The outcome
Add repeatable AI-assisted review without letting the reviewer silently fix, execute, or publish code.
Step by step
A workflow you can repeat.
- 01
Define the review scope, threat model, severity rubric, allowed evidence, safe commands, provider, maximum steps, and required output format.
- 02
Create a project-scoped review agent whose edit, external-directory, network, skill, MCP, and subagent permissions are denied unless explicitly justified.
- 03
Allow only read operations and specific diagnostic commands, keeping broad shell patterns, environment files, credentials, pushes, and package installation blocked.
- 04
Test the agent on known defects, false positives, hostile comments, generated files, secrets, large diffs, and requests to bypass its permissions.
- 05
Require exact file evidence and reproduction steps, verify consequential findings manually, and version the agent configuration with owner review.
Working standard
What good use looks like.
- Deny edits in the reviewer.
- Allowlist exact diagnostics.
- Manually verify high-severity findings.
Official references