feat(voice): live dictation via macOS 26 SpeechAnalyzer by Tryanks · Pull Request #233 · Tryanks/tcode

Tryanks · 2026-08-20T18:58:51Z

What

Voice input for the composer: a mic toggle button that streams on-device speech transcription into the message editor as you speak. macOS 26+ only — on every other platform/version the button does not exist (compile-time cfg + runtime is_supported()).

Research behind the approach: docs/voice-input-research.md. Plan: docs/plans/voice-input.md.

How

crates/voice (new) — tcode-voice: a Swift shim (swift/shim.swift, C ABI via @_cdecl) around SpeechAnalyzer/SpeechTranscriber, compiled and linked by build.rs only when targeting macOS with an SDK ≥ 26 (TCODE_VOICE_FORCE_STUB=1 forces the stub). Swift owns the whole audio path: AVAudioEngine mic tap → AVAudioConverter → AnalyzerInput stream; locale assets are installed on demand via AssetInventory. Events (Ready | Volatile | Final | Error | Ended) cross the FFI on arbitrary threads with a serialized, single-terminal contract.
Composer UI — mic button next to the send button (both row layouts). Click to start (spinner while assets/mic warm up), speech streams into the editor at an anchored insertion point: each volatile hypothesis rewrites the volatile tail in place, finals commit. Click again / Esc stops and keeps the text; submit, thread switch, or typing during dictation also stops it. All offsets are UTF-8 bytes (zh-CN safe).
Locale — follows the app locale: zh → zh_CN, otherwise en_US.
Packaging — NSMicrophoneUsageDescription + NSSpeechRecognitionUsageDescription added to the release Info.plist.

Test plan

cargo run -p tcode-voice --example file_dictation -- <aiff> zh_CN — 21 volatile events, correct final, clean exit (fixture from say -v Tingting)
cargo check --workspace, cargo clippy — clean (only the pre-existing block v0.1.6 future-incompat note)
cargo test -p tcode-ui — 254 passed, incl. locale key-set test and the new transcript-edit unit test
TCODE_VOICE_FORCE_STUB=1 build — stub path compiles, is_supported() false; non-macOS cfg path checked warning-free under -D warnings
Shim symbols present in the final tcode binary; is_supported() true at runtime on macOS 26.5
Live mic dictation in the running app (needs a human to approve the mic TCC prompt on first use)

Out of scope (per plan)

Settings section, cloud/BYOK backends, other platforms (Windows/Linux need a local ASR engine — see research doc), and conversational voice mode.

🤖 Generated with Claude Code

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Adds a mic toggle to the composer that streams on-device speech transcription into the editor. macOS 26+ only: the Swift shim (SpeechAnalyzer/SpeechTranscriber, AVAudioEngine capture) is compiled by build.rs when the SDK allows and the button does not exist elsewhere. Volatile hypotheses rewrite an anchored range in place; final results commit. Esc, submit, thread switch or typing stops the session. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The mic path stamped every AnalyzerInput with an explicit bufferStartTime and stopped via finalizeAndFinish(through: <cumulative time>). With its own timestamps the analyzer never reported results, and the targeted finalize waited on a timeline position that never arrived, so stop() hung and Ended was never delivered. Feed untimestamped buffers (the analyzer tracks the live timeline itself) and finalize through end of input, as validated against a standalone probe. Also treat an empty buffer from the priming rate converter as skippable rather than an error, and keep a mic_dictation example for live testing. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The shim reports a perceptual input level (RMS over a -50 dBFS window) at the audio callback rate as a new cosmetic Level event; the mic button renders it as a danger-tinted glow with fast-attack/slow-decay smoothing, so the user can see their speech is being picked up. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ze flush Two compounding reasons dictation text never appeared: - SpeechTranscriber with plain volatileResults buffers every result until finalization — nothing streams while speaking. DictationTranscriber's progressiveLongDictation preset (the dictation-purpose module; 53 locales incl. zh_CN/en_US) streams hypotheses every few hundred ms, verified against real audio driven through the exact mic pipeline. - The UI dropped the session and event pump immediately on stop, so even the finalization flush was discarded. Stop is now graceful (pump lives until the terminal event lands the flushed text); submit, destination switches and user edits abort instead, where a late flush would land in the wrong place. Also keeps a stderr diagnostic for mic authorization/terminal events. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…heck crates/voice/build.rs invokes xcrun/swiftc at build time; the Command::new boundary exists for runtime process spawning, which build scripts cannot route through the approved helpers. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…tub builds Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Tryanks · 2026-08-20T18:58:51Z

What

Voice input for the composer: a mic toggle button that streams on-device speech transcription into the message editor as you speak. macOS 26+ only — on every other platform/version the button does not exist (compile-time cfg + runtime is_supported()).

Research behind the approach: docs/voice-input-research.md. Plan: docs/plans/voice-input.md.

