Skip to content

todo.md

Latest commit

 

History

History
64 lines (44 loc) · 6.35 KB

todo.md

File metadata and controls

64 lines (44 loc) · 6.35 KB
Raw

Parley — session todos

Source: codebase + UI audit conducted this session. Each item carries file:line pointers and a severity.

Done this session

  • src/web/api.js:251 — gate PATCH /todos/:id behind requireAdmin. Action-item toggle now admin-only; non-admin receives 403.
  • src/pipeline/orchestrator.js:88 — seedTodos createdAt = meeting.started_at (not now()). Long transcription queues no longer drag action items onto "today" in the timeline.
  • src/voice/meeting-manager.js:73-82 — fail-loud .catch on finalize escape: logs stack, flips processing → transcription_failed (only if still processing), re-throws so done awaiters still see rejection. Resolves "stuck Processing card between restarts".
  • src/pipeline/retry.js:73 — seedTodos createdAt = meeting.started_at (consistency with orchestrator path).
  • src/web/api.js:231 — merge seedTodos createdAt = target.started_at (same alignment).
  • web/src/pages/ActionItems.jsx:69-83 + :124-127 — wrap toggle in try/catch; surface toggleErr banner. Non-admin 403 path now shows "Only admins can edit action items." instead of uncaught rejection.
  • web/src/pages/Search.jsx:1 + :8-23 — Fragment import added; highlight loops over all matches, not just first.
  • src/store/db.js + src/pipeline/orchestrator.js — atomically replace meeting utterances and persist transcript completeness; failed replacement rolls back without leaving a partial transcript.
  • src/pipeline/retry.js — retranscribe retained PCM for incomplete or legacy-unknown transcript state; known-complete transcripts still re-summarize without STT.
  • docs/audit-2026-07-06.md — reconcile audit #7 with the actual transactional persistence and PCM-aware retry fix.
  • Repository DB hygiene verified — *.db, *.db-shm, and *.db-wal are already ignored, and git log --all -- meetings.db is clean.

Verified: npm test 270/270 pass. npm --prefix web run build 64 modules, 282 kB bundle, clean. Sidecar: 10/10 pass (one known Starlette deprecation warning).

Open — backend / pipeline

High

  • src/pipeline/orchestrator.js:91-102 — persist delivered/deliveryError to the meetings table (audit #14, half-closed). Returned but never stored; /meetings/:id can't surface "notes saved but not posted" or offer a re-post. Fix: schema migration adds delivered/delivery_error columns; orchestrator writes them; expose via /meetings/:id; add POST /meetings/:id/repost reusing postNotes. ~3 files: db.js, orchestrator.js, api.js; 1 UI line in Reading.jsx.
  • src/pipeline/retry.js:75 — log delivery failures on the dashboard retry's resummarize path. Today deliver(...).catch(() => {}) swallows with no log, while the orchestrator logs the same class (orchestrator.js:99). Pick one behavior.
  • src/store/db.js — retention + VACUUM + scheduled sweep (audit #17). meetings/utterances/FTS grow unbounded for done meetings. reconcileOnBoot runs once at process start only. Add pruneOlderThan(days) callable from a setInterval in index.js + a POST /system/prune admin route.

Medium

  • src/web/auth.js:141 — requirePasswordChanged fails open when req.user is undefined. Mirrors pre-fix shape of requireAdmin (audit #22). Safe today because server.js:26 always mounts attachUser first, but a future bare-mount on the standalone server reintroduces the default-cred takeover surface. Fix: same req.authResolved gating as requireAdmin.
  • src/store/db.js:53 — todos_dedup UNIQUE INDEX lets duplicates in when meeting_id is NULL (SQLite NULLs are distinct in UNIQUE indexes). Every seedTodos call today sets meeting_id, so it never fires; add an explicit comment for future maintainers noting the invariant.
  • src/pipeline/transcribe.js:30 — round-robin worker assignment stalls on a hung convert/stt. Minor at default concurrency (2/4); skip unless you scale up.

Open — scripts / project hygiene

  • Drop one of the duplicated reprocessing script pairs. scripts/reprocess-meeting.mjs ≡ scripts/resummarize.js; scripts/retranscribe-meeting.mjs ≡ scripts/retranscribe.js. bot.js:77 failure message points at reprocess-meeting.mjs; the dashboard Retry button (api.js:167) covers the same path. Pick one set, delete the other, update bot.js message to point at the dashboard primarily.

Open — UI (audit, intentionally not fixed this round)

Low / polish

  • web/src/pages/Reading.jsx:205 — delete-then-navigate to /meetings leaves a stale list. Meetings.jsx useEffect only fires on guildId change. Fix requires a global meetings cache or a manual refresh signal — bigger refactor.
  • web/src/GuildContext.jsx:11-15 — fetches guilds once at mount. New guilds added in Discord after page load never appear until refresh (bot-controller emits no SSE). Audit #36 category.
  • web/src/components/Markdown.jsx — single-pass line parser, no nested formatting (e.g. bold link inside a list item). LLM answers are simple; dependency-free intent preserved. react-markdown migration optional.
  • web/src/components/Layout.jsx GuildSwitcher — no loading vs empty distinction. Shows "No servers" both while loading and when truly empty. Minor cosmetic.
  • web/src/pages/ForcePasswordChange.jsx:20 — client-side min length hardcoded at 8. Matches server PASSWORD_MIN today; drift risk if server bumps the threshold. Tolerable for self-hosted.
  • web/src/pages/Analytics.jsx — top-N hardcoded at 8. Could expose a "show all" toggle. Polish only.
  • web/src/components/ui.jsx:117 StatusPill — only knows 5 statuses. Unknown states (e.g. 'empty') fall through to a generic "Unknown" pill. Today 'empty' meetings get deleted by bot.js:65 before they reach the UI, so this is theoretical.

Verification checklist (re-run before any PR)

  • npm test → 270/270 pass.
  • npm --prefix web run build → clean, 64 modules.
  • stt_sidecar/.venv/bin/python -m pytest stt_sidecar/test_server.py -q → 10/10 pass; one known Starlette deprecation warning.
  • Manual smoke: start bot + sidecar, record a short meeting, confirm notes deliver; toggle an action item as admin (works) and as non-admin (403 + UI banner); retry a failed meeting from the dashboard.