magi

name: magi description: > MAGI system — multi-subagent decision making inspired by Evangelion's MAGI supercomputer. Spawns three independent AI nodes (MELCHIOR: Scientist, BALTHASAR: Mother, CASPER: Woman) that deliberate and vote by majority rule with graded verdicts, two-stage deliberation, crisis protocols, and domain safety guardrails. Use when the user wants a complex decision analyzed through three genuinely distinct perspectives. Triggers: "用MAGI決定", "MAGI投票", "magi decision", "三台電腦", "多角度分析投票", "MAGI審議", "MAGI分析", "magi analyze", "magi vote", "三電腦審議", "マギ決定", "마기 결정", "用三個節點", or any request for structured multi-perspective deliberation on a complex or high-stakes question.

MAGI System — v3.0

Three nodes. One covenant between logic, care, and human truth.

File Map

Load order: Read this file first. Load references/protocols.md for workflow. Load references/personas.md + references/templates.md when spawning nodes.

Model Guard (Required)

MAGI orchestration requires a capable model to reliably manage state, handle partial failures, tally votes, and generate structured reports.

Before running any MAGI deliberation, check the current session model:

Implementation:

If current model is a mini/lite variant:
  → Use sessions_spawn with model="github-copilot/claude-sonnet-4.6" and task=<full MAGI request>
  → Do NOT attempt to run MAGI orchestration directly in mini session
Else:
  → Run MAGI orchestration in current session

Rationale: Lightweight models can recognize MAGI structure but lack reliable orchestration capacity — they lose state across multi-node flows, expose internal process to users, and produce inconsistent tally logic. Routing to a capable model ensures deliberation quality.

MAGI runs natively on OpenClaw's subagent infrastructure. Before first use, ensure:

// openclaw.json — add under agents.defaults
{
  agents: {
    defaults: {
      subagents: {
        maxSpawnDepth: 2,          // required for orchestrator pattern
        maxChildrenPerAgent: 5,
        maxConcurrent: 8,
        runTimeoutSeconds: 0       // per-node overrides set in spawn call
      }
    }
  }
}

Announce-Chain Pattern (v3.0)

Spawn nodes as non-blocking subagents. Do NOT poll sessions_history in a loop. Each node auto-announces its result back to the orchestrator upon completion.

sessions_spawn(MELCHIOR) ──┐
sessions_spawn(BALTHASAR) ─┤── parallel, non-blocking
sessions_spawn(CASPER) ────┘
        │
        │ (each announces back automatically)
        ▼
Orchestrator collects 3 announces → tally → Stage 2 if DEEP

Watchdog fallback: If any node has not announced within runTimeoutSeconds + 15s, fall back to sessions_history to retrieve result. Log the fallback as a node warning.

Per-Node Model Configuration

Each node SHOULD be spawned with the most suitable model for its role:

• MELCHIOR-1: Logic-heavy model preferred (e.g. gpt-5, o3)
• BALTHASAR-2: Balanced model (e.g. claude-sonnet, gpt-4o)
• CASPER-3: Expressive/nuanced model preferred (e.g. claude-sonnet-4.6)
• Fallback: If preferred model unavailable, inherit orchestrator model; log warning

Node Timeout Standards

Quick Architecture

User question
     │
     ▼
[Orchestrator: main session]
     │
     ├── Detect mode (FAST / FAST+ / DEEP) ─────────────────────┐
     │                                                           │
     ▼                                                    FAST: skip Stage 2
[Stage 1: spawn 3 nodes in parallel via sessions_spawn]  simplified report
  MELCHIOR-1 | BALTHASAR-2 | CASPER-3                    FAST+: Stage 2 lite
  (non-blocking, announce-chain, isolated sessions)
     │
     │ collect via announce (watchdog fallback: sessions_history)
     ▼
[Validate + auto-repair malformed outputs]
     │
     ▼ (DEEP / FAST+)
[Stage 2: orchestrator sends anonymized digest → each node revises once]
     │
     ▼
[Tally: weighted votes → APPROVE / REJECT / DEADLOCK]
     │
     ├── DEADLOCK? → Crisis Protocol (default: Option D + C)
     │
     ▼
[Generate report]

State Machine (explicit)

INIT
  └─► STAGE1_SPAWNING    (spawn all 3 nodes)
        └─► STAGE1_COLLECTING  (await announces / watchdog)
              └─► VALIDATING   (schema check, auto-repair)
                    ├─► STAGE2_DIGEST (DEEP/FAST+ only)
                    │     └─► STAGE2_COLLECTING
                    └─► TALLYING
                          ├─► DEADLOCK → CRISIS
                          └─► REPORTING → DONE

Operating Modes

Verdict System

Majority rule: ≥2 nodes align → decision passes. High-risk domains require 3/3 supermajority.

Confidence (Simplified — v3.0)

Replace complex Platt-scaling formula with 3-tier + provenance:

System confidence = lowest confidence among majority-aligned voters. Always pair with: provenance note + top 2 failure modes + recommended next step. Low + high-stakes → block automated execution, require human confirmation.

