Catch pre-header fetch and stream rejections in LocalLLMService._doPullModel and map disposed or aborted pulls to ErrorCode.LLMProcessingCancelled (and unexpected failures to ErrorCode.LLMServerUnreachable). Convert disposed fetch aborts and JSON body aborts in LocalSTTService._fetchDownload and downloadModel into ErrorCode.STTModelDownloadCancelled while wrapping network or timeout errors as ErrorCode.STTModelDownloadFailed.
Cancel pending result-popup measure timers on hideResultPopup; track CaptionService aux-model load state, stop deduplication, and refine abort epochs across restarts; forward per-call AbortSignal through VoiceModeService and LlmGateway to LocalLLMService and abort pull streams plus availability polling on dispose; clear aborted response pointers and gate stale mic start errors in VoiceConversationService; guard AudioCaptureService stopping resume against exited SoX processes and cleanup failed starts; abort LocalSTTService downloadModel polling on dispose; and fail closed when desktopCapturer.getSources rejects in caption loopback setup.
The local model follows the polish instructions only most of the time: in
repeated runs it added a [00:00] or [화자 1] marker that was never in the
audio, and it turned "사실 거의 근간" into "사실 근간". Both change what the
meeting record says, so they are now enforced in code instead of hoped for
in the prompt: markers absent from the original chunk are stripped, and
meaning qualifiers the speaker said are restored from that chunk.
The instructions themselves were re-measured on the model: a fixed removal
list, an explicit qualifier example and a marker ban keep the filler cleanup
without touching ordinary words, and the same wording is used by the app,
the tests and the probe. Unknown-word repair still depends on the model.
Learned text becomes part of a prompt, so a dictionary word or phrase
containing a newline could add its own instruction line. Every value that
reaches the recognition hint or the polish block is now flattened to one
line and length-capped, and a test proves an injected line stays plain text.
Also cover the invariants that keep the correction safe (random inputs,
Hangul untouched, idempotent, no token growth, large transcript) and the
file-transcription wiring that attaches the learned hint to every chunk and
corrects each chunk and segment before it is stored.
Move the polish instructions into their own module so the prompt the app
sends, the unit tests and the local-model probe use the same text. Tests
now pin the two rules that matter: fix a misheard name or in-house term to
the learned spelling, and never swap an ordinary word on similarity alone.
Verified against the local model with the real learned corpus: 하람록에 was
rewritten to HaramLog에 and the dictionary pronunciation 디에로 to D3RO,
while a natural sentence with no learned term came back with no word
substitutions. Recorded the trap that this model returns empty content
unless thinking is disabled.
The first cut replaced any Hangul word within one jamo of a learned term.
Measured against the real local corpus (three stored meeting transcripts)
it changed 286 of 3,005 words: 하면 became 라면, 이제 became 이게, 매일 became
메일. Most of what the app learns from typing is ordinary grammar, so
distance alone cannot separate a misheard word from a correct one.
Deterministic correction now covers only dictionary pronunciations and the
Latin identifiers the user types, preserving their casing. Korean terms
still bias recognition through the prompt hint and guide the polish
prompt. Re-measured on the same transcripts: 0 of 3,005 words changed.
Hangul-layout typing artifacts (dkssudgktpdy) are dropped from the hint.
The tree carried uncommitted work from several areas (web push, mobile
screens, admin charts, test-hygiene cleanup, CI and map docs). It is landed
as one commit so the release pipeline can start from a clean tree; the
desktop transcript-correction behaviour shipped in the previous commit and
is unchanged here.
Meeting and caption transcription never used what the app had already
learned: speech recognition ran with a fixed Korean prompt, the refine
prompt had no glossary, refined lines stopped inside the caption service,
and the saved transcript was a raw join of segments. Names and jargon the
user types every day were therefore re-mangled on every recording.
The learned corpus (user dictionary plus personal phrases) now feeds one
glossary that biases recognition, fixes mistyped words deterministically
right after each line is finalized, and is handed to the local model when
it polishes a line. Refined lines also reach the meeting transcript and
screen through a new segment-updated event.
Corrections never touch words the user actually wrote, refuse ties between
equally likely candidates, and keep particles intact.