How

crates/voice (new) — tcode-voice: a Swift shim (swift/shim.swift, C ABI via @_cdecl) around SpeechAnalyzer/SpeechTranscriber, compiled and linked by build.rs only when targeting macOS with an SDK ≥ 26 (TCODE_VOICE_FORCE_STUB=1 forces the stub). Swift owns the whole audio path: AVAudioEngine mic tap → AVAudioConverter → AnalyzerInput stream; locale assets are installed on demand via AssetInventory. Events (Ready | Volatile | Final | Error | Ended) cross the FFI on arbitrary threads with a serialized, single-terminal contract.
Composer UI — mic button next to the send button (both row layouts). Click to start (spinner while assets/mic warm up), speech streams into the editor at an anchored insertion point: each volatile hypothesis rewrites the volatile tail in place, finals commit. Click again / Esc stops and keeps the text; submit, thread switch, or typing during dictation also stops it. All offsets are UTF-8 bytes (zh-CN safe).
Locale — follows the app locale: zh → zh_CN, otherwise en_US.
Packaging — NSMicrophoneUsageDescription + NSSpeechRecognitionUsageDescription added to the release Info.plist.

Test plan

cargo run -p tcode-voice --example file_dictation -- <aiff> zh_CN — 21 volatile events, correct final, clean exit (fixture from say -v Tingting)
cargo check --workspace, cargo clippy — clean (only the pre-existing block v0.1.6 future-incompat note)
cargo test -p tcode-ui — 254 passed, incl. locale key-set test and the new transcript-edit unit test
TCODE_VOICE_FORCE_STUB=1 build — stub path compiles, is_supported() false; non-macOS cfg path checked warning-free under -D warnings
Shim symbols present in the final tcode binary; is_supported() true at runtime on macOS 26.5
Live mic dictation in the running app (needs a human to approve the mic TCC prompt on first use)

Out of scope (per plan)

Settings section, cloud/BYOK backends, other platforms (Windows/Linux need a local ASR engine — see research doc), and conversational voice mode.

🤖 Generated with Claude Code

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Adds a mic toggle to the composer that streams on-device speech transcription into the editor. macOS 26+ only: the Swift shim (SpeechAnalyzer/SpeechTranscriber, AVAudioEngine capture) is compiled by build.rs when the SDK allows and the button does not exist elsewhere. Volatile hypotheses rewrite an anchored range in place; final results commit. Esc, submit, thread switch or typing stops the session. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The mic path stamped every AnalyzerInput with an explicit bufferStartTime and stopped via finalizeAndFinish(through: <cumulative time>). With its own timestamps the analyzer never reported results, and the targeted finalize waited on a timeline position that never arrived, so stop() hung and Ended was never delivered. Feed untimestamped buffers (the analyzer tracks the live timeline itself) and finalize through end of input, as validated against a standalone probe. Also treat an empty buffer from the priming rate converter as skippable rather than an error, and keep a mic_dictation example for live testing. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The shim reports a perceptual input level (RMS over a -50 dBFS window) at the audio callback rate as a new cosmetic Level event; the mic button renders it as a danger-tinted glow with fast-attack/slow-decay smoothing, so the user can see their speech is being picked up. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ze flush Two compounding reasons dictation text never appeared: - SpeechTranscriber with plain volatileResults buffers every result until finalization — nothing streams while speaking. DictationTranscriber's progressiveLongDictation preset (the dictation-purpose module; 53 locales incl. zh_CN/en_US) streams hypotheses every few hundred ms, verified against real audio driven through the exact mic pipeline. - The UI dropped the session and event pump immediately on stop, so even the finalization flush was discarded. Stop is now graceful (pump lives until the terminal event lands the flushed text); submit, destination switches and user edits abort instead, where a late flush would land in the wrong place. Also keeps a stderr diagnostic for mic authorization/terminal events. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…heck crates/voice/build.rs invokes xcrun/swiftc at build time; the Command::new boundary exists for runtime process spawning, which build scripts cannot route through the approved helpers. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…tub builds Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Tryanks · 2026-08-20T18:58:51Z

What

Voice input for the composer: a mic toggle button that streams on-device speech transcription into the message editor as you speak. macOS 26+ only — on every other platform/version the button does not exist (compile-time cfg + runtime is_supported()).

Research behind the approach: docs/voice-input-research.md. Plan: docs/plans/voice-input.md.

How

crates/voice (new) — tcode-voice: a Swift shim (swift/shim.swift, C ABI via @_cdecl) around SpeechAnalyzer/SpeechTranscriber, compiled and linked by build.rs only when targeting macOS with an SDK ≥ 26 (TCODE_VOICE_FORCE_STUB=1 forces the stub). Swift owns the whole audio path: AVAudioEngine mic tap → AVAudioConverter → AnalyzerInput stream; locale assets are installed on demand via AssetInventory. Events (Ready | Volatile | Final | Error | Ended) cross the FFI on arbitrary threads with a serialized, single-terminal contract.
Composer UI — mic button next to the send button (both row layouts). Click to start (spinner while assets/mic warm up), speech streams into the editor at an anchored insertion point: each volatile hypothesis rewrites the volatile tail in place, finals commit. Click again / Esc stops and keeps the text; submit, thread switch, or typing during dictation also stops it. All offsets are UTF-8 bytes (zh-CN safe).
Locale — follows the app locale: zh → zh_CN, otherwise en_US.
Packaging — NSMicrophoneUsageDescription + NSSpeechRecognitionUsageDescription added to the release Info.plist.

