wtf is agent-ledger?

Agent Ledger is a tool that makes AI coding agents like Codex or Claude Code review their own work and leave a clear paper trail before a human looks at it. Before making any changes, the agent has to state what it plans to do and what files it is allowed to touch. After it makes the change, the tool runs tests and checks, has the agent look over its own findings, lets it fix the safe and obvious problems, and then writes up a short summary for the person reviewing the work. It installs as a skill inside Codex or Claude Code, along with a small command line tool and an optional GitHub Action, and runs entirely on your own machine rather than sending anything to an outside service. Installing takes one command using the skills installer, which detects which coding agent you are using automatically. Once installed, you invoke it by typing a short command like slash agent-ledger followed by the task you want done, and it figures out a sensible scope and which checks apply on its own, only asking questions when something is genuinely unclear. Every run produces a set of files: a machine readable review record, a list of issues that were found and fixed, a short summary meant for a human, a pass, warn, or block style status for reviewing a pull request, and a visual timeline you can open and replay in a browser. There is also a smaller, redacted version of this evidence meant for sharing, though the tool notes that its redaction is best effort and should be checked before publishing. Agent Ledger deliberately limits what it will fix automatically. It will not touch things like authentication, secrets, permissions, billing, deployments, or anything destructive, leaving those for a human to decide on. The project is explicit that it does not make an AI model smarter and does not replace testing, code review, or human judgment. It only tries to make the agent's own execution more disciplined and its results easier to check. There is also a GitHub Action version that captures the changes made by a command you give it and posts the review directly into a pull request's job summary.