The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from pull request descriptions, review comments, bot outputs, and CI logs.
Ingestion points: PR metadata and content are read via gh pr view, gh pr checks, and GraphQL queries in scripts/status.mjs.
Boundary markers: The skill contains an explicit 'Authority boundary' section instructing the agent to treat this data as untrusted evidence rather than instructions, and it strictly forbids merging based on external triggers.
Capability inventory: The skill can execute git push, perform file edits, and run gh CLI commands.
Sanitization: Instructions explicitly warn against pasting untrusted text into shell commands and require local checks before pushing changes.
[COMMAND_EXECUTION]: The skill executes shell commands using the gh (GitHub) and git CLI tools. It also runs an internal Node.js script (scripts/status.mjs) using execFileSync to collect PR data and compute readiness verdicts.
[EXTERNAL_DOWNLOADS]: The skill fetches reference documentation (e.g., evidence-ladder.md and process-skill-composition.md) from the author's public GitHub repository (github.com/csark0812/toolbox).

babysit-pr