Test plan

cargo run -p tcode-voice --example file_dictation -- <aiff> zh_CN — 21 volatile events, correct final, clean exit (fixture from say -v Tingting)
cargo check --workspace, cargo clippy — clean (only the pre-existing block v0.1.6 future-incompat note)
cargo test -p tcode-ui — 254 passed, incl. locale key-set test and the new transcript-edit unit test
TCODE_VOICE_FORCE_STUB=1 build — stub path compiles, is_supported() false; non-macOS cfg path checked warning-free under -D warnings
Shim symbols present in the final tcode binary; is_supported() true at runtime on macOS 26.5
Live mic dictation in the running app (needs a human to approve the mic TCC prompt on first use)

Out of scope (per plan)

Settings section, cloud/BYOK backends, other platforms (Windows/Linux need a local ASR engine — see research doc), and conversational voice mode.

🤖 Generated with Claude Code

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Adds a mic toggle to the composer that streams on-device speech transcription into the editor. macOS 26+ only: the Swift shim (SpeechAnalyzer/SpeechTranscriber, AVAudioEngine capture) is compiled by build.rs when the SDK allows and the button does not exist elsewhere. Volatile hypotheses rewrite an anchored range in place; final results commit. Esc, submit, thread switch or typing stops the session. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The mic path stamped every AnalyzerInput with an explicit bufferStartTime and stopped via finalizeAndFinish(through: <cumulative time>). With its own timestamps the analyzer never reported results, and the targeted finalize waited on a timeline position that never arrived, so stop() hung and Ended was never delivered. Feed untimestamped buffers (the analyzer tracks the live timeline itself) and finalize through end of input, as validated against a standalone probe. Also treat an empty buffer from the priming rate converter as skippable rather than an error, and keep a mic_dictation example for live testing. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The shim reports a perceptual input level (RMS over a -50 dBFS window) at the audio callback rate as a new cosmetic Level event; the mic button renders it as a danger-tinted glow with fast-attack/slow-decay smoothing, so the user can see their speech is being picked up. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ze flush Two compounding reasons dictation text never appeared: - SpeechTranscriber with plain volatileResults buffers every result until finalization — nothing streams while speaking. DictationTranscriber's progressiveLongDictation preset (the dictation-purpose module; 53 locales incl. zh_CN/en_US) streams hypotheses every few hundred ms, verified against real audio driven through the exact mic pipeline. - The UI dropped the session and event pump immediately on stop, so even the finalization flush was discarded. Stop is now graceful (pump lives until the terminal event lands the flushed text); submit, destination switches and user edits abort instead, where a late flush would land in the wrong place. Also keeps a stderr diagnostic for mic authorization/terminal events. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…heck crates/voice/build.rs invokes xcrun/swiftc at build time; the Command::new boundary exists for runtime process spawning, which build scripts cannot route through the approved helpers. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…tub builds Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Tryanks · 2026-08-20T18:58:51Z

What

Voice input for the composer: a mic toggle button that streams on-device speech transcription into the message editor as you speak. macOS 26+ only — on every other platform/version the button does not exist (compile-time cfg + runtime is_supported()).

Research behind the approach: docs/voice-input-research.md. Plan: docs/plans/voice-input.md.

How

crates/voice (new) — tcode-voice: a Swift shim (swift/shim.swift, C ABI via @_cdecl) around SpeechAnalyzer/SpeechTranscriber, compiled and linked by build.rs only when targeting macOS with an SDK ≥ 26 (TCODE_VOICE_FORCE_STUB=1 forces the stub). Swift owns the whole audio path: AVAudioEngine mic tap → AVAudioConverter → AnalyzerInput stream; locale assets are installed on demand via AssetInventory. Events (Ready | Volatile | Final | Error | Ended) cross the FFI on arbitrary threads with a serialized, single-terminal contract.
Composer UI — mic button next to the send button (both row layouts). Click to start (spinner while assets/mic warm up), speech streams into the editor at an anchored insertion point: each volatile hypothesis rewrites the volatile tail in place, finals commit. Click again / Esc stops and keeps the text; submit, thread switch, or typing during dictation also stops it. All offsets are UTF-8 bytes (zh-CN safe).
Locale — follows the app locale: zh → zh_CN, otherwise en_US.
Packaging — NSMicrophoneUsageDescription + NSSpeechRecognitionUsageDescription added to the release Info.plist.

