AI Note Taker slices 1–2: notes generation, in-browser recorder, and action-item → task loop-closure

[vjvarada]

Builds the AI Note Taker from slice 0's upload→transcript foundation into a working record → notes → tasks / email loop. Rebased onto current main (the earlier migration 94→95 collision was fixed independently by #185).

Slice 1a — notes generation

• templates.py — meeting-notes template as data → prompt compiler with grounding + anti-injection rules + markdown renderer.
• summaries.py — notes on acb_llm (strict JSON), single-pass or map-reduce for long transcripts (per-chunk cap logged). Writes summary_md/json + meeting_note + draft action_item rows with segment_ids grounding. POST /summarize, GET/PUT /note, GET /actions, GET /templates.
• pipeline.py auto-chains notes after transcription; events.py + Next proxy → per-meeting SSE progress.

Slice 1b — the in-browser recorder

• recordings.py — chunked start→chunk?seq= (idempotent, gap-checked append)→complete.
• recorder.ts — MeetingRecorder: getUserMedia, feature-detected mime, a retry-with-backoff uploader (in-memory offline buffer) that flushes before complete; level meter.
• session/[id] — the "studio" screen: consent line, canvas waveform, mono timer, pause/stop. Library "Record" button live.

Slice 2 — loop-closure into Tasks + Email

• actions.py — approve inserts a LOCAL gtd_items task with an origin={kind:meeting,…} link + resulting_task_id (idempotent); reject; approve-all bulk ≥ threshold; every approval audited. Meeting detail action rail is interactive triage with deep-links into /tasks.
• Attendees — meeting.attendees JSONB (migration 96_note_taker_attendees.sql), inline chip editor, PUT /attendees.
• share.py — POST /share/email/draft LLM-drafts a recap from the notes → a HITL compose modal (editable to/subject/body + account picker) that sends via the existing /email/send. Draft-only server-side.

Verification

• 9 notes unit tests; full suite 1647 passed. All 17 /notes routes register; frontend tsc clean. Migration chain 00→96 + idempotency validated on a clean PG16+pgvector instance; gtd_items columns verified against migrations 48/65.

Not yet

Share-to-chat (last slice-2 item), then slice 3 — self-hosted STT + open speaker diarization.

🤖 Generated with Claude Code

https://claude.ai/code/session_016xdbRumGveUBFRsPh3Duu8

[vjvarada]

Unit-tests failure is a base-branch migration collision, not this PR

The failing check is test_migration_prefixes::test_migration_numeric_prefixes_are_unique:

duplicate migration number(s): {'93': ['93_rename_reply_to_needs_reply.sql', '93_gtd_planning_prefs.sql']}

This is inherited from main, not introduced here:

• This PR adds zero migrations — its own note-taker migrations are 95–98 and don't collide.
• main HEAD (f0c7919) already carries two 93_* and two 94_* files:
  • 93_gtd_planning_prefs.sql (calendar) + 93_rename_reply_to_needs_reply.sql (fix(infra): renumber rename migration to 93 (number collision) #170)
  • 94_email_voice_profile.sql (feat(email): voice & writing-style profile learned from past mail (mig 94) #176) + 94_gtd_day_templates.sql (calendar)
• So main is independently red on this check; my PR only trips it because CI tests the merge of this branch onto main.

I'm not renumbering those four already-merged migrations from other PRs here — that's out of scope for this branch and risks DB drift against what's already applied. This needs a dedicated fix on main (renumber the later of each colliding pair). Once main is green on this check I'll rebase and re-run; this PR's own suite is otherwise green (1700 passed locally, incl. the new Deepgram STT tests).

---

Generated by Claude Code