@teambrain/distill

The TeamBrain distiller: session records → proposed memories, as a pull request.

The governance gate. Its output is always a proposal to humans — it has no merge rights and never writes to the brain.

Pipeline (run by tb distill, typically on a weekly CI schedule):

1. Collect — new records on the teambrain/sessions branch since the last watermark, plus merged-PR metadata via gh.
2. Cluster — repeated struggle signals: the same paths fought across ≥2 sessions, repeated failing commands, memory_search queries with no hits, agent-proposed candidates.
3. Draft — one structured-output LLM call per cluster against a versioned prompt; invalid drafts are discarded, never silently "fixed".
4. Dedup & conflict — embedding similarity ≥0.85 drops duplicates; pairwise contradiction checks set supersedes and flag the PR.
5. Gate & PR — score by evidence × novelty, cap at 10 candidates, open a teambrain/proposals-<date> PR with a reviewer-friendly summary table.

Provider-agnostic: LLM access goes through a small Provider interface (Anthropic driver + a fixture-based FakeProvider for tests); the model is pinned in brain.yaml. This package is the only place in TeamBrain that calls an LLM.

npm install @teambrain/distill

Part of TeamBrain — most users want @teambrain/cli and its ci-templates/.

Apache-2.0