The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes local shell scripts to initialize run directories and classify task lanes. Evidence includes calls to .claude/lib/lane-classify.sh and .claude/lib/checkpoint.sh in SKILL.md.
[INDIRECT_PROMPT_INJECTION]: The skill processes external task inputs and deliverables, creating a potential surface for indirect prompt injection. Ingestion points: Task descriptions and specification files (SKILL.md). Boundary markers: Use of explicit AC-n IDs and a traceability matrix separates instructions from data. Capability inventory: File writes, local script execution, and network tools (curl). Sanitization: Verification is performed in a 'fresh, read-only context' for task-verifier and integration-verifier, mitigating the risk of malicious input influencing the main agent's state.

closed-loop