# Mission: Git Worktrees

## Why

To run several AI coding agents in parallel — each in its own isolated checkout — and then confidently land the finished work back onto the main branch without asking an agent to do it. Today the "spin one up" half sort of works; the "put the work back and clean up" half is a black box, so finished work piles up in directories that never get integrated.

## Success looks like

- Can explain, unprompted, what a worktree actually is and why two agents in two worktrees don't collide
- Can start a worktree for a new task from a cold terminal, without looking anything up
- Can take a finished worktree's work and land it on `main` — knowing exactly which directory to run each command from
- Can fully tear down a finished worktree: directory, branch, and stale metadata, with nothing orphaned
- Can read `git worktree list` and say what every line means
- Knows the names of the things well enough to search for answers and to give agents precise instructions

## Constraints

- Git experience is delegation-shaped: comfortable asking an agent to "merge this" / "make a PR" / "create a branch", but has not built the underlying mental model
- Terminology is the acute gap — not knowing what things are *called* blocks self-rescue
- Prefers the mental model taught first, so the commands follow from it instead of being memorized
- Lessons must stay short; one tangible win each

## Out of scope

- Rebase strategy, conflict-resolution technique, and git internals beyond what worktrees require
- Agent orchestration tooling itself (how to run the agents) — this workspace is about the git layer underneath
- Submodules (worktree support for them is incomplete and it is a distraction)