The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct the agent to perform file system operations and version control tasks, specifically using rm -rf to clean up temporary plan directories and various git commands (e.g., git branch --show-current) to manage development branches.
[DYNAMIC_EXECUTION]: The prompt template for the code quality reviewer includes a command to verify npm peer dependencies using node -e. This involves executing a string as JavaScript to inspect package metadata during the review process.
[INDIRECT_PROMPT_INJECTION]: The skill's orchestration logic is driven by external content read from disk, including manifest.json, plan.md, and task-specific briefing files. This design creates a surface where the agent's execution flow and subagent dispatches are determined by data that could be influenced by untrusted external sources if the plan directory is not strictly controlled.
Ingestion points: The orchestrator reads task lists, wave structures, and detailed briefing instructions from the .claude/plans/ directory.
Boundary markers: The skill uses explicit instructions in its review prompts (e.g., spec-reviewer.md) telling agents not to trust implementation reports and to verify work independently from the source code.
Capability inventory: The skill possesses the ability to spawn new subagents with custom prompts, delete directories, and modify code via implementation agents.
Sanitization: A mandatory two-stage review process is defined to catch deviations or malicious additions before tasks are marked as complete.

plan-execution