The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to handle external data that could contain malicious instructions aimed at overriding agent behavior or exfiltrating data.
Ingestion points: The workflow involves reading issue bodies, git diffs, automated-review output, PR comments, and handover documentation as described in SKILL.md.
Boundary markers: The instructions explicitly mention an 'untrusted-content rule' and command the agent to treat external content as 'evidence only' rather than instructions.
Capability inventory: The skill utilizes shell commands including git (for diffing and worktree management) and the GitHub CLI gh (for reviews, API interaction, and merging).
Sanitization: There are specific instructions to redact credential values and local file paths, and a requirement for human authorization (auto_merge_authorized) for critical operations.
[COMMAND_EXECUTION]: The skill relies on system-level commands to perform its primary functions.
Evidence: The workflow steps in SKILL.md prescribe the use of git diff, git worktree, gh pr review, gh api, and gh pr merge --squash --auto.

subagent-pipeline