try {
+    await new Promise<void>((resolve) => terminal.write(data, resolve));
+    // Every row, scrollback included (sessionScreen's terminalScreenText reads only the visible screen of a live one).
+    const shown = rowsText();
+    if (terminal.buffer.active.type === "alternate" || !leftScreen) return shown;
+    return shown ? `${leftScreen}\n${shown}` : leftScreen;
+  } finally {
+    leaving.dispose();
+    terminal.dispose();
+  }
+}
diff --git a/src/main/services/shutdownSteps.ts b/src/main/services/shutdownSteps.ts
new file mode 100644
index 000000000..670249712
--- /dev/null
+++ b/src/main/services/shutdownSteps.ts
@@ -0,0 +1,13 @@
+export interface ShutdownStep {
+  name: string;
+  run(): void | Promise<unknown>;
+}
+
+/** Cleanup remains ordered, but one failed save must not skip independent services. */
+export async function runShutdownSteps(steps: readonly ShutdownStep[],
+  warn: (message: string, error: unknown) => void = console.warn): Promise<void> {
+  for (const step of steps) {
+    try { await step.run(); }
+    catch (error) { warn(`CanvasTTY shutdown: ${step.name} failed.`, error); }
+  }
+}
diff --git a/src/preload/index.ts b/src/preload/index.ts
index 32f38cf65..dbde3cf91 100644
--- a/src/preload/index.ts
+++ b/src/preload/index.ts
@@ -254,6 +254,7 @@ const api: CanvasTTYApi = {
     status: () => ipcRenderer.invoke(IPC.githubAuthStatus),
     start: () => ipcRenderer.invoke(IPC.githubAuthStart),
     signOut: () => ipcRenderer.invoke(IPC.githubAuthSignOut),
+    cancel: () => ipcRenderer.invoke(IPC.githubAuthCancel),
     openUrl: (url: string) => ipcRenderer.invoke(IPC.githubAuthOpenUrl, url)
   },
   browser: {
diff --git a/src/renderer/src/App.tsx b/src/renderer/src/App.tsx
index cb72f6485..362f1bd2c 100644
--- a/src/renderer/src/App.tsx
+++ b/src/renderer/src/App.tsx
@@ -56,6 +56,7 @@ import {
 } from "./lib/skinStyles";
 import { TitleBar } from "./components/TitleBar";
 import { Toast } from "./components/Toast";
+import { pluginServiceNoticeKind, pluginServiceNoticeSection, pluginServiceNoticeText } from "../../shared/pluginServiceNotice.ts";
 import { environmentOptions } from "./features/launcher/LaunchOptionsSection";
 import { UpdateNotice } from "./components/UpdateNotice";
 import { resolveAppearanceSettings } from "./features/settings/appearanceSettings";
@@ -362,6 +363,7 @@ export function App(): React.JSX.Element {
   const [launchPosition, setLaunchPosition] = useState<Point | null>(null);
   const [settingsOpen, setSettingsOpen] = useState(false);
   const [openUpdatesRequest, setOpenUpdatesRequest] = useState(0);
+  const [openSectionRequest, setOpenSectionRequest] = useState<{ section: "agents" | "plugins"; id: number } | null>(null);
   const [shortcutReferenceOpen, setShortcutReferenceOpen] = useState(false);
   const [usageHistoryOpen, setUsageHistoryOpen] = useState(false);
   /** App root outside the canvas transform; overlays portal here to inherit the theme tokens. */
@@ -373,6 +375,7 @@ export function App(): React.JSX.Element {
   const [fullscreenSessionId, setFullscreenSessionId] = useState<string | null>(null);
   const [pendingTerminalUrl, setPendingTerminalUrl] = useState<string | null>(null);
   const [toast, setToast] = useState<string | null>(null);
+  const [toastAction, setToastAction] = useState<{ label: string; run(): void } | null>(null);
   const [updateStatus, setUpdateStatus] = useState<UpdateStatus>({ type: "idle" });
   const [dismissedUpdateNotice, setDismissedUpdateNotice] = useState<string | null>(null);
   const [updateNoticePending, setUpdateNoticePending] = useState(false);
@@ -391,7 +394,29 @@ export function App(): React.JSX.Element {
     fullscreen: false
   });
 
-  const showToast = useCallback((message: string): void => setToast(message), []);
+  const showToast = useCallback((message: string): void => {
+    setToastAction(null);
+    setToast(message);
+  }, []);
+  // A plugin page could not reach its service because its native code or module is off: say so with the way there.
+  const showPluginError = useCallback((message: string): void => {
+    const notice = pluginServiceNoticeKind(message);
+    if (!notice) {
+      showToast(message);
+      return;
+    }
+    const section = pluginServiceNoticeSection(notice);
+    setToast(pluginServiceNoticeText(notice, settings.locale));
+    setToastAction({
+      label: pluginServiceNoticeText("open", settings.locale),
+      run: () => {
+        setToast(null);
+        setToastAction(null);
+        setSettingsOpen(true);
+        setOpenSectionRequest((current) => ({ section, id: (current?.id ?? 0) + 1 }));
+      }
+    });
+  }, [settings.locale, showToast]);
   const fileSessionOptions = useMemo<FileSessionOption[]>(
     () => sessions.map((session) => ({ sessionId: session.id, label: session.title })),
     [sessions]
@@ -468,9 +493,13 @@ export function App(): React.JSX.Element {
 
   useEffect(() => {
     if (!toast) return;
-    const timer = window.setTimeout(() => setToast(null), 2_600);
+    // A toast with a button stays long enough to reach it.
+    const timer = window.setTimeout(() => {
+      setToast(null);
+      setToastAction(null);
+    }, toastAction ? 8_000 : 2_600);
     return () => window.clearTimeout(timer);
-  }, [toast]);
+  }, [toast, toastAction]);
 
   useEffect(() => {
     const unsubscribe = window.canvasTTY.window.onState(setWindowState);
@@ -2005,7 +2034,7 @@ export function App(): React.JSX.Element {
           onHomeGridSizeChange={changeHomeGridSize}
           onFinishHomeEdit={() => void finishHomeEditor()}
           onResetHomeLayout={resetHomeLayout}
-          onPluginError={showToast}
+          onPluginError={showPluginError}
           onPluginCanvasBoundsChange={changePluginCanvasBounds}
           onDisposePluginCanvas={disposePluginCanvas}
           onFocusPluginCanvas={focusPluginCanvas}
@@ -2111,6 +2140,7 @@ export function App(): React.JSX.Element {
         <SettingsPanel
           open={settingsOpen}
           openUpdatesRequest={openUpdatesRequest}
+          openSectionRequest={openSectionRequest}
           settings={settings}
           agentAvailability={agentAvailability}
           onRecheckAgentClis={recheckAgentClis}
@@ -2180,7 +2210,7 @@ export function App(): React.JSX.Element {
       {materials.snapshot.loadError && (
         <div className="materials-load-error" role="alert">{t(settings.locale, "materialsLoadFailed")}</div>
       )}
-      <Toast message={toast} />
+      <Toast message={toast} action={toastAction} />
     </div>
   );
 }
diff --git a/src/renderer/src/components/ShortcutReference.tsx b/src/renderer/src/components/ShortcutReference.tsx
index 2f736c8af..8752d086a 100644
--- a/src/renderer/src/components/ShortcutReference.tsx
+++ b/src/renderer/src/components/ShortcutReference.tsx
@@ -1,6 +1,6 @@
 import { useEffect, useRef } from "react";
 import type { AppSettings } from "../../../shared/contracts";
-import { displayCanvasNavigationBinding } from "../lib/shortcuts";
+import { displayCanvasNavigationBinding, displayShortcutBinding } from "../lib/shortcuts";
 import { UiIcon } from "./UiIcon";
 import { t, type TranslationKey } from "../lib/i18n";
 
@@ -11,9 +11,7 @@ export function ShortcutReference({ settings, onClose }: {
   const dialog = useRef<HTMLDialogElement>(null);
   const { locale, shortcuts: bindings } = settings;
   const mac = window.canvasTTY.window.isMacOS;
-  const display = (binding: string) => binding.replace("Meta", mac ? "Command" : "Super")
-    .replace("Comma", ",").replace(/ArrowUp/g, "↑").replace(/ArrowDown/g, "↓")
-    .replace(/ArrowLeft/g, "←").replace(/ArrowRight/g, "→") || t(locale, "disabled");
+  const display = (binding: string) => displayShortcutBinding(binding, mac) || t(locale, "disabled");
   const groups: { title: TranslationKey; shortcuts: [string[], TranslationKey][] }[] = [
     { title: "keyboardCanvas", shortcuts: [
       [[bindings.home], "homeShortcut"],
@@ -22,7 +20,7 @@ export function ShortcutReference({ settings, onClose }: {
       [[bindings.commandPalette], "commandPalette"],
       [[bindings.openSettings], "settings"],
       [[bindings.toggleDetail], "keyboardToggleDetail"],
-      [["Shift + drag"], "marqueeSelectionHint"]
+      [[t(locale, "shortcutShiftDrag")], "marqueeSelectionHint"]
     ] },
     { title: "keyboardNavigation", shortcuts: [
       [[bindings.focusUp], "keyboardFocusUp"],
diff --git a/src/renderer/src/components/Toast.tsx b/src/renderer/src/components/Toast.tsx
index 190760921..f25700112 100644
--- a/src/renderer/src/components/Toast.tsx
+++ b/src/renderer/src/components/Toast.tsx
@@ -1,7 +1,14 @@
 interface ToastProps {
   message: string | null;
+  /** One button next to the text (for example "Open settings"); the toast takes clicks only while it has one. */
+  action?: { label: string; run(): void } | null;
 }
 
-export function Toast({ message }: ToastProps): React.JSX.Element {
-  return <div className={`toast ${message ? "toast--visible" : ""}`}>{message}</div>;
+export function Toast({ message, action }: ToastProps): React.JSX.Element {
+  return (
+    <div className={`toast ${message ? "toast--visible" : ""}${message && action ? " toast--action" : ""}`} role={message && action ? "alert" : undefined}>
+      {message}
+      {message && action && <button className="toast__action" type="button" onClick={action.run}>{action.label}</button>}
+    </div>
+  );
 }
diff --git a/src/renderer/src/features/home/HomeZone.tsx b/src/renderer/src/features/home/HomeZone.tsx
index 3366a235a..7aa4c0ed7 100644
--- a/src/renderer/src/features/home/HomeZone.tsx
+++ b/src/renderer/src/features/home/HomeZone.tsx
@@ -533,8 +533,11 @@ function LimitRow({ row, locale, now }: { row: HomeLimitRow; locale: LocaleId; n
       {row.window ? (
         <span className="limit-row__metric">
           <span className="limit-row__summary">
-            <strong>{formatPercent(row.window.usedPercent, locale)}</strong>
-            <span>{formatResetCountdown(row.window.resetsAt, now, locale)}</span>
+            <span className="limit-row__share">
+              <strong>{formatPercent(row.window.usedPercent, locale)}</strong>
+              <span>{t(locale, "limitUsed")} · {limitLeft(row.window.usedPercent, locale)}</span>
+            </span>
+            <span className="limit-row__countdown">{formatResetCountdown(row.window.resetsAt, now, locale)}</span>
           </span>
           <span className="limit-row__track" aria-hidden="true">
             <i className="limit-row__fill" style={{ width: `${row.window.usedPercent}%` }} />
@@ -552,7 +555,8 @@ function limitMeta(row: HomeLimitRow, locale: LocaleId): string {
 
   const details = [
     formatLimitDuration(row.window.windowMinutes, locale),
-    `${formatPercent(row.window.usedPercent, locale)} ${t(locale, "limitUsed")}`,
+    t(locale, "limitUsedPercent").replace("{percent}", formatPercent(row.window.usedPercent, locale)),
+    limitLeft(row.window.usedPercent, locale),
     formatReset(row.window.resetsAt, locale)
   ];
   if (row.state === "stale") {
@@ -572,6 +576,11 @@ function formatReset(resetsAt: number, locale: LocaleId): string {
   return `${t(locale, "limitResetsAt")} ${formatted}`;
 }
 
+/** What is left of a window whose used share is `usedPercent` (both are always shown, never a bare percent). */
+function limitLeft(usedPercent: number, locale: LocaleId): string {
+  return t(locale, "limitLeftPercent").replace("{percent}", formatPercent(Math.max(0, 100 - usedPercent), locale));
+}
+
 function formatPercent(value: number, locale: LocaleId): string {
   return new Intl.NumberFormat(locale === "ru" ? "ru-RU" : "en-GB", {
     maximumFractionDigits: 1
diff --git a/src/renderer/src/features/plugins/PluginFrame.tsx b/src/renderer/src/features/plugins/PluginFrame.tsx
index 407230343..9747e8b64 100644
--- a/src/renderer/src/features/plugins/PluginFrame.tsx
+++ b/src/renderer/src/features/plugins/PluginFrame.tsx
@@ -17,6 +17,7 @@ import {
 import { createFrameReplyGate } from "./frameReplies";
 import { pluginVisibilityMessage } from "../../../../shared/pluginVisibility";
 import { isProviderId } from "../../../../shared/providerCatalog.ts";
+import { pluginServiceNoticeKind, pluginServiceNoticeText } from "../../../../shared/pluginServiceNotice.ts";
 
 const storageListeners = new Map<string, Set<(key: string, value: unknown) => void>>();
 
@@ -176,7 +177,10 @@ export function PluginFrame({
           value
         });
       }).catch((error: unknown) => {
-        const description = safeError(error);
+        // Native code off / module off: the page and the host notice get the localized text with its way out.
+        const raw = safeError(error);
+        const notice = pluginServiceNoticeKind(raw);
+        const description = notice ? pluginServiceNoticeText(notice, locale) : raw;
         onError(description);
         if (!mayReply(frame.current?.contentWindow, servedPlugin.current)) return;
         postToFrame(frame.current, {
@@ -190,7 +194,7 @@ export function PluginFrame({
     };
     window.addEventListener("message", receive);
     return () => window.removeEventListener("message", receive);
-  }, [canvasInstanceId, captureCanvasWheelOverWidgets, context, limits, onCanvasWheel, onError, onFocus, onHoverChange, onOpenLauncher, onMascotDrag, plugin, replies, sessions]);
+  }, [canvasInstanceId, captureCanvasWheelOverWidgets, context, limits, locale, onCanvasWheel, onError, onFocus, onHoverChange, onOpenLauncher, onMascotDrag, plugin, replies, sessions]);
 
   useEffect(() => {
     postToFrame(frame.current, { source: "canvastty-host", type: "context", value: context });
diff --git a/src/renderer/src/features/plugins/PluginSettingsSection.tsx b/src/renderer/src/features/plugins/PluginSettingsSection.tsx
index 666c01bc2..537b90e36 100644
--- a/src/renderer/src/features/plugins/PluginSettingsSection.tsx
+++ b/src/renderer/src/features/plugins/PluginSettingsSection.tsx
@@ -2,6 +2,7 @@ import { useEffect, useLayoutEffect, useRef, useState } from "react";
 import type {
   AppSettings,
   GithubAuthStatus,
+  GithubDeviceFlowState,
   GithubDeviceFlowStart,
   GithubPluginSearchResult,
   InstalledPlugin,
@@ -181,6 +182,7 @@ export function PluginSettingsSection({
   const [githubCode, setGithubCode] = useState<GithubDeviceFlowStart | null>(null);
   const [githubBusy, setGithubBusy] = useState(false);
   const [codeCopied, setCodeCopied] = useState(false);
+  const githubFlowNotice = githubFlowNoticeText(locale, githubStatus?.deviceFlowState);
   const [showcasePreviews, setShowcasePreviews] = useState<Record<string, PluginInstallPreview>>({});
   const [showcaseManifests, setShowcaseManifests] = useState<Record<string, PluginManifest>>({});
   const [selectedShowcase, setSelectedShowcase] = useState<string | null>(null);
@@ -197,28 +199,27 @@ export function PluginSettingsSection({
     // The main process owns the device flow; this only reads its status, so it pauses while Settings is
     // closed and picks up (sign-in finished or code expired) as soon as Settings opens again.
     const stop = pollWhileOpen(open, githubCode.interval * 1000, async () => {
-      if (Date.now() >= githubCode.expiresAt) {
-        if (!cancelled) {
-          setGithubCode(null);
-          setError(t(locale, "githubAuthExpired"));
-        }
-        return STOP_POLLING;
-      }
       // A transient status read failure is retried on the next tick.
       const status = await window.canvasTTY.githubAuth.status();
       if (cancelled) return STOP_POLLING;
-      setGithubStatus(status);
-      if (status.authorized) {
+      if (status.authorized || isGithubFlowOutcome(status.deviceFlowState)) {
+        setGithubStatus(status);
+        setGithubCode(null);
+        return STOP_POLLING;
+      }
+      if (Date.now() >= githubCode.expiresAt) {
+        setGithubStatus({ ...status, deviceFlowState: "expired" });
         setGithubCode(null);
         return STOP_POLLING;
       }
+      setGithubStatus(status);
       return undefined;
     }, { immediate: false });
     return () => {
       cancelled = true;
       stop();
     };
-  }, [githubCode, locale, open]);
+  }, [githubCode, open]);
 
   const inspect = async (): Promise<void> => {
     if (busy || sourceUrl.trim().length === 0) return;
@@ -365,10 +366,22 @@ export function PluginSettingsSection({
     if (githubBusy) return;
     setGithubBusy(true);
     setError(null);
+    let flow: GithubDeviceFlowStart;
+    try {
+      flow = await window.canvasTTY.githubAuth.start();
+    } catch (reason) {
+      // The IPC error is main-process text ("Error invoking remote method …", HTTP codes): it goes to the log, and the
+      // GitHub section shows its own localized line (not configured, or sign-in failed).
+      console.warn("GitHub sign-in could not start.", reason);
+      const status = await window.canvasTTY.githubAuth.status().catch(() => null);
+      setGithubStatus(status && !status.configured ? status
+        : { configured: true, authorized: false, login: null, tokenExpiresAt: null, ...status, deviceFlowState: "failed" });
+      setGithubBusy(false);
+      return;
+    }
     try {
-      const flow = await window.canvasTTY.githubAuth.start();
       setGithubCode(flow);
-      setGithubStatus({ configured: true, authorized: false, login: null, tokenExpiresAt: null });
+      setGithubStatus({ configured: true, authorized: false, login: null, tokenExpiresAt: null, deviceFlowState: "pending" });
       await openGithubAuthorization(flow, target);
     } catch (reason) {
       setError(errorMessage(reason, t(locale, "githubAuthNotConfigured")));
@@ -387,7 +400,8 @@ export function PluginSettingsSection({
         configured: current?.configured ?? false,
         authorized: false,
         login: null,
-        tokenExpiresAt: null
+        tokenExpiresAt: null,
+        deviceFlowState: "idle"
       }));
       setGithubCode(null);
     } finally {
@@ -395,6 +409,24 @@ export function PluginSettingsSection({
     }
   };
 
+  const runGithubCancel = async (): Promise<void> => {
+    if (githubBusy) return;
+    setGithubBusy(true);
+    setError(null);
+    try {
+      await window.canvasTTY.githubAuth.cancel();
+      // The poll may have finished first; the status says which outcome won.
+      setGithubStatus(await window.canvasTTY.githubAuth.status());
+      setGithubCode(null);
+      setCodeCopied(false);
+    } catch (reason) {
+      console.warn("GitHub sign-in could not be cancelled.", reason);
+      setError(githubSigninText(locale, "cancelFailed"));
+    } finally {
+      setGithubBusy(false);
+    }
+  };
+
   const selectShowcaseTile = (fullName: string): void => {
     if (busy) return;
     setSelectedShowcase(fullName);
@@ -833,6 +865,12 @@ export function PluginSettingsSection({
             </span>
           )}
         </h3>
+        {!githubAuthorized && !githubCode && githubFlowNotice && (
+          // A cancelled sign-in is the person's own choice; a denied, expired or failed one is an error.
+          githubStatus?.deviceFlowState === "cancelled"
+            ? <p className="plugin-github-unavailable" role="status">{githubFlowNotice}</p>
+            : <p className="plugin-settings__error" role="alert">{githubFlowNotice}</p>
+        )}
         {!githubAuthorized && !githubCode && <p className="plugin-github-optional">{t(locale, "githubAuthOptional")}</p>}
         {githubStatus?.authorized ? (
           <div>
@@ -882,6 +920,9 @@ export function PluginSettingsSection({
                 {t(locale, "githubAuthOpenExternal")}
               </button>
             </div>
+            <button type="button" className="plugin-github-signout" disabled={githubBusy} onClick={() => void runGithubCancel()}>
+              {githubSigninText(locale, "cancel")}
+            </button>
           </div>
         ) : githubStatus && !githubConfigured ? (
           <p className="plugin-github-unavailable" role="status">{t(locale, "githubAuthNotConfigured")}</p>
@@ -1154,3 +1195,39 @@ function errorMessage(error: unknown, fallback: string): string {
     .replace(/[\u0000-\u001f\u007f]/g, " ")
     .slice(0, 320);
 }
+
+function isGithubFlowOutcome(state: GithubDeviceFlowState | undefined): boolean {
+  return state === "denied" || state === "expired" || state === "failed" || state === "cancelled";
+}
+
+// The sign-in outcome lines live here, in the lazily loaded Settings chunk, not in the startup string table.
+const GITHUB_SIGNIN_TEXT = {
+  ru: {
+    denied: "Запрос на вход в GitHub отклонён. Можно попробовать снова.",
+    failed: "Не удалось завершить вход в GitHub. Можно попробовать снова.",
+    cancelled: "Вход в GitHub отменён.",
+    cancel: "Отменить вход",
+    cancelFailed: "Не удалось отменить вход в GitHub. Подробности записаны в журнал."
+  },
+  en: {
+    denied: "The GitHub sign-in request was denied. You can try again.",
+    failed: "GitHub sign-in could not be completed. You can try again.",
+    cancelled: "GitHub sign-in was cancelled.",
+    cancel: "Cancel sign-in",
+    cancelFailed: "GitHub sign-in could not be cancelled. Details are in the log."
+  }
+} satisfies Record<LocaleId, Record<string, string>>;
+
+function githubSigninText(locale: LocaleId, key: keyof typeof GITHUB_SIGNIN_TEXT.en): string {
+  return GITHUB_SIGNIN_TEXT[locale === "ru" ? "ru" : "en"][key];
+}
+
+function githubFlowNoticeText(locale: LocaleId, state: GithubDeviceFlowState | undefined): string | null {
+  switch (state) {
+    case "denied": return githubSigninText(locale, "denied");
+    case "expired": return t(locale, "githubAuthExpired");
+    case "failed": return githubSigninText(locale, "failed");
+    case "cancelled": return githubSigninText(locale, "cancelled");
+    default: return null;
+  }
+}
diff --git a/src/renderer/src/features/settings/SettingsPanel.tsx b/src/renderer/src/features/settings/SettingsPanel.tsx
index db4619189..e1ae02ff4 100644
--- a/src/renderer/src/features/settings/SettingsPanel.tsx
+++ b/src/renderer/src/features/settings/SettingsPanel.tsx
@@ -84,6 +84,7 @@ import {
 } from "../../lib/providers";
 import { shortcutFromKeyboardEvent, shortcutFromPointerEvent } from "../../lib/shortcuts";
 import { t, type TranslationKey } from "../../lib/i18n";
+import { canvasNavigationText } from "./canvasNavigationText";
 import { formatBytes } from "../materials/materialCardModel";
 import {
   createTerminalBorderSkinPreviewStyleController,
@@ -172,6 +173,8 @@ const DEFAULT_PROFILE_LABELS: Record<DefaultLaunchProfile, TranslationKey> = {
 interface SettingsPanelProps {
   open: boolean;
   openUpdatesRequest: number;
+  /** Opens a section from outside (a plugin notice's "Open settings"); "agents" scrolls to Extension native code. */
+  openSectionRequest?: { section: "agents" | "plugins"; id: number } | null;
   settings: AppSettings;
   agentAvailability: AgentCliAvailability | null;
   onRecheckAgentClis(): Promise<void>;
@@ -208,6 +211,7 @@ interface SettingsPanelProps {
 export function SettingsPanel({
   open,
   openUpdatesRequest,
+  openSectionRequest = null,
   settings,
   agentAvailability,
   onRecheckAgentClis,
@@ -289,6 +293,19 @@ export function SettingsPanel({
     if (open && contentRef.current) contentRef.current.scrollTop = 0;
   }, [open, section]);
 
+  const scrollToNativeCode = useRef(false);
+  useEffect(() => {
+    if (!openSectionRequest) return;
+    scrollToNativeCode.current = openSectionRequest.section === "agents";
+    setSection(openSectionRequest.section);
+  }, [openSectionRequest]);
+  // Once, after the section is shown (and its scroll reset): bring Extension native code into view.
+  useEffect(() => {
+    if (!open || section !== "agents" || !scrollToNativeCode.current) return;
+    scrollToNativeCode.current = false;
+    contentRef.current?.querySelector<HTMLElement>(".plugin-services")?.scrollIntoView({ block: "start" });
+  }, [open, openSectionRequest, section]);
+
   const openAgentInstall = (provider: AgentProviderId): void => {
     const url = PROVIDERS[provider].installUrl;
     if (!url) return;
@@ -626,20 +643,20 @@ export function SettingsPanel({
                     ))}
                   </div>
                 ))}
-                <SettingGroup label={t(locale, "canvasWheelCapture")} description={t(locale, "canvasWheelCaptureDescription")}>
+                <SettingGroup label={canvasNavigationText(locale, "canvasWheelCapture")} description={canvasNavigationText(locale, "canvasWheelCaptureDescription")}>
                   <Segmented value={settings.canvasWheelCaptureMode}
-                    options={[["off", "Off"], ["always", "On"], ["key", "Key"]]}
+                    options={[["off", t(locale, "off")], ["always", t(locale, "on")], ["key", canvasNavigationText(locale, "canvasWheelCaptureKey")]]}
                     onChange={(value) => changeCanvasWheelCaptureMode(value as CanvasWheelCaptureMode)} />
                   {settings.canvasWheelCaptureMode === "key" && (
                     <CanvasNavigationShortcutEditor open={open} locale={locale}
-                      label={t(locale, "canvasWheelOverride")} binding={settings.canvasWheelOverride}
+                      label={canvasNavigationText(locale, "canvasWheelOverride")} binding={settings.canvasWheelOverride}
                       actionShortcuts={Object.values(settings.shortcuts)} allowDisable={false}
                       onCaptureStart={() => { setCapturing(null); setShortcutError(null); }}
                       onChange={(canvasWheelOverride) => onChange({ keyboardPreset: "custom", canvasWheelOverride })} />
                   )}
-                  {canvasOverrideBindingsMatch && <p className="shortcut-editor__warning">{t(locale, "canvasOverrideBindingsMatch")}</p>}
+                  {canvasOverrideBindingsMatch && <p className="shortcut-editor__warning">{canvasNavigationText(locale, "canvasOverrideBindingsMatch")}</p>}
                 </SettingGroup>
-                <SettingGroup label={t(locale, "canvasNavigationOverride")} description={t(locale, "canvasNavigationOverrideDescription")}>
+                <SettingGroup label={t(locale, "canvasNavigationOverride")} description={canvasNavigationText(locale, "canvasNavigationOverrideDescription")}>
                   <CanvasNavigationShortcutEditor open={open} locale={locale}
                     label={t(locale, "canvasNavigationOverride")} binding={settings.canvasNavigationOverride}
                     actionShortcuts={Object.values(settings.shortcuts)} allowDisable
diff --git a/src/renderer/src/features/settings/canvasNavigationText.ts b/src/renderer/src/features/settings/canvasNavigationText.ts
new file mode 100644
index 000000000..b7d32db34
--- /dev/null
+++ b/src/renderer/src/features/settings/canvasNavigationText.ts
@@ -0,0 +1,28 @@
+import type { LocaleId } from "../../../../shared/contracts";
+
+// The canvas wheel and navigation settings' text. Only Settings shows it, so it lives here, in the lazily loaded
+// Settings chunk, rather than in the startup string table.
+const CANVAS_NAVIGATION_TEXT = {
+  ru: {
+    canvasWheelCapture: "Перехват колеса и жеста масштаба над виджетами",
+    canvasWheelCaptureDescription: "«Выключено» оставляет обычную прокрутку колесом и жест масштаба виджету с фокусом ввода, включая открытую страницу браузера. «Включено», а также «По комбинации» при нажатой комбинации, передают их канвасу; жест масштаба и Ctrl/Command + прокрутка над браузером всегда масштабируют канвас. Клики и перетаскивание не перехватываются.",
+    canvasWheelCaptureKey: "По комбинации",
+    canvasWheelOverride: "Комбинация для колеса и жеста масштаба",
+    canvasOverrideBindingsMatch: "Комбинации совпадают: полный перехват имеет приоритет и также забирает перетаскивание.",
+    canvasNavigationOverrideDescription: "Пока комбинация или кнопка мыши удерживается, прокрутка колесом, жест масштаба и перетаскивание над любым виджетом принадлежат канвасу. Перетаскивание с нажатым колёсиком работает всегда."
+  },
+  en: {
+    canvasWheelCapture: "Wheel/pinch capture over widgets",
+    canvasWheelCaptureDescription: "Off leaves plain wheel/pinch with the focused input widget, including a live Browser page. On and an active binding route it to the canvas; Browser pinch and Ctrl/Command + scroll always zoom the canvas. Clicks and drag are not captured.",
+    canvasWheelCaptureKey: "Key",
+    canvasWheelOverride: "Wheel/pinch shortcut",
+    canvasOverrideBindingsMatch: "The shortcuts match: full navigation takes priority and also captures drag.",
+    canvasNavigationOverrideDescription: "While the shortcut or mouse button is held, wheel, pinch, and drag over every widget belong to the canvas. Middle-button drag always pans."
+  }
+} satisfies Record<LocaleId, Record<string, string>>;
+
+export type CanvasNavigationTextKey = keyof typeof CANVAS_NAVIGATION_TEXT.en;
+
+export function canvasNavigationText(locale: LocaleId, key: CanvasNavigationTextKey): string {
+  return CANVAS_NAVIGATION_TEXT[locale === "ru" ? "ru" : "en"][key];
+}
diff --git a/src/renderer/src/features/usage/UsageHistoryView.tsx b/src/renderer/src/features/usage/UsageHistoryView.tsx
index a5e06e05d..83847feba 100644
--- a/src/renderer/src/features/usage/UsageHistoryView.tsx
+++ b/src/renderer/src/features/usage/UsageHistoryView.tsx
@@ -23,6 +23,7 @@ import {
   formatTimeZone,
   formatTokens,
   providerName,
+  quotaShareText,
   shortScope,
   usageText
 } from "./usageText.ts";
@@ -264,7 +265,7 @@ function QuotaWindow({ quota, format }: { quota: WindowReport; format: Format })
       {quota.accounts.length === 0 && quota.latest !== null && (
         <p className="usage-history__note">
           {fill(text.noObservationsInPeriod, {
-            percent: format.points(quota.latest.percent),
+            share: quotaShareText(format.locale, quota.latest.percent),
             time: format.time(quota.latest.at),
             account: quota.latest.scope ? shortScope(quota.latest.scope) : text.accountUnknown
           })}
@@ -305,7 +306,7 @@ function AccountQuota({ account, format }: { account: AccountSeries; format: For
         </div>
         <div>
           <dt>{text.latest}</dt>
-          <dd>{account.latest ? fill(text.latestValue, { percent: format.points(account.latest.percent), time: format.time(account.latest.at) }) : "—"}</dd>
+          <dd>{account.latest ? fill(text.latestValue, { share: quotaShareText(format.locale, account.latest.percent), time: format.time(account.latest.at) }) : "—"}</dd>
         </div>
         <div>
           <dt>{text.reset}</dt>
diff --git a/src/renderer/src/features/usage/usageText.ts b/src/renderer/src/features/usage/usageText.ts
index 8b96bb9ed..1470a0380 100644
--- a/src/renderer/src/features/usage/usageText.ts
+++ b/src/renderer/src/features/usage/usageText.ts
@@ -45,14 +45,15 @@ const RU = {
   account: "Аккаунт (хеш)",
   accountUnknown: "аккаунт неизвестен",
   latest: "Последнее наблюдение",
-  latestValue: "{percent}% · {time}",
+  latestValue: "{share} · {time}",
+  quotaShare: "использовано {used}% · осталось {left}%",
   reset: "Сброс окна",
   resetUnknown: "неизвестен",
   resetsInPeriod: "Сбросов окна за период: {count}",
   resetPassed: "Сброс уже наступил после последнего наблюдения — текущее значение неизвестно.",
   windowStale: "Последнее наблюдение {age} назад — ряд устарел, новые изменения не измеряются.",
   windowNeverObserved: "Нет наблюдений до конца периода.",
-  noObservationsInPeriod: "За период наблюдений нет. Последнее до него: {percent}% · {time} · {account}.",
+  noObservationsInPeriod: "За период наблюдений нет. Последнее до него: {share} · {time} · {account}.",
   unknownProviderName: "Провайдер не определён",
   measured: "Измерено",
   measuredValue: "+{points} п. п.",
@@ -206,14 +207,15 @@ const EN: UsageText = {
   account: "Account (hash)",
   accountUnknown: "account unknown",
   latest: "Latest observation",
-  latestValue: "{percent}% · {time}",
+  latestValue: "{share} · {time}",
+  quotaShare: "{used}% used · {left}% left",
   reset: "Window reset",
   resetUnknown: "unknown",
   resetsInPeriod: "Window resets in period: {count}",
   resetPassed: "The reset has passed since the latest observation — the current value is unknown.",
   windowStale: "Latest observation {age} ago — the series is stale and new changes are not measured.",
   windowNeverObserved: "No observations up to the end of the period.",
-  noObservationsInPeriod: "No observations in this period. Latest before it: {percent}% · {time} · {account}.",
+  noObservationsInPeriod: "No observations in this period. Latest before it: {share} · {time} · {account}.",
   unknownProviderName: "Provider not identified",
   measured: "Measured",
   measuredValue: "+{points} pp",
@@ -329,6 +331,12 @@ export function fill(template: string, values: Record<string, string | number>):
   return template.replace(/\{(\w+)\}/g, (match, key: string) => (key in values ? String(values[key]) : match));
 }
 
+/** A quota observation (the share of the window already used) with what is left of it: never a bare percent. */
+export function quotaShareText(locale: LocaleId, usedPercent: number): string {
+  const used = Math.min(100, Math.max(0, usedPercent));
+  return fill(usageText(locale).quotaShare, { used: formatPoints(locale, used), left: formatPoints(locale, 100 - used) });
+}
+
 export function formatPoints(locale: LocaleId, value: number): string {
   return new Intl.NumberFormat(locale, { maximumFractionDigits: 2, minimumFractionDigits: 0 }).format(value);
 }
diff --git a/src/renderer/src/features/workspace/WorkspaceCanvas.tsx b/src/renderer/src/features/workspace/WorkspaceCanvas.tsx
index 6fef8d6a7..f445f5ee6 100644
--- a/src/renderer/src/features/workspace/WorkspaceCanvas.tsx
+++ b/src/renderer/src/features/workspace/WorkspaceCanvas.tsx
@@ -30,7 +30,7 @@ import type {
 } from "../../../../shared/contracts";
 import { UiIcon } from "../../components/UiIcon";
 import { t } from "../../lib/i18n";
-import { displayCanvasNavigationBinding, isRenameInputTarget, isShortcutCaptureTarget, matchesPhysicalOrLayoutKey, matchesShortcut, shouldKeepNativeKeyboardInput } from "../../lib/shortcuts";
+import { displayCanvasNavigationBinding, displayShortcutBinding, isRenameInputTarget, isShortcutCaptureTarget, matchesPhysicalOrLayoutKey, matchesShortcut, shouldKeepNativeKeyboardInput } from "../../lib/shortcuts";
 import { BrowserCard } from "../browser/BrowserCard";
 import type { FileBrowserCardProps, FileSessionOption } from "../files/FileBrowserCard";
 import { fileCardRuntimeFor, type FileCardRuntimeState } from "../files/fileCardRuntime";
@@ -1671,11 +1671,11 @@ export function WorkspaceCanvas(props: WorkspaceCanvasProps): React.JSX.Element
             )}
             {settings.showShortcutHints && settings.shortcutHintsPlacement === placement && (
               <aside className="shortcut-hints" aria-label={t(settings.locale, "keyboardShortcuts")}>
-                <div><kbd>{settings.shortcuts.home}</kbd><span>{t(settings.locale, "homeShortcut")}</span></div>
-                <div><kbd>{settings.shortcuts.renameWindow}</kbd><span>{t(settings.locale, "renameWindow")}</span></div>
-                <div><kbd>{settings.shortcuts.toggleFullscreen.replace("Meta", window.canvasTTY.window.isMacOS ? "Command" : "Super")}</kbd><span>{t(settings.locale, "toggleFullscreen")}</span></div>
-                <div><kbd>{settings.shortcuts.focusUp}</kbd><span>{t(settings.locale, "keyboardFocusUp")}</span></div>
-                <div><kbd>Shift + drag</kbd><span>{t(settings.locale, "marqueeSelectionHint")}</span></div>
+                {([["home", "homeShortcut"], ["renameWindow", "renameWindow"], ["toggleFullscreen", "toggleFullscreen"],
+                  ["focusUp", "keyboardFocusUp"]] as const).map(([action, label]) => (
+                  <div key={action}><kbd>{displayShortcutBinding(settings.shortcuts[action], window.canvasTTY.window.isMacOS)}</kbd><span>{t(settings.locale, label)}</span></div>
+                ))}
+                <div><kbd>{t(settings.locale, "shortcutShiftDrag")}</kbd><span>{t(settings.locale, "marqueeSelectionHint")}</span></div>
                 {settings.canvasWheelCaptureMode === "key" && settings.canvasWheelOverride !== null && (
                   <div><kbd>{displayCanvasNavigationBinding(settings.canvasWheelOverride, window.canvasTTY.window.isMacOS)}</kbd>
                     <span>{t(settings.locale, "canvasWheelOverrideHint")}</span></div>
@@ -1687,7 +1687,7 @@ export function WorkspaceCanvas(props: WorkspaceCanvasProps): React.JSX.Element
                 <button className="shortcut-hints__more" type="button" data-interactive="true"
                   aria-haspopup="dialog" onClick={props.onOpenShortcutReference}>
                   {t(settings.locale, "keyboardShortcuts")}
-                  <UiIcon name="app-window" size={14} />
+                  <UiIcon name="app-window" size="1.27em" />
                 </button>
               </aside>
             )}
diff --git a/src/renderer/src/lib/i18n.ts b/src/renderer/src/lib/i18n.ts
index b1a05d81e..bf65bce59 100644
--- a/src/renderer/src/lib/i18n.ts
+++ b/src/renderer/src/lib/i18n.ts
@@ -35,7 +35,9 @@ const ru = {
   clock: "Часы",
   launcher: "Панель запуска",
   modelLimits: "Лимиты моделей",
-  limitUsed: "Использовано",
+  limitUsed: "использовано",
+  limitUsedPercent: "использовано {percent}",
+  limitLeftPercent: "осталось {percent}",
   limitStale: "Устарело",
   limitLoading: "Получаем лимит…",
   limitUnavailable: "Данные недоступны",
@@ -492,16 +494,11 @@ const ru = {
   zoomSensitivity: "Чувствительность зума",
   useScrollWheelToZoom: "Зум обычным скроллом",
   useScrollWheelToZoomDescription: "Если выключено, обычный скролл перемещает канвас, а Ctrl/Command + скролл и pinch масштабируют.",
-  canvasWheelCapture: "Перехват wheel/pinch над виджетами",
-  canvasWheelCaptureDescription: "Off оставляет обычный wheel/pinch focused input-виджету, включая live Browser page. On и активный бинд передают события канвасу; pinch и Ctrl/Command + скролл над Browser всегда масштабируют canvas. Клики и drag не перехватываются.",
-  canvasWheelOverride: "Комбинация wheel/pinch",
-  canvasWheelOverrideHint: "Wheel/pinch поверх виджетов",
+  canvasWheelOverrideHint: "Колесо и жест масштаба поверх виджетов",
   canvasOverrideShortcutInvalid: "Нужен модификатор, комбинация с ним или Mouse3/Mouse4/Mouse5.",
   canvasNavigationOverride: "Перехват навигации канвасом",
-  canvasNavigationOverrideDescription: "Пока комбинация или кнопка мыши удерживается, wheel, pinch и drag над любым виджетом принадлежат канвасу. Drag колёсиком работает всегда.",
   canvasNavigationCaptureHint: "Удержите модификатор, нажмите комбинацию или Mouse3/Mouse4/Mouse5. Escape отменяет запись.",
   canvasNavigationOverrideHint: "Навигация поверх виджетов",
-  canvasOverrideBindingsMatch: "Комбинации совпадают: полный перехват имеет приоритет и также забирает drag.",
   disabled: "Отключено",
   terminalWheelDirection: "Направление колеса терминала",
   copyOnSelect: "Копировать при выделении",
@@ -770,7 +767,8 @@ const ru = {
   terminalSearchClose: "Закрыть поиск",
   fitCanvas: "Показать всё",
   focusWindowHint: "Фокус на соседнее окно",
-  marqueeSelectionHint: "Shift + тянуть — выделить группу",
+  marqueeSelectionHint: "Выделить группу",
+  shortcutShiftDrag: "Shift + перетаскивание",
   remarkAnchorStep: "Замечание к шагу",
   remarkAnchorPage: "Замечание к странице",
   remarkAnchorTime: "Замечание ко времени",
@@ -1067,7 +1065,9 @@ const en: Record<keyof typeof ru, string> = {
   clock: "Clock",
   launcher: "Launcher",
   modelLimits: "Model limits",
-  limitUsed: "Used",
+  limitUsed: "used",
+  limitUsedPercent: "{percent} used",
+  limitLeftPercent: "{percent} left",
   limitStale: "Stale",
   limitLoading: "Checking limit…",
   limitUnavailable: "Data unavailable",
@@ -1524,16 +1524,11 @@ const en: Record<keyof typeof ru, string> = {
   zoomSensitivity: "Zoom sensitivity",
   useScrollWheelToZoom: "Use scroll wheel to zoom",
   useScrollWheelToZoomDescription: "When off, plain scrolling pans the canvas while Ctrl/Command + scroll and pinch zoom.",
-  canvasWheelCapture: "Wheel/pinch capture over widgets",
-  canvasWheelCaptureDescription: "Off leaves plain wheel/pinch with the focused input widget, including a live Browser page. On and an active binding route it to the canvas; Browser pinch and Ctrl/Command + scroll always zoom the canvas. Clicks and drag are not captured.",
-  canvasWheelOverride: "Wheel/pinch shortcut",
   canvasWheelOverrideHint: "Wheel/pinch over widgets",
   canvasOverrideShortcutInvalid: "Use a modifier, a modifier-based chord, or Mouse3/Mouse4/Mouse5.",
   canvasNavigationOverride: "Canvas navigation override",
-  canvasNavigationOverrideDescription: "While the shortcut or mouse button is held, wheel, pinch, and drag over every widget belong to the canvas. Middle-button drag always pans.",
   canvasNavigationCaptureHint: "Hold a modifier, press a chord, or press Mouse3/Mouse4/Mouse5. Escape cancels capture.",
   canvasNavigationOverrideHint: "Navigate over widgets",
-  canvasOverrideBindingsMatch: "The shortcuts match: full navigation takes priority and also captures drag.",
   disabled: "Disabled",
   terminalWheelDirection: "Terminal wheel direction",
   copyOnSelect: "Copy on selection",
@@ -1802,7 +1797,8 @@ const en: Record<keyof typeof ru, string> = {
   terminalSearchClose: "Close search",
   fitCanvas: "Fit to content",
   focusWindowHint: "Focus the neighbouring window",
-  marqueeSelectionHint: "Shift + drag selects a group",
+  marqueeSelectionHint: "Select a group",
+  shortcutShiftDrag: "Shift + drag",
   remarkAnchorStep: "Remark on step",
   remarkAnchorPage: "Remark on page",
   remarkAnchorTime: "Remark on the time",
diff --git a/src/renderer/src/lib/shortcuts.ts b/src/renderer/src/lib/shortcuts.ts
index 0b037c770..c1044a5ff 100644
--- a/src/renderer/src/lib/shortcuts.ts
+++ b/src/renderer/src/lib/shortcuts.ts
@@ -123,6 +123,15 @@ export function handleMacNativeSelectAll(event: KeyboardEvent, isMacOS: boolean)
   return true;
 }
 
