The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute multiple shell commands to manage the development lifecycle.
Evidence: Uses git for worktree management and commits, and gh (GitHub CLI) for PR creation, commenting, and merging (gh pr merge --squash --auto).
[INDIRECT_PROMPT_INJECTION]: The skill operates on external data sources which could contain malicious instructions designed to influence the subagents' behavior.
Ingestion points: Reads issue bodies in Step 1 and PR comments via gh api in Step 4 from SKILL.md.
Boundary markers: Step 3 and the Guardrails section suggest using 'self-contained prompts' and 'explicit out-of-scope boundaries', though no specific delimiters (like XML tags) are prescribed for the untrusted content.
Capability inventory: The subagents have the ability to write code, commit to the repository, push to remote branches, and merge PRs into protected branches (e.g., staging) as detailed in Steps 1, 4, and 7 of SKILL.md.
Sanitization: The workflow mitigates risk by employing a 'Reviewer subagent' (Step 3) to verify semantic logic and a 'Final OCR/CI settling gate' (Step 5) to validate changes before merging.

subagent-pipeline