On-Call

On-call in Eridian Ops is shift state plus an assistant that can query logs, services, and the incident timeline. It does not take production actions unless you mark a tool high_risk and pass dual control.

Current shift

GET /v1/ops/on-call/shift
Authorization: Bearer eridian_sk_...
X-Eridian-Project: prj_ops_001

{
  "object": "eridian.ops.shift",
  "rotation": "payments-primary",
  "primary": { "id": "usr_oncall", "until": 1718416800 },
  "secondary": { "id": "usr_backup", "until": 1718416800 },
  "handoff_notes": "Ledger p95 watch. Do not raise budget without dual control."
}

Rotations are imported from PagerDuty or Opsgenie. You can also PUT a rotation on Enterprise.

Assistant

POST /v1/ops/on-call/query
Authorization: Bearer eridian_sk_...

{
  "project_id": "prj_ops_001",
  "incident_id": "inc_2847",
  "model": "auto",
  "question": "What is the current error rate on payments-api versus 1 hour ago?",
  "tools": ["logs", "metrics", "services"]
}

{
  "id": "ocq_12",
  "object": "eridian.ops.oncall.query",
  "model": "gemini",
  "answer": "payments-api 5xx is 6.2% now versus 0.4% at 13:00Z.",
  "citations": [
    { "type": "metric", "name": "http.5xx.rate", "service_id": "svc_payments_api" }
  ],
  "eridian": { "route": "gemini", "latency_ms": 1800 }
}

Without citations the client should show a warning. Empty retrieval is a valid answer.

Follow-the-sun

Enterprise contracts can staff human on-call for Eridian from Dubai, Hong Kong, and South Africa. That is our gateway. Your rotation remains yours.

Do not send Customer Data to a WhatsApp group. Use the assistant and the incident object.

Guardrails

• The assistant cannot kubectl delete, disable PII, or change residency.
• Tools that mutate (silence an alert, ack PagerDuty) require high_risk: true on the template and dual control when the project says so.
• During a SEV-1 on Eridian itself, follow platform Runbooks first.

Handoff

POST /v1/ops/on-call/handoff

{
  "from_id": "usr_oncall",
  "to_id": "usr_backup",
  "notes": "inc_2847 still open. Ledger rollback complete. Watching 5xx."
}

Handoff notes are part of the audit trail. Keep them short and free of secrets.

See Log Analysis for the query engine the assistant uses.