+/**
+ * An action shortcut as the hints panel and the shortcut reference show it: arrows as glyphs, Meta as Command or
+ * Super. Key by key, so the Comma key is not looked for inside "Command".
+ */
+export function displayShortcutBinding(binding: string, isMacOS: boolean): string {
+  const names: Record<string, string> = { Meta: isMacOS ? "Command" : "Super", Comma: ",", ArrowUp: "↑", ArrowDown: "↓", ArrowLeft: "←", ArrowRight: "→" };
+  return binding.split("+").map((key) => names[key] ?? key).join("+");
+}
+
 export function displayCanvasNavigationBinding(binding: string, isMacOS: boolean): string {
   if (!isMacOS) return binding;
   return binding.split("+").map((part) => {
diff --git a/src/renderer/src/styles/app.css b/src/renderer/src/styles/app.css
index f4786c22b..3b72f650e 100644
--- a/src/renderer/src/styles/app.css
+++ b/src/renderer/src/styles/app.css
@@ -190,6 +190,9 @@ button { border: 0; }
 .limit-row__summary { min-width: 0; display: flex; align-items: baseline; justify-content: space-between; gap: 10px; }
 .limit-row__summary strong { flex: 0 0 auto; overflow: hidden; color: var(--text); font: 900 27px/1 var(--font-mono); letter-spacing: -.045em; text-overflow: ellipsis; white-space: nowrap; }
 .limit-row__summary > span { min-width: 0; overflow: hidden; color: var(--text-muted-on-dark); font: 800 14px/1 var(--font-mono); text-overflow: ellipsis; white-space: nowrap; }
+.limit-row__summary > .limit-row__share { flex: 1 1 auto; display: flex; align-items: baseline; gap: 8px; }
+.limit-row__share > span { min-width: 0; overflow: hidden; font-size: 12px; font-weight: 700; text-overflow: ellipsis; white-space: nowrap; }
+.limit-row__summary > .limit-row__countdown { flex: 0 0 auto; }
 .limit-row__track { position: relative; width: 100%; height: 6px; overflow: hidden; border-radius: 999px; background: rgba(255,255,255,.12); }
 .limit-row__fill { position: absolute; inset: 0 auto 0 0; max-width: 100%; border-radius: inherit; background: rgba(248,247,241,.78); }
 .limit-row__empty { overflow: hidden; color: var(--text-muted-on-dark); font-size: 13px; font-weight: 800; text-overflow: ellipsis; white-space: nowrap; }
@@ -203,6 +206,7 @@ button { border: 0; }
   .limit-row__summary { gap: 7px; }
   .limit-row__summary strong { font-size: 22px; }
   .limit-row__summary > span { font-size: 11px; }
+  .limit-row__share > span { font-size: 10px; }
   .limit-row__track { height: 4px; }
   .limit-row__empty { font-size: 11px; }
 }
@@ -324,11 +328,11 @@ button { border: 0; }
 .canvas-controls { height: 44px; display: flex; align-items: center; padding: 4px; border-radius: 12px; color: white; background: rgba(40,41,52,.94); box-shadow: var(--shadow-sm); pointer-events: auto; }
 .canvas-controls button { width: 36px; height: 36px; display: grid; place-items: center; border-radius: 8px; color: white; background: transparent; cursor: pointer; transition: background .15s ease; }
 .canvas-controls button:hover { background: rgba(255,255,255,.13); }
-.shortcut-hints { min-width: 224px; padding: 9px 11px; display: grid; gap: 7px; border-radius: 12px; color: white; background: rgba(40,41,52,.9); box-shadow: var(--shadow-sm); pointer-events: none; }
-.shortcut-hints > div { display: grid; grid-template-columns: minmax(52px, auto) 1fr; align-items: center; gap: 10px; }
-.shortcut-hints kbd { min-width: 52px; padding: 5px 7px; border: 1px solid rgba(255,255,255,.14); border-radius: 7px; color: white; background: rgba(255,255,255,.1); font: 800 11px var(--font-mono); text-align: center; }
-.shortcut-hints span { color: rgba(255,255,255,.66); font-size: 11px; font-weight: 750; }
-.shortcut-hints__more { display: flex; align-items: center; justify-content: space-between; gap: 12px; padding: 7px 0 2px; border-top: 1px solid rgba(255,255,255,.12); outline: 0; color: white; background: transparent; cursor: pointer; pointer-events: auto; font-size: 11px; font-weight: 750; }
+.shortcut-hints { min-width: 20.36em; padding: .82em 1em; display: grid; gap: .64em; border-radius: 1.09em; color: white; background: rgba(40,41,52,.9); box-shadow: var(--shadow-sm); pointer-events: none; font-size: calc(11px * var(--ui-scale, 1)); }
+.shortcut-hints > div { display: grid; grid-template-columns: minmax(4.73em, auto) 1fr; align-items: center; gap: .91em; }
+.shortcut-hints kbd { min-width: 4.73em; padding: .45em .64em; border: 1px solid rgba(255,255,255,.14); border-radius: .64em; color: white; background: rgba(255,255,255,.1); font: 800 1em var(--font-mono); text-align: center; }
+.shortcut-hints span { color: rgba(255,255,255,.66); font-size: 1em; font-weight: 750; }
+.shortcut-hints__more { display: flex; align-items: center; justify-content: space-between; gap: 1.09em; padding: .64em 0 .18em; border-top: 1px solid rgba(255,255,255,.12); outline: 0; color: white; background: transparent; cursor: pointer; pointer-events: auto; font-size: 1em; font-weight: 750; }
 .shortcut-hints__more:hover { color: var(--secondary); }
 .shortcut-hints__more:focus-visible { outline: 2px solid var(--secondary); outline-offset: 3px; border-radius: 3px; }
 .shortcut-reference-dialog { width: min(1200px, calc(100vw - 40px)); max-width: none; max-height: none; padding: 0; overflow: visible; border: 1px solid color-mix(in srgb, var(--text) 12%, transparent); border-radius: var(--radius-xl); color: var(--text); background: var(--surface); box-shadow: var(--shadow-lg); font-size: clamp(10px, calc(1.5vh * var(--ui-scale)), calc(13px * var(--ui-scale))); }
@@ -901,6 +905,8 @@ button { border: 0; }
 .segmented__button--active { color: var(--text-dark); background: var(--primary); }
 .segmented__button:focus-visible { box-shadow: inset 0 0 0 2px color-mix(in srgb, currentColor 42%, transparent); }
 .setting-group__control > .segmented + .shortcut-editor { margin-top: .6em; }
+/* The canvas navigation rows sit in the narrow control column: their (longer localized) labels wrap instead of being cut off. */
+.setting-group__control .shortcut-editor__row > span { overflow-wrap: break-word; text-overflow: clip; white-space: normal; line-height: 1.3; }
 .swatch-choices { display: grid; grid-template-columns: repeat(5, minmax(0, 1fr)); gap: .4em; }
 .swatch-choices--four { grid-template-columns: repeat(4, minmax(0, 1fr)); }
 .swatch-choice { min-width: 0; min-height: 4.6em; padding: .48em; display: grid; grid-template-rows: 2em auto; align-items: center; gap: .45em; border: .08em solid transparent; border-radius: .72em; outline: 0; color: var(--text-muted-on-dark); background: var(--surface-soft); cursor: pointer; font-size: .76em; font-weight: 700; transition: color .15s ease, background .15s ease, border-color .15s ease, box-shadow .15s ease; }
@@ -1061,34 +1067,34 @@ button { border: 0; }
 .home-grid-size strong { font: 850 12px var(--font-mono); }
 .plugin-settings { display: grid; grid-template-columns: minmax(0, 1fr); gap: 2px; }
 .plugin-install-row { display: grid; grid-template-columns: minmax(0, 1fr) auto; gap: 7px; }
-.plugin-install-row input { min-width: 0; height: 45px; padding: 0 13px; border: 1px solid rgba(45,45,56,.14); outline: 0; border-radius: 10px; color: var(--text-dark); background: rgba(255,255,255,.58); font: 700 12px var(--font-mono); user-select: text; }
+.plugin-install-row input { min-width: 0; height: 45px; padding: 0 13px; border: 1px solid rgba(45,45,56,.14); outline: 0; border-radius: 10px; color: var(--text-dark); background: rgba(255,255,255,.58); font: 700 calc(12px * var(--ui-scale, 1)) var(--font-mono); user-select: text; }
 .plugin-install-row input:focus { border-color: var(--surface); box-shadow: 0 0 0 3px color-mix(in srgb, var(--surface) 15%, transparent); }
-.plugin-install-row button, .plugin-preview__actions button, .installed-plugin__header button, .plugin-contribution button, .home-widget-list__row button { min-height: 36px; padding: 7px 11px; border-radius: 9px; color: var(--text-dark); background: rgba(255,255,255,.56); cursor: pointer; font-size: 11px; font-weight: 850; }
+.plugin-install-row button, .plugin-preview__actions button, .installed-plugin__header button, .plugin-contribution button, .home-widget-list__row button { min-height: 36px; padding: 7px 11px; border-radius: 9px; color: var(--text-dark); background: rgba(255,255,255,.56); cursor: pointer; font-size: calc(11px * var(--ui-scale, 1)); font-weight: 850; }
 .plugin-install-row button:disabled, .plugin-preview__actions button:disabled, .installed-plugin__header button:disabled, .plugin-contribution button:disabled, .home-widget-list__row button:disabled { cursor: not-allowed; opacity: .5; }
 .plugin-preview { margin-top: 9px; padding: 14px; border-radius: 13px; color: white; background: var(--surface); }
 .plugin-preview header { display: flex; align-items: center; justify-content: flex-start; gap: 12px; }
 .plugin-preview header > span { min-width: 0; display: flex; flex-direction: column; gap: 3px; }
-.plugin-preview header strong { font-size: 17px; }
-.plugin-preview header span { color: var(--text-muted-on-dark); font: 750 11px var(--font-mono); }
-.plugin-preview p, .installed-plugin > p { margin: 8px 0 11px; font-size: 12px; line-height: 1.45; opacity: .78; }
+.plugin-preview header strong { font-size: calc(17px * var(--ui-scale, 1)); }
+.plugin-preview header span { color: var(--text-muted-on-dark); font: 750 calc(11px * var(--ui-scale, 1)) var(--font-mono); }
+.plugin-preview p, .installed-plugin > p { margin: 8px 0 11px; font-size: calc(12px * var(--ui-scale, 1)); line-height: 1.45; opacity: .78; }
 .plugin-preview__actions { margin-top: 13px; display: flex; justify-content: flex-end; gap: 7px; }
 .plugin-preview__actions .plugin-primary-action { color: var(--text-dark); background: var(--primary); }
-.plugin-permissions { display: grid; gap: 6px; font-size: 11px; line-height: 1.35; }
+.plugin-permissions { display: grid; gap: 6px; font-size: calc(11px * var(--ui-scale, 1)); line-height: 1.35; }
 .plugin-permissions > strong { opacity: .72; }
 .plugin-permissions > span, .plugin-permissions li { opacity: .68; }
 .plugin-permissions ul { margin: 0; padding-left: 18px; display: grid; gap: 3px; }
-.plugin-hook-warning { margin-top: 9px; padding: 10px 11px; display: flex; flex-direction: column; gap: 4px; border: 1px solid rgba(239,208,132,.35); border-radius: 10px; color: #ffe7a9; background: rgba(96,62,34,.48); font-size: 10px; line-height: 1.4; }
-.plugin-hook-warning strong { font-size: 10px; }
+.plugin-hook-warning { margin-top: 9px; padding: 10px 11px; display: flex; flex-direction: column; gap: 4px; border: 1px solid rgba(239,208,132,.35); border-radius: 10px; color: #ffe7a9; background: rgba(96,62,34,.48); font-size: calc(10px * var(--ui-scale, 1)); line-height: 1.4; }
+.plugin-hook-warning strong { font-size: calc(10px * var(--ui-scale, 1)); }
 .plugin-hook-warning span { opacity: .82; }
 .plugin-hook-warning ul { margin: 4px 0 0; padding: 0; display: grid; gap: 4px; list-style: none; }
 .plugin-hook-warning li { min-width: 0; display: flex; align-items: baseline; justify-content: space-between; gap: 8px; }
-.plugin-hook-warning code { overflow-wrap: anywhere; font: 750 9px var(--font-mono); opacity: .82; text-align: right; }
+.plugin-hook-warning code { overflow-wrap: anywhere; font: 750 calc(9px * var(--ui-scale, 1)) var(--font-mono); opacity: .82; text-align: right; }
 .plugin-modules { margin: 10px 0 0; padding: 9px 10px; display: grid; gap: 7px; border: 1px solid rgba(255,255,255,.12); border-radius: 10px; }
-.plugin-modules legend { padding: 0 5px; font-size: 11px; font-weight: 850; opacity: .72; }
+.plugin-modules legend { padding: 0 5px; font-size: calc(11px * var(--ui-scale, 1)); font-weight: 850; opacity: .72; }
 .plugin-modules label { display: flex; align-items: flex-start; gap: 8px; cursor: pointer; }
 .plugin-modules label > span { min-width: 0; display: grid; flex: 1; gap: 2px; }
-.plugin-modules label strong { font-size: 12px; }
-.plugin-modules label small { font-size: 10px; line-height: 1.3; opacity: .68; }
+.plugin-modules label strong { font-size: calc(12px * var(--ui-scale, 1)); }
+.plugin-modules label small { font-size: calc(10px * var(--ui-scale, 1)); line-height: 1.3; opacity: .68; }
 .plugin-settings__error { margin: 9px 0 0; color: #8f3939; font-size: 12px; font-weight: 800; line-height: 1.4; }
 .plugin-settings__empty { margin: 0; padding: 17px; border-radius: 12px; background: rgba(255,255,255,.34); font-size: 12px; text-align: center; opacity: .65; }
 .plugin-icon { width: 34px; height: 34px; flex: 0 0 auto; border-radius: 8px; object-fit: cover; background: rgba(255,255,255,.12); }
@@ -1192,13 +1198,13 @@ button { border: 0; }
 .installed-plugin__details .plugin-modules { margin: 8px 0; }
 .installed-plugin__header > span.installed-plugin__title { min-width: 0; display: flex; flex-direction: row; align-items: center; gap: 10px; }
 .installed-plugin__header > span.installed-plugin__title > span { min-width: 0; display: flex; flex-direction: column; gap: 3px; }
-.installed-plugin__header > span strong { overflow: hidden; font-size: 16px; text-overflow: ellipsis; white-space: nowrap; }
-.installed-plugin__header > span small { color: var(--text-muted-on-dark); font: 700 10px var(--font-mono); display: flex; align-items: baseline; gap: 5px; }
-.installed-plugin__version-arrow { color: var(--primary); font-size: 11px; }
+.installed-plugin__header > span strong { overflow: hidden; font-size: calc(16px * var(--ui-scale, 1)); text-overflow: ellipsis; white-space: nowrap; }
+.installed-plugin__header > span small { color: var(--text-muted-on-dark); font: 700 calc(10px * var(--ui-scale, 1)) var(--font-mono); display: flex; align-items: baseline; gap: 5px; }
+.installed-plugin__version-arrow { color: var(--primary); font-size: calc(11px * var(--ui-scale, 1)); }
 .installed-plugin__version-new { color: var(--primary); }
-.installed-plugin__host { display: block; margin-top: 2px; font: 700 10px var(--font-mono); }
+.installed-plugin__host { display: block; margin-top: 2px; font: 700 calc(10px * var(--ui-scale, 1)) var(--font-mono); }
 .installed-plugin__author-row { display: flex; align-items: center; justify-content: space-between; gap: 10px; margin-bottom: 8px; }
-.installed-plugin__author { color: var(--text-muted-on-dark); font-size: 14px; font-weight: 750; opacity: .85; }
+.installed-plugin__author { color: var(--text-muted-on-dark); font-size: calc(14px * var(--ui-scale, 1)); font-weight: 750; opacity: .85; }
 .installed-plugin__repo-actions { display: inline-flex; gap: 6px; }
 .installed-plugin__header .installed-plugin__host, .plugin-showcase-tile__host { font-weight: 700; }
 .installed-plugin__header > div { display: flex; gap: 6px; }
@@ -1209,8 +1215,8 @@ button { border: 0; }
 .plugin-contribution-list { margin-top: 10px; display: grid; gap: 6px; }
 .plugin-contribution { min-height: 51px; padding: 7px 7px 7px 11px; display: grid; grid-template-columns: minmax(0, 1fr) auto; align-items: center; gap: 10px; border-radius: 10px; background: var(--surface-soft); }
 .plugin-contribution > span { min-width: 0; display: flex; flex-direction: column; gap: 3px; }
-.plugin-contribution > span strong { overflow: hidden; font-size: 12px; text-overflow: ellipsis; white-space: nowrap; }
-.plugin-contribution > span small { color: var(--text-muted-on-dark); font-size: 10px; }
+.plugin-contribution > span strong { overflow: hidden; font-size: calc(12px * var(--ui-scale, 1)); text-overflow: ellipsis; white-space: nowrap; }
+.plugin-contribution > span small { color: var(--text-muted-on-dark); font-size: calc(10px * var(--ui-scale, 1)); }
 
 /* Settings v3: iteration 7 two-pane structure with iteration 9 token roles. */
 .settings-panel .setting-group:not(.setting-group--field) { padding: 1.1em 0; border-bottom: .08em solid color-mix(in srgb, var(--text) 9%, transparent); }
@@ -1320,6 +1326,8 @@ button { border: 0; }
 .update-notice__dismiss:hover, .update-notice__dismiss:focus-visible { color: var(--text); background: var(--surface-soft); }
 .toast { position: fixed; z-index: 9999; left: 50%; bottom: 24px; max-width: min(540px, calc(100% - 40px)); padding: 12px 17px; border-radius: 11px; color: white; background: rgba(37,38,48,.94); box-shadow: var(--shadow-md); opacity: 0; transform: translate(-50%, 18px); pointer-events: none; transition: all .2s ease; }
 .toast--visible { opacity: 1; transform: translate(-50%, 0); }
+.toast--action { display: flex; align-items: center; gap: 12px; pointer-events: auto; }
+.toast__action { flex: 0 0 auto; padding: 6px 8px; border-radius: 6px; color: var(--text-dark); background: var(--primary); cursor: pointer; font-size: 11px; font-weight: 700; }
 .materials-load-error { position: fixed; z-index: 9998; left: 50%; bottom: 76px; width: max-content; max-width: min(540px, calc(100% - 40px)); padding: 12px 17px; border-radius: 11px; color: var(--text); background: var(--surface); box-shadow: inset 3px 0 0 var(--danger), var(--shadow-md); transform: translateX(-50%); }
 /* Shown until the critical snapshot is in, usually for a few frames only. It keeps the window's own native
    background (main/index.ts backgroundColor) so startup does not flash a dark panel between the empty window and
@@ -1368,9 +1376,9 @@ button { border: 0; }
 .terminal-card__search input:focus { border-color: var(--primary); }
 .terminal-card__search button { width: 26px; height: 26px; display: grid; place-items: center; border-radius: 6px; color: rgba(255,255,255,.78); background: rgba(255,255,255,.07); cursor: pointer; }
 .terminal-card__search button:hover { color: white; background: rgba(255,255,255,.16); }
-.terminal-card__menu { position: absolute; z-index: 18; top: calc(var(--card-header-height) + 6px); right: 10px; padding: 6px 8px; border: 1px solid rgba(255,255,255,.14); border-radius: 9px; color: white; background: rgba(28,30,40,.96); box-shadow: var(--shadow-md); font-size: 12px; font-weight: 700; }
-.terminal-card__menu label { display: flex; align-items: center; gap: 8px; padding: 4px 2px; cursor: pointer; }
-.terminal-card__menu-action { display: block; width: 100%; padding: 5px 4px; border-radius: 6px; color: inherit; background: transparent; font: inherit; text-align: left; cursor: pointer; }
+.terminal-card__menu { position: absolute; z-index: 18; top: calc(var(--card-header-height) + .5em); right: .84em; padding: .5em .67em; border: 1px solid rgba(255,255,255,.14); border-radius: .75em; color: white; background: rgba(28,30,40,.96); box-shadow: var(--shadow-md); font-size: calc(12px * var(--ui-scale, 1)); font-weight: 700; }
+.terminal-card__menu label { display: flex; align-items: center; gap: .67em; padding: .33em .17em; cursor: pointer; }
+.terminal-card__menu-action { display: block; width: 100%; padding: .42em .33em; border-radius: .5em; color: inherit; background: transparent; font: inherit; text-align: left; cursor: pointer; }
 .terminal-card__menu-action:hover, .terminal-card__menu-action:focus-visible { background: rgba(255,255,255,.1); }
 .terminal-card__plugin-badge { flex: 0 0 auto; min-width: 0; max-width: 14em; overflow: hidden; padding: .15em .55em; border-radius: .5em; background: rgba(255,255,255,.14); font: 700 .7em var(--font-mono); text-overflow: ellipsis; white-space: nowrap; }
 .terminal-card__plugin-badge--info { background: color-mix(in srgb, #5b9cf5 38%, transparent); }
diff --git a/src/shared/contracts.ts b/src/shared/contracts.ts
index e406aa1c2..6f5f67f02 100644
--- a/src/shared/contracts.ts
+++ b/src/shared/contracts.ts
@@ -1136,8 +1136,12 @@ export interface GithubAuthStatus {
   authorized: boolean;
   login: string | null;
   tokenExpiresAt: number | null;
+  /** Outcome of the most recent device sign-in; "idle" when none is pending or to report. */
+  deviceFlowState: GithubDeviceFlowState;
 }
 
+export type GithubDeviceFlowState = "idle" | "pending" | "denied" | "expired" | "failed" | "cancelled";
+
 export interface GithubDeviceFlowStart {
   userCode: string;
   verificationUri: string;
@@ -2049,6 +2053,7 @@ export interface CanvasTTYApi {
   githubAuth: {
     status(): Promise<GithubAuthStatus>;
     start(): Promise<GithubDeviceFlowStart>;
+    cancel(): Promise<void>;
     signOut(): Promise<void>;
     openUrl(url: string): Promise<void>;
   };
@@ -2245,6 +2250,7 @@ export const IPC = {
   githubAuthStatus: "github-auth:status",
   githubAuthStart: "github-auth:start",
   githubAuthSignOut: "github-auth:sign-out",
+  githubAuthCancel: "github-auth:cancel",
   githubAuthOpenUrl: "github-auth:open-url",
   terminalList: "terminal:list",
   terminalOpenFile: "terminal:open-file",
diff --git a/src/shared/pluginServiceNotice.ts b/src/shared/pluginServiceNotice.ts
new file mode 100644
index 000000000..4bbf9d570
--- /dev/null
+++ b/src/shared/pluginServiceNotice.ts
@@ -0,0 +1,48 @@
+import type { LocaleId } from "./contracts.ts";
+
+/**
+ * Why a plugin's own UI cannot reach its service although the service exists. The main process refuses with one of
+ * these fixed texts (plugins see them as the request's error); the renderer shows the localized text with a button
+ * that opens the settings section where the person can change it. A service that is trusted but not running keeps the
+ * supervisor's own error ("Plugin service is not running." with the last error, if any).
+ */
+export const PLUGIN_NATIVE_CODE_OFF_ERROR =
+  "This plugin's native code is turned off. Turn it on in Settings → Agents → Extension native code.";
+export const PLUGIN_SERVICE_MODULE_OFF_ERROR =
+  "The plugin module that runs this service is turned off. Turn it on in Settings → Plugins.";
+
+export type PluginServiceNoticeKind = "native-code-off" | "module-off";
+
+const NOTICE_TEXT: Record<LocaleId, Record<PluginServiceNoticeKind | "open", string>> = {
+  ru: {
+    "native-code-off": "Нативный код этого плагина выключен — включите его в Настройки → Агенты → Нативный код расширений.",
+    "module-off": "Модуль этого плагина с нативным кодом выключен — включите его в Настройки → Плагины.",
+    open: "Открыть настройки"
+  },
+  en: {
+    "native-code-off": "This plugin's native code is turned off — turn it on in Settings → Agents → Extension native code.",
+    "module-off": "The plugin module that runs this service is turned off — turn it on in Settings → Plugins.",
+    open: "Open settings"
+  }
+};
+
+/** The notice an error text stands for (the main process's fixed text or an already localized one), or null. */
+export function pluginServiceNoticeKind(message: string): PluginServiceNoticeKind | null {
+  const text = message.trim();
+  if (text === PLUGIN_NATIVE_CODE_OFF_ERROR || text === NOTICE_TEXT.ru["native-code-off"] || text === NOTICE_TEXT.en["native-code-off"]) {
+    return "native-code-off";
+  }
+  if (text === PLUGIN_SERVICE_MODULE_OFF_ERROR || text === NOTICE_TEXT.ru["module-off"] || text === NOTICE_TEXT.en["module-off"]) {
+    return "module-off";
+  }
+  return null;
+}
+
+export function pluginServiceNoticeText(kind: PluginServiceNoticeKind | "open", locale: LocaleId): string {
+  return (NOTICE_TEXT[locale] ?? NOTICE_TEXT.en)[kind];
+}
+
+/** The settings section that holds the switch for this notice. */
+export function pluginServiceNoticeSection(kind: PluginServiceNoticeKind): "agents" | "plugins" {
+  return kind === "module-off" ? "plugins" : "agents";
+}
diff --git a/tests/agent-ambient-keys.test.mjs b/tests/agent-ambient-keys.test.mjs
new file mode 100644
index 000000000..81ba49641
--- /dev/null
+++ b/tests/agent-ambient-keys.test.mjs
@@ -0,0 +1,280 @@
+import assert from "node:assert/strict";
+import { mkdtemp, realpath, rm } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import test from "node:test";
+import { AgentControlService } from "../src/main/services/AgentControlService.ts";
+import { AgentBrowserBridge } from "../src/main/services/agent-browser/AgentBrowserBridge.ts";
+import { AgentRuntimeBridge } from "../src/main/services/agent-runtime/AgentRuntimeBridge.ts";
+import { isControlProvider } from "../src/main/services/agent-control/controlCapabilities.ts";
+import { PluginSessions } from "../src/main/services/PluginSessions.ts";
+import { AgentIsolation } from "../src/main/services/isolation/AgentIsolation.ts";
+import { AGENT_PASS_ENV, agentEnvironment, shellEnvironment } from "../src/main/services/ambientProviderKeys.ts";
+import { TerminalManager } from "../src/main/services/TerminalManager.ts";
+import { availableRegistry, fakeSpawner } from "./helpers/terminal.mjs";
+
+const at = { x: 0, y: 0 };
+// Fake values only: the test sets them on its own process, never reads the person's environment.
+const FAKE = {
+  OPENAI_API_KEY: "fake-oa",
+  OPENAI_ADMIN_KEY: "fake-oa-adm",
+  ANTHROPIC_API_KEY: "fake-an",
+  ANTHROPIC_ADMIN_KEY: "fake-an-adm",
+  OPENROUTER_API_KEY: "fake-or",
+  GEMINI_API_KEY: "fake-gemini",
+  GITLAB_ADMIN_TOKEN: "fake-gl-adm",
+  ANTHROPIC_AUTH_TOKEN: "fake-at",
+  CLAUDE_CODE_OAUTH_TOKEN: "fake-oauth",
+  AWS_BEARER_TOKEN_BEDROCK: "fake-bedrock",
+  AWS_ACCESS_KEY_ID: "fake-akid",
+  AWS_SECRET_ACCESS_KEY: "fake-sak",
+  AWS_SESSION_TOKEN: "fake-st",
+  GITHUB_TOKEN: "fake-gh",
+  GH_TOKEN: "fake-gh2",
+  SSH_AUTH_SOCK: "/fake/agent",
+  XAI_API_KEY: "fake-xai",
+  HF_TOKEN: "fake-hf",
+  NPM_TOKEN: "fake-npm",
+  HTTPS_PROXY: "http://proxy:3128",
+  no_proxy: "localhost",
+  NODE_EXTRA_CA_CERTS: "/fake/ca.pem",
+  LC_ALL: "C.UTF-8",
+  EDITOR: "vi",
+  CANVASTTY_GITHUB_TOKEN: "fake-ctgh"
+};
+const ADMIN = ["OPENAI_ADMIN_KEY", "ANTHROPIC_ADMIN_KEY", "GITLAB_ADMIN_TOKEN"];
+const ORDINARY = ["OPENAI_API_KEY", "ANTHROPIC_API_KEY", "OPENROUTER_API_KEY", "GEMINI_API_KEY"];
+/** Credentials that are not provider API keys by name; a delegated card gets none of them (allowlist). */
+const OTHER_CREDENTIALS = ["ANTHROPIC_AUTH_TOKEN", "CLAUDE_CODE_OAUTH_TOKEN", "AWS_BEARER_TOKEN_BEDROCK", "AWS_ACCESS_KEY_ID",
+  "AWS_SECRET_ACCESS_KEY", "AWS_SESSION_TOKEN", "GITHUB_TOKEN", "GH_TOKEN", "SSH_AUTH_SOCK", "XAI_API_KEY", "HF_TOKEN", "NPM_TOKEN"];
+/** What a delegated card still needs to run and reach the network. */
+const KEPT = ["PATH", "HTTPS_PROXY", "no_proxy", "NODE_EXTRA_CA_CERTS", "LC_ALL", "EDITOR"];
+const SECRETS = [...ADMIN, ...ORDINARY, ...OTHER_CREDENTIALS];
+const SECRET_VALUES = [...SECRETS, "CANVASTTY_GITHUB_TOKEN"].map((name) => FAKE[name]);
+
+/** Sets fake keys (and optionally the opt-in) on this process for the test, restoring what was there. */
+function withFakeKeys(t, extra = {}) {
+  const names = [...Object.keys(FAKE), AGENT_PASS_ENV];
+  const saved = Object.fromEntries(names.map((name) => [name, process.env[name]]));
+  t.after(() => {
+    for (const [name, value] of Object.entries(saved)) {
+      if (value === undefined) delete process.env[name];
+      else process.env[name] = value;
+    }
+  });
+  delete process.env[AGENT_PASS_ENV];
+  Object.assign(process.env, FAKE, extra);
+}
+
+async function project(t) {
+  const base = await realpath(await mkdtemp(join(tmpdir(), "ctty-keys-")));
+  t.after(() => rm(base, { recursive: true, force: true }));
+  return base;
+}
+
+function present(env, names) {
+  return names.filter((name) => env[name] !== undefined);
+}
+
+function manager(t, calls, isolation) {
+  const terminals = new TerminalManager(() => undefined, availableRegistry(), undefined, undefined, true, fakeSpawner(calls));
+  t.after(() => terminals.shutdown());
+  if (isolation) terminals.configureIsolation(isolation);
+  return terminals;
+}
+
+/** A layer that always applies and records the environment it wraps. */
+function recordingLayer(userData, wrapped) {
+  return {
+    containment: () => true,
+    decide: (input) => new AgentIsolation({ userDataPath: userData, enabled: () => true, platform: "darwin", exists: () => true }).decide(input),
+    wrap: (launch) => {
+      wrapped.push(launch.env);
+      return { command: "/usr/bin/sandbox-exec", args: ["-f", "/p.sb", launch.command, ...launch.args], env: { ...launch.env, TMPDIR: "/s/" }, cleanup: () => undefined };
+    }
+  };
+}
+
+test("(a) a plain shell card the person starts keeps their environment, provider and admin keys included", async (t) => {
+  withFakeKeys(t);
+  const cwd = await project(t);
+  const calls = [];
+  manager(t, calls).create({ provider: "terminal", profile: "normal", cwd, position: at });
+  const env = calls.at(-1).options.env;
+  assert.deepEqual(present(env, SECRETS), SECRETS);
+  assert.equal(env.CANVASTTY_GITHUB_TOKEN, undefined, "CanvasTTY's own token is the app's, not the shell's");
+});
+
+test("a plain shell card a plugin starts is delegated: its owner drives it, so it gets only the allowlist", async (t) => {
+  withFakeKeys(t);
+  const cwd = await project(t);
+  const calls = [];
+  const terminals = manager(t, calls);
+  const sessions = new PluginSessions({ terminals, notify: () => true });
+  await sessions.handle("p.shell", "svc", "sessions.create", { provider: "terminal", cwd }, ["sessions:launch"]);
+  assert.equal(calls.length, 1);
+  const env = calls.at(-1).options.env;
+  assert.deepEqual(present(env, SECRETS), [], "no admin key, provider key or other credential");
+  assert.deepEqual(present(env, KEPT), KEPT, "what it needs to run and reach the network stays");
+  // Directly through the manager too (whatever the caller), keyed on the owning plugin.
+  terminals.create({ provider: "terminal", profile: "normal", cwd, position: at }, { origin: "plugin", ownerPluginId: "p.shell" });
+  assert.deepEqual(present(calls.at(-1).options.env, SECRETS), []);
+});
+
+test("an orchestrator cannot start a plain shell card, neither through spawn_agent nor its control connection", async (t) => {
+  const cwd = await project(t);
+  const calls = [];
+  const terminals = manager(t, calls);
+  const control = new AgentControlService(terminals, { containment: () => terminals.containment() });
+  const orchestrator = terminals.create({ provider: "codex", profile: "normal", cwd, position: at, role: "orchestrator" });
+  const before = calls.length;
+  assert.throws(() => control.spawn({ parentSessionId: orchestrator.id, provider: "terminal", cwd }), /Unknown agent provider/u);
+  assert.equal(calls.length, before);
+  assert.equal(isControlProvider("terminal"), false);
+});
+
+test("(b) an agent card the person starts keeps ordinary provider keys but never admin keys", async (t) => {
+  withFakeKeys(t);
+  const cwd = await project(t);
+  const calls = [];
+  const terminals = manager(t, calls);
+  for (const provider of ["claude", "codex", "opencode", "qwen"]) {
+    terminals.create({ provider, profile: "normal", cwd, position: at });
+    const env = calls.at(-1).options.env;
+    assert.deepEqual(present(env, ADMIN), [], `${provider}: no admin key`);
+    assert.equal(env.CANVASTTY_GITHUB_TOKEN, undefined, `${provider}: no CanvasTTY credential`);
+    assert.deepEqual(present(env, ORDINARY), ORDINARY, `${provider}: its sign-in by API key keeps working`);
+    assert.deepEqual(present(env, OTHER_CREDENTIALS), OTHER_CREDENTIALS, `${provider}: otherwise unchanged for the person's own card`);
+    assert.equal(env.OPENAI_API_KEY, FAKE.OPENAI_API_KEY);
+  }
+});
+
+test("(c) an agent card under isolation: no admin keys reach the layer or the process; the person's own card keeps ordinary keys", async (t) => {
+  withFakeKeys(t);
+  const cwd = await project(t);
+  const calls = [];
+  const wrapped = [];
+  const terminals = manager(t, calls, recordingLayer(join(cwd, "u"), wrapped));
+  const card = terminals.create({ provider: "claude", profile: "auto", cwd, position: at });
+  assert.deepEqual(card.isolation, { state: "on", layer: "seatbelt" });
+  assert.equal(calls.at(-1).command, "/usr/bin/sandbox-exec");
+  for (const env of [wrapped.at(-1), calls.at(-1).options.env]) {
+    assert.deepEqual(present(env, ADMIN), []);
+    assert.deepEqual(present(env, ORDINARY), ORDINARY);
+  }
+});
+
+test("(d) a subagent gets no ambient provider key, admin key or other credential, isolated or not", async (t) => {
+  withFakeKeys(t);
+  const cwd = await project(t);
+  for (const layer of ["isolated", "no layer"]) {
+    const calls = [];
+    const wrapped = [];
+    const isolation = layer === "isolated"
+      ? recordingLayer(join(cwd, "u"), wrapped)
+      : new AgentIsolation({ userDataPath: join(cwd, "u"), enabled: () => true, platform: "win32" });
+    const terminals = manager(t, calls, isolation);
+    const control = new AgentControlService(terminals, { containment: () => terminals.containment() });
+    const orchestrator = terminals.create({ provider: "codex", profile: "auto", cwd, position: at, role: "orchestrator" });
+    assert.deepEqual(present(calls.at(-1).options.env, ORDINARY), ORDINARY, `${layer}: the person's orchestrator keeps them`);
+    const child = await control.spawn({ parentSessionId: orchestrator.id, provider: "claude", cwd });
+    assert.equal(child.role, "subagent");
+    const env = calls.at(-1).options.env;
+    assert.deepEqual(present(env, SECRETS), [], `${layer}: the subagent`);
+    if (layer === "isolated") assert.deepEqual(present(wrapped.at(-1), SECRETS), []);
+    assert.deepEqual(present(env, KEPT), KEPT, "what it needs to run and reach the network stays");
+  }
+});
+
+test("the person's explicit opt-in passes named keys to every agent card, subagents included", async (t) => {
+  withFakeKeys(t, { [AGENT_PASS_ENV]: "OPENAI_ADMIN_KEY, ANTHROPIC_API_KEY" });
+  const cwd = await project(t);
+  const calls = [];
+  const terminals = manager(t, calls, new AgentIsolation({ userDataPath: join(cwd, "u"), enabled: () => true, platform: "win32" }));
+  const control = new AgentControlService(terminals, { containment: () => terminals.containment() });
+  const top = terminals.create({ provider: "codex", profile: "normal", cwd, position: at, role: "orchestrator" });
+  assert.deepEqual(present(calls.at(-1).options.env, ADMIN), ["OPENAI_ADMIN_KEY"]);
+  await control.spawn({ parentSessionId: top.id, provider: "claude", cwd });
+  assert.deepEqual(present(calls.at(-1).options.env, SECRETS), ["OPENAI_ADMIN_KEY", "ANTHROPIC_API_KEY"]);
+});
+
+test("agentEnvironment: names compare case-insensitively (Windows); admin credentials of any provider; delegated is an allowlist", () => {
+  const base = { Path: "C:\\bin", openai_api_key: "x", Openai_Admin_Key: "y", GITHUB_TOKEN: "g", AZURE_OPENAI_API_KEY: "z",
+    XAI_ADMIN_API_KEY: "w", ADMIN_TOKEN: "a", SOME_ADMIN_ACCESS_TOKEN: "b", HOME: "/h", SystemRoot: "C:\\Windows",
+    https_proxy: "http://p", Lc_Ctype: "UTF-8", XDG_CONFIG_HOME: "/h/.config", CODEX_HOME: "/h/.codex", MY_SERVICE_PASSWORD: "p" };
+  assert.deepEqual(agentEnvironment(base, { delegated: false }), { Path: "C:\\bin", openai_api_key: "x", GITHUB_TOKEN: "g",
+    AZURE_OPENAI_API_KEY: "z", HOME: "/h", SystemRoot: "C:\\Windows", https_proxy: "http://p", Lc_Ctype: "UTF-8",
+    XDG_CONFIG_HOME: "/h/.config", CODEX_HOME: "/h/.codex", MY_SERVICE_PASSWORD: "p" });
+  const delegated = { Path: "C:\\bin", HOME: "/h", SystemRoot: "C:\\Windows", https_proxy: "http://p", Lc_Ctype: "UTF-8",
+    XDG_CONFIG_HOME: "/h/.config", CODEX_HOME: "/h/.codex" };
+  assert.deepEqual(agentEnvironment(base, { delegated: true }), delegated);
+  assert.deepEqual(agentEnvironment({ ...base, [AGENT_PASS_ENV]: "openai_admin_key GITHUB_TOKEN" }, { delegated: true }),
+    { ...delegated, Openai_Admin_Key: "y", GITHUB_TOKEN: "g" });
+  // CanvasTTY's own credentials reach no agent card, the person's own included; its other variables are not inherited
+  // by a delegated card (the launch sets what a card needs).
+  const app = { CANVASTTY_GITHUB_TOKEN: "t", CANVASTTY_API_KEY: "k", CANVASTTY_CODEX_KEYBOARD: "{}", CANVASTTY_HELPERS: "node" };
+  assert.deepEqual(agentEnvironment(app, { delegated: false }), { CANVASTTY_CODEX_KEYBOARD: "{}", CANVASTTY_HELPERS: "node" });
+  assert.deepEqual(agentEnvironment(app, { delegated: true }), {});
+  assert.deepEqual(shellEnvironment({ ...app, GITHUB_TOKEN: "g" }), { CANVASTTY_CODEX_KEYBOARD: "{}", CANVASTTY_HELPERS: "node", GITHUB_TOKEN: "g" });
+});
+
+/** The real launch builders (agent runtime hooks, browser/orchestration MCP bridge), both built over process.env. */
+function builders(root) {
+  const helper = { command: "/app/CanvasTTY", args: ["/app/helper.mjs"], env: { ELECTRON_RUN_AS_NODE: "1" } };
+  const browserGateway = {
+    isEnabled: true,
+    setEnabled() {},
+    registerAgent: (input) => ({ agentId: "a", connectionId: "c", terminalSessionId: input.terminalSessionId, provider: input.provider,
+      capabilityToken: "b".repeat(40), address: "/tmp/b.sock" }),
+    revokeTerminalSession() {},
+    holdPendingForTerminal() {}
+  };
+  const runtimeGateway = {
+    registerSession: (id, provider) => ({ address: "/tmp/r.sock", terminalSessionId: id, provider, capabilityToken: "r".repeat(40) }),
+    revokeTerminalSession() {},
+    currentStatus: () => null
+  };
+  const browser = new AgentBrowserBridge(browserGateway, { helper, orchestrationHelper: helper, providerClis: availableRegistry(),
+    runtimeDirectory: join(root, "rt-b"), hermesHomeDirectory: join(root, "hermes"), kimiHomeDirectory: join(root, "kimi"),
+    probeKimiPerRunConfig: () => true });
+  const runtime = new AgentRuntimeBridge(runtimeGateway, { helper, runtimeDirectory: join(root, "rt-r"),
+    openCodePluginPath: "/app/opencode-plugin.mjs", kimiHomeDirectory: join(root, "kimi"), hermesHomeDirectory: join(root, "hermes"),
+    grokHomeDirectory: join(root, "grok"), coreHooksEnabled: true });
+  return { browser, runtime };
+}
+
+test("no launch builder re-injects what the filter removed: the env handed to the PTY of a delegated card holds no ambient credential", async (t) => {
+  // The person's OpenCode inline config carries a provider key; the builders read it from the app's environment.
+  withFakeKeys(t, { OPENCODE_CONFIG_CONTENT: JSON.stringify({ provider: { x: { options: { apiKey: "fake-oc" } } } }) });
+  t.after(() => { delete process.env.OPENCODE_CONFIG_CONTENT; });
+  const cwd = await project(t);
+  const calls = [];
+  const { browser, runtime } = builders(cwd);
+  const terminals = new TerminalManager(() => undefined, availableRegistry(), browser, runtime, true, fakeSpawner(calls));
+  t.after(() => terminals.shutdown());
+  const control = new AgentControlService(terminals, { containment: () => terminals.containment() });
+  const orchestrator = terminals.create({ provider: "opencode", profile: "normal", cwd, position: at, role: "orchestrator" });
+  const own = calls.at(-1).options.env;
+  assert.match(own.OPENCODE_CONFIG_CONTENT, /fake-oc/u, "the person's own card keeps their OpenCode config");
+  assert.match(own.OPENCODE_CONFIG_CONTENT, /opencode-plugin\.mjs/u);
+  assert.equal(own.CANVASTTY_GITHUB_TOKEN, undefined);
+  assert.deepEqual(present(own, ADMIN), []);
+  for (const provider of ["opencode", "claude", "codex", "qwen"]) {
+    await control.spawn({ parentSessionId: orchestrator.id, provider, cwd });
+    const env = calls.at(-1).options.env;
+    const text = JSON.stringify(env);
+    for (const value of [...SECRET_VALUES, "fake-oc"]) assert.ok(!text.includes(value), `${provider} subagent: ${value} is not in its env`);
+    assert.deepEqual(present(env, [...SECRETS, "CANVASTTY_GITHUB_TOKEN"]), [], provider);
+    if (provider === "opencode") {
+      const config = JSON.parse(env.OPENCODE_CONFIG_CONTENT);
+      assert.ok(config.plugin.some((entry) => entry.endsWith("/opencode-plugin.mjs")), "its lifecycle plugin is still there");
+      assert.ok(config.mcp.canvastty_browser, "its browser server is still there");
+      assert.equal(config.provider, undefined, "the person's inline config is not handed to a delegated agent");
+    }
+  }
+  // A plugin-started plain shell, through the same manager.
+  const sessions = new PluginSessions({ terminals, notify: () => true });
+  await sessions.handle("p.shell", "svc", "sessions.create", { provider: "terminal", cwd }, ["sessions:launch"]);
+  const shell = JSON.stringify(calls.at(-1).options.env);
+  for (const value of [...SECRET_VALUES, "fake-oc"]) assert.ok(!shell.includes(value), `plugin shell: ${value}`);
+});
diff --git a/tests/agent-control.test.mjs b/tests/agent-control.test.mjs
index 37bd58742..c75afbf89 100644
--- a/tests/agent-control.test.mjs
+++ b/tests/agent-control.test.mjs
@@ -709,7 +709,7 @@ test("send appends submit unless told otherwise and rejects exited sessions", as
   terminals.disposeAll();
 });
 
-test("observe returns a capped terminal tail and result reflects exit state", () => {
+test("observe returns a capped terminal tail and result reflects exit state", async () => {
   const { terminals, control } = serviceFixture();
   const parent = terminals.create({
     provider: "codex",
@@ -717,11 +717,11 @@ test("observe returns a capped terminal tail and result reflects exit state", ()
     profile: "normal",
     position: { x: 0, y: 0 }
   });
-  const observation = control.observe(parent.id);
+  const observation = await control.observe(parent.id);
   assert.equal(observation.sessionId, parent.id);
   assert.equal(observation.output.length <= 8_192, true);
 
-  const running = control.result(parent.id);
+  const running = await control.result(parent.id);
   assert.equal(running.state, "running");
   assert.equal(running.exitCode, null);
   terminals.disposeAll();
@@ -746,8 +746,8 @@ test("cancel disposes the subagent and plain terminals are not agents", async ()
   assert.equal(terminals.list().some((session) => session.id === child.id), false);
 
   assert.throws(() => control.send(terminal.id, "text"), /not agents/u);
-  assert.throws(() => control.observe(terminal.id), /not agents/u);
-  assert.throws(() => control.result(terminal.id), /not agents/u);
+  await assert.rejects(control.observe(terminal.id), /not agents/u);
+  await assert.rejects(control.result(terminal.id), /not agents/u);
   terminals.disposeAll();
 });
 
