Parallel - Concurrent Execution

Run multiple Ralph loops concurrently for independent tasks.

Quick Start

/parallel "fix auth errors" "fix api errors" "fix ui errors"
ralph parallel task1 task2 task3

When to Use

Good for Parallel

• •Independent file changes
• •Multiple module fixes
• •Batch reviews
• •Different analysis types

Must Be Sequential

• •Dependent changes
• •Same file modifications
• •Order-dependent operations
• •Shared state changes

Workflow

1. Spawn Parallel Agents

# Launch multiple background agents
Task:
  subagent_type: "code-reviewer"
  model: "sonnet"
  run_in_background: true
  prompt: "Review auth module"

Task:
  subagent_type: "code-reviewer"
  model: "sonnet"
  run_in_background: true
  prompt: "Review api module"

Task:
  subagent_type: "code-reviewer"
  model: "sonnet"
  run_in_background: true
  prompt: "Review ui module"

2. Monitor Progress

# Check each task
TaskOutput:
  task_id: "<auth-task>"
  block: false

TaskOutput:
  task_id: "<api-task>"
  block: false

3. Aggregate Results

# Wait for all to complete
TaskOutput:
  task_id: "<auth-task>"
  block: true

TaskOutput:
  task_id: "<api-task>"
  block: true

Parallel Patterns

Review Pattern

# Parallel reviews with different focus
/parallel "security review src/" "performance review src/" "quality review src/"

Fix Pattern

# Parallel fixes for different modules
/parallel "fix auth errors" "fix api errors" "fix db errors"

Analysis Pattern

# Parallel analysis tasks
/parallel "analyze complexity" "analyze coverage" "analyze dependencies"

Isolation

Each parallel task runs with:

• •Separate context (context: fork)
• •Independent iteration counter
• •Own quality gates
• •Isolated file access

Result Aggregation

All Succeed

• •Aggregate changes
• •Run global gates
• •VERIFIED_DONE

Partial Success

• •Report failures
• •Keep successful changes
• •Retry failed tasks

All Fail

• •Report all errors
• •Analyze patterns
• •Sequential retry

Integration

• •Used for independent sub-tasks
• •Each parallel task follows Ralph Loop
• •Results feed back to orchestrator

Anti-Patterns

• •Never run parallel on same files
• •Never exceed 5 concurrent agents
• •Never ignore partial failures
• •Never skip aggregation step