ralph-eats-beads

name	ralph-eats-beads
description	Autonomous AI coding loop using bd for task management with parallel tmux subagents. Ships features while you sleep.
license	MIT

Ralph BD - Parallel Autonomous Coding Loop

Contract: skills-lib-lkv (Ralph Contract v1)

Ralph runs Claude agents in a loop, picking tasks from bd ready, implementing them, and closing them on success. This version uses tmux for observability and supports parallel subagents.

Architecture

ralph (tmux session)
├── parent (window 0) - orchestrator Claude
├── watchdog (window 1) - health monitor Claude (optional)
├── BD-001 (window 2) - subagent
├── BD-002 (window 3) - subagent
└── BD-003 (window 4) - subagent

Prerequisites

• bd initialized with issues/epic
• tmux installed
• Claude CLI (claude command available)

Quick Start

Option A: Manual Plan (you create issues)

# 1. Create your epic and issues
bd create --type epic "Feature: User Authentication"
bd create "Add login form" --parent BD-001 --priority 1
bd create "Add email validation" --parent BD-001 --priority 2
bd create "Add auth server action" --parent BD-001 --priority 3

# 2. Start Ralph (25 max iterations)
scripts/ralph.sh 25

# 3. Attach to watch
tmux attach -t ralph

Option B: AI-Generated Plan (planner creates issues)

# 1. Let Claude create the plan from requirements
scripts/planner.sh requirements.md

# Or inline:
scripts/planner.sh <<< "Add OAuth2 authentication with Google and GitHub providers"

# Or pipe from anywhere:
cat feature-spec.md | scripts/planner.sh

# 2. Watch the planner
tmux attach -t ralph-plan

# 3. Once plan is created, implement it
scripts/ralph.sh 25

Planning vs Implementation

Script	Session	Purpose
planner.sh	ralph-plan	Recursive loop - creates/refines bd issues until plan is complete
ralph.sh	ralph	Parallel loop - implements issues from bd ready

Workflow:

Requirements → planner.sh (iterates) → bd issues → ralph.sh (parallel) → Implemented code

Both loops support MAX_ITERATIONS - planner defaults to 3, implementation defaults to 10.

How It Works

Key principle: Bash manages the iteration loop, each iteration gets fresh Claude context.

ralph.sh (bash loop)
├── Iteration 1
│   ├── Spawn fresh Claude parent
│   ├── Parent spawns subagents, then exits
│   ├── Bash waits for subagent windows to close
│   ├── bd sync
│   └── Check bd ready → continue or done
├── Iteration 2
│   └── (repeat with fresh context)
└── ...

1. Bash checks bd ready for available work
2. Bash spawns fresh Claude parent for this iteration
3. Parent runs bd prime, analyzes parallelizability
4. Parent spawns subagent tmux windows (3-5 concurrent), then exits
5. Bash waits for all subagent windows to close
6. Each subagent implements ONE issue, commits, closes issue
7. Bash syncs bd and loops until no work remains

Observer Commands

# Attach to see everything
tmux attach -t ralph

# Tile all panes
tmux select-layout -t ralph tiled

# Watch specific agent
tmux select-window -t ralph:BD-123

# Capture agent output
tmux capture-pane -t ralph:BD-123 -p -S -1000 > agent.log

# Kill stuck agent
tmux kill-window -t ralph:BD-123

# Kill entire session
tmux kill-session -t ralph

Planner Details

The planner (scripts/planner.sh) runs a recursive loop that iterates until the bd plan fully covers the input requirements:

while iteration < MAX_ITERATIONS:
    1. Read input plan file
    2. Check current bd state (existing issues)
    3. Compare: what's in the plan but missing from bd?
    4. If nothing missing → exit with summary
    5. Create missing issues with dependencies
    6. Repeat

Each iteration the planner:

1. Re-reads the plan - Keeps the source of truth in focus
2. Audits bd state - What stories exist? What's missing?
3. Fills gaps - Creates issues for uncovered sections
4. Sets dependencies - Ensures correct build order

This ensures complex plans with many sections get fully decomposed into implementable stories.