@@ -805,8 +805,8 @@ test("an orchestrator tool call looks its sessions up by id: no other card's scr
   assert.equal(control.status(child.id).id, child.id);
   assert.equal(control.isInSubtree(parent.id, child.id), true);
   assert.deepEqual(control.children(parent.id).map((session) => session.id), [child.id]);
-  control.observe(child.id);
-  control.result(child.id);
+  await control.observe(child.id);
+  await control.result(child.id);
   assert.throws(() => control.status("missing"), /does not exist/u);
 
   assert.deepEqual(copied, [child.id, child.id], "only the observed card's own scrollback is read, and no list() snapshot of every card");
diff --git a/tests/agent-isolation.test.mjs b/tests/agent-isolation.test.mjs
index 89b2266aa..c1497ef2b 100644
--- a/tests/agent-isolation.test.mjs
+++ b/tests/agent-isolation.test.mjs
@@ -180,6 +180,27 @@ test("bubblewrap: read-only root, writable project, tmpfs over what may not be r
   assert.deepEqual(args.slice(-5), ["--chdir", w.project, "--", "/usr/bin/claude", "--x"]);
 });
 
+test("the paths: Kimi Code's own home (KIMI_CODE_HOME, default ~/.kimi-code) is Kimi's, and no other CLI's", async (t) => {
+  const w = await world(t);
+  const pool = join(w.userData, "account-homes", "kimi-a");
+  const env = { ...w.env, KIMI_CODE_HOME: pool };
+  const kimi = isolationPaths({ provider: "kimi", cwd: w.project, sessionTemp: join(w.temp, "s"), env, userDataPath: w.userData, sessionId: "s1" });
+  assert.ok(kimi.writable.includes(pool) && !kimi.unreadable.includes(pool), "a moved Kimi Code home stays Kimi's own");
+  const codex = isolationPaths({ provider: "codex", cwd: w.project, sessionTemp: join(w.temp, "s"), env, userDataPath: w.userData, sessionId: "s1" });
+  assert.ok(codex.unreadable.includes(pool), "another CLI cannot read a moved Kimi Code home");
+  const plain = isolationPaths({ provider: "codex", cwd: w.project, sessionTemp: join(w.temp, "s"), env: w.env, userDataPath: w.userData, sessionId: "s1" });
+  assert.ok(plain.unreadable.includes(join(w.home, ".kimi-code")), "Kimi Code's default sign-in folder is hidden from other CLIs");
+  assert.ok(plain.unreadable.includes(join(w.home, ".kimi")), "and the older kimi-cli's");
+  const share = isolationPaths({ provider: "kimi", cwd: w.project, sessionTemp: join(w.temp, "s"), env: { ...w.env, KIMI_SHARE_DIR: join(w.base, "kimi-share") }, userDataPath: w.userData, sessionId: "s1" });
+  assert.ok(share.writable.includes(join(w.base, "kimi-share")), "KIMI_SHARE_DIR moves the older kimi-cli's home");
+  // Kimi's config.toml holds its permission rules and hooks: its folders are writable, but that file stays the person's.
+  const own = isolationPaths({ provider: "kimi", cwd: w.project, sessionTemp: join(w.temp, "s"), env: w.env, userDataPath: w.userData, sessionId: "s1" });
+  assert.ok(own.writable.includes(join(w.home, ".kimi-code")));
+  assert.ok(own.protectedWrites.includes(join(w.home, ".kimi-code", "config.toml")), "Kimi Code's permission settings stay the person's");
+  assert.ok(own.protectedWrites.includes(join(w.home, ".kimi", "config.toml")), "and kimi-cli's");
+  assert.ok(kimi.protectedWrites.includes(join(pool, "config.toml")), "and a moved home's");
+});
+
 test("the paths: another CLI's home moved by its own variable is unreadable, never this CLI's own", async (t) => {
   const w = await world(t);
   const custom = (name) => join(w.base, "creds", name);
diff --git a/tests/agent-runtime-account-homes.test.mjs b/tests/agent-runtime-account-homes.test.mjs
new file mode 100644
index 000000000..62d2bf6d8
--- /dev/null
+++ b/tests/agent-runtime-account-homes.test.mjs
@@ -0,0 +1,453 @@
+import assert from "node:assert/strict";
+import { lstat, mkdir, mkdtemp, readFile, readdir, rm, stat, symlink, writeFile, chmod } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import test from "node:test";
+
+import { AgentRuntimeBridge } from "../src/main/services/agent-runtime/AgentRuntimeBridge.ts";
+import { LaunchPipeline } from "../src/main/services/LaunchPipeline.ts";
+import { ProviderRuntimeLaunchAdapters } from "../src/main/services/agent-runtime/ProviderRuntimeLaunch.ts";
+import { ProviderLaunchAdapters } from "../src/main/services/agent-browser/ProviderLaunch.ts";
+import { AgentBrowserBridge } from "../src/main/services/agent-browser/AgentBrowserBridge.ts";
+import { accountHookHome, contributedHookHome } from "../src/main/services/accountHomes.ts";
+import { TerminalManager } from "../src/main/services/TerminalManager.ts";
+
+const helper = Object.freeze({
+  command: "/opt/CanvasTTY Agent/electron",
+  args: ["/opt/CanvasTTY Agent/agent-runtime/hook-helper.mjs"],
+  env: { ELECTRON_RUN_AS_NODE: "1" }
+});
+
+const GROK_FILE = ["hooks", "canvastty-runtime-hooks.json"];
+const KIMI_MARKER = "# >>> CanvasTTY lifecycle hooks >>>";
+const PERSON_KIMI = "# the person's own settings\ndefault_model = \"kimi-k2\"\n\n[[hooks]]\nevent = \"Stop\"\ncommand = \"notify-send done\"\n";
+
+/** userData with plugin-data/<plugin>/homes/<account>, the default CLI homes, and adapters bound to them. */
+async function fixture(t, { root: withRoot = true } = {}) {
+  const base = await mkdtemp(join(tmpdir(), "canvastty-account-homes-"));
+  t.after(() => rm(base, { recursive: true, force: true }));
+  const pluginData = join(base, "userData", "plugin-data");
+  const accountHome = (plugin, account) => join(pluginData, plugin, "homes", account);
+  const defaultGrok = join(base, "home", ".grok");
+  const defaultKimi = join(base, "home", ".kimi-code");
+  await Promise.all([
+    mkdir(accountHome("accounts", "a"), { recursive: true }),
+    mkdir(accountHome("accounts", "b"), { recursive: true }),
+    mkdir(defaultGrok, { recursive: true }),
+    mkdir(defaultKimi, { recursive: true })
+  ]);
+  const adapters = () => new ProviderRuntimeLaunchAdapters({
+    helper,
+    runtimeDirectory: join(base, "runtime"),
+    openCodePluginPath: join(base, "opencode-plugin.mjs"),
+    grokHomeDirectory: defaultGrok,
+    kimiHomeDirectory: defaultKimi,
+    hermesHomeDirectory: join(base, "home", ".hermes"),
+    ...(withRoot ? { accountHomesRoot: pluginData } : {}),
+    platform: "linux"
+  });
+  return { base, pluginData, accountHome, defaultGrok, defaultKimi, adapters };
+}
+
+const exists = (path) => lstat(path).then(() => true, () => false);
+const markers = (text) => text.split(KIMI_MARKER).length - 1;
+
+test("accountHookHome: GROK_HOME for Grok, KIMI_CODE_HOME for Kimi, nothing for the other CLIs", () => {
+  assert.equal(accountHookHome("grok", { GROK_HOME: "/data/plugin-data/a/homes/x/" }), "/data/plugin-data/a/homes/x");
+  assert.equal(accountHookHome("kimi", { KIMI_CODE_HOME: "/data/plugin-data/a/homes/k" }), "/data/plugin-data/a/homes/k");
+  // The older kimi-cli's home holds no Kimi Code hooks.
+  assert.equal(accountHookHome("kimi", { KIMI_SHARE_DIR: "/data/plugin-data/a/homes/k" }), undefined);
+  assert.equal(accountHookHome("grok", { KIMI_CODE_HOME: "/x" }), undefined);
+  assert.equal(accountHookHome("claude", { CLAUDE_CONFIG_DIR: "/x", GROK_HOME: "/y" }), undefined);
+  assert.equal(accountHookHome("codex", { CODEX_HOME: "/x" }), undefined);
+  assert.equal(accountHookHome("grok", undefined), undefined);
+  assert.equal(accountHookHome("grok", { GROK_HOME: "" }), undefined);
+});
+
+test("contributedHookHome: only the contributor's own data folder, only the hook home variable", () => {
+  const dataDir = "/data/plugin-data/accounts";
+  assert.equal(contributedHookHome("grok", dataDir, "GROK_HOME", `${dataDir}/homes/a/`), `${dataDir}/homes/a`);
+  assert.equal(contributedHookHome("kimi", dataDir, "KIMI_CODE_HOME", `${dataDir}/homes/k`), `${dataDir}/homes/k`);
+  assert.equal(contributedHookHome("grok", dataDir, "GROK_HOME", "/data/plugin-data/other/homes/a"), undefined);
+  assert.equal(contributedHookHome("grok", dataDir, "GROK_HOME", dataDir), undefined);
+  assert.equal(contributedHookHome("grok", dataDir, "GROK_HOME", `${dataDir}/../other/homes/a`), undefined);
+  assert.equal(contributedHookHome("grok", dataDir, "GROK_HOME", "homes/a"), undefined);
+  assert.equal(contributedHookHome("grok", dataDir, "KIMI_CODE_HOME", `${dataDir}/homes/a`), undefined);
+  assert.equal(contributedHookHome("kimi", dataDir, "KIMI_SHARE_DIR", `${dataDir}/homes/a`), undefined);
+  assert.equal(contributedHookHome("claude", dataDir, "CLAUDE_CONFIG_DIR", `${dataDir}/homes/a`), undefined);
+  assert.equal(contributedHookHome("grok", undefined, "GROK_HOME", `${dataDir}/homes/a`), undefined);
+});
+
+test("Grok: an account home gets the status hooks, the default home is left alone, and the release removes them", async (t) => {
+  const { accountHome, defaultGrok, adapters } = await fixture(t);
+  const home = accountHome("accounts", "a");
+  await mkdir(join(home, "hooks"), { recursive: true });
+  await writeFile(join(home, "hooks", "person.json"), "{\"hooks\":{}}\n");
+  await writeFile(join(home, "config.toml"), "model = \"grok-4\"\n");
+  const launch = adapters().prepare("grok", "card-grok", true, false, undefined, undefined, home);
+  const written = JSON.parse(await readFile(join(home, ...GROK_FILE), "utf8"));
+  assert.match(written.hooks.UserPromptSubmit[0].hooks[0].command, /working.*UserPromptSubmit/u);
+  assert.match(written.hooks.Notification[0].hooks[0].command, /needs_approval.*Notification/u);
+  assert.equal(written.hooks.Notification[0].matcher, "permission_prompt");
+  assert.equal((await stat(join(home, ...GROK_FILE))).mode & 0o777, 0o600);
+  assert.equal(await exists(join(defaultGrok, ...GROK_FILE)), false, "the default ~/.grok is not touched");
+  launch.releaseConfiguration();
+  assert.equal(await exists(join(home, ...GROK_FILE)), false);
+  assert.equal(await readFile(join(home, "hooks", "person.json"), "utf8"), "{\"hooks\":{}}\n");
+  assert.equal(await readFile(join(home, "config.toml"), "utf8"), "model = \"grok-4\"\n");
+});
+
+test("Kimi: an account home's config.toml gets one owned block; the person's settings and file mode are kept and restored", async (t) => {
+  const { accountHome, defaultKimi, adapters } = await fixture(t);
+  const home = accountHome("accounts", "a");
+  const config = join(home, "config.toml");
+  await writeFile(config, PERSON_KIMI);
+  await chmod(config, 0o640);
+  await writeFile(join(defaultKimi, "config.toml"), "default_model = \"default\"\n");
+  const launch = adapters().prepare("kimi", "card-kimi", true, false, undefined, undefined, home);
+  const during = await readFile(config, "utf8");
+  assert.ok(during.startsWith(PERSON_KIMI), "the person's settings stay first and unchanged");
+  assert.equal(markers(during), 1);
+  assert.match(during, /event = "TurnStarted"\ncommand = ".*working.*TurnStarted/u);
+  assert.equal((await stat(config)).mode & 0o777, 0o640);
+  assert.equal(await readFile(join(defaultKimi, "config.toml"), "utf8"), "default_model = \"default\"\n");
+  assert.equal(await exists(join(defaultKimi, ".canvastty-runtime-kimi.json")), false);
+  launch.releaseConfiguration();
+  assert.equal(await readFile(config, "utf8"), PERSON_KIMI);
+  assert.equal(await exists(join(home, ".canvastty-runtime-kimi.json")), false);
+  assert.deepEqual(await readdir(join(home, ".canvastty-runtime-backups")).catch(() => []), [], "no backup is left");
+});
+
+test("idempotent: two cards in one account home share one overlay; another home gets its own; a crashed overlay is refreshed", async (t) => {
+  const { accountHome, adapters } = await fixture(t);
+  const home = accountHome("accounts", "a");
+  const other = accountHome("accounts", "b");
+  const config = join(home, "config.toml");
+  await writeFile(config, PERSON_KIMI);
+  const runtime = adapters();
+  const first = runtime.prepare("kimi", "card-1", true, false, undefined, undefined, home);
+  const second = runtime.prepare("kimi", "card-2", true, false, undefined, undefined, `${home}/`);
+  const third = runtime.prepare("kimi", "card-3", true, false, undefined, undefined, other);
+  assert.equal(markers(await readFile(config, "utf8")), 1);
+  assert.equal(markers(await readFile(join(other, "config.toml"), "utf8")), 1);
+  first.releaseConfiguration();
+  first.releaseConfiguration();
+  assert.equal(markers(await readFile(config, "utf8")), 1, "still in use by card-2");
+  second.releaseConfiguration();
+  assert.equal(await readFile(config, "utf8"), PERSON_KIMI);
+  third.releaseConfiguration();
+  assert.equal(await exists(join(other, "config.toml")), false, "a file CanvasTTY created is removed again");
+
+  // An app that ended without releasing (a crash) leaves its overlay; the next launch there replaces it, not adds one.
+  runtime.prepare("kimi", "card-crashed", true, false, undefined, undefined, home);
+  runtime.prepare("grok", "card-crashed-grok", true, false, undefined, undefined, home);
+  const restarted = adapters();
+  const again = restarted.prepare("kimi", "card-again", true, false, undefined, undefined, home);
+  const grokAgain = restarted.prepare("grok", "card-again-grok", true, false, undefined, undefined, home);
+  assert.equal(markers(await readFile(config, "utf8")), 1);
+  assert.ok((await readFile(config, "utf8")).startsWith(PERSON_KIMI));
+  again.releaseConfiguration();
+  grokAgain.releaseConfiguration();
+  assert.equal(await readFile(config, "utf8"), PERSON_KIMI);
+  assert.equal(await exists(join(home, ...GROK_FILE)), false);
+});
+
+test("symbolic links: a linked account home, a linked folder above it, or a linked config file is refused and nothing is written", async (t) => {
+  const { base, pluginData, accountHome, defaultGrok, defaultKimi, adapters } = await fixture(t);
+  const outside = join(base, "outside");
+  await mkdir(join(outside, "hooks"), { recursive: true });
+  await writeFile(join(outside, "config.toml"), "outside = true\n");
+  const runtime = adapters();
+
+  const linkedHome = accountHome("accounts", "linked");
+  await symlink(outside, linkedHome);
+  for (const provider of ["grok", "kimi"]) {
+    assert.throws(() => runtime.prepare(provider, `card-${provider}`, true, false, undefined, undefined, linkedHome),
+      /missing or crosses a symbolic link/u);
+  }
+
+  await symlink(join(pluginData, "accounts"), join(pluginData, "alias"));
+  assert.throws(() => runtime.prepare("grok", "card-alias", true, false, undefined, undefined, join(pluginData, "alias", "homes", "a")),
+    /missing or crosses a symbolic link/u);
+
+  const kimiHome = accountHome("accounts", "a");
+  await symlink(join(outside, "config.toml"), join(kimiHome, "config.toml"));
+  assert.throws(() => runtime.prepare("kimi", "card-config", true, false, undefined, undefined, kimiHome),
+    /symbolic link where CanvasTTY writes/u);
+
+  const grokHome = accountHome("accounts", "b");
+  await symlink(join(outside, "hooks"), join(grokHome, "hooks"));
+  assert.throws(() => runtime.prepare("grok", "card-hooks", true, false, undefined, undefined, grokHome),
+    /symbolic link where CanvasTTY writes/u);
+
+  assert.deepEqual((await readdir(outside)).sort(), ["config.toml", "hooks"]);
+  assert.equal(await readFile(join(outside, "config.toml"), "utf8"), "outside = true\n");
+  assert.deepEqual(await readdir(join(outside, "hooks")), []);
+  assert.deepEqual(await readdir(defaultGrok), []);
+  assert.deepEqual(await readdir(defaultKimi), []);
+  assert.equal(await exists(join(kimiHome, ".canvastty-runtime-kimi.json")), false);
+});
+
+test("a new account's home that does not exist yet is created (private, no link followed) and gets the overlays", async (t) => {
+  const { accountHome, defaultGrok, defaultKimi, adapters } = await fixture(t);
+  const runtime = adapters();
+  for (const provider of ["grok", "kimi"]) {
+    // A first sign-in: the contributor names the account's home before the CLI ever created it.
+    const home = accountHome("accounts", `new-${provider}`);
+    assert.equal(await exists(home), false);
+    const launch = runtime.prepare(provider, `card-${provider}`, true, false, undefined, undefined, home);
+    assert.equal((await stat(home)).mode & 0o777, 0o700, `${provider}: created private`);
+    if (provider === "grok") assert.equal(await exists(join(home, ...GROK_FILE)), true);
+    else assert.match(await readFile(join(home, "config.toml"), "utf8"), /CanvasTTY lifecycle hooks/u);
+    launch.releaseConfiguration();
+  }
+  assert.deepEqual(await readdir(defaultGrok), []);
+  assert.deepEqual(await readdir(defaultKimi), []);
+});
+
+test("homes the launch does not own are never written: outside plugin data, a whole plugin's folder, relative, or no root", async (t) => {
+  const { base, pluginData, defaultGrok, defaultKimi, adapters } = await fixture(t);
+  const elsewhere = join(base, "elsewhere");
+  await mkdir(elsewhere, { recursive: true });
+  const runtime = adapters();
+  for (const candidate of [elsewhere, join(pluginData, "accounts"), pluginData, "relative/home", `${pluginData}/accounts/../accounts/homes/a`]) {
+    const launch = runtime.prepare("grok", "card", true, false, undefined, undefined, candidate);
+    const kimi = runtime.prepare("kimi", "card-k", true, false, undefined, undefined, candidate);
+    assert.equal(await exists(join(defaultGrok, ...GROK_FILE)), true, `${candidate}: the default home, as before`);
+    assert.match(await readFile(join(defaultKimi, "config.toml"), "utf8"), /CanvasTTY lifecycle hooks/u);
+    launch.releaseConfiguration();
+    kimi.releaseConfiguration();
+  }
+  assert.deepEqual(await readdir(elsewhere), []);
+  assert.deepEqual(await readdir(join(pluginData, "accounts", "homes", "a")), []);
+
+  const noRoot = await fixture(t, { root: false });
+  const home = noRoot.accountHome("accounts", "a");
+  const launch = noRoot.adapters().prepare("grok", "card", true, false, undefined, undefined, home);
+  assert.deepEqual(await readdir(home), []);
+  assert.equal(await exists(join(noRoot.defaultGrok, ...GROK_FILE)), true);
+  launch.releaseConfiguration();
+});
+
+test("default homes are unchanged without an account home, and hooks off writes nowhere", async (t) => {
+  const { accountHome, defaultGrok, defaultKimi, adapters } = await fixture(t);
+  const runtime = adapters();
+  const grok = runtime.prepare("grok", "card-grok");
+  const kimi = runtime.prepare("kimi", "card-kimi");
+  assert.equal(await exists(join(defaultGrok, ...GROK_FILE)), true);
+  assert.equal(markers(await readFile(join(defaultKimi, "config.toml"), "utf8")), 1);
+  grok.releaseConfiguration();
+  kimi.releaseConfiguration();
+  assert.equal(await exists(join(defaultGrok, ...GROK_FILE)), false);
+  assert.equal(await exists(join(defaultKimi, "config.toml")), false);
+
+  const home = accountHome("accounts", "a");
+  const off = runtime.prepare("grok", "card-off", false, false, undefined, undefined, home);
+  assert.deepEqual(await readdir(home), []);
+  off.releaseConfiguration();
+  // Claude Code and Codex take their hooks as arguments; their homes are never written either way.
+  for (const provider of ["claude", "codex"]) {
+    const launch = runtime.prepare(provider, `card-${provider}`, true, false, undefined, undefined, home);
+    assert.ok(launch.args.length > 0);
+    launch.releaseConfiguration();
+  }
+  assert.deepEqual(await readdir(home), []);
+});
+
+const PERSON_MCP = `${JSON.stringify({ mcpServers: { person: { transport: "stdio", command: "/usr/bin/true" } } }, null, 2)}\n`;
+const BROWSER_RULE = "# CanvasTTY temporary browser permission begin";
+
+const browserHelper = Object.freeze({ command: "/opt/CanvasTTY Agent/electron", args: ["/opt/CanvasTTY Agent/agent-browser/helper.mjs"], env: { ELECTRON_RUN_AS_NODE: "1" } });
+const orchestrationHelper = Object.freeze({ command: "/opt/CanvasTTY Agent/electron", args: ["/opt/CanvasTTY Agent/agent-browser/orchestration.mjs"], env: { ELECTRON_RUN_AS_NODE: "1" } });
+
+const registry = Object.freeze({
+  get: (provider) => ({ state: "available", provider, executable: `/resolved/${provider}`, launcher: "native", environment: { PATH: "/usr/bin" }, checked: [] }),
+  snapshot: () => ({})
+});
+
+function browserAdapters(fix, perRun, withRoot = true) {
+  return new ProviderLaunchAdapters({
+    helper: browserHelper,
+    orchestrationHelper,
+    providerClis: registry,
+    runtimeDirectory: join(fix.base, "browser-runtime"),
+    hermesHomeDirectory: join(fix.base, "home", ".hermes"),
+    kimiHomeDirectory: fix.defaultKimi,
+    probeKimiPerRunConfig: () => perRun,
+    ...(withRoot ? { accountHomesRoot: fix.pluginData } : {})
+  });
+}
+
+test("Kimi tools (per-run MCP file): the account home gets the permission rule, the default home nothing; restored after", async (t) => {
+  const fix = await fixture(t);
+  const home = fix.accountHome("accounts", "a");
+  await writeFile(join(home, "config.toml"), PERSON_KIMI);
+  await writeFile(join(fix.defaultKimi, "config.toml"), "default_model = \"default\"\n");
+  const adapters = browserAdapters(fix, true);
+  const launch = adapters.prepare("kimi", "connection-1", { orchestration: true, accountHome: home });
+  const mcpFile = launch.args[launch.args.indexOf("--mcp-config-file") + 1];
+  const servers = JSON.parse(await readFile(mcpFile, "utf8")).mcpServers;
+  assert.deepEqual(Object.keys(servers).sort(), ["canvastty_agents", "canvastty_browser"]);
+  const during = await readFile(join(home, "config.toml"), "utf8");
+  assert.ok(during.startsWith(PERSON_KIMI));
+  assert.equal(during.split(BROWSER_RULE).length - 1, 1);
+  assert.equal(await readFile(join(fix.defaultKimi, "config.toml"), "utf8"), "default_model = \"default\"\n");
+  // A second card in the same home shares the rule; another home gets its own.
+  const second = adapters.prepare("kimi", "connection-2", { orchestration: true, accountHome: `${home}/` });
+  assert.equal((await readFile(join(home, "config.toml"), "utf8")).split(BROWSER_RULE).length - 1, 1);
+  launch.releaseConfiguration();
+  assert.equal((await readFile(join(home, "config.toml"), "utf8")).split(BROWSER_RULE).length - 1, 1);
+  second.releaseConfiguration();
+  assert.equal(await readFile(join(home, "config.toml"), "utf8"), PERSON_KIMI);
+  assert.deepEqual((await readdir(home)).sort(), ["config.toml"]);
+});
+
+test("Kimi tools (shared mcp.json): the account home's mcp.json gets both servers next to the person's; restored after", async (t) => {
+  const fix = await fixture(t);
+  const home = fix.accountHome("accounts", "a");
+  await writeFile(join(home, "mcp.json"), PERSON_MCP);
+  const launch = browserAdapters(fix, false).prepare("kimi", "connection-1", { orchestration: true, accountHome: home });
+  assert.equal(launch.args.includes("--mcp-config-file"), false);
+  const servers = JSON.parse(await readFile(join(home, "mcp.json"), "utf8")).mcpServers;
+  assert.deepEqual(Object.keys(servers).sort(), ["canvastty_agents", "canvastty_browser", "person"]);
+  assert.match(await readFile(join(home, "config.toml"), "utf8"), /CanvasTTY temporary browser permission begin/u);
+  assert.deepEqual(await readdir(fix.defaultKimi), [], "the default home is not touched");
+  launch.releaseConfiguration();
+  assert.equal(await readFile(join(home, "mcp.json"), "utf8"), PERSON_MCP);
+  assert.deepEqual((await readdir(home)).sort(), ["mcp.json"]);
+});
+
+test("Kimi tools: symbolic links are refused, and a home the launch does not own falls back to the default home", async (t) => {
+  const fix = await fixture(t);
+  const outside = join(fix.base, "outside");
+  await mkdir(outside, { recursive: true });
+  await writeFile(join(outside, "mcp.json"), PERSON_MCP);
+  const home = fix.accountHome("accounts", "a");
+  await symlink(join(outside, "mcp.json"), join(home, "mcp.json"));
+  const adapters = browserAdapters(fix, false);
+  assert.throws(() => adapters.prepare("kimi", "c", { accountHome: home }), /symbolic link where CanvasTTY writes/u);
+  const linked = fix.accountHome("accounts", "linked");
+  await symlink(outside, linked);
+  assert.throws(() => adapters.prepare("kimi", "c", { accountHome: linked }), /missing or crosses a symbolic link/u);
+  assert.equal(await readFile(join(outside, "mcp.json"), "utf8"), PERSON_MCP);
+  assert.deepEqual(await readdir(outside), ["mcp.json"]);
+
+  const elsewhere = join(fix.base, "elsewhere");
+  await mkdir(elsewhere, { recursive: true });
+  for (const [candidate, runtime] of [[elsewhere, adapters], [fix.accountHome("accounts", "b"), browserAdapters(fix, false, false)]]) {
+    const launch = runtime.prepare("kimi", "c", { accountHome: candidate });
+    assert.ok(JSON.parse(await readFile(join(fix.defaultKimi, "mcp.json"), "utf8")).mcpServers.canvastty_browser);
+    launch.releaseConfiguration();
+    assert.deepEqual(await readdir(candidate), []);
+  }
+});
+
+/** A card a launch contributor (data folder `plugin-data/<owner>`) starts with `variable` = `home`, through TerminalManager. */
+async function contributedCard(t, fix, { provider, variable, home, owner = "accounts" }) {
+  const pipeline = new LaunchPipeline({
+    contributors: () => [{ pluginId: owner, pluginName: "Accounts", serviceId: "launcher",
+      launch: { appliesTo: [provider], fields: [{ key: "on", label: "On", kind: "boolean", default: true }] }, secrets: false,
+      dataDir: join(fix.pluginData, owner) }],
+    call: async () => ({ env: { [variable]: home } }),
+    secret: async () => null,
+    runsRoot: join(fix.base, "runs"),
+    timeoutMs: 1_000
+  });
+  const gateway = {
+    registerSession: (terminalSessionId, kind) => ({ address: join(fix.base, "rt.sock"), terminalSessionId, provider: kind, capabilityToken: "t" }),
+    revokeTerminalSession() {},
+    currentStatus: () => null
+  };
+  const bridge = new AgentRuntimeBridge(gateway, {
+    helper,
+    runtimeDirectory: join(fix.base, "runtime"),
+    openCodePluginPath: join(fix.base, "opencode-plugin.mjs"),
+    grokHomeDirectory: fix.defaultGrok,
+    kimiHomeDirectory: fix.defaultKimi,
+    hermesHomeDirectory: join(fix.base, "home", ".hermes"),
+    accountHomesRoot: fix.pluginData,
+    platform: "linux"
+  });
+  const browserGateway = {
+    isEnabled: true,
+    registerAgent: (input) => ({ agentId: "agent", connectionId: `connection-${input.terminalSessionId}`, terminalSessionId: input.terminalSessionId,
+      provider: input.provider, capabilityToken: "c", address: join(fix.base, "browser.sock"), authenticated: new Promise(() => {}) }),
+    revokeTerminalSession() {},
+    holdPendingForTerminal() {}
+  };
+  const browser = new AgentBrowserBridge(browserGateway, {
+    helper: browserHelper,
+    providerClis: registry,
+    runtimeDirectory: join(fix.base, "browser-runtime"),
+    hermesHomeDirectory: join(fix.base, "home", ".hermes"),
+    kimiHomeDirectory: fix.defaultKimi,
+    probeKimiPerRunConfig: () => false,
+    accountHomesRoot: fix.pluginData
+  });
+  const calls = [];
+  const spawner = (command, args, options) => {
+    const exits = [];
+    calls.push({ command, args, options, exit: (code) => exits.forEach((listener) => listener({ exitCode: code })) });
+    return {
+      pid: 41_000 + calls.length, process: command, write() {}, resize() {}, kill() {}, pause() {}, resume() {},
+      onData() { return { dispose() {} }; },
+      onExit(listener) { exits.push(listener); return { dispose() {} }; }
+    };
+  };
+  const manager = new TerminalManager(() => undefined, registry, browser, bridge, true, spawner);
+  t.after(() => manager.shutdown());
+  manager.configureLaunchPipeline(pipeline);
+  const created = manager.create({ provider, profile: "normal", cwd: process.cwd(), position: { x: 0, y: 0 }, launchOptions: { [owner]: {} } });
+  // Grok starts once the card reports its measured grid.
+  manager.resize(created.id, 100, 30);
+  const deadline = Date.now() + 5_000;
+  while (calls.length === 0 && Date.now() < deadline) await new Promise((resolve) => setTimeout(resolve, 10));
+  assert.equal(calls.length, 1, JSON.stringify(manager.list().map((card) => [card.status, card.failureDetails])));
+  assert.equal(calls[0].options.env[variable], home);
+  return { exit: () => calls[0].exit(0) };
+}
+
+async function until(predicate) {
+  const deadline = Date.now() + 5_000;
+  while (!(await predicate()) && Date.now() < deadline) await new Promise((resolve) => setTimeout(resolve, 10));
+  assert.ok(await predicate());
+}
+
+test("a Grok card a contributor starts with GROK_HOME in its own account home gets the hooks there, through TerminalManager", async (t) => {
+  const fix = await fixture(t);
+  const home = fix.accountHome("accounts", "a");
+  const card = await contributedCard(t, fix, { provider: "grok", variable: "GROK_HOME", home });
+  assert.equal(await exists(join(home, ...GROK_FILE)), true);
+  assert.equal(await exists(join(fix.defaultGrok, ...GROK_FILE)), false);
+  card.exit();
+  await until(async () => !(await exists(join(home, ...GROK_FILE))));
+});
+
+test("a Kimi card in its own account home gets status hooks and its browser/agents tools there, through TerminalManager", async (t) => {
+  const fix = await fixture(t);
+  const home = fix.accountHome("accounts", "a");
+  await writeFile(join(home, "config.toml"), PERSON_KIMI);
+  await writeFile(join(home, "mcp.json"), PERSON_MCP);
+  const card = await contributedCard(t, fix, { provider: "kimi", variable: "KIMI_CODE_HOME", home });
+  const config = await readFile(join(home, "config.toml"), "utf8");
+  assert.ok(config.startsWith(PERSON_KIMI));
+  assert.equal(markers(config), 1);
+  assert.equal(config.split(BROWSER_RULE).length - 1, 1);
+  assert.ok(JSON.parse(await readFile(join(home, "mcp.json"), "utf8")).mcpServers.canvastty_browser);
+  assert.deepEqual(await readdir(fix.defaultKimi), []);
+  card.exit();
+  await until(async () => (await readFile(join(home, "config.toml"), "utf8")) === PERSON_KIMI);
+  await until(async () => (await readFile(join(home, "mcp.json"), "utf8")) === PERSON_MCP);
+});
+
+test("a home in another plugin's data is not the launch's account home: nothing is written there, the default home is used", async (t) => {
+  const fix = await fixture(t);
+  const home = fix.accountHome("accounts", "a");
+  for (const [provider, variable] of [["grok", "GROK_HOME"], ["kimi", "KIMI_CODE_HOME"]]) {
+    const card = await contributedCard(t, fix, { provider, variable, home, owner: "other" });
+    assert.deepEqual(await readdir(home), [], provider);
+    if (provider === "grok") assert.equal(await exists(join(fix.defaultGrok, ...GROK_FILE)), true);
+    else assert.match(await readFile(join(fix.defaultKimi, "config.toml"), "utf8"), /CanvasTTY lifecycle hooks/u);
+    card.exit();
+  }
+});
diff --git a/tests/agent-runtime-gateway.test.mjs b/tests/agent-runtime-gateway.test.mjs
index 61ec999da..67e8bc0df 100644
--- a/tests/agent-runtime-gateway.test.mjs
+++ b/tests/agent-runtime-gateway.test.mjs
@@ -158,6 +158,32 @@ test("RuntimeGateway: a late start of an earlier turn does not take over from th
   assert.equal(gateway.currentStatus("terminal-turns"), "idle");
 });
 
+test("RuntimeGateway: a turn whose start hook was lost still counts after an idle, but not while another turn runs", POSIX_RUNTIME_GATEWAY_TEST, async (t) => {
+  const root = await fixture(t);
+  const signals = [];
+  const gateway = new RuntimeGateway({ runtimeDirectory: root, onSignal: (id, signal) => signals.push({ id, signal }) });
+  await gateway.start();
+  t.after(() => gateway.close());
+  const capability = gateway.registerSession("terminal-lost-start", "codex");
+  await send(capability.address, message(capability, "working", "UserPromptSubmit", "turn-1"));
+  // An event of another turn while turn 1 runs is out of order.
+  await send(capability.address, message(capability, "working", "PostToolUse", "turn-x"));
+  await send(capability.address, message(capability, "idle", "Stop", "turn-1"));
+  // Turn 2's UserPromptSubmit never arrived: its own events still report it.
+  await send(capability.address, message(capability, "working", "PostToolUse", "turn-2"));
+  assert.equal(gateway.currentStatus("terminal-lost-start"), "working");
+  // A late event of the earlier turn stays stale.
+  await send(capability.address, message(capability, "idle", "Stop", "turn-1"));
+  await send(capability.address, message(capability, "idle", "Stop", "turn-2"));
+  // After that idle, an id already seen does not open a turn again.
+  await send(capability.address, message(capability, "working", "PostToolUse", "turn-1"));
+  await new Promise((resolve) => setImmediate(resolve));
+  assert.deepEqual(signals.map(({ signal }) => `${signal.state}:${signal.event}:${signal.turnId}`), [
+    "working:UserPromptSubmit:turn-1", "idle:Stop:turn-1", "working:PostToolUse:turn-2", "idle:Stop:turn-2"
+  ]);
+  assert.equal(gateway.currentStatus("terminal-lost-start"), "idle");
+});
+
 test("RuntimeGateway propagates threadId for codex sessions with canonical UUID", POSIX_RUNTIME_GATEWAY_TEST, async (t) => {
   const root = await fixture(t);
   const signals = [];
diff --git a/tests/agent-runtime-provider-launch.test.mjs b/tests/agent-runtime-provider-launch.test.mjs
index c56a310b2..c3966d872 100644
--- a/tests/agent-runtime-provider-launch.test.mjs
+++ b/tests/agent-runtime-provider-launch.test.mjs
@@ -46,6 +46,7 @@ test("Claude and Codex receive automatic lifecycle hooks without prompt or respo
   assert.equal(settings.showStatusInTerminalTab, true);
   assert.match(settings.hooks.UserPromptSubmit[0].hooks[0].command, /working.*UserPromptSubmit/u);
   assert.match(settings.hooks.PermissionRequest[0].hooks[0].command, /needs_approval.*PermissionRequest/u);
+  assert.match(settings.hooks.PostToolUseFailure[0].hooks[0].command, /working.*PostToolUseFailure/u);
   assert.match(settings.hooks.Stop[0].hooks[0].command, /idle.*Stop/u);
   assert.match(settings.hooks.Stop[0].hooks[0].command, /^ELECTRON_RUN_AS_NODE='1' /u);
   assert.equal(JSON.stringify(settings).includes('"prompt":'), false);
@@ -497,3 +498,25 @@ function runtimeOptionsFor(root, qwenSystemSettingsPath, pluginHooks) {
     environment: {}
   };
 }
+
+test("with status hooks off a result-capturing subagent gets no lease and no lifecycle hooks", async (t) => {
+  const root = await fixture(t);
+  const registrations = [];
+  const gateway = {
+    registerSession(id, provider, capture, _grant, decisions) {
+      registrations.push({ id, capture, decisions });
+      return { address: join(root, "runtime.sock"), terminalSessionId: id, provider, capabilityToken: "token" };
+    },
+    revokeTerminalSession() {},
+    currentStatus: () => null
+  };
+  const bridge = new AgentRuntimeBridge(gateway, { ...runtimeOptionsFor(root), coreHooksEnabled: false });
+  for (const provider of ["codex", "opencode"]) {
+    const launch = bridge.prepareLaunch({ terminalSessionId: provider, provider, cwd: root, captureResult: true, decisions: false });
+    assert.equal(launch.environment[AGENT_RUNTIME_ENV.capabilityToken], undefined, `${provider}: turning status hooks off means no hooks`);
+    assert.deepEqual(launch.args.filter((arg) => /^hooks\./u.test(arg)), []);
+    assert.equal(launch.environment.OPENCODE_CONFIG_CONTENT, undefined);
+    launch.cleanup();
+  }
+  assert.deepEqual(registrations, []);
+});
diff --git a/tests/browser-engine-tabs.test.mjs b/tests/browser-engine-tabs.test.mjs
index eec0d8139..7e3f9d288 100644
--- a/tests/browser-engine-tabs.test.mjs
+++ b/tests/browser-engine-tabs.test.mjs
@@ -162,10 +162,12 @@ async function harness(t, { providers } = {}) {
   t.after(() => engines.dispose());
   let request = 0;
   const run = (actor, type, args = {}) => core.execute(actor, { type, requestId: `r${++request}`, ...args });
-  const ready = async (tabId) => {
+  // `after`: wait for a navigation the page starts by itself (a click), which the engine reports only after the
+  // command answered; until its Page.frameNavigated arrives the tab still shows the old document as ready.
+  const ready = async (tabId, { after } = {}) => {
     for (let attempt = 0; attempt < 100; attempt += 1) {
       const tab = host.getTab(tabId);
-      if (tab?.status === "ready") return tab;
+      if (tab?.status === "ready" && (after === undefined || tab.documentRevision > after)) return tab;
       await new Promise((resolve) => setTimeout(resolve, 10));
     }
     throw new Error("tab never became ready");
@@ -233,8 +235,10 @@ test("engines without layout: observe skips the fake geometry and clicks go thro
   const methods = cdpMethods(h.engine);
   assert.equal(methods.includes("Input.dispatchMouseEvent"), false, "no coordinate clicks");
   assert.equal(methods.includes("DOM.getBoxModel"), false, "no box models asked for");
-  await h.ready(opened.tabId);
-  assert.ok(h.host.getTab(opened.tabId).documentRevision > before, "the click navigated: a new revision");
+  // The click answers before the engine reports the navigation it started, as Chromium's does.
+  const navigated = await h.ready(opened.tabId, { after: before });
+  assert.ok(navigated.documentRevision > before, "the click navigated: a new revision");
+  assert.equal(navigated.url, DOCS);
   const hovered = await h.run(agent(), "browser_observe", { tabId: opened.tabId });
   assert.equal(hovered.ok, true);
 });
diff --git a/tests/claude-http-hooks.test.mjs b/tests/claude-http-hooks.test.mjs
index fb649ea87..94791bd48 100644
--- a/tests/claude-http-hooks.test.mjs
+++ b/tests/claude-http-hooks.test.mjs
@@ -262,7 +262,7 @@ test("Claude's lifecycle hooks become HTTP hooks except SessionStart; the decisi
   assert.equal(settings.hooks.PreToolUse[0].hooks[0].type, "command");
   assert.match(settings.hooks.PreToolUse[0].hooks[0].command, /permission-gate\.mjs' 'pretool'/u);
   for (const [event, state] of [["UserPromptSubmit", "working"], ["PermissionRequest", "needs_approval"], ["PostToolUse", "working"],
-    ["Stop", "idle"], ["StopFailure", "idle"], ["SessionEnd", "idle"], ["Notification", "needs_approval"]]) {
+    ["PostToolUseFailure", "working"], ["Stop", "idle"], ["StopFailure", "idle"], ["SessionEnd", "idle"], ["Notification", "needs_approval"]]) {
     const hook = settings.hooks[event][0].hooks[0];
     assert.deepEqual(hook, {
       type: "http",
diff --git a/tests/decision-hooks.test.mjs b/tests/decision-hooks.test.mjs
index b7a4a2308..2afb7fda2 100644
--- a/tests/decision-hooks.test.mjs
+++ b/tests/decision-hooks.test.mjs
@@ -345,7 +345,7 @@ test("OpenCode: the guard throws a deny, remembers an allow and answers OpenCode
   const sent = [];
   const decisions = createOpenCodeDecisions({ client, env, send: async (_address, message) => { sent.push(message); return next; } });
   await assert.rejects(decisions.guard({ tool: "write", callID: "c1" }, { args: { filePath: "/etc/hosts", content: "x" } }), /Not outside the project/u);
-  assert.equal(sent[0].toolName, "edit");
+  assert.equal(sent[0].toolName, "Write", "a whole-file write keeps its own name");
   next = { behavior: "allow", message: "" };
   await decisions.guard({ tool: "bash", callID: "c2" }, { args: { command: "ls" } });
   assert.equal(await decisions.permissionAsked({ id: "r1", tool: { callID: "c2" } }), true);
@@ -358,9 +358,27 @@ test("OpenCode: the guard throws a deny, remembers an allow and answers OpenCode
   assert.equal(sent.length, 3, "tools that neither run commands nor write files are not sent");
   assert.equal(createOpenCodeDecisions({ env: { ...env, [OPENCODE_DECISIONS_ENV]: "0" } }).enabled, false);
   assert.deepEqual(guardedCall("apply_patch", { patchText: "*** Begin Patch" }), { toolName: "apply_patch", toolInput: { patch: "*** Begin Patch" } });
-  // multiedit: every edit's new text reaches the decision, as Claude's MultiEdit tool_input does.
-  assert.deepEqual(guardedCall("multiedit", { filePath: "/p/a.ts", edits: [{ oldString: "a", newString: "TOKEN=1" }, { oldString: "b", newString: "rm -rf /" }, { newString: 5 }] }),
-    { toolName: "edit", toolInput: { file_path: "/p/a.ts", content: "TOKEN=1\nrm -rf /" } });
+  // OpenCode's file tools go out in Claude Code's shapes: Write is the whole new file; Edit and MultiEdit carry the
+  // text they replace as well as the new text (OpenCode run 15: edit {filePath, oldString, newString, replaceAll}).
+  assert.deepEqual(guardedCall("write", { filePath: "/p/new.mjs", content: "export const a = 1;\n" }), { toolName: "Write", toolInput: { file_path: "/p/new.mjs", content: "export const a = 1;\n" } });
+  assert.deepEqual(guardedCall("edit", { filePath: "/p/a.ts", oldString: "const a = 1;", newString: "const a = 2;", replaceAll: false }),
+    { toolName: "Edit", toolInput: { file_path: "/p/a.ts", old_string: "const a = 1;", new_string: "const a = 2;", replace_all: false } });
+  assert.deepEqual(guardedCall("edit", { filePath: "/p/a.ts", oldString: "x", newString: "" }),
+    { toolName: "Edit", toolInput: { file_path: "/p/a.ts", old_string: "x", new_string: "" } }, "a deletion keeps its empty new text; replaceAll absent stays absent");
+  assert.deepEqual(guardedCall("multiedit", { filePath: "/p/a.ts", edits: [{ filePath: "/p/a.ts", oldString: "a", newString: "TOKEN=1" },
+    { oldString: "b", newString: "rm -rf /", replaceAll: true }, { newString: 5 }, null] }),
+  { toolName: "MultiEdit", toolInput: { file_path: "/p/a.ts", edits: [{ old_string: "a", new_string: "TOKEN=1" },
+    { old_string: "b", new_string: "rm -rf /", replace_all: true }] } });
+  assert.equal(guardedCall("edit", { oldString: "a", newString: "b" }), null, "no path, nothing to check");
+  // Through the host: base protection still finds the path, and the service gets the edit's kind, path and both texts.
+  const calls = [];
+  const host = hooks({ services: [service("p")], answers: { p: null }, calls });
+  const edit = guardedCall("edit", { filePath: join(project, "a.ts"), oldString: "a", newString: "b" });
+  assert.equal((await host.decide("s1", request(edit.toolName, edit.toolInput), live())).behavior, "none");
+  assert.deepEqual(calls[0].params.tool, { name: "Edit", kind: "edit", command: null, paths: [join(project, "a.ts")] });
+  assert.deepEqual(calls[0].params.input, { file_path: join(project, "a.ts"), old_string: "a", new_string: "b" });
+  const outside = guardedCall("multiedit", { filePath: "/etc/hosts", edits: [{ oldString: "a", newString: "b" }] });
+  assert.equal((await host.decide("s1", request(outside.toolName, outside.toolInput), live())).behavior, "deny", "MultiEdit outside the project");
 });
 
