Hybrid SAST + Gemini/Claude LLM Security Reviewer for Agent & Developer PRs
Two tools with partial schemas and descriptions. 'audit' has 6 parameters with types and descriptions, but descriptions are brief (10-50 chars, below the 50-200 char baseline for LLM optimization). 'init-hook' has 3 parameters with types and descriptions, similarly brief. Neither tool description explains WHEN to use it vs alternatives, or what the output structure is. No documented output schemas. Error handling is absent, no guidance on what to do if Gemini API fails, git operations fail, or hook installation fails. Tool names lack clarity: 'audit' is generic (audit what? against what?); 'init-hook' uses a hyphenated noun rather than verb_noun pattern. Parameter descriptions lack format constraints, ranges, or examples of valid values.
Run a security audit against a target git branch diff
Install a local git hook to run security audits automatically before commit or push
Tool names lack action verbs. 'audit' is generic; 'init-hook' is a noun phrase. Should be 'run_security_audit' and 'install_git_hook' to clarify intent.
Parameter descriptions are too brief (10-50 chars vs 50-200 char baseline). E.g., 'Target git base branch to compare against' lacks format constraints, valid values, or when to override the default. LLMs cannot infer constraints from short descriptions.
No output schema documented. LLMs cannot plan downstream calls or extract results. What does 'audit' return? JSON? Markdown? Exit code only?
API keys exposed as parameters. 'anthropic-api-key' is a tool parameter, this leaks secrets into agent logs and traces. Use server-side secret injection via environment variables.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 39 | 2026-07-28+ | v2 |
No error handling guidance. If Gemini API fails, git diff fails, or hook installation fails, the tool returns nothing actionable. LLMs need 'retry this', 'ask the user', or 'this is fatal'.