bug(finish): a merged worktree stays on disk until a second call removes it #691

New issue

Open

opened 2026-10-03 15:41:17 +00:00 by hexajon · 0 comments

hexajon commented

2026-10-03 15:41:17 +00:00

Owner

Problem

When I finish a branch that lives in a linked worktree, madt_finish merges the PR and leaves the worktree on disk. Removing it takes a second call, madt_worktrees action=remove. In a batch of 40 small changes across repos, that is 40 extra calls, and a skipped call leaves a stale worktree and branch behind.

Ask

After the merge lands, madt_finish removes the linked worktree that has the finished branch checked out, but only when the ADR 0029 proof holds:

git finds the worktree clean, and
the forge's API confirms that the commits of its branch are reachable from the forge's default branch.

A failed forge read, a dirty tree, or a missing proof leaves the worktree in place, and the result says why.

Decision records

ADR 0029 already classifies a clean worktree remove as machine-proof. This change uses the same proof and adds no new destructive path.
ADR 0019 says the foreign flow ends at madt_finish plus worktree cleanup with madt_worktrees action=remove. This change folds the cleanup into finish, so ADR 0019 needs a dated amendment to that sentence.

Cases

Merged now: the proof runs, and the worktree is removed.
Auto-merge scheduled (the forge merges when checks pass): nothing is merged yet, so the worktree stays, and the result says the merge is pending.
Merged with squash or rebase: the branch commits are not on the default branch, only their content. ADR 0029 lets prune use a content proof for that case, but its worktree row names commits only. Open question: does worktree removal accept the content proof too?

Acceptance

A finish from a clean linked worktree, merged with a merge commit, removes the worktree once the forge proof passes.
A finish whose merge is only scheduled leaves the worktree and says the merge is pending.
With the forge out of reach, the worktree stays, and the result says the proof failed.
A dirty worktree stays.

Depends on #666.

## Problem When I finish a branch that lives in a linked worktree, `madt_finish` merges the PR and leaves the worktree on disk. Removing it takes a second call, `madt_worktrees action=remove`. In a batch of 40 small changes across repos, that is 40 extra calls, and a skipped call leaves a stale worktree and branch behind. ## Ask After the merge lands, `madt_finish` removes the linked worktree that has the finished branch checked out, but only when the ADR 0029 proof holds: - git finds the worktree clean, and - the forge's API confirms that the commits of its branch are reachable from the forge's default branch. A failed forge read, a dirty tree, or a missing proof leaves the worktree in place, and the result says why. ## Decision records - ADR 0029 already classifies a clean worktree remove as machine-proof. This change uses the same proof and adds no new destructive path. - ADR 0019 says the foreign flow ends at `madt_finish` plus worktree cleanup with `madt_worktrees action=remove`. This change folds the cleanup into finish, so ADR 0019 needs a dated amendment to that sentence. ## Cases - Merged now: the proof runs, and the worktree is removed. - Auto-merge scheduled (the forge merges when checks pass): nothing is merged yet, so the worktree stays, and the result says the merge is pending. - Merged with squash or rebase: the branch commits are not on the default branch, only their content. ADR 0029 lets prune use a content proof for that case, but its worktree row names commits only. Open question: does worktree removal accept the content proof too? ## Acceptance - A finish from a clean linked worktree, merged with a merge commit, removes the worktree once the forge proof passes. - A finish whose merge is only scheduled leaves the worktree and says the merge is pending. - With the forge out of reach, the worktree stays, and the result says the proof failed. - A dirty worktree stays. Depends on #666.

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

sixfold-space/madtea#691

No description provided.

Rows
Columns