Before submitting#
Area#
apps/desktop
Steps to reproduce#
- In a running Claude turn, the agent calls
t3_thread_configure on its own thread, changing only effort (e.g. high -> xhigh).
- The agent keeps working in the same turn.
Expected behavior#
The new effort applies to the rest of the running turn, as Claude Code's own /effort does, keeping the same process and prompt cache.
Actual behavior#
The running turn keeps the old effort; it applies only when a new turn starts. An agent needing more effort for its next step must queue a message to itself and end its turn: an extra turn, extra tokens and a wait.
Source: thread.model-selection.set only stores the selection (Orchestrator.ts). The only live call, live.query.applyFlagSettings(claudeLiveSelectionFlagSettings(...)) in ClaudeAdapterV2.ts, runs in the turn-start path. Suggested fix: when the thread has a live query and canApplyClaudeSelectionLive holds, call applyFlagSettings at configure time too.
Impact#
Major degradation or frequent failure
Version or commit#
0.1.9-nightly.20261111.4515 (source read at e4d86abca9)
Environment#
macOS, desktop app, Claude provider
Logs or stack traces#
Screenshots, recordings, or supporting files#
No response
Workaround#
Queue a message to the same thread and end the turn.