Per-session git worktree (Phase 0)

kube-coder isolates sessions at the terminal level (each task gets its own tmux session + ~/.claude-tasks/<id>/) but everything still runs in the same /home/dev filesystem. So two sessions on the same repo collide three ways:

• git tree — branch/checkout/index thrash, a dirty working tree
• build output — both write ./dist, ./build; Kaniko reads its context from the shared path
• dev-server ports — EADDRINUSE on :3000 (ports are per-pod, not per-session)

This skill fixes all three for the current session with no infra change: it gives you a dedicated git worktree (own checkout + branch) at a path outside the repo, and leases a free port for any server you start. Worktrees share the main clone's object store, so they're cheap on disk — only working files are duplicated.

The helper is worktree.sh next to this file. Invoke it as bash "$CLAUDE_SKILL_DIR/worktree.sh" … (or with its absolute path under /home/dev/kube-coder/.claude/skills/worktree/).

Start an isolated session (the common path)

Run from inside the repo you want to work on (cd there first). $ARGUMENTS is an optional human slug; if omitted it derives one from the current task id (KC_TASK_ID) or a random token, so it's always unique.

WT=/home/dev/kube-coder/.claude/skills/worktree/worktree.sh
eval "$(bash "$WT" new "$ARGUMENTS")"   # creates worktree+branch+port, sets KC_WT / PORT
cd "$KC_WT"                              # <-- this cd PERSISTS across your Bash calls

After this, do all work for the change from $KC_WT. The human summary (path / branch / port / preview URL) is printed to stderr — note the port, because exported env vars do not survive between separate Bash tool calls (the cd does). The worktree also records everything in $KC_WT/.kc-worktree.json, so you can re-read the port later:

jq -r .port "$KC_WT/.kc-worktree.json"   # or read .branch / .path

What you get

• worktree dir: /home/dev/.worktrees/<repo>/<slug>/
• branch: kc/<slug> (branched from current HEAD, or a base-ref: new <slug> <base>)
• leased port in 3000–3999 (skips kube-coder's reserved ports and other leases)

Running a dev server / preview

Bind the leased port explicitly (env doesn't persist between calls), then open it through the app-proxy — no port config in the dashboard needed:

PORT=$(jq -r .port "$KC_WT/.kc-worktree.json")
cd "$KC_WT" && PORT=$PORT npm run dev      # or: python -m http.server "$PORT", etc.
# preview at:  /api/app-proxy/<port>/

Builds

Because the worktree is its own path, docker-build (Kaniko) and local build output (./dist, ./build) no longer collide with other sessions. Run builds with the worktree as context — e.g. docker-build -t img:tag "$KC_WT".

List / tear down

bash "$WT" list                 # worktrees + leased ports for the current repo
bash "$WT" rm <slug>            # remove the worktree DIR; the branch is KEPT
bash "$WT" rm <slug> --force    # also drop uncommitted changes in it

rm deliberately keeps kc/<slug> — that branch is your change. Push/PR it as usual, then delete with the git branch -D line rm prints. rm refuses to delete a dirty worktree unless you pass --force.

Footguns

1. Uncommitted work stays behind. A worktree branches from HEAD (the committed state). Anything uncommitted on your original checkout is not carried in. Commit or stash first if you need it.
2. cd persists, env does not. In this harness the working directory survives between Bash calls but exported vars don't. Rely on cd "$KC_WT" and re-read the port from .kc-worktree.json when you need it; don't assume $PORT is still set in a later call.
3. Disk. Worktrees share .git, but a fresh node_modules / .venv / build tree per worktree is not shared and can fill the 20–50Gi PVC. Keep a handful live at once; rm ones you're done with.
4. One repo, one clone. All worktrees for a repo register under that clone's .git/worktrees. Don't rm -rf a worktree dir by hand — use rm (which runs git worktree remove + prune) so git's bookkeeping stays consistent.
5. Not for the boot-time kube-coder pull. start.sh does git pull --ff-only origin main on /home/dev/kube-coder at boot. Working in a kc/<slug> worktree keeps main clean so that pull still fast-forwards — which is the point. Don't check out main inside a worktree.

Scope (why this is "Phase 0")

Pure userland, works today, teaches us the naming/cleanup ergonomics. It does not yet integrate with the Task API or dashboard — the worktree/branch/port aren't shown in the task UI and aren't auto-created when a task starts, and the port lease is advisory (first-come). Phase 1 moves this into server.py's create_task() (record worktree in task.json, clean up in delete_task, surface it in NewTaskForm/task detail). Sanity-check the helper before relying on it:

bash -n /home/dev/kube-coder/.claude/skills/worktree/worktree.sh