Ralph Handoff

Ralph Handoff Protocol

This skill defines how agents communicate handoff requests in single-agent orchestration mode. Only ONE agent is active at a time - the watchdog process monitors agent output and orchestrates handoffs.

Overview

Instead of 3 agents polling simultaneously (consuming tokens continuously), this model:

1. Runs ONE agent at a time
2. Agent outputs handoff phrase when work is complete or another agent is needed
3. Watchdog detects handoff, gracefully stops current agent, starts next agent
4. Context is passed to next agent via state files + handoff message

Handoff Phrase Format

When you need another agent to take over, output this exact format:

HANDOFF:agent_name:base64_context

Where:

• agent_name is one of: pm, developer, qa
• base64_context is a Base64-encoded JSON object with handoff details

Example

HANDOFF:developer:eyJmcm9tIjoicG0iLCJyZWFzb24iOiJ0YXNrX2Fzc2lnbm1lbnQiLCJ0YXNrIjp7ImlkIjoiZmVhdC0wMDEiLCJ0aXRsZSI6IkFkZCB1c2VyIGF1dGgiLCJwcmlvcml0eSI6ImhpZ2gifX0=

Context JSON Structure

The context object should include:

{
  "from": "pm", // Your agent name
  "reason": "task_assignment", // Why handoff is needed
  "timestamp": "2026-01-20T...", // When you're handing off
  "task": {
    // Task-specific details
    "id": "feat-001",
    "title": "Add user authentication",
    "action": "implement", // What the next agent should do
    "notes": "Focus on JWT tokens" // Any additional context
  }
}

Handoff Reasons

Before Handoff: Save State

CRITICAL: Before outputting a handoff phrase, you MUST:

1. Save all state to files:

   • Update prd.json.session with current status
   • Update prd.json.items[{taskId}] with task progress
   • Update prd.json.agents.{agent} with your agent status
   • Commit any code changes (Developer only)

2. Signal readiness:

   AGENT_READY_FOR_HANDOFF
   

3. Output handoff phrase:

   HANDOFF:next_agent:context
   

The watchdog waits up to 30 seconds for AGENT_READY_FOR_HANDOFF before forcefully stopping you.

Helper: Encoding Context

PowerShell:

$context = @{ from = "pm"; reason = "task_assignment"; task = @{ id = "feat-001" } }
$json = $context | ConvertTo-Json -Compress
$bytes = [System.Text.Encoding]::UTF8.GetBytes($json)
$base64 = [System.Convert]::ToBase64String($bytes)
Write-Host "HANDOFF:developer:$base64"

Bash:

context='{"from":"pm","reason":"task_assignment","task":{"id":"feat-001"}}'
encoded=$(echo -n "$context" | base64 -w0)
echo "HANDOFF:developer:$encoded"

Receiving Handoff Context

When you start and receive handoff context, you'll see:

## HANDOFF CONTEXT (FROM PREVIOUS AGENT)
You are receiving control from another agent. Here is the context:

From: pm
Reason: task_assignment
Task Details: {"id":"feat-001","title":"Add user auth"}

Please acknowledge this handoff and continue the work accordingly.
Read prd.json.session and prd.json.items for full state.

Your first action should be:

1. Acknowledge the handoff: "Received handoff from PM for task feat-001"
2. Read prd.json.session, prd.json.agents, and prd.json.items to get full context
3. Begin the assigned work

Completion Protocol

When ALL PRD items have passes: true, the PM agent outputs:

<promise>RALPH_COMPLETE</promise>

This signals the watchdog to end the session gracefully.

Key Differences from Polling Mode

Error Handling

If you encounter an error that prevents work:

1. Save any partial state
2. Output error context in handoff:

   {
     "from": "developer",
     "reason": "error",
     "task": { "id": "feat-001" },
     "error": "Build failed: missing dependency X"
   }
   

3. Handoff to PM for resolution:

   HANDOFF:pm:base64_error_context