Run your own tools and scripts in a worktree

Each worktree terminal is a real shell in the worktree’s folder, so make, just, your own scripts and any agent CLI run there exactly as they do in any other terminal. The difference is that once an agent has launched there, the folder holds the team context that agent asked for: CLAUDE.md, AGENTS.md, skills and anything else in its context, written as real files and hidden from git.

To run several agents at once, open a worktree per story and split the terminal area into tiles. Each agent works in its own folder and branch, and you watch them side by side, with live diffs.

What the terminal gives you

When CodeSense opens a terminal for a worktree, it starts a login zsh (zsh -l) with the worktree as its working directory.

• Your zsh config loads. CodeSense points ZDOTDIR at a small wrapper directory (~/.codesense/shell/zsh) whose files source your own .zshenv, .zprofile, .zshrc and .zlogin. Your aliases, functions, PATH additions and tools all apply.
• CodeSense adds a little. ~/.codesense/bin goes first on PATH, and the shell gets a few CODESENSE_* variables, such as CODESENSE_TAB_SESSION_ID.
• The prompt is a Starship preset by default. With Shell integration on (the default), CodeSense renders its bundled Starship prompt using the Compact preset. To keep the prompt your own config sets, open Settings, find Terminal prompt and choose Plain (“Your default shell prompt (no Starship)”).

Coming

The terminal always starts zsh today, whatever $SHELL says. If you use bash or fish, your zsh config (if any) loads instead of your usual shell’s.

Unlocks with support for your own login shell.

Run your scripts in a worktree

1. Launch a story’s worktree (see Launch your first agent), or open the terminal of a worktree you already have.

2. Check that the context is there. CodeSense copies an agent’s context in when the agent launches in the worktree; a terminal you open on your own doesn’t trigger that copy:

   ls .claude/skills
   # add-endpoint  release-notes  review-pr  write-migration

3. Run what you normally run. In the Ledgerline sample, a worktree for story LL-214 might run:

   npm test                      # the repo's own scripts
   just migrate                  # your task runner
   ./scripts/dev.sh              # or anything you keep in the repo

Anything you start here, including another agent CLI, reads the same CLAUDE.md, AGENTS.md and .claude/skills/ as every other agent on the story. Nothing extra shows up in git status, because CodeSense lists the paths it copied in .git/info/exclude.

Headless runs, pre-commit hooks and CI-style scripts you start here read the same context files, because they’re plain files in the folder. No special integration is involved.

Coming

A command typed in the first second or two can run before .claude/ is written. Until the terminal waits for the context write, check with ls first, as in step 2.

Unlocks with the terminal waiting for context to land.

When the window closes, and when you quit

• Closing the window keeps CodeSense running in the background, so its terminals and the processes in them keep going.
• Quitting CodeSense ends its terminal sessions. The next launch opens fresh shells.

Watch several agents at once

Split the terminal area into tiles to watch several worktrees together. See Terminals for tabs, split tiles and the agents list.

Related

• Context in a worktree and reverse sync: how context lands in the worktree, and reverse sync (in Beta).
• Worktrees: how a story becomes a worktree and a branch.