desktop: show PTT/voice questions in chat instantly, before the reply

[kodjima33]

What

Voice/PTT questions now appear in the main chat page the instant they're transcribed — the same immediacy as a typed message — instead of only after the assistant finishes replying.

Why

Typed messages append optimistically and show immediately; voice turns were recorded to chat only at turn end (recordVoiceTurn → recordCompletedTurn), appended together with the finished reply, so the question stayed invisible until the model answered.

How

• ChatProvider.beginVoiceUserMessage — optimistically append the transcribed question the instant it's known (UI-only, no persist), returning the bubble.
• recordCompletedTurn(earlyUserMessageId:) — reconcile against THIS turn's bubble by explicit per-turn id (update text if the transcript was language-corrected) instead of appending a duplicate. Per-turn id, not a shared global: voice completions run from async tasks and can land out of order vs the next turn's start, so a shared id would corrupt a newer turn's bubble.
• RealtimeHubController — fire the early append at the earliest reliable per-provider signal (OpenAI's final input transcript; Gemini's first reply chunk, since it never marks its input transcript final), capturing the bubble id synchronously before the language-ID task. Skip the early show when the provider transcript classifies as a language the user didn't pick (likely a Gemini misdetect) — the completed turn's on-device correction shows the right text instead of flashing a wrong-language bubble.

Tests

• 8 new unit tests on the real ChatProvider (instant-show, no-duplicate, corrected-text reconcile, unchanged legacy path, empty guard, consecutive turns, and two interleaved/out-of-order race cases).
• Updated two DesktopCoordinatorService contract-string tests to the new recordVoiceTurn/recordVoiceAgentHandoff signatures (more specific, not weakened).

Verification

Exercised the real hub path via ptt_test_turn (real turn, PCM audio + forced provider transcript) with English selected: forced provider transcript イヤホン を つけて (ja) → saved chat bubble corrected to "What is the capital of France?" (used_local_transcript: true). The instant-show and the misdetect gate both confirmed live.

Follow-up (separate)

The misdetect correction only runs when a voice language is explicitly selected. Users who see "English ✓" via the settings fallback but never saved it get no correction — a migration to persist that fallback is tracked as follow-up work in the languages workstream.

🤖 Generated with Claude Code

[cubic-dev-ai]

1 issue found across 6 files

Confidence score: 3/5

• In desktop/macos/Desktop/Sources/FloatingControlBar/RealtimeHubController.swift, hubDidReceiveAudio/hubDidEmitText can exit early when suppressAssistantOutputForCurrentTurn is set, so Gemini handoff or suppressed-output turns may skip the new instant-display flow and make voice questions appear to stall or not surface promptly. Before merging, route these suppressed/handoff paths through the same instant-display trigger (or an equivalent fallback) and verify with a targeted handoff/suppressed-turn test run.

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="desktop/macos/Desktop/Sources/FloatingControlBar/RealtimeHubController.swift">

<violation number="1" location="desktop/macos/Desktop/Sources/FloatingControlBar/RealtimeHubController.swift:1003">
P2: Voice questions can still miss the new “instant display” path on Gemini handoff/suppressed-output turns, because `hubDidReceiveAudio`/`hubDidEmitText` return early on `suppressAssistantOutputForCurrentTurn` before running `showUserQuestionEarly()`. Showing the user bubble before that suppression guard would preserve the no-meta-output behavior while keeping question display immediate.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

P2: Voice questions can still miss the new “instant display” path on Gemini handoff/suppressed-output turns, because hubDidReceiveAudio/hubDidEmitText return early on suppressAssistantOutputForCurrentTurn before running showUserQuestionEarly(). Showing the user bubble before that suppression guard would preserve the no-meta-output behavior while keeping question display immediate.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At desktop/macos/Desktop/Sources/FloatingControlBar/RealtimeHubController.swift, line 1003:

<comment>Voice questions can still miss the new “instant display” path on Gemini handoff/suppressed-output turns, because `hubDidReceiveAudio`/`hubDidEmitText` return early on `suppressAssistantOutputForCurrentTurn` before running `showUserQuestionEarly()`. Showing the user bubble before that suppression guard would preserve the no-meta-output behavior while keeping question display immediate.</comment>

<file context>
@@ -986,6 +998,9 @@ final class RealtimeHubController: NSObject, RealtimeHubSessionDelegate {
       if !text.isEmpty { turnTranscript = text }
+      // OpenAI marks the input transcript final before/as the model starts replying —
+      // surface the question in chat now instead of waiting for the whole turn.
+      showUserQuestionEarly()
     } else {
       turnTranscript += text
</file context>