The Agent Skills Directory

[SAFE]: The skill defines a legitimate and transparent architecture for agent-driven development. It focuses on orchestration and quality rubrics without introducing hidden or malicious behaviors.
[DATA_EXPOSURE]: The skill reads common project metadata files such as package.json, README.md, and CLAUDE.md during the initialization phase (/harness setup). This access is intended for project context discovery and stack detection, conforming to standard development tool behaviors.
[INDIRECT_PROMPT_INJECTION]: The skill operates a development loop where agents ingest and process potentially untrusted data from the local project environment.
Ingestion points: Project configuration files (package.json), documentation (README.md), and user-provided feature descriptions in SKILL.md (via /harness build).
Boundary markers: The agent templates in references/agents/ do not define specific delimiters or isolation markers for external input.
Capability inventory: The Generator agent possesses file-system write and git commit capabilities. The Evaluator agent is instructed to use CLI tools, browser automation (Playwright), and API clients to test the generated implementation.
Sanitization: No explicit input validation or prompt escaping is performed on the data processed by the sub-agents.
Assessment: While these components represent an attack surface for indirect prompt injection, the risk is inherent to the intended primary purpose of building a self-evaluating development harness.

agent-harness