Quickstart: attribution

1. Install and add one repo. Run npx repoops, then point the wizard at a local git repo. Full walkthrough: Install.
2. Run one session the way you already do. Nothing changes about how you work. RepoOps reads what Claude Code writes to disk; any other agent can report itself over the ingest API.
3. Open the accountability ledger. Each session shows its length, the files it touched, and what it cost. From a defect, the causal trail walks back to the session and prompt that introduced it.

An answer arrives with a confidence band rather than a bare claim. Exact and strong count as proof. Weak reads as "possibly" and never counts toward anything. The git identity on a machine is never treated as evidence about a finding, because a commit author says who committed, not who wrote.

Cost is summed per session on the same key, so a finding arrives with the spend that produced it. A model the rate card cannot price reads as unpriced rather than as zero.

Once you have one attributed session, three things become available with almost no extra setup: check the bill against what your provider charged, score your agent setup, and have a fix built for something the ledger found.

Attribution breadth covers which signals carry which evidence. Causal trail covers the walk from a production error back to a prompt.