Test plan

cargo run -p tcode-voice --example file_dictation -- <aiff> zh_CN — 21 volatile events, correct final, clean exit (fixture from say -v Tingting)
cargo check --workspace, cargo clippy — clean (only the pre-existing block v0.1.6 future-incompat note)
cargo test -p tcode-ui — 254 passed, incl. locale key-set test and the new transcript-edit unit test
TCODE_VOICE_FORCE_STUB=1 build — stub path compiles, is_supported() false; non-macOS cfg path checked warning-free under -D warnings
Shim symbols present in the final tcode binary; is_supported() true at runtime on macOS 26.5
Live mic dictation in the running app (needs a human to approve the mic TCC prompt on first use)

Out of scope (per plan)

Settings section, cloud/BYOK backends, other platforms (Windows/Linux need a local ASR engine — see research doc), and conversational voice mode.

🤖 Generated with Claude Code

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Adds a mic toggle to the composer that streams on-device speech transcription into the editor. macOS 26+ only: the Swift shim (SpeechAnalyzer/SpeechTranscriber, AVAudioEngine capture) is compiled by build.rs when the SDK allows and the button does not exist elsewhere. Volatile hypotheses rewrite an anchored range in place; final results commit. Esc, submit, thread switch or typing stops the session. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The mic path stamped every AnalyzerInput with an explicit bufferStartTime and stopped via finalizeAndFinish(through: <cumulative time>). With its own timestamps the analyzer never reported results, and the targeted finalize waited on a timeline position that never arrived, so stop() hung and Ended was never delivered. Feed untimestamped buffers (the analyzer tracks the live timeline itself) and finalize through end of input, as validated against a standalone probe. Also treat an empty buffer from the priming rate converter as skippable rather than an error, and keep a mic_dictation example for live testing. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The shim reports a perceptual input level (RMS over a -50 dBFS window) at the audio callback rate as a new cosmetic Level event; the mic button renders it as a danger-tinted glow with fast-attack/slow-decay smoothing, so the user can see their speech is being picked up. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ze flush Two compounding reasons dictation text never appeared: - SpeechTranscriber with plain volatileResults buffers every result until finalization — nothing streams while speaking. DictationTranscriber's progressiveLongDictation preset (the dictation-purpose module; 53 locales incl. zh_CN/en_US) streams hypotheses every few hundred ms, verified against real audio driven through the exact mic pipeline. - The UI dropped the session and event pump immediately on stop, so even the finalization flush was discarded. Stop is now graceful (pump lives until the terminal event lands the flushed text); submit, destination switches and user edits abort instead, where a late flush would land in the wrong place. Also keeps a stderr diagnostic for mic authorization/terminal events. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…heck crates/voice/build.rs invokes xcrun/swiftc at build time; the Command::new boundary exists for runtime process spawning, which build scripts cannot route through the approved helpers. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…tub builds Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Tryanks · 2026-08-20T18:58:51Z

What

Voice input for the composer: a mic toggle button that streams on-device speech transcription into the message editor as you speak. macOS 26+ only — on every other platform/version the button does not exist (compile-time cfg + runtime is_supported()).

Research behind the approach: docs/voice-input-research.md. Plan: docs/plans/voice-input.md.

How

crates/voice (new) — tcode-voice: a Swift shim (swift/shim.swift, C ABI via @_cdecl) around SpeechAnalyzer/SpeechTranscriber, compiled and linked by build.rs only when targeting macOS with an SDK ≥ 26 (TCODE_VOICE_FORCE_STUB=1 forces the stub). Swift owns the whole audio path: AVAudioEngine mic tap → AVAudioConverter → AnalyzerInput stream; locale assets are installed on demand via AssetInventory. Events (Ready | Volatile | Final | Error | Ended) cross the FFI on arbitrary threads with a serialized, single-terminal contract.
Composer UI — mic button next to the send button (both row layouts). Click to start (spinner while assets/mic warm up), speech streams into the editor at an anchored insertion point: each volatile hypothesis rewrites the volatile tail in place, finals commit. Click again / Esc stops and keeps the text; submit, thread switch, or typing during dictation also stops it. All offsets are UTF-8 bytes (zh-CN safe).
Locale — follows the app locale: zh → zh_CN, otherwise en_US.
Packaging — NSMicrophoneUsageDescription + NSSpeechRecognitionUsageDescription added to the release Info.plist.

Test plan

cargo run -p tcode-voice --example file_dictation -- <aiff> zh_CN — 21 volatile events, correct final, clean exit (fixture from say -v Tingting)
cargo check --workspace, cargo clippy — clean (only the pre-existing block v0.1.6 future-incompat note)
cargo test -p tcode-ui — 254 passed, incl. locale key-set test and the new transcript-edit unit test
TCODE_VOICE_FORCE_STUB=1 build — stub path compiles, is_supported() false; non-macOS cfg path checked warning-free under -D warnings
Shim symbols present in the final tcode binary; is_supported() true at runtime on macOS 26.5
Live mic dictation in the running app (needs a human to approve the mic TCC prompt on first use)

Out of scope (per plan)

Settings section, cloud/BYOK backends, other platforms (Windows/Linux need a local ASR engine — see research doc), and conversational voice mode.

