feat(mobile): add offline iPhone voice input

[rynfar]

Fourth of the mobile batch. Adopted from upstream pingdotgg/t3code#8614 (352710d49). Follows #223, #224, #231.

Mobile gains on-device voice input: hold the mic in the composer, speak, and the transcript commits into the draft. Brings a voice-input controller to client-runtime, native transcription, the dictation UI, and two patched native deps.

The composer restructure, settled

This batch has cost a conflict in every cherry-pick because upstream's #8793 restructure arrives as context in each one. It turns out to be two independent changes, and treating them as one decision was the mistake:

• ComposerSurface — animate borderRadius on a shared value, absolute glass layer, bounded collapsed radius. Touches nothing in the toolbar. Adopted (ceaae36aa); our file is now structurally identical to upstream there, so these hunks should stop conflicting.
• The toolbar row — drop ComposerToolbarScroller for a fixed flex row. Declined, permanently. ComposerToolbarScroller is upstream's own component and they still ship it; they stopped using it because their toolbar holds four controls. Ours holds sixteen, because ControlPillMenu — which does not exist upstream at all — carries Refine, session goal, context window, agent count, the input queue, depth, resources, and reload.

Counts unchanged throughout: scroller 3, ControlPillMenu 13, QuickQuestionTrigger 2, ContextWindowIndicator 3.

Three defects I introduced, and fixed

The dictation controls had to be placed by hand rather than ported, since upstream places them inside the restructure we declined. Review found three ways that stranded the user. All 36 ported files were byte-identical to upstream — every defect was in the hand-written JSX, and none was caught by typecheck, lint, or 1042 tests.

The editor was never frozen.NewTaskDraftScreen passes readOnly={voiceInput.freezesEditor}; the thread composer did not. The keyboard stayed live during recording, and one keystroke makes resolveTranscriptCommit see a changed draft and discard the whole transcript as stale — up to five minutes of speech, silently. It also made every native read-only guard this branch adds dead code on that surface.

Send vanished on any dictation error. Gated on isVoiceInputPresented rather than voicePresentation.showsSend. Those look interchangeable but diverge in exactly one phase: error has showsSend: trueand a non-null statusLabel. So a denied permission, "no speech detected", or the stale-draft error above left no send control. Worse on entry, since dispose() no-ops in the error phase and a stale error survives navigating away and back.

Stop was unreachable for the whole dictation window. Placed inside ComposerToolbarScroller, which is the else branch of the dictation ternary — so across preparing → recording → transcribing → error there was no way to stop a running agent.

Also fixed: blocksSubmission guards on both submission entry points (canSend is derived above voiceInput and structurally cannot include it); a 4px spacer overflowing ComposerDictationToolbar's fixed 44px box and clipping the collapsed strip; the dropped pointerEvents="none" on the glass layer.

The showsSend divergence ships with a regression test — it is the one defect that is a pure predicate rather than JSX placement, and the one most likely to be reintroduced by someone reasoning that the two predicates look the same.

Pylon branding

Two leaks caught, both user-visible. app.config.ts carried upstream's "Allow T3 Code to use your microphone for voice input" — iOS renders that verbatim in the permission dialog, and the camera permission two lines below already said Pylon. Plus two T3 Code strings in docs/user/composer.md and docs/internals/voice-input.md.

No speech-recognition permission is needed, verified rather than assumed: @react-native-ai/apple uses SpeechAnalyzer/SpeechTranscriber, Apple's on-device framework, not SFSpeechRecognizer. That is what makes this offline.

Verification

Typecheck clean, lint clean, 1044 tests passing. Native build green with AppleLLM 0.12.0 and ExpoAudio.

This needs a native rebuild, not an OTA — two new native modules plus native Swift changes in t3-composer-editor/ios. runtimeVersion.policy is fingerprint, which covers new native deps and patches/, so a JS update cannot land on a stale binary.

Simulator pass verified the unavailable path only: with voice unavailable the composer renders correctly, full toolbar, no gap where the mic would be. The dictation UI itself has zero simulator coverage — AppleTranscription.isAvailable() is a native check whose own error string reads "requires a supported device with iOS 26 or later", so the on-device model is absent on a simulator. Mic, status row, cancel, and the toolbar swap need a device pass before this merges.

Reviewed and integrated with Claude Opus 5 in Claude Code.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.7 KiB	13.7 KiB	+8 B (+0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	−3 B (−0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.8 KiB	6.8 KiB	+11 B (+0.2%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	58.7 KiB	58.7 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	11	11	0 (0.0%)	21	✅
Claude	Total thread wire	13.5 KiB	13.8 KiB	+230 B (+1.7%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+2 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.6 KiB	6.8 KiB	+228 B (+3.4%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.0 KiB	59.5 KiB	+1.5 KiB (+2.6%)	66.4 KiB	✅
Claude	Live turn messages	9	11	+2 (+22.2%)	21	✅

Baseline: c908bb2 · PR result: b21cb8f · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.5 KiB
• Claude decoded thread snapshot: 110.2 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[rynfar]

Fourth of the mobile batch. Adopted from upstream pingdotgg/t3code#8614 (352710d49). Follows #223, #224, #231.

