The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes git and gh CLI tools for repository management, metadata retrieval, and fetching pull request information. These operations are restricted to standard development workflows.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from pull request bodies, linked issues, and CodeRabbit comments.\n
Ingestion points: Data enters the agent context via gh pr view and gh api calls in SKILL.md.\n
Boundary markers: The skill uses <HARD-GATE> markers and explicit warnings to treat external data as "claims only."\n
Capability inventory: The agent can execute git, gh, and pnpm commands, and run agent-authored reproduction scripts.\n
Sanitization: The instructions explicitly forbid executing commands, scripts, or test invocations found in the untrusted data.\n- [DYNAMIC_EXECUTION]: The agent is instructed to author and run its own reproduction scripts in isolated git worktrees. This is a controlled execution of dynamically created code used for validation.\n- [EXTERNAL_DOWNLOADS]: The skill may trigger pnpm install within a pull request worktree. Risks are mitigated by a mandatory manual review (Gate 0) if sensitive files like package.json or lockfiles have been modified in the PR.

bugfix-pr