parallel

name	parallel
description	Orchestrate parallel Claude Code sessions in separate git worktrees. Use when the user says "parallel", "spin up tasks", "run these in parallel", or invokes /parallel.
allowed-tools	["Bash(bash *launch.sh *)","Bash(bash *cleanup.sh *)","Bash(bash *status.sh *)","Bash(bash *merge.sh *)","Bash(mkdir -p *)"]

/parallel

Context

• Task arguments: $ARGUMENTS
• Current branch: !git branch --show-current
• Existing worktrees: !git worktree list

Instructions

Step 0 — Read repo conventions

Before gathering tasks, learn the project's commit and coding conventions:

1. Read the project's CLAUDE.md (if it exists) for commit style, excluded files, and coding rules
2. Run git log --oneline -5 to see recent commit message style
3. Note anything agents should know: commit format, scopes, files to avoid, coding standards

This context feeds into Step 2 — it is not for prescribing exact commit messages.

Step 1 — Gather tasks

If task slugs were passed as arguments (e.g. /parallel cleanup-bash cleanup-git), use them directly — skip to composing prompts for each slug.

Otherwise, ask the user to list them. For each task:

• Derive a short slug (lowercase, hyphens, no spaces) — e.g., add-auth, fix-pagination
• Capture or compose a clear, self-contained prompt describing what the agent should do

If the user provides only one task, suggest claude --worktree as a simpler alternative. Proceed with the script if they still want /parallel.

Step 2 — Write prompt files

Create a session directory and write one prompt file per task:

prompts_dir="${CLAUDE_CONFIG_DIR:-$HOME/.claude}/parallel/sessions/$(date +%Y-%m-%dT%H-%M-%S)"
mkdir -p "$prompts_dir"

For each task, use the Write tool to create <prompts_dir>/<slug>.md containing the agent's prompt. Each prompt should be actionable and self-contained — the agent won't have the original conversation context.

Prompt structure:

• Describe the work clearly — what to change, why, and any constraints
• Include relevant conventions discovered in Step 0 as context (e.g. "This repo uses Conventional Commits with scopes matching the config area. TODO.md is untracked — do not commit it.")
• Tell agents to commit each change atomically
• Do not prescribe exact commit messages, types, or scopes — let the agent decide based on the context provided
• Mention files that are excluded or out of scope

Step 3 — Launch

Pass all task slugs and the prompts directory to the launch script:

bash ${CLAUDE_SKILL_DIR}/scripts/launch.sh --prompts-dir "$prompts_dir" <slug-1> <slug-2> [...]

The script handles everything: creates worktrees as sibling directories, launches agents with their prompts in a new tmux window (or prints manual commands if not in tmux), and outputs the merge plan.

Step 4 — Monitor (optional)

Check agent progress with the status script:

bash ${CLAUDE_SKILL_DIR}/scripts/status.sh

It auto-discovers all task/* branches, checks for commits since base, and detects open tmux panes. Pass task names to check a subset:

bash ${CLAUDE_SKILL_DIR}/scripts/status.sh <slug-1> <slug-2>

Step 5 — Merge

Once agents are done, merge their branches back. The merge script handles worktree removal, stashing, rebasing/merging, and branch cleanup in one go:

bash ${CLAUDE_SKILL_DIR}/scripts/merge.sh <slug-1> <slug-2> [...]

Default strategy is rebase (linear history). For merge commits:

bash ${CLAUDE_SKILL_DIR}/scripts/merge.sh --strategy merge <slug-1> <slug-2>

If a branch conflicts, the script aborts that branch cleanly and continues with the rest. Resolve conflicts manually, then clean up leftovers with cleanup.sh.

To keep branches after merging (e.g. for further review):

bash ${CLAUDE_SKILL_DIR}/scripts/merge.sh --no-cleanup <slug-1> <slug-2>

Notes

• The full lifecycle is: launch → monitor → merge → cleanup — each step has a script.
• The launch script is idempotent — it skips worktrees that already exist and panes that are already open.
• Worktree paths are siblings to the repo to avoid untracked directory noise.
• If --prompts-dir is omitted, agents launch with bare claude (no initial prompt).
• status.sh detects running agents by checking tmux pane directories — it may not detect agents that have exited but whose pane is still open.
• merge.sh handles worktree removal automatically — no need to remove them manually before merging.
• To discard all worktrees and branches without merging: cleanup.sh <slug-1> <slug-2> ...
• Prompt files are kept in $CLAUDE_CONFIG_DIR/parallel/sessions/ (defaults to ~/.claude/parallel/sessions/) for audit trails and re-runs.
• Set PARALLEL_MAX_PANES to override the default limit of 16 panes per tmux window.