consensus-voting（共识投票）

Consensus Voting Skill

Step 1: Define Voting Parameters

Set up the voting session:

voting_session:
  topic: 'Which database to use for the new service'
  options:
    - PostgreSQL
    - MongoDB
    - DynamoDB
  quorum: 3 # Minimum votes required
  threshold: 0.6 # 60% agreement needed
  weights:
    database-architect: 2.0 # Expert gets 2x weight
    security-architect: 1.0
    devops: 1.5

Step 2: Collect Votes

Gather agent recommendations:

## Vote Collection

### database-architect (weight: 2.0)

- Vote: PostgreSQL
- Rationale: Strong ACID guarantees, mature ecosystem
- Confidence: 0.9

### security-architect (weight: 1.0)

- Vote: PostgreSQL
- Rationale: Better encryption at rest, audit logging
- Confidence: 0.8

### devops (weight: 1.5)

- Vote: DynamoDB
- Rationale: Managed service, auto-scaling
- Confidence: 0.7

Step 3: Calculate Consensus

Apply weighted voting:

PostgreSQL: (2.0 * 0.9) + (1.0 * 0.8) = 2.6
DynamoDB:   (1.5 * 0.7) = 1.05
MongoDB:    0

Total weight: 4.5
PostgreSQL: 2.6 / 4.5 = 57.8%
DynamoDB:   1.05 / 4.5 = 23.3%

Threshold: 60% → No clear consensus

Step 4: Resolve Conflicts

When no consensus is reached:

Strategy 1: Expert Override

• If domain expert has strong opinion (>0.8 confidence), defer to expert

Strategy 2: Discussion Round

• Ask dissenting agents to respond to majority arguments
• Re-vote after discussion

Strategy 3: Escalation

• Present options to user with pros/cons from each agent
• Let user make final decision

Step 5: Document Decision

Record the final decision:

## Decision Record

### Topic

Which database to use for the new service

### Decision

PostgreSQL

### Voting Summary

- PostgreSQL: 57.8% (2 votes)
- DynamoDB: 23.3% (1 vote)
- Consensus: NOT REACHED (below 60% threshold)

### Resolution Method

Expert override - database-architect (domain expert)
had 0.9 confidence in PostgreSQL

### Dissenting Opinion

DevOps preferred DynamoDB for operational simplicity.
Mitigation: Will use managed PostgreSQL (RDS) to
reduce operational burden.

### Decision Date

2026-01-23

</execution_process>

<best_practices>

1. Quorum Required: Don't decide without minimum participation
2. Weight by Expertise: Domain experts get more influence
3. Document Dissent: Record minority opinions for future reference
4. Clear Thresholds: Define what constitutes consensus upfront
5. Escalation Path: Have a process for unresolved conflicts

</best_practices>

The architect wants microservices but the developer prefers monolith.
Resolve this conflict.

Voting Process:

## Voting: Architecture Style

### Votes

- architect: Microservices (weight 1.5, confidence 0.8)
- developer: Monolith (weight 1.0, confidence 0.9)
- devops: Microservices (weight 1.0, confidence 0.6)

### Calculation

Microservices: (1.5 _ 0.8) + (1.0 _ 0.6) = 1.8
Monolith: (1.0 \* 0.9) = 0.9

Microservices: 66.7% → CONSENSUS REACHED

### Decision

Microservices, with modular monolith as migration path

### Dissent Mitigation

Start with modular monolith, extract services incrementally
to address developer's maintainability concerns.

</usage_example>

Dual-Completion Gate Protocol

Prevents premature closure of council or multi-agent tasks by requiring 2+ agents to independently confirm task completion before the session closes.

Problem

A single agent signaling "done" can produce incomplete results -- the agent may have finished its own subtask but the overall task is not complete. The dual-completion gate requires consensus on completion itself.

