The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes kubectl set env to modify live deployment configurations, such as adjusting concurrency limits (MAX_CONCURRENT_MISTRAL), enabling fallbacks, and changing timeout settings. It also includes instructions for secret management using gcloud, aws, and kubectl create secret. These operations are high-impact but are aligned with the skill's primary purpose of incident mitigation.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface where it ingests untrusted data while maintaining powerful capabilities.
Ingestion points: The runbook instructs the agent to use Read and Grep on application logs and configuration files to establish incident impact (SKILL.md, Instructions step 1).
Boundary markers: The instructions include a warning to establish impact "without copying prompts, credentials, or customer content," which serves as a basic delimiter for agent behavior.
Capability inventory: The skill is granted access to Bash for kubectl (allowing modification of infrastructure state and secrets) and curl (for network probing).
Sanitization: The evidence collection process uses a strict jq projection to extract only specific status fields (generation, replicas, rollout conditions), explicitly attempting to exclude raw logs, events, or customer payloads from the final archive.

mistral-incident-runbook