The Agent Skills Directory

[SAFE]: The skill establishes strict operational boundaries, declaring itself as "planning-only" and explicitly forbidding the spawning of workers, execution of task code, or integration of changes.\n- [COMMAND_EXECUTION]: The skill includes a Python script (scripts/test_planner_contract.py) designed for static verification of its own instructional content. The script uses standard libraries (unittest, pathlib) to perform internal text comparisons and does not interact with the network or sensitive file system paths.\n- [INDIRECT_PROMPT_INJECTION]: The skill has a defined surface for indirect prompt injection. Ingestion points: the skill reads repository instructions, Git state, relevant source/test boundaries, and existing task contracts (SKILL.md). Boundary markers: instructions mandate using minimal necessary context, included/excluded absolute paths, and independent approval evidence. Capability inventory: the skill is strictly planning-only; no subprocess calls, filesystem writes, or network operations are provided for task execution. Sanitization: the planner requires an integration owner to inspect actual artifacts and diffs rather than trusting success summaries. This attack surface is appropriately mitigated.\n- [PRIVILEGE_ESCALATION]: The instructions contain multiple prohibitions against unauthorized actions such as silent installation, delegation, or bypassing loading gates, ensuring the agent operates under the principle of least privilege.

multi-agent-planner-acrazie