[LukasDoubleU]

Part of #13

T4 Code now builds on upstream v0.1.9 (24e77473ec). The branch merges the tag with --no-ff, so its history keeps t4 and v0.1.9 as parents, then adds the ultracode fix, a stale generated file, the upkeep bookkeeping and the fixes for code that merged cleanly but no longer compiled.

Upstream range

• v0.1.8..v0.1.9: 198 upstream commits.
• Migrations: upstream adds 081_ProjectionMessageSearchText. The merged list is exactly upstream's, plus the fork's in-place change to 069_ThreadMessageSearch. The fork has no migration ids of its own, so nothing collides.
• Workflows: upstream adds .github/workflows/mobile-dev-client.yml (new) and changes ci.yml, mobile-eas-preview-fork.yml, mobile-eas-preview.yml, mobile-eas-production.yml and mobile-showcase-screenshots.yml. None are touched here. The new workflow has a pull_request trigger for its own path, so this PR may start a run; both its jobs require github.repository == 'pingdotgg/t3code' and skip on the fork. It still needs disabling after the push.

Commits

• Merge upstream v0.1.9: 32 conflicted files, resolved below.
• fix(mobile): the Catppuccin theme gets the guest accent color: regenerating the mobile theme variables (as the upkeep regenerates generated files) adds the fuchsia accent to the fork's Catppuccin theme. It was already missing on t4; guest messages, guest names in the queue and the Request status use it. routeTree.gen.ts regenerates identically; highlights.generated.ts is upstream's.
• fix(server): busy accounts recognize the ultracode toggle: see below.
• chore(t4): build on v0.1.9 and version 0.1.10-t4.N: upstream.json (base: "v0.1.9"), the t4-release.yml example and the README. First release: t4-v0.1.10-t4.1. The release order test gains a two-digit patch.
• fix: the fork's catch up, palette and minimap build on v0.1.9: code that merged without conflicts but failed to typecheck; listed under Semantic conflicts.

Conflicts and how they were resolved

Default: upstream's code, with the fork's behavior re-applied in the smallest form.

Thread search and migration 081 (#22553)

• messageSearchIndex.ts: upstream's search_text column, readable-text triggers and recreateMessageSearchTriggers are taken. The fork's indexedMessageCondition (every finished user and assistant message, including user-role ones agents, schedules or the system wrote) stays, so the update trigger keeps the fork's form without the createdBy comparison, now on upstream's readable text.
• ThreadSearch.ts: the fork's sentBy column next to upstream's messageSearchText.
• How 081 meets the fork's index: 081 adds search_text, recreates the triggers from the current code (so with the fork's condition) and backfills search_text only for messages holding a citation link. It doesn't rebuild the index: rows already indexed stay, and backfilled rows are re-indexed by the update trigger under the fork's condition. On a fresh database, 069 and 074 create the triggers with the fork's condition and 081 recreates them the same way. t3 doctor's rebuild uses the same condition.
• Test: the 081 migration test now also inserts an agent-written plain message and an agent-written quote at migration 80; after 81 both are found, the quote by its readable text.

Account switching (#22755)

• Orchestrator.ts merged cleanly apart from wakeWorkStartedAt, which now takes the projection (upstream) next to the fork's origin: dispatchedRunOrigin(...).
• Upstream's OrchestratorProviderSelectionBusyError refuses thread.model-selection.set and provider.switch while a run is active. The fork's failover moves a thread with a message.dispatch on the next instance once its turn is over, so the refusal never applies to it; the fork's SWITCH_ENDS_BACKGROUND_WORK refusal is unchanged. Upstream's runtimeLayer.test.ts dropped two rows that changed instances mid-run; the fork's tests (AccountFailover, UsageLimitRecoveryWorker, ProviderSwitch.integration, ClaudeSessionCarry, ProviderSessionTransitionPolicy) pass unchanged.
• ProviderTurnStartService.ts: the fork's providerMessageText (citations, reply references, reactions) with upstream's OpenCode native goals (#33758). The Claude session carry (feat(server): account moves keep the Claude session instead of a summary #151) merged cleanly.

OrchestratorMcpService.ts

• Upstream's #33295 makes loadCaller and loadScopedThread read only the fields a tool names. The fork's pinned-thread listing passes []; send to thread passes ["runs"] and keeps the fresh-start refusal. nextAccount (ranked among the caller's allowed instances, busy instances passed in) merged cleanly.

linkRequest.ts (#34139)

• Upstream now drops loopback addresses from every machine's list, including the Connect address, which is what the fork's own filter did. The file is now upstream's exactly; the fork's test for it stays and passes.

