The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses bash and jq for local automation, including updating specs/state.json, modifying TODO.md through the Edit tool, and managing git commits. These operations are scoped to the project's local directories.
[REMOTE_CODE_EXECUTION]: Facilitates complex planning tasks by spawning a specialized planner-agent subagent via the Task tool. This architectural pattern separates the coordination logic from the actual planning and artifact generation.
[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface through the ingestion and processing of project-level metadata.
Ingestion points: Reads task-specific information from specs/state.json and .return-meta.json (generated by the subagent).
Boundary markers: Absent for variable interpolation within shell scripts or for the prompts sent to the subagent.
Capability inventory: Includes bash for system calls, the Edit tool for file manipulation, and the Task tool for process delegation.
Sanitization: Uses jq --arg for most status updates to mitigate injection risk, though some task-number lookups use direct shell interpolation within jq filters.

skill-planner