The Agent Skills Directory

[Command Execution]: The skill utilizes command-line tools such as git, gh (GitHub CLI), and ruff to perform its primary functions. These are used for issue enumeration, branch management, code linting, and PR creation. The orchestrator delegates these actions to isolated subagents and verifies the results through deterministic gates.
[Indirect Prompt Injection Considerations]: The skill processes untrusted input in the form of GitHub issue reports and PR data, which enter the context during the triage and fixing phases. The architecture implements a robust evidence chain: ingestion via issue bodies is gated by mandatory reproduction on HEAD, strategic alignment checks via the apm-ceo persona, and functional verification through CI and regression tests. These boundary markers and sanitization steps significantly reduce the risk of the agent following malicious instructions embedded in issue data.
[Subagent Orchestration]: The skill fans out tasks to specialized subagents for triage, alignment, and fixing. This compartmentalization helps isolate the processing of potentially malicious issue descriptions and ensures that destructive actions (like pushing code) are subject to multiple layers of automated verification.
[Environment Isolation]: The skill uses git worktree to ensure that each mutating operation happens in a dedicated directory. This prevents race conditions and cross-contamination between parallel tasks, which is a significant operational and security best practice for automated code modification.

batch-bug-shepherd