@onenomad/cortex-pipeline-meeting

Three-pass meeting transcript extraction pipeline.

Per [ADR-004] model routing:

1. Pass 1 — structural (task: structural) Local model extracts JSON structure: participants, topics, decisions, action items, direct quotes.
2. Pass 2 — synthesis (task: synthesis) Quality-critical model merges structure with retrieved project context; rewrites action items with owners + implicit due dates.
3. Pass 3 — brief (task: brief) Local model generates a human-facing markdown brief.

Output: one memory per decision, one per action item, one for the full brief, plus transcript chunks. All share the same source_id (via #suffix) so Engram dedups re-runs.

Prompts live in src/prompts/ as markdown — see ADR-007. Edit them, re-run the fixture harness, iterate without touching code.

Fixtures

tests/fixtures/*.txt — sanitized sample transcripts for golden testing. The harness runs all three passes against a stub LLM that records calls + replays fixture responses, proving the prompt plumbing and shape of the multi-memory output.