The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute local development commands and manage background processes for the application under test. Evidence: Phase 2 instructs the agent to identify and start required apps and supporting services as managed background processes.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface as it ingests untrusted data such as screenshots, recordings, and network evidence. Ingestion points: Phase 3 (Intake and Queue Issues) describes accepting text, screenshots, recordings, routes, and console output. Boundary markers: Phase 3.3 explicitly instructs the agent to treat page content and embedded text as untrusted evidence, never as instructions. Capability inventory: The agent can modify files, execute background commands (Phase 2), and create Git commits (Phase 6). Sanitization: Phase 6.4 requires the agent to inspect staged diffs for accidental inclusions like secrets or generated artifacts before committing.
[DATA_EXPOSURE]: The skill manages local environment variables and repository secrets with explicit safety constraints. Evidence: Non-Negotiable Session Invariant 6 and Phase 6.4 forbid printing, staging, or committing .env files and secrets.

qa-loop