fix(desktop): stop voice status leaking across session switch (#46194)

[harjothkhara]

What does this PR do?

Stops the desktop composer's voice status from leaking across a session switch. After switching sessions, the prior session's in-flight voice conversation (listening) and the global read-aloud playback status ($voicePlayback → Preparing audio / Speaking) were not reset, so session B's status area showed session A's carried-over voice state.

Both switch paths are handled:

• Warm switch (composer stays mounted): a colocated hook, useEndVoiceOnSessionSwitch, fires on a real session change and tears the voice state down — it disables the conversation (so the idle effect can't auto-restart listening under the new session) and calls conversation.end() (cancels recording, stops playback). The direct end() is load-bearing for the read-aloud-only case, where flipping voiceConversationActive is a no-op and won't trigger the conversation hook's own enabled→false teardown.
• Cold switch (resuming a session with no warm runtime unmounts the composer): the component-local conversation state dies with the unmount, but the module-level read-aloud singleton survives, so useStopVoicePlaybackOnUnmount stops it on teardown.

The switch guard skips the initial mount and the null→id persist of the session already in view (mirrors the composer's existing placeholder-reset), treating undefined and null as the same "no session". Leaving a session (id→null) and moving between two real sessions both count as a switch.

Related Issue

#46194 (scoped — see below; the issue should stay open for the remaining async leak).

This fixes the displayed voice-status slice only. The other symptoms named in the issue are already handled or out of scope here:

• Queued follow-up state is already keyed per session on main ($queuedPromptsBySession, fix(desktop): stop stranding queued prompts across backend bounces #45414/fix(desktop): per-thread composer drafts on decoupled lifecycle (salvage #43660, supersedes #43939) #43959).
• The transcript path is already session-keyed (use-message-stream.ts); the second-hand "both sessions show the same content" report is not reproducible on current main.
• Remaining (follow-up): a dictation or in-flight voice turn whose transcribe/submit resolves after the switch can still land in the newly active session (the stale-submit cross-binding via ensureSessionState → rewriteOptimistic). That needs epoch-guarding the async voice continuations and is intentionally left out of this focused fix.

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

• New apps/desktop/src/app/chat/composer/hooks/use-voice-session-reset.ts — useEndVoiceOnSessionSwitch (warm) + useStopVoicePlaybackOnUnmount (cold).
• Wired both into apps/desktop/src/app/chat/composer/index.tsx (the ChatBar composer).
• use-voice-session-reset.test.ts — 9 cases.

How to Test

npm --workspace apps/desktop run test:ui -- src/app/chat/composer/hooks/use-voice-session-reset.test.ts

Manual: start a read-aloud (or voice conversation) in session A, switch to session B before it settles — B's composer no longer shows A's Preparing audio/Speaking/listening state. Repeat with a cold resume of B (no warm runtime) to exercise the unmount path.

Checklist

• My code follows the project's style guidelines
• I have added tests that prove my fix is effective
• New and existing unit tests pass locally (tsc + eslint clean; only the pre-existing trigger-popover.test.tsx @assistant-ui/tap packaging failure is unrelated)
• Tested on platform: macOS

🤖 Generated with Claude Code

[liuhao1024]

✅ Verified — Desktop voice status leaking across session switch

Reviewed the diff for correctness of the session-switch detection logic and teardown coverage.

• Warm switch guard (useEndVoiceOnSessionSwitch): Correctly skips the initial mount (no previous session) and the null → id transition (session already in view). Treats undefined and null as the same "no session" state to avoid false-positive switches at startup. Fires once per real session-to-session transition — confirmed by test coverage for A→B, B→C, C→A, and id→null.
• Cold switch guard (useStopVoicePlaybackOnUnmount): Cleanup function in useEffect([], []) stops module-level read-aloud playback on unmount. Correct pattern for singleton teardown.
• Callback freshness: onSwitch is in the useEffect dependency array, so stale closures are avoided. Test confirms the latest callback is used.
• Integration: Hooks wired into ChatBar component alongside the existing voice conversation teardown — the comment explains the load-bearing conversation.end() call for the read-aloud-only case.

Clean fix with thorough test coverage. No issues found.

[harjothkhara]

@teknium1 — green/mergeable fix for #46194: the desktop composer's voice status (in-flight listening + the global read-aloud $voicePlayback) was leaking across a session switch, so session B showed session A's carried-over voice state. Covers both paths — warm switch (colocated useEndVoiceOnSessionSwitch teardown) and cold switch (unmount stop for the module-level read-aloud singleton) — scoped to the desktop composer. Verified clean by @liuhao1024. A review when convenient would be appreciated.

[harjothkhara]

@teknium1 @OutThisLife gentle review ping here. This was opened on June 14, so it is about 6 days old now.

Importance: this is a focused desktop session-isolation bug fix. It stops voice status from leaking across a session switch, with dedicated hook coverage, so the UI does not show stale recording/voice state for the wrong conversation.

[konsisumer]

I opened #61448 to consolidate this with related work across several related issues — it aims to fold in your change alongside the others so the maintainer can land one clean PR. Flagging for visibility; no action needed from you, and your PR stays open.