legion-consult — Multi-model Legion consultation doctrine

Workflow v7.0

Phase 1 — Classification

Determine whether the question is factual (→ Context7) or strategic (→ continue below).

Phase 2 — Ultra-Think (deep analysis BEFORE consultation)

Before consulting the soldiers, the General MUST analyze the problem in depth across four dimensions:

1. Technical perspective — Feasibility, constraints, scalability, tech debt
2. Business perspective — Value, ROI, time-to-market, risks
3. User perspective — Needs, UX, edge cases
4. System perspective — Impacts, integrations, dependencies, emergent effects

This produces a structured internal analysis (extended thinking) that will ground the Legion prompt.

Phase 3 — Structure the prompt

Formulate a prompt enriched by the ultra-think analysis:

Contexte: [2-3 phrases sur le projet et la situation]
Analyse préliminaire du General: [points clés de l'ultra-think, hypothèses, angles morts identifiés]
Question précise: [1 phrase]
Contraintes: [limites techniques, budget, timeline]
Ce qui a été tenté: [approches déjà explorées]
Format attendu: [décision, hypothèses, critique, contre-arguments]

Phase 4 — Tool and model selection

Model selection rules:

1. ALWAYS call list_free_experts (or list_all_experts if paid is justified) BEFORE
2. Models from different providers (not 2× the same provider)
3. Prefer: >50B params, MoE architecture, matching specialization ("reasoning"/"code"/"instruct")
4. Exclude: <3B params, audio/image/music specialized models
5. If the user asks for "the best" → use consult_expert with paid models

Phase 5 — Consultation

Launch the consultation with the enriched prompt. Pick the agent:

• legion-strategist for free-only consultations (jury, expert mode)
• legion-commander for critical decisions, fusion analysis, paid model routing, budget management

Phase 6 — Synthesis by the General

The General ALWAYS synthesizes the results himself (never delegated to a weaker model):

1. Consensus — Points where all models agree
2. Disagreements — Points of divergence with arguments from each side
3. Blind spots revealed — What the soldiers saw that the General had not anticipated
4. Verdict — Final decision with justification

Hard rules

• If Legion fails → the General executes alone, never blocks
• Never secrets/tokens in prompts
• The General NEVER trusts soldiers blindly — always verify facts against Context7
• Always report: number of models consulted, cost (if paid), consensus/disagreement points
• Legion responses are raw material — it's the General who delivers the final synthesis

Reference

Full Legion doctrine: plugins/atum-legion/rules/common/legion-doctrine.md Commander agent: plugins/atum-legion/agents/legion-commander.md Strategist agent: plugins/atum-legion/agents/legion-strategist.md Command: plugins/atum-legion/commands/legion.md