 test("manifest: decide needs decision:provide, lists pre-tool, and at most one service per plugin decides", () => {
diff --git a/tests/github-auth-integration.test.mjs b/tests/github-auth-integration.test.mjs
index 7035b7eea..078a00a1e 100644
--- a/tests/github-auth-integration.test.mjs
+++ b/tests/github-auth-integration.test.mjs
@@ -9,6 +9,8 @@ const settingsPath = new URL(
   "../src/renderer/src/features/plugins/PluginSettingsSection.tsx",
   import.meta.url
 );
+const ipcPath = new URL("../src/main/ipc/registerIpc.ts", import.meta.url);
+const preloadPath = new URL("../src/preload/index.ts", import.meta.url);
 
 test("GitHub sign-in supports optional build input and both browser routes", async () => {
   const [config, main, releaseWorkflow, settings] = await Promise.all([
@@ -31,3 +33,19 @@ test("GitHub sign-in supports optional build input and both browser routes", asy
   assert.match(settings, /else await window\.canvasTTY\.githubAuth\.openUrl\(url\)/);
   assert.match(settings, /const url = flow\.verificationUri/);
 });
+
+test("a pending GitHub sign-in can be cancelled through the trusted renderer only", async () => {
+  const [ipc, preload, settings] = await Promise.all([
+    readFile(ipcPath, "utf8"),
+    readFile(preloadPath, "utf8"),
+    readFile(settingsPath, "utf8")
+  ]);
+  const start = ipc.indexOf("ipcMain.handle(IPC.githubAuthCancel,");
+  assert.notEqual(start, -1);
+  const handler = ipc.slice(start, ipc.indexOf("ipcMain.", start + 1));
+  assert.match(handler, /assertMainRenderer\(event, getMainWindow\)/);
+  assert.match(handler, /githubAuth\.cancelDeviceFlow\(\)/);
+  assert.match(preload, /cancel: \(\) => ipcRenderer\.invoke\(IPC\.githubAuthCancel\)/);
+  assert.match(settings, /window\.canvasTTY\.githubAuth\.cancel\(\)/);
+  assert.match(settings, /githubSigninText\(locale, "cancel"\)/);
+});
diff --git a/tests/github-auth.test.mjs b/tests/github-auth.test.mjs
index 793039552..9f53255de 100644
--- a/tests/github-auth.test.mjs
+++ b/tests/github-auth.test.mjs
@@ -1,5 +1,5 @@
 import assert from "node:assert/strict";
-import { mkdtemp, readFile, readdir, rm } from "node:fs/promises";
+import { mkdtemp, readFile, readdir, rm, writeFile } from "node:fs/promises";
 import { tmpdir } from "node:os";
 import test from "node:test";
 import { GithubAuthService } from "../src/main/services/GithubAuthService.ts";
@@ -63,7 +63,8 @@ test("GitHub device flow persists an encrypted session atomically without reques
       configured: true,
       authorized: true,
       login: "howdeploy",
-      tokenExpiresAt: null
+      tokenExpiresAt: null,
+      deviceFlowState: "idle"
     });
     assert.equal(await restored.getToken(), "access");
     await service.signOut();
@@ -157,7 +158,8 @@ test("signOut cancels an in-flight device flow and prevents a late token from be
       configured: true,
       authorized: false,
       login: null,
-      tokenExpiresAt: null
+      tokenExpiresAt: null,
+      deviceFlowState: "idle"
     });
     await assert.rejects(() => readFile(`${userData}/github-oauth.json`, "utf8"), { code: "ENOENT" });
   } finally {
@@ -211,7 +213,8 @@ test("reports an unconfigured OAuth client before starting device flow", async (
       configured: false,
       authorized: false,
       login: null,
-      tokenExpiresAt: null
+      tokenExpiresAt: null,
+      deviceFlowState: "idle"
     });
     await assert.rejects(
       () => service.startDeviceFlow(),
@@ -271,3 +274,260 @@ test("device polling survives network errors and timeouts, backs off, and honour
     await rm(userData, { recursive: true, force: true });
   }
 });
+
+const deviceCodeAnswer = () => Response.json({
+  device_code: "test-device-code",
+  user_code: "TEST-CODE",
+  verification_uri: "https://github.com/login/device",
+  expires_in: 900,
+  interval: 1
+});
+
+function serviceWith(userData, pollAnswer, options = {}) {
+  const requests = { polls: 0, profile: 0 };
+  const fetcher = async (url, init = {}) => {
+    if (String(url).endsWith("/login/device/code")) return deviceCodeAnswer();
+    if (String(url).endsWith("/login/oauth/access_token")) {
+      requests.polls += 1;
+      return pollAnswer(requests.polls, init);
+    }
+    if (String(url) === "https://api.github.com/user") {
+      requests.profile += 1;
+      return Response.json({ login: "fixture-user" });
+    }
+    return new Response("missing", { status: 404 });
+  };
+  const service = new GithubAuthService(userData, "client-id", {
+    fetcher,
+    safeStorage,
+    delay: async () => undefined,
+    ...options
+  });
+  return { service, requests };
+}
+
+test("access_denied, expired_token and unknown provider errors stop polling with a reported outcome", async () => {
+  for (const [answer, expected] of [
+    [() => Response.json({ error: "access_denied" }), "denied"],
+    [() => Response.json({ error: "expired_token" }), "expired"],
+    [() => Response.json({ error: "access_denied" }, { status: 400 }), "denied"],
+    [() => Response.json({ error: "device_flow_disabled" }), "failed"],
+    [() => Response.json({ error: "incorrect_client_credentials" }, { status: 400 }), "failed"],
+    [() => new Response("forbidden", { status: 403 }), "failed"]
+  ]) {
+    const userData = await mkdtemp(`${tmpdir()}/canvastty-github-auth-${expected}-`);
+    try {
+      const { service, requests } = serviceWith(userData, answer);
+      await service.startDeviceFlow();
+      await waitFor(async () => (await service.status()).deviceFlowState !== "pending");
+      const status = await service.status();
+      assert.equal(status.authorized, false);
+      assert.equal(status.deviceFlowState, expected);
+      // Give a runaway poll loop a chance to show itself.
+      for (let i = 0; i < 5; i += 1) await new Promise((resolve) => setImmediate(resolve));
+      assert.equal(requests.polls, 1, `${expected}: polling stopped after the terminal answer`);
+      assert.equal(requests.profile, 0);
+      await assert.rejects(() => readFile(`${userData}/github-oauth.json`, "utf8"), { code: "ENOENT" });
+    } finally {
+      await rm(userData, { recursive: true, force: true });
+    }
+  }
+});
+
+test("cancelDeviceFlow stops polling, reports cancelled and keeps an existing session", async () => {
+  const userData = await mkdtemp(`${tmpdir()}/canvastty-github-auth-cancel-flow-`);
+  const existing = { accessToken: "existing-fx", refreshToken: null, expiresAt: null, login: "existing-user" };
+  const storedFile = JSON.stringify({ data: Buffer.from(JSON.stringify(existing)).toString("base64") });
+  await writeFile(`${userData}/github-oauth.json`, storedFile);
+  const pollSignals = [];
+  const releases = [];
+  try {
+    const { service, requests } = serviceWith(
+      userData,
+      // The fetcher ignores its AbortSignal on purpose: a late answer must still be discarded.
+      () => new Promise((resolve) => { releases.push(resolve); }),
+      {
+        delay: (_duration, signal) => {
+          pollSignals.push(signal);
+          return signal.aborted ? Promise.reject(new DOMException("cancelled", "AbortError")) : Promise.resolve();
+        }
+      }
+    );
+    await service.load();
+    await service.startDeviceFlow();
+    await waitFor(() => releases.length === 1);
+    assert.equal((await service.status()).deviceFlowState, "pending");
+
+    service.cancelDeviceFlow();
+    assert.equal(pollSignals[0].aborted, true);
+    assert.equal((await service.status()).deviceFlowState, "cancelled");
+
+    releases[0](Response.json({ access_token: "late-fx", token_type: "bearer", scope: "" }));
+    for (let i = 0; i < 5; i += 1) await new Promise((resolve) => setImmediate(resolve));
+
+    const status = await service.status();
+    assert.equal(status.deviceFlowState, "cancelled");
+    assert.equal(status.login, "existing-user");
+    assert.equal(requests.polls, 1, "no poll after cancel");
+    assert.equal(requests.profile, 0, "the late token was not used");
+    assert.equal(await service.getToken(), "existing-fx");
+    assert.equal(await readFile(`${userData}/github-oauth.json`, "utf8"), storedFile);
+
+    // Cancelling again, or with nothing pending, changes nothing.
+    service.cancelDeviceFlow();
+    assert.equal((await service.status()).deviceFlowState, "cancelled");
+  } finally {
+    await rm(userData, { recursive: true, force: true });
+  }
+});
+
+test("cancelDeviceFlow after a completed sign-in keeps it", async () => {
+  const userData = await mkdtemp(`${tmpdir()}/canvastty-github-auth-cancel-done-`);
+  try {
+    const { service } = serviceWith(userData, () => Response.json({ access_token: "done-fx", token_type: "bearer", scope: "" }));
+    await service.startDeviceFlow();
+    await waitFor(async () => (await service.status()).authorized);
+    const completed = await service.status();
+    assert.equal(completed.deviceFlowState, "idle");
+    service.cancelDeviceFlow();
+    assert.deepEqual(await service.status(), completed);
+    assert.equal(await service.getToken(), "done-fx");
+    await service.signOut();
+  } finally {
+    await rm(userData, { recursive: true, force: true });
+  }
+});
+
+test("slow_down adds five seconds to this and every later poll", async () => {
+  const userData = await mkdtemp(`${tmpdir()}/canvastty-github-auth-slow-down-`);
+  const waits = [];
+  const answers = [
+    { error: "authorization_pending" },
+    { error: "slow_down" },
+    { error: "authorization_pending" },
+    { error: "slow_down" },
+    { error: "authorization_pending" },
+    { access_token: "slow-fx", token_type: "bearer", scope: "" }
+  ];
+  try {
+    const { service } = serviceWith(userData, () => Response.json(answers.shift()), {
+      delay: async (ms) => { waits.push(ms); }
+    });
+    await service.startDeviceFlow();
+    await waitFor(async () => (await service.status()).authorized);
+    assert.deepEqual(waits, [1_000, 1_000, 6_000, 6_000, 11_000, 11_000]);
+    await service.signOut();
+  } finally {
+    await rm(userData, { recursive: true, force: true });
+  }
+});
+
+test("transient poll errors retry with back-off, reset on a valid answer, and give up after a bounded run", async () => {
+  const userData = await mkdtemp(`${tmpdir()}/canvastty-github-auth-bounded-`);
+  let clock = 1_000_000;
+  const waits = [];
+  const transient = [
+    () => { throw new TypeError("fetch failed"); },
+    () => new Response("unavailable", { status: 503 }),
+    () => new Response("too many", { status: 429 }),
+    () => new Response("not json", { status: 200 }),
+    () => new Response("bad gateway", { status: 502 })
+  ];
+  // Four failures, one pending answer (resets the count), then five failures in a row.
+  const answers = [...transient.slice(0, 4), () => Response.json({ error: "authorization_pending" }), ...transient];
+  const warn = console.warn;
+  console.warn = () => undefined;
+  try {
+    const { service, requests } = serviceWith(userData, () => answers.shift()(), {
+      now: () => clock,
+      delay: async (ms) => { waits.push(ms); clock += ms; }
+    });
+    await service.startDeviceFlow();
+    await waitFor(async () => (await service.status()).deviceFlowState !== "pending");
+    assert.equal((await service.status()).deviceFlowState, "failed");
+    assert.equal(requests.polls, 10);
+    assert.equal(answers.length, 0);
+    assert.deepEqual(waits, [1_000, 2_000, 4_000, 8_000, 16_000, 16_000, 32_000, 60_000, 60_000, 60_000]);
+  } finally {
+    console.warn = warn;
+    await rm(userData, { recursive: true, force: true });
+  }
+});
+
+test("the device code expires on the host clock even while GitHub keeps answering pending", async () => {
+  const userData = await mkdtemp(`${tmpdir()}/canvastty-github-auth-elapsed-`);
+  let clock = 1_000_000;
+  try {
+    const { service, requests } = serviceWith(userData, () => Response.json({ error: "authorization_pending" }), {
+      now: () => clock,
+      pollTimeoutMs: 2_500,
+      delay: async (ms) => { clock += ms; }
+    });
+    await service.startDeviceFlow();
+    await waitFor(async () => (await service.status()).deviceFlowState !== "pending");
+    assert.equal((await service.status()).deviceFlowState, "expired");
+    assert.equal(requests.polls, 2);
+  } finally {
+    await rm(userData, { recursive: true, force: true });
+  }
+});
+
+test("signOut clears a reported sign-in outcome", async () => {
+  const userData = await mkdtemp(`${tmpdir()}/canvastty-github-auth-signout-outcome-`);
+  try {
+    const { service } = serviceWith(userData, () => Response.json({ error: "access_denied" }));
+    await service.startDeviceFlow();
+    await waitFor(async () => (await service.status()).deviceFlowState === "denied");
+    await service.signOut();
+    assert.equal((await service.status()).deviceFlowState, "idle");
+  } finally {
+    await rm(userData, { recursive: true, force: true });
+  }
+});
+
+test("a profile request that times out after the token arrived reports a failed sign-in, not a stuck pending one", async () => {
+  const userData = await mkdtemp(`${tmpdir()}/canvastty-github-auth-profile-timeout-`);
+  const fetcher = async (url, init = {}) => {
+    if (String(url).endsWith("/login/device/code")) return deviceCodeAnswer();
+    if (String(url).endsWith("/login/oauth/access_token")) return Response.json({ access_token: "access", token_type: "bearer", scope: "" });
+    // The profile request hangs until the service's own request timeout aborts it.
+    return new Promise((_resolve, reject) => {
+      init.signal.addEventListener("abort", () => reject(new DOMException("The operation was aborted.", "AbortError")), { once: true });
+    });
+  };
+  const warn = console.warn;
+  console.warn = () => undefined;
+  try {
+    const service = new GithubAuthService(userData, "client-id", {
+      fetcher, safeStorage, requestTimeoutMs: 10, delay: async () => undefined
+    });
+    await service.startDeviceFlow();
+    await waitFor(async () => (await service.status()).deviceFlowState !== "pending");
+    const status = await service.status();
+    assert.equal(status.deviceFlowState, "failed");
+    assert.equal(status.authorized, false);
+    await assert.rejects(() => readFile(`${userData}/github-oauth.json`, "utf8"), { code: "ENOENT" });
+  } finally {
+    console.warn = warn;
+    await rm(userData, { recursive: true, force: true });
+  }
+});
+
+test("HTTP 429 and 5xx answers with an OAuth error body are retried like other transient answers", async () => {
+  for (const status of [429, 503]) {
+    const userData = await mkdtemp(`${tmpdir()}/canvastty-github-auth-transient-json-`);
+    try {
+      const { service, requests } = serviceWith(userData, (poll) => (poll === 1
+        ? Response.json({ error: "temporarily_unavailable" }, { status })
+        : Response.json({ access_token: "access", token_type: "bearer", scope: "" })));
+      await service.startDeviceFlow();
+      await waitFor(async () => (await service.status()).deviceFlowState !== "pending" || (await service.status()).authorized);
+      const result = await service.status();
+      assert.equal(result.authorized, true, `HTTP ${status}: the sign-in went on after the transient answer`);
+      assert.equal(requests.polls, 2);
+      await service.signOut();
+    } finally {
+      await rm(userData, { recursive: true, force: true });
+    }
+  }
+});
diff --git a/tests/github-signin-errors-ui.test.mjs b/tests/github-signin-errors-ui.test.mjs
new file mode 100644
index 000000000..db7e6edd9
--- /dev/null
+++ b/tests/github-signin-errors-ui.test.mjs
@@ -0,0 +1,72 @@
+// The GitHub sign-in buttons show localized outcomes; main-process error text (IPC prefix, HTTP codes) only goes to the log.
+import assert from "node:assert/strict";
+import test from "node:test";
+import { findAll, importWithFakeReact, tick } from "./helpers/fake-react.mjs";
+
+const section = await importWithFakeReact("src/renderer/src/features/plugins/PluginSettingsSection.tsx", "PluginSettingsSection");
+
+function harness(t, locale, githubAuth) {
+  const oldWindow = globalThis.window;
+  const warnings = [];
+  t.mock.method(console, "warn", (...args) => { warnings.push(args); });
+  globalThis.window = { setTimeout, clearTimeout, canvasTTY: { githubAuth, appVersion: async () => "1.8.1" } };
+  t.after(() => { section.__unmount(); globalThis.window = oldWindow; });
+  section.__reset();
+  const none = async () => ({});
+  const props = { settings: { locale }, plugins: [], open: false, onPreviewPlugin: none, onInstallPlugin: none, onSearchPlugins: async () => [],
+    onShowcasePlugins: async () => [], onFetchPluginIcons: none, onPreviewManifests: none, onOpenBrowser: none, onCheckPluginUpdates: none,
+    onUpdatePlugin: none, onSetPluginModules: none, onSetPluginEnabled: none, onUninstallPlugin: none, onOpenPluginContribution: none };
+  let tree;
+  const render = () => { section.__flush(); tree = section.__render(section.PluginSettingsSection, props); return tree; };
+  const settle = async () => { for (let i = 0; i < 6; i++) { render(); await tick(); } render(); };
+  const button = (label) => findAll(tree, (node) => node.type === "button" && node.props?.children === label)[0];
+  return { render, settle, button, warnings, tree: () => tree };
+}
+
+const ipcFailure = new Error("Error invoking remote method 'github-auth:start': Error: GitHub device flow failed with HTTP 404.");
+
+test("a sign-in that cannot start shows the localized failure in the GitHub block and logs the details (ru)", async (t) => {
+  const status = { configured: true, authorized: false, login: null, tokenExpiresAt: null, deviceFlowState: "idle" };
+  const ui = harness(t, "ru", { status: async () => status, start: async () => { throw ipcFailure; } });
+  await ui.settle();
+  const signIn = ui.button("Войти в системном браузере");
+  assert.ok(signIn, "the sign-in button");
+  signIn.props.onClick();
+  await ui.settle();
+  const json = JSON.stringify(ui.tree());
+  assert.match(json, /Не удалось завершить вход в GitHub/u);
+  assert.doesNotMatch(json, /Error invoking|github-auth:start|HTTP 404/u);
+  assert.ok(ui.warnings.some(([, reason]) => reason === ipcFailure), "the details are logged");
+});
+
+test("a build without an OAuth client shows its own line, not the raw error", async (t) => {
+  let configured = true;
+  const ui = harness(t, "en", {
+    status: async () => ({ configured, authorized: false, login: null, tokenExpiresAt: null, deviceFlowState: "idle" }),
+    start: async () => { configured = false; throw new Error("Error invoking remote method 'github-auth:start': Error: GitHub OAuth is not configured (missing client id)."); }
+  });
+  await ui.settle();
+  ui.button("Sign in with the system browser").props.onClick();
+  await ui.settle();
+  const json = JSON.stringify(ui.tree());
+  assert.match(json, /this build has no OAuth Client ID/u);
+  assert.doesNotMatch(json, /Error invoking|missing client id/u);
+});
+
+test("a failed, denied or expired sign-in reads as an error; a cancelled one as a plain status line", async (t) => {
+  const line = (tree, text) => findAll(tree, (node) => node.type === "p" && node.props?.children === text)[0];
+  for (const [state, text, error] of [
+    ["failed", "Не удалось завершить вход в GitHub. Можно попробовать снова.", true],
+    ["denied", null, true],
+    ["expired", null, true],
+    ["cancelled", null, false]
+  ]) {
+    const ui = harness(t, "ru", { status: async () => ({ configured: true, authorized: false, login: null, tokenExpiresAt: null, deviceFlowState: state }) });
+    await ui.settle();
+    const notice = text ? line(ui.tree(), text)
+      : findAll(ui.tree(), (node) => node.type === "p" && /plugin-github-unavailable|plugin-settings__error/u.test(node.props?.className ?? ""))[0];
+    assert.ok(notice, state);
+    assert.equal(notice.props.className, error ? "plugin-settings__error" : "plugin-github-unavailable", state);
+    assert.equal(notice.props.role, error ? "alert" : "status", state);
+  }
+});
diff --git a/tests/helpers/terminal.mjs b/tests/helpers/terminal.mjs
index c36d9d628..08a8b35c4 100644
--- a/tests/helpers/terminal.mjs
+++ b/tests/helpers/terminal.mjs
@@ -25,6 +25,18 @@ export function availableRegistry() {
  * returns a fake PTY: `emitData`/`emitExit` drive its listeners, `lastResize`
  * holds the last size, and `onWrite(data, options)` sees what was written.
  */
+/**
+ * What OpenCode prints once its prompt has focus: the alternate screen, its hidden cursor while it draws, then the
+ * cursor shown in the input field (TerminalManager types queued input only after that for OpenCode).
+ */
+export const OPENCODE_INPUT_READY = "\x1b[?1049h\x1b[?2004h\x1b[?25l  Ask anything\x1b[?25h";
+
+/** Shows OpenCode's input field on the fake PTY OpenCode was spawned on, once that PTY exists. */
+export async function openCodeInputReady(calls, index = calls.length) {
+  for (let i = 0; i < 200 && calls.length <= index; i++) await new Promise((resolve) => setTimeout(resolve, 5));
+  calls[index].process.emitData(OPENCODE_INPUT_READY);
+}
+
 export function fakeSpawner(calls, { pidBase = 20_000, onWrite } = {}) {
   return (command, args, options) => {
     let dataListener = () => undefined;
@@ -53,3 +65,20 @@ export function fakeSpawner(calls, { pidBase = 20_000, onWrite } = {}) {
     return process;
   };
 }
+
+/** Codex's empty task composer with its dim placeholder (automated Codex prompts wait for it). */
+export const CODEX_COMPOSER_READY = "\x1b[2J\x1b[H>_ OpenAI Codex\r\n\r\n› \x1b[2mAsk Codex to do anything\x1b[22m\r\n";
+
+/**
+ * fakeSpawner whose PTYs show their CLI's input field shortly after they start: OpenCode's cursor in its full-screen
+ * view, Codex's empty task composer (input waits for those for OpenCode and for automated Codex prompts).
+ */
+export function inputReadySpawner(calls, options) {
+  const spawner = fakeSpawner(calls, options);
+  return (command, args, spawnOptions) => {
+    const process = spawner(command, args, spawnOptions);
+    const ready = /(?:^|\/)codex$/u.test(command) ? CODEX_COMPOSER_READY : OPENCODE_INPUT_READY;
+    setTimeout(() => process.emitData(ready), 10);
+    return process;
+  };
+}
diff --git a/tests/launch-cleanup-order.test.mjs b/tests/launch-cleanup-order.test.mjs
new file mode 100644
index 000000000..f18e5c115
--- /dev/null
+++ b/tests/launch-cleanup-order.test.mjs
@@ -0,0 +1,211 @@
+import assert from "node:assert/strict";
+import { mkdtemp, rm, stat } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import test from "node:test";
+import { LaunchPipeline } from "../src/main/services/LaunchPipeline.ts";
+import { TerminalManager } from "../src/main/services/TerminalManager.ts";
+import { availableRegistry, fakeSpawner } from "./helpers/terminal.mjs";
+
+function deferred() {
+  let resolve;
+  const promise = new Promise((done) => { resolve = done; });
+  return { promise, resolve };
+}
+
+function cleanupPipeline(runs, serviceCalls = []) {
+  return new LaunchPipeline({
+    contributors: () => [{
+      pluginId: "fixture.cleanup",
+      pluginName: "Cleanup fixture",
+      serviceId: "launch",
+      secrets: false,
+      launch: { fields: [{ key: "enabled", label: "Enabled", kind: "boolean", default: true }] }
+    }],
+    call: async (...args) => {
+      serviceCalls.push(args);
+      return { env: { FIXTURE_FILE: "{launchFiles}/run.txt" }, files: [{ relPath: "run.txt", content: "temporary launch data" }] };
+    },
+    secret: async () => null,
+    runsRoot: runs,
+    timeoutMs: 2000
+  });
+}
+
+const createCard = (terminals, cwd) => terminals.create({
+  provider: "codex",
+  profile: "normal",
+  cwd,
+  position: { x: 0, y: 0 },
+  launchOptions: { "fixture.cleanup": { enabled: true } }
+});
+
+/**
+ * One card with a launch plugin. Each run's cleanup is held until the test releases it, and every step of the
+ * child (run folder) and parent (card folder) removals is recorded in `events`.
+ */
+async function fixture(t, { rejectChild = false } = {}) {
+  const root = await mkdtemp(join(tmpdir(), "ctty-cleanup-order-"));
+  const runs = join(root, "runs");
+  const rows = [];
+  const calls = [];
+  const events = [];
+  const spawned = [deferred(), deferred()];
+  const forgotten = deferred();
+  const pipeline = cleanupPipeline(runs);
+  const prepare = pipeline.prepare.bind(pipeline);
+  const forget = pipeline.forgetSession.bind(pipeline);
+  pipeline.prepare = async (context) => {
+    const contribution = await prepare(context);
+    assert.equal(contribution.ok, true);
+    const cleanup = contribution.cleanup;
+    const index = rows.length;
+    const row = { started: deferred(), release: deferred(), calls: 0 };
+    rows.push(row);
+    contribution.cleanup = async () => {
+      row.calls++;
+      events.push(`child${index}:start`);
+      row.started.resolve();
+      try {
+        await row.release.promise;
+        if (rejectChild) throw new Error("fixture child cleanup failure");
+        await cleanup();
+      } finally {
+        events.push(`child${index}:settled`);
+      }
+    };
+    return contribution;
+  };
+  let forgetCalls = 0;
+  pipeline.forgetSession = (id, runDirectories) => {
+    forgetCalls++;
+    events.push("parent:start");
+    const pending = forget(id, runDirectories).then(() => { events.push("parent:done"); });
+    forgotten.resolve(pending);
+    return pending;
+  };
+  const fake = fakeSpawner(calls);
+  const terminals = new TerminalManager(() => undefined, availableRegistry(), undefined, undefined, true, (...args) => {
+    const pty = fake(...args);
+    spawned[calls.length - 1]?.resolve();
+    return pty;
+  });
+  terminals.configureLaunchPipeline(pipeline);
+  t.after(async () => {
+    for (const row of rows) row.release.resolve();
+    terminals.disposeAll();
+    await forgotten.promise;
+    await rm(root, { recursive: true, force: true });
+  });
+  const session = createCard(terminals, root);
+  await spawned[0].promise;
+  return { root, runs, rows, calls, events, spawned, forgotten, terminals, session, forgetCalls: () => forgetCalls };
+}
+
+for (const exited of [false, true]) {
+  for (const rejectChild of [false, true]) {
+    test(`parent launch cleanup waits for ${exited ? "previous exit" : "active disposal"} cleanup (${rejectChild ? "rejection" : "success"})`, { timeout: 5000 }, async (t) => {
+      const f = await fixture(t, { rejectChild });
+      if (exited) f.calls[0].process.emitExit(1);
+      f.terminals.dispose(f.session.id);
+      f.terminals.dispose(f.session.id);
+      await f.rows[0].started.promise;
+      assert.equal(f.forgetCalls(), 0, "parent removal must not overlap a held child removal");
+      assert.deepEqual(f.events, ["child0:start"]);
+
+      f.rows[0].release.resolve();
+      await f.forgotten.promise;
+      assert.deepEqual(f.events, ["child0:start", "child0:settled", "parent:start", "parent:done"]);
+      assert.equal(f.rows[0].calls, 1);
+      assert.equal(f.forgetCalls(), 1, "disposal remains idempotent");
+      await assert.rejects(stat(join(f.runs, f.session.id)), { code: "ENOENT" }, "parent still cleaned after child failure");
+    });
+  }
+}
+
+test("restart retains every pending run cleanup and invokes each captured callback once before forgetting the card", { timeout: 5000 }, async (t) => {
+  const f = await fixture(t);
+  f.calls[0].process.emitExit(1);
+  await f.rows[0].started.promise;
+  f.terminals.restart(f.session.id);
+  await f.spawned[1].promise;
+  assert.equal(f.rows.length, 2);
+
+  f.terminals.dispose(f.session.id);
+  assert.equal(f.rows[1].calls, 0, "the newer cleanup queues behind the old launch callback");
+  assert.equal(f.forgetCalls(), 0);
+  // A late notification from the old PTY must neither consume nor overwrite the new launch cleanup.
+  f.calls[0].process.emitExit(1);
+  f.rows[0].release.resolve();
+  await f.rows[1].started.promise;
+  assert.equal(f.forgetCalls(), 0, "parent waits for both run directories");
+
+  f.rows[1].release.resolve();
+  await f.forgotten.promise;
+  assert.deepEqual(f.events, ["child0:start", "child0:settled", "child1:start", "child1:settled", "parent:start", "parent:done"]);
+  assert.deepEqual(f.rows.map((row) => row.calls), [1, 1]);
+  assert.equal(f.forgetCalls(), 1);
+  await assert.rejects(stat(join(f.runs, f.session.id)), { code: "ENOENT" });
+});
+
+test("a launch for a card closed in the same tick never asks its plugins or spawns", { timeout: 5000 }, async (t) => {
+  const root = await mkdtemp(join(tmpdir(), "ctty-cleanup-order-"));
+  const serviceCalls = [];
+  const calls = [];
+  const pipeline = cleanupPipeline(join(root, "runs"), serviceCalls);
+  const forgotten = [];
+  const forget = pipeline.forgetSession.bind(pipeline);
+  pipeline.forgetSession = (...args) => {
+    const pending = forget(...args);
+    forgotten.push(pending);
+    return pending;
+  };
+  const terminals = new TerminalManager(() => undefined, availableRegistry(), undefined, undefined, true, fakeSpawner(calls));
+  terminals.configureLaunchPipeline(pipeline);
+  t.after(async () => {
+    terminals.disposeAll();
+    await rm(root, { recursive: true, force: true });
+  });
+
+  const session = createCard(terminals, root);
+  const internal = terminals.sessions.get(session.id);
+  const tasks = [...internal.launchTasks];
+  terminals.dispose(session.id);
+  await Promise.allSettled(tasks);
+  await Promise.allSettled(forgotten);
+
+  assert.equal(serviceCalls.length, 0, "no plugin prepares a launch for a card that is already closed");
+  assert.equal(calls.length, 0);
+  await assert.rejects(stat(join(root, "runs", session.id)), { code: "ENOENT" });
+});
+
+test("a failed launch-file removal at close is reported with the card id", { timeout: 5000 }, async (t) => {
+  const root = await mkdtemp(join(tmpdir(), "ctty-cleanup-order-"));
+  const calls = [];
+  const pipeline = cleanupPipeline(join(root, "runs"));
+  const failure = new Error("fixture forget failure");
+  const forgotten = deferred();
+  pipeline.forgetSession = async () => {
+    forgotten.resolve();
+    throw failure;
+  };
+  const warnings = [];
+  t.mock.method(console, "warn", (...args) => { warnings.push(args); });
+  const terminals = new TerminalManager(() => undefined, availableRegistry(), undefined, undefined, true, fakeSpawner(calls));
+  terminals.configureLaunchPipeline(pipeline);
+  t.after(async () => {
+    terminals.disposeAll();
+    await rm(root, { recursive: true, force: true });
+  });
+
+  const session = createCard(terminals, root);
+  await Promise.allSettled([...terminals.sessions.get(session.id).launchTasks]);
+  terminals.dispose(session.id);
+  await forgotten.promise;
+  for (let i = 0; i < 20 && !warnings.length; i++) await new Promise((done) => setImmediate(done));
+
+  const warning = warnings.find(([message]) => String(message).includes(session.id));
+  assert.ok(warning, "the close reports which card's launch files stayed behind");
+  assert.ok(warning[1] instanceof AggregateError);
+  assert.deepEqual(warning[1].errors, [failure]);
+});
diff --git a/tests/launch-contributors.test.mjs b/tests/launch-contributors.test.mjs
index df2491a1e..8e47ab77f 100644
--- a/tests/launch-contributors.test.mjs
+++ b/tests/launch-contributors.test.mjs
@@ -310,11 +310,11 @@ test("secret env comes from the plugin's own secrets, reaches the child, and is
   calls[0].print(`env: PROVIDER_API_KEY=${SECRET} PLAIN=visible\n`);
 
   const control = new AgentControlService(manager);
-  const observed = control.observe(created.id).output;
+  const observed = (await control.observe(created.id)).output;
   assert.match(observed, /PROVIDER_API_KEY=<redacted:secret> PLAIN=visible/u);
   assert.doesNotMatch(observed, new RegExp(SECRET, "u"));
   calls[0].exit(1);
-  assert.doesNotMatch(control.result(created.id).output, new RegExp(SECRET, "u"));
+  assert.doesNotMatch((await control.result(created.id)).output, new RegExp(SECRET, "u"));
   assert.doesNotMatch(card(manager, created.id).failureDetails ?? "", new RegExp(SECRET, "u"));
 
   await manager.shutdown();
@@ -430,3 +430,22 @@ test("end to end: the example service prepares a launch over JSON-RPC", async (t
   const empty = await pipeline.prepare(context({ greeting: "" }));
   assert.match(empty.reason, /Launch Env: Value is empty/u);
 });
+
+test("a plugin-started agent drops ambient provider keys, yet the key its launch contribution names explicitly arrives", async (t) => {
+  // Fake values only, set on this test process.
+  const saved = { OPENAI_API_KEY: process.env.OPENAI_API_KEY, OPENAI_ADMIN_KEY: process.env.OPENAI_ADMIN_KEY, ANTHROPIC_API_KEY: process.env.ANTHROPIC_API_KEY };
+  t.after(() => { for (const [key, value] of Object.entries(saved)) { if (value === undefined) delete process.env[key]; else process.env[key] = value; } });
+  Object.assign(process.env, { OPENAI_API_KEY: "fake-oa", OPENAI_ADMIN_KEY: "fake-oa-adm", ANTHROPIC_API_KEY: "fake-an" });
+  const { pipeline } = await pipelineFixture(t, {
+    contributors: [contributor("p.keys", { secrets: true })],
+    answers: { "p.keys": { secretEnv: { OPENAI_API_KEY: "api-key" } } },
+    secrets: { "p.keys/api-key": SECRET }
+  });
+  const { manager, calls } = await managerFixture(t, pipeline);
+  manager.create({ provider: "codex", profile: "normal", cwd, position: at, launchOptions: { "p.keys": {} } }, { ownerPluginId: "p.keys" });
+  await waitFor(() => calls.length === 1);
+  const env = calls[0].options.env;
+  assert.equal(env.OPENAI_API_KEY, SECRET, "the contribution's own key");
+  assert.equal(env.ANTHROPIC_API_KEY, undefined, "no ambient provider key for a delegated agent");
+  assert.equal(env.OPENAI_ADMIN_KEY, undefined, "never an admin key");
+});
diff --git a/tests/launch-file-lifecycle.test.mjs b/tests/launch-file-lifecycle.test.mjs
new file mode 100644
index 000000000..7131f2ff5
--- /dev/null
+++ b/tests/launch-file-lifecycle.test.mjs
@@ -0,0 +1,282 @@
+import assert from "node:assert/strict";
+import { mkdtemp, readdir, rm, stat } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { basename, join } from "node:path";
+import test from "node:test";
+import { EnvironmentRegistry } from "../src/main/services/EnvironmentRegistry.ts";
+import { LaunchPipeline } from "../src/main/services/LaunchPipeline.ts";
+import { TerminalManager } from "../src/main/services/TerminalManager.ts";
+import { persistedTerminalSession } from "../src/main/services/TerminalSessionStore.ts";
+import { availableRegistry, fakeSpawner } from "./helpers/terminal.mjs";
+
+function deferred() {
+  let resolve;
+  const promise = new Promise((done) => { resolve = done; });
+  return { promise, resolve };
+}
+
+async function fixture(t, { spawnFailure = false } = {}) {
+  const root = await mkdtemp(join(tmpdir(), "ctty-launch-files-"));
+  const runs = join(root, "runs");
+  const calls = [];
+  const pipelines = [];
+  const gates = [];
+  const disposals = [];
+  const fake = fakeSpawner(calls);
+  const terminals = new TerminalManager(() => undefined, availableRegistry(), undefined, undefined, true, (...args) => {
+    if (spawnFailure) throw new Error("fixture PTY spawn failure");
+    return fake(...args);
+  });
+  // Closing a card starts its run-file cleanup in the background; the tests await that same promise.
+  const forget = terminals.forgetLaunchFiles.bind(terminals);
+  terminals.forgetLaunchFiles = (session) => {
+    const pending = forget(session);
+    disposals.push(pending);
+    return pending;
+  };
+  const gate = () => {
+    const result = deferred();
+    gates.push(result);
+    return result;
+  };
+
+  function pipeline({ holdPrepare = false, holdForget = false, failCleanup = false } = {}) {
+    const started = deferred();
+    const release = gate();
+    const forgetStarted = deferred();
+    const forgetRelease = gate();
+    const events = [];
+    const rows = [];
+    if (!holdPrepare) release.resolve();
+    if (!holdForget) forgetRelease.resolve();
+    const value = new LaunchPipeline({
+      contributors: () => [{
+        pluginId: "fixture.files",
+        pluginName: "Files",
+        serviceId: "launch",
+        secrets: false,
+        launch: { fields: [{ key: "enabled", label: "Enabled", kind: "boolean" }] }
+      }],
+      call: async () => {
+        started.resolve();
+        await release.promise;
+        return { env: { FIXTURE_FILE: "{launchFiles}/run.txt" }, files: [{ relPath: "run.txt", content: "temporary" }] };
+      },
+      secret: async () => null,
+      runsRoot: runs,
+      timeoutMs: 4000
+    });
+    const prepare = value.prepare.bind(value);
+    const remove = value.forgetSession.bind(value);
+    value.prepare = async (context) => {
+      const result = await prepare(context);
+      assert.equal(result.ok, true);
+      const row = { file: result.env.FIXTURE_FILE, runDirectory: result.runDirectory, cleanupCalls: 0 };
+      const cleanup = result.cleanup;
+      rows.push(row);
+      result.cleanup = async () => {
+        row.cleanupCalls++;
+        events.push("child");
+        if (failCleanup) throw new Error("fixture cleanup failure");
+        await cleanup();
+      };
+      return result;
+    };
+    value.forgetSession = async (id, runDirectories) => {
+      events.push("parent:start");
+      forgetStarted.resolve();
+      await forgetRelease.promise;
+      await remove(id, runDirectories);
+      events.push("parent:done");
+    };
+    const result = { value, started, release, forgetStarted, forgetRelease, events, rows };
+    pipelines.push(result);
+    return result;
+  }
+
+  const create = (extra = {}) => terminals.create({
+    provider: "codex",
+    profile: "normal",
+    cwd: root,
+    position: { x: 0, y: 0 },
+    launchOptions: { "fixture.files": { enabled: true } },
+    ...extra
+  });
+  const drain = async () => { await Promise.allSettled(disposals); };
+  t.after(async () => {
+    for (const item of gates) item.resolve();
+    terminals.disposeAll();
+    await drain();
+    await rm(root, { recursive: true, force: true });
+  });
+  return { root, runs, calls, terminals, pipeline, create, gate, drain, disposals };
+}
+
+for (const failCleanup of [false, true]) {
+  test(`closing during preparation drains its late files before parent deletion (${failCleanup ? "failed" : "successful"} child cleanup)`, { timeout: 5000 }, async (t) => {
+    const f = await fixture(t);
+    const p = f.pipeline({ holdPrepare: true, failCleanup });
+    f.terminals.configureLaunchPipeline(p.value);
+    const session = f.create();
+    await p.started.promise;
+    f.terminals.dispose(session.id);
+    assert.deepEqual(p.events, []);
+
+    p.release.resolve();
+    await f.drain();
+    assert.deepEqual(p.events, ["child", "parent:start", "parent:done"]);
+    assert.equal(p.rows[0].cleanupCalls, 1);
+    assert.equal(f.calls.length, 0);
+    // A failed child cleanup is retried by the parent step, which still removes only this card's own run folder.
+    await assert.rejects(stat(join(f.runs, session.id)), { code: "ENOENT" });
+  });
+}
+
+test("restart tracks both pending preparations and sequentially forgets their original pipelines after reconfiguration", { timeout: 5000 }, async (t) => {
+  const f = await fixture(t);
+  const a = f.pipeline({ holdPrepare: true, holdForget: true });
+  const b = f.pipeline({ holdPrepare: true });
+  const c = f.pipeline();
+  f.terminals.configureLaunchPipeline(a.value);
+  const session = f.create();
+  await a.started.promise;
+  // Force a restart while its original plugin response is still in flight, as restoration can do.
+  const internal = f.terminals.sessions.get(session.id);
+  internal.metadata.status = "failed";
+  internal.metadata.exitCode = 1;
+  f.terminals.configureLaunchPipeline(b.value);
+  f.terminals.restart(session.id);
+  await b.started.promise;
+  f.terminals.configureLaunchPipeline(c.value);
+  f.terminals.dispose(session.id);
+
+  a.release.resolve();
+  await Promise.resolve();
+  assert.deepEqual(a.events, [], "parent cannot precede pending preparation");
+  b.release.resolve();
+  await a.forgetStarted.promise;
+  assert.deepEqual(a.events, ["child", "parent:start"]);
+  assert.deepEqual(b.events, ["child"], "same parent cannot be removed concurrently by another pipeline");
+  assert.deepEqual(c.events, [], "current configuration is not the owner");
+
+  a.forgetRelease.resolve();
+  await f.drain();
+  assert.deepEqual(b.events, ["child", "parent:start", "parent:done"]);
+  assert.deepEqual([a.rows[0].cleanupCalls, b.rows[0].cleanupCalls], [1, 1]);
+  assert.equal(f.calls.length, 0);
+  await assert.rejects(stat(join(f.runs, session.id)), { code: "ENOENT" });
+});
+
+for (const outcome of ["success", "refusal", "throw"]) {
+  test(`close while environment wrapping ${outcome} releases planned capability and files once`, { timeout: 5000 }, async (t) => {
+    const f = await fixture(t);
+    const p = f.pipeline();
+    f.terminals.configureLaunchPipeline(p.value);
+    const wrapping = deferred();
+    const release = f.gate();
+    let plannedCleanups = 0;
+    const plan = f.terminals.planSpawn.bind(f.terminals);
+    f.terminals.planSpawn = (...args) => {
+      const result = plan(...args);
+      assert.ok(!("failure" in result));
+      const cleanup = result.cleanup;
+      result.cleanup = () => {
+        plannedCleanups++;
+        cleanup();
+      };
+      return result;
+    };
+    const registry = new EnvironmentRegistry({
+      providers: () => [{
+        pluginId: "fixture.env",
+        pluginName: "Environment",
+        serviceId: "env",
+        secrets: false,
+        kinds: [{ kind: "test", label: "Test", fields: [] }]
+      }],
+      call: async (_plugin, _service, method) => method.endsWith(".prepare") ? { ref: { id: "one" }, label: "Test" } : {},
+      secret: async () => null
+    });
+    registry.wrap = async (_environment, input) => {
+      wrapping.resolve();
+      await release.promise;
+      if (outcome === "throw") throw new Error("fixture wrapper threw");
+      if (outcome === "refusal") return { ok: false, reason: "fixture refused" };
+      return { ok: true, ...input.launch, secrets: [] };
+    };
+    f.terminals.configureEnvironments(registry);
+    const session = f.create({ environment: { pluginId: "fixture.env", kind: "test" } });
+    await wrapping.promise;
+    await stat(p.rows[0].file);
+    f.terminals.dispose(session.id);
+    assert.deepEqual(p.events, []);
+
+    release.resolve();
+    await f.drain();
+    assert.equal(plannedCleanups, 1);
+    assert.equal(p.rows[0].cleanupCalls, 1);
+    assert.equal(f.calls.length, 0);
+    assert.deepEqual(p.events, ["child", "parent:start", "parent:done"]);
+    await assert.rejects(stat(join(f.runs, session.id)), { code: "ENOENT" });
+  });
+}
+
+test("spawn refusal releases unadopted files and capabilities before eventual parent cleanup", { timeout: 5000 }, async (t) => {
+  const f = await fixture(t, { spawnFailure: true });
+  const p = f.pipeline();
+  f.terminals.configureLaunchPipeline(p.value);
+  let cleanups = 0;
+  const plan = f.terminals.planSpawn.bind(f.terminals);
+  f.terminals.planSpawn = (...args) => {
+    const result = plan(...args);
+    const cleanup = result.cleanup;
+    result.cleanup = () => {
+      cleanups++;
+      cleanup();
+    };
+    return result;
+  };
+  const session = f.create();
+  const internal = f.terminals.sessions.get(session.id);
+  await Promise.all([...internal.launchTasks]);
+  assert.equal(internal.metadata.status, "failed");
+  assert.equal(cleanups, 1);
+  assert.equal(p.rows[0].cleanupCalls, 1);
+  assert.deepEqual(p.events, ["child"]);
+
+  f.terminals.dispose(session.id);
+  await f.drain();
+  assert.deepEqual(p.events, ["child", "parent:start", "parent:done"]);
+  await assert.rejects(stat(join(f.runs, session.id)), { code: "ENOENT" });
+});
+
+test("a delayed close keeps the run files of a card brought back under the same id", { timeout: 5000 }, async (t) => {
+  const f = await fixture(t);
+  const p = f.pipeline({ holdForget: true });
+  f.terminals.configureLaunchPipeline(p.value);
+  const session = f.create();
+  const first = f.terminals.sessions.get(session.id);
+  await Promise.all([...first.launchTasks]);
+  assert.equal(f.calls.length, 1);
+  const record = persistedTerminalSession(first.metadata, first.threadId, first.extras);
+  const firstRun = p.rows[0].runDirectory;
+
+  // Close the card; its folder removal is held, as a slow disk or a busy Windows file would hold it.
+  f.terminals.dispose(session.id);
+  await p.forgetStarted.promise;
+
+  // The same card comes back (restore under its saved id) and prepares a new run before the old removal finishes.
+  await f.terminals.restoreSessionRecords([{ ...record, lastState: "running" }], "continue");
+  const second = f.terminals.sessions.get(session.id);
+  assert.ok(second && second !== first);
+  await Promise.all([...(second.launchTasks ?? [])]);
+  assert.equal(p.rows.length, 2);
+  const secondRun = p.rows[1].runDirectory;
+  assert.notEqual(secondRun, firstRun);
+
+  p.forgetRelease.resolve();
+  await f.drain();
+  await assert.rejects(stat(firstRun), { code: "ENOENT" }, "the closed card's own run folder is removed");
+  await stat(p.rows[1].file);
+  assert.deepEqual(await readdir(join(f.runs, session.id)), [basename(secondRun)], "the newer card's run folder is kept");
+});
diff --git a/tests/launch-readiness.test.mjs b/tests/launch-readiness.test.mjs
index 47768047e..70fafc5be 100644
--- a/tests/launch-readiness.test.mjs
+++ b/tests/launch-readiness.test.mjs
@@ -14,9 +14,9 @@ import { controlRequest } from "../scripts/canvastty-control.mjs";
 import { AgentControlService } from "../src/main/services/AgentControlService.ts";
 import { LaunchPipeline } from "../src/main/services/LaunchPipeline.ts";
 import { PluginSessions } from "../src/main/services/PluginSessions.ts";
-import { TerminalManager } from "../src/main/services/TerminalManager.ts";
+import { INPUT_READY_HOOK_GRACE_MS, TerminalManager } from "../src/main/services/TerminalManager.ts";
 import { AgentControlGateway } from "../src/main/services/agent-control/AgentControlGateway.ts";
-import { ScopedOrchestrationHandler } from "../src/main/services/agent-browser/OrchestrationTools.ts";
+import { AGENT_TOOL_DELIVERY_MS, ScopedOrchestrationHandler } from "../src/main/services/agent-browser/OrchestrationTools.ts";
 
