Context
Every self-hosted repo's contribution quality -- both for human contributors and for our own review AI at src/services/ai-review.ts -- is bounded by how accurate and current its CLAUDE.md/AGENT.md and any repo-specific skill files are. Hand-written versions of these files go stale within weeks of active development: module boundaries shift, test commands change, contribution workflows get rewritten, and nobody circles back to update prose docs. The self-host stack already runs RAG/embeddings (src/review/rag.ts) and deterministic signal extraction (src/signals) against every registered repo at review time, and that same infrastructure sits idle the rest of the time. It is a natural, already-paid-for substrate for keeping these files grounded in the actual codebase instead of a snapshot from whenever someone last got around to updating them.
This roadmap adds a capability to derive a repo profile from that infrastructure, generate CLAUDE.md/AGENT.md and, where warranted, scoped skill files from it, and keep them current on a schedule -- always through a reviewable pull request, never a direct write.
Design principles
- Never silent-write to a target repo. Every generation run produces a diff delivered as an opened pull request (or, at minimum, a diff a maintainer explicitly reviews and approves) -- this mirrors how the rest of the review engine never takes unreviewed write actions against a repo it does not own outright.
- Default OFF, opt-in per repo. No repo gets this behavior until a maintainer explicitly enables it in its config.
- Must not leak private repo config, gate settings, signals, or scoring into the generated file -- the output is public-facing documentation, not an internal-state dump.
- Must degrade gracefully when RAG is not configured or not yet indexed for a repo -- skip generation and log the reason rather than emit a low-quality or empty file.
Child issues
Context
Every self-hosted repo's contribution quality -- both for human contributors and for our own review AI at src/services/ai-review.ts -- is bounded by how accurate and current its CLAUDE.md/AGENT.md and any repo-specific skill files are. Hand-written versions of these files go stale within weeks of active development: module boundaries shift, test commands change, contribution workflows get rewritten, and nobody circles back to update prose docs. The self-host stack already runs RAG/embeddings (src/review/rag.ts) and deterministic signal extraction (src/signals) against every registered repo at review time, and that same infrastructure sits idle the rest of the time. It is a natural, already-paid-for substrate for keeping these files grounded in the actual codebase instead of a snapshot from whenever someone last got around to updating them.
This roadmap adds a capability to derive a repo profile from that infrastructure, generate CLAUDE.md/AGENT.md and, where warranted, scoped skill files from it, and keep them current on a schedule -- always through a reviewable pull request, never a direct write.
Design principles
Child issues