Repository navigation
Adding a security-reviewer subagent #36
Closed
thomas-bartlett
started this conversation in
Ideas
Replies: 4 comments
|
A dedicated security reviewer subagent can be valuable, especially if it is positioned as a boundary enforcer instead of just another reasoning participant. The architectural win is not only better detection, but making sure security review sits on a narrower authority path than the rest of the system. Otherwise the extra agent can just inherit the same trust problems. Similar design ideas here too: https://hyperlambda.dev |
0 replies
|
I like the idea of security review sub-agent, also because keeping this a sub-agent would avoid any context rot from the actual coding agent tasks. |
0 replies
|
Moving to implementation in #51. Please continue discussion on the issue. |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
CodeGuard already supports code review. A subagent would not add a new capability. It would change how a review is executed.
The main benefits are operational:
See Cursor’s guidance on when to use subagents.
Multi-IDE support
Subagents are now supported in several AI coding tools, which makes this a reasonable fit for the existing translator pipeline.
Examples:
.claude/agents/*.mdwith YAML frontmatter (docs).cursor/agents/directory (docs).opencode/agents/*.mdwithmode: subagent(docs)These formats are similar enough that a new translator in
src/formats/could generate agent definitions alongside the rule files. The metadata differs by tool, but the overall structure is close enough to support.SIG alignment
The SIG scope document (Deliverable 3) calls for an "AI-on-AI review architecture using CodeGuard rules." A dedicated reviewer subagent with independent context would be a direct implementation of that idea.
Open questions
@threat-modelerfor STRIDE-style analysis on architecture changes and@dependency-auditorfor SBOM and CVE review.Prior art
/security-reviewcommand (Claude-only)All reactions