🤖 Generated with Claude Code

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Adds a mic toggle to the composer that streams on-device speech transcription into the editor. macOS 26+ only: the Swift shim (SpeechAnalyzer/SpeechTranscriber, AVAudioEngine capture) is compiled by build.rs when the SDK allows and the button does not exist elsewhere. Volatile hypotheses rewrite an anchored range in place; final results commit. Esc, submit, thread switch or typing stops the session. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The mic path stamped every AnalyzerInput with an explicit bufferStartTime and stopped via finalizeAndFinish(through: <cumulative time>). With its own timestamps the analyzer never reported results, and the targeted finalize waited on a timeline position that never arrived, so stop() hung and Ended was never delivered. Feed untimestamped buffers (the analyzer tracks the live timeline itself) and finalize through end of input, as validated against a standalone probe. Also treat an empty buffer from the priming rate converter as skippable rather than an error, and keep a mic_dictation example for live testing. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The shim reports a perceptual input level (RMS over a -50 dBFS window) at the audio callback rate as a new cosmetic Level event; the mic button renders it as a danger-tinted glow with fast-attack/slow-decay smoothing, so the user can see their speech is being picked up. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ze flush Two compounding reasons dictation text never appeared: - SpeechTranscriber with plain volatileResults buffers every result until finalization — nothing streams while speaking. DictationTranscriber's progressiveLongDictation preset (the dictation-purpose module; 53 locales incl. zh_CN/en_US) streams hypotheses every few hundred ms, verified against real audio driven through the exact mic pipeline. - The UI dropped the session and event pump immediately on stop, so even the finalization flush was discarded. Stop is now graceful (pump lives until the terminal event lands the flushed text); submit, destination switches and user edits abort instead, where a late flush would land in the wrong place. Also keeps a stderr diagnostic for mic authorization/terminal events. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…heck crates/voice/build.rs invokes xcrun/swiftc at build time; the Command::new boundary exists for runtime process spawning, which build scripts cannot route through the approved helpers. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…tub builds Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Tryanks · 2026-08-20T18:58:51Z

What

Voice input for the composer: a mic toggle button that streams on-device speech transcription into the message editor as you speak. macOS 26+ only — on every other platform/version the button does not exist (compile-time cfg + runtime is_supported()).

Research behind the approach: docs/voice-input-research.md. Plan: docs/plans/voice-input.md.

How

crates/voice (new) — tcode-voice: a Swift shim (swift/shim.swift, C ABI via @_cdecl) around SpeechAnalyzer/SpeechTranscriber, compiled and linked by build.rs only when targeting macOS with an SDK ≥ 26 (TCODE_VOICE_FORCE_STUB=1 forces the stub). Swift owns the whole audio path: AVAudioEngine mic tap → AVAudioConverter → AnalyzerInput stream; locale assets are installed on demand via AssetInventory. Events (Ready | Volatile | Final | Error | Ended) cross the FFI on arbitrary threads with a serialized, single-terminal contract.
Composer UI — mic button next to the send button (both row layouts). Click to start (spinner while assets/mic warm up), speech streams into the editor at an anchored insertion point: each volatile hypothesis rewrites the volatile tail in place, finals commit. Click again / Esc stops and keeps the text; submit, thread switch, or typing during dictation also stops it. All offsets are UTF-8 bytes (zh-CN safe).
Locale — follows the app locale: zh → zh_CN, otherwise en_US.
Packaging — NSMicrophoneUsageDescription + NSSpeechRecognitionUsageDescription added to the release Info.plist.

Test plan

cargo run -p tcode-voice --example file_dictation -- <aiff> zh_CN — 21 volatile events, correct final, clean exit (fixture from say -v Tingting)
cargo check --workspace, cargo clippy — clean (only the pre-existing block v0.1.6 future-incompat note)
cargo test -p tcode-ui — 254 passed, incl. locale key-set test and the new transcript-edit unit test
TCODE_VOICE_FORCE_STUB=1 build — stub path compiles, is_supported() false; non-macOS cfg path checked warning-free under -D warnings
Shim symbols present in the final tcode binary; is_supported() true at runtime on macOS 26.5
Live mic dictation in the running app (needs a human to approve the mic TCC prompt on first use)

Out of scope (per plan)

Settings section, cloud/BYOK backends, other platforms (Windows/Linux need a local ASR engine — see research doc), and conversational voice mode.

