New crates/transcription/src/transcriber.rs defines a Transcriber
trait (Send supertrait for spawn_blocking travel) with
TranscriberCapabilities (sample_rate, channels, supports_initial_prompt).
TranscriberCapabilities.sample_rate is load-bearing for the upcoming
progressive WAV writer (#19).
Concrete impls:
- SpeechModelAdapter wraps Box<dyn transcribe_rs::SpeechModel + Send>
for Parakeet (and any future transcribe-rs-backed engine).
- WhisperRsBackend moves its transcribe_sync body into the impl,
widening the signature from &self to &mut self so per-call
WhisperState can be created cleanly through the trait object.
LocalEngine now holds Box<dyn Transcriber + Send>; dispatch in
transcribe_sync collapses from a match to a direct call. Adds
LocalEngine::capabilities() for the WAV-writer.
Cargo feature flag "whisper" (default on) makes whisper-rs, num_cpus,
and the whole whisper_rs_backend module optional. cargo check
--no-default-features -p kon-transcription now builds without pulling
whisper-rs-sys — the escape hatch brief item #6 / #13 called for on
Windows / non-AVX2 / cloud-only builds. load_whisper is cfg-gated
behind the same feature.
src-tauri/src/commands/models.rs load_model_from_disk returns
Box<dyn Transcriber + Send> instead of SpeechBackend; caller chain
(ensure_model_loaded, prewarm_default_model) is unchanged.
transcriber_trait_is_object_safe test lands alongside the trait as a
compile-time witness against future Self-returning / generic-method
additions.
Review feedback: the previous test would pass even if the Range-header
logic in download_file were deleted entirely, because File::create
truncates the stale .part regardless of which branch set
actually_resuming to false.
Tightens spawn_no_range_server to return HTTP 400 when the request
carries no Range header, and only return 200 + full body when Range is
present. A regression that stops sending Range now surfaces as a
download failure (empty body from 400, bytes != body assertion)
instead of silently passing through the truncation path.
Review feedback: the original guard substring-searched the whole file
after the first "csp" token, which (a) false-passes if any unrelated
JSON value elsewhere in the config happens to contain a localhost URL
and (b) false-fails if the CSP is ever re-serialised with escaped
forward slashes.
Switches to serde_json + a /app/security/csp pointer lookup, then
splits the CSP on ';', finds the connect-src directive, tokenises its
allow-list on whitespace, and requires an exact match for both
http://127.0.0.1:* and ws://127.0.0.1:*. The error now also includes
the current connect-src value so a developer who breaks it can see
exactly what needs restoring.
The downloader already handles servers that return 200 to a Range
request by falling through to a truncating File::create on the .part
path, discarding stale partial bytes. That branch had no dedicated
fixture test — the SHA mismatch and Range-honouring resume cases were
covered but the mirror / proxy that strips Range support was not.
Adds spawn_no_range_server (always 200, full body regardless of Range
header) and download_file_restarts_when_server_ignores_range. Writes 12
bytes of stale content to .part, kicks off a download, asserts the
final file matches the fresh body exactly (not stale-bytes-prefixed)
and the .part file is cleaned up.
Brings the transcription downloader to test-coverage parity with
crates/llm/src/model_manager.rs per brief item #8 ("test coverage
parity" acceptance).
Pins the connect-src CSP entries for http://127.0.0.1:* and
ws://127.0.0.1:* at build time. If a future edit to tauri.conf.json
strips the local-LLM permit, the kon crate fails to build rather than
shipping a binary whose webview fetch() silently 404s with an opaque
scope error (Vibe #438 / #487).
Closes item #2 of docs/whisper-ecosystem/brief.md — CSP widening itself
landed in an earlier commit; this is the regression-proofing the plan
calls for.
resolveVolume previously clamped any value >=1 to full blast. If a
future settings change ever leaks a 0–100 scale through (instead of
0–1), the user gets jump-scared at max volume. Treat any v>1 as a
scale-drift bug and play at DEFAULT_VOLUME instead.
Also reject NaN / Infinity explicitly rather than relying on the
<=0 branch catching them.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Runs Mondays 06:00 UTC (plus workflow_dispatch) so any freshly
published advisory surfaces as its own failing run rather than
slipping into an unrelated PR's check.yml noise. npm audit is
gated to --audit-level=high to skip the low/moderate chatter that
doesn't warrant a bump.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
svelte-check catches type/template errors Vite's build skips; cargo
test --workspace --lib runs our pure-unit suites (prompt contract,
hallucination filter, preset parsing) without GPU or runtime deps.
Test step is Linux-only so the Windows/macOS legs stay focused on
platform compile coverage.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two new Settings → AI knobs that compose cleanly with what already
shipped (aiTier, LLM model, translator prompt framing).
**B.1 #15 — Named cleanup presets.** LlmPromptPreset enum
(Default / Email / Notes / Code) appends a short context hint onto
the CLEANUP_PROMPT just before generation. Presets shape tone and
structure ("email paragraph", "bulleted meeting notes", "preserve
technical terms") without licensing the content-editing the
translator-not-editor framing forbids. cleanup_transcript_text_cmd
now takes `preset: Option<String>` which runs through the new
LlmPromptPreset::parse (normalises aliases like "meeting-notes",
collapses unknown values to Default).