Mobile gains on-device voice input: hold the mic in the composer, speak, and the transcript commits into the draft. Brings a voice-input controller to client-runtime, native transcription, the dictation UI, and two patched native deps.

The composer restructure, settled

This batch has cost a conflict in every cherry-pick because upstream's #8793 restructure arrives as context in each one. It turns out to be two independent changes, and treating them as one decision was the mistake:

• ComposerSurface — animate borderRadius on a shared value, absolute glass layer, bounded collapsed radius. Touches nothing in the toolbar. Adopted (ceaae36aa); our file is now structurally identical to upstream there, so these hunks should stop conflicting.
• The toolbar row — drop ComposerToolbarScroller for a fixed flex row. Declined, permanently. ComposerToolbarScroller is upstream's own component and they still ship it; they stopped using it because their toolbar holds four controls. Ours holds sixteen, because ControlPillMenu — which does not exist upstream at all — carries Refine, session goal, context window, agent count, the input queue, depth, resources, and reload.

Counts unchanged throughout: scroller 3, ControlPillMenu 13, QuickQuestionTrigger 2, ContextWindowIndicator 3.

Three defects I introduced, and fixed

The dictation controls had to be placed by hand rather than ported, since upstream places them inside the restructure we declined. Review found three ways that stranded the user. All 36 ported files were byte-identical to upstream — every defect was in the hand-written JSX, and none was caught by typecheck, lint, or 1042 tests.

The editor was never frozen.NewTaskDraftScreen passes readOnly={voiceInput.freezesEditor}; the thread composer did not. The keyboard stayed live during recording, and one keystroke makes resolveTranscriptCommit see a changed draft and discard the whole transcript as stale — up to five minutes of speech, silently. It also made every native read-only guard this branch adds dead code on that surface.

Send vanished on any dictation error. Gated on isVoiceInputPresented rather than voicePresentation.showsSend. Those look interchangeable but diverge in exactly one phase: error has showsSend: trueand a non-null statusLabel. So a denied permission, "no speech detected", or the stale-draft error above left no send control. Worse on entry, since dispose() no-ops in the error phase and a stale error survives navigating away and back.

Stop was unreachable for the whole dictation window. Placed inside ComposerToolbarScroller, which is the else branch of the dictation ternary — so across preparing → recording → transcribing → error there was no way to stop a running agent.

Also fixed: blocksSubmission guards on both submission entry points (canSend is derived above voiceInput and structurally cannot include it); a 4px spacer overflowing ComposerDictationToolbar's fixed 44px box and clipping the collapsed strip; the dropped pointerEvents="none" on the glass layer.

The showsSend divergence ships with a regression test — it is the one defect that is a pure predicate rather than JSX placement, and the one most likely to be reintroduced by someone reasoning that the two predicates look the same.

Pylon branding

Two leaks caught, both user-visible. app.config.ts carried upstream's "Allow T3 Code to use your microphone for voice input" — iOS renders that verbatim in the permission dialog, and the camera permission two lines below already said Pylon. Plus two T3 Code strings in docs/user/composer.md and docs/internals/voice-input.md.

No speech-recognition permission is needed, verified rather than assumed: @react-native-ai/apple uses SpeechAnalyzer/SpeechTranscriber, Apple's on-device framework, not SFSpeechRecognizer. That is what makes this offline.

Verification

Typecheck clean, lint clean, 1044 tests passing. Native build green with AppleLLM 0.12.0 and ExpoAudio.

This needs a native rebuild, not an OTA — two new native modules plus native Swift changes in t3-composer-editor/ios. runtimeVersion.policy is fingerprint, which covers new native deps and patches/, so a JS update cannot land on a stale binary.

Simulator pass verified the unavailable path only: with voice unavailable the composer renders correctly, full toolbar, no gap where the mic would be. The dictation UI itself has zero simulator coverage — AppleTranscription.isAvailable() is a native check whose own error string reads "requires a supported device with iOS 26 or later", so the on-device model is absent on a simulator. Mic, status row, cancel, and the toolbar swap need a device pass before this merges.

Reviewed and integrated with Claude Opus 5 in Claude Code.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.7 KiB	13.7 KiB	+8 B (+0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	−3 B (−0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.8 KiB	6.8 KiB	+11 B (+0.2%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	58.7 KiB	58.7 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	11	11	0 (0.0%)	21	✅
Claude	Total thread wire	13.5 KiB	13.8 KiB	+230 B (+1.7%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+2 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.6 KiB	6.8 KiB	+228 B (+3.4%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.0 KiB	59.5 KiB	+1.5 KiB (+2.6%)	66.4 KiB	✅
Claude	Live turn messages	9	11	+2 (+22.2%)	21	✅

Baseline: c908bb2 · PR result: b21cb8f · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.5 KiB
• Claude decoded thread snapshot: 110.2 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[rynfar]

Fourth of the mobile batch. Adopted from upstream pingdotgg/t3code#8614 (352710d49). Follows #223, #224, #231.