🤖 Generated with Claude Code

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Adds a mic toggle to the composer that streams on-device speech transcription into the editor. macOS 26+ only: the Swift shim (SpeechAnalyzer/SpeechTranscriber, AVAudioEngine capture) is compiled by build.rs when the SDK allows and the button does not exist elsewhere. Volatile hypotheses rewrite an anchored range in place; final results commit. Esc, submit, thread switch or typing stops the session. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The mic path stamped every AnalyzerInput with an explicit bufferStartTime and stopped via finalizeAndFinish(through: <cumulative time>). With its own timestamps the analyzer never reported results, and the targeted finalize waited on a timeline position that never arrived, so stop() hung and Ended was never delivered. Feed untimestamped buffers (the analyzer tracks the live timeline itself) and finalize through end of input, as validated against a standalone probe. Also treat an empty buffer from the priming rate converter as skippable rather than an error, and keep a mic_dictation example for live testing. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The shim reports a perceptual input level (RMS over a -50 dBFS window) at the audio callback rate as a new cosmetic Level event; the mic button renders it as a danger-tinted glow with fast-attack/slow-decay smoothing, so the user can see their speech is being picked up. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ze flush Two compounding reasons dictation text never appeared: - SpeechTranscriber with plain volatileResults buffers every result until finalization — nothing streams while speaking. DictationTranscriber's progressiveLongDictation preset (the dictation-purpose module; 53 locales incl. zh_CN/en_US) streams hypotheses every few hundred ms, verified against real audio driven through the exact mic pipeline. - The UI dropped the session and event pump immediately on stop, so even the finalization flush was discarded. Stop is now graceful (pump lives until the terminal event lands the flushed text); submit, destination switches and user edits abort instead, where a late flush would land in the wrong place. Also keeps a stderr diagnostic for mic authorization/terminal events. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…heck crates/voice/build.rs invokes xcrun/swiftc at build time; the Command::new boundary exists for runtime process spawning, which build scripts cannot route through the approved helpers. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…tub builds Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Tryanks · 2026-08-20T18:58:51Z

What

Voice input for the composer: a mic toggle button that streams on-device speech transcription into the message editor as you speak. macOS 26+ only — on every other platform/version the button does not exist (compile-time cfg + runtime is_supported()).

Research behind the approach: docs/voice-input-research.md. Plan: docs/plans/voice-input.md.

How

crates/voice (new) — tcode-voice: a Swift shim (swift/shim.swift, C ABI via @_cdecl) around SpeechAnalyzer/SpeechTranscriber, compiled and linked by build.rs only when targeting macOS with an SDK ≥ 26 (TCODE_VOICE_FORCE_STUB=1 forces the stub). Swift owns the whole audio path: AVAudioEngine mic tap → AVAudioConverter → AnalyzerInput stream; locale assets are installed on demand via AssetInventory. Events (Ready | Volatile | Final | Error | Ended) cross the FFI on arbitrary threads with a serialized, single-terminal contract.
Composer UI — mic button next to the send button (both row layouts). Click to start (spinner while assets/mic warm up), speech streams into the editor at an anchored insertion point: each volatile hypothesis rewrites the volatile tail in place, finals commit. Click again / Esc stops and keeps the text; submit, thread switch, or typing during dictation also stops it. All offsets are UTF-8 bytes (zh-CN safe).
Locale — follows the app locale: zh → zh_CN, otherwise en_US.
Packaging — NSMicrophoneUsageDescription + NSSpeechRecognitionUsageDescription added to the release Info.plist.

Test plan

cargo run -p tcode-voice --example file_dictation -- <aiff> zh_CN — 21 volatile events, correct final, clean exit (fixture from say -v Tingting)
cargo check --workspace, cargo clippy — clean (only the pre-existing block v0.1.6 future-incompat note)
cargo test -p tcode-ui — 254 passed, incl. locale key-set test and the new transcript-edit unit test
TCODE_VOICE_FORCE_STUB=1 build — stub path compiles, is_supported() false; non-macOS cfg path checked warning-free under -D warnings
Shim symbols present in the final tcode binary; is_supported() true at runtime on macOS 26.5
Live mic dictation in the running app (needs a human to approve the mic TCC prompt on first use)

Out of scope (per plan)

Settings section, cloud/BYOK backends, other platforms (Windows/Linux need a local ASR engine — see research doc), and conversational voice mode.

🤖 Generated with Claude Code

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Adds a mic toggle to the composer that streams on-device speech transcription into the editor. macOS 26+ only: the Swift shim (SpeechAnalyzer/SpeechTranscriber, AVAudioEngine capture) is compiled by build.rs when the SDK allows and the button does not exist elsewhere. Volatile hypotheses rewrite an anchored range in place; final results commit. Esc, submit, thread switch or typing stops the session. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The mic path stamped every AnalyzerInput with an explicit bufferStartTime and stopped via finalizeAndFinish(through: <cumulative time>). With its own timestamps the analyzer never reported results, and the targeted finalize waited on a timeline position that never arrived, so stop() hung and Ended was never delivered. Feed untimestamped buffers (the analyzer tracks the live timeline itself) and finalize through end of input, as validated against a standalone probe. Also treat an empty buffer from the priming rate converter as skippable rather than an error, and keep a mic_dictation example for live testing. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The shim reports a perceptual input level (RMS over a -50 dBFS window) at the audio callback rate as a new cosmetic Level event; the mic button renders it as a danger-tinted glow with fast-attack/slow-decay smoothing, so the user can see their speech is being picked up. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ze flush Two compounding reasons dictation text never appeared: - SpeechTranscriber with plain volatileResults buffers every result until finalization — nothing streams while speaking. DictationTranscriber's progressiveLongDictation preset (the dictation-purpose module; 53 locales incl. zh_CN/en_US) streams hypotheses every few hundred ms, verified against real audio driven through the exact mic pipeline. - The UI dropped the session and event pump immediately on stop, so even the finalization flush was discarded. Stop is now graceful (pump lives until the terminal event lands the flushed text); submit, destination switches and user edits abort instead, where a late flush would land in the wrong place. Also keeps a stderr diagnostic for mic authorization/terminal events. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…heck crates/voice/build.rs invokes xcrun/swiftc at build time; the Command::new boundary exists for runtime process spawning, which build scripts cannot route through the approved helpers. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…tub builds Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Tryanks · 2026-08-20T18:58:51Z