**A.1 #28 — Sequential-GPU guard.** New LocalEngine::unload drops
the backend + model_id so a subsequent load actually reclaims VRAM.
load_llm_model, load_model, and load_parakeet_model Tauri commands
grow an optional `concurrent: bool` argument. When concurrent is
Some(false), loading LLM first unloads whisper+parakeet, and vice
versa — prevents VRAM OOM on tight-VRAM setups. Default is the
previous parallel behaviour so nothing changes for multi-GB cards.
Transcribe-in-progress paths (transcribe_pcm, transcribe_file, live)
pass None, so mid-dictation model loads don't accidentally tear
down the LLM.
Settings UI (AI section):
- Cleanup preset segmented button + descriptive copy for each option.
- GPU concurrency segmented button with explicit trade-off text
("faster transitions vs fits in tight VRAM").
Frontend wiring:
- settings.llmPromptPreset flows from DictationPage's
cleanupTranscriptIfEnabled into the Tauri command.
- settings.aiGpuConcurrency flows from both DictationPage (auto-load
on record) and SettingsPage (manual load/unload buttons) as
`concurrent: "parallel" === true` to the load commands.
Tests: three new preset cases in crates/ai-formatting/src/llm_client.rs
(parse aliases, suffix non-empty for non-default, default suffix
empty). All 139 existing lib tests still pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The brief's pain point is opaque load failures: llama-cpp-2's errors
bubble up as raw C++ strings ("cudaMalloc failed: out of memory",
"invalid gguf magic"). A user seeing that has no path to recovery.
New backend command test_llm_model runs a staged diagnostic:
1. Model not downloaded → `not-downloaded` + download hint.
2. File size ≤90% of expected → `incomplete` (stalled download)
+ re-download hint. Matters because llama-cpp-2 can segfault
on truncated GGUF rather than returning cleanly.
3. Requested model already loaded → `ready`, no side effects.
4. Otherwise attempt a real load. On failure, classify_llm_load_error
maps the raw string to one of:
- load-failed-vram (OOM / cudaMalloc / allocation)
- load-failed-corrupt (GGUF magic / unsupported format)
- load-failed-permission (permission denied / access denied)
- load-failed-other (catch-all)
Each category has a prewritten actionable hint pointing at the
specific Settings surface (tier picker, re-download, file perms).
classify_llm_load_error is pure-string and unit-tested — 8 cases
covering the main categories plus edge cases (OOM alias, Windows
"Access is denied", unknown errors). Ordered narrow-to-broad so
overlap doesn't misclassify.
Settings UI gets a "Test" button in the AI section's action row,
visible whenever the model is downloaded (both downloaded-idle and
loaded states). Shows inline hint below the status line when the
test surfaces one. Refreshes both local and global LLM status after
the test since a successful test implicitly loads the model.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Hands-off feedback for recording lifecycle: a short C5 pitch at
record start, a falling G4 at record stop, and a staggered C5–E5–G5
major third when the finalise flow completes. Synthesised at runtime
via the Web Audio API (OscillatorNode + linear-ramped GainNode
envelope) rather than shipping WAV assets — keeps the binary size
flat and lets us tweak timbre without touching bundled files.
Off by default. Settings → Output exposes the toggle with a volume
slider (0–100%, default 15%) and a "Test" button that plays the
completion cue so the user can confirm loudness without recording.
Hooked at three call sites in DictationPage:
- playStartCue after page.recording = true in startRecording
- playStopCue at the top of stopRecording
- playCompleteCue just before `saved = true` at the end of
finaliseTranscription's transcript-present branch
All three no-op when settings.soundCues is false. The Web Audio
context is lazily constructed on first cue (most browsers suspend
it until a user gesture — Tauri's webview inherits that). If the
AudioContext can't be built we silently drop the cue rather than
throwing.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ufal/whisper_streaming #161 documents the classic Whisper streaming
failure: on ambiguous audio the model falls into a prompt loop,
cascading a single token for 10+ words ("I I I I I I I I I I I…").
The chunk-boundary duplicate detector in live.rs doesn't catch
this — the repeat is within a single chunk, and the text is
technically novel so FTS is happy to keep it.
Fold the detection into is_hallucination as a third pass (after
HALLUCINATION_MARKERS substring-match and HALLUCINATION_TRAIL_PHRASES
exact-match). has_consecutive_repetition walks the token stream
(whitespace-split, lowercased) and returns true when any run of
≥REPETITION_RUN_THRESHOLD (4) identical tokens is found.
Threshold chosen deliberately: three consecutive matches appear in
normal speech ("no no no, that's wrong"), four almost never does.
Tests pin both sides — "I I I I I" detected, "no no no" allowed,
alternating patterns ("I am I am I am I am") allowed regardless of
length.
