MetaHarness: devops vertical

A ready-made multi-agent scaffold for incident response on top of Claude Code. It provisions four specialized agents — responder, runbook-runner, escalator, and postmortem — wired together with tiered model routing, project-scoped settings, and a harness doctor health check. Use it to bootstrap an on-call assistant that triages alerts, executes runbooks, escalates when stuck, and writes the postmortem afterward.

This scaffold is for SREs, platform engineers, and on-call developers who want a working incident-response harness in one command instead of assembling agents, prompts, and config by hand. It is NOT a hosted service, an alerting system, or a replacement for PagerDuty / Opsgenie / Grafana OnCall — it is the Claude Code layer that sits next to those tools.

Quickstart

npx @metaharness/devops@latest my-bot
cd my-bot && npm install && npx harness doctor

Then launch Claude Code against the scaffold:

claude -p --plugin-dir my-bot "page: api-gateway 5xx spike, last 10m"

What you get

• agents/responder.md — first-touch triage agent (tier: sonnet). Reads the alert, pulls recent logs/metrics, proposes hypotheses.
• agents/runbook-runner.md — executes documented runbooks step-by-step (tier: haiku). Cheap, deterministic, fast.
• agents/escalator.md — decides when human escalation is warranted and drafts the page (tier: sonnet).
• agents/postmortem.md — writes a blameless postmortem from the incident transcript (tier: opus).
• .claude/settings.json — project-scoped Claude Code config with allow-listed tools (Read, Grep, Glob, Bash for read-only commands).
• mcp.json — pre-wired MCP server stubs for log search and metrics queries; edit to point at your stack.
• runbooks/ — example runbook templates the runbook-runner agent consumes.
• package.json with a harness bin shim so npx harness doctor and npx harness validate work out of the box.

Advanced

Run the bundled doctor to verify the scaffold is healthy:

$ npx harness doctor
[harness] node: v20.11.1  ok
[harness] claude code: detected
[harness] agents: 4 found (responder, runbook-runner, escalator, postmortem)
[harness] mcp.json: 2 servers configured
[harness] settings.json: valid
ok

Validate the agent frontmatter and tier assignments:

$ npx harness validate
[harness] agents/responder.md      tier=sonnet  tools=Read,Grep,Bash  ok
[harness] agents/runbook-runner.md tier=haiku   tools=Read,Bash       ok
[harness] agents/escalator.md      tier=sonnet  tools=Read,WebFetch   ok
[harness] agents/postmortem.md     tier=opus    tools=Read,Write      ok
ok

Drive the harness headlessly for CI or webhook integrations:

claude -p --plugin-dir my-bot --output-format json \
  "incident: redis OOMKilled in prod-us-east-1, alert id INC-4421"

Pipe a recent alert payload from your incident tool straight into the responder:

cat alert.json | claude -p --plugin-dir my-bot --append-system-prompt "You are the on-call responder. Start with the responder agent."

FAQ

Q: Does this replace PagerDuty / Opsgenie? A: No. It plugs into them. The escalator agent drafts the page; your existing tool delivers it. Wire the MCP server in mcp.json to your incident platform's API.

Q: Can I change which model each agent uses? A: Yes. Edit the model: field in the YAML frontmatter of each agents/*.md file. The defaults (haiku for runbook execution, sonnet for triage, opus for postmortems) follow the 3-tier cost/latency pattern but every choice is a one-line override.

Q: How do I add my own runbook? A: Drop a markdown file into runbooks/. The runbook-runner agent discovers them by glob and matches against the incident description. No registration step.

License

MIT. Built on metaharness (https://www.npmjs.com/package/metaharness).

Deep-dive

Full explainer gist: https://gist.github.com/ruvnet/7c2975652c9be393536a71cdafc358f3