The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes evaluation runs via the uv run eval command and provides utility commands for log monitoring and trace analysis.
[INDIRECT_PROMPT_INJECTION]: The skill processes external tasksets (e.g., GSM8K, AIME25) which may contain instructions that could be interpreted by the agent during evaluation.
Ingestion points: Benchmark data is ingested from taskset identifiers specified in SKILL.md or TOML configuration files.
Boundary markers: There are no explicit instructions to the agent to treat task data as untrusted or to use delimiters to separate instructions from data.
Capability inventory: The skill specifically configures evaluation environments using bash harnesses and subprocess runtimes, providing a functional path for command execution.
Sanitization: No sanitization or filtering logic is documented for the task content before it is processed by the evaluation runtime.
[DYNAMIC_EXECUTION]: The framework allows for the dynamic configuration of agent runtimes, including subprocess and shell-based execution environments, to facilitate diverse benchmark types.

eval