Sidebar, palette, shortcuts

• Sidebar.tsx: upstream moved the drag ghost out and dropped useTerminalFocus (#33129); the fork's thread memory imports and its New thread dialog for Add project stay. Upstream's snooze detail labels are applied to the fork's menu too.
• threadActionMenu.logic.ts: the fork's menu with More stays; its snooze presets use upstream's label plus detail (#24945).
• chatRouteCommands.ts and its test: Ctrl+N keeps opening the fork's new thread dialog; New thread here uses upstream's last-used project (readLastUsedProjectKey).
• CommandPalette.tsx: both sides' React imports.
• UsageLimitsPooled.tsx: upstream's reset credit choices with the fork's UsageNextAccount.
• Sidebar.browser.test.tsx, fakeEnvironment.ts: both sides' mocks and fields.

Notifications

• ThreadNotificationCoordinator.logic.ts: the fork's observeThreadNotification takes upstream's request ids and Monitor wake grace (#32520, #32719): attention keys carry the request id, a Monitor's held completion is not settled, and backgroundWake is passed. Needs you still replaces the completion it arrives with, and runs the user didn't start stay quiet. The fork's tests now expect the request id in the key.

Citations, timeline, minimap

• assistantCitationNavigation.ts: upstream's side chat routing (#33160), widened to the fork's reply references. The fork's useAssistantCitationSourceLink (used by the chip, sent quotes and reply references) now goes through assistantCitationRoute, so every source link opens a side chat's quote on its parent.
• assistantCitations.ts (contracts, shared): upstream's 4096-character ids (#28084) apply to reply references too, and the href limit uses upstream's formula for both link kinds.
• MessagesTimeline.tsx: the fork's mark capacity and tall-reply refresh next to upstream's restore refresh; upstream's chevrons replace MorphIcon (#33985) next to the fork's reply time.
• TimelineMinimap.tsx: the fork's grouped marks and previews with upstream's display-only preview, at-end Next (#33038) and hit strip (#32151).
• TimelineUserMessageRow.tsx: LinkIcon (#33985), emoticons next to upstream's ordered-list remark plugin (#33820).

Other

• ThreadDetailScreen.tsx (mobile): upstream's column width wrapper plus the fork's lastVisitedAt.
• DesktopApp.test.ts, DesktopWindow.test.ts: both sides' mocks and tests.
• agentAwareness.test.ts: both sides' shell fields.
• scripts/install.test.ts: upstream's 504 retry rows with the fork's 1.2.3-t4.1 release.
• docs/user/keybindings.md: both sides' commands. docs/user/source-control.md: the fork's clone text plus upstream's pasted page URLs, which the fork's clone dialog gets through normalizePastedCloneUrl.

Semantic conflicts (merged cleanly, failed to typecheck)

• TextGeneration.ts: upstream's generators pass the project to their fallback (#33741). Catch me up now passes the thread's project as well (ThreadCatchUpGenerationInput.projectId), and the OpenCode driver resolves its catch-up runtime as its other generators do.
• CommandPalette.tsx: the fork's scroll to an initial highlight passes upstream's new first-row flag.
• MessagesTimeline.tsx: the minimap reads the scroll node for upstream's at-end check, since the fork computes the visible range in a helper.
• chatRouteCommands.test.ts: drops a branch no row reaches.

Ultracode fix

Since v0.1.8, ultracode is its own boolean model option (ultracode: true, effort xhigh). isUltracode in ProviderSessionManager.ts only checked effort === "ultracode", so a session using the toggle never ranked its account as busy (#150). It now reads the toggle with getModelSelectionBooleanOptionValue and falls back to the legacy effort value, as withLegacyUltracodeEffort does. The busy test runs once per form; the toggle row failed before the fix. docs/user/usage.md and the comments now say "with Ultracode on" instead of "at effort ultracode"; the reason text ("busy: ultracode session") didn't name the effort.

Re-checked fork behaviors

Usage-limit work (#148 wakes before a reset queue behind the limited run, failover retries, SWITCH_ENDS_BACKGROUND_WORK; #150 busy ranking; #151 Claude session carry), nextAccount, message search with agent-written messages, Experts with folding quiet experts, Needs you never buried in Settled, shelves at the bottom, last-message sort, machine icon, expert row title width, thread icons and menu, quiet background turns and Needs you notifications, reply to a reply, the minimap for long threads, catch up, citations and reply references, the new thread dialog, and the fork guards: version scheme (0.1.10-t4.1 sorts above every 0.1.9-t4.N in newestCliReleaseVersion; the desktop and remote updaters use the same numeric semver compare; install-t4.sh takes GitHub's latest release and prunes backups by timestamp), update track hidden and refused, startup recovery to the fork's releases, star prompt hidden and sourceControl.star refused, triage refused, t3 install pointing at T4's scripts, t3 update with t4-v tags, and the Windows afterPack hook skipping win32 (#153). Nothing new in v0.1.9 stars or reports to upstream from fork builds; see the open question on mobile:install.