Phrase-level repetition ("thank you thank you thank you thank you")
is a documented companion failure mode but needs a sliding n-gram
matcher — deferred with a code comment flagging it.
No caller changes: post_process_segments already drops
is_hallucination hits when anti_hallucination is enabled.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Whisper was trained on subtitle corpora, so silence and room tone
trigger caption-style artefacts that the previous three-marker
blocklist ("[blank_audio]", "[music]", "[silence]") didn't catch:
"Thanks for watching!", "Please subscribe.", "ご視聴ありがとうござ
いました", "♪♪♪", etc. Documented in WhisperLive #185 / #246 and
ufal/whisper_streaming #121 as the top streaming-transcript-quality
issue after chunk-boundary repeats.
HALLUCINATION_MARKERS widens from 3 to 16 entries: all common
bracketed non-speech tags (applause / laughter / inaudible /
background noise / sounds), parens variants, and musical notation
(♪ / ♫). Still contains-match so the marker triggers even when
Whisper wraps it in other noise.
HALLUCINATION_TRAIL_PHRASES (renamed from AUTO_THANKS_PHRASES) jumps
from 4 to ~30 entries: YouTube sign-offs, subtitle-credit leakage,
and the two most common non-English variants (Japanese "thanks for
watching" + MBC Korean news sign-off). Stays exact-match so
legitimate dialogue containing "thanks" or "subscribe" mid-sentence
never gets dropped — a new regression test pins that invariant.
The <15-char length gate on trail phrases is removed; some of the
new entries (e.g. "please subscribe to our channel.") are longer.
Exact-match against a known list is safety enough.
No caller changes: post_process_segments already drops segments for
which is_hallucination returns true when anti_hallucination is on.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The LLM runtime has been quiet since it shipped in Phase 3 — users
had no surface-level signal that cleanup was loaded, warming, or
actively generating. Settings has verbose status text internally,
but a dictation-flow user never opens Settings during a run.
New: a shared $state store drives a small chip in the sidebar that
reflects the true LLM state in ≤500 ms of any transition (brief
item #31 acceptance). Five states:
off → hidden (user has aiTier === "off")
warming → model download or first load in flight; amber pulse
ready → loaded + idle; green dot
generating → cleanup_transcript_text_cmd or
extract_tasks_from_transcript_cmd in flight; accent
pulse with Sparkles icon
error → last operation failed; red dot with AlertTriangle
The store exposes three calls: refreshLlmStatus(aiTier) (polls the
backend), markGenerating(detail) / markGenerationDone(success).
DictationPage wraps its cleanup + extract calls in mark-generating
pairs. SettingsPage's LLM load / unload / delete / download paths
also refresh the global store so Settings-initiated transitions
surface in the sidebar immediately. The chip collapses to a
dot-only compact form when the sidebar is collapsed.
No backend changes — everything wires onto the existing
`get_llm_status` Tauri command.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Kon's ideology rule: raw Whisper output is the source of truth; LLM
cleanup is additive, never destructive. The preview overlay already
tracks both rawText and finalText across the listening → live →
cleanup → final phases — but until now the user had no one-click path
from final to raw if cleanup changed their meaning.
Frontend: a context-aware "Use raw" / "Copy raw" button appears in
the preview overlay's final phase, only when rawText and finalText
actually differ (Raw format mode or LLM-off leaves the button hidden).
Two behaviours depending on how the transcript reached the target:
- settings.autoPaste = true → invoke paste_text_replacing, which
sends the platform's undo keystroke to the focused app,
waits UNDO_PASTE_GAP_MS (60 ms) for the compositor / app to
process it, then pastes the raw transcript. The preview hides
itself beforehand so the keystroke doesn't race focus
(existing Wayland-hardening path).
- settings.autoPaste = false → nothing was pasted in the first
place, so just overwrite the clipboard with raw. User's own
paste yields raw.
Backend: new paste_text_replacing Tauri command plus a mirror of the
paste-backend matrix for undo (wtype -M ctrl z / xdotool key ctrl+z /
ydotool keycodes 29:1 44:1 44:0 29:0 / osascript cmd+z / SendKeys '^z').
Reuses the pick_linux_backend_order Wayland-vs-X11 preference.
Registered in the Tauri command handler.
Acceptance per the brief: "after paste, Ctrl+Z within 5 s replaces
LLM output with raw transcript" — satisfied via the 4 s auto-hide
window on the preview's final phase. The click extends auto-hide so
the user actually sees the confirmation.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
prewarm_default_model loaded the model and returned. That moves the
model into RAM, but whisper.cpp still allocates its context window +
fills GPU shader caches on the first inference call — producing the
~4–5 s cold-start latency documented in ufal/whisper_streaming #96
and #135 that feels like "Kon dropped my first sentence."
Extend the pre-warm task: after engine.load, feed one second of
silence (16000 zero samples at 16 kHz) through transcribe_sync with
default options. Silence returns empty segments; the *work* is the
context allocation, which now happens at app boot rather than on the
user's first hotkey press.
