Coding agents read whatever they can find. That includes the ADR you superseded, the plan you abandoned, and the second architecture document somebody wrote because they could not find the first. Canonly ranks by what the project actually trusts, not by what matches the query.
Connect over MCP: hybrid BM25 and vector retrieval fused with reciprocal rank fusion, then re-ranked by a trust score built from canonical status, source authority, human approval, freshness and unresolved conflicts. Relevance and trust are reported separately.
It never writes to your repository. No commits, no pull requests, no branches. An agent cannot approve its own work and cannot replace a document a person approved: it proposes, you decide.