Mobile gains on-device voice input: hold the mic in the composer, speak, and the transcript commits into the draft. Brings a voice-input controller to client-runtime, native transcription, the dictation UI, and two patched native deps.

The composer restructure, settled

This batch has cost a conflict in every cherry-pick because upstream's #8793 restructure arrives as context in each one. It turns out to be two independent changes, and treating them as one decision was the mistake:

• ComposerSurface — animate borderRadius on a shared value, absolute glass layer, bounded collapsed radius. Touches nothing in the toolbar. Adopted (ceaae36aa); our file is now structurally identical to upstream there, so these hunks should stop conflicting.
• The toolbar row — drop ComposerToolbarScroller for a fixed flex row. Declined, permanently. ComposerToolbarScroller is upstream's own component and they still ship it; they stopped using it because their toolbar holds four controls. Ours holds sixteen, because ControlPillMenu — which does not exist upstream at all — carries Refine, session goal, context window, agent count, the input queue, depth, resources, and reload.

Counts unchanged throughout: scroller 3, ControlPillMenu 13, QuickQuestionTrigger 2, ContextWindowIndicator 3.

Three defects I introduced, and fixed

The dictation controls had to be placed by hand rather than ported, since upstream places them inside the restructure we declined. Review found three ways that stranded the user. All 36 ported files were byte-identical to upstream — every defect was in the hand-written JSX, and none was caught by typecheck, lint, or 1042 tests.

The editor was never frozen.NewTaskDraftScreen passes readOnly={voiceInput.freezesEditor}; the thread composer did not. The keyboard stayed live during recording, and one keystroke makes resolveTranscriptCommit see a changed draft and discard the whole transcript as stale — up to five minutes of speech, silently. It also made every native read-only guard this branch adds dead code on that surface.

Send vanished on any dictation error. Gated on isVoiceInputPresented rather than voicePresentation.showsSend. Those look interchangeable but diverge in exactly one phase: error has showsSend: trueand a non-null statusLabel. So a denied permission, "no speech detected", or the stale-draft error above left no send control. Worse on entry, since dispose() no-ops in the error phase and a stale error survives navigating away and back.

Stop was unreachable for the whole dictation window. Placed inside ComposerToolbarScroller, which is the else branch of the dictation ternary — so across preparing → recording → transcribing → error there was no way to stop a running agent.

Also fixed: blocksSubmission guards on both submission entry points (canSend is derived above voiceInput and structurally cannot include it); a 4px spacer overflowing ComposerDictationToolbar's fixed 44px box and clipping the collapsed strip; the dropped pointerEvents="none" on the glass layer.

The showsSend divergence ships with a regression test — it is the one defect that is a pure predicate rather than JSX placement, and the one most likely to be reintroduced by someone reasoning that the two predicates look the same.

Pylon branding

Two leaks caught, both user-visible. app.config.ts carried upstream's "Allow T3 Code to use your microphone for voice input" — iOS renders that verbatim in the permission dialog, and the camera permission two lines below already said Pylon. Plus two T3 Code strings in docs/user/composer.md and docs/internals/voice-input.md.

No speech-recognition permission is needed, verified rather than assumed: @react-native-ai/apple uses SpeechAnalyzer/SpeechTranscriber, Apple's on-device framework, not SFSpeechRecognizer. That is what makes this offline.

Verification

Typecheck clean, lint clean, 1044 tests passing. Native build green with AppleLLM 0.12.0 and ExpoAudio.

This needs a native rebuild, not an OTA — two new native modules plus native Swift changes in t3-composer-editor/ios. runtimeVersion.policy is fingerprint, which covers new native deps and patches/, so a JS update cannot land on a stale binary.

Simulator pass verified the unavailable path only: with voice unavailable the composer renders correctly, full toolbar, no gap where the mic would be. The dictation UI itself has zero simulator coverage — AppleTranscription.isAvailable() is a native check whose own error string reads "requires a supported device with iOS 26 or later", so the on-device model is absent on a simulator. Mic, status row, cancel, and the toolbar swap need a device pass before this merges.

Reviewed and integrated with Claude Opus 5 in Claude Code.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.7 KiB	13.7 KiB	+8 B (+0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	−3 B (−0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.8 KiB	6.8 KiB	+11 B (+0.2%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	58.7 KiB	58.7 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	11	11	0 (0.0%)	21	✅
Claude	Total thread wire	13.5 KiB	13.8 KiB	+230 B (+1.7%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+2 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.6 KiB	6.8 KiB	+228 B (+3.4%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.0 KiB	59.5 KiB	+1.5 KiB (+2.6%)	66.4 KiB	✅
Claude	Live turn messages	9	11	+2 (+22.2%)	21	✅

Baseline: c908bb2 · PR result: b21cb8f · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.5 KiB
• Claude decoded thread snapshot: 110.2 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[rynfar]

Fourth of the mobile batch. Adopted from upstream pingdotgg/t3code#8614 (352710d49). Follows #223, #224, #231.