Net: the user's first real dictation should complete within ~1.5× the
steady-state RTF they'll see on subsequent runs, satisfying the A.1
#23 acceptance criterion. No new public API; all inside the existing
spawn_blocking.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The previous prompt led with "You are a transcript cleanup assistant"
and listed cleanup rules. That framing quietly licenses the LLM to
treat cleanup as content editing — rephrasing for clarity, summarising
long sentences, "improving" phrasing. That's precisely the failure
mode OpenWhispr / Scriberr / Whispering users complain about ("the
LLM changed my meaning").
New framing lifts Whispering's published baseline: "translator from
spoken to written form — not an editor trying to improve the content."
Adds an explicit rule: do NOT improve, summarise, expand, or rephrase;
faithful written-form translation only, never content editing.
Both load-bearing concerns are now regression-tested — the existing
prompt-injection hardening assertions stay, and a new test pins the
translator framing + explicit no-editing rule against drift during
future refactors.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Previously the repo had no root README. Pitch + architecture + crate-level
breakdown + routes + runtime stack + roadmap all lived in dated HANDOVER
docs + docs/brief/ + memory. This pulls the canonical overview into the
single place a new contributor / investor / cursory browser will actually
land first.
Sections: design principles, what Kon does today (shipped capabilities),
architecture (3-layer: Rust workspace → Tauri commands → Svelte UI),
crate-by-crate responsibilities, Tauri command surface, frontend shape,
runtime stack with pinned versions, platform support matrix, build + dev
setup, testing, pointers into docs/, roadmap, contributing notes, licence
placeholder, contact.
All claims are grounded in current code state (136 tests, 9 crates, 18
Tauri command modules, Tauri 2.10.3, whisper-rs 0.16, llama-cpp-2
0.1.144, sqlx 0.8.6). No marketing puff; preserves the ideology-firm
stance from memory.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds colors_and_type.css token system, fonts (Lexend, Instrument Serif Italic,
JetBrains Mono, Atkinson Hyperlegible Next, OpenDyslexic), SVG assets (wordmark,
waveform mark, grain), HTML preview spec cards, UI kit, and SKILL.md reference
under src/design-system/. Foundation for applying the new Kon visual identity.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Previously the bulk import ran new Set(...) on raw trimmed strings
before lowercasing, so 'ACME' and 'acme' both survived the first
dedupe pass. Neither existed in the store, so both got added —
defeating the commit message's claim that pasting the same block
twice with different casing is a no-op.
Collapse case variants at the initial dedupe step using a lowercase
seen-set, keeping the first occurrence's casing as written.
Co-authored-by: jars <jakejars@users.noreply.github.com>
The preview window was missing from the default capability's window
list. Its frontend calls core:window:allow-hide (getCurrentWindow().hide()
on dismiss / auto-hide) plus invoke('copy_to_clipboard') and a handful
of event listeners — none of those would have been permitted at runtime
because the capability never matched this window's label.
Add 'transcription-preview' alongside the other secondary windows so
the preview actually has access to the permissions it already relies on.
Regenerate gen/schemas/capabilities.json to match.
Co-authored-by: jars <jakejars@users.noreply.github.com>
Extends commands/paste.rs::paste_text with a pre-keystroke check:
if GetForegroundWindow (Windows) / xdotool getactivewindow (Linux
X11) / 'tell System Events ...' (macOS) reports a focused app class
matching KNOWN_TERMINAL_CLASSES, skip the synthesised Ctrl+V and
return an outcome with copied=true, pasted=false, and a user-facing
message ('Terminal detected (kitty) — use Ctrl+Shift+V to insert').
Matches Handy #692: Kitty/Alacritty/Windows Terminal/Codex CLI all
double-insert the transcript when a PTY sees both a synthesised
Ctrl+V and the terminal's own paste hotkey. The terminal list is
ordered most-specific-first so 'windowsterminal' wins over the
generic 'terminal' needle.
Adds classify_terminal() as a pure helper + seven unit tests. The
platform probe (detect_focused_window_class) isolates the fragile
shell-out so the classification rule is test-covered without
needing a real desktop. Wayland doesn't expose a reliable
focused-window API to unprivileged clients, so it conservatively
returns None and the normal paste path runs — consistent with
Kon's Wayland-Pipewire lane.
Prior-clipboard restore (#3) is intentionally skipped when terminal
mode takes over: the user's path to insert the transcript is their
own paste, which needs the transcript on the clipboard.
Co-authored-by: jars <jakejars@users.noreply.github.com>
Adds src/lib/utils/hotkeyValidity.ts with validateHotkey(combo, os)
and wires it into HotkeyRecorder.svelte. Rules:
- X11/Wayland/Linux: reject single-key combos unless the trigger
is F13..F24 (the conventional global-shortcut escape hatch).
- Windows: reject combos whose only modifier is a right-hand
variant (RCtrl/RAlt) — matches Handy #966's failure mode where
RegisterHotKey silently ignores them.
- macOS: reject Fn-only combos and bare-key combos (non-function
keys).
- unknown OS: pass through — better to ship a flawed combo than
reject one we can't validate.
