[screen-gd]

Problem

The redesign removed Model folders, so users could no longer reuse voice models downloaded by another app. Dictation also had no sound to indicate when recording started.

Change

Restore Model folders in Voice settings. T3 verifies models before loading them and never changes files in those folders.

Add an optional recording sound, off by default, with seven choices. It plays when recording starts; retrying saved audio stays silent. Available on web and desktop.

The sounds are unchanged from Hex. Its MIT license permits reuse with attribution. The full copyright and license notice are included in T3's web and desktop notices.

Scope and approval

Model folders restores the behavior from merged #30224 on the redesign branch. The recording sound is a separate addition; no maintainer approval to include it here is linked yet.

Verification

• Folder tests cover reusing verified models, rejecting corrupt files and leaving external files untouched.
• Changed-file checks passed, including 209 tests. All 11 Voice settings browser tests and the desktop build also passed.
• All seven sounds match Hex's originals; generated web and desktop notices contain the full license.
• One separate license test fails on an existing Windows path-format assertion. Audible recording-start playback and macOS/Linux were not checked with the current sounds.

Screenshots show the folder fix; they predate the sound picker.

Before	After

Made with GPT-6.1 Sol in Codex and Claude Opus 5.5 in Claude Code, through T3 Code.

[macroscopeapp]

Approvability

Verdict: Approved at 86cb4a6

Macroscope's review found this PR approvable — Adds optional model folders so speech can reuse verified catalog downloads without modifying external files. Existing behavior remains unchanged until folders are configured, and model deletion remains limited to app-owned storage.

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

[screen-gd]

Note

🤖 GPT-6.1 Sol responding on behalf of screen-gd

Updated to the latest redesign and resolved the conflicts. Model folders and the optional recording sounds are preserved. Hex's MIT attribution is included in the shipped notices. Checks passed; the description lists the remaining limits and approval needed for the sound cue.

This replaces the earlier progress updates. Current details and screenshots are in the PR description.

[binjamshid]

Restoring Model folders looks well contained: external files are SHA-256 verified before use, only the app-owned copy is ever removed, and the new endpoint sits behind AuthSettingsWriteScope. Most of my notes are about scope rather than logic.

1. medium, PR scope. The PR currently shows 100 commits / 295 files (+10.5k) against maria/redesign/6-workspace-panes, because the head merged a newer redesign than the base has. Only five commits (9934c4d, 56f7529, 8e78835, 20726eb, 751dfbf) are this change, so the voice work is hard to find among unrelated pane, sidebar, orchestration and mobile-theme diffs. A rebase onto the current base (or retargeting) would bring this down to the ~25 voice files and avoid reintroducing redesign commits if the base moves again.

2. low, apps/server/src/speech/SpeechModelStore.ts:83. verifiedExternal lives only in memory, so a shared model is fully re-hashed on the first transcription after every server restart. For large GGUF files this adds noticeable first-use latency each launch. That may be the right tradeoff for "never write into external folders", but a one-line comment, or a verification marker kept under T3's own state dir and keyed by path + identity, would make it explicit.

3. low, apps/server/src/speech/SpeechService.ts:976. Dedupe is only trim(), so /models and /models/ (or C:\models vs C:/models) are stored as two folders and both are scanned. path.normalize before the Set would fix it. Separately, updateModelDirectories always calls disposeModel(), even when the list is unchanged or only an unrelated folder was removed. That forces a cold reload on the next dictation. Skipping the dispose when the normalized list equals the current setting would avoid it.

4. low, apps/web/src/speech/useEnvironmentSpeechInput.ts:401-402. The cue plays when the controller emits recording. In controller.ts, that happens right after recorder.record(), so on laptop speakers the cue can end up at the start of the captured audio and the live stream. Hex starts its sound before capture begins. Playing it just before record(), or briefly delaying capture, would keep it out of the transcript input. This is worth checking once audibly, since the description notes playback wasn't verified.

5. low, apps/web/src/components/settings/VoiceSettingsPanel.tsx. The client-side valid regex accepts \\server without a share segment, but the server rejects that on win32 with invalid_command. The user then sees a generic "Could not update model folders". Either align the regex with the server's UNC check or surface the invalid-directory reason.

Process: the description already notes this bundles the folder restore with the new sound cue, and the cue has no linked approval. Splitting the cue into its own PR would let the restore land independently.

[screen-gd]

Note

🤖 Claude Opus 5.5 responding on behalf of screen-gd

Closing this. #34096, now in maria/redesign/6-workspace-panes, covers the model folder restore: Download checks Handy and the Hugging Face cache, and Use existing links a verified file or folder without copying or deleting it. Porting the folder setting on top would add a second way to do the same thing, so review points 2, 3 and 5 no longer apply.

The sound cue moved to #34265, rebuilt on the current base as one commit (18 files instead of 1,094). It also fixes point 4: the cue now plays and finishes before recorder.record(), and controller.test.ts asserts that order.