Mobile gains on-device voice input: hold the mic in the composer, speak, and the transcript commits into the draft. Brings a voice-input controller to client-runtime, native transcription, the dictation UI, and two patched native deps.

The composer restructure, settled

This batch has cost a conflict in every cherry-pick because upstream's #8793 restructure arrives as context in each one. It turns out to be two independent changes, and treating them as one decision was the mistake:

• ComposerSurface — animate borderRadius on a shared value, absolute glass layer, bounded collapsed radius. Touches nothing in the toolbar. Adopted (ceaae36aa); our file is now structurally identical to upstream there, so these hunks should stop conflicting.
• The toolbar row — drop ComposerToolbarScroller for a fixed flex row. Declined, permanently. ComposerToolbarScroller is upstream's own component and they still ship it; they stopped using it because their toolbar holds four controls. Ours holds sixteen, because ControlPillMenu — which does not exist upstream at all — carries Refine, session goal, context window, agent count, the input queue, depth, resources, and reload.

Counts unchanged throughout: scroller 3, ControlPillMenu 13, QuickQuestionTrigger 2, ContextWindowIndicator 3.

Three defects I introduced, and fixed

The dictation controls had to be placed by hand rather than ported, since upstream places them inside the restructure we declined. Review found three ways that stranded the user. All 36 ported files were byte-identical to upstream — every defect was in the hand-written JSX, and none was caught by typecheck, lint, or 1042 tests.

The editor was never frozen.NewTaskDraftScreen passes readOnly={voiceInput.freezesEditor}; the thread composer did not. The keyboard stayed live during recording, and one keystroke makes resolveTranscriptCommit see a changed draft and discard the whole transcript as stale — up to five minutes of speech, silently. It also made every native read-only guard this branch adds dead code on that surface.

Send vanished on any dictation error. Gated on isVoiceInputPresented rather than voicePresentation.showsSend. Those look interchangeable but diverge in exactly one phase: error has showsSend: trueand a non-null statusLabel. So a denied permission, "no speech detected", or the stale-draft error above left no send control. Worse on entry, since dispose() no-ops in the error phase and a stale error survives navigating away and back.

Stop was unreachable for the whole dictation window. Placed inside ComposerToolbarScroller, which is the else branch of the dictation ternary — so across preparing → recording → transcribing → error there was no way to stop a running agent.

Also fixed: blocksSubmission guards on both submission entry points (canSend is derived above voiceInput and structurally cannot include it); a 4px spacer overflowing ComposerDictationToolbar's fixed 44px box and clipping the collapsed strip; the dropped pointerEvents="none" on the glass layer.

The showsSend divergence ships with a regression test — it is the one defect that is a pure predicate rather than JSX placement, and the one most likely to be reintroduced by someone reasoning that the two predicates look the same.

Pylon branding

Two leaks caught, both user-visible. app.config.ts carried upstream's "Allow T3 Code to use your microphone for voice input" — iOS renders that verbatim in the permission dialog, and the camera permission two lines below already said Pylon. Plus two T3 Code strings in docs/user/composer.md and docs/internals/voice-input.md.

No speech-recognition permission is needed, verified rather than assumed: @react-native-ai/apple uses SpeechAnalyzer/SpeechTranscriber, Apple's on-device framework, not SFSpeechRecognizer. That is what makes this offline.

Verification

Typecheck clean, lint clean, 1044 tests passing. Native build green with AppleLLM 0.12.0 and ExpoAudio.

This needs a native rebuild, not an OTA — two new native modules plus native Swift changes in t3-composer-editor/ios. runtimeVersion.policy is fingerprint, which covers new native deps and patches/, so a JS update cannot land on a stale binary.

Simulator pass verified the unavailable path only: with voice unavailable the composer renders correctly, full toolbar, no gap where the mic would be. The dictation UI itself has zero simulator coverage — AppleTranscription.isAvailable() is a native check whose own error string reads "requires a supported device with iOS 26 or later", so the on-device model is absent on a simulator. Mic, status row, cancel, and the toolbar swap need a device pass before this merges.

Reviewed and integrated with Claude Opus 5 in Claude Code.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.7 KiB	13.7 KiB	+8 B (+0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	−3 B (−0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.8 KiB	6.8 KiB	+11 B (+0.2%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	58.7 KiB	58.7 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	11	11	0 (0.0%)	21	✅
Claude	Total thread wire	13.5 KiB	13.8 KiB	+230 B (+1.7%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+2 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.6 KiB	6.8 KiB	+228 B (+3.4%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.0 KiB	59.5 KiB	+1.5 KiB (+2.6%)	66.4 KiB	✅
Claude	Live turn messages	9	11	+2 (+22.2%)	21	✅

Baseline: c908bb2 · PR result: b21cb8f · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.5 KiB
• Claude decoded thread snapshot: 110.2 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[rynfar]

Fourth of the mobile batch. Adopted from upstream pingdotgg/t3code#8614 (352710d49). Follows #223, #224, #231.

