Worktree & PR
A worktree buys two things: the main working copy stays usable while the task
runs, and the untouched copy remains available as a baseline to compare
against. Both are worth real setup cost, and neither applies to every change.
This covers the isolation mechanics. How large the change itself should be is
scoped-change,
and it applies inside a worktree exactly as it does anywhere else.
Deciding
Consider a worktree when at least one holds: the task is large enough that a
half-finished state would block other work, several tasks need to progress on
the same repository at once, the result has to be compared against current
behavior, or the change is risky enough that abandoning it should cost
nothing.
Skip it when the change is small, self-contained, and reviewable in one pass.
Setting up an isolated copy for a typo fix or a one-line config edit costs
more than it saves, and the extra branch, directory, and PR are all overhead
that someone pays later. When the requester says to work directly on the
current branch, that decision is already made.
Running
- Branch from the integration branch, not from whatever is checked out.
The point of the isolated copy is a clean starting state; inheriting
unrelated in-progress edits forfeits it and makes the eventual diff
unreadable. Fetch first so the base is current, not a stale local ref.
- Leave the main working copy alone while the task runs. Editing both
defeats the isolation and destroys the baseline. If the task turns out to
need a change in the main tree, stop and say so rather than reaching
across.
- Parallel worktrees stay independent only while they stay disjoint. Two
copies of one repository editing the same files converge into a conflict
nobody scheduled, and the cost lands at merge time rather than now. Before
opening a second one, check what the first is touching.
- Name the artifacts' destination explicitly. Screenshots, builds, and
exports produced inside a worktree vanish with it. Anything the requester
needs to see must be written somewhere durable, or attached to the PR, and
the location has to be stated rather than assumed.
Before declaring it done
- Compare against the untouched baseline, not against expectation. The
reason to keep a clean copy is to run both and see the difference. For
behavior changes this means exercising the old path and the new one; for
visual changes it means the same view captured twice. "Should be
equivalent" is a claim, and the baseline is right there to check it.
- Account for what the tooling added. Generated output, formatter churn,
and dependency lock updates accumulate in an isolated copy without anyone
choosing them, and they are indistinguishable from intent once merged.
Either explain why each belongs in this branch, or drop it.
Landing
Follow the user's established authorization preferences. If committing,
pushing, or opening a pull request requires explicit authorization, complete
and validate the local changes, then hand over the exact commands until
authorized. When authorized, push the branch and open a pull request describing
motivation, the commands exercised, and anything reviewers must check by hand;
leave merging to the repository owner unless told otherwise. Say plainly which
parts are done, which are unverified, and which were deliberately left out,
since a worktree's isolation makes it easy to report an untested result as
finished. Once the branch has landed, remove the worktree; a stale copy of a
merged branch is a trap for the next session that opens it.
1---2name: worktree-pr3description: Decide whether a task deserves its own git worktree and, if chosen, run it end to end: branch from the integration branch rather than the current tree, keep the main working copy untouched while the task runs, compare the result against the untouched baseline before declaring it done, and prepare it for a pull request instead of merging. Use when a request says to do the work in a new worktree or branch, when a change is large or risky enough that the main tree should stay usable, when several tasks need to run in parallel on one repository, when a result has to be diffed against current behavior, and when deciding where screenshots, builds, and other artifacts produced in a worktree should end up.4license: Apache-2.05---67# Worktree & PR89A worktree buys two things: the main working copy stays usable while the task10runs, and the untouched copy remains available as a baseline to compare11against. Both are worth real setup cost, and neither applies to every change.1213This covers the isolation mechanics. How large the change itself should be is14[`scoped-change`](https://github.com/scarletkc/agents/blob/main/skills/scoped-change/SKILL.md),15and it applies inside a worktree exactly as it does anywhere else.1617## Deciding1819Consider a worktree when at least one holds: the task is large enough that a20half-finished state would block other work, several tasks need to progress on21the same repository at once, the result has to be compared against current22behavior, or the change is risky enough that abandoning it should cost23nothing.2425Skip it when the change is small, self-contained, and reviewable in one pass.26Setting up an isolated copy for a typo fix or a one-line config edit costs27more than it saves, and the extra branch, directory, and PR are all overhead28that someone pays later. When the requester says to work directly on the29current branch, that decision is already made.3031## Running3233- **Branch from the integration branch, not from whatever is checked out.**34 The point of the isolated copy is a clean starting state; inheriting35 unrelated in-progress edits forfeits it and makes the eventual diff36 unreadable. Fetch first so the base is current, not a stale local ref.37- **Leave the main working copy alone while the task runs.** Editing both38 defeats the isolation and destroys the baseline. If the task turns out to39 need a change in the main tree, stop and say so rather than reaching40 across.41- **Parallel worktrees stay independent only while they stay disjoint.** Two42 copies of one repository editing the same files converge into a conflict43 nobody scheduled, and the cost lands at merge time rather than now. Before44 opening a second one, check what the first is touching.45- **Name the artifacts' destination explicitly.** Screenshots, builds, and46 exports produced inside a worktree vanish with it. Anything the requester47 needs to see must be written somewhere durable, or attached to the PR, and48 the location has to be stated rather than assumed.4950## Before declaring it done5152- **Compare against the untouched baseline, not against expectation.** The53 reason to keep a clean copy is to run both and see the difference. For54 behavior changes this means exercising the old path and the new one; for55 visual changes it means the same view captured twice. "Should be56 equivalent" is a claim, and the baseline is right there to check it.57- **Account for what the tooling added.** Generated output, formatter churn,58 and dependency lock updates accumulate in an isolated copy without anyone59 choosing them, and they are indistinguishable from intent once merged.60 Either explain why each belongs in this branch, or drop it.6162## Landing6364Follow the user's established authorization preferences. If committing,65pushing, or opening a pull request requires explicit authorization, complete66and validate the local changes, then hand over the exact commands until67authorized. When authorized, push the branch and open a pull request describing68motivation, the commands exercised, and anything reviewers must check by hand;69leave merging to the repository owner unless told otherwise. Say plainly which70parts are done, which are unverified, and which were deliberately left out,71since a worktree's isolation makes it easy to report an untested result as72finished. Once the branch has landed, remove the worktree; a stale copy of a73merged branch is a trap for the next session that opens it.