Protocol

1. Track completion signals per agent: {agent_id, timestamp, signal: "complete"}
2. Require min_completions (default: 2) signals within a window_seconds time window (default: 60s)
3. If only 1 agent signals completion, send a "verification nudge" to remaining agents after nudge_after_seconds (default: 30s)
4. After fallback_timeout (default: 120s) with only 1 completion, accept single-agent completion with a warning: "single_agent_completion" flag

Configuration

completion_gate:
  min_completions: 2 # minimum agents that must signal done
  window_seconds: 60 # time window for completion consensus
  nudge_after_seconds: 30 # send nudge to remaining agents after first completion
  fallback_timeout: 120 # accept single completion after this timeout (with warning)

Pseudocode

completions = []

on_agent_complete(agent_id):
  completions.push({ agent_id, timestamp: now() })

  if completions.length >= min_completions:
    window_start = completions[0].timestamp
    window_end = completions[-1].timestamp
    if (window_end - window_start) <= window_seconds:
      return CLOSE_SESSION(status: "consensus_complete")

  if completions.length == 1:
    schedule_nudge(nudge_after_seconds)
    schedule_fallback(fallback_timeout)

on_nudge_timeout():
  send_to_remaining_agents("A team member has signaled completion. Please confirm if the task is done.")

on_fallback_timeout():
  if completions.length < min_completions:
    return CLOSE_SESSION(status: "single_agent_complete", warning: "single_agent_completion")

Integration with LLM Council

The dual-completion gate is invoked by the llm-council skill before closing a council session:

1. After Stage 3 synthesis, each model is asked: "Is this synthesis complete and accurate?"
2. Models respond with "complete" or "needs_revision"
3. The gate requires min_completions "complete" signals before closing
4. If gate fails (insufficient completions), chairman reviews and decides

Voting Protocols Table Update

Rules

• Always require quorum before deciding
• Weight votes by domain expertise
• Document dissenting opinions for future reference
• Require dual-agent completion consensus before closing council sessions

Related Workflow

This skill has a corresponding workflow for complex multi-agent scenarios:

• Workflow: .claude/workflows/consensus-voting-skill-workflow.md
• When to use workflow: For critical multi-agent decisions requiring Byzantine fault-tolerant consensus with Queen/Worker topology (architectural decisions, security reviews, technology selection)
• When to use skill directly: For simple voting scenarios or when integrating consensus into other workflows

Workflow Integration

This skill enables decision-making in multi-agent orchestration:

Router Decision: .claude/workflows/core/router-decision.md

• Router spawns multiple reviewers, then uses consensus to resolve conflicts
• Planning Orchestration Matrix triggers consensus voting for review phases

Artifact Lifecycle: .claude/workflows/core/skill-lifecycle.md

• Consensus voting determines artifact deprecation decisions
• Multiple maintainers vote on breaking changes

Related Workflows:

• swarm-coordination skill for parallel agent spawning before voting
• Enterprise workflows use consensus for design reviews
• Security reviews in .claude/workflows/enterprise/ require security-architect consensus

Iron Laws

1. NEVER accept a decision without meeting minimum quorum — decisions made without quorum are illegitimate; if quorum is not met, postpone the decision or escalate to human intervention.
2. ALWAYS weight votes by domain expertise — equal weights give a generalist developer the same influence as a domain expert; weight by expertise relevance to the decision domain.
3. NEVER discard dissenting opinions — minority perspectives contain the most important signal about edge cases and risks; document all rationales, including the losing side.
4. ALWAYS require re-vote with deliberation before escalating — a split vote without deliberation wastes the consensus mechanism; agents must share reasoning and vote again before escalating to human.
5. NEVER allow abstentions in critical decisions — abstentions on high-stakes decisions mean agents are avoiding responsibility; all participants must vote on CRITICAL/UNANIMOUS-threshold decisions.

Anti-Patterns

Memory Protocol (MANDATORY)

Before starting:

cat .claude/context/memory/learnings.md

After completing:

• New pattern -> .claude/context/memory/learnings.md
• Issue found -> .claude/context/memory/issues.md
• Decision made -> .claude/context/memory/decisions.md

ASSUME INTERRUPTION: Your context may reset. If it's not in memory, it didn't happen.