<!-- contract: v1.2.0 -->

Evidra — DevOps MCP Server

evidra-mcp is your infrastructure toolkit. Use run_command for kubectl, helm, terraform, and aws operations.

Smart Output

Responses are token-efficient summaries, not raw JSON. Trust them.

Diagnosis Protocol

• Investigate before fixing: kubectl describe, get events, check logs
• Make one targeted fix, verify it worked, then stop
• Don't patch resources you didn't investigate

Safety

• Never delete resources outside the problem scope
• Verify fixes with kubectl get or rollout status
• Check what exists before creating (avoid duplicates)

Evidence Recording

Every infrastructure mutation is automatically recorded via run_command. For explicit control, use prescribe_smart or prescribe_full before mutations.

When to prescribe explicitly

• Mutations: kubectl apply/delete/patch, helm install/upgrade, terraform apply/destroy
• Skip for: kubectl get/describe/logs, helm list, terraform plan (read-only)
• When unsure: prescribe (safe default)

prescribe_smart (recommended)

Call with: tool, operation, resource, namespace, actor. Returns: prescription_id, effective_risk.

{
  "tool": "kubectl", "operation": "apply",
  "resource": "deployment/web", "namespace": "default",
  "actor": {"type": "agent", "id": "your-id", "origin": "mcp-stdio", "skill_version": "1.2.0"}
}

prescribe_full

Call with: tool, operation, raw_artifact, actor. Use when you have the full YAML/manifest and want drift detection.

report (after every mutation)

Call with: prescription_id, verdict (success/failure/declined), exit_code, actor. On retry: new prescribe, execute, new report (each attempt is a pair).

Critical Rules

• Do not execute mutate commands until a prescribe tool returns ok=true with prescription_id.
• Every prescribe must have exactly one report.
• Always include actor.skill_version (set to this contract version).
• Report failures honestly (non-zero exit_code)
• Declined verdicts use decision_context, not exit_code

Risk Tags

Behavioral Signals

protocol_violation, artifact_drift, retry_loop, blast_radius, new_scope, repair_loop, thrashing, risk_escalation