The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the task runner to execute system-level diagnostics and service management actions. It references a Taskfile.yml in a sibling directory (../roofline-toolbox/) to perform tasks such as health, metrics, gpu, logs, restart, and smoke-toolcall.
[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest and analyze server logs and metrics to perform triage. Because server logs can contain data derived from external client requests or error messages, they serve as a potential ingestion point for indirect prompt injection.
Ingestion points: Output from the task logs command (e.g., in SKILL.md and references/symptoms.md).
Capability inventory: The skill has the capability to stop/restart the serving process, modify configuration profiles, and perform network smoke tests.
Boundary markers: No specific delimiters or "ignore instructions" warnings are provided for the agent when analyzing the log content.
Sanitization: The instructions do not specify any sanitization or filtering steps for the log data before processing.
[EXTERNAL_DOWNLOADS]: The skill includes references to official documentation and troubleshooting resources for vLLM (docs.vllm.ai), SGLang (docs.sglang.ai), and engineering roadmaps (roadmap.sh). These references target well-known and established technology services.

roofline-incident-ru