[zolotarev]

Friction

In a project where merge requests into the main branch are squash-merged, dpcli gc --remove printed "no merged branches" after the module was merged. The module branch and its task branch were never ancestors of the main branch, so git branch --merged did not list them, and their copies and branches had to be removed by hand.

Seen 1 time: one module, 2 branches (module branch + task branch), ledger entry F-1. A squash merge into the main branch is the default in that project, so every module ends this way.

Change

squash_merged() in commands/branches.py, used by gc:

• A branch counts as squash-merged when its net diff from the merge-base (a temporary commit-tree of its tree) has a patch-id equivalent in the main branch (git cherry prints -).
• Branches that are ancestors of a squash-merged branch also count (task branches merged into the module branch).
• A branch whose net diff is empty is never matched.
• Matched branches are deleted with branch -D, because -d refuses them. The gc_live_reason and dirty-copy checks still apply.
• Cost: a few git calls per unmerged local branch, and one dangling commit object each, collected by git gc. This is noted in the docstring.

Tests

tests/test_gc.py:

• test_squash_merged: a module branch and its task branch are removed after a squash; an unrelated branch is kept.
• test_squash_then_more_work_and_empty_diff_kept: a branch that got a commit after its squash is kept, and so is a branch whose net diff is empty.

Full suite: 207 OK. VERSION 35, CHANGELOG line.

[github-actions]

VERDICT: rework

• Not anonymized: the PR body names a project ("ashoo") and quotes the friction in Russian. Public repo rule: no project data. Rewrite the body as an English, anonymized conclusion.
• Evidence is thin: "Seen ×0 in 0 task(s)". The friction (squash MRs leave copies and branches that gc does not see) is plausible. Link a feedback issue or give a real count before this merges.
• lib/dpcli/commands/branches.py:~254: branch -D is now used for every squash-detected branch. Detection is a patch-id match, and the second pass treats any ancestor of a squashed branch as safe. Both are heuristics that decide whether work gets force-deleted. Add a test where the branch gets one more commit after the squash and must be kept. Add one for an empty-diff branch (tree equals merge-base tree), whose git cherry result is not obvious.
• squash_merged runs commit-tree once per unmerged local branch on every gc. That writes a dangling object each time and costs a few git calls per branch. Acceptable, but say so in the docstring, or skip branches that have a worktree with uncommitted work.
• Process is fine: VERSION bumped to 35 and a CHANGELOG line added. The CHANGELOG line is long, so shorten it. I did not run the tests in this review.

[zolotarev]

Rework done (76c44ee):

• The body is anonymized, in English, and gives the real count (1 module, 2 branches).
• New test for a branch with work after the squash (kept) and a branch with an empty net diff (kept, with an explicit guard).
• The docstring notes the cost and the dangling objects.
• The CHANGELOG line is shortened.

[github-actions]

VERDICT: accept

• Real friction (feedback label, ledger F-1): squash-merged module and task branches stayed because git branch --merged cannot see them. The fix is small and local to gc.
• Safe: -D is used only for branches whose whole net diff has a patch-id match in main (git cherry). Branches with later commits or an empty diff are never matched. gc_live_reason, the dirty-copy and symlink checks still apply. A false negative only keeps a branch.
• Tests: test_squash_merged and test_squash_then_more_work_and_empty_diff_kept cover the match, the ancestor (task branch) case, an unrelated branch, work after the squash, and an empty diff.
• Process: VERSION 35 and a CHANGELOG line are present, in English, with no project data.
• Nit: b not in out in the ancestor comprehension at lib/dpcli/commands/branches.py:216 is redundant. rest is already filtered, and the any(...) over out is enough.
• Nit: squash_merged runs on every gc call, including the dry run, at a few git calls per unmerged branch. That is acceptable, and the docstring says so.
• No tools/budgets.json change. The +21 lines look reasonable for the feature.