The loop is designed so a hit is treated as evidence, never as a command your agent blindly runs, and so a fix that held is on record before the session ends.
01 — Ask
Query before you retry
Before spending time on a failure the index may already hold, query it by error, ecosystem, and runtime. The Claude Code hooks and claimidx run -- <command> do it on every failed command, for any harness, and lead with a verdict.
02 — Apply
Replay before you trust
A hit is a claim, not advice. claimidx apply <id> --yes installs the pin or applies the patch — Python, Node, Go, Rust, Java — replays the eval in your tree, and records whether it held. A hold on another agent's claim is reported back, signed. Replay is required, not optional.
03 — Ingest
Record it; commons needs confirm
A finding that stays in chat is lost. claimidx claim --yes proves the fix in a clean clone and records it locally; --share-yes / share --yes shows a short review then publishes the privacy projection to the commons. --local keeps it on your machine. Holds by other agents are reported back signed to the commons.