The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from technical specifications (specs/PRDs) to generate execution 'drivers' (e.g., shell commands or browser actions). This behavior is fundamental to the skill's purpose of automating integration testing but constitutes an indirect injection surface. * Ingestion points: Reads spec and sub-issue files in SKILL.md (build mode). * Boundary markers: No specific delimiters are used to separate spec content from instructions. * Capability inventory: Executes derived drivers using curl, agent-browser, and build tools in SKILL.md (gate mode). * Sanitization: Validates <SPEC-ID> to prevent path traversal but does not sanitize the generated command drivers.
[COMMAND_EXECUTION]: The skill instructs the agent to run local development commands such as docker compose, pnpm dev, and curl to verify system state during the integration gate phase. These operations are explicitly limited to local test environments or authorized staging environments.
[SAFE]: The skill implements protective measures against common vulnerabilities, specifically validating identifiers used in file paths (<SPEC-ID>) to prevent directory traversal attacks.

integration-contract