What

Voice input for the composer: a mic toggle button that streams on-device speech transcription into the message editor as you speak. macOS 26+ only — on every other platform/version the button does not exist (compile-time cfg + runtime is_supported()).

Research behind the approach: docs/voice-input-research.md. Plan: docs/plans/voice-input.md.

How

crates/voice (new) — tcode-voice: a Swift shim (swift/shim.swift, C ABI via @_cdecl) around SpeechAnalyzer/SpeechTranscriber, compiled and linked by build.rs only when targeting macOS with an SDK ≥ 26 (TCODE_VOICE_FORCE_STUB=1 forces the stub). Swift owns the whole audio path: AVAudioEngine mic tap → AVAudioConverter → AnalyzerInput stream; locale assets are installed on demand via AssetInventory. Events (Ready | Volatile | Final | Error | Ended) cross the FFI on arbitrary threads with a serialized, single-terminal contract.
Composer UI — mic button next to the send button (both row layouts). Click to start (spinner while assets/mic warm up), speech streams into the editor at an anchored insertion point: each volatile hypothesis rewrites the volatile tail in place, finals commit. Click again / Esc stops and keeps the text; submit, thread switch, or typing during dictation also stops it. All offsets are UTF-8 bytes (zh-CN safe).
Locale — follows the app locale: zh → zh_CN, otherwise en_US.
Packaging — NSMicrophoneUsageDescription + NSSpeechRecognitionUsageDescription added to the release Info.plist.

Test plan

cargo run -p tcode-voice --example file_dictation -- <aiff> zh_CN — 21 volatile events, correct final, clean exit (fixture from say -v Tingting)
cargo check --workspace, cargo clippy — clean (only the pre-existing block v0.1.6 future-incompat note)
cargo test -p tcode-ui — 254 passed, incl. locale key-set test and the new transcript-edit unit test
TCODE_VOICE_FORCE_STUB=1 build — stub path compiles, is_supported() false; non-macOS cfg path checked warning-free under -D warnings
Shim symbols present in the final tcode binary; is_supported() true at runtime on macOS 26.5
Live mic dictation in the running app (needs a human to approve the mic TCC prompt on first use)

Out of scope (per plan)

Settings section, cloud/BYOK backends, other platforms (Windows/Linux need a local ASR engine — see research doc), and conversational voice mode.

🤖 Generated with Claude Code

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Adds a mic toggle to the composer that streams on-device speech transcription into the editor. macOS 26+ only: the Swift shim (SpeechAnalyzer/SpeechTranscriber, AVAudioEngine capture) is compiled by build.rs when the SDK allows and the button does not exist elsewhere. Volatile hypotheses rewrite an anchored range in place; final results commit. Esc, submit, thread switch or typing stops the session. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The mic path stamped every AnalyzerInput with an explicit bufferStartTime and stopped via finalizeAndFinish(through: <cumulative time>). With its own timestamps the analyzer never reported results, and the targeted finalize waited on a timeline position that never arrived, so stop() hung and Ended was never delivered. Feed untimestamped buffers (the analyzer tracks the live timeline itself) and finalize through end of input, as validated against a standalone probe. Also treat an empty buffer from the priming rate converter as skippable rather than an error, and keep a mic_dictation example for live testing. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

The shim reports a perceptual input level (RMS over a -50 dBFS window) at the audio callback rate as a new cosmetic Level event; the mic button renders it as a danger-tinted glow with fast-attack/slow-decay smoothing, so the user can see their speech is being picked up. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…ze flush Two compounding reasons dictation text never appeared: - SpeechTranscriber with plain volatileResults buffers every result until finalization — nothing streams while speaking. DictationTranscriber's progressiveLongDictation preset (the dictation-purpose module; 53 locales incl. zh_CN/en_US) streams hypotheses every few hundred ms, verified against real audio driven through the exact mic pipeline. - The UI dropped the session and event pump immediately on stop, so even the finalization flush was discarded. Stop is now graceful (pump lives until the terminal event lands the flushed text); submit, destination switches and user edits abort instead, where a late flush would land in the wrong place. Also keeps a stderr diagnostic for mic authorization/terminal events. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…heck crates/voice/build.rs invokes xcrun/swiftc at build time; the Command::new boundary exists for runtime process spawning, which build scripts cannot route through the approved helpers. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…tub builds Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Tryanksand others added 7 commits August 21, 2026 02:40

feat(voice): scaffold tcode-voice crate with pinned dictation API
82a9dfc
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

fix(voice): drop_non_drop clippy error in file_dictation example on s…
101781e
…tub builds Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Tryanks merged commit e8a5f1c into mainAug 20, 2026
3 checks passed

