The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides operational runbooks involving standard DevOps commands for infrastructure management and incident mitigation.
Evidence: Commands such as kubectl rollout undo, kubectl set env, kubectl patch, and git revert are defined for use by the agent during incident response scenarios.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to handle alerts and regressions that originate from untrusted or adversarial AI model outputs.
Ingestion points: Incident triggers and monitoring metrics (e.g., llm_hallucination_detected_total, llm_guardrail_violations_total) are derived from LLM interaction data.
Boundary markers: No explicit delimiters or boundary markers are defined in the response templates to separate instruction from processed data.
Capability inventory: The skill utilizes capabilities including kubectl, git, python, and curl to modify environment configurations.
Sanitization: No explicit sanitization or filtering of the triggering content is provided in the documentation.
[DYNAMIC_EXECUTION]: The skill references local script execution for quality assessment tasks.
Evidence: The Quality Regression Runbook involves executing python run_evals.py to assess model performance, which is a standard procedure in SRE workflows.

ai-sre-incident-response