 const at = { x: 0, y: 0 };
 const cwd = process.cwd();
@@ -252,3 +252,236 @@ test("plugin Codex sends wait beyond PTY creation and bracket multiline tasks an
   await sessions.handle("p.owner","svc","sessions.send",{sessionId,text:"draft",submit:false},permissions);
   assert.equal(call.written.at(-1),"\x1b[200~draft\x1b[201~");
 });
+
+// OpenCode drops every key that arrives while its full-screen view starts, so a prompt typed when its PTY exists is
+// lost and the agent sits at an empty prompt. Its prompt goes in only once its input field is up: a lifecycle hook,
+// or the control sequence that shows its text cursor on the alternate screen (never its human-readable screen text).
+
+const ALT_SCREEN_HIDDEN_CURSOR = "\x1b[?1049h\x1b[?2004h\x1b[?25l";
+const CURSOR_SHOWN = "\x1b[1 q\x1b[?25h";
+
+/**
+ * A slow-starting OpenCode stand-in: it draws its start-up for `startupMs` with the cursor hidden and drops every key
+ * written meanwhile (as OpenCode does), then shows its input field. `accepted` holds what reached that field.
+ */
+function slowOpenCode({ startupMs = 120, ready = true } = {}) {
+  const runs = [];
+  const spawn = () => {
+    let emit = () => undefined;
+    const exits = [];
+    const run = { written: [], accepted: [], dropped: [], ready: false, exit: (code) => exits.forEach((listener) => listener({ exitCode: code })) };
+    runs.push(run);
+    setTimeout(() => emit(ALT_SCREEN_HIDDEN_CURSOR), 5);
+    setTimeout(() => emit("  Ask anything… \"Fix broken tests\"   tab agents  ctrl+p commands"), startupMs / 2);
+    if (ready) setTimeout(() => { run.ready = true; emit(CURSOR_SHOWN); }, startupMs);
+    return {
+      pid: 70_000 + runs.length, process: "opencode", resize() {}, kill() {}, pause() {}, resume() {},
+      write(text) { run.written.push(text); (run.ready ? run.accepted : run.dropped).push(text); },
+      onData(listener) { emit = listener; return { dispose() {} }; },
+      onExit(listener) { exits.push(listener); return { dispose() {} }; }
+    };
+  };
+  return { runs, spawn };
+}
+
+function slowOpenCodeOrchestration(t, options) {
+  const cli = slowOpenCode(options);
+  let orchestratorStarted = false;
+  const terminals = new TerminalManager(() => {}, clis(), undefined, undefined, true, (...args) => {
+    if (!orchestratorStarted) { orchestratorStarted = true; return spawner([], { auto: false })(...args); }
+    return cli.spawn(...args);
+  });
+  t.after(() => terminals.disposeAll());
+  const control = new AgentControlService(terminals);
+  const handler = new ScopedOrchestrationHandler(control);
+  const orchestrator = terminals.create({ provider: "codex", profile: "normal", cwd, position: at, role: "orchestrator" });
+  return { cli, terminals, control, execute: (tool, args) => handler.execute(orchestrator.id, { tool, arguments: args }) };
+}
+
+test("spawn_agent: a slow-starting OpenCode gets its first prompt once its input field is up, exactly once", async (t) => {
+  const f = slowOpenCodeOrchestration(t, { startupMs: 150 });
+  const spawned = f.execute("spawn_agent", { provider: "opencode", cwd, prompt: "do part A" });
+  await new Promise((resolve) => setTimeout(resolve, 100));
+  assert.equal(await state(spawned), "pending", "the PTY exists and the home screen text is drawn, but the field is not up");
+  assert.deepEqual(f.cli.runs[0].written, []);
+  const result = await spawned;
+  assert.equal(typeof result.sessionId, "string");
+  assert.deepEqual(f.cli.runs[0].accepted, ["do part A\r"], "the prompt reached the input field");
+  assert.deepEqual(f.cli.runs[0].dropped, [], "nothing was typed while OpenCode started");
+
+  // send_to_agent to the running card goes in at once.
+  assert.deepEqual(await f.execute("send_to_agent", { sessionId: result.sessionId, prompt: "now part B" }), { sessionId: result.sessionId, sent: true });
+  assert.deepEqual(f.cli.runs[0].accepted, ["do part A\r", "now part B\r"]);
+});
+
+test("spawn_agent: an OpenCode that exits before its input field is up fails the call with the reason", async (t) => {
+  const f = slowOpenCodeOrchestration(t, { ready: false });
+  const spawned = f.execute("spawn_agent", { provider: "opencode", cwd, prompt: "never typed" });
+  await tick();
+  f.cli.runs[0].exit(1);
+  await assert.rejects(spawned, (error) => {
+    assert.match(error.bridgeError?.message ?? "", /was not delivered: The session/u);
+    return true;
+  });
+  assert.deepEqual(f.cli.runs[0].written, []);
+});
+
+test("OpenCode that never shows its input field gets no prompt and a bounded reason", async (t) => {
+  const cli = slowOpenCode({ ready: false });
+  const terminals = new TerminalManager(() => {}, clis(), undefined, undefined, true, cli.spawn);
+  t.after(() => terminals.disposeAll());
+  const session = terminals.create({ provider: "opencode", profile: "normal", cwd, position: at });
+  const started = Date.now();
+  const result = await terminals.deliverInput(session.id, "never\r", 200);
+  assert.equal(result.delivered, false);
+  assert.match(result.reason, /did not show its input field within 200 ms; nothing was typed/u);
+  assert.ok(Date.now() - started < 1_000);
+  assert.deepEqual(cli.runs[0].written, [], "its home screen text alone is not readiness");
+});
+
+test("OpenCode readiness: a cursor shown before the alternate screen does not count; one cut across chunks does", async (t) => {
+  const calls = [];
+  const terminals = new TerminalManager(() => {}, clis(), undefined, undefined, true, (command, args, options) => {
+    const pty = spawner(calls, { auto: false })(command, args, options);
+    return { ...pty, onData(listener) { calls.at(-1).emit = listener; return { dispose() {} }; } };
+  });
+  t.after(() => terminals.disposeAll());
+  const session = terminals.create({ provider: "opencode", profile: "normal", cwd, position: at });
+  const pending = terminals.deliverInput(session.id, "split\r", 2_000);
+  calls[0].emit("\x1b[?25h");
+  calls[0].emit("\x1b[?104");
+  calls[0].emit("9h\x1b[?25l drawing");
+  await tick();
+  assert.deepEqual(calls[0].written, []);
+  calls[0].emit("\x1b[?2");
+  await tick();
+  assert.deepEqual(calls[0].written, []);
+  calls[0].emit("5h");
+  assert.deepEqual(await pending, { delivered: true });
+  assert.deepEqual(calls[0].written, ["split\r"]);
+});
+
+test("OpenCode readiness: a lifecycle hook counts only after a grace, and sends made while it starts arrive in order", async (t) => {
+  const calls = [];
+  const terminals = new TerminalManager(() => {}, clis(), undefined, undefined, true, spawner(calls, { auto: false }));
+  t.after(() => terminals.disposeAll());
+  const session = terminals.create({ provider: "opencode", profile: "normal", cwd, position: at });
+  const first = terminals.deliverInput(session.id, "first\r", INPUT_READY_HOOK_GRACE_MS + 2_000);
+  const second = terminals.deliverInput(session.id, "second\r", INPUT_READY_HOOK_GRACE_MS + 2_000);
+  await tick();
+  assert.deepEqual(calls[0].written, []);
+  const hookAt = Date.now();
+  terminals.applyProviderSignal(session.id, { kind: "lifecycle", state: "idle", event: "session.created" });
+  await new Promise((resolve) => setTimeout(resolve, 200));
+  assert.deepEqual(calls[0].written, [], "a hook alone is not the input field");
+  assert.deepEqual([await first, await second], [{ delivered: true }, { delivered: true }]);
+  assert.ok(Date.now() - hookAt >= INPUT_READY_HOOK_GRACE_MS - 20, "typed only after the grace");
+  assert.deepEqual(calls[0].written, ["first\r", "second\r"]);
+});
+
+test("OpenCode readiness: a resumed card whose plugin reports before its input field is up gets the text on the cursor signal", async (t) => {
+  const calls = [];
+  const terminals = new TerminalManager(() => {}, clis(), undefined, undefined, true, (command, args, options) => {
+    const pty = spawner(calls, { auto: false })(command, args, options);
+    return { ...pty, onData(listener) { calls.at(-1).emit = listener; return { dispose() {} }; } };
+  });
+  t.after(() => terminals.disposeAll());
+  const session = terminals.create({ provider: "opencode", profile: "normal", cwd, position: at });
+  calls[0].emit("\x1b[?1049h\x1b[?25l");
+  // The plugin reports idle on session.updated about 0.6 s before typed keys reach the prompt.
+  terminals.applyProviderSignal(session.id, { kind: "lifecycle", state: "idle", event: "session.updated" });
+  const sent = terminals.deliverInput(session.id, "resumed\r", 5_000);
+  await new Promise((resolve) => setTimeout(resolve, 300));
+  assert.deepEqual(calls[0].written, [], "not typed on the hook");
+  const shownAt = Date.now();
+  calls[0].emit("\x1b[?25h");
+  assert.deepEqual(await sent, { delivered: true });
+  assert.ok(Date.now() - shownAt < INPUT_READY_HOOK_GRACE_MS / 2, "typed on the cursor signal, not after the grace");
+  assert.deepEqual(calls[0].written, ["resumed\r"]);
+});
+
+test("OpenCode readiness: a send cancelled while it waits behind another settles at once and is never typed", async (t) => {
+  const calls = [];
+  const terminals = new TerminalManager(() => {}, clis(), undefined, undefined, true, (command, args, options) => {
+    const pty = spawner(calls, { auto: false })(command, args, options);
+    return { ...pty, onData(listener) { calls.at(-1).emit = listener; return { dispose() {} }; } };
+  });
+  t.after(() => terminals.disposeAll());
+  const session = terminals.create({ provider: "opencode", profile: "normal", cwd, position: at });
+  const first = terminals.deliverInput(session.id, "first\r", 5_000);
+  const controller = new AbortController();
+  const queued = terminals.deliverInput(session.id, "queued\r", 5_000, controller.signal);
+  const third = terminals.deliverInput(session.id, "third\r", 5_000);
+  await tick();
+  const abortedAt = Date.now();
+  controller.abort();
+  assert.deepEqual(await queued, { delivered: false, reason: "The delivery was cancelled." });
+  assert.ok(Date.now() - abortedAt < 200, "settled without waiting for the send ahead of it");
+  calls[0].emit("\x1b[?1049h\x1b[?25h");
+  assert.deepEqual([await first, await third], [{ delivered: true }, { delivered: true }]);
+  assert.deepEqual(calls[0].written, ["first\r", "third\r"]);
+});
+
+test("OpenCode readiness is not awaited on Windows, whose pseudo console may not pass the cursor sequences", async (t) => {
+  const platform = Object.getOwnPropertyDescriptor(process, "platform");
+  const calls = [];
+  const terminals = new TerminalManager(() => {}, clis(), undefined, undefined, true, spawner(calls, { auto: false }));
+  t.after(() => terminals.disposeAll());
+  const session = terminals.create({ provider: "opencode", profile: "normal", cwd, position: at });
+  Object.defineProperty(process, "platform", { ...platform, value: "win32" });
+  try {
+    assert.deepEqual(await terminals.deliverInput(session.id, "windows\r", 200), { delivered: true });
+  } finally {
+    Object.defineProperty(process, "platform", platform);
+  }
+  assert.deepEqual(calls[0].written, ["windows\r"]);
+});
+
+test("spawn_agent and send_to_agent wait for delivery within a budget under the MCP client's deadline", async (t) => {
+  assert.ok(AGENT_TOOL_DELIVERY_MS <= 45_000);
+  const f = slowOpenCodeOrchestration(t, { ready: false });
+  const waits = [];
+  const { control } = f;
+  const spawn = control.spawn.bind(control);
+  control.spawn = (request, signal, promptWaitMs) => { waits.push(["spawn", promptWaitMs]); return spawn(request, signal, 150); };
+  const send = control.send.bind(control);
+  control.send = (id, text, submit, signal, waitMs) => { waits.push(["send", waitMs]); return send(id, text, submit, signal, 150); };
+  await assert.rejects(f.execute("spawn_agent", { provider: "opencode", cwd, prompt: "late" }), (error) => {
+    assert.match(error.bridgeError?.message ?? "", /did not show its input field within 150 ms; nothing was typed/u);
+    return true;
+  });
+  const [child] = f.terminals.listMetadata().filter((session) => session.role === "subagent");
+  await assert.rejects(f.execute("send_to_agent", { sessionId: child.id, prompt: "again" }));
+  assert.equal(waits[0][0], "spawn");
+  assert.ok(waits[0][1] > 0 && waits[0][1] <= AGENT_TOOL_DELIVERY_MS, String(waits[0][1]));
+  assert.deepEqual(waits[1], ["send", AGENT_TOOL_DELIVERY_MS]);
+});
+
+test("OpenCode readiness: cancellation and a restart while it starts never deliver the queued text", async (t) => {
+  const cli = slowOpenCode({ startupMs: 80 });
+  const terminals = new TerminalManager(() => {}, clis(), undefined, undefined, true, cli.spawn);
+  t.after(() => terminals.disposeAll());
+  const session = terminals.create({ provider: "opencode", profile: "normal", cwd, position: at });
+  const controller = new AbortController();
+  const canceled = terminals.deliverInput(session.id, "cancel\r", 2_000, controller.signal);
+  const replaced = terminals.deliverInput(session.id, "old launch\r", 2_000);
+  controller.abort();
+  assert.deepEqual(await canceled, { delivered: false, reason: "The delivery was cancelled." });
+  cli.runs[0].exit(0);
+  terminals.restart(session.id);
+  assert.equal((await replaced).delivered, false);
+  await new Promise((resolve) => setTimeout(resolve, 150));
+  assert.ok(cli.runs[1].ready, "the new launch became ready");
+  assert.ok(cli.runs.every((run) => run.written.length === 0), "neither launch got the text");
+});
+
+test("Claude and Codex still get input as soon as their PTY exists", async (t) => {
+  const calls = [];
+  const terminals = new TerminalManager(() => {}, clis(), undefined, undefined, true, spawner(calls, { auto: false }));
+  t.after(() => terminals.disposeAll());
+  for (const provider of ["claude", "codex"]) {
+    const session = terminals.create({ provider, profile: "normal", cwd, position: at });
+    assert.deepEqual(await terminals.deliverInput(session.id, `${provider} task\r`, 200), { delivered: true });
+  }
+  assert.deepEqual(calls.map((call) => call.written), [["claude task\r"], ["codex task\r"]]);
+});
diff --git a/tests/material-handoff-service.test.mjs b/tests/material-handoff-service.test.mjs
index f4797f486..8d28c0b50 100644
--- a/tests/material-handoff-service.test.mjs
+++ b/tests/material-handoff-service.test.mjs
@@ -1034,3 +1034,19 @@ test("a narrow ordinary Codex composer still receives a material handoff", async
     assert.ok(terminal.writes.length > 0);
   });
 });
+
+test("a handoff waits for the agent's CLI to take input instead of failing at once", async () => {
+  await withHandoffs(async ({ handoffs, terminal, draft }) => {
+    terminal.add({ provider: "claude", id: "s1" });
+    terminal.tui = claudeTui(terminal);
+    const waits = [];
+    const deliverInput = terminal.deliverInput.bind(terminal);
+    terminal.deliverInput = async (id, data, waitMs) => {
+      waits.push(waitMs);
+      return deliverInput(id, data);
+    };
+    assert.equal((await handoffs.send(draft("s1"))).handoff.delivery.state, "submitted");
+    assert.ok(waits.length >= 2, "the paste and its Enter");
+    assert.ok(waits.every((waitMs) => waitMs >= 10_000), `waits: ${waits.join(", ")}`);
+  });
+});
diff --git a/tests/opencode-result.test.mjs b/tests/opencode-result.test.mjs
index 1355f842a..a6f77225e 100644
--- a/tests/opencode-result.test.mjs
+++ b/tests/opencode-result.test.mjs
@@ -7,10 +7,12 @@ import test from "node:test";
 import { AGENT_RUNTIME_ENV, CAPTURE_RESULT_ENV, MAX_RESULT_CHARS } from "../src/agent-runtime/runtime-protocol.mjs";
 import { finalAnswer } from "../src/agent-runtime/opencode-final-answer.mjs";
 import { RuntimeGateway } from "../src/main/services/agent-runtime/RuntimeGateway.ts";
+import { applyRuntimeSignal } from "../src/main/services/agent-runtime/applyRuntimeSignal.ts";
+import { LaunchPipeline } from "../src/main/services/LaunchPipeline.ts";
 import { TerminalManager } from "../src/main/services/TerminalManager.ts";
 import { AgentControlService } from "../src/main/services/AgentControlService.ts";
 import { ScopedOrchestrationHandler } from "../src/main/services/agent-browser/OrchestrationTools.ts";
-import { availableRegistry, fakeSpawner } from "./helpers/terminal.mjs";
+import { availableRegistry, fakeSpawner, inputReadySpawner, openCodeInputReady } from "./helpers/terminal.mjs";
 
 const POSIX = { skip: process.platform === "win32" ? "Unix socket test; the Windows pipe transport has its own suite." : false };
 
@@ -55,9 +57,9 @@ await hooks.event({ event: { type: "session.status", properties: { sessionID: "s
 await hooks.event({ event: { type: "session.idle", properties: { sessionID: "ses_root" } } });
 `;
 
-function runFakeOpenCode(capability, answer, captureResult) {
+function runFakeOpenCode(capability, answer, captureResult, script = FAKE_OPENCODE, extraEnv = {}, onCapture = null) {
   const env = {
-    ...process.env,
+    ...process.env, ...extraEnv,
     PLUGIN_URL: new URL("../src/agent-runtime/opencode-plugin.mjs", import.meta.url).href,
     FAKE_ANSWER: JSON.stringify(answer),
     [AGENT_RUNTIME_ENV.address]: capability.address,
@@ -67,35 +69,43 @@ function runFakeOpenCode(capability, answer, captureResult) {
   };
   if (captureResult) env[CAPTURE_RESULT_ENV] = "1";
   else delete env[CAPTURE_RESULT_ENV];
-  const child = spawn(process.execPath, ["--input-type=module", "-e", FAKE_OPENCODE], { env, stdio: ["ignore", "ignore", "pipe"] });
+  const child = spawn(process.execPath, ["--input-type=module", "-e", script], { env, stdio: ["pipe", "pipe", "pipe"] });
+  const timeout = setTimeout(() => child.kill(), 5_000);
   let stderr = "";
   child.stderr.on("data", (chunk) => { stderr += chunk; });
-  return new Promise((resolve) => child.on("close", (code) => resolve({ code, stderr })));
+  return new Promise((resolve, reject) => {
+    // The script writes to stdout when its SDK read starts and waits on stdin until onCapture has run.
+    if (onCapture) child.stdout.once("data", () => {
+      Promise.resolve().then(onCapture).then(() => child.stdin.end("continue"), (error) => { child.kill(); reject(error); });
+    });
+    child.on("error", (error) => { clearTimeout(timeout); reject(error); });
+    child.on("close", (code) => { clearTimeout(timeout); resolve({ code, stderr }); });
+  });
 }
 
-async function setup(t) {
+async function setupTerminals(t) {
   const root = await realpath(await mkdtemp(join(tmpdir(), "ctty-ocr-")));
   t.after(() => rm(root, { recursive: true, force: true }));
   const calls = [];
-  const terminals = new TerminalManager(() => undefined, availableRegistry(), undefined, undefined, true, fakeSpawner(calls));
+  const terminals = new TerminalManager(() => undefined, availableRegistry(), undefined, undefined, true, inputReadySpawner(calls));
   t.after(() => terminals.disposeAll());
-  const runtimeDirectory = await mkdtemp(join(tmpdir(), "ctty-rt-"));
-  t.after(() => rm(runtimeDirectory, { recursive: true, force: true }));
-  // What main/index.ts does with each runtime signal.
-  const gateway = new RuntimeGateway({
-    runtimeDirectory,
-    onSignal: (id, signal) => {
-      terminals.applyProviderSignal(id, { kind: "lifecycle", state: signal.state });
-      if (signal.result) terminals.recordAnswer(id, signal.result);
-    }
-  });
-  await gateway.start();
-  t.after(() => gateway.close());
   const orchestrator = terminals.create({ provider: "codex", cwd: root, profile: "normal", position: { x: 0, y: 0 }, role: "orchestrator" });
   const control = new AgentControlService(terminals, { waitTiming: { checkMs: 5, settleMs: 300, quietMs: 2_000 } });
   const handler = new ScopedOrchestrationHandler(control);
   const child = await handler.execute(orchestrator.id, { id: "1", tool: "spawn_agent", arguments: { provider: "opencode", cwd: root } });
-  return { root, calls, terminals, gateway, handler, orchestrator, childId: child.sessionId };
+  return { root, calls, terminals, control, handler, orchestrator, childId: child.sessionId };
+}
+
+async function setup(t) {
+  const fixture = await setupTerminals(t);
+  const { terminals } = fixture;
+  const runtimeDirectory = await mkdtemp(join(tmpdir(), "ctty-rt-"));
+  t.after(() => rm(runtimeDirectory, { recursive: true, force: true }));
+  // The same glue main/index.ts runs for each runtime signal.
+  const gateway = new RuntimeGateway({ runtimeDirectory, onSignal: (id, signal) => applyRuntimeSignal(terminals, id, signal) });
+  await gateway.start();
+  t.after(() => gateway.close());
+  return { ...fixture, gateway };
 }
 
 test("an OpenCode subagent's final reply reaches wait_for_agent and get_agent_result, masked", POSIX, async (t) => {
@@ -147,11 +157,14 @@ test("an ordinary OpenCode card (not a subagent) never keeps an answer", async (
 test("wait_for_agent ignores an idle before the prompt's turn and returns that turn's answer", async (t) => {
   const root = await realpath(await mkdtemp(join(tmpdir(), "ctty-ocr-")));
   t.after(() => rm(root, { recursive: true, force: true }));
-  const terminals = new TerminalManager(() => undefined, availableRegistry(), undefined, undefined, true, fakeSpawner([]));
+  const calls = [];
+  const terminals = new TerminalManager(() => undefined, availableRegistry(), undefined, undefined, true, fakeSpawner(calls));
   t.after(() => terminals.disposeAll());
   const orchestrator = terminals.create({ provider: "codex", cwd: root, profile: "normal", position: { x: 0, y: 0 }, role: "orchestrator" });
   const handler = new ScopedOrchestrationHandler(new AgentControlService(terminals, { waitTiming: { checkMs: 5, settleMs: 10, quietMs: 60_000 } }));
-  const child = await handler.execute(orchestrator.id, { id: "1", tool: "spawn_agent", arguments: { provider: "opencode", cwd: root, prompt: "do part A" } });
+  const spawning = handler.execute(orchestrator.id, { id: "1", tool: "spawn_agent", arguments: { provider: "opencode", cwd: root, prompt: "do part A" } });
+  await openCodeInputReady(calls, 1);
+  const child = await spawning;
   const wait = (timeoutSeconds) => handler.execute(orchestrator.id, { id: "w", tool: "wait_for_agent", arguments: { sessionId: child.sessionId, timeoutSeconds } });
   // OpenCode's session.created reports idle before the prompt's turn starts.
   terminals.applyProviderSignal(child.sessionId, { kind: "lifecycle", state: "idle", event: "session.created" });
@@ -167,3 +180,318 @@ test("wait_for_agent ignores an idle before the prompt's turn and returns that t
   assert.equal(done.reason, "idle");
   assert.deepEqual(done.answer, { text: "part A done", truncated: false });
 });
+
+test("OpenCode turn correlation rejects preceding answers and continuations across submitted prompts", async (t) => {
+  const { terminals, childId } = await setupTerminals(t);
+  const apply = (state, event, turnId, answer) => {
+    const current = terminals.applyProviderSignal(childId, { kind: "lifecycle", state, event, requestId: turnId });
+    if (current && answer) terminals.recordAnswer(childId, { text: answer, truncated: false }, { turnId });
+    return current;
+  };
+  assert.equal(apply("working", "session.status:busy", "turn1"), true);
+  assert.equal(apply("idle", "session.idle", "turn1", "old answer"), true);
+  assert.equal(terminals.answer(childId).text, "old answer");
+  assert.deepEqual(await terminals.deliverInput(childId, "new task\r"), { delivered: true });
+  assert.equal(terminals.answer(childId), null, "a submitted prompt drops the previous answer");
+  assert.equal(apply("idle", "session.idle", "turn1", "delayed old answer"), false);
+  assert.equal(apply("working", "permission.replied", "turn1"), false, "an old continuation cannot acknowledge the new prompt");
+  assert.equal(terminals.getMetadata(childId).status, "working", "but it still moves the status of the turn that runs");
+  terminals.recordAnswer(childId, { text: "uncorrelated old answer", truncated: false });
+  assert.equal(terminals.answer(childId), null);
+  assert.equal(terminals.turnProgress(childId).turnStartedSincePrompt, false);
+  assert.equal(apply("working", "session.status:busy", "turn2"), true);
+  assert.equal(apply("idle", "session.idle", "turn1", "late old after new start"), false);
+  assert.equal(terminals.getMetadata(childId).status, "working", "a late signal of an earlier turn changes nothing");
+  assert.equal(apply("needs_approval", "permission.asked", "turn2"), true);
+  const generation = terminals.sessions.get(childId).inputGeneration ?? 0;
+  terminals.input(childId, "y\r");
+  assert.equal(terminals.sessions.get(childId).inputGeneration ?? 0, generation, "an approval answer is not a new task");
+  assert.equal(apply("working", "permission.replied", "turn2"), true);
+  assert.equal(apply("working", "session.status:retry", "turn2"), true);
+  assert.equal(apply("idle", "session.idle", "turn2", "new answer"), true);
+  assert.equal(terminals.answer(childId).text, "new answer");
+  apply("working", "session.status:busy", "turn3");
+  await terminals.deliverInput(childId, "interrupt current work\r");
+  const starts = terminals.sessions.get(childId).turnStarts;
+  assert.equal(apply("working", "session.status:busy", "turn4"), true);
+  assert.equal(terminals.sessions.get(childId).turnStarts, starts + 1, "a new turn id is a new turn even while the previous one was working");
+  assert.equal(terminals.turnProgress(childId).turnStartedSincePrompt, true);
+  assert.equal(apply("idle", "session.idle", "turn3", "interrupted answer"), false);
+  assert.equal(terminals.answer(childId), null);
+});
+
+test("with status hooks off the plugin reports nothing and the wait says the answer cannot be captured", POSIX, async (t) => {
+  const { terminals, gateway, handler, orchestrator, childId } = await setup(t);
+  terminals.setLifecycleHooksEnabled(false);
+  const signals = [];
+  const apply = terminals.applyProviderSignal.bind(terminals);
+  terminals.applyProviderSignal = (id, signal, ...args) => { signals.push(signal); return apply(id, signal, ...args); };
+  // A lease that exists for another reason (the decision hook) still carries no lifecycle report and no answer.
+  const capability = gateway.registerSession(childId, "opencode", true);
+  const run = await runFakeOpenCode(capability, "never captured", true, FAKE_OPENCODE, { CANVASTTY_LIFECYCLE_HOOKS_ENABLED: "0" });
+  assert.equal(run.code, 0, run.stderr);
+  assert.deepEqual(signals, []);
+  assert.equal(terminals.getMetadata(childId).status, "unavailable");
+  assert.equal(terminals.answer(childId), null);
+  const waited = await handler.execute(orchestrator.id, { id: "w", tool: "wait_for_agent", arguments: { sessionId: childId, timeoutSeconds: 5 } });
+  assert.equal(waited.reason, "quiet");
+  assert.equal(waited.answer, undefined);
+  assert.equal(waited.answerUnavailable.reason, "status_hooks_off");
+  assert.match(waited.answerUnavailable.message, /status hooks are off[\s\S]*Read its output/u);
+  const result = await handler.execute(orchestrator.id, { id: "r", tool: "get_agent_result", arguments: { sessionId: childId } });
+  assert.equal(result.answer, undefined);
+  assert.deepEqual(result.answerUnavailable, waited.answerUnavailable);
+  // Turning them on again applies to agents started or restarted afterwards, not to this launch.
+  terminals.setLifecycleHooksEnabled(true);
+  assert.equal(terminals.applyProviderSignal(childId, { kind: "lifecycle", state: "idle", event: "session.idle", requestId: "t" }), false);
+  const later = await handler.execute(orchestrator.id, { id: "r2", tool: "get_agent_result", arguments: { sessionId: childId } });
+  assert.equal(later.answerUnavailable?.reason, "status_hooks_off");
+});
+
+test("OpenCode async final answer keeps its ending turn and drops a read superseded by a new turn", POSIX, async (t) => {
+  const { terminals, gateway, childId } = await setup(t);
+  const capability = gateway.registerSession(childId, "opencode", true);
+  const signals = [];
+  const original = terminals.applyProviderSignal.bind(terminals);
+  terminals.applyProviderSignal = (id, signal, ...args) => { signals.push(signal); return original(id, signal, ...args); };
+  const script = `
+    const { CanvasTTYLifecycle } = await import(process.env.PLUGIN_URL);
+    let resolveOld, started; const capturing = new Promise((done) => started = done);
+    let reads = 0;
+    const client = { session: { messages: async () => {
+      if (reads++ === 0) { started(); return new Promise((done) => resolveOld = done); }
+      return { data: [{ info: { role: "assistant" }, parts: [{ type: "text", text: "second answer" }] }] };
+    } } };
+    const hooks = await CanvasTTYLifecycle({ client });
+    const event = (type, rest = {}) => hooks.event({ event: { type, properties: { sessionID: "root", ...rest } } });
+    await event("session.created", { info: { id: "root" } });
+    await event("session.status", { status: { type: "busy" } });
+    await event("session.status", { status: { type: "idle" } });
+    const old = event("session.idle"); await capturing;
+    await event("session.status", { status: { type: "busy" } });
+    resolveOld({ data: [{ info: { role: "assistant" }, parts: [{ type: "text", text: "old delayed answer" }] }] });
+    await old;
+    await event("permission.asked"); await event("permission.replied");
+    await event("session.status", { status: { type: "retry" } });
+    await event("session.status", { status: { type: "idle" } });
+    await event("session.idle");
+  `;
+  const run = await runFakeOpenCode(capability, "unused", true, script);
+  assert.equal(run.code, 0, run.stderr);
+  const starts = signals.filter((signal) => signal.event === "session.status:busy");
+  assert.equal(starts.length, 2);
+  assert.notEqual(starts[0].requestId, starts[1].requestId, "each turn has its own id");
+  assert.equal(signals.filter((signal) => signal.event === "session.status:idle").length, 0, "no idle before the answer was read");
+  const ends = signals.filter((signal) => signal.event === "session.idle");
+  assert.equal(ends.length, 1, "a superseded answer read reports neither idle nor an answer");
+  assert.equal(ends[0].requestId, starts[1].requestId);
+  assert.equal(signals.find((signal) => signal.event === "permission.replied").requestId, starts[1].requestId);
+  assert.equal(signals.find((signal) => signal.event === "session.status:retry").requestId, starts[1].requestId);
+  assert.equal(terminals.answer(childId).text, "second answer");
+});
+
+test("an answer read that finishes after the next prompt was submitted is not that prompt's answer", POSIX, async (t) => {
+  const { terminals, gateway, handler, orchestrator, childId } = await setup(t);
+  const capability = gateway.registerSession(childId, "opencode", true);
+  const script = `
+    const { CanvasTTYLifecycle } = await import(process.env.PLUGIN_URL);
+    const client = { session: { messages: async () => {
+      process.stdout.write("capturing");
+      await new Promise((done) => process.stdin.once("data", done));
+      return { data: [{ info: { role: "assistant" }, parts: [{ type: "text", text: "late preceding answer" }] }] };
+    } } };
+    const hooks = await CanvasTTYLifecycle({ client });
+    const event = (type, rest = {}) => hooks.event({ event: { type, properties: { sessionID: "root", ...rest } } });
+    await event("session.created", { info: { id: "root" } });
+    await event("session.status", { status: { type: "busy" } });
+    await event("session.idle");
+  `;
+  const run = await runFakeOpenCode(capability, "unused", true, script, {}, () => terminals.deliverInput(childId, "next task\r"));
+  assert.equal(run.code, 0, run.stderr);
+  const result = await handler.execute(orchestrator.id, { id: "poll", tool: "get_agent_result", arguments: { sessionId: childId } });
+  assert.equal(result.answer, undefined);
+  assert.equal(terminals.turnProgress(childId).turnStartedSincePrompt, false, "the late idle does not acknowledge the next prompt");
+});
+
+test("approval continuations and repeated working events stay in the same turn", async (t) => {
+  const { terminals, childId } = await setupTerminals(t);
+  const apply = (state, event, turnId) => terminals.applyProviderSignal(childId, { kind: "lifecycle", state, event, requestId: turnId });
+  assert.equal(apply("working", "session.status:busy", "turn1"), true);
+  const starts = terminals.sessions.get(childId).turnStarts;
+  assert.equal(apply("working", "session.status:busy", "turn1"), true);
+  assert.equal(apply("working", "session.status:retry", "turn1"), true);
+  assert.equal(terminals.sessions.get(childId).turnStarts, starts, "a repeated working or retry is not another start");
+  assert.equal(apply("needs_approval", "permission.asked", "turn1"), true);
+  assert.equal(terminals.getMetadata(childId).status, "needs_approval");
+  const generation = terminals.sessions.get(childId).inputGeneration ?? 0;
+  await terminals.deliverInput(childId, "y\r");
+  assert.equal(terminals.sessions.get(childId).inputGeneration ?? 0, generation, "an approval answer keeps the current turn");
+  assert.equal(apply("working", "permission.replied", "turn1"), true);
+  assert.equal(terminals.turnProgress(childId).turnStartedSincePrompt, true, "the continuation acknowledges the approval");
+  const continued = terminals.sessions.get(childId).turnStarts;
+  assert.equal(apply("working", "session.status:retry", "turn1"), true);
+  assert.equal(terminals.sessions.get(childId).turnStarts, continued);
+  assert.equal(apply("idle", "session.idle", "turn1"), true);
+  terminals.recordAnswer(childId, { text: "answer after approval", truncated: false }, { turnId: "turn1" });
+  assert.equal(terminals.answer(childId).text, "answer after approval");
+  assert.equal(terminals.getMetadata(childId).status, "idle");
+  terminals.input(childId, "next task\r");
+  assert.equal(terminals.sessions.get(childId).inputGeneration, generation + 1);
+  assert.equal(apply("working", "permission.replied", "turn1"), false);
+});
+
+test("a result-capturing OpenCode does not report idle while its final answer is still being read", POSIX, async (t) => {
+  const { terminals, gateway, handler, orchestrator, childId } = await setup(t);
+  const capability = gateway.registerSession(childId, "opencode", true);
+  const script = `
+    const { CanvasTTYLifecycle } = await import(process.env.PLUGIN_URL);
+    const client = { session: { messages: async () => {
+      process.stdout.write("capturing"); await new Promise((done) => process.stdin.once("data", done));
+      return { data: [{ info: { role: "assistant" }, parts: [{ type: "text", text: "captured after delay" }] }] };
+    } } };
+    const hooks = await CanvasTTYLifecycle({ client });
+    const event = (type, rest = {}) => hooks.event({ event: { type, properties: { sessionID: "root", ...rest } } });
+    await event("session.created", { info: { id: "root" } });
+    await event("session.status", { status: { type: "busy" } });
+    await event("session.status", { status: { type: "idle" } });
+    await event("session.idle");
+  `;
+  const run = await runFakeOpenCode(capability, "unused", true, script, {}, () => {
+    assert.equal(terminals.getMetadata(childId).status, "working");
+    assert.equal(terminals.answer(childId), null);
+  });
+  assert.equal(run.code, 0, run.stderr);
+  const waited = await handler.execute(orchestrator.id, { id: "w", tool: "wait_for_agent", arguments: { sessionId: childId, timeoutSeconds: 5 } });
+  assert.equal(waited.reason, "idle");
+  assert.equal(waited.answer.text, "captured after delay");
+});
+
+for (const failure of ["error", "timeout"]) test(`a captured OpenCode turn still ends, without an answer, after an SDK read ${failure}`, POSIX, async (t) => {
+  const { terminals, gateway, childId } = await setup(t);
+  const accepted = [];
+  const apply = terminals.applyProviderSignal.bind(terminals);
+  terminals.applyProviderSignal = (id, signal, ...args) => { accepted.push(signal); return apply(id, signal, ...args); };
+  const capability = gateway.registerSession(childId, "opencode", true);
+  const script = `
+    const { CanvasTTYLifecycle } = await import(process.env.PLUGIN_URL);
+    const client = { session: { messages: async () => { ${failure === "error" ? 'throw new Error("fixture failed");' : "return new Promise(() => {});"} } } };
+    const hooks = await CanvasTTYLifecycle({ client });
+    const event = (type, rest = {}) => hooks.event({ event: { type, properties: { sessionID: "root", ...rest } } });
+    await event("session.created", { info: { id: "root" } }); await event("session.status", { status: { type: "busy" } });
+    await event("session.status", { status: { type: "idle" } }); await event("session.idle");
+  `;
+  const run = await runFakeOpenCode(capability, "unused", true, script);
+  assert.equal(run.code, 0, run.stderr);
+  assert.equal(accepted.filter((signal) => signal.event === "session.status:idle").length, 0);
+  assert.equal(accepted.filter((signal) => signal.event === "session.idle").length, 1, "the turn end is still reported");
+  assert.equal(terminals.answer(childId), null);
+  assert.equal(terminals.getMetadata(childId).status, "idle");
+});
+
+/** Feeds one runtime signal through the glue main uses, as the gateway would hand it over. */
+const signalFrom = (terminals, id) => (state, event, turnId, answer) => applyRuntimeSignal(terminals, id,
+  { state, event, turnId, ...(answer ? { result: { text: answer, truncated: false } } : {}) });
+
+test("a person's Enter in the middle of a turn still lets that turn end the wait, without an answer", async (t) => {
+  const { terminals, handler, orchestrator, childId } = await setupTerminals(t);
+  const signal = signalFrom(terminals, childId);
+  assert.deepEqual(await terminals.deliverInput(childId, "orchestrator task\r"), { delivered: true });
+  signal("working", "session.status:busy", "turn1");
+  // The person presses Enter in the subagent's card while it works: the turn's answer is no longer for that prompt.
+  terminals.input(childId, "\r");
+  const waiting = handler.execute(orchestrator.id, { id: "w", tool: "wait_for_agent", arguments: { sessionId: childId, timeoutSeconds: 5 } });
+  setTimeout(() => signal("idle", "session.idle", "turn1", "answer to the earlier prompt"), 20);
+  const waited = await waiting;
+  assert.equal(waited.reason, "idle", "the wait does not hang until its timeout");
+  assert.equal(waited.status, "idle");
+  assert.equal(waited.answer, undefined);
+  assert.equal(waited.answerUnavailable, undefined, "status hooks are on: nothing says the answer cannot be captured");
+});
+
+test("an answer that arrives as the agent exits is kept when it belongs to the current turn", async (t) => {
+  const { calls, terminals, handler, orchestrator, childId } = await setupTerminals(t);
+  const signal = signalFrom(terminals, childId);
+  await terminals.deliverInput(childId, "last task\r");
+  signal("working", "session.status:busy", "turn1");
+  calls.at(-1).process.emitExit(0);
+  assert.equal(terminals.getMetadata(childId).status, "done");
+  signal("idle", "session.idle", "turn1", "final words");
+  assert.equal(terminals.getMetadata(childId).status, "done", "the exit stays");
+  const result = await handler.execute(orchestrator.id, { id: "r", tool: "get_agent_result", arguments: { sessionId: childId } });
+  assert.equal(result.state, "done");
+  assert.equal(result.answer?.text, "final words");
+  // A late answer of an earlier turn still does not count after the exit.
+  signal("idle", "session.idle", "turn0", "stale");
+  assert.equal(terminals.answer(childId)?.text, "final words");
+});
+
+test("an Enter with nothing typed in an idle card keeps the stored answer", async (t) => {
+  const { terminals, childId } = await setupTerminals(t);
+  const signal = signalFrom(terminals, childId);
+  await terminals.deliverInput(childId, "task\r");
+  signal("working", "session.status:busy", "turn1");
+  signal("idle", "session.idle", "turn1", "kept answer");
+  const generation = terminals.sessions.get(childId).inputGeneration;
+  terminals.input(childId, "\r");
+  assert.equal(terminals.answer(childId)?.text, "kept answer", "an empty Enter submits nothing");
+  assert.equal(terminals.sessions.get(childId).inputGeneration, generation);
+  // Typed key by key, as a person types: the Enter after it submits a new task.
+  for (const key of ["n", "e", "x", "t"]) terminals.input(childId, key);
+  terminals.input(childId, "\r");
+  assert.equal(terminals.answer(childId), null);
+  assert.equal(terminals.sessions.get(childId).inputGeneration, generation + 1);
+});
+
+test("every input write goes through writeInput, which keeps the turn and answer bookkeeping", async (t) => {
+  const { terminals, childId } = await setupTerminals(t);
+  const signal = signalFrom(terminals, childId);
+  const writes = [];
+  const writeInput = terminals.writeInput;
+  assert.equal(typeof writeInput, "function", "one write path for person and agent input");
+  terminals.writeInput = function (...args) { writes.push(args[3]); return writeInput.apply(this, args); };
+  await terminals.deliverInput(childId, "task\r");
+  terminals.input(childId, "x");
+  assert.deepEqual(writes, ["task\r", "x"], "input() and deliverInput write through it");
+  terminals.writeInput = writeInput;
+  signal("working", "session.status:busy", "turn1");
+  signal("idle", "session.idle", "turn1", "kept answer");
+  // A writer that held input and replays it later (outside input()) uses the same method: its Enter submits a task.
+  const session = terminals.sessions.get(childId);
+  const generation = session.inputGeneration;
+  for (const key of ["n", "e", "x", "t", "\r"]) assert.equal(terminals.writeInput(childId, session, session.process, key), true);
+  assert.equal(terminals.answer(childId), null, "the replayed prompt drops the earlier answer");
+  assert.equal(session.inputGeneration, generation + 1);
+  // Never into a process the card no longer runs.
+  assert.equal(terminals.writeInput(childId, session, { write: () => assert.fail("stale process") }, "late\r"), false);
+  assert.equal(session.inputGeneration, generation + 1);
+});
+
+for (const [atCreate, beforeSpawn] of [[false, true], [true, false]]) {
+  test(`a deferred launch takes status hooks as they are when it is planned (${atCreate ? "on" : "off"} at create, ${beforeSpawn ? "on" : "off"} at launch)`, async (t) => {
+    const root = await realpath(await mkdtemp(join(tmpdir(), "ctty-ocr-deferred-")));
+    t.after(() => rm(root, { recursive: true, force: true }));
+    let release;
+    const prepared = new Promise((resolve) => { release = resolve; });
+    const pipeline = new LaunchPipeline({
+      contributors: () => [{ pluginId: "fixture.hold", pluginName: "Hold", serviceId: "launch", secrets: false,
+        launch: { fields: [{ key: "on", label: "On", kind: "boolean", default: true }] } }],
+      call: async () => { await prepared; return {}; },
+      secret: async () => null,
+      runsRoot: join(root, "runs"),
+      timeoutMs: 4_000
+    });
+    const calls = [];
+    const terminals = new TerminalManager(() => undefined, availableRegistry(), undefined, undefined, atCreate, fakeSpawner(calls));
+    t.after(() => { release(); terminals.disposeAll(); });
+    terminals.configureLaunchPipeline(pipeline);
+    const card = terminals.create({ provider: "opencode", cwd: root, profile: "normal", position: { x: 0, y: 0 },
+      launchOptions: { "fixture.hold": { on: true } } }, { captureResult: true });
+    terminals.setLifecycleHooksEnabled(beforeSpawn);
+    release();
+    const deadline = Date.now() + 4_000;
+    while (calls.length === 0 && Date.now() < deadline) await new Promise((resolve) => setTimeout(resolve, 5));
+    assert.equal(calls.length, 1, "the launch spawned");
+    assert.equal(terminals.answerUnavailable(card.id), beforeSpawn ? null : "status_hooks_off");
+    assert.equal(terminals.applyProviderSignal(card.id, { kind: "lifecycle", state: "working", event: "session.status:busy", requestId: "t1" }), beforeSpawn);
+  });
+}
diff --git a/tests/opencode-subagent-ask.test.mjs b/tests/opencode-subagent-ask.test.mjs
new file mode 100644
index 000000000..21b9b20e5
--- /dev/null
+++ b/tests/opencode-subagent-ask.test.mjs
@@ -0,0 +1,192 @@
+/**
+ * An OpenCode subagent on a launch contributor's model keeps `bash` on "ask" (the rule of 566cb4f7: a third-party
+ * model's own judgement is not a safety boundary). Its ask must reach someone who may answer it:
+ * - a decision service's deny fails the call before OpenCode asks; an allow answers OpenCode's prompt once;
+ * - with no verdict, OpenCode's own prompt stays and the card turns needs_approval: Needs attention lists it, the
+ *   orchestrator's wait_for_agent returns it with the prompt, the orchestrator cannot answer it, and the person's
+ *   answer resumes the subagent and the orchestrator's next wait.
+ * The real OpenCode plugin module, runtime gateway, decision hooks, TerminalManager and orchestration tools run here;
+ * no OpenCode, no model. HOME is whatever the test runner's fake HOME is.
+ */
+import assert from "node:assert/strict";
+import { mkdtemp, realpath, rm } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import test from "node:test";
+import { AGENT_RUNTIME_ENV, OPENCODE_DECISIONS_ENV } from "../src/agent-runtime/runtime-protocol.mjs";
+import { AgentControlService, APPROVAL_REFUSAL } from "../src/main/services/AgentControlService.ts";
+import { DecisionHooks } from "../src/main/services/DecisionHooks.ts";
+import { LaunchPipeline } from "../src/main/services/LaunchPipeline.ts";
+import { TerminalManager } from "../src/main/services/TerminalManager.ts";
+import { RuntimeGateway } from "../src/main/services/agent-runtime/RuntimeGateway.ts";
+import { ScopedOrchestrationHandler } from "../src/main/services/agent-browser/OrchestrationTools.ts";
+import { attentionSessions } from "../src/renderer/src/features/home/attentionQueue.ts";
+import { availableRegistry, fakeSpawner } from "./helpers/terminal.mjs";
+
+const POSIX = { skip: process.platform === "win32" ? "Unix sockets." : false };
+const until = async (check, what, ms = 4_000) => {
+  const deadline = Date.now() + ms;
+  while (!check()) {
+    if (Date.now() > deadline) throw new Error(`timed out waiting for ${what}`);
+    await new Promise((resolve) => setTimeout(resolve, 5));
+  }
+};
+
+function runtime(decisions) {
+  return {
+    prepareLaunch: () => ({ args: [], environment: decisions ? { CANVASTTY_RUNTIME_DECISIONS: "1" } : {}, decisions, cleanup() {} }),
+    currentStatus: () => null
+  };
+}
+
+/** A launch contributor that puts OpenCode on its own model, as the Assistant's Auto router does. */
+async function modelContributor(t) {
+  const runsRoot = await mkdtemp(join(tmpdir(), "ctty-oc-ask-runs-"));
+  t.after(() => rm(runsRoot, { recursive: true, force: true }));
+  return new LaunchPipeline({
+    contributors: () => [{ pluginId: "p.router", pluginName: "Router", serviceId: "svc", secrets: false,
+      launch: { fields: [{ key: "on", label: "On", kind: "boolean", default: true }], delegable: true } }],
+    call: async () => ({ thirdPartyModel: true }),
+    secret: async () => null,
+    runsRoot,
+    timeoutMs: 500
+  });
+}
+
+async function cards(t, { thirdPartyModel }) {
+  const root = await realpath(await mkdtemp(join(tmpdir(), "ctty-oc-ask-")));
+  t.after(() => rm(root, { recursive: true, force: true }));
+  const calls = [];
+  const writes = [];
+  const terminals = new TerminalManager(() => undefined, availableRegistry(), undefined, runtime(true), true,
+    fakeSpawner(calls, { onWrite: (data, options) => writes.push({ data, options }) }));
+  terminals.configureBaseProtection(() => true);
+  t.after(() => terminals.disposeAll());
+  if (thirdPartyModel) terminals.configureLaunchPipeline(await modelContributor(t));
+  const options = thirdPartyModel ? { launchOptions: { "p.router": { on: true } } } : {};
+  const launch = async (request) => {
+    const before = calls.length;
+    const created = terminals.create({ provider: "opencode", cwd: root, profile: "auto", position: { x: 0, y: 0 }, ...options, ...request });
+    await until(() => calls.length > before, "the launch");
+    const inline = calls.at(-1).options.env.OPENCODE_CONFIG_CONTENT;
+    return { id: created.id, env: calls.at(-1).options.env, bash: JSON.parse(inline).agent.build.permission.bash,
+      metadata: () => terminals.getMetadata(created.id) };
+  };
+  const person = await launch({});
+  const orchestrator = await launch({ role: "orchestrator" });
+  const subagent = await launch({ role: "subagent", parentSessionId: orchestrator.id });
+  return { root, terminals, writes, person, orchestrator, subagent };
+}
+
+test("a subagent on a launch contributor's model keeps bash asking; person cards and an own-model OpenCode are unchanged", async (t) => {
+  const routed = await cards(t, { thirdPartyModel: true });
+  for (const card of [routed.person, routed.orchestrator, routed.subagent]) {
+    assert.equal(card.bash, "ask", "a third-party model's shell commands ask, in every card");
+    assert.equal(card.metadata().autoDowngraded, true, "and the card says auto runs as edits only");
+  }
+  assert.equal(routed.subagent.metadata().role, "subagent");
+  const own = await cards(t, { thirdPartyModel: false });
+  for (const card of [own.person, own.orchestrator, own.subagent]) {
+    assert.equal(card.bash, "allow", "OpenCode on its own model under base protection: the guard is the gate, as before");
+    assert.equal(card.metadata().autoDowngraded, undefined);
+  }
+});
+
+test("the subagent's shell ask: decisions answer OpenCode's prompt, otherwise it reaches the person, not the orchestrator", POSIX, async (t) => {
+  const { terminals, writes, orchestrator, subagent, root } = await cards(t, { thirdPartyModel: true });
+  // What the decision service answers for each command; null is no opinion.
+  const verdicts = { "npm test": { verdict: "allow", reason: "tests" }, "rm -rf build": { verdict: "deny", reason: "not that" }, ls: null };
+  const decide = new DecisionHooks({
+    baseProtection: () => true,
+    services: () => [{ pluginId: "p.review", pluginName: "Review", serviceId: "decide", mayAllow: true }],
+    call: async (_plugin, _service, _method, params) => verdicts[params.tool.command] ?? null,
+    session: (id) => id === subagent.id ? { provider: "opencode", role: "subagent", cwd: root, configDirs: [], profile: "auto" } : null
+  });
+  const runtimeDirectory = await mkdtemp(join(tmpdir(), "ctty-oc-ask-rt-"));
+  const gateway = new RuntimeGateway({
+    runtimeDirectory,
+    onPermissionRequest: (id, request, signal) => decide.decide(id, request, signal),
+    // As src/main/index.ts applies it.
+    onSignal: (id, signal) => terminals.applyProviderSignal(id, { kind: "lifecycle", state: signal.state, event: signal.event,
+      ...(signal.turnId ? { requestId: signal.turnId } : {}), ...(signal.threadId ? { threadId: signal.threadId } : {}) })
+  });
+  await gateway.start();
+  t.after(async () => { await gateway.close(); await rm(runtimeDirectory, { recursive: true, force: true }); });
+  const capability = gateway.registerSession(subagent.id, "opencode", false, undefined, true);
+
+  // The subagent's OpenCode loads CanvasTTY's plugin with the environment its launch gives it.
+  assert.equal(subagent.env[OPENCODE_DECISIONS_ENV], "1", "the launch installs the guard");
+  Object.assign(process.env, {
+    [OPENCODE_DECISIONS_ENV]: "1",
+    [AGENT_RUNTIME_ENV.address]: capability.address,
+    [AGENT_RUNTIME_ENV.terminalSessionId]: subagent.id,
+    [AGENT_RUNTIME_ENV.provider]: "opencode",
+    [AGENT_RUNTIME_ENV.capabilityToken]: capability.capabilityToken
+  });
+  const { CanvasTTYLifecycle } = await import("../src/agent-runtime/opencode-plugin.mjs");
+  const replies = [];
+  const plugin = await CanvasTTYLifecycle({ client: { permission: { reply: async (body) => { replies.push(body); return { response: { ok: true } }; } } } });
+  const root_ = "ses_root01";
+  const event = (type, properties) => plugin.event({ event: { type, properties } });
+  await event("session.created", { info: { id: root_ } });
+  await event("session.status", { sessionID: root_, status: { type: "busy" } });
+  await until(() => subagent.metadata().status === "working", "the subagent working");
+  const bash = async (callID, command) => {
+    await plugin["tool.execute.before"]({ tool: "bash", sessionID: root_, callID }, { args: { command } });
+    // The guard let the call go on: OpenCode now checks bash, which asks.
+    await event("permission.asked", { id: `per_${callID}`, sessionID: root_, permission: "bash", patterns: [command], tool: { messageID: "msg_1", callID } });
+  };
+
+  // A decision service's allow answers OpenCode's prompt once; nobody else is asked.
+  await bash("c1", "npm test");
+  assert.deepEqual(replies, [{ requestID: "per_c1", reply: "once" }]);
+  assert.equal(subagent.metadata().status, "working", "no needs_approval for an answered prompt");
+  // Its deny fails the call with the reason before OpenCode asks anything.
+  await assert.rejects(plugin["tool.execute.before"]({ tool: "bash", sessionID: root_, callID: "c2" }, { args: { command: "rm -rf build" } }),
+    /Review.*not that/u);
+  assert.equal(replies.length, 1);
+
+  // No verdict: OpenCode's own prompt stays, for the person. CanvasTTY never answers it.
+  await bash("c3", "ls");
+  assert.equal(replies.length, 1, "no verdict is never turned into an allow");
+  await until(() => subagent.metadata().status === "needs_approval", "needs_approval");
+  assert.deepEqual(attentionSessions(terminals.list()).map((session) => session.id), [subagent.id], "Needs attention lists the subagent");
+
+  // The orchestrator learns at once, with the prompt; it cannot answer it.
+  const control = new AgentControlService(terminals, { waitTiming: { checkMs: 5, settleMs: 20, quietMs: 60_000 } });
+  const handler = new ScopedOrchestrationHandler(control, null, { cli: () => "available", limits: () => ({ maxDepth: 2, maxSubagents: 8 }) });
+  const call = (tool, args) => handler.execute(orchestrator.id, { id: "r", tool, arguments: args });
+  const first = await call("wait_for_agent", { sessionId: subagent.id, timeoutSeconds: 5 });
+  assert.equal(first.reason, "needs_approval");
+  assert.ok(first.waitedMs < 1_000);
+  const typed = writes.length;
+  for (const prompt of ["y", "allow"]) {
+    await assert.rejects(call("send_to_agent", { sessionId: subagent.id, prompt }), (error) => {
+      assert.equal(error.bridgeError?.code ?? error.code, "INVALID_REQUEST");
+      assert.equal(error.message, APPROVAL_REFUSAL);
+      return true;
+    });
+  }
+  assert.equal(writes.length, typed, "nothing reached the subagent's terminal");
+
+  // The next wait waits for the person instead of returning the same prompt again.
+  const again = call("wait_for_agent", { sessionId: subagent.id, timeoutSeconds: 5 });
+  let settled = false;
+  void again.then(() => { settled = true; }, () => { settled = true; });
+  await new Promise((resolve) => setTimeout(resolve, 150));
+  assert.equal(settled, false, "still waiting while the prompt waits for the person");
+  // The person answers OpenCode's prompt in the card; OpenCode replies, runs `ls` and ends the turn.
+  await event("permission.replied", { sessionID: root_, requestID: "per_c3", reply: "once" });
+  await until(() => subagent.metadata().status === "working", "the subagent working again");
+  assert.deepEqual(attentionSessions(terminals.list()), [], "the answered card leaves Needs attention");
+  await event("session.status", { sessionID: root_, status: { type: "idle" } });
+  const done = await again;
+  assert.equal(done.reason, "idle");
+  assert.equal(replies.length, 1, "the person answered, not CanvasTTY");
+
+  // A new prompt after that is reported at once again.
+  await event("session.status", { sessionID: root_, status: { type: "busy" } });
+  await bash("c4", "ls");
+  await until(() => subagent.metadata().status === "needs_approval", "the second prompt");
+  assert.equal((await call("wait_for_agent", { sessionId: subagent.id, timeoutSeconds: 5 })).reason, "needs_approval");
+});
diff --git a/tests/orchestration-gateway.test.mjs b/tests/orchestration-gateway.test.mjs
index e2f0a524a..ace1c16c2 100644
--- a/tests/orchestration-gateway.test.mjs
+++ b/tests/orchestration-gateway.test.mjs
@@ -370,6 +370,7 @@ test("a spawn_agent canceled while it was starting closes the agent it created",
   const control = {
     status: () => ({ role: "orchestrator", provider: "codex" }),
     profileFor: () => ({ profile: "normal", inherited: true }),
+    subagentLaunchOptions: (_parent, _provider, launchOptions) => ({ launchOptions, sources: {} }),
     spawn: async () => {
       controller.abort();
       return { id: "child-1", provider: "codex", status: "running", title: "worker" };
diff --git a/tests/orchestration-providers-wait.test.mjs b/tests/orchestration-providers-wait.test.mjs
index ebe189141..dbeb0c662 100644
--- a/tests/orchestration-providers-wait.test.mjs
+++ b/tests/orchestration-providers-wait.test.mjs
@@ -2,6 +2,8 @@ import assert from "node:assert/strict";
 import test from "node:test";
 import {
   AGENT_PROVIDER_IDS,
+  CLIENT_SAFE_AGENT_WAIT_SECONDS,
+  DEFAULT_AGENT_WAIT_SECONDS,
   MAX_AGENT_WAIT_SECONDS,
   ORCHESTRATION_TOOL_DEFINITIONS,
   ORCHESTRATION_TOOL_NAMES,
@@ -22,10 +24,14 @@ test("the catalog lists list_providers first and wait_for_agent with a bounded t
   assert.deepEqual(definition("list_providers").inputSchema, { type: "object", properties: {}, required: [], additionalProperties: false });
   const wait = definition("wait_for_agent").inputSchema;
   assert.deepEqual(wait.required, ["sessionId"]);
-  assert.equal(wait.properties.timeoutSeconds.maximum, 600);
-  assert.equal(MAX_AGENT_WAIT_SECONDS, 600);
-  assert.equal(validateOrchestrationArguments("wait_for_agent", { sessionId: "a", timeoutSeconds: 600 }).ok, true);
-  assert.match(validateOrchestrationArguments("wait_for_agent", { sessionId: "a", timeoutSeconds: 601 }).error, /above the maximum/u);
+  assert.equal(wait.properties.timeoutSeconds.maximum, 100);
+  assert.equal(MAX_AGENT_WAIT_SECONDS, 100);
+  assert.ok(DEFAULT_AGENT_WAIT_SECONDS <= CLIENT_SAFE_AGENT_WAIT_SECONDS && CLIENT_SAFE_AGENT_WAIT_SECONDS < 60,
+    "a call without timeoutSeconds, and every call from a 60-second MCP client, ends before that client gives up");
+  assert.equal(validateOrchestrationArguments("wait_for_agent", { sessionId: "a", timeoutSeconds: 100 }).ok, true);
+  assert.match(validateOrchestrationArguments("wait_for_agent", { sessionId: "a", timeoutSeconds: 101 }).error, /above the maximum/u);
+  assert.match(definition("wait_for_agent").description, /at most 100; unless this session is Claude Code, a call returns after at most 50 seconds/u);
+  assert.match(definition("wait_for_agent").description, /A timeout means it is still running: read output as its progress and call wait_for_agent again/u);
   assert.equal(validateOrchestrationArguments("list_providers", {}).ok, true);
   assert.match(validateOrchestrationArguments("list_providers", { x: 1 }).error, /Unexpected argument/u);
   assert.match(definition("list_providers").description, /Never search the filesystem/u);
@@ -217,7 +223,7 @@ test("wait_for_agent stops at once when the call is canceled", async () => {
   const controller = new AbortController();
   const started = Date.now();
   setTimeout(() => controller.abort(), 20);
-  await assert.rejects(call(handler, "wait_for_agent", { sessionId: "child", timeoutSeconds: 600 }, controller.signal),
+  await assert.rejects(call(handler, "wait_for_agent", { sessionId: "child", timeoutSeconds: 100 }, controller.signal),
     (error) => error.bridgeError?.code === "CANCELED");
   assert.ok(Date.now() - started < 1_000);
   const aborted = new AbortController();
@@ -261,7 +267,7 @@ test("after a prompt, an idle before the turn starts does not end wait_for_agent
   const { tmpdir } = await import("node:os");
   const { join } = await import("node:path");
   const { TerminalManager } = await import("../src/main/services/TerminalManager.ts");
-  const { availableRegistry, fakeSpawner } = await import("./helpers/terminal.mjs");
+  const { availableRegistry, fakeSpawner, openCodeInputReady } = await import("./helpers/terminal.mjs");
   const root = await realpath(await mkdtemp(join(tmpdir(), "ctty-turn-")));
   t.after(() => rm(root, { recursive: true, force: true }));
   const calls = [];
@@ -277,7 +283,9 @@ test("after a prompt, an idle before the turn starts does not end wait_for_agent
   terminals.applyProviderSignal(idleChild.sessionId, { kind: "lifecycle", state: "idle" });
   assert.equal((await handler.execute(orchestrator.id, { id: "i", tool: "wait_for_agent", arguments: { sessionId: idleChild.sessionId } })).reason, "idle");
 
