Image: GitHub
Why it mattersA single agent that reviews its own code review has no outside check on it. Splitting the job keeps the pass or fail out of one model's hands.
A coding agent that reviews a coding agent's work is grading its own exam. kyy-logs released code-review-skill on 29 October. It is a free MIT skill. It splits the job across four separate probes, and each one writes down its own evidence.
The repository has 101 stars three days after it was created. The latest release is v0.1.0-beta.3, tagged on 31 October. The skill runs inside Claude Code with Agent Skills support, or inside Codex with custom-agent support. It installs with one line and the install command names the agent.
Four probes, each in a fresh context
The skill does not ask one agent to review the whole change. It runs four separate probes, each in a fresh context, with no memory of what the others saw. The names are behavior-review, test-coverage, repo-guard and impact-analysis. Behavior review checks whether the change does what the pull request says it does. Test coverage checks that the new behavior has a test. Repo guard checks the change against the project's own rules, written in the repository. Impact analysis walks the file graph and names the places a change could reach that the author did not touch.
Keeping them apart matters. A model asked to find a flaw in work it did a moment ago rarely finds one. Four fresh contexts cannot grade their own output.
Every finding has to carry its evidence
The README states the evidence contract directly. "Every finding records: location, summary, evidence, severity, status." A high-severity finding has to include the method signature or a concrete code path. There is no high finding that says "this looks risky" with no file and no line attached.
The skill never changes the code it is reviewing. The output is a human-readable report and a JSON record of every finding. A build server reads the JSON. It counts the high findings and the blockers, and decides whether the merge may go through. The gate is a number from a file, not a sentence from a model.
The one regression the author published
The README gives one measured run, on 29 October, against an eleven-case historical set. Five of five original behavior targets were hit. Four of four clean or fixed controls passed, meaning the probes did not block a change that was correct. Across the eleven cases, the probes produced 33 raw findings: 8 high, 20 medium and 5 low.
Eleven cases is a small sample, and the author ran it. The test shows the design holds on the cases the author chose. It does not show how another team's history would land. The honest read is that this is one run that passed its own check.
A team already using Claude Code or Codex on pull requests gets something new here. The review is broken into four smaller jobs, each with written evidence. The alternative is one long prose summary a human has to trust. The cost is one install line and four probe runs per review. The gain, if the design holds on a real codebase, is a number a build server can act on.
Source
code-review-skill on GitHub, kyy-logs.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
