feat: Codex voice conversations on web, desktop, and mobile

[jakeleventhal]

Screen.Recording.2026-09-22.at.8.13.50.PM.mov

ScreenRecording_09-22-2026.20-15-53_1.MP4

Codex 0.156 shipped voice conversations: you talk to a realtime voice model, it hands coding work to Codex in the thread, and it reads Codex's answer back. T3 Code had no way to do this.

How it works

• Audio stays off the T3 socket. The client owns the mic, speaker, and WebRTC peer (oai-events data channel), and audio flows directly between the client and OpenAI. The new provider.voice.session stream RPC carries only the SDP handshake, captions, and lifecycle, so voice costs the same locally, over a relay, or through a tunnel.
• The stream is the conversation. Unsubscribing or disconnecting stops the Codex realtime session, so a dropped client can't leave a billed session running.
• Server (CodexVoice.ts, CodexAdapterV2). Starts thread/realtime/start with the same V3 WebRTC, client-managed-handoff settings the Codex TUI uses. When a voice turn finishes, its final answer goes back through thread/realtime/appendSpeech; commentary and reasoning are never spoken.
• Idle-thread handoffs. Codex starts or steers the delegated turn itself, and that can't be turned off. Steering an active run needs nothing special. On an idle thread V2 has no provider-initiated run, so the adapter interrupts Codex's turn and, once it settles, offers a voice continuation. The request then runs as an ordinary user message. The tradeoff: Codex's own history keeps a stub of the interrupted turn. First-class provider-initiated runs would remove this.
• Clients. A shared VoiceModeController in client-runtime handles negotiation, mute, captions, mic/speaker levels, the connect timeout, and teardown. Browsers and react-native-webrtc share the same structural peer API.
• Web/desktop. Waveform button in the composer, /voice, F8 (voice.toggle), an unbound voice.toggleMute, command palette actions, and a voice strip with status, meters, mute/stop, and live captions. Captions are ephemeral.
• Desktop. Grants mic access to the app origin only, asks macOS on first use, and adds NSMicrophoneUsageDescription and the audio-input entitlement.
• Mobile. Adds react-native-webrtc (native rebuild required), Android RECORD_AUDIO, loudspeaker routing, and the same button and strip. Conversations stop on background or when leaving the thread.
• Settings and scope. Voice is chosen in Codex provider settings. It's Codex-only (ServerProvider.supportsVoice) and needs a thread whose provider session is running.

Verification

• Typecheck is clean in contracts, shared, client-runtime, server, web, desktop, and mobile.
• Focused tests: 188 server (voice module, V2 Codex adapter, continuation service), 7 client-runtime, 88 web, 74 desktop, 69 shared keybindings, and 5 mobile.
• Codex 0.156's app-server accepts the exact start params.
• Installed and launched on desktop and iPhone from a personal build. A full spoken round trip has not been verified yet, and no screenshots are attached.

🤖 Generated with Claude Code (Claude Opus 5.5, running in T3 Code)

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This PR adds a substantial cross-platform Codex voice workflow with WebRTC, microphone permissions, native entitlements, new server orchestration, and an authenticated streaming RPC. It also changes the default F8 shortcut and adds static-analysis suppressions, so the scope and side effects warrant human review.

^{You can add or adjust custom eligibility rules. Learn more.}

[macroscopeapp]

Ignore my attempted follow-up review submission; it was not posted. The submitted inline comment on ChatComposer.tsx is the only UI consistency finding from this check.

^{Posted via Macroscope — UI Consistency}