kube-coder: Skill for Claude Code

Per-session git worktree

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