The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill generates Python scripts such as evaluate.py and run_evo.py by snapshotting existing code and wrapping it in new logic. The workflow includes a 'smoke test' step where the agent is instructed to execute the generated evaluate.py script to confirm it works correctly.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted source code from the user's current working directory, creating an attack surface where malicious content in the source files could influence the agent's behavior.
Ingestion points: The agent 'inspects the current working directory' and snapshots the 'source subtree' as described in the Workflow section of SKILL.md.
Boundary markers: There are no instructions for the agent to use delimiters or ignore potential commands hidden within the source code files it reads.
Capability inventory: The skill possesses the ability to write multiple files to the local filesystem and execute shell commands to run evaluation tests.
Sanitization: The instructions do not include requirements for sanitizing or escaping the content of the source files before they are incorporated into the generated scripts.
[COMMAND_EXECUTION]: The workflow explicitly requires the execution of shell commands (specifically python evaluate.py) to verify the generated task. Additionally, for non-Python projects, the generated templates suggest using subprocess to execute candidate programs.

shinka-convert