Mobile gains on-device voice input: hold the mic in the composer, speak, and the transcript commits into the draft. Brings a voice-input controller to client-runtime, native transcription, the dictation UI, and two patched native deps.

The composer restructure, settled

This batch has cost a conflict in every cherry-pick because upstream's #8793 restructure arrives as context in each one. It turns out to be two independent changes, and treating them as one decision was the mistake:

• ComposerSurface — animate borderRadius on a shared value, absolute glass layer, bounded collapsed radius. Touches nothing in the toolbar. Adopted (ceaae36aa); our file is now structurally identical to upstream there, so these hunks should stop conflicting.
• The toolbar row — drop ComposerToolbarScroller for a fixed flex row. Declined, permanently. ComposerToolbarScroller is upstream's own component and they still ship it; they stopped using it because their toolbar holds four controls. Ours holds sixteen, because ControlPillMenu — which does not exist upstream at all — carries Refine, session goal, context window, agent count, the input queue, depth, resources, and reload.

Counts unchanged throughout: scroller 3, ControlPillMenu 13, QuickQuestionTrigger 2, ContextWindowIndicator 3.

Three defects I introduced, and fixed

The dictation controls had to be placed by hand rather than ported, since upstream places them inside the restructure we declined. Review found three ways that stranded the user. All 36 ported files were byte-identical to upstream — every defect was in the hand-written JSX, and none was caught by typecheck, lint, or 1042 tests.

The editor was never frozen.NewTaskDraftScreen passes readOnly={voiceInput.freezesEditor}; the thread composer did not. The keyboard stayed live during recording, and one keystroke makes resolveTranscriptCommit see a changed draft and discard the whole transcript as stale — up to five minutes of speech, silently. It also made every native read-only guard this branch adds dead code on that surface.

Send vanished on any dictation error. Gated on isVoiceInputPresented rather than voicePresentation.showsSend. Those look interchangeable but diverge in exactly one phase: error has showsSend: trueand a non-null statusLabel. So a denied permission, "no speech detected", or the stale-draft error above left no send control. Worse on entry, since dispose() no-ops in the error phase and a stale error survives navigating away and back.

Stop was unreachable for the whole dictation window. Placed inside ComposerToolbarScroller, which is the else branch of the dictation ternary — so across preparing → recording → transcribing → error there was no way to stop a running agent.

Also fixed: blocksSubmission guards on both submission entry points (canSend is derived above voiceInput and structurally cannot include it); a 4px spacer overflowing ComposerDictationToolbar's fixed 44px box and clipping the collapsed strip; the dropped pointerEvents="none" on the glass layer.

The showsSend divergence ships with a regression test — it is the one defect that is a pure predicate rather than JSX placement, and the one most likely to be reintroduced by someone reasoning that the two predicates look the same.

Pylon branding

Two leaks caught, both user-visible. app.config.ts carried upstream's "Allow T3 Code to use your microphone for voice input" — iOS renders that verbatim in the permission dialog, and the camera permission two lines below already said Pylon. Plus two T3 Code strings in docs/user/composer.md and docs/internals/voice-input.md.

No speech-recognition permission is needed, verified rather than assumed: @react-native-ai/apple uses SpeechAnalyzer/SpeechTranscriber, Apple's on-device framework, not SFSpeechRecognizer. That is what makes this offline.

Verification

Typecheck clean, lint clean, 1044 tests passing. Native build green with AppleLLM 0.12.0 and ExpoAudio.

This needs a native rebuild, not an OTA — two new native modules plus native Swift changes in t3-composer-editor/ios. runtimeVersion.policy is fingerprint, which covers new native deps and patches/, so a JS update cannot land on a stale binary.

Simulator pass verified the unavailable path only: with voice unavailable the composer renders correctly, full toolbar, no gap where the mic would be. The dictation UI itself has zero simulator coverage — AppleTranscription.isAvailable() is a native check whose own error string reads "requires a supported device with iOS 26 or later", so the on-device model is absent on a simulator. Mic, status row, cancel, and the toolbar swap need a device pass before this merges.

Reviewed and integrated with Claude Opus 5 in Claude Code.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.7 KiB	13.7 KiB	+8 B (+0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	−3 B (−0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.8 KiB	6.8 KiB	+11 B (+0.2%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	58.7 KiB	58.7 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	11	11	0 (0.0%)	21	✅
Claude	Total thread wire	13.5 KiB	13.8 KiB	+230 B (+1.7%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+2 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.6 KiB	6.8 KiB	+228 B (+3.4%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.0 KiB	59.5 KiB	+1.5 KiB (+2.6%)	66.4 KiB	✅
Claude	Live turn messages	9	11	+2 (+22.2%)	21	✅

Baseline: c908bb2 · PR result: b21cb8f · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.5 KiB
• Claude decoded thread snapshot: 110.2 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[rynfar]

Fourth of the mobile batch. Adopted from upstream pingdotgg/t3code#8614 (352710d49). Follows #223, #224, #231.