Acceptance waited for modifiers using only the uiohook-tracked set; a missed keyup
made it time out after 1.5 s and paste while Ctrl/Alt counted as held, so nothing
was inserted. Trust GetAsyncKeyState instead, and suppress new suggestions while an
accept is in flight.
Typed deltas were saved per snapshot (every 0.4-0.8 s), storing mid-word fragments.
Buffer them per field and save one diff on idle, sentence end, send or focus change;
drop short pieces, URLs, trailing jamo and truncated emoji.
The tip drew its done state on session-completed, before the async insert
ran and even when auto-insert was off or the text was empty. A failed insert
also left the screen-saver-level tip over the result popup's copy button.
session-completed now says whether an insert follows, VoiceModeService emits
text-inserted on success, and the tip hides first when the insert fails.
The feature is not wired to the UI yet but is reachable over IPC:
- planning called the premium proxy directly, sending speech to the cloud
even with the local backend selected and failing in local mode
- LLM-provided app names and key combos were spliced into cmd/PowerShell
strings; they are now validated, and combos map named keys (alt+tab no
longer types Alt+T, a, b)
- presets matched substrings and bare nouns ("commute" toggled mute,
"크롬 캐시 지워줘" opened Chrome); they now match whole words only
- a plan without safe=true is treated as unsafe and unsafe shell commands
are not run; URLs go through openExternalSafe
The chunk store reads whatever database is open at call time, so a query that
was awaiting the model returned an answer built from the previous account's
documents, and a large file being read was inserted into the new account's
knowledge base and synced there. Each operation now remembers the account
scope it started in and stops if it changed.
The lock desktop swallows keyup, so a hold-to-talk recording never ended,
the first shortcut after unlocking was dropped as auto-repeat, and Win from
Win+L made the next suggestion accept wait out the full 1.5s modifier timeout.
On powerMonitor lock-screen/suspend the chord machine now emits released for
emitted triggers and clears its state, and modifier tracking forgets held keys.
stop() during model loading saw no aux model yet and skipped the unload, and
the failure path (e.g. no microphone) never unloaded it, so the caption model
kept VRAM alongside the dictation model until the next session ended.
- Each document gets an indexing run number; a reindex, synced replacement or
delete makes the older run stop instead of embedding stale chunks, reporting
'indexed' early and leaving a progress bar for a deleted document
- Reindexing a file that is now empty throws like addDocument instead of
replacing the stored chunks with nothing
Each 30-second chunk received the previous chunk's text as initial_prompt,
so once one chunk looped on a word the loop became the next chunk's hint
and the rest of the file was the same word. A looping chunk is now
re-transcribed without the hint, looped text is never carried forward, and
runs of 4+ identical phrases or segments are collapsed to one.
Capture still used the old snapshot code: a file copied in Explorer looked
empty and was cleared, and Excel's own formats were dropped on restore. It
now uses the shared clipboard snapshot policy and leaves the clipboard
alone when it couldn't put everything back.
Chains were pinned to the premium service, so they always failed in
anonymous local mode and sent text to the cloud for signed-in users who had
chosen the local backend.
Cancel only aborted an in-flight LLM response, which does not exist yet
during STT, so the cancelled speech was answered a few seconds later. A
turn counter bumped by cancel/stop now discards the late transcript.
SQLite lower() only folds ASCII, so importing 'äpfel' next to an existing
'Äpfel' created a second local entry that later collided with the server's
unique index and silently overwrote one spelling. Import now compares with
the same JS identity key as add/update, including duplicates within the file.
While suggestions are on, typos are no longer replaced silently. The popup shows
the fix next to the candidate and applies it only when the candidate is accepted.
The per-minute request cap rises from 6 to 20 so suggestions keep appearing after
accepting one and continuing to type.
The PC-wide slowdown was not a leak but model residency: boot warmup loaded
gemma4:e4b (~11.6 GB RAM incl. 5.9 GB pinned host buffer, no mmap on Windows
CUDA) and kept it for 30m, next to the large-v3 sidecar. Measured on the HEAD
build: free RAM 7.9 GB -> 32 MB within 20 s; after this change 10.8 GB with no
Ollama model loaded until typing starts.
