Parliament

Parliament — Multi-Model Collaboration

Spawn parallel sub-agents across different LLM models, collect their outputs, and synthesize a unified report. Like a parliament where each model is a delegate with its own perspective.

When to Use

• Code review from multiple model perspectives
• Document/paper review with diverse viewpoints
• Decision-making that benefits from consensus
• Quality assurance through model disagreement detection
• Any task where "second opinions" from different models add value

Inputs

Available Models

⚠️ Model availability depends on configured API keys. Parliament gracefully handles unavailable models (see Error Handling).

⚠️ GLM-5 model ID is glm/claude-opus-4-6 — the glm/ prefix routes through MMKG, NOT Anthropic direct.

Default parliament: all 4 models. Quick mode: opus46 + one other.

Execution Protocol

When the user asks for a parliament review, follow these steps exactly:

Step 1: Prepare the Delegate Prompt

Read the file (if a path is given), then construct the delegate prompt:

You are one delegate in a multi-model parliament review. Other models are reviewing the same content independently.

Your role: Give your honest, thorough, independent analysis. Do not hedge or be diplomatic — be direct.

Task: {user's prompt}

Content to review:
{file content or inline content}

Structure your response as:
1. **Key Observations** — what stands out immediately
2. **Issues Found** — bugs, risks, gaps, errors (with severity: high/medium/low)
3. **Strengths** — what's done well
4. **Recommendations** — specific, actionable improvements
5. **Confidence** — your overall confidence level (high/medium/low) with reasoning

Step 2: Spawn All Delegates in Parallel

Use sessions_spawn for each model. Spawn ALL at once (do not wait sequentially):

sessions_spawn(
  task=<delegate prompt>,
  model="anthropic/claude-opus-4-6",
  label="parliament-opus46",
  mode="run",
  runTimeoutSeconds=120
)

Labels: parliament-opus46, parliament-gpt53, parliament-gemini, parliament-glm5.

Step 3: Collect Results

Sub-agents auto-announce when done. As each result arrives, note it. Proceed to synthesis when:

• All delegates have returned, OR
• Timeout reached and at least 2 delegates responded

Step 4: Synthesize (Host Agent Does This)

The host agent (you) produces the final Parliament Report by comparing all delegate outputs:

How to identify consensus: Points raised by 3+ delegates with similar conclusions.

How to identify divergence: Points where delegates reach different conclusions or assign different severity.

How to identify unique insights: Points raised by only 1 delegate that others missed.

How to derive recommendation: Weight by consensus strength. If 3/4 agree, follow the majority. If split 2/2, present both positions and recommend the more conservative option.

Step 5: Deliver Report

Output the report in this format:

# 🏛️ Parliament Report

**Task:** {summary}
**Models:** {list of models that responded}
**Date:** {timestamp}
**Quorum:** {N}/{total} delegates responded

---

## Delegate Responses

### 🟣 Opus 4.6
{response}

### 🟢 GPT-5.3 Codex
{response}

### 🔵 Gemini 3.1 Pro
{response}

### 🟡 GLM-5
{response}

---

## Synthesis

### ✅ Consensus (3+ delegates agree)
- {point 1}
- {point 2}

### ⚠️ Divergence (delegates disagree)
- {topic}: {model A} says X, {model B} says Y
- Recommendation: {which position and why}

### 💡 Unique Insights (only one delegate caught this)
- {model}: {insight}

### 📊 Confidence Matrix

| Model | Confidence | Top Concern |
|-------|-----------|-------------|
| Opus 4.6 | high/med/low | ... |
| GPT-5.3 | high/med/low | ... |
| Gemini | high/med/low | ... |
| GLM-5 | high/med/low | ... |

### 📋 Final Recommendation
{synthesized recommendation based on weighted consensus}

Error Handling

Minimum quorum for synthesis: 2 delegates.

Modes

Review Mode (default)

Independent analysis from each model. Best for code review, document review, architecture decisions.

Quick Mode

Only 2 models (opus46 + gemini). Faster, still catches most issues. Use when time matters more than breadth.

Debate Mode

Frame the task as a debate: "Take a strong position and argue for it." Synthesis highlights strongest arguments from each side. Best for decision-making with tradeoffs.

Permissions

• sessions_spawn — spawn sub-agent sessions with different models
• read — read files for review (when file paths provided)
• write — save the parliament report (optional)

Tips

• Parliament works best with specific, concrete tasks (not vague "review everything")
• For large files (>500 lines), split into sections and run parliament on each
• Pay special attention to divergence points — that's where the real value is
• If one model consistently disagrees, investigate why — it may have caught something others missed
• Check model availability first (/status) if you're unsure which API keys are configured