Security audit
Security checks for vulnerabilities and agentic risk
This is a disclosed code-review skill that reads diffs, runs tests/builds, and may use review helpers, with no evidence of hidden persistence, exfiltration, or automatic destructive behavior.
Install this if you want an active code-review workflow. Be aware that it may run your repository's tests/builds, fetch branch refs, create a temporary worktree, and use sub-agents to inspect diffs. Confirm before allowing any fixes or GitHub comment replies.
The trigger phrases are broad enough to auto-activate this skill for many common requests, causing it to run tests/builds and spawn sub-agents without a tightly scoped opt-in. That can lead to unintended command execution, resource consumption, and access to repository contents in contexts where the user only wanted lightweight advice.
L032 says the review pass is read-only on git state and forbids operations like checkout, stash, reset, and branch switching. In the same sentence it instructs using git worktree add /tmp/review-<sha> <sha>, which mutates repository state by adding a worktree, contradicting the stated read-only characterization.
The skill includes capability to post GitHub PR review replies via gh api, which exceeds a local code-review/reporting role and can modify remote collaboration state. In an automated or loosely supervised run, this broadens the blast radius from local analysis to external side effects, creating risk of unintended comment spam, disclosure of internal analysis, or acting on unverified findings.
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.
## Anti-patterns
- ❌ Editing files during the review pass. Report first, **always**.
- ❌ Fixing without asking, even for "obvious" issues. Obviousness is not consent.
- ❌ Skipping the test/build run because "the diff looks fine".
- ❌ Reporting a green build without actually running it.
- ❌ Blending scopes — one report mixing working-tree findings with branch-commit findings. Pick the scope (or run two clearly separated reports); branch scope follows [references/branch-review.md](references/branch-review.md).
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.
- ❌ Convening the lens council on a 15-line diff. Single-pass it.
- ❌ Spawning the lens agents serially instead of in parallel — one message, N agents.
- ❌ Skipping the critique round and publishing the raw union of lens findings. False-positive blockers erode trust fast.
- ❌ Applying an external reviewer's suggestion without checking it against the codebase — or opening the reply with "You're absolutely right!". Verify, then fix or push back with reasoning.
- ❌ Batch-applying six findings and running the tests once at the end. One finding, one validation.
- ❌ Switching branches or stashing during the review to "see the old version" — use `git show` or a throwaway worktree.
- ❌ Dumping the per-agent transcripts into the user's report. The council is internal deliberation — the user sees the synthesized report.
The trigger phrases are broad enough to auto-activate this skill for many common requests, causing it to run tests/builds and spawn sub-agents without a tightly scoped opt-in. That can lead to unintended command execution, resource consumption, and access to repository contents in contexts where the user only wanted lightweight advice.