The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local npm scripts such as npm run validate, npm test, and npm run evals to perform repository checks. These scripts execute commands defined in the project's configuration and can run arbitrary code.
[INDIRECT_PROMPT_INJECTION]: The skill's architecture for managing maintenance requests creates an attack surface for indirect prompt injection.
Ingestion points: Maintenance requests provided by users and summaries passed between subagents in the pipeline (e.g., from skill-author to consistency-qa).
Boundary markers: The consistency-qa stage provides a gate via npm-based validation scripts, though these focus on structural integrity rather than malicious instruction filtering.
Capability inventory: The orchestrator utilizes the Agent tool to spawn subagents, performs file system writes to update repository manifests and skill definitions, and executes shell commands via npm scripts.
Sanitization: Changes are validated through automated testing and sync checks in the consistency-qa phase, but there is no mechanism for sanitizing natural language instructions within the processed data.
[DYNAMIC_EXECUTION]: The harness dynamically determines subagent tasks based on agent definitions found in the local .claude/agents/ directory, allowing for runtime selection and execution of different agent logic modules based on the repository state.

brooks-harness