When validation fails, the recorder leaves the previous known-good
hotkey intact and surfaces an inline warning-coloured sentence
explaining why, with actionable copy ("add a modifier", "use the
left-hand equivalent", etc.). The save button disappears because
settings.globalHotkey is never written to — no state drift.
Matches Handy #917 / #1019 / #966 / #956, brief item #5.
Co-authored-by: jars <jakejars@users.noreply.github.com>
A rapid double-tap of the global hotkey, evdev autorepeat, or a
sticky-key compositor quirk (KDE's 'slow keys') can all deliver
the same press twice within ~100ms. Without a guard, the recording
toggles into and out of the same frame and the capture is lost.
Gates the evdev 'kon:hotkey-pressed' forwarder in +layout.svelte
behind a 120ms debounce (Date.now()-based; no timers, so no tail
latency for a legitimate single press). The debounce is intentionally
shorter than a deliberate double-press cadence but longer than any
autorepeat interval we've seen in the wild.
The audio-stream-warming half of brief item #4 (Handy #1143) lives
in Workstream A's Phase A.3 warm-up WAV; this covers the UX side.
Co-authored-by: jars <jakejars@users.noreply.github.com>
Extends commands/paste.rs so that when auto-paste fires we:
1. capture the user's existing clipboard text BEFORE overwriting
it with the transcript (via arboard::Clipboard::get_text — a
non-text clipboard returns None and the restore step is
skipped, which keeps images / files safe),
2. after the paste keystroke lands, sleep 300 ms,
3. restore the snapshot only if the clipboard still holds the
transcript we wrote — respects any Cmd+C the user did in the
300 ms window.
Moves the decision to a pure should_restore(current, transcript)
helper with four unit tests. The existing paste-matrix regression
tests are unchanged.
Matches Handy #921 (workstream-B brief item #3). Coexists with the
Wayland preview-hide dance already in this file, and doesn't run
when the paste keystroke fails (transcript stays on the clipboard
for the user's own paste).
Co-authored-by: jars <jakejars@users.noreply.github.com>
Adds src/lib/utils/settingsMigrations.ts exposing
loadSettingsWithMigration() / saveSettingsWithVersion() around a
{version, data} envelope in localStorage["kon_settings"]. The
migration chain is indexed by destination version so adding a v2
is one MIGRATIONS[2] = (prev) => next entry away from working.
Legacy bare-object settings blobs are treated as v0 and folded into
v1 identically to before — no user-facing reset — but an unreadable
blob now surfaces a single Settings reset toast instead of silently
dropping data.
Covers Handy #602 ('settings reset on update') and unblocks the
other B.2/B.3 SettingsState additions listed in
docs/whisper-ecosystem/workstream-B.md: every subsequent field
lands behind a MIGRATIONS step, so older Kon builds stay readable.
Co-authored-by: jars <jakejars@users.noreply.github.com>
Sequences the 13 B-scope items from docs/whisper-ecosystem/brief.md
into three phases (pre-emptive UX, feature pinches, LLM layer) with
stop-for-review boundaries between each.
Enumerates the Settings sections touched per item (net: +2 toggles,
+2 sub-cards, nothing invisible becomes visible), the new
SettingsState fields with defaults, the schema migration bump
(version 1 -> 2), and the explicit Workstream A dependencies +
stubbed fallbacks for each (#14 list_gpus, #30 streaming cleanup,
#31 llm-state-change event).
Co-authored-by: jars <jakejars@users.noreply.github.com>
Two follow-ups to the previous CI deps commit:
1. Linux: libclang-dev installs libclang.so under
/usr/lib/llvm-*/lib and /usr/lib/x86_64-linux-gnu, but
bindgen-0.72.1 does not probe the llvm-versioned directory by
default. Resolve the newest /usr/lib/llvm-*/lib candidate
dynamically and export LIBCLANG_PATH so bindgen finds it without
any clang-sys guesswork. Also install glslang-tools + spirv-tools
so find_package(Vulkan COMPONENTS glslc) succeeds against
ggml-vulkan's cmake step.
2. macOS: llama-cpp-sys-2 sets GGML_VULKAN=ON unconditionally when
the vulkan feature is enabled; cmake's find_package(Vulkan) then
fails on a vanilla macOS-latest runner because there is no
Vulkan SDK shipped by default. The LunarG macOS SDK ships as a
non-scriptable interactive installer, so we compose MoltenVK +
Vulkan loader + headers + shaderc via Homebrew formulae instead.
Export VULKAN_SDK=$(brew --prefix) and CMAKE_PREFIX_PATH=same so
both the build.rs branch and cmake's FindVulkan resolve headers /
libs / glslc from /opt/homebrew.
Applied symmetrically to check.yml and build.yml. Windows config is
unchanged.
Co-authored-by: jars <jakejars@users.noreply.github.com>
Two issues with the previous #12 approach, both caught by CI:
1. tauri-build rejects the '_comment' json field as unknown when
parsing tauri.windows.conf.json:
unknown field `_comment`, expected one of `$schema`,
`product-name`, `productName`, ...
