← Mission Control
◈ CoddexaCOUNCIL — Complete Deliberation Flow
Entry
User Submits Topic
Boss types a question or decision topic and clicks SUBMIT TO COUNCIL. A session is created instantly with a unique ID.
Output
session_id = cs-{date}-{hex} status: running topic_original saved
Phase 1a + 1b Auditors — Server & Architecture (parallel) run before classify
Step 1a — Server Auditor
Claude Haiku — Filesystem Scan
claude-haiku-4-5-20251001
Scanează /opt/coddexa/ pentru fișiere relevante topic-ului. Verifică porturi active, servicii rulate, dependențe disponibile, constrângeri de disk. Identifică oportunități de refolosire a codului existent.
Output
relevant_existing[ ] ports_in_use[ ] reuse_opportunities[ ] constraints[ ] server_summary
⟳
Step 1b — Architecture Auditor
Claude Sonnet — Vault & Context Scan
claude-sonnet-4-6
Citește vault/20-decisions, vault/30-patterns, active-state toate proiectele, session_index (ultimele 15 sesiuni). Identifică decizii arhitecturale relevante, riscuri de integrare cu ce există, open loops legate de topic.
Output
relevant_decisions[ ] applicable_patterns[ ] integration_risks[ ] related_projects[ ] arch_summary
Ambele rezultate → audit_context string
Injectate în: Classifier prompt · COUNCIL_PROMPT (toate cele 6 modele) · RESEARCH_QUERY_PROMPT_SEAT
Phase 1c Classifier — Difficulty & Prompt Refinement status: classifying · runs with full audit context
Classifier
DeepSeek Chat — Topic Analysis
deepseek-chat
Analyzes the raw topic and determines difficulty level. Reformulates the prompt for maximum clarity and context. Returns structured JSON.
Output
difficulty: easy | medium | hard topic_refined: clarified prompt layer_order computed
Route to Layer
Easy → Layer 1 (1 turn, no research) Medium → Layer 2 (2 turns, Gemini Flash research) Hard → Layer 3 (3 turns, Gemini Pro deep research)
Phase 2 Deep Research — Medium & Hard layers only status: researching
Architecture
6 Parallel Per-Seat Research Chains
Each of the 6 council seats runs its own independent research pipeline. DeepSeek formulates a query tailored to the seat's role and perspective, then Gemini executes deep research on that query. All 6 chains run concurrently via ThreadPoolExecutor.
⟳ Parallel — 6 independent DeepSeek → Gemini chains
Claude
Strategic Architect lens
deepseek → gemini
Gemini
Data Analyst lens
deepseek → gemini
OpenAI
Product Strategist lens
deepseek → gemini
DeepSeek
Risk Auditor lens
deepseek → gemini
Grok
Devil's Advocate lens
deepseek → gemini
Mistral
Execution Specialist lens
deepseek → gemini
Session output
seat_research.claude seat_research.gemini seat_research.openai seat_research.deepseek seat_research.grok seat_research.mistral cost_usd accumulated
Phase 3 Council Turn — 6 Models in Parallel Easy: 1 turn · Medium: up to 2 · Hard: up to 3
⟳ Parallel Execution — All 6 models simultaneously
Claude
Strategic Architect
Arhitectură pe termen lung, calitate implementare, design sistemic.
haiku-4-5 / sonnet-4-6 / opus-4-7
Gemini
Data Analyst
Date, metrici, tendințe de piață, dovezi cantitative.
flash-2.0 / flash-2.5 / pro-2.5
OpenAI
Product Strategist
Experiența utilizatorului, product-market fit, poziționare competitivă.
gpt-4o-mini / gpt-4o / gpt-4o
DeepSeek
Risk Auditor
Edge cases, moduri de eșec, datorii tehnice, riscuri ascunse.
deepseek-chat / reasoner / reasoner
Grok
Devil's Advocate
Ipoteze contestate, perspective contrare, unghiuri oarbe.
grok-3-mini / grok-3 / grok-3
Mistral
Execution Specialist
Implementare practică, resurse necesare, fezabilitate operațională.
small-latest / large-latest / large-latest
Each model returns
verdict: PROCEED | HOLD | MODIFY confidence: 0–100 pros[ ] cons[ ] risks[ ] summary: 1–2 sentences
Phase 4 Orchestrator — Consensus Evaluation & Decision status: orchestrating
Orchestrator
Claude — Synthesis & Consensus Check
claude-sonnet-4-6 (easy / medium)
claude-opus-4-7 (hard)
Evaluates all 6 model outputs. Checks for consensus: ≥3 models agree on same verdict AND average confidence ≥65%. On the final turn of a layer it is forced to produce a PLAN regardless of consensus.
Consensus condition
set(verdicts) == 1 unique value avg confidence ≥ 65% ≥ 3 models responded
PLAN Consensus reached or final turn forced. Orchestrator produces structured execution plan.
RETRY No consensus yet, turns remain in current layer. Prompt is refined, models run again.
ESCALATE No consensus after all turns in layer. Moves up to the next harder layer with better models.
Output
Execution Plan
Structured markdown plan with: Verdict (PROCEED/HOLD/MODIFY), rationale, key points, recommended path, risks. Session marked completed. Available in history.
Session state
status: completed execution_plan (markdown) completed_at timestamp cost_usd accumulated seat_research[6 seats] all turns + orchestrator preserved memory → vault/10-sessions/
Layer Comparison
Layer Trigger Max Turns Research Council Models Orchestrator
Easy Simple questions, low risk, clear answer 1 None Haiku · Flash-2.0 · GPT-4o-mini · DeepSeek-chat · Grok-3-mini · Mistral-small Claude Sonnet 4.6
Medium Trade-offs present, context required 2 6× DeepSeek-chat → Gemini Flash 2.5
(per-seat parallel)		 Sonnet · Flash-2.5 · GPT-4o · DeepSeek-reasoner · Grok-3 · Mistral-large Claude Sonnet 4.6
Hard Complex, ambiguous, high-stakes decisions 3 6× DeepSeek-reasoner → Gemini Pro 2.5
(per-seat parallel)		 Opus · Pro-2.5 · GPT-4o · DeepSeek-reasoner · Grok-3 · Mistral-large Claude Opus 4.7
Escalation & Retry Logic
Orchestrator Action Condition What happens
PLAN Consensus OR final turn of a layer Produces structured markdown execution plan. Session marked completed.
RETRY No consensus, turns remaining in current layer Orchestrator refines the prompt, next turn runs in same layer with updated context.
ESCALATE No consensus after all turns in layer Escalation recorded. Moves to next layer (Easy→Medium→Hard) with stronger models and deeper research.
FAILSAFE All 3 layers exhausted without PLAN Fallback plan generated from dominant verdict across all model responses.
Anthropic / Claude
Google / Gemini
OpenAI / GPT
DeepSeek
xAI / Grok
Mistral
Loop / retry boundary