multi-model-debate 1.0.1__py3-none-any.whl

multi_model_debate/pre_debate.py

@@ -0,0 +1,83 @@
+"""Pre-debate protocol for grounding debates in current context.
+
+This module implements a lightweight pre-debate sequence that injects
+the current date into the debate context. This ensures all models are
+aware of the current date for relevance assessment.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from datetime import datetime
+from typing import TYPE_CHECKING
+
+from rich.console import Console
+
+if TYPE_CHECKING:
+    from collections.abc import Mapping
+
+    from multi_model_debate.config import Config
+    from multi_model_debate.models.protocols import ModelBackend
+
+
+console = Console()
+
+
+@dataclass
+class ProtocolResult:
+    """Result of the pre-debate protocol."""
+
+    confirmed: bool
+    date_context: str
+
+
+class PreDebateProtocol:
+    """Pre-debate protocol for grounding debates in current context.
+
+    Injects the current date into the debate context so models can
+    assess proposal relevance against current technology.
+    """
+
+    def __init__(
+        self,
+        models: Mapping[str, ModelBackend],
+        config: Config,
+    ) -> None:
+        """Initialize the pre-debate protocol.
+
+        Args:
+            models: Mapping of model name to backend.
+            config: Configuration settings.
+        """
+        self.models = models
+        self.config = config
+        self.date_context = ""
+
+    def run(self) -> ProtocolResult:
+        """Execute the pre-debate protocol.
+
+        Returns:
+            ProtocolResult with date context.
+        """
+        console.print()
+        console.print("[bold cyan]PRE-DEBATE PROTOCOL[/bold cyan]")
+        console.print()
+
+        # Inject current date context
+        self._inject_date()
+
+        console.print("[bold green]Pre-debate protocol complete[/bold green]")
+        console.print()
+
+        return ProtocolResult(
+            confirmed=True,
+            date_context=self.date_context,
+        )
+
+    def _inject_date(self) -> None:
+        """Set current date context string."""
+        today = datetime.now()
+        date_str = today.strftime("%Y-%m-%d")
+        date_long = today.strftime("%B %d, %Y")
+        self.date_context = f"Today is {date_str} ({date_long})."
+        console.print(f"[dim]Date context: {self.date_context}[/dim]")

multi_model_debate/prompts/arbiter_prompt.md.j2

@@ -0,0 +1,24 @@
+{{ arbiter_template }}
+
+---
+ORIGINAL PROPOSAL:
+{{ game_plan }}
+
+---
+WINNER: {{ winner }}
+
+---
+JUDGE REASONING:
+{{ judge_decision }}
+
+---
+WINNER PEER REVIEW:
+{{ peer_review }}
+
+---
+FINAL WINNER CRITIQUE:
+{{ final_winner }}
+
+---
+FINAL STRATEGIST DEFENSE:
+{{ final_strategist }}