The schema is strict, so the doc-comment has to live elsewhere.
2. tauri-build's bundle.resources list is resolved at build-script
(cargo check) time, not at 'tauri build' time. With the DLLs
intentionally gitignored for licensing reasons (see the dir's
README), every cargo check run on Windows would fail.
Fix: delete tauri.windows.conf.json entirely. The intent of #12 —
'runtime falls back to CPU when Vulkan is absent' — is already
live in src-tauri/src/commands/models.rs::detect_active_compute_device,
unchanged.
Rewrite resources/windows/README.md to document a cargo tauri build
--resource ... invocation for the release engineer. That's the only
invocation that needs the DLLs present; everyone else (including
CI's cargo check) doesn't go near them.
This matches how Kon already handles CI/release split elsewhere
(macOS code-sign certs, Windows code-sign certs, etc. all stay out
of tauri.conf.json for the same reason).
Co-authored-by: jars <jakejars@users.noreply.github.com>
Both whisper-rs-sys and llama-cpp-sys-2 run bindgen at build time,
which requires libclang.so/dylib/dll resolvable from the loader.
None of the three GitHub runners ship this out of the box today:
- ubuntu-22.04: bindgen-0.72.1 panics with 'couldn't find any
valid shared libraries matching libclang.so*'
- macos-latest: same panic against libclang.dylib
- windows-latest: choco already installed llvm here, but libclang
was on an unset LIBCLANG_PATH, so bindgen still couldn't find it
And llama-cpp-sys-2's vulkan feature (wired on by whisper-rs' vulkan
feature → whisper-rs-sys + its own shared ggml build) hard-panics on
Windows when VULKAN_SDK is unset, and needs libvulkan.so linkable on
Linux.
Changes, applied symmetrically to check.yml and build.yml:
- Linux: add libclang-dev, clang, libvulkan-dev to apt-get install.
- macOS: brew install llvm, set LIBCLANG_PATH to brew --prefix
llvm /lib so bindgen can load libclang.dylib.
- Windows: choco install vulkan-sdk, set VULKAN_SDK to the
newest-version directory under C:\VulkanSDK (resolved
dynamically so a minor-version bump doesn't hardcode-break
anything), set LIBCLANG_PATH to the llvm bin dir.
Unblocks the per-push cargo check job on main, phase4-systems-f7d0,
and phase4-ux-f7d0; also unblocks the release build. The Windows
Vulkan SDK install is the new long pole (~2 min on a cold runner)
but is needed unconditionally while the vulkan feature is on in
crates/transcription/Cargo.toml.
Co-authored-by: jars <jakejars@users.noreply.github.com>
src/lib/utils/textMeasure.ts imports @chenglou/pretext but the
package was only present in the local node_modules — npm reported it
as 'extraneous'. npm ci in CI refuses to install anything that isn't
declared, so vite build fell over at the Rollup resolve step:
[vite]: Rollup failed to resolve import '@chenglou/pretext' from
'src/lib/utils/textMeasure.ts'.
Pins to the version that was already installed locally (0.0.5) and
regenerates package-lock.json. npm run build now completes cleanly
through the SvelteKit / Vite / adapter-static pipeline.
Co-authored-by: jars <jakejars@users.noreply.github.com>
Adds src-tauri/tauri.windows.conf.json with a bundle.resources list
for vulkan-1.dll, libssl-3-x64.dll, and libcrypto-3-x64.dll shipped
side-by-side with kon.exe, plus a src-tauri/resources/windows/README.md
explaining how release engineers populate the directory (licensing
constraints keep it manual rather than scripted).
The runtime fallback is already live from commit A.1 #1: if the
Vulkan loader is missing after launch, emit_runtime_warnings() fires
a runtime-warning event (kind: vulkan-loader-missing) and
get_runtime_capabilities() reports activeComputeDevice=cpu with a
reason. The app starts and transcribes on CPU — degraded but never
broken — so the acceptance criterion 'launches cleanly on a VM with
no GPU' holds by construction.
Matches Whispering #840/#829 and Buzz #1459 pain patterns.
Co-authored-by: jars <jakejars@users.noreply.github.com>
Adds src-tauri/src/commands/power.rs exposing a PowerAssertion RAII
guard that macOS uses to pin NSProcessInfo.beginActivityWithOptions
around long-running work. Wired into:
- run_live_session (entire live-dictation lifetime)
- cleanup_transcript_text_cmd's spawn_blocking body (LLM run)
Non-macOS targets get a no-op guard so callers don't have to #cfg
the call sites. The actual Objective-C bridge to NSProcessInfo is
stubbed (begin_activity returns Err so the guard logs a warning
instead of silently pretending); the stub doesn't regress recording
or LLM behaviour on macOS — it just means App Nap is not yet
suppressed, which matches today's behaviour. Full objc2 integration
is a follow-up that can introduce objc2 cleanly in its own commit.
Matches Whispering #549/#559 pain-pattern; acceptance text ("10
minute background recording completes unattended") is satisfied
once the bridge is finished, and nothing regresses today.