Verification

Heavy steps ran one at a time under a shared lock, vitest with 2 workers. No dev server; browser tests ran headless in vitest.

• vp i: lockfile unchanged.
• Typecheck: contracts, shared, client-runtime, server (t3), web, desktop and mobile pass, plus provider-core and provider-opencode.
• Usage limits: AccountFailover, runtimeLayer, UsageLimitRecoveryWorker, ProviderSessionManager, ClaudeSessionCarry, ProviderSessionTransitionPolicy, ProviderSwitch.integration, nextAccountInstances, OrchestratorMcpNextAccount, TaskCancelUsageLimit.integration (10 files, 296 tests) and shared usageNextAccount: all pass.
• Search and migrations: 069_ThreadMessageSearch, 081_ProjectionMessageSearchText, ThreadSearch (38 tests) pass; Migrations, reconcileV2PreviewMigration, DatabaseMaintenance, initializeV2Database pass.
• Server conflict areas: Orchestrator (control reads, migration, receipt race), ProviderTurnStartService, ThreadMessageIntake, ThreadCatchUpService, OrchestratorMcpService (+ activity, waitWake, toolkit integration), threadAccess, toolkits/core, linkOrigin, McpHttpServer, text generation: 17 files, 254 tests pass.
• Fork guards: star sign-in, triage, t3 update, t3 install, self-update, pinned runtime, service preflight (8 files, 71 tests); scripts/install.test.ts and scripts/t4/install-t4.test.ts pass (5 PowerShell-only cases skip on Linux).
• Packages: contracts (33 files), shared (citations, awareness, release order, model, usage), client-runtime (linkRequest, projects), provider-opencode driver: pass.
• Web: 16 logic files (440 tests) and 18 browser files (120 tests) for the sidebar, palette, shortcuts, notifications, citations, timeline, minimap and usage: pass.
• Mobile: 298 files (3391 tests) pass. Desktop: 112 of 113 files pass; scripts/browser-secret-native.test.mjs needs pkg-config and libsecret headers, which this machine lacks.
• ProjectionStore.test.ts: 53 of 54 pass. Upstream's new "reads a fixed SQL turn-item window" test (#33305) calls StatementSync.stat, which needs Node 26; this machine runs Node 24.
• Evals: no API key here, so no paid run. mcp-tool-use --dry-run builds the three WebMCP page tool cases that upstream's wording change (#26570) targets; --context: system text 8854 chars, tool descriptions 16670, schemas 52718 across 72 tools. The fork changed no model-facing text in this upkeep.
• vp run check:changed origin/t4 (steps one at a time, VITEST_MAX_WORKERS=2, under the lock): unused patches, format, lint, knip, graphql, typecheck (41 packages), node tests and the React Compiler check pass. The test step ran 349 files: 346 passed, 3 failed, none from this branch:
  • AgentContextBudget.test.ts: fails as it does on t4. Measured 7129 / 8857 / 51638 bytes (Claude append / Codex context / MCP tool list) against upstream's committed 5597 / 7327 / 48217. The fork's overhead is unchanged from t4 (+1532 / +1530 / +3421).
  • ProjectionStore.test.ts "reads a fixed SQL turn-item window": Node 26 only, as above.
  • GitVcsDriverCore.test.ts "makes background upstream status fetches non-interactive": this machine's global core.sshCommand overrides the test's GIT_SSH; it passes with GIT_CONFIG_GLOBAL=/dev/null, as on t4.
• fix(build): Windows payload stays within the file limit #153's scripts/lib/server-compile-cache.test.ts passes.

Open questions

• mobile:install (new upstream dev tooling, scripts/mobile-native-client.ts) defaults T3_NATIVE_CLIENT_REPOSITORY to pingdotgg/t3code: with gh installed, it asks upstream's API whether HEAD is there and could dispatch upstream's mobile-dev-client.yml (only with permission, which the fork lacks). It neither stars nor reports and isn't part of any build, so I left it; a fork default would be a one-line change.
• mobile-dev-client.yml needs disabling after the push (see Workflows).
• The ProjectionStore window test and the desktop libsecret test need Node 26 and pkg-config; CI should run both.

Upkeep by Claude Opus 5.5 in Claude Code.

🤖 Generated with Claude Code