The Agent Skills Directory

[COMMAND_EXECUTION]: The skill manages and runs inference server processes (vLLM, SGLang, or custom Docker containers) as defined in local profile configuration files. The command execution is handled via bash -c in scripts/serve.sh.
[DYNAMIC_EXECUTION]: Configuration profiles stored in the profiles/ directory are sourced as shell scripts in scripts/serve.sh to initialize the environment and execution parameters for the inference server.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to interact with local or remote OpenAI-compatible API endpoints (defined by INFER_URL) for monitoring, benchmarking, and health checks using curl and Python's urllib library.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external LLM API responses for metrics gathering and tool-calling validation, which presents a technical attack surface.
Ingestion points: API responses from the configured INFER_URL are processed in scripts/health.sh, scripts/bench.py, scripts/metrics.py, and scripts/smoke_toolcall.py.
Boundary markers: The skill relies on standard JSON structure for API responses rather than explicit natural language delimiters.
Capability inventory: The skill possesses process management capabilities (via serve.sh and stop.sh) and broad network access for API interaction.
Sanitization: External content is parsed using standard Python JSON libraries (json.load, json.loads), which provides basic protection against malformed or malicious payloads.

roofline-toolbox