Skip to content

MOBIUS OPERATOR RUNBOOK

Mobius Operator Runbook (v0.1)

Last Updated: November 23, 2025
Maintainer: Operations


Purpose

Day-of-operations guide for Mobius operators to monitor GI/MII health, respond to incidents, and onboard engines safely.

Quick Mental Model

  • Engines (muscle) → Sentinels (brain) → GI Engine (heart) → Civic Ledger (memory) → Orchestrator (nervous system)
  • Operator focus:
  • Watch the heart rate (GI / MII)
  • Watch the blood pressure (error rates / latency)
  • Intervene when stress appears

What to Watch

A) DVA.LITE (new service)

Endpoints: - GET /health → returns { "status": "ok" } - GET /metrics → Prometheus metrics: - mobius_mii_current - mobius_mii_trend{direction="up|down|flat"} - POST /alerts/check → runs threshold check and sends alert if GI < 0.95

Recommended dashboard tiles: - MII current (large gauge) - MII 24h trend line - Count of alerts over last 24h

B) Broker Multi-Engine Dashboard

Surface: 1. GI distribution by engine combo - Average GI for local, antigravity-first, multi-engine 2. Engine error rates - antigravity_error_rate, openai_error_rate, claude_error_rate, deepseek_error_rate 3. Latency - P50 / P95 latency per engine 4. Decisions - % ok, % needs_human_review, % reject


Thresholds & Colors

  • Green: GI ≥ 0.98 — System healthy
  • Yellow: 0.97 ≤ GI < 0.98 — Watch trends; check engine errors
  • Orange: 0.95 ≤ GI < 0.97 — Investigate engines/task types dragging GI
  • Red (CRITICAL): GI < 0.95 — DVA.LITE should fire alert; consider safestop on risky flows or specific engines

Start-of-Day Checklist (Operator Ritual)

  1. Check DVA.LITE
  2. GET /healthok
  3. Confirm mobius_mii_current on metrics
  4. Scan GI trend (last 24h)
  5. Note any extended time in Orange/Red
  6. Review alerts
  7. Look for "GI Threshold Breached" events overnight
  8. Check engine status
  9. Error rates < ~2–3%
  10. Latency roughly stable vs yesterday
  11. Smoke test Broker
  12. Call /v1/deliberate with low-risk query using routingMode=local
  13. Call /v1/deliberate with low-risk query using routingMode=antigravity-first
  14. Expect status="ok" and GI ≥ 0.98
  15. Log to Ops ledger (markdown or issue):
  16. 2025-11-23T10:xxZ — Start-of-day check: GI=0.991, no overnight alerts, all engines green.

When GI Drops < 0.95 (Incident Playbook)

Triggers: DVA.LITE /alerts/check fires, Broker responses return needs_human_review repeatedly, or GI chart shows Red.

  1. Identify the source
  2. Filter recent deliberations by routingMode, engineId, and task type (code / policy / search / etc.)
  3. Look for a single engine with lower GI or a task channel causing drops
  4. Contain
  5. If a single engine (e.g., DeepSeek): set ENGINE_CONFIG.enabled = false and redeploy Broker
  6. If task type (e.g., code exec tools): temporarily block that tool in Antigravity (allowedTools limited)
  7. If noisy inputs overall: increase proportion of local routing; require human_required for sensitive categories (law/health)
  8. Human-in-the-loop
  9. For each flagged deliberation, have a Custodian label correct / incorrect / unsafe / low-evidence
  10. Store corrections (e.g., corrections/ or DB) for future training
  11. Document the incident
  12. Time window, root cause hypothesis, mitigation, re-enable decision, follow-ups (e.g., add extra Sentinel policy)

Adding a New Engine Safely (Onboarding)

  1. Register in ENGINE_CONFIG with enabled: false.
  2. Create client (e.g., similar to antigravityClient.ts) returning EngineResult with engineId, answer, riskFlags, latencyMs.
  3. Run shadow mode: routingMode=multi-engine; new engine runs in parallel but is never final.
  4. Measure GI with/without participation; check if GI is raised or lowered.
  5. If stable, allow primary for certain task types with guardrails and human review for new categories.

Human Override & Safestop

  • Per-engine switch: ENGINE_X_ENABLED=false
  • Global risky-flow switch: ENABLE_CODE_EXEC=false
  • Manual safestop script/button should:
  • Disable specific engines/tools
  • Force all /v1/deliberate to routingMode="local"
  • Broadcast "Mobius in Safe Mode" to ops (Discord/Telegram)

Safe Mode: Local Sentinels only; GI still computed; no external tools.


Suggested Ops Dashboard Layout

  • Top row: MII gauge, 24h GI trend, alerts last 24h
  • Middle row: Engine latency & error bar charts; decision types pie chart (ok / human_required / reject)
  • Bottom row: Table of last 20 deliberations (prompt hash/category, engines used, GI score, decision, ledger attestation link)