Parallel Worktree Session

Overview

One branch. One worktree. One agent. This is the core constraint that makes parallel development work without file conflicts, port collisions, or context pollution between features.

Core principle: Spin up isolated workspaces first, coordinate after.

Phase 1 — Plan: Feature Decomposition

Before creating anything, define what you're parallelizing:

1. List features — name each by what it does, not how
2. Verify independence — features that modify the same files are NOT good candidates
3. Choose execution mode (see decision table below)
4. Build a tracking table:

Feature	Branch	Worktree Path	Mode	Port
myapp-auth	feature/auth	.worktrees/auth	multi-window	3001
myapp-payments	feature/payments	.worktrees/payments	multi-window	3002

Naming convention: {project}-{feature-slug} — e.g., myapp-auth, api-rate-limiting

Multi-Window vs Subagent Mode

digraph mode_decision {
    "Features independent?" [shape=diamond];
    "Need full Claude context per feature?" [shape=diamond];
    "Short-lived tasks (<30 min)?" [shape=diamond];
    "Use multi-window" [shape=box];
    "Use subagent mode" [shape=box];
    "Reconsider parallelization" [shape=box];

    "Features independent?" -> "Need full Claude context per feature?" [label="yes"];
    "Features independent?" -> "Reconsider parallelization" [label="no"];
    "Need full Claude context per feature?" -> "Use multi-window" [label="yes"];
    "Need full Claude context per feature?" -> "Short-lived tasks (<30 min)?" [label="no"];
    "Short-lived tasks (<30 min)?" -> "Use subagent mode" [label="yes"];
    "Short-lived tasks (<30 min)?" -> "Use multi-window" [label="no"];
}

	Multi-Window	Subagent Mode
Isolation	Full — separate Claude processes	Shared parent context
Best for	Long-running features, deep context	Parallel tasks, batch work
Setup	cd <path> && claude per window	Task tool dispatch
Max concurrent	4-5 (coordination cost rises beyond)	Bounded by Task concurrency

Phase 2 — Spin Up

REQUIRED SUB-SKILL: Use superpowers:using-git-worktrees for each feature. That skill handles: directory selection, .gitignore verification, dependency setup, and baseline test verification.

Multi-window mode — after each worktree is created, open a new terminal per feature:

cd .worktrees/auth && claude
cd .worktrees/payments && claude

Subagent mode — dispatch via Task tool:

Task("Implement auth feature in worktree at .worktrees/auth on branch feature/auth.
     Goal: [specific acceptance criteria]
     Constraints: Do not modify files outside src/auth/
     Return: Summary of changes and test results")

Port offset rule: Stagger dev servers to avoid conflicts:

• Main: 3000, Feature 1: 3001, Feature 2: 3002

Phase 3 — Track

Run this to see worktree status at a glance:

git worktree list

Conflict risk check — detect if two worktrees touched the same file:

# Compare changes in two worktrees against main
git -C .worktrees/auth diff main --name-only > /tmp/auth-files.txt
git -C .worktrees/payments diff main --name-only > /tmp/payments-files.txt
comm -12 <(sort /tmp/auth-files.txt) <(sort /tmp/payments-files.txt)

Sync cadence: At the start of each session, pull main into each active worktree:

git -C .worktrees/auth pull origin main

Worktrees that drift >1 week from main accumulate merge conflicts exponentially. Sync daily on active work.

Phase 4 — Merge

REQUIRED SUB-SKILL: Use superpowers:finishing-a-development-branch per feature when ready.

Merge order matters when multiple features finish close together:

1. Run the conflict check from Phase 3 first
2. Merge the feature with fewer shared files first
3. After first merge, re-run conflict check against updated main before merging second feature
4. If both features modified the same file: resolve before either merge

Phase 5 — Cleanup

After each feature is merged/discarded:

# ALWAYS use this — never rm -rf a worktree directory
git worktree remove .worktrees/auth

# After all removals, prune stale metadata
git worktree prune

Safety check before removing:

git -C .worktrees/auth status  # Must show clean working tree
git -C .worktrees/auth log origin/main..HEAD  # Must show no unpushed commits

Manual rm -rf on a worktree corrupts git's internal metadata. git worktree remove updates .git/worktrees/ correctly.

Quick Reference

Phase	Command
List worktrees	git worktree list
Add worktree	git worktree add .worktrees/auth -b feature/auth
Sync with main	git -C .worktrees/auth pull origin main
Conflict check	comm -12 <(sort auth-files.txt) <(sort payments-files.txt)
Safe remove	git worktree remove .worktrees/auth
Prune metadata	git worktree prune

Common Mistakes

Mistake	Fix
Same port for all dev servers	Offset: 3001, 3002, 3003
Checking out same branch in two worktrees	Each worktree requires a unique branch
rm -rf on worktree directory	Always use git worktree remove
Letting worktrees drift weeks from main	Pull main at session start
Merging without conflict check	Run file overlap check before second merge
>5 concurrent worktrees	Coordination cost exceeds parallelization benefit

Integration

Delegates to:

• superpowers:using-git-worktrees — Phase 2: single worktree creation with safety checks
• superpowers:finishing-a-development-branch — Phase 4: merge/PR/cleanup per feature
• superpowers:dispatching-parallel-agents — Subagent dispatch pattern for Phase 2

Called by:

• superpowers:brainstorming — when approved design splits into parallel implementation tracks
• superpowers:dispatching-parallel-agents — when tasks require filesystem isolation