Co-authored-by: jars <jakejars@users.noreply.github.com>
Ports the kon-llm model_manager resume pattern the rest of the way
into kon-transcription::model_manager:
- download() now validates an existing complete file against its
sha256 before skipping; a hash mismatch removes the file and
re-fetches, instead of serving a corrupt file to whisper.cpp.
- download_file() now distinguishes 206 Partial Content, 200 OK
(resume silently ignored by server), and other statuses, rather
than treating any non-206 as 'just use it as a fresh start'.
200-on-Range is handled by discarding the partial and starting
over cleanly.
- New tests: download_file_resumes_from_partial_and_verifies_sha
(TcpListener fixture, same shape as kon-llm's), and
download_file_fails_on_sha_mismatch_and_cleans_part_file.
- sha256_of_file helper + unit test for the existing-file guard.
Dev-deps: tempfile + tokio(net/io-util/macros). Total workspace
lib-test count: 116 → 123.
Co-authored-by: jars <jakejars@users.noreply.github.com>
Extends get_runtime_capabilities() with three new fields:
- activeComputeDevice: {kind, label, reason} — "GPU (Vulkan)" on
the happy path, "CPU (fallback)" with a reason when the Vulkan
loader is absent at runtime. libloading::Library::new probes
libvulkan.so.1 / vulkan-1.dll / libMoltenVK.dylib per target OS.
- cpuFeatures: { avx2, avx512f, fma, sse4_2, neon, hasGgmlBaseline }
sourced from the new probe_cpu_features() helper. hasGgmlBaseline
is the one flag the Settings banner actually reads.
- parallelModeAvailable: placeholder false until Phase A.4 lands
the real GPU VRAM probe + GpuGuard semaphore permit logic.
Adds emit_runtime_warnings(&AppHandle) called once at setup() after
prewarm_default_model. Emits a runtime-warning event with kind
"avx2-missing" or "vulkan-loader-missing" so Workstream B can
render a dismissible Settings banner without polling capabilities.
Contract matches docs/whisper-ecosystem/workstream-A.md §Item #1 for
Workstream B to consume without waiting for Phase A.2 to land the
real whisper_print_system_info bridge.
Co-authored-by: jars <jakejars@users.noreply.github.com>
Extends hardware::CpuInfo with a CpuFeatures struct populated via
std::is_x86_feature_detected! on x86_64 and an architectural
assumption for aarch64 (NEON). Adds has_ggml_baseline() so callers
can cheaply ask 'will whisper.cpp / llama.cpp ship a fast path on
this CPU?' without knowing the arch-specific rule.
The point of #7 is giving the runtime a way to surface a clear
"non-AVX2 fallback" warning before the user hits a wall of silent
slowness. The banner itself ships in the next commit as part of
get_runtime_capabilities (item #1). Tests cover the baseline helper
on both x86 and non-x86 targets.
Co-authored-by: jars <jakejars@users.noreply.github.com>
Adds a build.rs guard that parses Cargo.lock and panics on Windows
if the tokenizers crate ever appears in the workspace dependency
tree, mirroring the MSVC C-runtime conflict that broke Whispering
v7.11.0 when they linked whisper-rs-sys + tokenizers in the same
binary.
On non-Windows hosts the guard downgrades to a cargo:warning so
cross-compilation or CI from Linux surfaces the issue before a
Windows build attempt actually panics.
No tokenizers crate is in the tree today; the guard is preemptive.
If we ever legitimately need HF tokenizers on Windows, the escape
hatch is an out-of-process sidecar (separate CRT).
Co-authored-by: jars <jakejars@users.noreply.github.com>
Adds http://127.0.0.1:* and ws://127.0.0.1:* to the connect-src
allowlist so future BYO-cloud LLM integrations (Ollama on 11434,
llama.cpp server on 8080, LM Studio on 1234, etc.) can fetch(...)
without tripping the CSP.
Kon's bundled LLM (llama-cpp-2) is in-process and does not need HTTP,
but the localhost surface is the standard external LLM transport and
pre-approving it here lets Workstream B wire Ollama's Test Connection
(item #27) without re-spinning the CSP. No tauri-plugin-http is in
use today — when that lands the scope allowlist goes in capabilities.
Matches the pain pattern from Vibe #438 (Tauri scope blocking
127.0.0.1 LLM endpoints).
Co-authored-by: jars <jakejars@users.noreply.github.com>
Sequences the 18 A-scope items from docs/whisper-ecosystem/brief.md
into four phases (pre-emptive patches, engine abstraction, streaming
correctness, LLM guard) with stop-for-review boundaries between each.
Lists the new command + event contracts (#1 activeComputeDevice,
#14 list_gpus/set_preferred_gpu, #24 tentative segment flag, #28
parallel-mode toggle) that Workstream B will consume.
Co-authored-by: jars <jakejars@users.noreply.github.com>
Both workflows had `workspaces: src-tauri -> target`, which tells
rust-cache the workspace lives at src-tauri/ and its target dir is
src-tauri/target. Neither is true: the workspace is defined at the repo
root (Cargo.toml:1–3 — members = ["src-tauri", "crates/*"]), so cargo
walks up and puts artifacts at ./target, not ./src-tauri/target.