Mobile gains on-device voice input: hold the mic in the composer, speak, and the transcript commits into the draft. Brings a voice-input controller to client-runtime, native transcription, the dictation UI, and two patched native deps.

The composer restructure, settled

This batch has cost a conflict in every cherry-pick because upstream's #8793 restructure arrives as context in each one. It turns out to be two independent changes, and treating them as one decision was the mistake:

• ComposerSurface — animate borderRadius on a shared value, absolute glass layer, bounded collapsed radius. Touches nothing in the toolbar. Adopted (ceaae36aa); our file is now structurally identical to upstream there, so these hunks should stop conflicting.
• The toolbar row — drop ComposerToolbarScroller for a fixed flex row. Declined, permanently. ComposerToolbarScroller is upstream's own component and they still ship it; they stopped using it because their toolbar holds four controls. Ours holds sixteen, because ControlPillMenu — which does not exist upstream at all — carries Refine, session goal, context window, agent count, the input queue, depth, resources, and reload.

Counts unchanged throughout: scroller 3, ControlPillMenu 13, QuickQuestionTrigger 2, ContextWindowIndicator 3.

Three defects I introduced, and fixed

The dictation controls had to be placed by hand rather than ported, since upstream places them inside the restructure we declined. Review found three ways that stranded the user. All 36 ported files were byte-identical to upstream — every defect was in the hand-written JSX, and none was caught by typecheck, lint, or 1042 tests.

The editor was never frozen.NewTaskDraftScreen passes readOnly={voiceInput.freezesEditor}; the thread composer did not. The keyboard stayed live during recording, and one keystroke makes resolveTranscriptCommit see a changed draft and discard the whole transcript as stale — up to five minutes of speech, silently. It also made every native read-only guard this branch adds dead code on that surface.

Send vanished on any dictation error. Gated on isVoiceInputPresented rather than voicePresentation.showsSend. Those look interchangeable but diverge in exactly one phase: error has showsSend: trueand a non-null statusLabel. So a denied permission, "no speech detected", or the stale-draft error above left no send control. Worse on entry, since dispose() no-ops in the error phase and a stale error survives navigating away and back.

Stop was unreachable for the whole dictation window. Placed inside ComposerToolbarScroller, which is the else branch of the dictation ternary — so across preparing → recording → transcribing → error there was no way to stop a running agent.

Also fixed: blocksSubmission guards on both submission entry points (canSend is derived above voiceInput and structurally cannot include it); a 4px spacer overflowing ComposerDictationToolbar's fixed 44px box and clipping the collapsed strip; the dropped pointerEvents="none" on the glass layer.

The showsSend divergence ships with a regression test — it is the one defect that is a pure predicate rather than JSX placement, and the one most likely to be reintroduced by someone reasoning that the two predicates look the same.

Pylon branding

Two leaks caught, both user-visible. app.config.ts carried upstream's "Allow T3 Code to use your microphone for voice input" — iOS renders that verbatim in the permission dialog, and the camera permission two lines below already said Pylon. Plus two T3 Code strings in docs/user/composer.md and docs/internals/voice-input.md.

No speech-recognition permission is needed, verified rather than assumed: @react-native-ai/apple uses SpeechAnalyzer/SpeechTranscriber, Apple's on-device framework, not SFSpeechRecognizer. That is what makes this offline.

Verification

Typecheck clean, lint clean, 1044 tests passing. Native build green with AppleLLM 0.12.0 and ExpoAudio.

This needs a native rebuild, not an OTA — two new native modules plus native Swift changes in t3-composer-editor/ios. runtimeVersion.policy is fingerprint, which covers new native deps and patches/, so a JS update cannot land on a stale binary.

Simulator pass verified the unavailable path only: with voice unavailable the composer renders correctly, full toolbar, no gap where the mic would be. The dictation UI itself has zero simulator coverage — AppleTranscription.isAvailable() is a native check whose own error string reads "requires a supported device with iOS 26 or later", so the on-device model is absent on a simulator. Mic, status row, cancel, and the toolbar swap need a device pass before this merges.

Reviewed and integrated with Claude Opus 5 in Claude Code.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.7 KiB	13.7 KiB	+8 B (+0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	−3 B (−0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.8 KiB	6.8 KiB	+11 B (+0.2%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	58.7 KiB	58.7 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	11	11	0 (0.0%)	21	✅
Claude	Total thread wire	13.5 KiB	13.8 KiB	+230 B (+1.7%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+2 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.6 KiB	6.8 KiB	+228 B (+3.4%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.0 KiB	59.5 KiB	+1.5 KiB (+2.6%)	66.4 KiB	✅
Claude	Live turn messages	9	11	+2 (+22.2%)	21	✅

Baseline: c908bb2 · PR result: b21cb8f · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.5 KiB
• Claude decoded thread snapshot: 110.2 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[rynfar]

Fourth of the mobile batch. Adopted from upstream pingdotgg/t3code#8614 (352710d49). Follows #223, #224, #231.

