fix: robustness cleanup — error handling, resource leaks, config hygiene by TuurHart · Pull Request #3 · TuurHart/Skrift

TuurHart · 2026-04-02T18:54:30Z

Summary

Backend: Replace bare except: and print() calls with proper logging across 5 files (system.py, files.py, transcription.py, sanitisation.py, enhancement.py). Errors that were silently swallowed now surface in logs.
Frontend: Fix EventSource leak on sidebar unmount, add cancellation support to transcription polling, clean up polling intervals on unmount, remove debug console.logs from drag-drop handler, fix non-null assertions.
Electron: Add backendProc.on('error') handler so spawn failures surface immediately instead of a silent 60s timeout.
Gitignore: Stop tracking user-specific files (user_settings.json, names.json, batch_state.json) — they stay on disk but are no longer in the repo.

Test plan

Start backend — no import/startup errors
Start frontend (npm run dev:electron) — app loads normally
Upload a file, transcribe, enhance — full pipeline works
Check Electron dev console for leaked EventSource or interval warnings
npx tsc --noEmit passes (verified locally)

🤖 Generated with Claude Code

…king Backend: replace bare except clauses and print() with proper logger calls, log ffmpeg pass-1 failures, warn on names.json parse errors. Frontend: fix EventSource leak on sidebar unmount, add cancellation support to waitForTranscription polling, clean up transcription polling interval on unmount, remove debug console.logs from drag-drop, fix non-null assertions. Electron: surface backend spawn errors immediately instead of silent 60s timeout. Gitignore: stop tracking user_settings.json, names.json, and batch_state.json. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…ibe (conv bug-hunt #1–#3,#5) Fixes the conversation/diarization → name-linking → Obsidian export defects on the Mac (CONVERSATION_BUGHUNT_PROMPT.md). User decisions locked; see backlog.md. - #1 over-canonicalization: inline name mentions now link as Obsidian alias-display `[[Canonical|spoken]]` (every mention) — the SPOKEN word ("Tuur") is preserved, not replaced by the canonical/short. New `Sanitiser.processConversation`. - #2 speaker labels: a speaker's FIRST turn header → full `[[Canonical]]`, every later header → plain short name `**Tuur:**` (no link). Fixes the "every header full canonical" bug (voice-only people with empty aliases were skipped before the demote). - #3 fragmentation: `SpeakerTranscript.parse`/`mergeAdjacentTurns` ported to the desktop; processConversation merges consecutive same-speaker turns. - isAttributed line-anchored + requires ≥2 DISTINCT speakers — a hand-formatted `**Pros:**`/`**Cons:**` body is no longer mis-read as a conversation (was skipping copy-edit on plain notes). - BatchRunner: conversations route to processConversation; the Mac-diarize path now emits PLAIN headers so phone-synced and Mac-diarized memos render identically. - Pipe-aware link identity (the regression trap from introducing `|alias`): linkTarget / hasCanonicalLink / linkDisplay; BodyTextView.person(matchingCore:); resolver first-mention detection; unlink/relink restore the spoken display word. - #5: Re-transcribe + Redo-copy-edit hidden for diarized memos (would destroy the turns — the phone uploads no segments/word-timings, so the `**Name:**` text is the only copy); redo(.copyEdit) keeps a conversation verbatim defensively. Gate: 255 UnitTests green (+7 new conversation tests); full SkriftDesktop build (-skipMacroValidation) SUCCEEDED. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… turns (conv bug-hunt #3,#4,#5) - #5 data-loss: MemoSaver.diarizeIntoTurns marks the diarized transcript transcriptUserEdited=true. A phone conversation's `**Name:**` turns are the only copy (segments/word-timings aren't uploaded); without this, a noisy multi-speaker take with ASR confidence < 0.7 was UNtrusted → silently re-ASR'd on the Mac at ingest, destroying the turns. Same rationale as the quote-capture / append paths. - #3/#4 fusion hardening (SpeakerFusion, byte-identical on both apps): - minTurnWords default 2→3: a 1–2 word wrong burst flanked by the same speaker on both sides is now smoothed away (less fragmentation / mid-run mis-attribution). - gap-word assignment uses the NEAREST SEGMENT BOUNDARY, not the nearest midpoint — a word at a long segment's edge was being snapped to a shorter neighbour whose midpoint happened to be closer. (#4 boundary-word mis-snap.) Mid-sentence mis-splits remain bounded by Sortformer segment quality; the Mac same-speaker merge (processConversation) + manual reassign are the backstops. Gate: mobile 383 unit tests green (+2 fusion cases); desktop 255 UnitTests green. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… [[img]] marker Two of the three reported video-from-Photos bugs (the third, playback, landed separately): - Source glyph (#3): a video import is neither a share-capture (it HAS audio) nor a book-capture, so it carried no source marker and looked like a plain voice memo. Added MemoMetadata.sourceType (free-form String, additive/optional, value "video" via MemoMetadata.Source.video; set in processVideo incl. the no-audio-track failure path) → Memo.isVideoImport → a video.fill leading glyph + a "Video" chip in the list row AND the detail header chips. Mobile-only: NOT added to UploadMetadata, so the Mac contract is byte-unchanged (the glyph is a phone concern; the full cross-app taxonomy stays deferred). - Row snippet (#2 real wart): the untitled-row snippet showed the raw "[[img_001]]" marker — a video transcript always opens with the frame marker, which filled the whole snippet line. MemoCard.snippet now strips [[img_NNN]] markers (titled rows already used the marker-stripped firstTranscriptLine). Thumbnail aspect (#2 literal claim): NOT reproducible — verified via -seedVideoMemo (a landscape 16:9 frame with a centered circle stays a perfect circle, not an ellipse, in BOTH the row thumb and the detail embed). The square thumb is a conventional center-crop, not a distortion; every display path already aspect-fills + clips. Added VideoMemoUITests (seeds the frame, asserts the Video chip, screenshots row + detail to /tmp/skrift-video-shots) as the gate. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

