Thinking Parliament

Overview

Orchestrates multi-agent deliberation for problems that exceed single-perspective capacity. Combines CDO cognitive design patterns with axiom-style multi-turn thinking and skill catalog discovery.

Quick Decision Tree

Problem presented? │ ├─→ Confidence < 40%? │ └─→ FULL PARLIAMENT - Multi-agent deliberation (see below) │ ├─→ Confidence 40-60%? │ └─→ RESONANCE MODE - Skill discovery + power combos │ ├─→ Confidence 60-85%? │ └─→ STANDARD - Single-agent with skill hints │ └─→ Confidence > 85%? └─→ DIRECT EXECUTION - No deliberation needed

Confidence Assessment

Before routing, assess problem confidence:

CONFIDENCE FACTORS: ├─ Domain familiarity (have I solved this before?) ├─ Information completeness (are requirements clear?) ├─ Risk level (what happens if wrong?) ├─ Reversibility (can we iterate?) └─ Stakeholder alignment (shared understanding?)

Score: Sum factors 0-100%

Full Parliament Mode (<40% confidence)

When: Strategic decisions, system architecture, multi-stakeholder impact.

Phase 0: Context Gathering (lev-get pre-step)

Before dispatching agents, gather context with semantic search:

Extract keywords from problem statement

KEYWORDS=$(echo "$PROBLEM" | grep -oE '\b[A-Za-z]{4,}\b' | sort -u | head -10 | tr '\n' ' ')

Search across all relevant indexes

lev get "$KEYWORDS" --indexes codebase,documentation,memory,skills > "$SESSION_DIR/00-context.md"

Include prior art and related decisions

lev get "decision $KEYWORDS" --indexes memory,sessions >> "$SESSION_DIR/00-context.md"

Why pre-step: Agents deliberate with shared context, not in a vacuum. Memory index surfaces past decisions. Skills index finds relevant frameworks.

Fallback (if lev get slow/unavailable):

Timeout after 30 seconds, proceed with reduced context

timeout 30s lev get "$KEYWORDS" --indexes codebase,memory > "$SESSION_DIR/00-context.md" 2>/dev/null || { echo "# Context gathering timed out - proceeding with problem statement only" > "$SESSION_DIR/00-context.md" echo "Note: Parliament will deliberate with less context. Results may need manual validation." >> "$SESSION_DIR/00-context.md" }

Phase 1: Workspace Setup

SESSION_DIR="./tmp/parliament-$(date +%Y%m%d-%H%M%S)" mkdir -p "$SESSION_DIR" echo "Parliament session: $SESSION_DIR"

Phase 2: Multi-Modal Agent Dispatch

Deploy 5 agents with distinct models for genuine perspective diversity:

Agent Role Perspective Model Rationale

A1 Advocate Strongest case FOR openrouter/openai/gpt-5.2-pro

Frontier reasoning, "think hard" support

A2 Critic Strongest case AGAINST openrouter/google/gemini-3-flash-preview

1M context, different training data

A3 Systems Second-order effects openrouter/x-ai/grok-4-fast

2M context (!), fast inference

A4 Pragmatist Implementation reality claude-opus-4-5

Deep reasoning for edge cases

A5 Wild Card Unconsidered alternatives openrouter/deepseek/deepseek-v3.2

Value king, unconventional reasoning

Dispatch Pattern (parallel via AI SDK):

All agents dispatch simultaneously with shared context

CONTEXT=$(cat "$SESSION_DIR/00-context.md")

Using lev exec with different providers

lev exec "Role: Advocate. $PROBLEM\n\nContext:\n$CONTEXT" --model=openai/gpt-5.2-pro --adapter=ai-sdk > "$SESSION_DIR/advocate.md" & lev exec "Role: Critic. $PROBLEM\n\nContext:\n$CONTEXT" --model=google/gemini-3-flash-preview --adapter=ai-sdk > "$SESSION_DIR/critic.md" & lev exec "Role: Systems. $PROBLEM\n\nContext:\n$CONTEXT" --model=x-ai/grok-4-fast --adapter=ai-sdk > "$SESSION_DIR/systems.md" & lev exec "Role: Pragmatist. $PROBLEM\n\nContext:\n$CONTEXT" --model=claude-opus-4-5-20251101 --adapter=claude-agent-sdk > "$SESSION_DIR/pragmatist.md" & lev exec "Role: Wild Card. $PROBLEM\n\nContext:\n$CONTEXT" --model=deepseek/deepseek-v3.2 --adapter=ai-sdk > "$SESSION_DIR/wildcard.md" &

wait # All 5 run in parallel

Why multi-modal: Synthetic diversity from one model creates correlated blind spots. Different training data = genuinely different failure modes. Claude catches what Gemini misses and vice versa.