Mobile gains on-device voice input: hold the mic in the composer, speak, and the transcript commits into the draft. Brings a voice-input controller to client-runtime, native transcription, the dictation UI, and two patched native deps.

The composer restructure, settled

This batch has cost a conflict in every cherry-pick because upstream's #8793 restructure arrives as context in each one. It turns out to be two independent changes, and treating them as one decision was the mistake:

• ComposerSurface — animate borderRadius on a shared value, absolute glass layer, bounded collapsed radius. Touches nothing in the toolbar. Adopted (ceaae36aa); our file is now structurally identical to upstream there, so these hunks should stop conflicting.
• The toolbar row — drop ComposerToolbarScroller for a fixed flex row. Declined, permanently. ComposerToolbarScroller is upstream's own component and they still ship it; they stopped using it because their toolbar holds four controls. Ours holds sixteen, because ControlPillMenu — which does not exist upstream at all — carries Refine, session goal, context window, agent count, the input queue, depth, resources, and reload.

Counts unchanged throughout: scroller 3, ControlPillMenu 13, QuickQuestionTrigger 2, ContextWindowIndicator 3.

Three defects I introduced, and fixed

The dictation controls had to be placed by hand rather than ported, since upstream places them inside the restructure we declined. Review found three ways that stranded the user. All 36 ported files were byte-identical to upstream — every defect was in the hand-written JSX, and none was caught by typecheck, lint, or 1042 tests.

The editor was never frozen.NewTaskDraftScreen passes readOnly={voiceInput.freezesEditor}; the thread composer did not. The keyboard stayed live during recording, and one keystroke makes resolveTranscriptCommit see a changed draft and discard the whole transcript as stale — up to five minutes of speech, silently. It also made every native read-only guard this branch adds dead code on that surface.

Send vanished on any dictation error. Gated on isVoiceInputPresented rather than voicePresentation.showsSend. Those look interchangeable but diverge in exactly one phase: error has showsSend: trueand a non-null statusLabel. So a denied permission, "no speech detected", or the stale-draft error above left no send control. Worse on entry, since dispose() no-ops in the error phase and a stale error survives navigating away and back.

Stop was unreachable for the whole dictation window. Placed inside ComposerToolbarScroller, which is the else branch of the dictation ternary — so across preparing → recording → transcribing → error there was no way to stop a running agent.

Also fixed: blocksSubmission guards on both submission entry points (canSend is derived above voiceInput and structurally cannot include it); a 4px spacer overflowing ComposerDictationToolbar's fixed 44px box and clipping the collapsed strip; the dropped pointerEvents="none" on the glass layer.

The showsSend divergence ships with a regression test — it is the one defect that is a pure predicate rather than JSX placement, and the one most likely to be reintroduced by someone reasoning that the two predicates look the same.

Pylon branding

Two leaks caught, both user-visible. app.config.ts carried upstream's "Allow T3 Code to use your microphone for voice input" — iOS renders that verbatim in the permission dialog, and the camera permission two lines below already said Pylon. Plus two T3 Code strings in docs/user/composer.md and docs/internals/voice-input.md.

No speech-recognition permission is needed, verified rather than assumed: @react-native-ai/apple uses SpeechAnalyzer/SpeechTranscriber, Apple's on-device framework, not SFSpeechRecognizer. That is what makes this offline.

Verification

Typecheck clean, lint clean, 1044 tests passing. Native build green with AppleLLM 0.12.0 and ExpoAudio.

This needs a native rebuild, not an OTA — two new native modules plus native Swift changes in t3-composer-editor/ios. runtimeVersion.policy is fingerprint, which covers new native deps and patches/, so a JS update cannot land on a stale binary.

Simulator pass verified the unavailable path only: with voice unavailable the composer renders correctly, full toolbar, no gap where the mic would be. The dictation UI itself has zero simulator coverage — AppleTranscription.isAvailable() is a native check whose own error string reads "requires a supported device with iOS 26 or later", so the on-device model is absent on a simulator. Mic, status row, cancel, and the toolbar swap need a device pass before this merges.

Reviewed and integrated with Claude Opus 5 in Claude Code.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.7 KiB	13.7 KiB	+8 B (+0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	−3 B (−0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.8 KiB	6.8 KiB	+11 B (+0.2%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	58.7 KiB	58.7 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	11	11	0 (0.0%)	21	✅
Claude	Total thread wire	13.5 KiB	13.8 KiB	+230 B (+1.7%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+2 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.6 KiB	6.8 KiB	+228 B (+3.4%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.0 KiB	59.5 KiB	+1.5 KiB (+2.6%)	66.4 KiB	✅
Claude	Live turn messages	9	11	+2 (+22.2%)	21	✅

Baseline: c908bb2 · PR result: b21cb8f · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.5 KiB
• Claude decoded thread snapshot: 110.2 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[rynfar]

Fourth of the mobile batch. Adopted from upstream pingdotgg/t3code#8614 (352710d49). Follows #223, #224, #231.

