The Agent Skills Directory

Standard Command Execution: The skill utilizes standard developer tools like git and the GitHub CLI (gh) to perform its primary function of managing repositories and issues.
Evidence: Found in SKILL.md and assets/acceptance-observer.md, where the agent is instructed to use git worktree, gh issue list, and gh pr create to manage development workflows.
Human-in-the-Loop Security Architecture: A central feature of the skill is the 'Phase 2' consolidated review, which serves as a mandatory security and quality gate.
Evidence: SKILL.md and assets/confidence-gate-rubric.md define a protocol where all triage decisions must be ratified by a human maintainer before any implementation sub-agents are spawned. The rubric explicitly routes issues involving 'security-surface' or 'auth-surface' to escalation by default.
Indirect Prompt Injection Surface: The skill ingests data from GitHub issue bodies, which are potentially provided by untrusted external users. This creates a surface for indirect prompt injection attacks.
Evidence: assets/triage-prompt.md ingests the ISSUE_BODY verbatim. This is mitigated by the 'confidence-gate' and human review phase, which are designed to detect and block suspicious or overly complex instructions embedded in issue descriptions.
Controlled Sub-agent Execution: The skill manages a complex fan-out architecture using sub-agents for specialized tasks like architecture planning and code implementation.
Evidence: assets/model-routing.md and assets/solution-pipeline-prompt.md detail how sub-agents are constrained to specific models and personas based on the stakes of the task, ensuring that high-stakes planning uses more capable, reasoning-heavy models while maintaining isolation between execution environments (worktrees).

apm-issue-autopilot