Watch — karajan-watch

karajan-watch is the watchman of the Karajan family: a shared RAG over ALL your organization’s repos, fed on every merge — and, on top of it, two answers no single-repo view can give you: which other repos may this change break? and which documentation did it just make stale?

Three layers, strictly separated

Layer	What it is	Where
Engine	karajan-rag — index/query/serve, context strategies, sensitivity policy + PII redaction	Open source, npm
Product	karajan-watch — generic orchestration: reusable ingestion, impact and drift workflows. Knows no organization.	Open source, npm
Deployment	One private repo per organization: repo list, thresholds, infra, credentials, calibration golden set	Yours

The dependency always points upward — the product never learns who deploys it. Like nginx and your nginx.conf: the product belongs to everyone, the instance is yours.

Honest by design

Impact analysis produces a risk ranking with evidence — retrieval similarity, git co-change history, literal contract matches (HTTP routes, event topics, SQL tables) and a policy-gated LLM verdict. It is never sold as a calibrated “probability”: heuristics are labeled as heuristics, contracts are labeled as evidence, and you calibrate the thresholds against a golden set of your own real incidents.

Start here

Design and phases The problem, the three layers and the F1/F2/F3 roadmap.

Configuration The strict contract between product and deployment repo.

Ingestion on merge The reusable workflow that keeps the shared corpus fresh.

Cross-repo impact Four signals, one ranking with evidence on every merge.

Documentation drift Which sections went stale — with literal proof.

Eval and calibration precision/recall@k against your own incident history.

Contract with the engine What watch consumes from karajan-rag, and the upstream gaps.