-  const child = await handler.execute(orchestrator.id, { id: "1", tool: "spawn_agent", arguments: { provider: "opencode", cwd: root, prompt: "do part A" } });
+  const spawning = handler.execute(orchestrator.id, { id: "1", tool: "spawn_agent", arguments: { provider: "opencode", cwd: root, prompt: "do part A" } });
+  await openCodeInputReady(calls, 2);
+  const child = await spawning;
   // The CLI's startup idle arrives after the prompt was typed but before its turn starts.
   terminals.applyProviderSignal(child.sessionId, { kind: "lifecycle", state: "idle" });
   const early = await wait(1);
@@ -303,3 +311,188 @@ test("the skill tells the orchestrator to pass a model the person names", async
   const skill = await readFile(new URL("../agent/orchestrator/SKILL.md", import.meta.url), "utf8");
   assert.match(skill, /If the person names a model, pass it as `model`/u);
 });
+
+test("wait_for_agent returns before the orchestrator's MCP client gives up, as a timeout that says the agent still runs", async () => {
+  for (const [provider, requested, expectedMs] of [
+    ["opencode", 100, 50_000], ["codex", 100, 50_000], ["opencode", undefined, 50_000], ["opencode", 20, 20_000], ["claude", 100, 100_000]
+  ]) {
+    const { terminals, control, handler } = setup();
+    terminals.sessions.get("orch").provider = provider;
+    let received;
+    const real = control.waitFor.bind(control);
+    control.waitFor = (id, request) => { received = request.timeoutMs; return real(id, { ...request, timeoutMs: 30 }); };
+    terminals.sessions.get("child").buffer = "step 3 of 5";
+    const result = await call(handler, "wait_for_agent", { sessionId: "child", ...(requested ? { timeoutSeconds: requested } : {}) });
+    assert.equal(received, expectedMs, `${provider} asking for ${requested ?? "the default"}`);
+    assert.deepEqual([result.reason, result.status, result.output], ["timeout", "working", "step 3 of 5"]);
+    assert.match(result.message, /Still running\. Read output for its progress and call wait_for_agent again\./u);
+  }
+  const { terminals, handler } = setup();
+  terminals.sessions.get("child").status = "idle";
+  assert.equal((await call(handler, "wait_for_agent", { sessionId: "child" })).message, undefined, "only a timeout says to call again");
+});
+
+/** A recorded OpenCode-style session: alternate screen, absolute cursor moves, colors, a spinner and in-place redraws. */
+function openCodeRedraws() {
+  const e = "\u001b";
+  const at = (row, column) => `${e}[${row};${column}H`;
+  const parts = [
+    `${e}[?1049h${e}[?25l${e}[?2004h${e}]0;opencode${e}\\${e}[2J${e}[H`,
+    `${at(1, 1)}${e}[48;2;24;24;37m${e}[38;2;205;214;244m┃ opencode  build  gpt-5.5${e}[K${e}[0m`,
+    `${at(3, 3)}${e}[1mYou${e}[22m`,
+    `${at(4, 3)}${e}[38;5;252mFix the failing Button test${e}[39m`,
+    `${at(6, 3)}${e}[38;5;245mThinking${e}[0m`
+  ];
+  // The spinner turns in place many times while the model works.
+  for (let turn = 0; turn < 120; turn++) parts.push(`${at(6, 12)}${"⠋⠙⠹⠸⠼⠴⠦⠧⠇⠏"[turn % 10]}${at(40, 1)}${e}[7m working ${turn}s ${e}[27m${e}[K`);
+  parts.push(
+    `${at(6, 1)}${e}[2K${at(6, 3)}${e}[38;2;166;227;161mThe Button test passes now.${e}[0m`,
+    `${at(7, 3)}Used key sk-secret-value for the request.`,
+    `${at(8, 3)}${e}[38;5;111m│${e}[0m src/Button.tsx ${e}[32m+3${e}[0m ${e}[31m-1${e}[0m`,
+    `${at(40, 1)}${e}[2K${e}[7m ctrl+p commands ${e}[27m`
+  );
+  return parts.join("");
+}
+
+test("wait_for_agent and get_agent_result return a TUI's screen as plain text, not its escape sequences", async () => {
+  const { terminals, handler } = setup();
+  const child = terminals.sessions.get("child");
+  setTimeout(() => {
+    child.buffer = openCodeRedraws();
+    child.outputOffset = child.buffer.length;
+    child.status = "idle";
+  }, 20);
+  const waited = await call(handler, "wait_for_agent", { sessionId: "child", timeoutSeconds: 5 });
+  const result = await call(handler, "get_agent_result", { sessionId: "child" });
+  for (const output of [waited.output, result.output]) {
+    assert.equal(/[\u0000-\u0008\u000b-\u001f\u007f]/u.test(output), false, "no control characters or escape sequences");
+    assert.deepEqual(output.split("\n"), [
+      "┃ opencode  build  gpt-5.5",
+      "",
+      "  You",
+      "  Fix the failing Button test",
+      "",
+      "  The Button test passes now.",
+      "  Used key [redacted] for the request.",
+      "  │ src/Button.tsx +3 -1",
+      ...Array.from({ length: 31 }, () => ""),
+      " ctrl+p commands"
+    ]);
+    assert.ok(!output.includes("sk-secret") && !output.includes("Thinking") && !output.includes("working"), "masked; only what the screen shows now");
+  }
+});
+
+test("the plain-text output stays bounded and masked for a long line-based transcript", async () => {
+  const { terminals, control } = setup();
+  const child = terminals.sessions.get("child");
+  child.buffer = Array.from({ length: 4_000 }, (_value, index) => `\u001b[32mline ${index}\u001b[0m with key sk-secret-value\r\n`).join("");
+  child.outputOffset = child.buffer.length;
+  child.status = "idle";
+  const result = await control.result("child");
+  assert.ok(result.output.length <= 8_192);
+  assert.ok(result.output.endsWith("line 3999 with key [redacted]"), result.output.slice(-80));
+  assert.ok(!result.output.includes("\u001b") && !result.output.includes("sk-secret"));
+});
+
+/** Invisible SGR resets that pad a stream to an exact length: any length from 6 up is a sum of 3s and 4s. */
+function sgrPadding(length) {
+  const e = "\u001b";
+  let fours = 0;
+  while ((length - 4 * fours) % 3 !== 0) fours++;
+  return `${e}[0m`.repeat(fours) + `${e}[m`.repeat((length - 4 * fours) / 3);
+}
+
+test("a scrollback window that starts inside an escape sequence leaves no fragment of it on the screen", async () => {
+  const { terminals, handler } = setup();
+  const child = terminals.sessions.get("child");
+  const e = "\u001b";
+  const at = (row, column) => `${e}[${row};${column}H`;
+  const start = "18;9Hof specialists who had internalized the machine.";
+  const redraws = [];
+  for (let turn = 0; turn < 900; turn++) redraws.push(`${at(40, 1)}${e}[7m working ${turn}s ${e}[27m${e}[K`);
+  const end = `${at(3, 3)}The first response was the assembler.`;
+  const body = start + redraws.join("") + end;
+  // The last 65,536 characters (the window the screen is rendered from) begin right after the ESC [ of a cursor move.
+  child.buffer = `${"earlier output\r\n".repeat(50)}${e}[` + body + sgrPadding(65_536 - body.length);
+  child.outputOffset = child.buffer.length;
+  child.status = "idle";
+  const waited = await call(handler, "wait_for_agent", { sessionId: "child", timeoutSeconds: 5 });
+  const result = await call(handler, "get_agent_result", { sessionId: "child" });
+  for (const output of [waited.output, result.output]) {
+    assert.ok(!output.includes("18;9H"), output.slice(0, 80));
+    assert.ok(output.includes("  The first response was the assembler."), output);
+  }
+});
+
+test("exitLines are the exited TUI's screen as plain text, not its redraws joined together", async () => {
+  const { terminals, handler } = setup();
+  const child = terminals.sessions.get("child");
+  const e = "\u001b";
+  const at = (row, column) => `${e}[${row};${column}H`;
+  // Drawn once at the start and never again: still on the screen when the CLI exits, more than 16 KiB later.
+  const early = `${at(3, 3)}${e}[2mThinking${e}[0m${at(8, 3)}${e}[38;5;111m▣  Build · glm-5.3-flash${e}[0m`;
+  const turns = 1_300;
+  const progress = [];
+  for (let turn = 0; turn < turns; turn++) progress.push(`${at(2, 3)}${e}[38;2;166;227;161m${"■".repeat(turn % 7)}${"⬝".repeat(7 - (turn % 7))}${e}[0m`);
+  // A TUI redraws only the cells that changed: here a line written once, then its typos fixed in place.
+  const body = early + progress.join("") +
+    `${at(4, 3)}As explicit memoy address, and if a single instructon was` +
+    `${at(4, 19)}ry address, and if a single instruction was` +
+    `${at(6, 3)}${e}[31mError: the model stopped answering.${e}[0m`;
+  assert.ok(body.length > 16_384 && body.length < 65_000, String(body.length));
+  // The rendered window (the last 65,536 characters) begins with the `m` of a color sequence whose start was cut off.
+  child.buffer = `${"x".repeat(100)}${e}[38;2;1;2;3` + "m" + body + sgrPadding(65_536 - 1 - body.length);
+  child.outputOffset = child.buffer.length;
+  child.exitCode = 1;
+  child.status = "failed";
+  const waited = await call(handler, "wait_for_agent", { sessionId: "child", timeoutSeconds: 5 });
+  const observed = await call(handler, "observe_agent", { sessionId: "child" });
+  const result = await call(handler, "get_agent_result", { sessionId: "child" });
+  assert.equal(waited.reason, "failed");
+  const last = (turns - 1) % 7;
+  assert.deepEqual(waited.exitLines.split("\n"), [
+    `  ${"■".repeat(last)}${"⬝".repeat(7 - last)}`,
+    "  Thinking",
+    "  As explicit memory address, and if a single instruction was",
+    "  Error: the model stopped answering.",
+    "  ▣  Build · glm-5.3-flash"
+  ]);
+  assert.equal(observed.exitLines, waited.exitLines);
+  assert.equal(result.exitLines, waited.exitLines);
+});
+
+test("observe_agent returns the screen as plain text too, within its maxChars", async () => {
+  const { terminals, handler } = setup();
+  const child = terminals.sessions.get("child");
+  child.buffer = openCodeRedraws();
+  child.outputOffset = child.buffer.length;
+  child.status = "working";
+  const observed = await call(handler, "observe_agent", { sessionId: "child" });
+  assert.equal(/[\u0000-\u0008\u000b-\u001f\u007f]/u.test(observed.output), false, "no ESC or other control bytes");
+  assert.ok(observed.output.includes("  The Button test passes now.") && observed.output.includes("[redacted]"));
+  assert.ok(!observed.output.includes("sk-secret") && !observed.output.includes("working"));
+  const short = await call(handler, "observe_agent", { sessionId: "child", maxChars: 256 });
+  assert.ok(short.output.length <= 256 && short.output.endsWith(" ctrl+p commands"), short.output);
+});
+
+test("a TUI that left its full screen as it exited keeps its last screen in output and exitLines", async () => {
+  const { terminals, handler } = setup();
+  const child = terminals.sessions.get("child");
+  const e = "\u001b";
+  const at = (row, column) => `${e}[${row};${column}H`;
+  // A full-screen CLI shows why it stops on its own screen, then leaves the alternate screen as every TUI does on exit.
+  child.buffer = `${e}[?1049h${e}[2J${at(1, 1)}┃ opencode  build  glm-5.3-flash${at(6, 3)}${e}[31mError: Unexpected server error${e}[0m`
+    + `${e}[?25h${e}[?1049l` + "Session ended.\r\n";
+  child.outputOffset = child.buffer.length;
+  child.exitCode = 1;
+  child.status = "failed";
+  const waited = await call(handler, "wait_for_agent", { sessionId: "child", timeoutSeconds: 5 });
+  const result = await call(handler, "get_agent_result", { sessionId: "child" });
+  for (const answer of [waited, result]) {
+    assert.match(answer.exitLines, /Error: Unexpected server error/u);
+    assert.match(answer.exitLines, /Session ended\./u);
+    assert.ok(answer.exitLines.indexOf("Error: Unexpected") < answer.exitLines.indexOf("Session ended."), answer.exitLines);
+    assert.match(answer.output, /Error: Unexpected server error/u);
+    assert.equal(/[\u0000-\u0008\u000b-\u001f\u007f]/u.test(answer.output), false);
+  }
+});
diff --git a/tests/orchestrator-approval-answers.test.mjs b/tests/orchestrator-approval-answers.test.mjs
new file mode 100644
index 000000000..11e546b80
--- /dev/null
+++ b/tests/orchestrator-approval-answers.test.mjs
@@ -0,0 +1,217 @@
+import assert from "node:assert/strict";
+import { randomUUID } from "node:crypto";
+import { mkdir, mkdtemp, realpath, rm } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { dirname, join } from "node:path";
+import { setTimeout as delay } from "node:timers/promises";
+import test from "node:test";
+import { controlRequest } from "../scripts/canvastty-control.mjs";
+import { AgentControlService, APPROVAL_REFUSAL, DelegationRefusal } from "../src/main/services/AgentControlService.ts";
+import { AgentControlGateway, PERSON_ANSWERS_REFUSAL, codexApprovalPrompt } from "../src/main/services/agent-control/AgentControlGateway.ts";
+import { TerminalManager } from "../src/main/services/TerminalManager.ts";
+import { availableRegistry, fakeSpawner } from "./helpers/terminal.mjs";
+
+// An orchestrator never answers a subagent's approval or permission prompt: not with send_to_agent (MCP), and not with
+// choose, dismiss or send on its own control connection. Those prompts are the person's (or a decision service's the
+// person allowed, which answers before the CLI asks). Non-approval menus stay the orchestrator's, the person's own
+// connection and the card's own input are unchanged.
+
+const localSocket = { skip: process.platform === "win32" ? "Unix socket tests" : false };
+const at = { x: 0, y: 0 };
+const COMPOSER = "\x1b[2J\x1b[H>_ OpenAI Codex\r\nmodel: test\r\n\r\n› \x1b[2mAsk Codex to do anything\x1b[22m";
+const APPROVAL = "\x1b[2J\x1b[HWould you like to run the following command?\r\n\r\n  $ rm -rf build\r\n\r\n"
+  + "› 1. Yes, proceed (y)\r\n  2. Yes, and don't ask again for this command (p)\r\n"
+  + "  3. No, and tell Codex what to do differently (esc)\r\n\r\nPress enter to confirm or esc to cancel";
+const EDITS = "\x1b[2J\x1b[HWould you like to make the following edits?\r\n\r\n  src/a.ts (+1 -0)\r\n\r\n"
+  + "› 1. Yes, proceed (y)\r\n  2. Yes, and don't ask again for these files (a)\r\n"
+  + "  3. No, and tell Codex what to do differently (esc)\r\n\r\nPress enter to confirm or esc to cancel";
+const HOOKS = "\x1b[2J\x1b[HHooks need review\r\n› 1. Review hooks\r\n  2. Trust all and continue\r\n  3. Continue without hooks\r\n"
+  + "Press enter to confirm or esc to cancel";
+const MODEL = "\x1b[2J\x1b[HApproaching rate limits\r\nSwitch to a smaller model for lower credit usage?\r\n"
+  + "› 1. Switch model\r\n  2. Keep current model\r\n  3. Keep current model (never show again)\r\nPress enter to confirm or esc to go back";
+
+async function fixture(t) {
+  const root = await realpath(await mkdtemp(join(tmpdir(), "ctty-no-approve-")));
+  const userData = await realpath(await mkdtemp(join("/tmp", "ctty-no-approve-grant-")));
+  t.after(() => Promise.all([rm(root, { recursive: true, force: true }), rm(userData, { recursive: true, force: true })]));
+  const project = join(root, "project");
+  const inner = join(project, "src");
+  await mkdir(inner, { recursive: true });
+  const calls = [];
+  const writes = new Map();
+  let gateway;
+  const terminals = new TerminalManager((channel, payload) => gateway?.observe(channel, payload), availableRegistry(), undefined, undefined, true,
+    fakeSpawner(calls, { onWrite: (data, options) => {
+      const call = calls.find((entry) => entry.options === options);
+      if (call) writes.set(call, [...(writes.get(call) ?? []), data]);
+    } }));
+  const control = new AgentControlService(terminals, { waitTiming: { checkMs: 10, settleMs: 20, quietMs: 200 } });
+  gateway = new AgentControlGateway({ userDataPath: userData, terminals, lifecycleEnabled: () => true,
+    spawnSubagent: (request) => control.spawn(request) });
+  const appConnection = await gateway.start();
+  terminals.setControlConnection({ connectionPath: appConnection, cliPath: "/cli.mjs", grant: (id) => gateway.grantSession(id) });
+  t.after(async () => { await gateway.close(); await terminals.shutdown(); });
+  const orchestrator = terminals.create({ provider: "codex", profile: "auto", cwd: project, position: at, role: "orchestrator" });
+  const connectionPath = calls.at(-1).options.env.CANVASTTY_CONTROL_CONNECTION;
+  assert.ok(connectionPath && connectionPath !== appConnection, "the orchestrator has its own connection");
+  const agentClient = join(dirname(connectionPath), "controller.json");
+  const agent = (method, params = {}, requestId = randomUUID()) =>
+    controlRequest({ connectionPath, clientPath: agentClient, method, params, requestId, timeoutMs: 10_000 });
+  const person = (method, params = {}, requestId = randomUUID()) =>
+    controlRequest({ connectionPath: appConnection, clientPath: join(userData, "person-client.json"), method, params, requestId, timeoutMs: 10_000 });
+  // The card a create just launched (each launch spawns one PTY).
+  const callOf = (id) => {
+    const call = calls.at(-1);
+    assert.equal(terminals.getMetadata(id)?.exitCode, null);
+    return call;
+  };
+  const written = (call) => writes.get(call) ?? [];
+  return { root, project, inner, terminals, control, gateway, orchestrator, calls, agent, person, callOf, written };
+}
+
+/** Prints a screen into the card and waits until its controller's view shows it. */
+async function show(request, id, call, data, expected) {
+  call.process.emitData(data);
+  for (let i = 0; i < 60; i++) {
+    const screen = await request("screen", { sessionId: id });
+    if (screen.text.includes(expected)) return screen;
+    await delay(5);
+  }
+  assert.fail(`the screen did not show ${expected}`);
+}
+
+const needsPerson = (error) => error.code === "NEEDS_PERSON" && error.message === PERSON_ANSWERS_REFUSAL;
+
+test("Codex's approval prompts are told apart from its other menus", () => {
+  for (const screen of [APPROVAL, EDITS, "Would you like to grant these permissions?\n› 1. Yes\n  2. No",
+    "Do you want to approve network access to \"example.com\"?\n› 1. Yes, just this once\n  2. No",
+    "docs needs your approval.\n› 1. Yes, provide the requested info\n  2. No, but continue without it",
+    "› 1. Yes, proceed (y)\n  2. No (esc)"]) {
+    assert.equal(codexApprovalPrompt(screen), true, screen);
+  }
+  for (const screen of [HOOKS, MODEL, COMPOSER, "Do you trust the contents of this directory?\n› 1. Yes, continue\n  2. No, quit",
+    "Update available!\n› 1. Update now\n  2. Skip\n  3. Skip until next version"]) {
+    assert.equal(codexApprovalPrompt(screen), false, screen);
+  }
+});
+
+test("an orchestrator's control connection cannot choose, dismiss or type into a subagent's approval prompt; the person still answers it", localSocket, async (t) => {
+  const f = await fixture(t);
+  const created = await f.agent("create", { provider: "codex", cwd: f.inner });
+  const id = created.session.id;
+  assert.equal(f.terminals.getMetadata(id).parentSessionId, f.orchestrator.id);
+  assert.equal(created.capabilities.menus, true);
+  const call = f.callOf(id);
+
+  // A task the orchestrator legitimately sends, at an empty composer.
+  f.terminals.applyProviderSignal(id, { kind: "lifecycle", state: "idle" });
+  await show(f.agent, id, call, COMPOSER, "Ask Codex");
+  await f.agent("send", { sessionId: id, text: "clean the build" });
+  assert.equal(f.written(call).length, 1);
+  f.terminals.applyProviderSignal(id, { kind: "lifecycle", state: "working" });
+
+  // Codex shows its approval prompt before its PermissionRequest hook's report arrived: the screen alone refuses.
+  const early = await show(f.agent, id, call, APPROVAL, "Yes, proceed");
+  assert.equal(early.interaction.options.length, 3, "the menu is still shown to the orchestrator");
+  await assert.rejects(f.agent("choose", { sessionId: id, choice: 1, revision: early.interaction.revision }), needsPerson);
+  await assert.rejects(f.agent("dismiss", { sessionId: id, revision: early.revision }), needsPerson);
+  await assert.rejects(f.agent("send", { sessionId: id, text: "y" }), needsPerson, "text (and its Enter) would answer it too");
+  // The hook's report: the card is needs_approval.
+  f.terminals.applyProviderSignal(id, { kind: "lifecycle", state: "needs_approval" });
+  const screen = await f.agent("screen", { sessionId: id });
+  for (const choice of [1, 2, 3]) {
+    await assert.rejects(f.agent("choose", { sessionId: id, choice, revision: screen.interaction.revision }), needsPerson, `choice ${choice}`);
+  }
+  await assert.rejects(f.agent("dismiss", { sessionId: id, revision: screen.revision }), needsPerson, "Escape is an answer too");
+  await assert.rejects(f.agent("send", { sessionId: id, text: "y" }), needsPerson);
+  // Not retryable: the same request id replays the refusal instead of writing.
+  const again = randomUUID();
+  await assert.rejects(f.agent("choose", { sessionId: id, choice: 1, revision: screen.interaction.revision }, again), needsPerson);
+  await assert.rejects(f.agent("choose", { sessionId: id, choice: 1, revision: screen.interaction.revision }, again), needsPerson);
+  assert.equal(f.written(call).length, 1, "nothing reached the prompt");
+  // Reading stays open, and stopping its own turn is not an answer: the asked command does not run.
+  assert.equal((await f.agent("status", { sessionId: id })).session.status, "needs_approval");
+  await f.agent("interrupt", { sessionId: id });
+  assert.deepEqual(f.written(call).slice(1), ["\x03"]);
+
+  // The person answers in the card (its own input path), whatever the orchestrator was refused.
+  f.terminals.input(id, "y");
+  assert.deepEqual(f.written(call).slice(2), ["y"]);
+  f.terminals.applyProviderSignal(id, { kind: "lifecycle", state: "working" });
+  f.terminals.applyProviderSignal(id, { kind: "lifecycle", state: "idle", event: "Stop" });
+  // Edits ask the same way.
+  const edits = await show(f.agent, id, call, EDITS, "make the following edits");
+  await assert.rejects(f.agent("choose", { sessionId: id, choice: 1, revision: edits.interaction.revision }), needsPerson);
+  assert.equal(f.written(call).length, 3);
+});
+
+test("an orchestrator still resolves menus that are not approvals, and the person's own connection still chooses approvals", localSocket, async (t) => {
+  const f = await fixture(t);
+  const { session } = await f.agent("create", { provider: "codex", cwd: f.inner });
+  const call = f.callOf(session.id);
+  f.terminals.applyProviderSignal(session.id, { kind: "lifecycle", state: "idle" });
+  const hooks = await show(f.agent, session.id, call, HOOKS, "Hooks need review");
+  await f.agent("choose", { sessionId: session.id, choice: 3, revision: hooks.interaction.revision });
+  const model = await show(f.agent, session.id, call, MODEL, "Approaching rate limits");
+  await f.agent("choose", { sessionId: session.id, choice: 2, revision: model.interaction.revision });
+  const notice = await show(f.agent, session.id, call, "\x1b[2J\x1b[HHooks\r\n6 hooks reviewed\r\nPress esc to close", "Press esc");
+  await f.agent("dismiss", { sessionId: session.id, revision: notice.revision });
+  assert.deepEqual(f.written(call), ["\x1b[B\x1b[B\r", "\x1b[B\r", "\x1b"]);
+  // A worker with no approval pending takes the orchestrator's task as before.
+  await show(f.agent, session.id, call, COMPOSER, "Ask Codex");
+  await f.agent("send", { sessionId: session.id, text: "next task" });
+  assert.equal(f.written(call).length, 4);
+
+  // The person's own automation (the app-wide connection) is the person: its worker's approval prompt is its to choose.
+  const own = await f.person("create", { provider: "codex", profile: "normal", cwd: f.inner, title: "Person's worker" });
+  const ownCall = f.callOf(own.session.id);
+  f.terminals.applyProviderSignal(own.session.id, { kind: "lifecycle", state: "needs_approval" });
+  const approval = await show(f.person, own.session.id, ownCall, APPROVAL, "Yes, proceed");
+  await f.person("choose", { sessionId: own.session.id, choice: 3, revision: approval.interaction.revision });
+  assert.deepEqual(f.written(ownCall), ["\x1b[B\x1b[B\r"]);
+  // And the orchestrator cannot reach the person's worker at all.
+  await assert.rejects(f.agent("choose", { sessionId: own.session.id, choice: 1, revision: approval.interaction.revision }),
+    (error) => error.code === "SESSION_NOT_FOUND");
+});
+
+test("send_to_agent is refused at every CLI's approval or permission prompt and allowed once the person answered", async (t) => {
+  const root = await realpath(await mkdtemp(join(tmpdir(), "ctty-no-approve-send-")));
+  t.after(() => rm(root, { recursive: true, force: true }));
+  const calls = [];
+  const writes = new Map();
+  const terminals = new TerminalManager(() => undefined, availableRegistry(), undefined, undefined, true, fakeSpawner(calls, {
+    onWrite: (data, options) => writes.set(options, [...(writes.get(options) ?? []), data])
+  }));
+  t.after(() => terminals.shutdown());
+  const control = new AgentControlService(terminals);
+  const orchestrator = terminals.create({ provider: "claude", profile: "normal", cwd: root, position: at, role: "orchestrator" });
+  calls.forEach((entry) => { entry.claimed = true; });
+  // Each maps its CLI's own approval signal to needs_approval: Claude Code, Codex, Qwen and Kimi PermissionRequest,
+  // Grok's permission_prompt notification, Hermes pre_approval_request, OpenCode permission.asked and question.asked.
+  for (const provider of ["claude", "codex", "grok", "kimi", "qwen", "hermes", "opencode"]) {
+    const child = terminals.create({ provider, profile: "normal", cwd: root, position: at, role: "subagent", parentSessionId: orchestrator.id });
+    if (provider === "grok") terminals.resize(child.id, 80, 24); // Grok starts with the renderer-measured grid.
+    let call;
+    for (let i = 0; i < 100 && !call; i++) {
+      call = calls.find((entry) => entry.options.cwd === root && JSON.stringify([entry.command, entry.args]).includes(`/resolved/${provider}`) && !entry.claimed);
+      if (!call) await delay(5);
+    }
+    assert.ok(call, `${provider}: launched`);
+    call.claimed = true;
+    terminals.applyProviderSignal(child.id, { kind: "lifecycle", state: "needs_approval" });
+    for (const prompt of ["y", "1", "allow", "yes, and don't ask again"]) {
+      assert.throws(() => control.send(child.id, prompt), (error) => error instanceof DelegationRefusal && error.message === APPROVAL_REFUSAL, `${provider}: ${prompt}`);
+      assert.throws(() => control.send(child.id, prompt, false), DelegationRefusal, `${provider}: ${prompt} without Enter`);
+    }
+    assert.deepEqual(writes.get(call.options) ?? [], [], `${provider}: nothing reached its prompt`);
+    // The person answers in the card; the subagent works and finishes its turn.
+    terminals.input(child.id, "1");
+    terminals.applyProviderSignal(child.id, { kind: "lifecycle", state: "working" });
+    terminals.applyProviderSignal(child.id, { kind: "lifecycle", state: "idle", event: "Stop" });
+    if (provider === "codex") call.process.emitData(COMPOSER);
+    await control.send(child.id, "next task");
+    const sent = writes.get(call.options);
+    assert.equal(sent?.[0], "1", `${provider}: the person's answer`);
+    assert.equal(sent.length, 2, `${provider}: the orchestrator's next task once the person answered`);
+  }
+});
diff --git a/tests/plugin-launch-choices.test.mjs b/tests/plugin-launch-choices.test.mjs
index 884ea5bb6..1f024e90b 100644
--- a/tests/plugin-launch-choices.test.mjs
+++ b/tests/plugin-launch-choices.test.mjs
@@ -158,7 +158,9 @@ test("spawn_agent takes plugin launch options and hands them to the launch", ()
     get: (id) => (id === "orch" ? parent : undefined),
     getMetadata: (id) => (id === "orch" ? parent : null),
     listMetadata: () => [parent],
-    create: (request) => { created.push(request); return { id: "child", ...request, status: "starting", title: "c" }; }
+    create: (request) => { created.push(request); return { id: "child", ...request, status: "starting", title: "c" }; },
+    launchOptionsOf: () => undefined,
+    modelSetByOf: () => []
   };
   const control = new AgentControlService(terminals);
   control.spawn({ parentSessionId: "orch", provider: "codex", cwd: process.cwd(), launchOptions: { "canvastty-accounts": { account: "ollama" } } });
diff --git a/tests/plugin-service-notice.test.mjs b/tests/plugin-service-notice.test.mjs
new file mode 100644
index 000000000..f5adc3872
--- /dev/null
+++ b/tests/plugin-service-notice.test.mjs
@@ -0,0 +1,206 @@
+import assert from "node:assert/strict";
+import { createHash } from "node:crypto";
+import { cp, mkdir, mkdtemp, readFile, rm, writeFile } from "node:fs/promises";
+import { createRequire } from "node:module";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import test from "node:test";
+import { fileURLToPath } from "node:url";
+import { build } from "esbuild";
+import { PluginManager } from "../src/main/services/PluginManager.ts";
+import { PluginServiceSupervisor } from "../src/main/services/PluginServiceSupervisor.ts";
+import {
+  PLUGIN_NATIVE_CODE_OFF_ERROR,
+  PLUGIN_SERVICE_MODULE_OFF_ERROR,
+  pluginServiceNoticeKind,
+  pluginServiceNoticeSection,
+  pluginServiceNoticeText
+} from "../src/shared/pluginServiceNotice.ts";
+
+// A plugin page whose service cannot run because the person has not turned its native code (or its module) on gets
+// an actionable notice instead of "Plugin service is not running."; a trusted service that crashed keeps the generic
+// text, now with its last error.
+const root = fileURLToPath(new URL("..", import.meta.url));
+const example = new URL("../examples/plugins/service-echo/", import.meta.url);
+const exampleManifest = JSON.parse(await readFile(new URL("canvastty.plugin.json", example), "utf8"));
+const sha256 = (content) => createHash("sha256").update(content).digest("hex");
+const source = (path) => readFile(new URL(`../${path}`, import.meta.url), "utf8");
+
+function supervisor(manager, options = {}) {
+  return new PluginServiceSupervisor({
+    command: process.execPath,
+    hostVersion: "9.9.9",
+    locale: () => "en",
+    requestTimeoutMs: 2_000,
+    stopGraceMs: 300,
+    restartDelaysMs: [30],
+    host: {
+      storageGet: (pluginId, key) => manager.storageGet(pluginId, key),
+      storageSet: (pluginId, key, value) => manager.storageSet(pluginId, key, value),
+      emit: () => undefined
+    },
+    ...options
+  });
+}
+
+/** What registerIpc's requestPluginService does: the manager's check, then the supervisor. */
+async function pageRequest(manager, instance, pluginId, serviceId, method, params) {
+  manager.assertService(pluginId, serviceId);
+  return instance.request(pluginId, serviceId, method, params);
+}
+
+test("native code off: the page's request names the switch, and the notice is localized with a way to Agents", async (t) => {
+  const userData = await mkdtemp(join(tmpdir(), "canvastty-service-notice-"));
+  const manager = new PluginManager(userData, async (_url, destination) => { await cp(example, destination, { recursive: true }); });
+  const instance = supervisor(manager);
+  manager.setServiceObserver((specs) => instance.sync(specs));
+  t.after(async () => {
+    await instance.dispose();
+    await manager.dispose();
+    await rm(userData, { recursive: true, force: true });
+  });
+  await manager.load();
+  const { manifest } = await manager.install((await manager.previewInstall("https://github.com/example/service-echo")).token);
+
+  await assert.rejects(pageRequest(manager, instance, manifest.id, "echo", "echo", { text: "x" }), (error) => {
+    assert.equal(error.message, PLUGIN_NATIVE_CODE_OFF_ERROR);
+    return true;
+  });
+  const kind = pluginServiceNoticeKind(PLUGIN_NATIVE_CODE_OFF_ERROR);
+  assert.equal(kind, "native-code-off");
+  assert.equal(pluginServiceNoticeSection(kind), "agents");
+  assert.equal(
+    pluginServiceNoticeText(kind, "ru"),
+    "Нативный код этого плагина выключен — включите его в Настройки → Агенты → Нативный код расширений."
+  );
+  assert.match(pluginServiceNoticeText(kind, "en"), /native code is turned off — turn it on in Settings → Agents → Extension native code/u);
+  assert.equal(pluginServiceNoticeText("open", "ru"), "Открыть настройки");
+  assert.equal(pluginServiceNoticeText("open", "en"), "Open settings");
+  // The renderer recognizes its own localized text again (the page reports it, the host shows the notice).
+  assert.equal(pluginServiceNoticeKind(pluginServiceNoticeText(kind, "ru")), "native-code-off");
+  assert.equal(pluginServiceNoticeKind(pluginServiceNoticeText(kind, "en")), "native-code-off");
+
+  // Trusted: the same request reaches the service.
+  await manager.setNativeCodeTrusted(manifest.id, true);
+  assert.equal((await pageRequest(manager, instance, manifest.id, "echo", "echo", { text: "x" })).count, 1);
+  // Disabling revokes the trust; the disabled plugin itself is refused as before.
+  await manager.setEnabled(manifest.id, false);
+  await manager.setEnabled(manifest.id, true);
+  assert.throws(() => manager.assertService(manifest.id, "echo"), (error) => error.message === PLUGIN_NATIVE_CODE_OFF_ERROR);
+  assert.throws(() => manager.assertService(manifest.id, "other"), /does not exist/u);
+});
+
+test("a service whose module is off says so and points to Plugins", async (t) => {
+  const userData = await mkdtemp(join(tmpdir(), "canvastty-service-notice-modules-"));
+  const files = {
+    "apps/echo.html": await readFile(new URL("apps/echo.html", example)),
+    "services/echo.mjs": await readFile(new URL("services/echo.mjs", example)),
+    "services/extra.mjs": Buffer.from("process.stdin.resume();\n")
+  };
+  const asset = (path) => ({ path, bytes: files[path].length, sha256: sha256(files[path]) });
+  const manifest = {
+    ...exampleManifest,
+    coreFiles: [asset("apps/echo.html"), asset("services/echo.mjs")],
+    modules: [{ id: "extra", title: "Extra", defaultSelected: false, permissions: [], files: [asset("services/extra.mjs")] }],
+    services: [...exampleManifest.services, { id: "extra", title: "Extra", entry: "services/extra.mjs", module: "extra" }]
+  };
+  const manager = new PluginManager(
+    userData,
+    async (_url, destination) => {
+      await mkdir(destination, { recursive: true });
+      await writeFile(join(destination, "canvastty.plugin.json"), JSON.stringify(manifest));
+    },
+    async (_url, destination, requested) => {
+      for (const file of requested) {
+        await mkdir(join(destination, file.path, ".."), { recursive: true });
+        await writeFile(join(destination, file.path), files[file.path]);
+      }
+    }
+  );
+  t.after(async () => {
+    await manager.dispose();
+    await rm(userData, { recursive: true, force: true });
+  });
+  await manager.load();
+  const installed = await manager.install((await manager.previewInstall("https://github.com/example/service-echo")).token, []);
+  await manager.setNativeCodeTrusted(installed.manifest.id, true);
+  manager.assertService(installed.manifest.id, "echo");
+  assert.throws(() => manager.assertService(installed.manifest.id, "extra"), (error) => error.message === PLUGIN_SERVICE_MODULE_OFF_ERROR);
+  const kind = pluginServiceNoticeKind(PLUGIN_SERVICE_MODULE_OFF_ERROR);
+  assert.equal(kind, "module-off");
+  assert.equal(pluginServiceNoticeSection(kind), "plugins");
+  assert.match(pluginServiceNoticeText(kind, "ru"), /Модуль этого плагина с нативным кодом выключен — включите его в Настройки → Плагины/u);
+});
+
+test("a trusted service that crashed keeps the generic text, with its last error, and no settings notice", async (t) => {
+  const dir = await mkdtemp(join(tmpdir(), "canvastty-service-notice-crash-"));
+  const instance = supervisor({ storageGet: async () => null, storageSet: async () => undefined }, { maxRestarts: 0 });
+  t.after(async () => { await instance.dispose(); await rm(dir, { recursive: true, force: true }); });
+  const entry = [
+    'import { createInterface } from "node:readline";',
+    'createInterface({ input: process.stdin }).on("line", (line) => {',
+    '  const m = JSON.parse(line);',
+    '  if (m.method === "crash") process.exit(3);',
+    '  if (m.method === "ping") process.stdout.write(JSON.stringify({ jsonrpc: "2.0", id: m.id, result: "pong" }) + "\\n");',
+    "});"
+  ].join("\n");
+  const entryPath = join(dir, "probe.mjs");
+  await writeFile(entryPath, entry);
+  await instance.sync([{ pluginId: "com.example.a", serviceId: "probe", root: dir, entryPath, sha256: sha256(entry), dataDir: join(dir, "data"), permissions: [] }]);
+  assert.equal(await instance.request("com.example.a", "probe", "ping", null), "pong");
+  await assert.rejects(instance.request("com.example.a", "probe", "crash", null), /stopped/u);
+  assert.equal(instance.report("com.example.a").services[0].state, "failed");
+  await assert.rejects(instance.request("com.example.a", "probe", "ping", null), (error) => {
+    assert.match(error.message, /^Plugin service is not running\. Last error: The service stopped unexpectedly \(exit code 3\) too often/u);
+    assert.equal(pluginServiceNoticeKind(error.message), null, "a real crash is not a settings notice");
+    return true;
+  });
+  // A service the supervisor never ran (no record) keeps the plain text.
+  await assert.rejects(instance.request("com.example.b", "probe", "ping", null), (error) => error.message === "Plugin service is not running.");
+});
+
+test("the toast shows the notice's button only when it has one", async (t) => {
+  const outdir = await mkdtemp(join(tmpdir(), "canvastty-toast-"));
+  t.after(() => rm(outdir, { recursive: true, force: true }));
+  await build({
+    stdin: {
+      contents: [
+        'export { Toast } from "./src/renderer/src/components/Toast";',
+        'export { createElement } from "react";',
+        'export { renderToStaticMarkup } from "react-dom/server";'
+      ].join("\n"),
+      resolveDir: root,
+      loader: "ts"
+    },
+    jsx: "automatic", bundle: true, platform: "node", format: "cjs", outfile: join(outdir, "toast.cjs"), logLevel: "silent",
+    define: { "process.env.NODE_ENV": '"production"' }
+  });
+  const { Toast, createElement, renderToStaticMarkup } = createRequire(import.meta.url)(join(outdir, "toast.cjs"));
+  const plain = renderToStaticMarkup(createElement(Toast, { message: "Saved" }));
+  assert.match(plain, /class="toast toast--visible"/u);
+  assert.doesNotMatch(plain, /<button/u);
+  const text = pluginServiceNoticeText("native-code-off", "ru");
+  const actionable = renderToStaticMarkup(createElement(Toast, { message: text, action: { label: pluginServiceNoticeText("open", "ru"), run: () => undefined } }));
+  assert.match(actionable, /toast--action/u);
+  assert.match(actionable, /<button class="toast__action" type="button">Открыть настройки<\/button>/u);
+  assert.ok(actionable.includes(text));
+});
+
+test("the notice is wired: frames localize it, App shows it with Open settings, Settings opens Extension native code", async () => {
+  const [frame, app, settings, ipc] = await Promise.all([
+    source("src/renderer/src/features/plugins/PluginFrame.tsx"),
+    source("src/renderer/src/App.tsx"),
+    source("src/renderer/src/features/settings/SettingsPanel.tsx"),
+    source("src/main/ipc/registerIpc.ts")
+  ]);
+  assert.match(frame, /pluginServiceNoticeKind\(raw\)/u);
+  assert.match(frame, /pluginServiceNoticeText\(notice, locale\)/u);
+  assert.match(app, /onPluginError=\{showPluginError\}/u);
+  assert.doesNotMatch(app, /onPluginError=\{showToast\}/u);
+  assert.match(app, /openSectionRequest=\{openSectionRequest\}/u);
+  assert.match(app, /<Toast message=\{toast\} action=\{toastAction\} \/>/u);
+  assert.match(settings, /setSection\(openSectionRequest\.section\)/u);
+  assert.match(settings, /querySelector<HTMLElement>\("\.plugin-services"\)/u);
+  // The service request path still checks the manager before the supervisor.
+  assert.match(ipc, /plugins\.assertService\(pluginId, serviceId\);\n\s+return pluginServices\.request/u);
+});
diff --git a/tests/quota-used-left-labels.test.mjs b/tests/quota-used-left-labels.test.mjs
new file mode 100644
index 000000000..c4321d2da
--- /dev/null
+++ b/tests/quota-used-left-labels.test.mjs
@@ -0,0 +1,100 @@
+import assert from "node:assert/strict";
+import { readFile } from "node:fs/promises";
+import { fileURLToPath } from "node:url";
+import test from "node:test";
+import { build } from "esbuild";
+import { normalizeClaudeLimits, normalizeCodexLimits } from "../src/main/services/LimitsService.ts";
+import { selectHomeModel } from "../src/renderer/src/features/home/homeModel.ts";
+import { fill, quotaShareText, usageText } from "../src/renderer/src/features/usage/usageText.ts";
+
+// The real HOME row, rendered without Electron (same harness as codex-api-limits-ui.test.mjs).
+const bundle = await build({
+  stdin: {
+    contents: 'export { LimitRow } from "./src/renderer/src/features/home/HomeZone.tsx"; export { createElement } from "react"; export { renderToStaticMarkup } from "react-dom/server";',
+    resolveDir: fileURLToPath(new URL("..", import.meta.url)), loader: "ts"
+  },
+  plugins: [{ name: "test-limit-row", setup(build) {
+    build.onLoad({ filter: /[/\\]HomeZone\.tsx$/ }, async ({ path }) => ({
+      contents: await readFile(path, "utf8") + "\nexport { LimitRow };", loader: "tsx"
+    }));
+  } }],
+  bundle: true, platform: "node", format: "esm", jsx: "automatic", write: false,
+  loader: { ".svg": "dataurl", ".png": "dataurl", ".ico": "dataurl", ".css": "empty" }, logLevel: "silent",
+  banner: { js: `import { createRequire } from "node:module"; const require = createRequire(${JSON.stringify(import.meta.url)});` }
+});
+const { LimitRow, createElement, renderToStaticMarkup } = await import(
+  "data:text/javascript;base64," + Buffer.from(bundle.outputFiles[0].text).toString("base64")
+);
+
+const now = 1_786_100_000_000;
+const seconds = (ms) => Math.floor(ms / 1_000);
+
+// `account/rateLimits/read` of `codex app-server`: usedPercent is the share already spent.
+const codexPayload = {
+  rateLimits: {
+    limitId: "codex",
+    limitName: null,
+    primary: { usedPercent: 9, windowDurationMins: 300, resetsAt: seconds(now + 2 * 3_600_000) },
+    secondary: { usedPercent: 9, windowDurationMins: 10_080, resetsAt: seconds(now + 5 * 86_400_000) }
+  },
+  rateLimitsByLimitId: {}
+};
+// The Claude OAuth usage endpoint: utilization is the share already spent, too.
+const claudePayload = {
+  five_hour: { utilization: 9, resets_at: new Date(now + 2 * 3_600_000).toISOString() },
+  seven_day: { utilization: 9, resets_at: new Date(now + 5 * 86_400_000).toISOString() }
+};
+
+/** What a person reads: text nodes only, not title/aria-label attributes. */
+const visibleText = (html) => html.replace(/<[^>]*>/g, " ").replace(/&nbsp;/g, " ").replace(/\s+/g, " ").trim();
+const attribute = (html, name) => html.match(new RegExp(`${name}="([^"]*)"`))?.[1] ?? "";
+
+const render = (provider, windows, locale) => {
+  const snapshot = { fetchedAt: now, providers: [
+    { provider, source: provider === "codex" ? "codex-app-server" : "claude-usage-api", state: "available", checkedAt: now, windows }
+  ] };
+  const row = selectHomeModel([], snapshot, "ready", now, [provider]).limitRows[0];
+  return renderToStaticMarkup(createElement(LimitRow, { row, locale, now }));
+};
+
+test("the Codex and Claude normalizers keep 9 % spent as usedPercent 9 (not 91)", () => {
+  assert.deepEqual(normalizeCodexLimits(codexPayload).map((window) => window.usedPercent), [9, 9]);
+  assert.deepEqual(normalizeClaudeLimits(claudePayload).map((window) => window.usedPercent), [9, 9]);
+});
+
+for (const [locale, used, left] of [
+  ["en", /9% used/, /91% left/],
+  ["ru", /использовано/, /осталось 91%/]
+]) {
+  for (const [provider, windows] of [["codex", normalizeCodexLimits(codexPayload)], ["claude", normalizeClaudeLimits(claudePayload)]]) {
+    test(`HOME ${provider} row in ${locale} says on screen that 9 % is used and 91 % is left`, () => {
+      const html = render(provider, windows, locale);
+      const text = visibleText(html);
+      assert.match(text, /9\s?%/u, text);
+      assert.match(text, used, `the visible row must say the share is used: ${text}`);
+      assert.match(text, left, `the visible row must say what is left: ${text}`);
+      assert.doesNotMatch(text, locale === "en" ? /\b9% left/ : /осталось 9%/u, text);
+      // The hover/screen-reader line says both too.
+      const label = attribute(html, "aria-label");
+      assert.match(label, locale === "en" ? /9% used/ : /использовано 9%/u, label);
+      assert.match(label, locale === "en" ? /91% left/ : /осталось 91%/u, label);
+      // The bar is the used share.
+      assert.match(html, /limit-row__fill" style="width:9%"/);
+    });
+  }
+}
+
+test("usage history says the latest quota observation is used, with what is left, in both languages", () => {
+  assert.equal(quotaShareText("en", 9), "9% used · 91% left");
+  assert.equal(quotaShareText("ru", 9), "использовано 9% · осталось 91%");
+  assert.equal(quotaShareText("ru", 12.5), "использовано 12,5% · осталось 87,5%");
+  assert.equal(quotaShareText("en", 140), "100% used · 0% left");
+  for (const locale of ["en", "ru"]) {
+    const text = usageText(locale);
+    const latest = fill(text.latestValue, { share: quotaShareText(locale, 9), time: "T" });
+    assert.ok(latest.includes(quotaShareText(locale, 9)), latest);
+    const none = fill(text.noObservationsInPeriod, { share: quotaShareText(locale, 9), time: "T", account: "A" });
+    assert.ok(none.includes(quotaShareText(locale, 9)), none);
+    assert.ok(!/\{percent\}/u.test(text.latestValue + text.noObservationsInPeriod), "no bare percent template left");
+  }
+});
diff --git a/tests/renderer-bundle-splitting.test.mjs b/tests/renderer-bundle-splitting.test.mjs
index 91a34d321..fbb90fd86 100644
--- a/tests/renderer-bundle-splitting.test.mjs
+++ b/tests/renderer-bundle-splitting.test.mjs
@@ -1,5 +1,5 @@
 import assert from "node:assert/strict";
-import { mkdtemp, rm } from "node:fs/promises";
+import { mkdtemp, readFile, rm } from "node:fs/promises";
 import { tmpdir } from "node:os";
 import { join } from "node:path";
 import test from "node:test";
@@ -84,3 +84,32 @@ test("Settings and the launch/link dialogs are split out of the app's static imp
 