@mainactor

…e + doc refresh Cleanup pass after the audit (adversarial review + hygiene + doc-consistency): - Reconcile non-blocking (correctness, the audit's one moderate finding): the launch / app-active / CloudKit-import triggers now dispatch the sweep as `Task { @mainactor }` instead of a synchronous call, so they never block App.init() or the notification handler — the window renders first. SwiftData stays on the main actor by design (the codebase funnels all SwiftData onto main to avoid cross-context corruption; a background context was the wrong fix). The sweep dedups, so only a one-time large first-enable backlog could jank main — off-main blob materialization noted as a deferred follow-up. - .gitignore: add build-release/ (a device/release derived-data output dir, same as the already-ignored build/ and build-device/) so it stops showing as untracked. - Docs refreshed to match reality: MAC_CLOUDKIT_PLAN.md RESUME (all pushed; device-verified — 73 memos synced, 32-byte APNs token; remaining = one Process→write-back round-trip + prod promotion) + hard-steps #1-#3 marked DONE for dev (capability, team signing, push). FEATURES: the CloudKit-push row flipped 🟡→✅ for BOTH apps (mobile since 53451a6, desktop since 6642af4, device-verified), and the reconcile row notes the async triggers + device-verification. Audit cleared everything else: container isolation, gating, contexts, write-back idempotency/ no-orphan, Memo.make round-trip, dedup, no force-unwraps/fatalErrors, phone push fully ready. Gate: desktop full -skipMacroValidation build SUCCEEDED (reconcile change is App-only). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix: robustness cleanup — error handling, resource leaks, config hygiene#3

fix: robustness cleanup — error handling, resource leaks, config hygiene#3
TuurHart wants to merge 1 commit into
mainfrom
robustness-cleanup

TuurHart commented Apr 2, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant