[collissimo]

Orchestrator mode had no way to choose reasoning effort: a new orchestrator took its model's default, and its subagents took whatever the project's default model carried. An orchestrator also refused to do a task itself when asked.

• Settings → Threads → Orchestrator: "Orchestrator effort" (what a new orchestrator starts on) and "Subagent effort" (what its subagents start on). A running orchestrator's own effort is still changed in its composer.
• Per chat: the orchestrator icon beside the chat title is now a menu, "Subagent effort in this chat". It overrides the default for that chat; "Default" follows Settings.
• The server applies an effort only when the worker's model offers it, so a value from another provider never reaches a model that lacks it. Prompt-only values such as ultrathink are not offered.
• Orchestrator instructions: when the user tells the orchestrator to do a task itself, it does, for that task.

REASONING_EFFORT_OPTION_IDS moved from client-runtime to @t3tools/shared/model so the server can use it.

Checked: the unit test in packages/shared/src/orchestratorMode.test.ts covers default, per-chat override, an effort the model lacks, and inheriting the orchestrator's model; in a dev server both settings saved and the chat menu saved its per-chat value. check:changed passes except the React Compiler coverage test for DictationHistoryDialog.tsx, which this PR does not touch.

Not checked: a real delegated subagent starting on the chosen effort end to end, the instruction change against a model (no eval covers orchestrator mode yet), mobile (orchestrator mode has no mobile UI). No screenshots attached.

Claude Fable 5.2, Claude Code

🤖 Generated with Claude Code