The Agent Skills Directory

[SAFE]: The skill follows secure patterns for automated code analysis and modification. It employs a multi-step verification process including test execution and human-in-the-loop review for all non-trivial changes. While the skill processes untrusted code (Indirect Prompt Injection surface), it mitigates this through structured subagent prompts and narrow capability scopes.
[PROMPT_INJECTION]: Analyzed as a potential surface (Category 8) given the ingestion of untrusted diffs and code. (1) Ingestion points: Ingests code content via git diff, gh pr diff, and workspace file reads. (2) Boundary markers: Uses specific reference files (references/security.md, etc.) and 'What NOT to flag' rules to delimit analysis. (3) Capability inventory: CLI tools (git, gh), test runners (pytest, npm test, etc.), and subagent creation. (4) Sanitization: Findings are categorized by severity and safety, verified by specialized subagents, and require human approval for merging.
[COMMAND_EXECUTION]: The skill executes local version control tools (git, gh) and project-specific test runners. This execution is necessary for the stated purpose and is constrained by the verification discipline which fallbacks to human review if safety criteria are not met.

autofix