Result: the cache action was saving an empty (or wrong) directory on
every run. Every CI run on every OS effectively started from a cold
build, which is the actual reason the Windows job appeared to compile
sqlx from scratch every push — it was compiling sqlx from scratch every
push.
Point the cache at the real workspace root. While here, drop the
`working-directory: src-tauri` on the cargo check step so the command
runs from the workspace root too; cargo finds the same workspace either
way, but running from root is consistent with the cache's view.
Expected impact: Windows check job drops from ~15–25 min cold-every-time
to ~2–3 min on warm runs, matching Linux/macOS behaviour.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
sqlx 0.8's default feature set pulls in `any`, `macros`, `migrate`, and
`json`. Grepping the workspace confirms none of these are used — the
code calls sqlx::query() / query_scalar() at runtime, implements its own
migration sequencing in crates/storage/src/migrations.rs, is sqlite-only
(no `any` needed), and never derives FromRow / applies sqlx proc-macros.
Dropping them keeps only what's needed: runtime-tokio + sqlite.
Why it matters disproportionately on Windows: the `macros` feature pulls
sqlx-macros → sqlx-macros-core → proc-macro2 / syn / quote / async-trait
/ url / heck / dotenvy / sha2 / filetime. Each proc-macro crate on
Windows MSVC compiles to a .dll with a full linker invocation (slower
ABI than Linux/macOS proc-macro .so). Net: tens of seconds shaved off
every cold-cache CI run, compounding with the cache-path fix in the next
commit.
kon-mcp was already lean (default-features = false); matching that shape
across the workspace now.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Specs the subset of the five-phase GPU kernel tuning roadmap that ships
without requiring ggml-dedup or agentic-search prerequisites:
- Phase 1 — Advanced → GPU Tuning settings panel (GGML env var toggles,
applied at startup before threads spawn).
- Phase 2 — kon-bench local autotuning CLI. Subprocess-based grid search
over env vars, outputs a ranked gpu-profile.toml.
- Phase 3-lite — kon-configs community repo. Manual-PR workflow (no CI
replay), fingerprint-matched fetch from Kon Settings.
Total ~7–10 days of focused work; captures roughly 85% of the eventual
value of the full roadmap. Phases 4–5 (custom SPIR-V drops + agentic
autotune) stay pinned in memory.
Includes the UX spec for the "one-click auto-optimise" flow: community
config check first (~15 s end-to-end), local benchmark fallback
(~8 min backgrounded), opt-in share-back via browser PR. Non-GPU users
see a clean "tuning doesn't apply" card with no nag.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Cloud agents (Codex / Opus in Cursor) can't read the author's local
Claude memory directory. This doc folds the non-negotiable ideology,
architectural decisions, what's already shipped, what's intentionally
deferred, and the file-ownership fence between Workstream A and B into
a single in-repo reference.
Kickoff prompts now point agents at docs/whisper-ecosystem/kon-context.md
as the primary read BEFORE brief.md, so the agents understand what NOT
to re-implement (already-shipped features) and what NOT to touch (the
other workstream's territory, the ggml dedup interim, etc.).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Survey of 10 OSS Whisper projects (whisper.cpp, whisper-rs, Handy, Buzz,
Whispering/Epicenter, faster-whisper, WhisperLive, whisper_streaming,
Scriberr, Vibe, OpenWhispr). Pins the cross-repo pain pattern matrix,
feature inventory with priority tags, streaming-specific findings, LLM
formatting findings, and a 31-item atomic task backlog — all URL-sourced.
Lives in the repo as the canonical reference for the phase-4 implementation
pass. Both Cursor workstreams (Codex for systems + streaming, Opus for UX
+ LLM layer) read this at kickoff.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Settings → Vocabulary gets a "Bulk add from a list…" disclosure under the
single-term row. Expanding reveals a textarea; paste newline- or
comma-separated terms, hit Import, and the page loops addTerm for each
entry the active profile doesn't already have.
Dedupes case-insensitively against the existing term list so pasting the
same block twice is a no-op. Skipped + failed counts surface via toast;
persistent errors (any failing term) also land in vocabularyError so the
inline panel explains what went wrong.
Covers OpenWhispr issue #460 — one-at-a-time entry becomes friction past
roughly ten terms. No backend changes; addTerm is already in profilesStore.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Without this, every secondary window (preview overlay, task float,
transcript viewer) opened at whatever spot Tauri / the compositor picked,
which was especially noticeable on Wayland where placement hints are
advisory. Main window's position was also lost on restart.
Registering tauri_plugin_window_state in the builder gives automatic
per-window-label save + restore. State lives in app-data/window-state.json;
fresh installs still fall back to the builder defaults (no changes to
inner_size on any of the four windows). Covers OpenWhispr issue #605 and
the broader UX pain.
No frontend changes — the plugin is purely backend. Regenerated ACL
manifests / desktop + linux schemas pick up the plugin registration.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
