Consensus Workflow Instructions

Consensus Workflow Instructions

CRITICAL: These are your operational instructions. Read fully before any complex task.

What This Is

You have access to a consensus server at http://localhost:3000 that gets independent reviews from ChatGPT and Grok before you proceed. This catches errors and edge cases.

Mandatory Rules

1. Call the consensus server for EVERY iteration - no exceptions
2. Display ALL consensus details to the user EVERY time (scores, feedback, status)
3. Score must EXCEED 0.95 (0.95 exactly is not enough)
4. Maximum 10 rounds per task - if reached, stop and ask user
5. Never skip consensus for complex tasks without user permission

When You Must Seek Consensus

• Plans with 3+ steps
• Code over 50 lines
• Architecture decisions
• Complex algorithms
• Security-sensitive code
• Anything user explicitly requests

When You May Skip Consensus

• Simple one-liner fixes
• Typo corrections
• Adding comments
• Trivial changes

Server Endpoints

Health check:

curl http://localhost:3000/health

Consensus request:

curl -s -X POST "http://localhost:3000/consensus" \
  -H "Content-Type: application/json" \
  -d '{
    "session_id": "unique-id-per-task",
    "task": "Description of what you are building",
    "phase": "plan",
    "current_output": "Your plan or code as a string",
    "system_prompt": "Context for reviewers",
    "threshold": 0.95
  }'

Override (when user approves):

curl -s -X POST "http://localhost:3000/consensus/override" \
  -H "Content-Type: application/json" \
  -d '{"session_id": "task-123", "action": "proceed"}'

Request Fields

• session_id (required): Unique ID for this task, reuse across iterations of same phase
• task (required): What you are building
• phase (required): "plan", "code", or "verify"
• current_output (required): Your plan or code to review
• system_prompt (optional): Context for reviewers
• threshold (optional): Default 0.95

Response Fields

• status: "approved", "iterating", "stalled", or "error"
• approved: boolean
• avg_score: the average score from all reviewers
• round: current round number
• max_rounds: maximum rounds (10)
• agent_results: array with each agent's score and feedback
• suggestions: array of improvement suggestions

Your Workflow

Phase 1: Plan

1. Receive task from user
2. Generate a structured plan (numbered steps)
3. Call POST /consensus with phase: "plan"
4. IMMEDIATELY display full consensus details to user:
   • Round number
   • Each agent's name, score, and feedback
   • Average score
   • Status (approved or not)
5. If not approved:
   • List specific changes you are making based on feedback
   • Revise plan
   • Call consensus again (same session_id) - THIS IS MANDATORY
   • Display results to user again
6. If approved: Proceed to code phase

Phase 2: Code

1. Write the code based on approved plan
2. Call POST /consensus with phase: "code"
3. Display full consensus details to user
4. Iterate until approved or stalle