model-consensus-mcip

name	model-consensus-mcip
description	Orchestrate multi-model consensus (Gemini 3 Pro/Flash, Sonnet 4.6, Codex 5.3, Opus 4.6) for i2i/MCIP design, protocol analysis, and implementation decisions in OpenClaw. Use when you need higher-confidence outputs by collecting independent model opinions, reconciling disagreements, and producing a single consensus recommendation with rationale and confidence.

Model Consensus MCIP

Run consensus through your real i2i MCIP stack, not a generic orchestration.

When to Use

• You need a high-confidence decision for MCIP/i2i architecture, protocol behaviour, or rollout choices.
• You want independent model perspectives reconciled into one execution-ready recommendation.
• You need protocol-native consensus output (HIGH|MEDIUM|LOW|NONE|CONTRADICTORY) instead of ad-hoc voting.
• You need a safe "no consensus" outcome when evidence is weak or contradictory.

Canonical Protocol Sources (read first)

• /home/generic/dev/i2i/RFC-MCIP.md (protocol spec)
• /home/generic/dev/i2i/i2i/protocol.py (runtime API)
• /home/generic/dev/i2i/i2i/schema.py (message/result schema)
• /home/generic/dev/i2i/config.json (default model set)

Treat these as source of truth for levels, fields, and behaviour.

Workflow

1. Normalize the question into MCIP terms

• Frame as a consensus_query task.
• Preserve protocol constraints: safety, liveness, interoperability, implementation cost, rollout risk.

2. Run consensus via real i2i protocol

• Prefer using scripts/run_mcip_consensus.py.
• Execute from repo root:
  • cd /home/generic/dev/i2i
  • uv run python /home/generic/dev/skills/model-consensus-mcip/scripts/run_mcip_consensus.py --query "..."
• This calls Protocol.consensus_query(...) from the i2i library.

3. Model panel selection

• Default to i2i config model set when available.
• If user asks for OpenClaw aliases (gemini-pro, sonnet, codex, opus), map to i2i-compatible IDs:
  • gemini-pro -> gemini-3-flash-preview (or configured Gemini model)
  • sonnet -> claude-sonnet-4-5-20250929 (or configured Sonnet model)
  • codex -> gpt-5.2 (proxy for OpenAI coding-capable panel member)
  • opus -> use only as tie-breaker in synthesis stage if configured in i2i provider/model list
• Never invent model IDs; verify against configured providers.

4. Reconcile with protocol-native outputs

• Use MCIP consensus levels exactly: HIGH | MEDIUM | LOW | NONE | CONTRADICTORY.
• Treat LOW/NONE/CONTRADICTORY as no reliable consensus.
• Include task-aware metadata when present (consensus_appropriate, task_category, confidence_calibration).

5. Produce execution-ready recommendation

• Include protocol impact, implementation steps, validation, and rollback trigger.
• If confidence is weak, return required follow-up experiments instead of forcing a decision.

Output Contract

Use this structure:

## Consensus Decision
- Recommendation:
- Confidence:
- Why this wins:

## MCIP / i2i Impact
- Consensus safety:
- Cross-chain interoperability impact:
- Failure domain changes:

## Implementation Plan
1.
2.
3.

## Validation
- Unit tests:
- Simulation/fuzz:
- Cross-chain integration tests:
- Adversarial tests:

## Open Risks
- Risk:
  - Mitigation:
  - Owner:

Model Routing Rules

• Default panel: gemini-pro + sonnet high + codex high
• Add opus high only when:
  • split decision remains after reconciliation
  • safety-critical disagreement exists
  • economic/security assumptions conflict
• For lightweight checks, use gemini-flash + sonnet medium only.

Implementation Notes for OpenClaw

• Primary path: run real i2i MCIP (Protocol.consensus_query) instead of manual orchestration.
• Secondary path: use independent model panel prompts only when i2i runtime is unavailable.
• Do not leak one model’s answer into another model’s prompt before first-pass outputs are collected.
• If consensus level is LOW/NONE/CONTRADICTORY, return “no consensus” with required follow-up experiments.

References

• Use references/consensus-prompt-template.md for independent panel fallback template.
• Use references/mcip-binding.md for concrete protocol binding rules and field mapping.
• Use scripts/run_mcip_consensus.py to execute consensus on the real i2i implementation.
• Use scripts/score_consensus.py to aggregate model panel JSON when running fallback orchestration.