Mobile gains on-device voice input: hold the mic in the composer, speak, and the transcript commits into the draft. Brings a voice-input controller to client-runtime, native transcription, the dictation UI, and two patched native deps.

The composer restructure, settled

This batch has cost a conflict in every cherry-pick because upstream's #8793 restructure arrives as context in each one. It turns out to be two independent changes, and treating them as one decision was the mistake:

• ComposerSurface — animate borderRadius on a shared value, absolute glass layer, bounded collapsed radius. Touches nothing in the toolbar. Adopted (ceaae36aa); our file is now structurally identical to upstream there, so these hunks should stop conflicting.
• The toolbar row — drop ComposerToolbarScroller for a fixed flex row. Declined, permanently. ComposerToolbarScroller is upstream's own component and they still ship it; they stopped using it because their toolbar holds four controls. Ours holds sixteen, because ControlPillMenu — which does not exist upstream at all — carries Refine, session goal, context window, agent count, the input queue, depth, resources, and reload.

Counts unchanged throughout: scroller 3, ControlPillMenu 13, QuickQuestionTrigger 2, ContextWindowIndicator 3.

Three defects I introduced, and fixed

The dictation controls had to be placed by hand rather than ported, since upstream places them inside the restructure we declined. Review found three ways that stranded the user. All 36 ported files were byte-identical to upstream — every defect was in the hand-written JSX, and none was caught by typecheck, lint, or 1042 tests.

The editor was never frozen.NewTaskDraftScreen passes readOnly={voiceInput.freezesEditor}; the thread composer did not. The keyboard stayed live during recording, and one keystroke makes resolveTranscriptCommit see a changed draft and discard the whole transcript as stale — up to five minutes of speech, silently. It also made every native read-only guard this branch adds dead code on that surface.

Send vanished on any dictation error. Gated on isVoiceInputPresented rather than voicePresentation.showsSend. Those look interchangeable but diverge in exactly one phase: error has showsSend: trueand a non-null statusLabel. So a denied permission, "no speech detected", or the stale-draft error above left no send control. Worse on entry, since dispose() no-ops in the error phase and a stale error survives navigating away and back.

Stop was unreachable for the whole dictation window. Placed inside ComposerToolbarScroller, which is the else branch of the dictation ternary — so across preparing → recording → transcribing → error there was no way to stop a running agent.

Also fixed: blocksSubmission guards on both submission entry points (canSend is derived above voiceInput and structurally cannot include it); a 4px spacer overflowing ComposerDictationToolbar's fixed 44px box and clipping the collapsed strip; the dropped pointerEvents="none" on the glass layer.

The showsSend divergence ships with a regression test — it is the one defect that is a pure predicate rather than JSX placement, and the one most likely to be reintroduced by someone reasoning that the two predicates look the same.

Pylon branding

Two leaks caught, both user-visible. app.config.ts carried upstream's "Allow T3 Code to use your microphone for voice input" — iOS renders that verbatim in the permission dialog, and the camera permission two lines below already said Pylon. Plus two T3 Code strings in docs/user/composer.md and docs/internals/voice-input.md.

No speech-recognition permission is needed, verified rather than assumed: @react-native-ai/apple uses SpeechAnalyzer/SpeechTranscriber, Apple's on-device framework, not SFSpeechRecognizer. That is what makes this offline.

Verification

Typecheck clean, lint clean, 1044 tests passing. Native build green with AppleLLM 0.12.0 and ExpoAudio.

This needs a native rebuild, not an OTA — two new native modules plus native Swift changes in t3-composer-editor/ios. runtimeVersion.policy is fingerprint, which covers new native deps and patches/, so a JS update cannot land on a stale binary.

Simulator pass verified the unavailable path only: with voice unavailable the composer renders correctly, full toolbar, no gap where the mic would be. The dictation UI itself has zero simulator coverage — AppleTranscription.isAvailable() is a native check whose own error string reads "requires a supported device with iOS 26 or later", so the on-device model is absent on a simulator. Mic, status row, cancel, and the toolbar swap need a device pass before this merges.

Reviewed and integrated with Claude Opus 5 in Claude Code.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.7 KiB	13.7 KiB	+8 B (+0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	−3 B (−0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.8 KiB	6.8 KiB	+11 B (+0.2%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	58.7 KiB	58.7 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	11	11	0 (0.0%)	21	✅
Claude	Total thread wire	13.5 KiB	13.8 KiB	+230 B (+1.7%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+2 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.6 KiB	6.8 KiB	+228 B (+3.4%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.0 KiB	59.5 KiB	+1.5 KiB (+2.6%)	66.4 KiB	✅
Claude	Live turn messages	9	11	+2 (+22.2%)	21	✅

Baseline: c908bb2 · PR result: b21cb8f · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 109.5 KiB
• Claude decoded thread snapshot: 110.2 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.