- single LOCAL_LLM_KEEP_ALIVE (10m) for every local request; no boot warmup
- unload this app's resident Ollama models (keep_alive 0) on quit
- sidecar POST /unload for the aux slot; captions release it on stop
- partial dictation copies only the tail window, not the whole recording
- cap the voice-conversation listen buffer at 5 minutes
- mic test: generation guard so re-entry cannot orphan lease/timers
- startRealtime: single-flight and removeChannel to stop duplicate bindings
- meeting WAV save writes header and PCM without a concat copy
Docs: AI-01 policy note, GAP-PERF-01 with remaining work.
Typo auto-correct (INPUT-19)
- TypoCorrectionService + core typo-correction: 2-set Korean layout
mistypes both ways, always-wrong spellings, near-misses of the user's
frequent words, optional model check re-validated in core; applied by
backspace N + paste and cancelled on new keys or window change.
Prediction quality (INPUT-20/21)
- gemma models are sent raw with the user's unfinished sentence pre-filled
into the model turn, so the model truly continues the text (mid-word
joins and spacing exact). Other models keep the instruction prompt.
- Context: text before the caret (opening lines and a fact list of names,
dates, amounts outside the 400-char window), messages just sent in the
same chat window (memory only, 30 min), personal-graph follow-ups and
IDF-ranked related sentences with Korean stemming.
- Guards: first sentence only, restated-prefix stripping (also fixes the
instruction path), repeat drop, no invented dates/names/numbers
(one retry), plain -다 register, no emoji unless the user uses them,
search-box keywords. Fill loop mixes instruction prompts for variety.
- Opt-in real-model probe: tests/integration/suggestion-quality.ollama.test.ts
Suggestions now actually show up
- Terminal (Claude Code): caret trust from the input line, not the whole
screen (status line below the box blocked every keystroke); Korean
composition counts as typing at the end; a grey placeholder filling the
empty box at once is not user text.
- Chrome: UIA focus flicker while typing no longer resets "edited".
- KakaoTalk/RichEdit: trailing paragraph mark ignored for edit position
and stripped so accepted text keeps its separating space.
- While the overlay is visible the suggestion shortcuts are captured so
Ctrl+Alt+Enter no longer also types a newline; released when hidden.
- nut-js preloaded after boot (first accept waited 4.8 s).
CI run 121 failed at "버전·계약 정본 대조": earlier refactors moved update
gating from UpdateService into update-policy.ts (evaluateUpdateOffer ties
decideUpdate + isWithinRollout), the differential-download switch into
update-adapters.ts, and the Forgejo re-release guard into
lib/immutable-package-guard.mjs (sha256 comparison, abort on conflict).
The checks still looked for the old markers in the old files. They now
verify the same guarantees where the code lives, and the self-test's
negative case targets assertNoImmutableConflicts.
Also clears the lint gate: a control-character regex in Keycap (range now
starts at U+0020) and unused helpers in two red-team tests.
Batch of red-team r3 fixes that were in the working tree before the
2026-09-28 design overhaul, committed as one unit with their tests.
- desktop main: STT timeouts and sidecar, voice recording store, sync
(credentials, audio, knowledge reindex, push gates), runtime
provisioner, update policy, AltGr keybindings, voice-command policy,
dictionary file codec/limits, meeting transcript condensing and a
local recording ledger so interrupted-session recovery only closes
meetings this device recorded (a phone's live meeting is left alone).
- mobile: login CSRF via implicit token callbacks rejected, account
deletion/retention, durable queue retention, knowledge realtime
without unfiltered DELETE, meeting re-record failure paths, cloud STT
client, preferences store/resync.
- core: text chunking splits long unbroken transcripts to fit, template
field policy, dictionary limits, meeting markdown inline handling.
- server: payple webhook policy and cancellation order scope, meeting
document generation quota, team RPC null-role guard, unified LLM
quota in-flight accounting, knowledge chunk vector index, meeting
re-record failure paths (migrations 20260929*).
- ci: portable/runtime feed gates, update-policy schema, Forgejo file
delete and alias planning.
Four older tests are updated to the new contracts rather than the old
behavior: token-pair auth callbacks are rejected, knowledge realtime no
longer subscribes to DELETE, long transcript lines are split, and
meeting recovery requires the local recording ledger for empty rows.