   );
 });
+
+// Text that only Settings shows is kept next to it (not in the startup string table), so it loads with the Settings
+// chunk. Each entry: a module that must not be statically reachable from App.js, and an English string from it that
+// must not appear in any chunk App.js loads at startup.
+const SETTINGS_ONLY_TEXT = [
+  ["src/renderer/src/features/settings/canvasNavigationText.ts", "Wheel/pinch capture over widgets"],
+  ["src/renderer/src/features/plugins/PluginSettingsSection.tsx", "GitHub sign-in could not be cancelled"]
+];
+
+test("text that only Settings shows stays out of the startup bundle", async () => {
+  const outdir = await mkdtemp(join(tmpdir(), "canvastty-bundle-text-"));
+  try {
+    const { metafile } = await build({
+      entryPoints: [appEntry], bundle: true, splitting: true, platform: "browser", format: "esm", outdir, write: true,
+      metafile: true, logLevel: "silent",
+      loader: { ".svg": "text", ".css": "empty", ".png": "empty", ".ttf": "empty", ".woff2": "empty", ".ico": "empty" }
+    });
+    const { paths: eagerPaths } = staticallyReachableBytes(metafile, outputFor(metafile, "App.js"));
+    const eagerText = (await Promise.all([...eagerPaths].map((path) => readFile(path, "utf8")))).join("\n");
+    for (const [module, text] of SETTINGS_ONLY_TEXT) {
+      const holders = Object.entries(metafile.outputs).filter(([, output]) => Object.keys(output.inputs).some((input) => input.endsWith(module)));
+      assert.ok(holders.length > 0, `${module} is bundled`);
+      for (const [path] of holders) assert.ok(!eagerPaths.has(path), `${module} must only load with Settings`);
+      assert.ok(!eagerText.includes(text), `"${text}" must not be in the startup bundle`);
+    }
+  } finally {
+    await rm(outdir, { recursive: true, force: true });
+  }
+});
diff --git a/tests/ru-shortcut-hints-i18n.test.mjs b/tests/ru-shortcut-hints-i18n.test.mjs
new file mode 100644
index 000000000..578d04e75
--- /dev/null
+++ b/tests/ru-shortcut-hints-i18n.test.mjs
@@ -0,0 +1,119 @@
+import assert from "node:assert/strict";
+import { readFile } from "node:fs/promises";
+import test from "node:test";
+import { t as tableText } from "../src/renderer/src/lib/i18n.ts";
+import { canvasNavigationText } from "../src/renderer/src/features/settings/canvasNavigationText.ts";
+
+// The canvas wheel/navigation settings keep their text next to Settings (out of the startup string table).
+const SETTINGS_ONLY = new Set(["canvasWheelCapture", "canvasWheelCaptureDescription", "canvasWheelCaptureKey", "canvasWheelOverride",
+  "canvasOverrideBindingsMatch", "canvasNavigationOverrideDescription"]);
+const t = (locale, key) => (SETTINGS_ONLY.has(key) ? canvasNavigationText(locale, key) : tableText(locale, key));
+
+const canvasPath = new URL("../src/renderer/src/features/workspace/WorkspaceCanvas.tsx", import.meta.url);
+const referencePath = new URL("../src/renderer/src/components/ShortcutReference.tsx", import.meta.url);
+const settingsPath = new URL("../src/renderer/src/features/settings/SettingsPanel.tsx", import.meta.url);
+
+// Latin words that may stay untranslated in Russian hints: keyboard and mouse key
+// names, plus product names and the HOME zone label that the UI shows verbatim.
+const allowedLatin = new Set([
+  "Shift", "Command", "Super", "Ctrl", "Control", "Alt", "Option", "Meta",
+  "Enter", "Escape", "Esc", "Space", "Tab", "Backspace", "Delete",
+  "PageUp", "PageDown", "Mouse3", "Mouse4", "Mouse5",
+  "HOME", "Codex", "CLI", "TUI"
+]);
+
+function latinWords(text) {
+  return [...text.matchAll(/[A-Za-z][A-Za-z0-9]*/g)].map((match) => match[0])
+    .filter((word) => !allowedLatin.has(word) && !/^F\d{1,2}$/.test(word));
+}
+
+async function hintKeys() {
+  const [canvas, reference] = await Promise.all([readFile(canvasPath, "utf8"), readFile(referencePath, "utf8")]);
+  const aside = canvas.slice(canvas.indexOf('<aside className="shortcut-hints"'), canvas.indexOf("</aside>"));
+  assert.ok(aside.length > 0, "canvas shortcut hints panel found");
+  const keys = new Set();
+  for (const match of aside.matchAll(/t\(settings\.locale, "(\w+)"\)/g)) keys.add(match[1]);
+  for (const match of reference.matchAll(/\], "(\w+)"\]/g)) keys.add(match[1]);
+  for (const match of reference.matchAll(/title: "(\w+)"/g)) keys.add(match[1]);
+  for (const match of reference.matchAll(/t\(locale, "(\w+)"\)/g)) keys.add(match[1]);
+  return { keys: [...keys], aside, reference };
+}
+
+test("Russian shortcut hints contain no English words besides key names", async () => {
+  const { keys } = await hintKeys();
+  assert.ok(keys.includes("canvasWheelOverrideHint"));
+  assert.ok(keys.includes("canvasNavigationOverrideHint"));
+  assert.ok(keys.includes("marqueeSelectionHint"));
+  assert.ok(keys.length >= 30, `collected ${keys.length} hint keys`);
+  const offenders = keys.map((key) => [key, t("ru", key), latinWords(t("ru", key))])
+    .filter(([, , words]) => words.length > 0);
+  assert.deepEqual(offenders, []);
+});
+
+test("shortcut hint key caps do not hard-code English gesture words", async () => {
+  const { aside, reference } = await hintKeys();
+  for (const source of [aside, reference]) {
+    const literals = [...source.matchAll(/<kbd>([^<{]+)<\/kbd>|\[\["([^"\]]+)"\]/g)]
+      .map((match) => match[1] ?? match[2]);
+    for (const literal of literals) assert.deepEqual(latinWords(literal), [], `kbd literal ${literal}`);
+  }
+  assert.equal(t("ru", "shortcutShiftDrag"), "Shift + перетаскивание");
+  assert.equal(t("en", "shortcutShiftDrag"), "Shift + drag");
+});
+
+test("the Russian canvas wheel settings are translated, including the capture mode labels", async () => {
+  const settings = await readFile(settingsPath, "utf8");
+  const start = settings.indexOf('label={canvasNavigationText(locale, "canvasWheelCapture")}');
+  assert.ok(start > 0, "the wheel capture setting is found");
+  const group = settings.slice(start, settings.indexOf("</SettingGroup>", start));
+  const options = /options=\{(\[\[[\s\S]*?\]\])\}/u.exec(group)?.[1] ?? "";
+  assert.ok(options.includes('"off"') && options.includes('"always"') && options.includes('"key"'), options);
+  assert.deepEqual([...options.matchAll(/\["\w+", "([^"]*)"\]/gu)].map((match) => match[1]), [], "no hard-coded option labels");
+  const keys = new Set(["canvasWheelCapture", "canvasWheelCaptureDescription", "canvasWheelOverride", "canvasWheelCaptureKey"]);
+  for (const match of group.matchAll(/(?:\bt|canvasNavigationText)\(locale, "(\w+)"\)/gu)) keys.add(match[1]);
+  const offenders = [...keys].map((key) => [key, latinWords(t("ru", key))]).filter(([, words]) => words.length > 0);
+  assert.deepEqual(offenders, []);
+  assert.equal(t("en", "canvasWheelCaptureKey"), "Key");
+});
+
+test("the Russian canvas navigation override setting is translated", async () => {
+  const settings = await readFile(settingsPath, "utf8");
+  const start = settings.indexOf('label={t(locale, "canvasNavigationOverride")}');
+  assert.ok(start > 0, "the navigation override setting is found");
+  const group = settings.slice(start, settings.indexOf("</SettingGroup>", start));
+  const keys = new Set(["canvasNavigationOverride", "canvasNavigationOverrideDescription", "canvasNavigationCaptureHint"]);
+  for (const match of group.matchAll(/(?:\bt|canvasNavigationText)\(locale, "(\w+)"\)/gu)) keys.add(match[1]);
+  const offenders = [...keys].map((key) => [key, latinWords(t("ru", key))]).filter(([, words]) => words.length > 0);
+  assert.deepEqual(offenders, []);
+});
+
+test("Russian hints and canvas settings name the drag gesture with one word", async () => {
+  const { keys } = await hintKeys();
+  const all = [...keys, "canvasWheelCaptureDescription", "canvasNavigationOverrideDescription", "canvasOverrideBindingsMatch"];
+  const other = all.filter((key) => /тян(уть|ите)|протаскив/iu.test(t("ru", key)));
+  assert.deepEqual(other, []);
+  // The key cap names the gesture; its label only says what it does.
+  for (const locale of ["ru", "en"]) assert.ok(!t(locale, "marqueeSelectionHint").includes(t(locale, "shortcutShiftDrag")), locale);
+  assert.equal(t("ru", "marqueeSelectionHint"), "Выделить группу");
+  assert.equal(t("en", "marqueeSelectionHint"), "Select a group");
+  assert.match(t("ru", "canvasNavigationOverrideDescription"), /перетаскивание/u);
+});
+
+test("the hints panel shows action shortcuts with the same key formatter as the shortcut reference", async () => {
+  const { displayShortcutBinding } = await import("../src/renderer/src/lib/shortcuts.ts");
+  assert.equal(displayShortcutBinding("Alt+ArrowUp", false), "Alt+↑");
+  assert.equal(displayShortcutBinding("Meta+Comma", true), "Command+,");
+  assert.equal(displayShortcutBinding("Meta+Shift+F", false), "Super+Shift+F");
+  const { aside, reference } = await hintKeys();
+  assert.doesNotMatch(aside, /<kbd>\{settings\.shortcuts\./u, "no raw binding in a hint key cap");
+  assert.match(aside, /displayShortcutBinding\(settings\.shortcuts\[action\]/u);
+  for (const action of ["home", "renameWindow", "toggleFullscreen", "focusUp"]) assert.match(aside, new RegExp(`\\["${action}", "\\w+"\\]`, "u"), action);
+  assert.match(reference, /displayShortcutBinding\(/u);
+});
+
+test("long labels in the canvas navigation rows of Settings wrap instead of being cut off", async () => {
+  const css = await readFile(new URL("../src/renderer/src/styles/app.css", import.meta.url), "utf8");
+  const rule = /\.setting-group__control \.shortcut-editor__row > span \{([^}]*)\}/u.exec(css)?.[1] ?? "";
+  assert.match(rule, /white-space: normal/u);
+  assert.match(rule, /text-overflow: clip/u);
+});
diff --git a/tests/runtime-result-transport.test.mjs b/tests/runtime-result-transport.test.mjs
new file mode 100644
index 000000000..010f56428
--- /dev/null
+++ b/tests/runtime-result-transport.test.mjs
@@ -0,0 +1,129 @@
+import assert from "node:assert/strict";
+import { spawn } from "node:child_process";
+import { realpathSync } from "node:fs";
+import { mkdtemp, rm } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import test from "node:test";
+import { fileURLToPath } from "node:url";
+import { AGENT_RUNTIME_ENV, CAPTURE_RESULT_ENV } from "../src/agent-runtime/runtime-protocol.mjs";
+import { AgentControlService } from "../src/main/services/AgentControlService.ts";
+import { AgentRuntimeBridge } from "../src/main/services/agent-runtime/AgentRuntimeBridge.ts";
+import { RuntimeGateway } from "../src/main/services/agent-runtime/RuntimeGateway.ts";
+import { applyRuntimeSignal } from "../src/main/services/agent-runtime/applyRuntimeSignal.ts";
+import { TerminalManager } from "../src/main/services/TerminalManager.ts";
+import { availableRegistry, inputReadySpawner } from "./helpers/terminal.mjs";
+
+const POSIX = { timeout: 15_000, skip: process.platform === "win32" ? "Real Unix socket roundtrip; the Windows pipe transport has its own suite." : false };
+const helper = fileURLToPath(new URL("../src/agent-runtime/hook-helper.mjs", import.meta.url));
+const plugin = fileURLToPath(new URL("../src/agent-runtime/opencode-plugin.mjs", import.meta.url));
+
+/** Runs a hook or plugin process the way the agent CLI would, with the launch's environment. */
+function run(args, env, input = "") {
+  const child = spawn(process.execPath, args, { env: { ...process.env, ...env }, stdio: ["pipe", "ignore", "pipe"] });
+  const timeout = setTimeout(() => child.kill(), 5_000);
+  let stderr = "";
+  child.stderr.on("data", (chunk) => { stderr += chunk; });
+  child.stdin.end(input);
+  return new Promise((resolve, reject) => {
+    child.on("error", reject);
+    child.on("close", (code) => { clearTimeout(timeout); code === 0 ? resolve() : reject(new Error(stderr || `exit ${code}`)); });
+  });
+}
+
+const openCodeTurn = (answer) => `
+  const { CanvasTTYLifecycle } = await import(${JSON.stringify(new URL("../src/agent-runtime/opencode-plugin.mjs", import.meta.url).href)});
+  const hooks = await CanvasTTYLifecycle({ client: { session: { messages: async () => ({ data: [{ info: { role: "assistant" }, parts: [{ type: "text", text: ${JSON.stringify(answer)} }] }] }) } } });
+  const event = (type, rest = {}) => hooks.event({ event: { type, properties: { sessionID: "worker-root", ...rest } } });
+  await event("session.created", { info: { id: "worker-root" } });
+  await event("session.status", { status: { type: "busy" } });
+  await event("session.status", { status: { type: "idle" } });
+  await event("session.idle");
+`;
+
+/** A real gateway, bridge and hook helper/plugin, with fake CLI PTYs; signals go through main's own glue. */
+async function fixture(t, statusHooks) {
+  const root = realpathSync(await mkdtemp(join(tmpdir(), "ctty-rt-")));
+  const calls = [];
+  let terminals;
+  const gateway = new RuntimeGateway({ runtimeDirectory: root, onSignal: (id, signal) => applyRuntimeSignal(terminals, id, signal) });
+  await gateway.start();
+  const bridge = new AgentRuntimeBridge(gateway, {
+    runtimeDirectory: join(root, "configs"), helper: { command: process.execPath, args: [helper] }, openCodePluginPath: plugin,
+    coreHooksEnabled: statusHooks, environment: {},
+    kimiHomeDirectory: join(root, "kimi"), hermesHomeDirectory: join(root, "hermes"), grokHomeDirectory: join(root, "grok")
+  });
+  terminals = new TerminalManager(() => undefined, availableRegistry(), undefined, bridge, statusHooks, inputReadySpawner(calls));
+  t.after(async () => {
+    terminals.disposeAll();
+    await gateway.close();
+    await rm(root, { recursive: true, force: true });
+  });
+  const control = new AgentControlService(terminals, { waitTiming: { checkMs: 5, settleMs: 0, quietMs: 200 } });
+  const parent = terminals.create({ provider: "codex", profile: "normal", cwd: root, role: "orchestrator", position: { x: 0, y: 0 } });
+  // Status hooks are one switch: the bridge (hooks, leases) and the cards (status) follow it together, as in main.
+  const setStatusHooks = (enabled) => {
+    bridge.setCoreHooksEnabled(enabled);
+    terminals.setLifecycleHooksEnabled(enabled);
+  };
+  return { root, calls, gateway, bridge, terminals, control, parent, setStatusHooks };
+}
+
+async function finishTurn(provider, env, answer) {
+  if (provider === "codex") {
+    await run([helper, "working", "UserPromptSubmit"], env, JSON.stringify({ turn_id: "fixture-turn" }));
+    await run([helper, "idle", "Stop"], env, JSON.stringify({ turn_id: "fixture-turn", last_assistant_message: answer }));
+  } else {
+    await run(["--input-type=module", "-e", openCodeTurn(answer)], env);
+  }
+}
+
+for (const provider of ["codex", "opencode"]) {
+  test(`${provider} subagent's answer reaches wait_for_agent through its hooks while status hooks are on`, POSIX, async (t) => {
+    const f = await fixture(t, true);
+    const worker = await f.control.spawn({ parentSessionId: f.parent.id, provider, cwd: f.root });
+    const env = f.calls.at(-1).options.env;
+    assert.equal(env[CAPTURE_RESULT_ENV], "1");
+    assert.ok(env[AGENT_RUNTIME_ENV.capabilityToken]);
+    await f.terminals.deliverInput(worker.id, "worker task\r");
+    await finishTurn(provider, env, "worker final answer");
+    const waited = await f.control.waitFor(worker.id, { timeoutMs: 2_000 });
+    assert.equal(waited.reason, "idle");
+    assert.equal(waited.answer?.text, "worker final answer");
+    assert.equal(waited.answerUnavailable, undefined);
+
+    // Turning status hooks off revokes the lease: this launch has no hooks any more, and the result says so.
+    f.setStatusHooks(false);
+    assert.equal(f.gateway.currentStatus(worker.id), null);
+    assert.equal((await f.control.result(worker.id)).answer?.text, "worker final answer", "an answer captured earlier stays");
+    await f.terminals.deliverInput(worker.id, "next task\r");
+    const after = await f.control.result(worker.id);
+    assert.equal(after.answer, undefined);
+    assert.equal(after.answerUnavailable?.reason, "status_hooks_off");
+  });
+
+  test(`${provider} subagent started with status hooks off gets no hooks, and the wait says its answer cannot be captured`, POSIX, async (t) => {
+    const f = await fixture(t, false);
+    const worker = await f.control.spawn({ parentSessionId: f.parent.id, provider, cwd: f.root });
+    const launch = f.calls.at(-1);
+    assert.equal(launch.options.env[AGENT_RUNTIME_ENV.capabilityToken], undefined, "no runtime lease");
+    assert.deepEqual(launch.args.filter((arg) => typeof arg === "string" && /^hooks\./u.test(arg)), [], "no lifecycle hooks");
+    await f.terminals.deliverInput(worker.id, "worker task\r");
+    const waited = await f.control.waitFor(worker.id, { timeoutMs: 2_000 });
+    assert.equal(waited.reason, "quiet");
+    assert.equal(waited.status, "unavailable");
+    assert.equal(waited.answer, undefined);
+    assert.equal(waited.answerUnavailable?.reason, "status_hooks_off");
+    assert.match(waited.answerUnavailable.message, /status hooks are off/u);
+
+    // Turning them on applies to agents started or restarted afterwards.
+    f.setStatusHooks(true);
+    assert.equal((await f.control.result(worker.id)).answerUnavailable?.reason, "status_hooks_off");
+    launch.process.emitExit(0);
+    f.terminals.restart(worker.id);
+    const restarted = f.calls.at(-1);
+    assert.notEqual(restarted, launch);
+    assert.ok(restarted.options.env[AGENT_RUNTIME_ENV.capabilityToken], "the restarted launch has its hooks");
+    assert.equal((await f.control.result(worker.id)).answerUnavailable, undefined);
+  });
+}
diff --git a/tests/secret-redaction.test.mjs b/tests/secret-redaction.test.mjs
index f3b74d5c2..63ca2ea36 100644
--- a/tests/secret-redaction.test.mjs
+++ b/tests/secret-redaction.test.mjs
@@ -163,7 +163,7 @@ test("the vault hands every value it reads or writes to the registry; agent-read
     readBuffer: () => ({ buffer: `env OPENAI=${value}\n${secrets.github}\n` }),
     redactSecrets: (text) => manager.redactSecrets(text)
   });
-  const observed = control.observe("s1").output;
+  const observed = (await control.observe("s1")).output;
   assert.equal(observed.includes(value) || observed.includes(secrets.github), false);
   assert.match(observed, /env OPENAI=<redacted:secret>/u);
 });
@@ -204,17 +204,17 @@ test("observe_agent and get_agent_result mask the whole buffer before cutting th
   const control = new AgentControlService(f.terminals);
   // The 300-character observation starts 12 characters before the secret ends.
   f.print(`${"a".repeat(1_000)}${PLAIN_SECRET}${"b".repeat(288)}`);
-  const observed = control.observe(f.card.id, 300).output;
+  const observed = (await control.observe(f.card.id, 300)).output;
   assert.ok(observed.length <= 300, "the bound is kept");
   assert.deepEqual(fragments(observed), []);
   // get_agent_result's 8 192-character tail, cut inside the secret the same way.
   f.print(`${PLAIN_SECRET}${"c".repeat(8_192 - 10)}`);
-  const result = control.result(f.card.id).output;
+  const result = (await control.result(f.card.id)).output;
   assert.ok(result.length <= 8_192);
   assert.deepEqual(fragments(result), []);
   // Wrapped secrets are still masked whole.
   f.print(`\r\n${PLAIN_SECRET.slice(0, 15)}\r\n${PLAIN_SECRET.slice(15)}\r\n`);
-  assert.deepEqual(fragments(control.observe(f.card.id, 8_192).output), []);
+  assert.deepEqual(fragments((await control.observe(f.card.id, 8_192)).output), []);
 });
 
 test("plugin screen text and failure details mask the whole buffer before cutting", async (t) => {
@@ -318,8 +318,8 @@ test("observe_agent, get_agent_result and the plugin screen mask a bounded windo
   f.print(scrollback(250_000));
   f.print(`${PLAIN_SECRET}${"g".repeat(200)}`);
   const control = new AgentControlService(f.terminals);
-  assert.deepEqual(fragments(control.observe(f.card.id).output), []);
-  assert.deepEqual(fragments(control.result(f.card.id).output), []);
+  assert.deepEqual(fragments((await control.observe(f.card.id)).output), []);
+  assert.deepEqual(fragments((await control.result(f.card.id)).output), []);
   const { PluginSessions } = await import("../src/main/services/PluginSessions.ts");
   const screens = [];
   const sessions = new PluginSessions({ terminals: f.terminals, notify: (_p, _s, _m, event) => { if (event.screen !== undefined) screens.push(event.screen); return true; } });
@@ -330,7 +330,8 @@ test("observe_agent, get_agent_result and the plugin screen mask a bounded windo
   await new Promise((resolve) => setTimeout(resolve, 20));
   assert.ok(screens.length > 0);
   assert.deepEqual(fragments(screens.join("\n")), []);
-  assert.ok(masked.length >= 3 && Math.max(...masked) < 40_000, `masked ${masked} characters per call`);
+  // get_agent_result renders its plain text from a 64 KiB masked window (plus the tail margin); never the whole scrollback.
+  assert.ok(masked.length >= 3 && Math.max(...masked) < 90_000, `masked ${masked} characters per call`);
 });
 
