RFC: knowledge intake — merged-PR and bugfix learnings return as reviewed knowledge PRs #111

Background

The knowledge tree has a written intake contract (knowledge/AGENTS.md: PR-learning produces only executable rules in the nearest owner's rules.md, validators + release audit gate every batch) — but no automated feeder. Learnings evaporate at the two moments they are freshest:

1. a vllm-omni PR merges (review discussion + final diff carry generalizable lessons), and
2. a copilot pr_debug run lands a verified fix (root cause + verification stop at run-scoped debug memory and never reach the human-curated tree).

The governing constraint is the knowledge plane's safety model: agents propose, humans promote. A pull request against this repo is the promotion gate, so the design delivers validator-clean proposals as PRs and nothing else.

Design

Three decoupled stages (full record: doc/RFC-knowledge-intake.md on the PR below):

• Recording (cheap, every watch cycle) — the reviewbot scans all merged PRs straight from GitHub's closed pulls (drafts and untracked PRs included, durable overlap-safe cursor) and ingests copilot bugfix drop files (pr.harvest_debug_knowledge writes one JSON per landed, non-dry-run fix). Rows wait in a knowledge_intake SQLite table so a distiller outage never loses an event.
• Distillation (at most once a day) — the model only proposes: catalog-constrained rule sections as structured output, sources fenced as untrusted data. Code applies append-only (new ## <rule-id> sections + updated: bump, never a rewrite); check_knowledge_tree.py + check_wiki_lint.py gate the batch fail-closed, with a bounded attempt cap parking poison rows visibly.
• Publication (fork-based, double-gated) — the date-scoped branch pushes to the bot's fork and the PR targets this repo; merging is the human promotion act. Shadow by default; posting requires POST_MODE=review and KNOWLEDGE_PR_ENABLED plus a configured fork. Same-day reruns build on the day's branch and resume failed publication.

Implementation

• zuiho-kai/omni-reviewbot#21 — intake ledger + merged-PR/drop scanner
• zuiho-kai/omni-reviewbot#22 — daily distiller + fork PR publication (stacked on Preserve Strict findings for inline GitHub reviews #21)
• pr-debug: drop a landed-fix record for external knowledge intake #110 — pr.harvest_debug_knowledge step + executor/resume e2e + doc/RFC-knowledge-intake.md

End-to-end tests pin the watch-cycle flow (reviewbot) and the crash-then-resume checkpoint path (copilot); the cross-repo drop-record contract is pinned by deliberately duplicated fixtures on each side.

Rollout

KNOWLEDGE_INTAKE_ENABLED (recording + shadow batches) → wire KNOWLEDGE_INTAKE_DIR = COPILOT_INTAKE_DIR (bugfix channel, shared host) → create the bot's fork + KNOWLEDGE_FORK_SLUG → KNOWLEDGE_PR_ENABLED under POST_MODE=review. Rollback at any stage is turning the flag off; nothing in the tree changes without a merged PR.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EwmmhpCK9pfEsWWzUBx3GB