The Agent Skills Directory

[COMMAND_EXECUTION]: The framework's context injection hook (templates/hooks/context-injector.py) executes git commands via subprocess.check_output to retrieve branch names and recent commit history, which is used to provide relevant project state to the agent.\n- [DYNAMIC_EXECUTION]: The skill employs import('datetime') within its scaffolding script to insert current timestamps into generated documentation. It also configures a project-level hook system in .claude/settings.json that triggers local Python scripts during specific lifecycle events.\n- [INDIRECT_PROMPT_INJECTION]: The skill establishes an environment where multiple agents process external data from the project codebase, feature specifications, and user requirements.\n
Ingestion points: Reads contents from docs/specs/, docs/contracts/, and project source files to inform planning and building tasks.\n
Boundary markers: Uses instructional 'Golden Principles' to define agent behavior and constraints, though it lacks strict data-level isolation.\n
Capability inventory: Agents have access to high-privilege tools including file system modification, shell command execution (within permitted list), and browser automation via Playwright.\n
Sanitization: No evidence of automated sanitization or filtering of ingested project data before it is presented to the LLM agents.

harness-engineering