mundi-orch-multi-llm-consensus

name	mundi-orch-multi-llm-consensus
description	Cross-model consensus — spawn parallel calls to Claude + Gemini + GPT on the same question, aggregate via vote or merge, return a decision with divergence flagged. Use when a high-stakes decision needs genuine model-level diversity (not just framing variation within one provider). Distinct from /mundi:consensus which is Claude-only framing variation.
allowed-tools	Read, Write, Edit, Bash, Task, WebFetch

Mundi Orch — Multi-LLM Consensus

Overview

Most of Mundi Princeps decision workflows run on Claude. That means every "consensus" is really Claude-agreeing-with-itself. This orchestrator addresses the blind spot: takes a question, fans out in parallel to Claude + Gemini + GPT (and optionally local Ollama), aggregates responses via vote / merge / divergence-flag.

Paired with /mundi:consensus (Claude-only) — this is the cross-provider upgrade path.

When to use

• Strategic architecture decisions (one-way doors, major spend, hiring)
• Risk review on a recommendation Claude made (independent verification)
• Ethics / legal / compliance calls where provider-level bias matters
• Pre-mortem: "what could go wrong" — diversity of priors maximizes blind-spot coverage

Do NOT use for:

• Routine code tasks (cheap, single-model is fine)
• Questions with clear deterministic answers (single model + verification is enough)
• Low-stakes reversible decisions

Architecture

                    ┌─── Claude (Sonnet 4.6 + Opus 4.7)
   Question ───┬───┼─── Gemini (3.1 Pro, Files API)
                │   ├─── GPT (via Codex bridge or OpenAI API)
                │   └─── Ollama (optional local, Llama 3 / Qwen)
                │
                └──▶ Aggregator (Opus scribe)
                       │
                       ▼
                  Consensus report
                       {decision, divergence_level, outlier_flags, ...}

Pipeline (5 steps)

[1] Prepare prompt    → same question, same format expectation, templated
[2] Parallel fan-out  → parallel calls, 60s timeout each, capture structured JSON responses
[3] Diversity check   → detect mode collapse (all say the same thing) → if so, flag and retry with adversarial framing
[4] Aggregator scribe → Opus reads all responses, produces consensus + outlier + divergence
[5] Artifact persist  → docs/decisions/<YYYY-MM-DD>-<topic>-multi-llm.md

I/O contract (MWP)

state_reads:

• ~/Mundi Princeps/config/api_keys.json — Anthropic, OpenAI/Codex, Gemini keys
• ~/.claude/skills/claude-agent-sdk/SKILL.md — Claude SDK invocation pattern
• ~/.claude/skills/video-to-action/scripts/analyze-youtube.sh — Gemini curl pattern
• ~/.claude/skills/saraev-cc-codex-diversify/ — Codex/GPT bridge pattern

state_writes:

• docs/decisions/<YYYY-MM-DD>-<topic>/ — artifact dir
  • round1/claude.md, round1/gemini.md, round1/gpt.md — raw responses
  • consensus.md — aggregated verdict with divergence flags
  • metadata.json — timing, tokens, confidence

Composition

provider	invocation
Claude	claude-agent-sdk skill or /mundi:debate-style direct call
Gemini	curl via Gemini 3.1 Pro API, Files API for long context
GPT	Codex CLI bridge (saraev-cc-codex-diversify) OR OpenAI API
Ollama	subprocess ollama run <model> "..." — optional, for local diversity
aggregator	Opus scribe, fixed-role prompt: "quote + tally + flag divergence, do not re-argue"

Aggregation rules

Divergence levels:
- LOW       — all providers agree on top-line decision + reasoning
- MEDIUM    — agree on decision, differ on rationale/tradeoffs
- HIGH      — differ on decision itself
- OUTLIER   — N-1 agree, 1 dissents strongly → surface dissent prominently

Output structure:
- Decision: <recommended path, if any>
- Consensus rationale: <what they all agree on>
- Divergence: <where and why they split, by provider>
- Outlier flag: <if present>
- Assumptions load-bearing: [a1, a2, a3]
- Reversal trigger: <observation that would flip the decision>

Failure modes

failure	recovery
One provider API down	proceed with available 2; flag degraded: missing <provider>
All 3 agree too strongly	possible echo; re-run round 2 with "steelman the opposing view" instruction
Aggregator produces low-confidence output	run a 4th provider; if still low, surface "no consensus"
Rate-limit	exponential backoff; never retry a 429 within 30s

Invocation

Skill(skill="mundi-orch-multi-llm-consensus", {
  question: "Should we open-source the capiq-mcp tool registration layer?",
  providers: ["claude", "gemini", "gpt"],
  aggregator: "opus",
  format: "structured" | "prose",
  save_artifact: true
})

Output contract

{
  "question": "...",
  "providers_called": ["claude", "gemini", "gpt"],
  "providers_succeeded": ["claude", "gemini", "gpt"],
  "divergence_level": "LOW|MEDIUM|HIGH|OUTLIER",
  "decision": "...",
  "consensus_rationale": "...",
  "divergence_by_provider": {"claude": "...", "gemini": "...", "gpt": "..."},
  "outlier": null | {"provider": "...", "view": "..."},
  "load_bearing_assumptions": [...],
  "reversal_trigger": "...",
  "artifact_path": "docs/decisions/..."
}

Cross-references

• Distinct from: /mundi:consensus (Claude-only stochastic framing variation), /mundi:debate (Claude-only chat.json debate)
• Related: mundi-orch-multi-llm-debate (adversarial version), mundi-orch-multi-llm-route (task→provider classifier), stochastic-multi-agent-consensus
• Plugin skills: linkdrop-x-tom-doerr-multi-llm-council-deliberation (external /council pattern — 18-persona cross-model)
• Memory: feedback_linkdrop_faithfulness.md (no invented content in aggregator)
• KB: docs/knowledge-base/wiki/01-ai-development-and-agents/INDEX.md
• Plan source: docs/plans/2026-04-19-phase-3-final.md line 862

Safety

• Provider diversity required. If only Claude responds, this degrades to /mundi:consensus — do not pretend.
• Aggregator is a scribe, not an arbiter. Lock its role: quote + tally + flag divergence, never re-argue.
• Not a substitute for human judgment — output is an input to your decision.