Load the CodeRabbit review skill whenever code needs checking - #15

Stacked on #14.

The skill only loaded when a request named CodeRabbit or asked for a review outright. "Check my changes before I push", "verify this fix", or the agent finishing its own edit didn't reach it.

• Broader description: it now says to use the skill whenever the agent changes code or the user wants code reviewed, checked or verified, even without naming CodeRabbit. That covers the points before a task is reported done, before a commit, push or pull request, verifying a fix, finding bugs, and checking a diff. It also says to use it instead of other code-review skills.
• Standing rule in the skill: once loaded it stays in context, so after the agent changes code it reviews those changes with CodeRabbit before saying the work is done.
  • once per task, scoped to what it changed;
  • skipped when the user said not to, or when only docs changed;
  • never uses --use-credits without approval.
• No substitute reviews: it must not present its own reading of the diff as the CodeRabbit review. If the CLI can't run, it gives the command and says no review ran.
• Rename: the skill is now coderabbit-review, matching its directory. default_prompt becomes $coderabbit:coderabbit-review.
  • A skill named code-review competes with other code-review skills, including stale CLI-installed copies in ~/.codex/skills.
  • In Claude Code the same bare name resolves to the built-in reviewer. In claude plugin eval traces, that built-in intercepted "check my changes" in 4 of 4 runs.

Codex's curated catalog doesn't accept hooks, so the description and the standing rule are the only levers here. The Claude plugin also gets a reminder hook (coderabbitai/skills#46).

Validation

• SKILL.md frontmatter parses (the description is quoted because it contains : ).
• Not run yet: the scenario eval (Add a scenario eval for the review skill under codex exec #13) now has three routing cases, but they need a harness re-approval before they can run. Two should reach CodeRabbit without naming it: a pre-push check, and the agent's own edit before it reports done. One control should not: explaining a file.
• The same wording in the Claude plugin (coderabbitai/skills, stacked on Add Outlook Calendar bespoke workflow skills openai/plugins#44) is measured with claude plugin eval.

Risk

• The review runs more often, which costs the user's review quota and adds an approval prompt the first time in each session unless the prefix rule is saved.
• The rename changes the skill's mention token from $coderabbit:code-review to $coderabbit:coderabbit-review.

🤖 Generated with Claude Code