gemini-runner-audit

# Gemini Runner Audit Skill Use this skill when you need Gemini to run in an isolated subprocess and return structured results without polluting the main agent context. ## When to use - Deep code audit (security, correctness, architecture) - Large repo analysis where clean parent context is important - Multi-turn Gemini analysis using `resume_session` ## Tool to call - MCP tool name: `gemini_audit` - Required argument: `task` - Optional arguments: - `cwd` (defaults to current working directory) - `model` - `mode`: `default | auto_edit | yolo | plan` - `timeout_seconds` - `resume_session` ## Default execution policy 1. Set `cwd` to current repo root unless user explicitly requests another path. 2. Start with `mode=default` and `timeout_seconds=1800` for heavy audits. 3. For short checks, use 120-300 seconds. 4. If the response includes a `Session ID`, reuse it for follow-up questions. ## Recommended prompts - Security: - `Run a comprehensive security audit of this repository and rank findings by severity.` - Bug hunt: - `Find concrete logic bugs and race conditions with file/line references.` - Docs quality: - `Evaluate README and docs for gaps, inconsistencies, and missing operational steps.` ## Output handling Always parse and surface: - `Status` - `Session ID` - `Summary` / `Full Response` - `Tool Uses` - `Errors` - `Artifacts` path (`result.json`) If status is `error` or `timeout`, include clear next action: - increase timeout - narrow task scope - retry with same `resume_session` if applicable ## Guardrails - Do not stream human mode for automation. - Do not suppress Gemini errors; report them explicitly. - Keep filesystem scope aligned with `cwd` to match user-visible project context.