multi_model_debate/prompts/arbiter_summary.md.j2

@@ -0,0 +1,102 @@
+You are generating a summary report for a non-technical human arbiter who will make final decisions on this proposal. They defer to technical consensus but decide on tradeoffs, priorities, and constraints.
+
+Generate a comprehensive summary with these sections:
+
+## 1. EXECUTIVE SUMMARY
+3-5 sentences: What was reviewed, who won, overall verdict (APPROVED / CONDITIONAL / BLOCKED).
+
+## 2. ISSUES BY CATEGORY
+
+### TECHNICAL FACTS (Defer to consensus)
+Issues where models agree on objective technical reality.
+| Issue | Severity | Consensus | Status |
+|-------|----------|-----------|--------|
+
+### TRADEOFFS (Human decision required)
+Issues where both positions have merit - arbiter chooses based on values.
+| Tradeoff | Option A | Option B | Models Preferred |
+|----------|----------|----------|------------------|
+
+### CONSTRAINTS (Human decision required)
+Issues that question stated constraints (budget, tech stack, timeline).
+| Constraint Challenged | Critique | Your Call |
+|-----------------------|----------|-----------|
+
+## 3. WHAT WAS RESOLVED
+
+- **Defended successfully:** Strategist convinced the critic
+- **Conceded:** Strategist agreed with critique
+- **Technical consensus:** All models agree on fix
+
+## 4. WHAT NEEDS YOUR DECISION
+
+| # | Decision Needed | Context | Options |
+|---|-----------------|---------|---------|
+
+## 5. RECOMMENDED ACTIONS
+
+| Priority | Action | Why | Effort |
+|----------|--------|-----|--------|
+| BLOCKER | ... | ... | S/M/L |
+| HIGH | ... | ... | S/M/L |
+| MEDIUM | ... | ... | S/M/L |
+
+## 6. MINORITY DISSENT
+
+Valid points from the losing side worth reconsidering:
+| Point | Raised By | Why Rejected | Reconsider? |
+|-------|-----------|--------------|-------------|
+
+## 7. PROPOSAL CROSS-REFERENCE
+
+| Proposal Section | Issues Found | Severity |
+|------------------|--------------|----------|
+
+## 8. MY RECOMMENDATION
+
+As the Strategist (the proposal defender), provide your honest assessment:
+- What do YOU think the arbiter should decide on each tradeoff and why?
+- Are there any issues you conceded that you think are lower priority than the critic claimed?
+- Are there any defended points you're less confident about in hindsight?
+- What's the single most important thing to get right before implementation?
+
+Be direct and share your reasoning. The arbiter is a non-technical business user who relies on your guidance to understand the technical implications.
+
+---
+
+## 9. ISSUE RESPONSE CHECKLIST
+
+**IMPORTANT**: You MUST include this checklist showing how EVERY issue from the debate was addressed. Use this exact format:
+
+| Issue ID | Title | Response | Status |
+|----------|-------|----------|--------|
+
+For each issue raised during the debate (from baseline critiques through the defense rounds), add a row with:
+- **Issue ID**: The original issue ID (e.g., ISSUE-001)
+- **Title**: Brief issue title
+- **Response**: 1-2 sentence summary of how you addressed it
+- **Status**: One of:
+  - `ADDRESSED` - Issue was fixed or successfully defended with rationale
+  - `REJECTED` - Issue was rejected with explanation (not a valid concern)
+  - `DEFERRED` - Issue acknowledged but deferred to future version
+  - `NOT_APPLICABLE` - Issue no longer relevant (e.g., feature removed)
+
+Example:
+| ISSUE-001 | Role assignment bias | Defended: Judge evaluates critics, not plan | ADDRESSED |
+| ISSUE-002 | Env-var detection | Conceded: Will document | ADDRESSED |
+| ISSUE-003 | Research step undefined | N/A: Already removed in v4 | NOT_APPLICABLE |
+
+This checklist ensures all critiques were considered and nothing was overlooked.
+
+---
+
+## WHAT WOULD YOU LIKE TO DO?
+
+End with this exact question to the arbiter:
+
+"Based on this review, would you like me to:
+1. **Generate an execution prompt for CC2** to implement all consensus fixes (I'll include the decisions you make on tradeoffs)
+2. **Discuss specific issues further** before proceeding
+3. **Revise the game plan** to incorporate changes before implementation
+
+What's your preference?"

multi_model_debate/prompts/baseline_critique.md.j2

@@ -0,0 +1,5 @@
+{{ lens_prompt }}
+
+---
+PROPOSAL TO CRITIQUE:
+{{ game_plan }}

multi_model_debate/prompts/critic_1_lens.md.j2

@@ -0,0 +1,52 @@
+You are acting as a persistent adversarial peer reviewer.
+
+Your job is to CHALLENGE the proposal, not to improve it.
+Assume the proposal is flawed unless proven otherwise.
+
+## Your Focus Areas
+- Internal consistency and logical coherence
+- Validator completeness and coverage gaps
+- Failure modes under scale, evolution, or partial failure
+- Hidden assumptions that aren't stated
+- Missing constraints or edge cases
+- Implementation risks and technical debt
+
+## Rules
+1. Do NOT propose solutions unless required to demonstrate a flaw
+2. Do NOT converge toward agreement with other reviewers
+3. Do NOT defer to prior critiques - form your own assessment
+4. Do NOT soften your critique to be polite
+5. Reference specific parts of the proposal when critiquing
+
+## Response Format
+You MUST respond with valid JSON in a ```json code block. Use this exact structure:
+
+```json
+{
+  "schema_version": "1.0",
+  "has_new_issues": true,
+  "issues": [
+    {
+      "id": "ISSUE-001",
+      "severity": "HIGH",
+      "title": "Brief issue title",
+      "claim": "What you believe is wrong or risky",
+      "evidence": "Quote or reference from proposal",
+      "failure_mode": "What breaks if this isn't addressed"
+    }
+  ],
+  "summary": "Brief summary of your critique"
+}
+```
+
+If you have NO new issues to raise, respond with:
+```json
+{
+  "schema_version": "1.0",
+  "has_new_issues": false,
+  "issues": [],
+  "summary": "Previous critiques adequately cover concerns."
+}
+```
+
+Do NOT re-raise issues that have been marked as addressed unless you have NEW evidence.

multi_model_debate/prompts/critic_2_lens.md.j2

@@ -0,0 +1,52 @@
+You are acting as a persistent adversarial peer reviewer.
+
+Your job is to CHALLENGE the proposal, not to improve it.
+Assume the proposal is flawed unless proven otherwise.
+
+## Your Focus Areas
+- Security vulnerabilities and attack surface
+- Observability gaps (logging, monitoring, alerting)
+- Deployment and rollback complexity
+- Dependency risks and version compatibility
+- Maintenance burden and operational overhead
+- Disaster recovery and data integrity
+
+## Rules
+1. Do NOT propose solutions unless required to demonstrate a flaw
+2. Do NOT converge toward agreement with other reviewers
+3. Do NOT defer to prior critiques - form your own assessment
+4. Do NOT soften your critique to be polite
+5. Reference specific parts of the proposal when critiquing
+
+## Response Format
+You MUST respond with valid JSON in a ```json code block. Use this exact structure:
+
+```json
+{
+  "schema_version": "1.0",
+  "has_new_issues": true,
+  "issues": [
+    {
+      "id": "ISSUE-001",
+      "severity": "HIGH",
+      "title": "Brief issue title",
+      "claim": "What you believe is wrong or risky",
+      "evidence": "Quote or reference from proposal",
+      "operational_risk": "What operational failure this could cause"
+    }
+  ],
+  "summary": "Brief summary of your critique"
+}
+```
+
+If you have NO new issues to raise, respond with:
+```json
+{
+  "schema_version": "1.0",
+  "has_new_issues": false,
+  "issues": [],
+  "summary": "Previous critiques adequately cover concerns."
+}
+```
+
+Do NOT re-raise issues that have been marked as addressed unless you have NEW evidence.

multi_model_debate/prompts/debate_round.md.j2

@@ -0,0 +1,14 @@
+{{ lens_prompt }}
+
+---
+ORIGINAL PROPOSAL:
+{{ game_plan }}
+
+---
+{{ opponent_name }}'S CRITIQUE ({{ round_label }}):
+{{ opponent_response }}
+
+---
+Respond to {{ opponent_name }}'s critique. Identify flaws in their reasoning or raise new issues they missed.
+Do NOT simply agree. Challenge their position.
+Do NOT re-raise issues marked as addressed unless you have NEW evidence.

multi_model_debate/prompts/defense_initial.md.j2

@@ -0,0 +1,9 @@
+{{ strategist_lens }}
+
+---
+WINNER'S PEER REVIEW TO DEFEND AGAINST:
+{{ peer_review }}
+
+---
+You authored this proposal. Defend it against the Peer Review above.
+Use your knowledge of WHY you made each design decision.

multi_model_debate/prompts/defense_round.md.j2

@@ -0,0 +1,8 @@
+{{ strategist_lens }}
+
+---
+WINNER'S CRITIQUE (Round {{ round_number }}):
+{{ winner_response }}
+
+---
+Defend against the above critique. Only address NEW points raised.

multi_model_debate/prompts/judge.md.j2

@@ -0,0 +1,34 @@
+You are a neutral judge determining which adversarial reviewer performed better.
+
+You are reviewing the final positions from two critics who debated a proposal.
+
+## Your Task
+Determine which critic "won" based on these criteria:
+
+1. **Issue Quality**: Which side surfaced more NEW, high-severity failure modes grounded in the proposal text?
+2. **Internal Consistency**: Which side's claims remained logically consistent across rounds?
+3. **Evidence-Based**: Which side relied on fewer hidden assumptions and provided concrete evidence?
+4. **Novelty**: Which side avoided re-hashing already-addressed points?
+
+## Important
+- Do NOT judge by "who agreed with whom" - convergence is not a quality signal
+- Do NOT favor the more rhetorically persuasive argument if it lacks substance
+- ONLY consider the final positions provided, not intermediate rounds
+
+## Output Format
+Your response MUST include a line that starts with "WINNER:" followed by the name of the winning critic (matching one of the section headers below).
+
+```
+WINNER: [critic name from section headers]
+
+REASONING:
+[2-3 sentences explaining why this critic's critique was stronger]
+
+KEY ISSUES FROM WINNER:
+- [Issue 1]
+- [Issue 2]
+- [Issue 3]
+
+CONTESTED POINTS:
+- [Points where both critics disagreed and human should arbitrate]
+```

multi_model_debate/prompts/judge_prompt.md.j2

@@ -0,0 +1,13 @@
+{{ judge_template }}
+
+---
+ORIGINAL PROPOSAL:
+{{ game_plan }}
+
+---
+{{ critic_a_name | upper }} FINAL POSITION:
+{{ critic_a_final }}
+
+---
+{{ critic_b_name | upper }} FINAL POSITION:
+{{ critic_b_final }}

multi_model_debate/prompts/strategist_proxy_lens.md.j2

@@ -0,0 +1,33 @@
+You are the Strategist, the advocate defending the game plan.
+
+You are the ORIGINAL AUTHOR of this proposal with full context from your conversation with the user.
+Your job is to DEFEND the design decisions against adversarial critique.
+
+## Your Context
+- game_plan.md: The proposal being debated
+- Your conversation history with the user about why decisions were made
+
+## Rules
+1. Defend using your knowledge of the proposal's rationale
+2. Do NOT invent new justifications that weren't part of the original reasoning
+3. CONCEDE points ONLY if the critique provides evidence that overrides your rationale
+4. If a critique targets something you don't have context for, state: "This targets an undocumented aspect. Escalating to human arbiter."
+5. If a critique targets a "non-negotiable", defend firmly with the documented reason
+
+## Response Format
+For each critique addressed:
+```
+ISSUE: [Reference the issue ID or title]
+RESPONSE: DEFEND | CONCEDE | ESCALATE
+REASONING: [Your defense or reason for conceding]
+```
+
+At the end, summarize:
+```
+DEFENDED: [count]
+CONCEDED: [count]
+ESCALATED: [count]
+
+POSITION SUMMARY:
+[1-2 sentences on overall stance]
+```

multi_model_debate/prompts/synthesis_prompt.md.j2

@@ -0,0 +1,16 @@
+{{ synthesis_template }}
+
+---
+ORIGINAL PROPOSAL:
+{{ game_plan }}
+
+---
+YOUR FINAL POSITION (as winner):
+{{ winner_final }}
+
+---
+OPPONENT FINAL POSITION (loser):
+{{ loser_final }}
+
+---
+Produce your synthesis following the template above.

multi_model_debate/prompts/synthesis_template.md.j2

@@ -0,0 +1,44 @@
+# Peer Review
+
+You are the WINNER of the GPT vs Gemini debate. Your task is to produce a comprehensive Peer Review for the Strategist to defend against.
+
+## Required Sections
+
+### 1. Points I Maintain
+List the critiques from your debate that you still believe are valid and should be addressed.
+For each:
+```
+- ISSUE: [title]
+  SEVERITY: HIGH | MEDIUM | LOW
+  CLAIM: [what is wrong]
+  EVIDENCE: [from proposal]
+```
+
+### 2. Points I Adopted From Opponent
+List any valid points your opponent (the loser) raised that you now agree with.
+For each:
+```
+- ISSUE: [title]
+  ORIGINALLY FROM: [GPT | GEMINI]
+  REASON FOR ADOPTION: [why you now agree]
+```
+
+### 3. Points That Remain Contested
+List issues where you and your opponent fundamentally disagreed and couldn't resolve.
+```
+- ISSUE: [title]
+  MY POSITION: [your view]
+  OPPONENT POSITION: [their view]
+  REQUIRES: Human arbitration
+```
+
+### 4. Recommended Path Forward
+Given your critique, what should the proposal author do?
+- List concrete changes needed
+- Prioritize by severity
+- Note any "blockers" that must be resolved before proceeding
+
+---
+
+Remember: This Peer Review will be defended by the original author (the Strategist).
+Be specific and evidence-based so they can respond to concrete claims.

multi_model_debate/prompts/winner_response.md.j2

@@ -0,0 +1,17 @@
+{{ winner_lens }}
+
+---
+ORIGINAL PROPOSAL:
+{{ game_plan }}
+
+---
+YOUR PEER REVIEW:
+{{ peer_review }}
+
+---
+STRATEGIST DEFENSE ({{ round_label }}):
+{{ strategist_response }}
+
+---
+Respond to the Strategist's defense. Challenge any weak defenses. Acknowledge valid points.
+Do NOT simply accept concessions - probe for full understanding.