The Agent Skills Directory

Indirect Prompt Injection Surface: The skill ingests untrusted data from GitHub issue bodies and titles into its triage and implementation phases (e.g., in assets/triage-prompt.md). This represents a potential surface for embedded instructions. However, the skill provides significant mitigation through a mandatory human-in-the-loop checkpoint in Phase 2, where a maintainer must review and approve all implementation plans before any code is written or labels are applied.
Command Line Operations: The orchestrator performs various system-level operations using tools such as git, gh (GitHub CLI), and uv (Python environment manager). These tools are used to manage local worktrees, create branches, and open pull requests within the microsoft/apm repository. The use of isolated Git worktrees for implementation tasks ensures that operations are scoped to specific issues and do not interfere with the primary working directory.
Skill Composition and Local Dependencies: The skill is designed to compose other local sibling skills, such as apm-triage-panel and shepherd-driver. It uses deterministic probes (e.g., test -f) to verify the presence of these components before execution. This modular approach relies on components within the same trusted local environment and follows established development patterns for complex agents.
Development and Testing Utilities: The package includes a Python script (scripts/run_evals.py) used for evaluating skill performance. This script uses standard libraries to perform keyword matching and regex-based scoring against pre-recorded fixtures. It is a utility for skill maintenance and does not perform network operations or dynamic code execution.

apm-issue-autopilot