Phase 3: Devil's Advocate Trigger

At >70% agreement → Trigger devil's advocate:

IF all agents agree on direction: └─→ Spawn contrarian agent └─→ Must argue opposite with full conviction └─→ Document in $SESSION_DIR/devils-advocate.md

Phase 4: Synthesis

Read all agent artifacts, produce:

• Common ground (where all agree)

• Genuine tensions (where experts differ)

• Decision framework (when to use which approach)

• Confidence calibration (typically lower than initial)

Resonance Mode (40-60% confidence)

When: Problem is scoped but approach unclear.

Skill Discovery

Search 568-skill catalog

node ~/lev/workshop/poc/lookup/cli.js find "<problem keywords>"

Browse by domain

node ~/lev/workshop/poc/lookup/cli.js list --tag=strategy node ~/lev/workshop/poc/lookup/cli.js list --tag=systems

Power Combo Discovery

Skills have complementsWell metadata. Chain them:

Example combo: Strategic Decision

decision-matrix + rice-scoring + reversibility-check

Example combo: Systems Analysis

systems-thinking + first-principles + cognitive-parliament

See: references/power-combos.md

CDO 5-Stage Cycle

All parliament work follows:

┌─────────────────────────────────────────────┐ │ 1. PLAN │ Define problem, scope, success │ │ 2. THINK │ Multi-agent exploration │ │ 3. EXECUTE │ Synthesize findings │ │ 4. REVIEW │ Validate against criteria │ │ 5. LEARN │ Update patterns, calibrate │ └─────────────────────────────────────────────┘

See: references/cdo-patterns.md

Axiom Workflow (Anti-Groupthink)

Multi-turn thinking with disk-based artifacts:

./tmp/parliament-{timestamp}/ ├── 00-input.md # Original problem ├── 01-perspectives.md # Initial agent views ├── 02-debate.md # FOR vs AGAINST ├── 03-synthesis.md # Integration ├── 04-decision.md # Framework └── FINAL.md # Actionable output

Why disk-based: Forces deliberation, prevents premature consensus, creates audit trail.

See: references/axiom-workflow.md

References

• references/cdo-patterns.md - CDO 5-stage cycle, merge strategies

• references/power-combos.md - Skill graph, complementsWell network

• references/axiom-workflow.md - Disk-based anti-groupthink

• references/parliamentary-protocol.md - Multi-agent deliberation rules

• references/confidence-routing.md - Routing thresholds and calibration

Quick Start Examples

Strategic Decision

User: "Should we migrate to microservices?"

1. Assess confidence → 35% (high impact, unclear path)
2. Route → Full Parliament
3. Dispatch: Advocate, Critic, Systems, Pragmatist, Wild Card
4. Synthesize: Boundary conditions, phased approach
5. Output: Decision framework with when-to-migrate criteria

Skill Discovery

User: "Help me evaluate these 5 vendors"

1. Assess confidence → 55% (scoped, approach unclear)
2. Route → Resonance Mode
3. Search: "evaluation comparison decision"
4. Find: decision-matrix, weighted-scoring, negotiation-leverage
5. Chain: decision-matrix → weighted-scoring → final-recommendation

Hidden Nugget Extraction

User: "What patterns am I missing in this design?"

1. Assess confidence → 45% (unknown unknowns)
2. Route → Resonance Mode + Parliament elements
3. Random skill sampling: lookup random --limit=10
4. Pattern matching against design
5. Surface unconsidered alternatives

Load references as needed based on problem complexity.

Technique Map

• Role definition - Clarifies operating scope and prevents ambiguous execution.

• Context enrichment - Captures required inputs before actions.

• Output structuring - Standardizes deliverables for consistent reuse.

• Step-by-step workflow - Reduces errors by making execution order explicit.

• Edge-case handling - Documents safe fallbacks when assumptions fail.

Technique Notes

These techniques improve reliability by making intent, inputs, outputs, and fallback paths explicit. Keep this section concise and additive so existing domain guidance remains primary.

Prompt Architect Overlay

Role Definition

You are the prompt-architect-enhanced specialist for lev-orch-thinking-parliament, responsible for deterministic execution of this skill's guidance while preserving existing workflow and constraints.

Input Contract

• Required: clear user intent and relevant context for this skill.

• Preferred: repository/project constraints, existing artifacts, and success criteria.

• If context is missing, ask focused questions before proceeding.

Output Contract

• Provide structured, actionable outputs aligned to this skill's existing format.

• Include assumptions and next steps when appropriate.

• Preserve compatibility with existing sections and related skills.

Edge Cases & Fallbacks

• If prerequisites are missing, provide a minimal safe path and request missing inputs.

• If scope is ambiguous, narrow to the highest-confidence sub-task.

• If a requested action conflicts with existing constraints, explain and offer compliant alternatives.