 /**
diff --git a/tests/shutdown-steps.test.mjs b/tests/shutdown-steps.test.mjs
new file mode 100644
index 000000000..fd0c46dd5
--- /dev/null
+++ b/tests/shutdown-steps.test.mjs
@@ -0,0 +1,76 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+import { runShutdownSteps } from "../src/main/services/shutdownSteps.ts";
+
+test("a failed shutdown step does not skip the steps after it", async () => {
+  const calls = [], warnings = [];
+  const failure = new Error("fixture close failed");
+  await assert.doesNotReject(runShutdownSteps([
+    { name: "Even G2", run: async () => { calls.push("Even G2"); throw failure; } },
+    { name: "sync step", run: () => { calls.push("sync step"); throw failure; } },
+    ...["terminal processes", "browser", "plugin services", "plugins", "materials"].map(name => ({ name, run: async () => { calls.push(name); } }))
+  ], (message, error) => warnings.push({ message, error })));
+  assert.deepEqual(calls, ["Even G2", "sync step", "terminal processes", "browser", "plugin services", "plugins", "materials"]);
+  assert.equal(warnings.length, 2);
+  assert.equal(warnings[0].error, failure);
+  assert.match(warnings[0].message, /Even G2 failed/u);
+});
+
+test("shutdown steps run in order and each waits for the previous one", async () => {
+  const events = [];
+  let release;
+  const held = new Promise(resolve => { release = resolve; });
+  const done = runShutdownSteps([
+    { name: "first", run: async () => { events.push("first:start"); await held; events.push("first:done"); } },
+    { name: "second", run: () => { events.push("second"); } },
+    { name: "absent service", run: () => undefined }
+  ], () => assert.fail("no step failed"));
+  await Promise.resolve();
+  assert.deepEqual(events, ["first:start"]);
+  release();
+  await done;
+  assert.deepEqual(events, ["first:start", "first:done", "second"]);
+});
+
+test("quitting still hangs up the terminals when the Even G2 companion fails to close", async () => {
+  const { readFile } = await import("node:fs/promises");
+  const { stripTypeScriptTypes } = await import("node:module");
+  const { runInNewContext } = await import("node:vm");
+  const main = (await readFile(new URL("../src/main/index.ts", import.meta.url), "utf8")).replaceAll("\r\n", "\n");
+  const start = main.indexOf("async function shutdownServices");
+  const source = main.slice(start, main.indexOf("\n}\n", start) + 2);
+  const calls = [];
+  // The usage-history save finishes only after the limits service aborts its requests.
+  let limitsAborted;
+  const aborted = new Promise(resolve => { limitsAborted = resolve; });
+  const service = (name, method, fail = false) => ({ [method]: async () => { calls.push(name); if (fail) throw new Error(`fixture ${name} failed`); } });
+  const warn = console.warn;
+  console.warn = () => undefined;
+  try {
+    const shutdownServices = runInNewContext(`${stripTypeScriptTypes(source)}; shutdownServices`, {
+      runShutdownSteps, console, clearTimeout, clearInterval, Promise, Error,
+      diagnostics: { record: () => undefined, flush: async () => { calls.push("diagnostics"); } },
+      mascotManager: { dispose: () => { calls.push("mascots"); } },
+      agentChatHistory: { dispose: () => { calls.push("chat history"); } },
+      agentControl: service("agent control", "close"),
+      updateTimer: null, updateInterval: null, browserRequests: new Map(),
+      evenG2: service("Even G2", "close", true),
+      terminalManager: { shutdown: async () => { calls.push("terminals"); }, waitForProcessExits: async () => 0 },
+      usageHistory: { dispose: () => { calls.push("usage history"); return aborted.then(() => { calls.push("usage history saved"); }); } },
+      limitsService: { dispose: () => { calls.push("limits"); limitsAborted(); } },
+      agentGateway: service("agent gateway", "close"),
+      runtimeGateway: service("runtime gateway", "close"),
+      browserService: service("browser", "dispose"),
+      pluginServices: service("plugin services", "dispose"),
+      pluginManager: service("plugins", "dispose"),
+      handoffService: service("handoff", "dispose"),
+      materialService: service("materials", "dispose")
+    });
+    await shutdownServices();
+  } finally {
+    console.warn = warn;
+  }
+  assert.deepEqual(calls, ["mascots", "chat history", "agent control", "Even G2", "terminals", "usage history", "limits",
+    "usage history saved", "agent gateway", "runtime gateway", "browser", "plugin services", "plugins", "handoff", "materials",
+    "diagnostics"]);
+});
diff --git a/tests/subagent-model-account.test.mjs b/tests/subagent-model-account.test.mjs
new file mode 100644
index 000000000..d70f8e39d
--- /dev/null
+++ b/tests/subagent-model-account.test.mjs
@@ -0,0 +1,346 @@
+import assert from "node:assert/strict";
+import { mkdtemp, rm } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import test from "node:test";
+import { AgentControlService } from "../src/main/services/AgentControlService.ts";
+import { ScopedOrchestrationHandler } from "../src/main/services/agent-browser/OrchestrationTools.ts";
+import { LaunchPipeline, modelSettingArgument } from "../src/main/services/LaunchPipeline.ts";
+import { TerminalManager } from "../src/main/services/TerminalManager.ts";
+import { TerminalSessionStore } from "../src/main/services/TerminalSessionStore.ts";
+import { availableRegistry, inputReadySpawner } from "./helpers/terminal.mjs";
+
+// A subagent spawned without launch options once ran on its CLI's own default model (for OpenCode the hosted OpenCode
+// Zen model), while its orchestrator ran on a model the person picked through a plugin (a model account). It now keeps
+// the orchestrator's options for every plugin the spawn does not name; naming a plugin is an explicit choice, and a
+// second model flag next to a plugin's own is refused instead of starting a CLI that crashes.
+
+const at = { x: 0, y: 0 };
+// Any plugin id: the core knows no plugin by name.
+const MODELS = "example-model-accounts";
+const EXTRAS = "example-extras";
+const OTHER = "example-other-models";
+const BARE = "example-no-fields";
+
+async function setup(t, { delegable = true, extrasDelegable = false, store } = {}) {
+  const root = await mkdtemp(join(tmpdir(), "ctty-sub-account-"));
+  t.after(() => rm(root, { recursive: true, force: true }));
+  const calls = [];
+  const prepared = [];
+  // What the person can change while the orchestrator runs: turn a plugin off, or a plugin update changes its fields.
+  const state = { gone: new Set(), extrasFields: [{ key: "flavor", label: "Flavor", kind: "text" }] };
+  const terminals = new TerminalManager(() => undefined, availableRegistry(), undefined, undefined, true, inputReadySpawner(calls));
+  t.after(() => store ? terminals.shutdown() : terminals.disposeAll());
+  terminals.configureLaunchPipeline(new LaunchPipeline({
+    contributors: () => [
+      { pluginId: MODELS, pluginName: "Model accounts", serviceId: "accounts", secrets: false,
+        launch: { fields: [{ key: "account", label: "Model account", kind: "text" }], ...(delegable ? { delegable: true } : {}) } },
+      // Options that do not decide the model.
+      { pluginId: EXTRAS, pluginName: "Extras", serviceId: "extras", secrets: false,
+        launch: { fields: state.extrasFields, ...(extrasDelegable ? { delegable: true } : {}) } },
+      // A plugin without fields: choosing it in the launcher stores {}.
+      { pluginId: BARE, pluginName: "No fields", serviceId: "bare", secrets: false, launch: { fields: [] } },
+      // A second plugin that also sets the model.
+      { pluginId: OTHER, pluginName: "Other models", serviceId: "other", secrets: false,
+        launch: { fields: [{ key: "pick", label: "Model", kind: "text" }], delegable: true } }
+    ].filter((contributor) => !state.gone.has(contributor.pluginId)),
+    call: async (pluginId, _service, method, params) => {
+      if (method !== "canvastty.launch.prepare") return null;
+      prepared.push({ pluginId, sessionId: params.sessionId, chosen: params.chosen, options: params.options });
+      if (pluginId === OTHER && params.chosen) return { env: {}, secretEnv: {}, args: [`--model=other/${params.options.pick}`], files: [] };
+      // An endpoint plugin: it routes the CLI to another vendor's service but passes no model argument.
+      if (pluginId === EXTRAS && params.chosen && params.options.flavor === "endpoint") return { env: {}, secretEnv: {}, args: [], files: [], thirdPartyModel: true };
+      return pluginId === MODELS && params.chosen && params.options.account && params.options.account !== "none"
+        ? { env: {}, secretEnv: {}, args: ["--model", `acct_${params.options.account}/glm-5.3-flash`], files: [] }
+        : { env: {}, secretEnv: {}, args: [], files: [] };
+    },
+    secret: async () => null, runsRoot: join(root, "runs"), timeoutMs: 2000
+  }));
+  if (store) {
+    terminals.configureSessionPersistence(new TerminalSessionStore(store), "continue");
+    await terminals.restorePersistedSessions();
+  }
+  const control = new AgentControlService(terminals);
+  const handler = new ScopedOrchestrationHandler(control, null,
+    { cli: () => "available", limits: () => null, models: () => ({ models: ["zai/glm-5.3"], checkedAt: Date.now() }), checkModel: async () => null },
+    {});
+  const orchestrator = (launchOptions, provider = "opencode") => terminals.create({ provider, profile: "normal", cwd: root, position: at, role: "orchestrator",
+    ...(launchOptions ? { launchOptions } : {}) });
+  const spawn = (parent, args) => handler.execute(parent.id, { id: `c-${Math.random()}`, tool: "spawn_agent", arguments: { provider: "opencode", cwd: root, prompt: "Small task", ...args } });
+  const settle = () => new Promise((resolve) => setTimeout(resolve, 50));
+  return { root, calls, prepared, state, terminals, control, handler, orchestrator, spawn, settle };
+}
+
+test("a subagent spawned without launch options runs on its orchestrator's model account, never on the CLI default", async (t) => {
+  const s = await setup(t);
+  const parent = s.orchestrator({ [MODELS]: { account: "glm-flash" } });
+  await s.settle();
+  assert.deepEqual(s.terminals.modelSetByOf(parent.id), [MODELS]);
+  const child = await s.spawn(parent, {});
+  await s.settle();
+  assert.deepEqual(child.servedBy, { provider: "opencode", model: null, launchOptions: { [MODELS]: "inherited" } });
+  const childPrepare = s.prepared.find((entry) => entry.sessionId === child.sessionId);
+  assert.deepEqual(childPrepare, { pluginId: MODELS, sessionId: child.sessionId, chosen: true, options: { account: "glm-flash" } });
+  const launched = s.calls.at(-1).args;
+  assert.ok(launched.includes("acct_glm-flash/glm-5.3-flash"), launched.join(" "));
+  assert.equal(launched.filter((arg) => arg === "--model").length, 1);
+  assert.deepEqual(s.terminals.launchOptionsOf(child.sessionId), { [MODELS]: { account: "glm-flash" } });
+});
+
+test("naming the plugin is an explicit choice; without any options the CLI's default serves the subagent", async (t) => {
+  const s = await setup(t);
+  const parent = s.orchestrator({ [MODELS]: { account: "glm-flash" } });
+  await s.settle();
+  const own = await s.spawn(parent, { launchOptions: { [MODELS]: { account: "none" } } });
+  await s.settle();
+  assert.deepEqual(own.servedBy.launchOptions, { [MODELS]: "explicit" });
+  assert.deepEqual(s.terminals.launchOptionsOf(own.sessionId), { [MODELS]: { account: "none" } });
+  assert.ok(!s.calls.at(-1).args.some((arg) => arg.startsWith("acct_")), "explicit none: no account model");
+  assert.deepEqual(s.terminals.modelSetByOf(own.sessionId), []);
+
+  const bare = await setup(t);
+  const plain = bare.orchestrator();
+  const child = await bare.spawn(plain, {});
+  await bare.settle();
+  assert.deepEqual(child.servedBy, { provider: "opencode", model: null, launchOptions: {} });
+  assert.equal(bare.terminals.launchOptionsOf(child.sessionId), undefined);
+});
+
+test("another value needs the plugin's delegable declaration; the orchestrator's own values do not", async (t) => {
+  const s = await setup(t, { delegable: false });
+  const parent = s.orchestrator({ [MODELS]: { account: "glm-flash" }, [EXTRAS]: { flavor: "mild" } });
+  await s.settle();
+  const inherited = await s.spawn(parent, {});
+  assert.deepEqual(inherited.servedBy.launchOptions, { [MODELS]: "inherited", [EXTRAS]: "inherited" });
+  assert.deepEqual(s.terminals.launchOptionsOf(inherited.sessionId), { [MODELS]: { account: "glm-flash" }, [EXTRAS]: { flavor: "mild" } });
+  const same = await s.spawn(parent, { launchOptions: { [MODELS]: { account: "glm-flash" } } });
+  assert.deepEqual(same.servedBy.launchOptions, { [MODELS]: "explicit", [EXTRAS]: "inherited" });
+  await assert.rejects(s.spawn(parent, { launchOptions: { [MODELS]: { account: "other-paid" } } }),
+    /Model accounts has not declared its launch options safe for an orchestrator to choose/u);
+});
+
+test("a subagent of another CLI cannot silently drop the options that set the orchestrator's model", async (t) => {
+  const s = await setup(t);
+  const parent = s.orchestrator({ [MODELS]: { account: "glm-flash" }, [EXTRAS]: { flavor: "mild" } });
+  await s.settle();
+  const refused = new RegExp(`model is set by its ${MODELS} launch options, which the person chose for opencode; a codex subagent would not get that model\\. Spawn a opencode subagent instead\\.`, "u");
+  await assert.rejects(s.spawn(parent, { provider: "codex" }), refused);
+  assert.throws(() => s.control.spawn({ parentSessionId: parent.id, provider: "codex", cwd: s.root }), refused);
+  // Naming the plugin for the other CLI is explicit; options that do not decide the model stay with their CLI.
+  const explicit = await s.spawn(parent, { provider: "codex", launchOptions: { [MODELS]: { account: "none" } } });
+  assert.deepEqual(explicit.servedBy.launchOptions, { [MODELS]: "explicit" });
+  assert.deepEqual(s.terminals.launchOptionsOf(explicit.sessionId), { [MODELS]: { account: "none" } });
+});
+
+test("AgentControlService.spawn applies the same inheritance for every caller", async (t) => {
+  const s = await setup(t);
+  const parent = s.orchestrator({ [MODELS]: { account: "glm-flash" } });
+  await s.settle();
+  const child = await s.control.spawn({ parentSessionId: parent.id, provider: "opencode", cwd: s.root });
+  await s.settle();
+  assert.deepEqual(s.terminals.launchOptionsOf(child.id), { [MODELS]: { account: "glm-flash" } });
+  assert.deepEqual(s.terminals.modelSetByOf(child.id), [MODELS]);
+});
+
+test("a subagent with an explicit model next to launch options that set the model is refused instead of a crashing double --model", async (t) => {
+  const s = await setup(t);
+  const parent = s.orchestrator({ [EXTRAS]: { flavor: "mild" } });
+  await s.settle();
+  const spawned = s.calls.length;
+  const child = await s.control.spawn({ parentSessionId: parent.id, provider: "opencode", cwd: s.root, model: "zai/glm-5.3",
+    launchOptions: { [MODELS]: { account: "glm-flash" } } });
+  await s.settle();
+  const failed = s.terminals.getMetadata(child.id);
+  assert.equal(failed.status, "failed");
+  assert.match(failed.failureDetails, /Model accounts sets the model of this launch, and the launch also asks for model zai\/glm-5\.3/u);
+  assert.equal(s.calls.length, spawned, "no CLI was started with two --model flags");
+
+  // Options that do not set the model take an explicit model as before.
+  const plain = await s.control.spawn({ parentSessionId: parent.id, provider: "opencode", cwd: s.root, model: "zai/glm-5.3" });
+  await s.settle();
+  assert.notEqual(s.terminals.getMetadata(plain.id).status, "failed");
+  assert.equal(s.calls.at(-1).args.filter((arg) => arg === "--model").length, 1);
+
+  // A subagent that would inherit them is refused before anything starts.
+  const accounts = s.orchestrator({ [MODELS]: { account: "glm-flash" } });
+  await s.settle();
+  await assert.rejects(s.spawn(accounts, { model: "zai/glm-5.3" }),
+    new RegExp(`inherits its orchestrator's ${MODELS} launch options, which set its model; it cannot also take model zai/glm-5\\.3`, "u"));
+});
+
+test("a subagent whose launch options come from two plugins that both set the model is refused", async (t) => {
+  const s = await setup(t);
+  const parent = s.orchestrator();
+  await s.settle();
+  const spawned = s.calls.length;
+  const child = await s.control.spawn({ parentSessionId: parent.id, provider: "opencode", cwd: s.root,
+    launchOptions: { [MODELS]: { account: "glm-flash" }, [OTHER]: { pick: "x" } } });
+  await s.settle();
+  const failed = s.terminals.getMetadata(child.id);
+  assert.equal(failed.status, "failed");
+  assert.match(failed.failureDetails, /Model accounts and Other models both set the model of this launch/u);
+  assert.equal(s.calls.length, spawned);
+});
+
+test("a person's own card keeps starting with a model and a plugin that sets the model, also when it is restored", async (t) => {
+  const s = await setup(t);
+  const card = s.terminals.create({ provider: "opencode", profile: "normal", cwd: s.root, position: at, model: "zai/glm-5.3",
+    launchOptions: { [MODELS]: { account: "glm-flash" } } });
+  await s.settle();
+  assert.notEqual(s.terminals.getMetadata(card.id).status, "failed");
+  assert.ok(s.calls.at(-1).args.includes("acct_glm-flash/glm-5.3-flash"));
+
+  const store = await mkdtemp(join(tmpdir(), "ctty-sub-account-store-"));
+  t.after(() => rm(store, { recursive: true, force: true }));
+  await new TerminalSessionStore(store).replace([{
+    id: "person-card", provider: "opencode", profile: "normal", role: "agent", title: "Agent", titleCustomized: false,
+    cwd: s.root, position: at, size: { width: 700, height: 430 }, lastState: "running", restore: true,
+    model: "zai/glm-5.3", options: { [MODELS]: { account: "glm-flash" } }
+  }]);
+  const restored = await setup(t, { store });
+  await restored.settle();
+  const metadata = restored.terminals.getMetadata("person-card");
+  assert.notEqual(metadata.status, "failed", metadata.failureDetails ?? "");
+  assert.equal(restored.calls.length, 1);
+  assert.ok(restored.calls[0].args.includes("acct_glm-flash/glm-5.3-flash"));
+  assert.deepEqual(restored.terminals.modelSetByOf("person-card"), [MODELS]);
+});
+
+test("a subagent of another CLI is never checked against options the person chose for its orchestrator's CLI", async (t) => {
+  const s = await setup(t, { delegable: false });
+  // EXTRAS is not delegable and does not set the model: an opencode orchestrator's choice is not a codex subagent's.
+  const parent = s.orchestrator({ [EXTRAS]: { flavor: "mild" } });
+  await s.settle();
+  const child = await s.spawn(parent, { provider: "codex" });
+  assert.deepEqual(child.servedBy.launchOptions, {});
+  assert.equal(s.terminals.launchOptionsOf(child.sessionId), undefined);
+  // Naming it for codex is the orchestrator's own choice, which this plugin does not allow.
+  await assert.rejects(s.spawn(parent, { provider: "codex", launchOptions: { [EXTRAS]: { flavor: "mild" } } }),
+    /Extras has not declared its launch options safe for an orchestrator to choose/u);
+});
+
+test("model-setting arguments are recognised in every form each CLI takes", () => {
+  const sets = (provider, ...args) => modelSettingArgument(provider, args);
+  for (const provider of ["opencode", "codex", "qwen", "kimi", "grok", "claude", "cursor"]) {
+    assert.equal(sets(provider, "--model", "x"), "--model", provider);
+    assert.equal(sets(provider, "--model=x"), "--model=x", provider);
+  }
+  for (const provider of ["opencode", "codex", "qwen", "kimi", "grok"]) {
+    assert.equal(sets(provider, "-m", "x"), "-m", provider);
+    assert.equal(sets(provider, "-mx"), "-mx", provider);
+  }
+  // Claude has no -m.
+  assert.equal(sets("claude", "-m", "x"), undefined);
+  // Codex: a profile, and configuration keys that pick the model or its provider.
+  assert.equal(sets("codex", "--profile", "work"), "--profile");
+  assert.equal(sets("codex", "--profile=work"), "--profile=work");
+  assert.equal(sets("codex", "-p", "work"), "-p");
+  assert.equal(sets("codex", "-pwork"), "-pwork");
+  assert.equal(sets("codex", "-c", "model=\"o3\""), "-c model=\"o3\"");
+  assert.equal(sets("codex", "-c", " model = o3"), "-c  model = o3");
+  assert.equal(sets("codex", "--config", "model_provider=acct"), "--config model_provider=acct");
+  assert.equal(sets("codex", "--config=profile=work"), "--config=profile=work");
+  assert.equal(sets("codex", "-cmodel=o3"), "-cmodel=o3");
+  assert.equal(sets("codex", "-c", "\"model\"=o3"), "-c \"model\"=o3");
+  // Other configuration keys and flags do not choose the model.
+  assert.equal(sets("codex", "-c", "model_reasoning_effort=high"), undefined);
+  assert.equal(sets("codex", "-c", "model"), undefined);
+  assert.equal(sets("codex", "--search", "-c", "tools.web_search=true"), undefined);
+  assert.equal(sets("opencode", "--profile", "x", "-p", "y"), undefined);
+  assert.equal(sets("opencode", "--models"), undefined);
+});
+
+test("without the delegable declaration a subagent gets exactly the inherited values, never the plugin's defaults", () => {
+  const pipeline = new LaunchPipeline({
+    contributors: () => [{ pluginId: EXTRAS, pluginName: "Extras", serviceId: "extras", secrets: false,
+      launch: { fields: [{ key: "a", label: "A", kind: "text", default: "da" }, { key: "b", label: "B", kind: "text", default: "db" },
+        { key: "c", label: "C", kind: "text" }] } }],
+    call: async () => null, secret: async () => null, runsRoot: join(tmpdir(), "ctty-unused-runs")
+  });
+  const inherited = { [EXTRAS]: { a: "1", b: "2", c: "" } };
+  const delegated = { delegated: true, inherited };
+  // A subset is completed from the orchestrator's values, not from the defaults.
+  assert.deepEqual(pipeline.normalizeOptions("opencode", { [EXTRAS]: { a: "1" } }, delegated), { [EXTRAS]: { a: "1", b: "2", c: "" } });
+  assert.deepEqual(pipeline.normalizeOptions("opencode", { [EXTRAS]: { a: "1", b: "2", c: "" } }, delegated), inherited);
+  const refused = /Extras has not declared its launch options safe for an orchestrator to choose/u;
+  assert.throws(() => pipeline.normalizeOptions("opencode", { [EXTRAS]: { a: "1", b: "x" } }, delegated), refused);
+  assert.throws(() => pipeline.normalizeOptions("opencode", { [EXTRAS]: { c: "set" } }, delegated), refused);
+  assert.throws(() => pipeline.normalizeOptions("opencode", { [EXTRAS]: { a: "1", extra: "1" } }, delegated), refused);
+  assert.throws(() => pipeline.normalizeOptions("opencode", { [EXTRAS]: {} }, delegated), refused);
+  // An orchestrator whose own values the plugin no longer accepts as they were is refused, not given defaults.
+  assert.throws(() => pipeline.normalizeOptions("opencode", { [EXTRAS]: { a: "1" } }, { delegated: true, inherited: { [EXTRAS]: { a: "1", b: "2" } } }),
+    /Extras launch options changed since this orchestrator started/u);
+  assert.throws(() => pipeline.normalizeOptions("opencode", { [EXTRAS]: { a: "1" } }, { delegated: true }), refused);
+});
+
+test("a plugin without fields chosen for the orchestrator is inherited as chosen, not refused", async (t) => {
+  const s = await setup(t, { delegable: false });
+  const parent = s.orchestrator({ [BARE]: {} });
+  await s.settle();
+  assert.notEqual(s.terminals.getMetadata(parent.id).status, "failed");
+  const child = await s.spawn(parent, {});
+  await s.settle();
+  assert.deepEqual(child.servedBy.launchOptions, { [BARE]: "inherited" });
+  assert.deepEqual(s.terminals.launchOptionsOf(child.sessionId), { [BARE]: {} });
+  assert.notEqual(s.terminals.getMetadata(child.sessionId).status, "failed");
+});
+
+test("a plugin turned off after the orchestrator started: left out unless it set the model, then refused", async (t) => {
+  const s = await setup(t, { delegable: false });
+  const parent = s.orchestrator({ [MODELS]: { account: "glm-flash" }, [EXTRAS]: { flavor: "mild" } });
+  await s.settle();
+  // Options that do not decide the model: the subagent starts without them.
+  s.state.gone.add(EXTRAS);
+  const child = await s.spawn(parent, {});
+  await s.settle();
+  assert.deepEqual(child.servedBy.launchOptions, { [MODELS]: "inherited" });
+  assert.deepEqual(s.terminals.launchOptionsOf(child.sessionId), { [MODELS]: { account: "glm-flash" } });
+  // The plugin that set the model: never a silent fall back to the CLI's default model.
+  s.state.gone.add(MODELS);
+  const spawned = s.calls.length;
+  await assert.rejects(s.spawn(parent, {}),
+    new RegExp(`model is set by its ${MODELS} launch options, and ${MODELS} is now disabled, removed or not trusted; a subagent would run on its CLI's own default model instead`, "u"));
+  assert.equal(s.calls.length, spawned);
+});
+
+test("a plugin whose fields changed after the orchestrator started refuses its subagents with that reason", async (t) => {
+  const s = await setup(t, { delegable: false });
+  const parent = s.orchestrator({ [EXTRAS]: { flavor: "mild" } });
+  await s.settle();
+  s.state.extrasFields.splice(0, 1, { key: "taste", label: "Taste", kind: "text" });
+  await assert.rejects(s.spawn(parent, {}), /Extras launch options changed since this orchestrator started, so its subagents cannot repeat the person's choice/u);
+  s.state.extrasFields.splice(0, 1, { key: "flavor", label: "Flavor", kind: "text" }, { key: "extra", label: "Extra", kind: "text", default: "on" });
+  await assert.rejects(s.spawn(parent, {}), /Extras launch options changed since this orchestrator started/u);
+});
+
+test("a subagent inherits a person's model plugin next to an endpoint plugin that passes no model argument", async (t) => {
+  const s = await setup(t, { delegable: false });
+  const parent = s.orchestrator({ [MODELS]: { account: "glm-flash" }, [EXTRAS]: { flavor: "endpoint" } });
+  await s.settle();
+  assert.notEqual(s.terminals.getMetadata(parent.id).status, "failed", "the person's own orchestrator starts");
+  const child = await s.spawn(parent, {});
+  await s.settle();
+  const metadata = s.terminals.getMetadata(child.sessionId);
+  assert.notEqual(metadata.status, "failed", metadata.failureDetails ?? "");
+  assert.equal(s.calls.at(-1).args.filter((arg) => arg === "--model").length, 1);
+});
+
+test("an endpoint plugin and a model plugin the orchestrator picks itself are still refused as two models", async (t) => {
+  // The person chose only the endpoint; the orchestrator adds a model account next to it.
+  const s = await setup(t);
+  const parent = s.orchestrator({ [EXTRAS]: { flavor: "endpoint" } });
+  await s.settle();
+  const spawned = s.calls.length;
+  await assert.rejects(s.spawn(parent, { launchOptions: { [MODELS]: { account: "glm-flash" } } }),
+    /Extras and Model accounts both set the model of this launch|Model accounts and Extras both set the model of this launch/u);
+  assert.equal(s.calls.length, spawned, "nothing started");
+
+  // The person chose only the model account; the orchestrator adds an endpoint next to it.
+  const e = await setup(t, { extrasDelegable: true });
+  const accounts = e.orchestrator({ [MODELS]: { account: "glm-flash" } });
+  await e.settle();
+  const before = e.calls.length;
+  await assert.rejects(e.spawn(accounts, { launchOptions: { [EXTRAS]: { flavor: "endpoint" } } }),
+    /Extras and Model accounts both set the model of this launch|Model accounts and Extras both set the model of this launch/u);
+  assert.equal(e.calls.length, before, "nothing started");
+});
diff --git a/tests/subagent-signal-exit.test.mjs b/tests/subagent-signal-exit.test.mjs
new file mode 100644
index 000000000..c434e05c9
--- /dev/null
+++ b/tests/subagent-signal-exit.test.mjs
@@ -0,0 +1,118 @@
+import assert from "node:assert/strict";
+import { execFileSync } from "node:child_process";
+import { existsSync } from "node:fs";
+import { mkdtemp, rm } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { basename, join } from "node:path";
+import test from "node:test";
+import pty from "node-pty";
+import { AgentControlService } from "../src/main/services/AgentControlService.ts";
+import { TerminalManager } from "../src/main/services/TerminalManager.ts";
+import { persistedTerminalSession } from "../src/main/services/TerminalSessionStore.ts";
+import { failureSummary } from "../src/renderer/src/features/home/failureSummary.ts";
+import { availableRegistry } from "./helpers/terminal.mjs";
+
+// A subagent whose CLI was killed with SIGKILL showed "done, exit 0", so wait_for_agent reported a crash as success:
+// node-pty reports a signal death as { exitCode: 0, signal: 9 }. These tests start real processes (inside sandbox-exec on macOS, as the
+// isolation layer wraps an agent there) and kill them.
+
+const at = { x: 0, y: 0 };
+const SANDBOX = "/usr/bin/sandbox-exec";
+const wrapped = process.platform === "darwin" && existsSync(SANDBOX);
+const skip = process.platform === "win32" ? "POSIX signals only" : false;
+
+/** Every launch becomes the given real command tree, wrapped in sandbox-exec where the isolation layer would. */
+function realSpawner(argv, spawned) {
+  return (_command, _args, options) => {
+    const [command, ...args] = wrapped ? [SANDBOX, "-p", "(version 1)(allow default)", ...argv] : argv;
+    const child = pty.spawn(command, args, { name: "xterm-256color", cols: 80, rows: 24, cwd: options.cwd, env: { PATH: "/usr/bin:/bin", HOME: options.cwd } });
+    spawned.push(child);
+    return child;
+  };
+}
+
+async function setup(t, argv) {
+  const root = await mkdtemp(join(tmpdir(), "ctty-signal-exit-"));
+  const spawned = [];
+  const terminals = new TerminalManager(() => undefined, availableRegistry(), undefined, undefined, true, realSpawner(argv, spawned));
+  t.after(async () => {
+    terminals.disposeAll();
+    for (const child of spawned) { try { process.kill(child.pid, "SIGKILL"); } catch { /* exited */ } }
+    await rm(root, { recursive: true, force: true });
+  });
+  const control = new AgentControlService(terminals);
+  const orchestrator = terminals.create({ provider: "opencode", profile: "normal", cwd: root, position: at, role: "orchestrator" });
+  return { root, spawned, terminals, control, orchestrator };
+}
+
+async function until(check, timeoutMs = 5_000) {
+  const deadline = Date.now() + timeoutMs;
+  while (Date.now() < deadline) {
+    const value = check();
+    if (value) return value;
+    await new Promise((resolve) => setTimeout(resolve, 25));
+  }
+  throw new Error("timed out");
+}
+
+const descendants = (pid) => {
+  try { return execFileSync("/usr/bin/pgrep", ["-P", String(pid)], { encoding: "utf8" }).trim().split("\n").filter(Boolean).map(Number); } catch { return []; }
+};
+
+test("a subagent whose CLI is killed by SIGKILL is failed (128+9) with the signal named, everywhere its exit is read", { skip }, async (t) => {
+  const s = await setup(t, ["/bin/sleep", "30"]);
+  const child = await s.control.spawn({ parentSessionId: s.orchestrator.id, provider: "opencode", cwd: s.root });
+  const live = s.spawned.at(-1);
+  // sandbox-exec execs the CLI: the PTY's own process is the CLI, not a wrapper that could exit 0 after it.
+  // (node-pty's spawn-helper and then sandbox-exec exec in place first.)
+  // macOS reports an executable path; Linux ps reports only the command name.
+  const command = await until(() => { const comm = execFileSync("/bin/ps", ["-o", "comm=", "-p", String(live.pid)], { encoding: "utf8" }).trim(); return basename(comm) === "sleep" ? comm : null; });
+  assert.equal(basename(command), "sleep");
+  assert.deepEqual(descendants(live.pid), [], "no process below the CLI");
+  process.kill(live.pid, "SIGKILL");
+  const ended = await until(() => { const m = s.terminals.getMetadata(child.id); return m.exitCode !== null ? m : null; });
+  assert.equal(ended.status, "failed");
+  assert.equal(ended.exitCode, 137);
+  assert.match(ended.failureDetails, /killed by signal SIGKILL \(9\)/u);
+  // wait_for_agent and get_agent_result read the same exit.
+  const waited = await s.control.waitFor(child.id, { timeoutMs: 1_000 });
+  assert.equal(waited.reason, "failed");
+  assert.equal(waited.status, "failed");
+  assert.equal(waited.exitCode, 137);
+  const result = await s.control.result(child.id);
+  assert.equal(result.state, "failed");
+  assert.equal(result.exitCode, 137);
+  // The saved session restores as failed, and HOME's failure summary shows the exit code with the signal line.
+  const saved = persistedTerminalSession(ended);
+  assert.equal(saved.lastState, "failed");
+  assert.equal(saved.exitCode, 137);
+  const summary = failureSummary(ended, "en");
+  assert.match(summary.headline, /137/u);
+  assert.match(summary.diagnostics, /^The process was killed by signal SIGKILL \(9\)\./u);
+});
+
+test("a wrapper shell whose CLI child is killed reports the signal as a non-zero exit: the subagent is failed", { skip }, async (t) => {
+  const s = await setup(t, ["/bin/sh", "-c", "/bin/sleep 30; exit $?"]);
+  const child = await s.control.spawn({ parentSessionId: s.orchestrator.id, provider: "opencode", cwd: s.root });
+  const shell = s.spawned.at(-1);
+  const cli = await until(() => descendants(shell.pid)[0]);
+  process.kill(cli, "SIGKILL");
+  const ended = await until(() => { const m = s.terminals.getMetadata(child.id); return m.exitCode !== null ? m : null; });
+  assert.equal(ended.status, "failed");
+  assert.equal(ended.exitCode, 137);
+  // The shell exited normally with 137: no signal reached node-pty, so nothing is named.
+  assert.doesNotMatch(ended.failureDetails ?? "", /killed by signal/u);
+  assert.equal((await s.control.waitFor(child.id, { timeoutMs: 1_000 })).reason, "failed");
+});
+
+test("a clean exit stays done", { skip }, async (t) => {
+  const s = await setup(t, ["/usr/bin/true"]);
+  const child = await s.control.spawn({ parentSessionId: s.orchestrator.id, provider: "opencode", cwd: s.root });
+  const ended = await until(() => { const m = s.terminals.getMetadata(child.id); return m.exitCode !== null ? m : null; });
+  assert.equal(ended.status, "done");
+  assert.equal(ended.exitCode, 0);
+  assert.equal(ended.failureDetails, null);
+  assert.equal((await s.control.waitFor(child.id, { timeoutMs: 1_000 })).reason, "done");
+  assert.equal((await s.control.result(child.id)).state, "done");
+  assert.equal(persistedTerminalSession(ended).lastState, "exited");
+});
diff --git a/tests/terminal-lifecycle.test.mjs b/tests/terminal-lifecycle.test.mjs
index 8ca7f9c16..fc1b3e545 100644
--- a/tests/terminal-lifecycle.test.mjs
+++ b/tests/terminal-lifecycle.test.mjs
@@ -290,15 +290,31 @@ test("session metadata revisions advance before lifecycle events cross IPC", asy
   assert.match(source, /metadata\.revision \+= 1;[\s\S]{0,120}?this\.emit\(IPC\.terminalSession/);
 });
 
-test("revoking lifecycle hooks makes live agent status unavailable until a restarted session gets a new parser", async () => {
+test("revoking lifecycle hooks makes live agent status unavailable until a restarted session gets a new parser", async (t) => {
   const source = await readFile(terminalManagerPath, "utf8");
-
-  assert.match(source, /setLifecycleHooksEnabled\(enabled: boolean\): void/);
-  assert.match(source, /this\.lifecycleHooksEnabled = next;\s*if \(next\) return;/);
-  assert.match(source, /session\.lifecycle = null;/);
-  assert.match(source, /session\.metadata\.provider === "terminal"/);
-  assert.match(source, /session\.metadata\.status = "unavailable";\s*this\.emitSession\(session\.metadata\)/);
-  assert.match(source, /if \(!this\.lifecycleHooksEnabled \|\| !session/);
+  const { TerminalManager } = await import("../src/main/services/TerminalManager.ts");
+  const { availableRegistry, fakeSpawner } = await import("./helpers/terminal.mjs");
+  const manager = new TerminalManager(() => undefined, availableRegistry(), undefined, undefined, true, fakeSpawner([]));
+  t.after(() => manager.disposeAll());
+  const session = manager.create({ provider: "codex", profile: "normal", cwd: process.cwd(), position: { x: 0, y: 0 } });
+  const shell = manager.create({ provider: "terminal", profile: "normal", cwd: process.cwd(), position: { x: 0, y: 0 } });
+  const shellStatus = manager.getMetadata(shell.id).status;
+  manager.applyProviderSignal(session.id, { kind: "lifecycle", state: "working" });
+  assert.equal(manager.getMetadata(session.id).status, "working");
+  manager.setLifecycleHooksEnabled(false);
+  assert.equal(manager.getMetadata(session.id).status, "unavailable");
+  assert.equal(manager.sessions.get(session.id).lifecycle, null);
+  assert.equal(manager.getMetadata(shell.id).status, shellStatus, "a plain terminal has no agent status to revoke");
+  for (const state of ["needs_approval", "working", "idle"]) {
+    manager.applyProviderSignal(session.id, { kind: "lifecycle", state });
+    assert.equal(manager.getMetadata(session.id).status, "unavailable", "a hook signal never restores a disabled status");
+  }
+  manager.setLifecycleHooksEnabled(true);
+  manager.applyProviderSignal(session.id, { kind: "lifecycle", state: "working" });
+  assert.equal(manager.getMetadata(session.id).status, "unavailable", "re-enabling applies only to new or restarted sessions");
+  const fresh = manager.create({ provider: "codex", profile: "normal", cwd: process.cwd(), position: { x: 0, y: 0 } });
+  manager.applyProviderSignal(fresh.id, { kind: "lifecycle", state: "working" });
+  assert.equal(manager.getMetadata(fresh.id).status, "working");
   assert.match(source, /session\.lifecycle = this\.lifecycleHooksEnabled\s*\? createProviderLifecycleParser/);
 });
 
@@ -306,8 +322,9 @@ test("only explicit turn-stop hooks set completed art and a new turn clears it",
   const [card, manager, main] = await Promise.all([
     readFile(terminalCardPath, "utf8"),
     readFile(terminalManagerPath, "utf8"),
-    readFile(new URL("../src/main/index.ts", import.meta.url), "utf8")
+    readFile(new URL("../src/main/services/agent-runtime/applyRuntimeSignal.ts", import.meta.url), "utf8")
   ]);
+  // main/index.ts hands every runtime signal to applyRuntimeSignal, which passes the hook's event on.
   assert.match(main, /state: signal\.state,\s*event: signal\.event/);
   assert.match(manager, /\["Stop", "StopFailure", "StopCancelled"\]\.includes\(signal\.event \?\? ""\)/);
   assert.match(manager, /nextStatus === "working" \? false : completed \|\| Boolean\(session\.metadata\.turnCompleted\)/);
@@ -377,7 +394,6 @@ test("failed PTYs preserve their final sanitized output as failure details", asy
 
   // Masked whole before the last lines are chosen (a cut inside a secret would leave its tail readable).
   assert.match(source, /terminalFailureDetails\(this\.redactSecrets\(current\.bufferChunks\.slice\(current\.bufferStart\)\.join\(""\)\)\)/);
-  assert.match(source, /current\.metadata\.failureDetails = exitCode === 0/);
 });
 
 function effectDependenciesContaining(source, marker) {
diff --git a/tests/terminal-session-store.test.mjs b/tests/terminal-session-store.test.mjs
index ab8dc7e06..fc778fecf 100644
--- a/tests/terminal-session-store.test.mjs
+++ b/tests/terminal-session-store.test.mjs
@@ -29,7 +29,8 @@ test("terminal window descriptors persist atomically without scrollback or envir
     const store = new TerminalSessionStore(directory);
     await store.replace([descriptor]);
     const raw = await readFile(store.filePath, "utf8");
-    assert.doesNotMatch(raw, /buffer|environment|capability|token/u);
+    // The checkout folder is the descriptor's cwd; leave it out so a path like ~/environment/... cannot match.
+    assert.doesNotMatch(raw.split(JSON.stringify(descriptor.cwd)).join('""'), /buffer|environment|capability|token/u);
 
     const reloaded = new TerminalSessionStore(directory);
     assert.deepEqual(await reloaded.load(), [descriptor]);
diff --git a/tests/ui-scale-chrome.test.mjs b/tests/ui-scale-chrome.test.mjs
new file mode 100644
index 000000000..28230b88c
--- /dev/null
+++ b/tests/ui-scale-chrome.test.mjs
@@ -0,0 +1,58 @@
+// UI chrome follows the root --ui-scale (docs/UI_CONTRACT.md): the card options menu, the canvas
+// shortcut-hints panel and the Settings → Plugins cards must not keep fixed pixel text sizes.
+import assert from "node:assert/strict";
+import { readFile } from "node:fs/promises";
+import test from "node:test";
+
+const css = await readFile(new URL("../src/renderer/src/styles/app.css", import.meta.url), "utf8");
+const canvas = await readFile(new URL("../src/renderer/src/features/workspace/WorkspaceCanvas.tsx", import.meta.url), "utf8");
+const escape = (text) => text.replace(/[.*+?^${}()|[\]\\]/gu, "\\$&");
+const rule = (selector) => css.match(new RegExp(`^${escape(selector)} \\{[^}]*\\}`, "mu"))?.[0] ?? "";
+const declaration = (block, property) => block.match(new RegExp(`(?:^|[;{ ])${property}: ([^;]+);`, "u"))?.[1];
+const scaledSource = String.raw`calc\(\d+(?:\.\d+)?px \* var\(--ui-scale, 1\)\)`;
+
+test("the card options menu scales its font with --ui-scale and sizes everything else from it", () => {
+  const menu = rule(".terminal-card__menu");
+  assert.match(menu, /font-size: calc\(12px \* var\(--ui-scale, 1\)\)/u);
+  for (const selector of [".terminal-card__menu", ".terminal-card__menu label", ".terminal-card__menu-action"]) {
+    const block = rule(selector);
+    assert.ok(block, selector);
+    for (const property of ["padding", "gap", "border-radius", "right"]) {
+      const value = declaration(block, property);
+      if (value) assert.doesNotMatch(value, /\dpx/u, `${selector} ${property} follows the menu's font size`);
+    }
+  }
+});
+
+test("the shortcut-hints panel scales its font with --ui-scale and sizes everything else from it", () => {
+  assert.match(rule(".shortcut-hints"), /font-size: calc\(11px \* var\(--ui-scale, 1\)\)/u);
+  for (const selector of [".shortcut-hints", ".shortcut-hints > div", ".shortcut-hints kbd", ".shortcut-hints span", ".shortcut-hints__more"]) {
+    const block = rule(selector);
+    assert.ok(block, selector);
+    for (const property of ["min-width", "padding", "gap", "border-radius", "grid-template-columns", "font", "font-size"]) {
+      const value = declaration(block, property);
+      if (value) assert.doesNotMatch(value.replace(new RegExp(scaledSource, "gu"), ""), /\dpx/u, `${selector} ${property} follows the panel's font size`);
+    }
+  }
+  assert.match(canvas, /className="shortcut-hints__more"[\s\S]{0,300}?<UiIcon name="app-window" size="[\d.]+em" \/>/u,
+    "the Keyboard shortcuts icon scales with the panel text");
+});
+
+test("Settings → Plugins text in the installed, preview, permission and contribution cards follows --ui-scale", () => {
+  const prefixes = [".plugin-install-row", ".plugin-preview", ".plugin-permissions", ".plugin-hook-warning", ".plugin-modules", ".installed-plugin", ".plugin-contribution"];
+  const blocks = css.split("\n").filter((line) => prefixes.some((prefix) => line.startsWith(prefix)) && line.includes("{"));
+  assert.ok(blocks.length > 20, "plugin settings rules were found");
+  let checked = 0;
+  for (const block of blocks) {
+    for (const property of ["font", "font-size"]) {
+      const value = declaration(block, property);
+      if (!value) continue;
+      checked += 1;
+      assert.doesNotMatch(value.replace(new RegExp(scaledSource, "gu"), ""), /\dpx/u, `${block.slice(0, block.indexOf("{")).trim()} ${property} follows --ui-scale`);
+    }
+  }
+  assert.ok(checked > 15, "plugin settings font sizes were checked");
+  for (const selector of [".plugin-permissions", ".plugin-contribution > span strong", ".plugin-contribution > span small"]) {
+    assert.match(declaration(rule(selector), "font-size") ?? "", new RegExp(scaledSource, "u"), selector);
+  }
+});