Git Worktrees + Celistra: Parallel Claude Code Sessions With Full Supervision

Tim Dietrich's writeup on parallel sub-agents calls git worktrees "the foundation of parallel agentic development." That's right — and the next layer up is the supervision question. Worktrees give the agents space; the daemon gives you the room to watch.

What worktrees give you

One repository, multiple checkouts. Each worktree is a real directory; each has its own .git indirection but shares the underlying object database. Two agents working in two worktrees of the same repo:

• Don't fight over node_modules (each worktree has its own)
• Don't fight over dist / build outputs
• Don't fight over generated files
• Can be on different branches simultaneously
• Share git objects on disk (cheap)

Setup

cd ~/code/myproject
git worktree add ../myproject-task-1 -b agent/task-1 main
git worktree add ../myproject-task-2 -b agent/task-2 main
git worktree add ../myproject-task-3 -b agent/task-3 main

Three worktrees, three new branches off main. Each agent gets a clean place to work.

Spawn three agents

celistrad ctl spawn \
  --workspace ~/code/myproject-task-1 \
  --auto-restart \
  'claude --print "Refactor src/auth/* to use the new session API. Detail in TASK.md"'

celistrad ctl spawn \
  --workspace ~/code/myproject-task-2 \
  --auto-restart \
  'claude --print "Find and fix every typing.Optional[X] = None pattern. Detail in TASK.md"'

celistrad ctl spawn \
  --workspace ~/code/myproject-task-3 \
  --auto-restart \
  'claude --print "Add tests for src/services/payment/. Detail in TASK.md"'

Each agent's --workspace is enforced by the sandbox profile — they can only write inside their own worktree. Even if Claude Code goes off the rails and tries to write to ~/code/myproject-task-2/src from the agent in worktree 1, the kernel says no.

Watch them

Dashboard fleet view shows three rows. Click any to attach. Each PTY is independent; the 64KB scrollback ring lets you drop in mid-task.

Pull results

cd ~/code/myproject
git fetch ../myproject-task-1 agent/task-1:agent/task-1
git fetch ../myproject-task-2 agent/task-2:agent/task-2
git fetch ../myproject-task-3 agent/task-3:agent/task-3

# Now the three branches are in your main repo, ready to review:
git log --oneline agent/task-1
git log --oneline agent/task-2
git log --oneline agent/task-3

Cleanup

# When done with a worktree:
cd ~/code/myproject
git worktree remove ../myproject-task-1
# Branch stays; if you want it gone too:
git branch -D agent/task-1  # only if merged or you really mean it

Why this is better than three terminals

• Persistence. Close the terminals, agents keep running. Daemon-managed.
• Sandboxing. Each agent fenced to its worktree. Even reckless Claude Code can't cross-contaminate.
• Audit. 30-day SQLite history records every spawn, every exit, every restart per agent.
• Mobile. Approval requests buzz your phone with the agent name. You always know which task is asking.
• Restart-on-crash. Model 500 doesn't kill the run; the daemon respawns.

Failure modes

Agents fighting over the lockfile. Don't share a lockfile — each worktree has its own package-lock.json / yarn.lock. If they all need to install at once, run npm install once on main and copy node_modules into each worktree before spawning, or accept the parallel install.

Three agents all want to bump the same shared file. They'll merge-conflict at the end. That's fine — you review three branches and merge in order. Worktrees don't make merging magical; they just make the work isolated until then.

Agent named "task-1" is actually rewriting the schema. The audit log + the diff from main shows you what each one did. If you don't recognize a change, that's the moment for the human to step in.

Cost arithmetic

Three Claude Code instances bill against your Anthropic plan as three sessions. If they each run for 30 minutes, you pay roughly 3x what one 30-minute run costs. The math works out when the three are actually independent — different parts of the codebase. It doesn't work out when they all fight over the same module; you've turned 30 minutes of work into 90 minutes of token spend.