The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from Git targets (branches, pull requests, or merge requests) and user-supplied 'free-form review briefs'. This content is analyzed by LLM-based 'Reviewers' to identify issues and propose code repairs.
Ingestion points: Git repository content (code, diffs), pull request metadata, and the user-provided review brief in SKILL.md.
Boundary markers: The skill enforces a 'report-only Reviewer boundary' during the evaluation phase and requires every agent in the tree to inherit these constraints.
Capability inventory: The skill has the ability to write to the local filesystem (repairs) and execute shell commands to run tests, linters, and type checks.
Sanitization: It includes a verification step where claims are treated as hypotheses and independently checked against the source, specifications, or observable behavior before being accepted as findings.
[COMMAND_EXECUTION]: The skill executes local system commands to verify code changes.
Evidence: Instructions in SKILL.md state: 'Run the smallest relevant tests, lint, and type checks. Exercise a changed output surface directly when existing tests do not observe it.'
Risk Mitigation: These operations are performed on an editable local checkout after a verification phase, intended to validate the correctness of generated repairs.

review-fix