The Agent Skills Directory

[PROMPT_INJECTION]: Indirect prompt injection surface. The scripts/agent_evaluator.py script processes untrusted user queries and file paths to generate a JSON payload containing 'IMPORTANT GUIDELINES'. These instructions are intended to be incorporated into the parent agent's context.
Ingestion points: User query and optional file path parameters passed as command-line arguments (sys.argv) to scripts/agent_evaluator.py.
Boundary markers: No explicit delimiters or boundary markers are used in the generated additional_context string to isolate the instructions from user-controlled input.
Capability inventory: While the script itself only outputs JSON to stdout, its recommendations control the activation of agents with broader capabilities, such as Git-Committer (git operations) and Instruction-Writer (file system modifications).
Sanitization: The script uses re.escape() for keyword matches, but performs regex-based intent matching directly on the raw user query, which could allow a crafted query to trigger specific agent activations or reasoning strings.
[SAFE]: No hardcoded credentials, sensitive file access, or network exfiltration patterns were detected.
[SAFE]: The skill relies entirely on the Python standard library. No external or unverifiable dependencies are required.
[SAFE]: Dynamic code execution in scripts/run_tests.py and scripts/agent_evaluator.show_test-cases.py is restricted to loading local test modules from the skill's own directory, which is a standard practice for unit testing and does not pose a remote execution risk.

agent-evaluator