oncall-runbook

When an incident fires, run a structured triage: fetch the incident, investigate across three independent angles in parallel (for_each fan-out), rank the findings, and produce a one-page summary for the oncall engineer.

What it demonstrates

• tool → agent chaining across two phases
• for_each fan-out on the investigate node (one agent call per dimension)
• array_append merge strategy for accumulating findings from parallel iterations
• Sequential synthesis: ranked findings → prose summary
• Multi-LLM compatibility (ANTHROPIC_API_KEY or OPENAI_API_KEY)

The graph

┌───────────────┐   ┌────────────────┐
│ fetch_incident│   │ seed_dimensions│
└──────┬────────┘   └───────┬────────┘
       │  triage phase      │
       └────────────────────┘
                │
                ▼
       ┌─────────────────┐
       │   investigate   │  ← for_each over dimensions (×3)
       └────────┬────────┘
                │  synthesis phase
                ▼
          ┌──────────┐
          │ prioritize│
          └─────┬────┘
                ▼
           ┌────────┐
           │ summary│
           └────────┘

Phases: triage → synthesis.

State schema

Field	Type	Merge	Description
incident	object	—	Raw incident payload
dimensions	array<object>	—	Investigation angles seeded by list_dimensions.mjs
findings	array<object>	array_append	Accumulated from each investigate iteration
prioritized_findings	array<object>	—	Ranked by the prioritize agent
summary	string	—	Oncall-facing one-pager

How it runs

export ANTHROPIC_API_KEY=sk-...   # or OPENAI_API_KEY=...
oe run examples/oncall-runbook --tui

No special CLIs required. The --tui flag shows a live dashboard of node statuses and elapsed time.

What happens

1. triage phase — fetch_incident loads fixtures/incident.json into state. seed_dimensions writes three investigation angles: service_health, infra_layer, dependency_failures.
2. fan-out — investigate runs once per dimension (three concurrent-ish calls in V1's sequential scheduler). Each call reads incident and the current $item dimension, returns structured findings[]. The array_append merge concatenates all three result arrays into one flat findings list.
3. synthesis phase — prioritize reads all findings + the incident and returns a ranked prioritized_findings[]. summary reads the incident + ranked findings and writes a human-readable summary string.

Sample final state:

$ oe state summary
[oncall one-pager prose...]

$ oe state prioritized_findings | head -5
[{"title": "Payment service 500 rate 12%", "priority": "P1"}, ...]

Fixtures

fixtures/incident.json contains a sample PagerDuty-style incident payload. Replace it with a real incident to triage a live event.

Try it: variations

1. Add a fourth dimension. Edit list_dimensions.mjs to return a fourth object, e.g. { key: "data_layer", focus: "database slow queries" }. The for_each fan-out automatically picks it up — no graph changes needed.

2. Point at a real PagerDuty API. Replace fetch_incident.mjs with a fetch call to the PagerDuty REST API (using the incident ID from --args). The rest of the graph is unchanged.

3. Add a Slack-post tool node. Append a post_summary tool node after summary that POSTs summary to a Slack webhook. Chain { from: summary, to: post_summary }.

Source

examples/oncall-runbook/