Ralph Spawn

/ralph.spawn [epic]

Create worktrees for all ready tasks in an epic and start parallel agent execution.

Prerequisites

CRITICAL: Run /ralph.preflight before spawning agents. If Linear MCP is not authenticated, spawn will BLOCK with recovery instructions.

Process

0. Pre-Flight Check (REQUIRED)

   Run /ralph.preflight to verify:

   • Linear MCP is authenticated (REQUIRED)
   • Git repository is valid (REQUIRED)
   • Cognee is available (OPTIONAL - will buffer locally if not)

   If pre-flight BLOCKS, fix the issue and resume with /ralph.resume.

1. Query Linear for ready tasks

   mcp__plugin_linear_linear__list_issues({
     "team": "floe",
     "project": "floe-{epic_number}-{slug}",
     "state": "backlog"
   })
   

2. Build dependency graph from blockedBy relations

3. Detect file overlaps to determine parallelization

   • Tasks modifying different files → parallel
   • Tasks modifying same files → sequential (waves)

4. Create worktrees for non-overlapping tasks

   git worktree add "../floe-agent-{epic}-{task_slug}" -b "feature/{epic}-{task}" main
   

5. Enable direnv in each worktree (CRITICAL for environment setup)

   cd "../floe-agent-{epic}-{task_slug}" && direnv allow
   

   Why: Without direnv allow, the worktree won't load .envrc and environment variables (like PYRIGHT_PYTHON_FORCE_VERSION) won't be set. This causes:

   • Type checking warnings/failures
   • Missing environment variables
   • Inconsistent development environment

6. Initialize agent state in each worktree

   • Copy .ralph/templates/plan.json → .agent/plan.json
   • Copy .ralph/templates/activity.md → .agent/activity.md
   • Copy .ralph/templates/PROMPT.md → .agent/PROMPT.md
   • Copy .specify/memory/constitution.md → .agent/constitution.md

7. Claim tasks in Linear

   mcp__plugin_linear_linear__update_issue({
     "id": task_id,
     "state": "started",
     "assignee": "me"
   })
   

8. Update manifest.json with active agents

Output

RALPH WIGGUM SPAWN - EP001

Created Worktrees: 4
Wave 1: [T001 (auth), T003 (catalog), T005 (tests)]
Wave 2: [T002 (models)] - depends on T001

Worktree Details:
- ../floe-agent-ep001-auth    T001  FLO-33  feature/ep001-auth
- ../floe-agent-ep001-catalog T003  FLO-35  feature/ep001-catalog
- ../floe-agent-ep001-tests   T005  FLO-37  feature/ep001-tests

Next: Run /ralph.status to monitor progress

Configuration

See .ralph/config.yaml:

• max_parallel_agents: Maximum concurrent worktrees
• worktree_pattern: Naming convention for worktrees

Agent Loop (Per Worktree)

Each agent runs the Ralph Wiggum pattern:

1. Read state - .agent/plan.json, .agent/activity.md
2. Identify next subtask - First with passes: false
3. Implement - Make atomic, focused changes
4. Run quality gates - lint, type, test, security, constitution
5. Update state - Mark subtask complete, commit, log
6. Repeat until all subtasks pass or BLOCKED

Error Handling

• If worktree exists: Remove and recreate
• If branch exists: Delete and recreate
• If Linear unavailable: Signal error, don't start
• If no ready tasks: Report "No ready tasks for epic"

Related Commands

• /ralph.status - Monitor agent progress
• /ralph.integrate - Prepare for PR after completion
• /ralph.cleanup - Remove worktrees after merge