feat: add local training mode review and export - #1
Open
djurcola-agent wants to merge 144 commits into
Open
Conversation
in the debug menu
If recording is triggered but nothing is said, the empty transcription was still sent to the post-processing LLM, which replies with a useless error message such as "you need to provide the transcription". Guard `post_process_transcription()` (the single funnel for all post-processing paths) to return early on empty or whitespace-only input, so blank recordings no longer trigger a model call. Fixes cjpais#1409 Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
…ed (cjpais#1447) On Linux the recording_overlay window is created at boot regardless of the overlay_position setting, so even with overlay_position: none (the Linux default) its hidden WebKit subprocess receives the mic-level events emit_levels broadcasts at ~24 Hz during recording. Each event drives WebKit C++ allocations that accumulate without bound, reaching multi-GB RSS over long sessions and ending in an OOM kill. Cache the overlay-enabled state in an AtomicBool, populated at startup and kept in sync from change_overlay_position_setting. emit_levels does a single relaxed atomic load and returns early when the overlay is disabled, so the audio path emits nothing for the default Linux config. When the overlay is enabled, replace the dual global broadcast with a single emit_to targeting the overlay window, halving the per-callback WebKit dispatch work. Fixes cjpais#1279
* Add Dutch translation file for application This commit adds a new Dutch translation file for the application, containing various UI strings and model descriptions. * Add Dutch language support
* move to auto timestamps for all models * remove comment, just use defaults
…ompt (cjpais#1603) * fix: gate whisper run extension on model arch, not Feature::InitialPrompt Non-whisper models such as Voxtral Small 24B advertise Feature::InitialPrompt but reject the whisper-kind run extension with INVALID_ARG, which made every offline transcription fail (cjpais#1601). Probe model.arch() == "whisper" once per run and gate on it: - the whisper run extension (custom words as initial_prompt) - the custom_words_already_prompted flag for the fuzzy fallback, so models like Voxtral still get fuzzy custom-word correction Complements cjpais#1602, which handles the timestamps half of the bug. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs: update comments per Copilot review nits Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Alberto Aldas <12873489+btoaldas@users.noreply.github.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* Update Italian translations * fix * format --------- Co-authored-by: CJ Pais <cj@cjpais.com>
Co-authored-by: Piste <5099194+Piste@users.noreply.github.com>
cjpais#1510) * fix(build): auto-fall-back to AI stub on Command Line Tools-only macOS The Command Line Tools SDK ships FoundationModels.framework, so the existing `framework_path.exists()` check selected the real Apple Intelligence Swift path even on toolchains that cannot compile it: the CLT `swiftc` has no FoundationModelsMacros plugin (full Xcode only), so the build hard-fails for anyone without Xcode installed (issue cjpais#1448). Detect a CLT-only toolchain via `xcode-select -p` and fall back to the stub automatically, emitting a cargo:warning that points at installing Xcode or the existing HANDY_FORCE_AI_STUB=1 explicit override. This lets a plain `cargo build` / `tauri dev` succeed out of the box on a stock developer machine while keeping the real path for full-Xcode builds. The auto-detect is skipped when SWIFTC is overridden, so the existing non-Xcode standalone-swift route (e.g. nixpkgs) is not forced to the stub. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * nits --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: CJ Pais <cj@cjpais.com>
…jpais#1444) Resolves cjpais#1279. - Throttle emit_levels to ~30 FPS (33ms) via atomic timestamp - Replace dual global .emit() with single .emit_to("recording_overlay") to eliminate redundant evaluate_script calls (~188/s → ~30/s) Co-authored-by: CJ Pais <cj@cjpais.com>
…alk (cjpais#1344) The FrameResampler was created once per audio session and reused across multiple recordings without clearing its internal buffers. The rubato FftFixedIn overlap buffers retained audio from the previous recording, causing the first ~30ms of each new recording to contain stale samples. This manifested as lost/corrupted words at the start of transcriptions and occasional stale text fragments (e.g. "ology.") from previous sessions leaking into new ones. Fix: add FrameResampler::reset() that clears in_buf, pending, and calls rubato's reset() to zero the FFT overlap buffers. Called on Cmd::Start alongside the existing processed_samples and VAD resets. Co-authored-by: Andrew Leech <andrew.leech@planetinnovation.com.au> Co-authored-by: CJ Pais <cj@cjpais.com>
…cjpais#1354) During app termination, macOS sends a quit event which triggers `exit()` → `__cxa_finalize_ranges` → Rust Drop handlers. If a mutex was poisoned by a prior panic anywhere in the transcription pipeline, calling `.lock().unwrap()` inside a Drop impl triggers a second panic. Rust treats panic-during-Drop as unrecoverable and calls `abort()`, crashing the app instead of exiting cleanly. Replace `.lock().unwrap()` with `match` + `into_inner()` in both `TranscriptionManager::drop()` and `LoadingGuard::drop()` so they gracefully recover poisoned mutexes during cleanup. Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* wip filler fixes * translations * add lid + use transcribe.cpp lid * minor tweaks
* fix(linux): avoid eager enigo requirement in paste path Dispatch paste methods before touching Enigo so PasteMethod::None and Linux-native paths no longer fail early when Enigo is unavailable. Reuse a shared Wayland-aware clipboard write helper for both paste and CopyToClipboard, enable Enigo's Wayland backend, and add regression tests for the non-Enigo None path. * fix(linux): remove Enigo Wayland feature The lazy Enigo paste fix does not require changing Enigo backend behavior. Restore the upstream feature set and remove its lockfile additions. * fix: preserve clipboard across paste fallback failures --------- Co-authored-by: CJ Pais <cj@cjpais.com>
Co-authored-by: Artem Shishkin <ArtemiyShisha@users.noreply.github.com>
…bar (cjpais#1659) * fix(theme): apply appearance setting to the overlay and macOS title bar The [data-theme] override blocks lived in App.css, which only the settings window loads. RecordingOverlay.css imports styles/theme.css alone, so the overlay never saw them and followed prefers-color-scheme instead. Move the blocks into theme.css, apply the stored theme in the overlay before render, and emit a theme-changed event so it updates live. Widen apply_window_theme from Windows-only to Windows and macOS. The main window is built with default decorations, so on macOS AppKit paints the title bar from NSApp.effectiveAppearance regardless of what the webview does. System clears the override so both follow the OS again. * clean up css --------- Co-authored-by: CJ Pais <cj@cjpais.com>
Co-authored-by: Mustafa Özkan <leongrphc@gmail.com>
Owner
|
Approved. |
Keep native overlay focus, hit testing, timeout provenance, and replacement transitions under the Rust owner. Preserve Custom Reading progression and training history provenance while preventing stale review operations from mutating newer state.
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
20 participants
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.
Before Submitting This PR
Human Written Description
TODO for Dan: Please replace this section with your own 2–3 sentence description of the problem this local Training Mode solves and why it matters. This placeholder is intentional to comply with Handy’s contribution requirements.
Related Issues/Discussions
Fixes #
Discussion: None. This is an experimental, local-only feature in an owner fork.
Community Feedback
None gathered. Handy is in feature freeze; this draft is for owner review and build-artifact testing, not an upstream merge request.
Testing
cargo fmt --checkcargo test managers::history::tests— 19 passedcargo test— 218 passedgit diff --checkFrontend package validation was not run because Bun and
node_moduleswere absent; no dependencies were installed.Screenshots/Videos (if applicable)
Not included; desktop interaction needs artifact-based testing.
AI Assistance
If AI was used: