AgentSkillsCN

consensus-coder

多模型共识编码工作流。当您需要同时调用多种大语言模型（Opus、Gemini、Codex），协同诊断编码问题、提出解决方案、探讨备选方案，并在实施前达成共识时，可选用此工作流。特别适合处理复杂的架构决策、重构策略，或是那些需要多元视角来提升解决方案质量的棘手算法问题。

SKILL.md
--- frontmatter
name: consensus-coder
description: Multi-model consensus coding workflow. Use when you need multiple LLMs (Opus, Gemini, Codex) to collaboratively diagnose a coding problem, propose solutions, debate alternatives, and reach consensus before implementation. Ideal for complex architectural decisions, refactoring strategies, or tricky algorithmic problems where diverse perspectives improve solution quality.

Consensus Coder

A collaborative multi-model workflow that achieves consensus on coding solutions through structured debate and voting.

How It Works

Phase 1: Diagnosis & Proposal (Opus)

  1. •Opus 4.5 analyzes the problem and proposes 3 distinct approaches
  2. •Each proposal includes: rationale, trade-offs, estimated complexity

Phase 2: Review & Vote (Gemini 2.5 + Codex)

  1. •Gemini 2.5 reviews all 3 proposals and casts a vote with reasoning
  2. •Codex reviews all 3 proposals and casts a vote with reasoning
  3. •Opus votes on its own proposals or agrees with a reviewer's choice

Phase 3: Consensus or Iteration

  • •If unanimous: Opus creates implementation plan → Auggie implements
  • •If split: The dissenting model(s) propose alternative solutions → revote
  • •Max 5 iterations: If no consensus after 5 rounds, escalate to you for human decision

Phase 4: Implementation (Auggie)

Once consensus achieved:

  1. •Opus creates detailed implementation plan
  2. •Auggie executes the plan with your Codebake tasks

Usage

Start a consensus session:

code
problem: "Your coding problem or refactoring challenge"
context: "Additional context (framework, constraints, goals)"

The workflow will:

  1. •Show each model's proposals and reasoning
  2. •Display vote tallies each round
  3. •Auto-iterate on disagreement
  4. •Ask for your decision if stuck after 5 rounds
  5. •Implement the winning solution

Voting Format

Each model votes using this structure:

code
**Vote:** [Proposal A | Proposal B | Proposal C]
**Confidence:** [High | Medium | Low]
**Reasoning:** [Why this proposal wins]
**Concerns:** [Any reservations or caveats]

Iteration Rules

  • •Unanimous (3-0): Move to implementation
  • •2-1 split: Dissenting model proposes alternative → revote
  • •No consensus after 5 rounds: Escalate to human decision (you decide)

When dissenting, the minority model must articulate:

  • •Specific concern with the winning proposal
  • •What their alternative addresses that others miss

Implementation

Auggie executes with:

  • •Opus's detailed plan as the spec
  • •Your Codebake workspace for task tracking
  • •Automatic error feedback loops

Token Cost

Expect ~8-12k tokens per consensus cycle (3 proposal reviews + voting + iteration logic). Multiple iterations can be expensive, so this works best for high-impact decisions.

See references/ for voting logic, prompt templates, and consensus rules.