feat(speech): on-demand offline speech engine via verified release asset (PRD 10.2.4) - #760
Merged
Merged
Conversation
…set (PRD 10.2.4) First safe slice of the AI footprint/optimization plan: a reliable acquisition path that fetches a redistributable component from QUILL's own pinned, SHA-256-verified GitHub release asset. The engine stays bundled in the installer, so this is a recovery/optional path and capability never depends on a download. - quill/core/release_assets.py (new, wx-free): pinned manifest (whisper.cpp -> our assets-v1 release, SHA-256), is_pinned() guard (refuses placeholder SHAs / moving refs), HTTPS-only retry + RESUMABLE download (HTTP Range), atomic verified install (temp -> verify -> copy), Safe-Mode gate, clean errors for graceful degradation. Verified end-to-end against the published asset (downloads, verifies, unpacks whisper-cli.exe + DLLs). - Tools > Speech > Download Offline Speech Engine... : download_offline_speech_engine() worker (mirrors download_ffmpeg) + menu item/id/binding. Installs into <data>/speech-engine, which resolve_whisper_executable() already searches. - GATE-9: new egress site audited. GATE-11: main_frame_speech/main_frame_menu rebaselined (handlers live with the other download_* siblings). - Published the MIT whisper.cpp v1.9.1 engine to the non-latest assets-v1 release. - Docs everywhere: CHANGELOG, release notes, user guide, main PRD (5.25c), planning roadmap + AI-Optimization PRD linkage. Regenerated HTML/EPUB for edited docs. Scope-limited per decision: only the MIT whisper.cpp engine is re-hosted; nothing is unbundled from the installer; license-unclear engines (DECtalk/eSpeak) and ffmpeg are NOT re-hosted. Tests: 6 unit tests (pin/guard/safe-mode/unknown/checksum-mismatch/unpack) pass; ruff + GATE-9 + GATE-11 green. Co-Authored-By: Claude Opus 4.8 <[email protected]>
…erization snapshot (#760) The on-demand engine download added a public MainFrame method; update the characterization fixture so the snapshot guard (and the unit-test suite) pass. Co-Authored-By: Claude Opus 4.8 <[email protected]>
accesswatch
added a commit
that referenced
this pull request
Jun 29, 2026
…downloads (PRD 10.2.4) (#764) * feat(packaging): unbundle Kokoro voices -> on-demand verified download (PRD 10.2.4) Proof-of-concept unbundle that actually shrinks the installer ~120 MB, taken safely. The Kokoro neural voices are no longer staged into the installer; they download on demand from QUILL's pinned, SHA-256-verified release asset (kokoro-models.zip on assets-v1) to %APPDATA%/kokoro-models, which the runtime prefers. Upgraders are protected automatically (verified): Inno only overlays new [Files] and never removes old ones, [InstallDelete] does not touch kokoro-models, and runtime resolution still checks {app}/kokoro-models -- so a user upgrading from a release that bundled Kokoro keeps their copy with nothing to re-download. Kokoro is "safe to unbundle": other read-aloud voices (DECtalk/eSpeak/Piper/SAPI) work offline meanwhile. - release_assets.py: add the pinned "kokoro" entry; add should_cancel/DownloadCancelled (cancel handled in-module, not via a swallowed progress-raise) and a `version` field (groundwork for newer-version notifications); the whisper handler uses should_cancel too. - _download_kokoro_models: route through release_assets; SMART re-download -- if the voices are already present it offers to replace rather than silently re-fetching ~120 MB, and declining keeps the existing copy. - build: stop staging Kokoro by default (still possible via --kokoro-dir for portable); remove the speechkokoro component + kokoro-models [Files] from the installer; iss regenerated; iss test updated to assert absence. - Docs everywhere (CHANGELOG, release notes, user guide, PRD 5.25c, roadmap). - Builds on #760 (release_assets). GATE-9/GATE-11 green; e2e verified against the published asset (downloads, verifies, unpacks the 88 MB model + 26 MB voices). Co-Authored-By: Claude Opus 4.8 <[email protected]> * feat(packaging): unbundle whisper.cpp too -> on-demand verified download (PRD 10.2.4) Per decision, fully unbundle the offline speech engine like Kokoro. whisper.cpp is no longer staged into the installer; fresh installs download the ~8 MB engine on demand from QUILL's pinned, SHA-256-verified release asset (assets-v1) to %APPDATA%/speech-engine, which the runtime searches. Protected at point of use (it is the default offline engine): the dictation pre-flight now OFFERS to download it ("Download it now? ~8 MB, verified") instead of just pointing at a menu, so a fresh install is never stranded. The provider's not-installed guidance points at Tools > Speech > Download Offline Speech Engine. Upgraders keep their existing {app}/tools/speech/whispercpp copy automatically (Inno never removes it; [InstallDelete] does not touch it; resolver still checks it). - build: stop auto-staging whisper (still stageable via --whisper-dir); remove the speechwhisper component + whispercpp [Files]; iss regenerated. - tests updated (bundledTools/speechAssets/iss assertions assert absence). - docs updated everywhere (CHANGELOG, release notes, user guide, PRD 5.25c). - GATE-9/GATE-11 green; main_frame_dictation_hotkeys rebaselined (+12 for the offer). Co-Authored-By: Claude Opus 4.8 <[email protected]> --------- Co-authored-by: Claude Opus 4.8 <[email protected]>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
The first safe slice of the AI footprint/optimization plan: a reliable, verified path for fetching a redistributable component from QUILL's own GitHub release asset, wired as Tools → Speech → Download Offline Speech Engine... The whisper.cpp engine stays bundled in the installer, so this is a recovery/optional path — capability never depends on a download.
Decision honored (no risk)
assets-v1release (tag does not matchv*, so it never triggers the release workflow or the update feed).Reliable acquisition (PRD 10.2.3/10.2.4)
quill/core/release_assets.py(new, wx-free):is_pinned()refuses placeholder SHAs and moving refs (latest/head/main/master).Installs into
<data>/speech-engine, whichresolve_whisper_executable()already searches.Verified end-to-end
Ran the real path against the published asset: it downloads from
…/releases/download/assets-v1/whisper-bin-x64.zip, SHA-256-verifies, and unpackswhisper-cli.exe+ 14 DLLs. Plus 6 unit tests (pin/guard, Safe Mode, unknown component, unpinned refusal, checksum mismatch, verify-then-unpack) — no network.UI
download_offline_speech_engine()mirrors the provendownload_ffmpeghandler (confirm → cancelable progress → worker thread → SHA-verified install), with the #748 whole-percent throttle. New "Download Offline Speech Engine..." item in Tools → Speech.Docs (all sources)
CHANGELOG, release notes (0.8.1 Beta 1), user guide (Offline transcription section), main PRD (§5.25c), planning roadmap (§1.2) + linkage to the AI-Optimization PRD. HTML/EPUB regenerated for every edited doc.
Gates
ruff clean; GATE-9 (no-silent-network) OK; GATE-11 budgets rebaselined (handlers live with the other
download_*siblings) with a dated note; mypy clean on the core module.🤖 Generated with Claude Code