Code review
A read-only review workflow: one portable skill plus one Copilot custom agent. The skill frames the review and delegates an evidence-backed pass to the agent. The agent also works on its own.
| Asset | Path | Use when |
|---|---|---|
adversarial-review skill | SKILL.md | Asking for an adversarial or defect-focused review of code, a diff, or a plan |
code-reviewer agent | code-reviewer.agent.md | Delegating a read-only specialist pass directly |
Install
copilot plugin marketplace add OWNER/REPO
copilot plugin install code-review@MARKETPLACE
Replace OWNER/REPO with the GitHub repository that hosts this marketplace
and MARKETPLACE with the name in its marketplace.json. See
Install plugins for VS Code.
What this teaches
- Skill plus agent: a skill can delegate to a bundled agent by its advertised
identifier. Copilot CLI names the installed agent
code-review:code-reviewer. A workspace-staged profile appears ascode-reviewer. - Read-only tools: the agent only gets
readandsearch. It never edits, runs code, commits, or posts comments. - No quota: a clean result is valid. The agent returns
No actionable findings.rather than inventing defects.
Try it
Use adversarial-review to review this diff without changing it.
Or select the agent directly with --agent code-review:code-reviewer or the
agent picker.
Customize
Tune the agent's checklist for your stack, and keep the skill's delegation step pointed at the agent's identifier. If you rename the agent, update the skill, evals, and this README together.
Host checks
Before a release, try these by hand in each host you support. Automated tests and evals don't prove that a host loads or shows everything.
- Confirm the
code-revieweragent is listed, and select it with--agent code-review:code-revieweror the agent picker. - Ask for an adversarial review of a small diff. Confirm the
adversarial-reviewskill runs, delegates to the agent, and nothing is edited. - Ask an unrelated question and confirm neither the skill nor the agent runs.