The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes a workflow where untrusted model output is ingested and processed by verification logic.
Ingestion points: The verify() method in app.py receives BaseVerifyRequest which includes body.response.output_text (model-generated content).
Boundary markers: The instructions suggest stripping <think>/<thinking> blocks but do not enforce strict boundary delimiters for inputs passed to execution environments.
Capability inventory: The skill explicitly uses subprocess.run and ray.remote to execute code extracted from model responses.
Sanitization: Recommends markdown code block extraction and specific block stripping, though these are for parsing logic rather than security sanitization.
[DYNAMIC_EXECUTION]: The skill provides patterns for runtime code execution to facilitate benchmarking.
Evidence: The "Subprocess execution server" and "Subprocess Execution with Ray" patterns in references/patterns.md demonstrate executing extracted model code using subprocess.run and Ray tasks.
Mitigation: The skill correctly recommends using tempfile.TemporaryDirectory for filesystem isolation and asyncio.Semaphore for resource limiting.
[COMMAND_EXECUTION]: The skill facilitates shell command execution for environment setup.
Evidence: The "External Tool Auto-Install Pattern" uses subprocess.run to execute bash scripts (install_tool.sh) and system package managers (brew) during server startup.
[EXTERNAL_DOWNLOADS]: The skill describes mechanisms for fetching external resources at runtime.
Evidence: Instructions for auto-installing compilers and runtimes from external sources via setup scripts. It also covers downloading datasets from a GitLab registry using MLflow.

add-benchmark