Input Formats

# From file
scripts/planner.sh requirements.md
scripts/planner.sh feature-spec.txt

# Inline (heredoc)
scripts/planner.sh << 'EOF'
Add user authentication:
- OAuth2 with Google and GitHub
- Session management with JWT
- Protected routes middleware
EOF

# One-liner
scripts/planner.sh <<< "Add dark mode with system preference detection"

# Pipe from anywhere
curl -s https://example.com/spec.md | scripts/planner.sh
gh issue view 123 --json body -q .body | scripts/planner.sh

What the Planner Creates

Epic: AUTH-001 "Feature: OAuth2 Authentication"
├── AUTH-002 "Add OAuth2 configuration" [ready]
├── AUTH-003 "Create Google provider" [depends: AUTH-002]
├── AUTH-004 "Create GitHub provider" [depends: AUTH-002]
├── AUTH-005 "Add callback handlers" [depends: AUTH-003, AUTH-004]
└── AUTH-006 "Add protected route middleware" [depends: AUTH-005]

Stories include:

• Clear acceptance criteria
• File paths to modify
• References to existing patterns
• Testing requirements

Watchdog (Active Health Monitor)

The watchdog is a separate Claude agent that monitors Ralph sessions and unsticks them automatically. Unlike the passive monitor, it takes action:

# Start watchdog for a ralph session (30 second check interval)
scripts/watchdog.sh ralph-myproject 30

# It runs in its own tmux window within the session
tmux select-window -t ralph-myproject:watchdog

The watchdog:

• Detects idle agents - Looks for Claude CLI prompts indicating waiting state
• Checks issue status - Queries bd to see if work is complete
• Nudges stuck agents - Sends tmux keystrokes to prompt them to continue
• Kills completed windows - Cleans up windows where the issue is closed
• Kicks the parent - Sends Enter if parent is idle with work remaining

This keeps Ralph flowing autonomously without human intervention.

Passive Monitor (Logging Only)

For unattended runs where you just want logging without intervention:

# Start monitor (checks every 20 min, logs to /tmp/ralph-monitor.log)
nohup scripts/monitor.sh 20 &

# Check status
tail -f /tmp/ralph-monitor.log

The passive monitor detects:

• STOPPED: Ralph session died unexpectedly
• STUCK: Only parent window but work remains (2 consecutive checks)
• COMPLETE: No ready work and only parent window
• RUNNING: Normal operation with active subagents

bd Workflow

Setup Work

# Create epic
bd create --type epic "Feature: Dark Mode"  # → BD-001

# Create stories
bd create "Add theme toggle" --parent BD-001 --priority 1
bd create "Add theme context" --parent BD-001 --priority 2
bd create "Update components" --parent BD-001 --priority 3

# Add dependencies (bd ready respects these)
bd dep add BD-004 BD-003  # BD-004 depends on BD-003

Monitor Progress

bd ready              # What's workable?
bd epic status BD-001 # Epic completion
bd blocked            # What's stuck?
bd activity           # Recent changes

After Completion

bd list --state closed --parent BD-001  # All closed?
bd show BD-002 --comments               # Review learnings
bd epic close-eligible                  # Close the epic

Best Practices

Right-Sized Stories

BAD:  "Build authentication system"
GOOD: "Add login form with email/password fields"
      "Add client-side email validation"
      "Create auth server action"

Clear Acceptance Criteria

In issue body:

Acceptance Criteria:
- [ ] Email field with validation
- [ ] Password field with show/hide
- [ ] typecheck passes
- [ ] tests pass

Learnings Accumulate

Agents add learnings via bd comments. Future iterations benefit:

bd show BD-005 --comments
bd search "migration pattern"

Not Suitable For

• Exploratory work without clear criteria
• Major refactors spanning entire codebase
• Security-critical code
• Anything needing human review before commit