Tryanks deleted the feat/voice-input branch August 20, 2026 19:33

Tryanks mentioned this pull request Aug 22, 2026
Roadmap: near-term priorities #125
Closed
12 tasks

Tryanksand others added 7 commits August 21, 2026 02:40

feat(voice): scaffold tcode-voice crate with pinned dictation API
82a9dfc
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

fix(voice): drop_non_drop clippy error in file_dictation example on s…
101781e
…tub builds Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Tryanks merged commit e8a5f1c into mainAug 20, 2026
3 checks passed

Tryanks deleted the feat/voice-input branch August 20, 2026 19:33

Tryanks mentioned this pull request Aug 22, 2026
Roadmap: near-term priorities #125
Closed
12 tasks

Tryanksand others added 7 commits August 21, 2026 02:40

feat(voice): scaffold tcode-voice crate with pinned dictation API
82a9dfc
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

fix(voice): drop_non_drop clippy error in file_dictation example on s…
101781e
…tub builds Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Tryanks merged commit e8a5f1c into mainAug 20, 2026
3 checks passed

Tryanks deleted the feat/voice-input branch August 20, 2026 19:33

Tryanks mentioned this pull request Aug 22, 2026
Roadmap: near-term priorities #125
Closed
12 tasks

Tryanksand others added 7 commits August 21, 2026 02:40

feat(voice): scaffold tcode-voice crate with pinned dictation API
82a9dfc
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

fix(voice): drop_non_drop clippy error in file_dictation example on s…
101781e
…tub builds Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Tryanks merged commit e8a5f1c into mainAug 20, 2026
3 checks passed

Tryanks deleted the feat/voice-input branch August 20, 2026 19:33

Tryanks mentioned this pull request Aug 22, 2026
Roadmap: near-term priorities #125
Closed
12 tasks

Tryanksand others added 7 commits August 21, 2026 02:40

feat(voice): scaffold tcode-voice crate with pinned dictation API
82a9dfc
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

fix(voice): drop_non_drop clippy error in file_dictation example on s…
101781e
…tub builds Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Tryanks merged commit e8a5f1c into mainAug 20, 2026
3 checks passed

Tryanks deleted the feat/voice-input branch August 20, 2026 19:33

Tryanks mentioned this pull request Aug 22, 2026
Roadmap: near-term priorities #125
Closed
12 tasks

Tryanksand others added 7 commits August 21, 2026 02:40

feat(voice): scaffold tcode-voice crate with pinned dictation API
82a9dfc
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

fix(voice): drop_non_drop clippy error in file_dictation example on s…
101781e
…tub builds Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Tryanks merged commit e8a5f1c into mainAug 20, 2026
3 checks passed

Tryanks deleted the feat/voice-input branch August 20, 2026 19:33

Tryanks mentioned this pull request Aug 22, 2026
Roadmap: near-term priorities #125
Closed
12 tasks

Tryanksand others added 7 commits August 21, 2026 02:40

feat(voice): scaffold tcode-voice crate with pinned dictation API
82a9dfc
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

fix(voice): drop_non_drop clippy error in file_dictation example on s…
101781e
…tub builds Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Tryanks merged commit e8a5f1c into mainAug 20, 2026
3 checks passed

Tryanks deleted the feat/voice-input branch August 20, 2026 19:33

Tryanks mentioned this pull request Aug 22, 2026
Roadmap: near-term priorities #125
Closed
12 tasks

Tryanksand others added 7 commits August 21, 2026 02:40

feat(voice): scaffold tcode-voice crate with pinned dictation API
82a9dfc
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

fix(voice): drop_non_drop clippy error in file_dictation example on s…
101781e
…tub builds Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Tryanks merged commit e8a5f1c into mainAug 20, 2026
3 checks passed

Tryanks deleted the feat/voice-input branch August 20, 2026 19:33

Tryanks mentioned this pull request Aug 22, 2026
Roadmap: near-term priorities #125
Closed
12 tasks

Conversation

Tryanks commented Aug 20, 2026

What

How

Test plan

Out of scope (per plan)

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

Tryanks commented Aug 20, 2026

What

How

Test plan

Out of scope (per plan)

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

Tryanks commented Aug 20, 2026

What

How

Test plan

Out of scope (per plan)

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

Tryanks commented Aug 20, 2026

What

How

Test plan

Out of scope (per plan)

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

Tryanks commented Aug 20, 2026

What

How

Test plan

Out of scope (per plan)

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant