The Agent Skills Directory

[PROMPT_INJECTION]: Indirect Prompt Injection Surface. The skill defines a workflow where agents process data from external sources, such as pull request descriptions and validation command outputs, which could contain instructions designed to influence agent behavior.
Ingestion points: pull request data (via gh pr view), repository file content, and validation/CI logs.
Boundary markers: The skill employs structured contracts (Verdict Contract and Scope Matrix) to organize data, which helps provide context but does not explicitly sanitize inputs against embedded instructions.
Capability inventory: The agents have the ability to execute code via codex, perform repository operations via git, and interact with GitHub via the gh CLI.
Sanitization: There is no evidence of input filtering or escaping for data retrieved from external sources.
[COMMAND_EXECUTION]: The skill instructions involve the use of various command-line utilities to manage the development lifecycle.
Evidence: Usage of git status, npm view, and npm ls for reconciliation and verification.
GitHub CLI: Usage of gh pr view and gh pr comment to interact with remote repositories.
Vendor Tools: Execution of codex exec for code modification and implementation tasks.

review-fix-signoff-loop