feat(sessions): Add steer and queue messaging modes

[charlesvien]

Problem

Messages sent while the agent was working could only queue until the turn finished, so there was no way to steer a run already in progress.

Closes #1056

Changes

• Add a Steer/Queue toggle to the composer, a mod+s shortcut and a "Default messaging mode" setting in General settings
• Steer Claude local turns natively by folding the message into the running turn at the next tool boundary (priority: "next")
• Fall back to interrupt-and-resend for Codex and cloud sessions
• Persist a per-task mode override layered over the global default
• Skip turn-state bookkeeping for steer messages so the running turn keeps its prompt lifecycle
• Thread steer through the agent prompt router, schema and workspace-server

How did you test this?

Manually

Automatic notifications

• Publish to changelog?
• Alert Sales and Marketing teams?

[github-actions]

React Doctor found 1 issue in 1 file · 1 warning.

1 warning

src/features/sessions/components/SessionView.tsx

• ⚠️ src/features/sessions/components/SessionView.tsx:631 JSX element passed as a prop jsx-no-jsx-as-prop

_{Reviewed by React Doctor for commit 5534cf3.}

[stamphog]

Complex feature touching agent mid-turn injection, session lifecycle state, and API contracts with zero reviews. The implementation looks architecturally sound, but the behavioral changes to core agent message handling warrant a human reviewer sign-off before auto-approval.

[greptile-apps]

Prompt To Fix All With AI

Fix the following 2 code review issues. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 2
packages/ui/src/features/sessions/hooks/useToggleMessagingMode.ts:29-35
**Queued messages silently dropped on flush error**

`void sessionService.sendPrompt(...)` discards any rejection, so if the session becomes unavailable mid-flush (session ends, network drop, `addDirectoryDialog` open) the messages are permanently lost with no user feedback. The messages were already dequeued before the sends, so there is no recovery path. Wrapping each call in a try/catch — or collecting the promises and surfacing a toast on failure — would prevent silent data loss.

### Issue 2 of 2
packages/ui/src/features/sessions/messagingModeStore.ts:14-27
**`modesByTaskId` grows indefinitely**

Every `setMode` call adds an entry keyed by `taskId` and the store is persisted to disk. There is no eviction — a user with many historical tasks will accumulate stale entries that are never read again. Adding a `clearMode(taskId)` action and calling it when a task is deleted (or sessions are cleaned up) would keep the persisted payload bounded.

_{Reviews (1): Last reviewed commit: "add steer and queue messaging modes" | Re-trigger Greptile}

[charlesvien]

• feat(sessions): Add steer and queue messaging modes #2663 👈 (View in Graphite)
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[stamphog]

The steer/queue implementation looks technically correct, but this is a T1d-complex feature (444 lines, 20 files) touching core agent mid-turn injection behavior, session lifecycle state (currentPromptId / isPromptPending bookkeeping), and API contracts across five packages — with zero human reviews on the current head. Both existing reviews are bot comments on an older commit.