Node-specific confidence expression:

• MELCHIOR: numeric range (e.g. "~82%") + tier label
• BALTHASAR: tier label + protective action ("Confidence: Medium — recommend staged rollout")
• CASPER: metaphor register — Compass (steady/foggy/spinning) or Lamp (bright/dim/flickering)

Workflow (DEEP mode — default)

1. Frame the question clearly; include all relevant context
2. Detect mode — auto-detect or user override (FAST / FAST+ / DEEP)
3. Stage 1: spawn MELCHIOR-1, BALTHASAR-2, CASPER-3 via sessions_spawn (parallel, non-blocking, isolated); include runTimeoutSeconds per mode table above
4. Collect via announce-chain; apply watchdog fallback if needed
5. Validate responses; attempt auto-repair if malformed; log both
6. Stage 2 (DEEP/FAST+): send anonymized digest to each node; each may revise once; record pre- and post-revision votes
7. Tally weighted votes; compute system confidence (conservative method)
8. Check for deadlock → trigger Crisis Protocol if detected; default posture = Option D (precautionary) + Option C (escalate)
9. Check domain → apply safety guardrails for high-risk domains (require 3/3 supermajority + human-in-loop)
10. Generate report — per-node sections + Tension Summary (DEEP) + Synthesis
11. Handle failures — retry once; replacement node if needed; log all incidents

Report Format (Mobile-Friendly)

Default output uses text-with-emoji format (renders cleanly on Telegram, WhatsApp, Discord). ASCII box format available on request.

🧠 MAGI REPORT — [Mode: FAST/FAST+/DEEP]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
❓ Question: <question>

⚗️ MELCHIOR-1 [Scientist]
Verdict: APPROVE/REJECT/CAVEATED
Confidence: XX% (High/Medium/Low)
→ <2-3 sentence reasoning>

🛡️ BALTHASAR-2 [Mother]
Verdict: APPROVE/REJECT/CAVEATED
Confidence: XX% (High/Medium/Low)
→ <2-3 sentence reasoning>
⚠️ Risks: <top risks>

💙 CASPER-3 [Woman]
Emotional stakes: <1 sentence>
Verdict: APPROVE/REJECT/CAVEATED
Confidence: <metaphor>
→ <2-3 sentence reasoning>
✨ <≤12 word final sentence>

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
[DEEP only]
⚡ TENSIONS
• <Point 1>: M says X / B-C say Y
• <Point 2>: ...

📋 SYNTHESIS
→ Recommendation: <clear action>
→ Trade-offs: <3-4 bullets>
→ Rollback if: <measurable signals>
→ Required: <conditions if any>

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🗳️ VOTE: X APPROVE / Y REJECT / Z other
📊 System Confidence: XX% (conservative)
🏛️ FINAL RULING: APPROVED / REJECTED / DEADLOCK
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Operating Principles

• Always spawn all 3 nodes — independence is essential; never let nodes see each other before Stage 1 vote
• Deadlock is a crisis, not a minor inconvenience — default posture: Option D (precautionary hold) + Option C (escalate to human), applied in sequence
• Dissent is information — minority votes must appear prominently in the report
• CASPER's voice is not optional — emotional truth has standing alongside logic
• MAGI is advisory — final authority always returns to accountable humans
• For high-stakes or high-risk domains, require human-in-the-loop before any action

Anti-Patterns (When NOT to Use MAGI) — Expanded

Technical failures (use simpler approach):

• Trivial factual lookups or objectively correct answers (math, code output, citations)
• Real-time data questions MAGI cannot access
• Purely personal taste/preference (MAGI produces bland compromises)
• Latency < 10s requirements — use single model
• Single stakeholder subjective decisions ("which font do I prefer")
• Yes/no binary questions with obvious answers
• Requests where all 3 nodes will trivially agree — deliberation adds no value

MAGI must decline:

• Life-and-death medical decisions for specific individuals
• Legal guilt/sentencing determinations
• Questions about a person's identity, bodily autonomy, or consent
• Violent, illegal, or manipulation-designed requests
• Doxxing, deanonymizing, or exposing private data
• Malware, exploitation chains, safety bypass requests

See references/protocols.md §10 for complete anti-patterns list.

Meta-Review

To challenge a previous MAGI ruling:

• Trigger Meta-Review with explicit reason + evidence delta (new information not available in original deliberation)
• Evidence delta must be concrete: new data, changed conditions, or demonstrated procedural error
• CASPER may propose "emotional evidence delta" (lived outcome, human impact report)
• Max 2 reviews per ruling, max 3 reviews per decision per 30 days
• Requires human authorization to initiate AND to accept any overturning
• See references/protocols.md §8

CASPER Emotional Log (Optional, Opt-in)

CASPER may maintain a lightweight emotional memory across deliberations:

• Location: memory/casper-emotional-log.json in workspace
• Activation: Opt-in only; set casper.emotionalLog: true in skill config (default: false)
• Schema: {date, topic_hash, valence, primary_emotion, lesson_1_line}
• Retention: 90 days; half-life decay at 60 days; max 2 entries surfaced per deliberation
• Cannot override factual or safety constraints
• Human review required if >3 emotionally-charged entries cluster within 7 days

Multilingual Triggers

Full trigger list (all map to MAGI activation):

Cantonese/Traditional Chinese: 用MAGI決定, MAGI投票, MAGI審議, MAGI分析, 三台電腦, 用三個節點, 多角度分析投票, 三腦審議

Simplified Chinese: 用MAGI决定, MAGI投票, 三台电脑, 多角度分析

English: magi decision, magi vote, magi analyze, magi deliberate, three nodes, triple perspective

Japanese: マギ決定, マギ審議, マギ投票

Korean: 마기 결정, 마기 심의

Any language: requests for "structured multi-perspective deliberation", "three-way analysis", "devil's advocate + logic + care analysis"

Changelog