[2023Anita]

Summary

wt list showed a stale 🤖/💬 when a branch was deleted outside wt (git worktree remove, git branch -D) and the same name was created again. The marker lives at worktrunk.state.<branch>.marker, keyed only by branch name, so the key survived the deletion. #4684 already clears that key when wt itself deletes the branch.

This is the outside-wt recreate case #4684 left open, and only that:

• wt list skips a marker when git's branch reflog says the name was created after the marker's set_at (delete-and-recreate).
• The creation time comes from git log -g --date=unix (the oldest branch: Created from … entry), so it works on any ref backend that git records a reflog for. No reflog file parsing.
• A set_at of 0 (the documented hand-written form) keeps the marker and returns before running git log -g, so hand-written markers cost no extra subprocess.
• The config key is left in place. wt config state marker get still returns it, and marker clear still removes it.
• No new command or flag, and no new guide prose. worktrunk.state.<branch>.vars.* is unchanged.

Limits

• There is no missing-branch check. Git refuses branch -D on a branch that a worktree has checked out. An unborn --orphan branch has no ref yet. A deleted-but-not-recreated name keeps its marker.
• When git has no creation entry the marker still shows, since that is not evidence the name was reused. That covers an expired reflog, an unborn --orphan branch, and bare repos, where core.logAllRefUpdates is off by default. In a bare repo the fix only applies when reflogs are enabled.

Whether this or a wt config state prune command (suggestion 3) is the better approach is still open on #4683 for a maintainer.

Reference #4683

Test

Integration: test_list_hides_marker_recreated_outside_wt sets a marker, deletes the branch with plain git, recreates it, and checks wt list no longer shows the old marker.

Unit tests cover a recreated branch, unknown branch age (set_at: 0, an expired reflog, and reflogs off), an unborn --orphan branch, a bare repo with and without a recorded reflog, and picking the oldest Created from entry from git log -g output.

[worktrunk-bot]

Bare repos still show the old marker. In a bare repo git leaves core.logAllRefUpdates off by default, so logs/refs/heads/<branch> is never written. branch_created_at then returns None and the marker shows. Repos using the reftable backend behave the same way, because their reflogs live inside the tables. I reproduced it with this branch's build: in a bare repo with linked worktrees, git worktree add -b feature, wt config state marker set 🤖, then git worktree remove + git branch -D feature, then recreate feature. wt list still shows 🤖 on the new feature. The bare layout is one Worktrunk supports directly, so the #4683 repro still happens there. The new help sentence says it doesn't. The inline suggestion drops that sentence. AGENTS.md → Documentation also says "A correctness fix that preserves usage needs no new guide prose".

The fix parses git's on-disk reflog format. branch_created_at hard-codes the files-backend path, and creation_timestamp_from_reflog_line re-implements git's reflog line format (zero OID, tz validation, committer-name splitting). AGENTS.md → "Other Tools Own Their Conventions" asks to query the tool instead of encoding its file format. Git answers this directly, and on every ref backend: git log -g --date=unix --format='%gd %gs' refs/heads/<branch> prints entries newest first, so the last one is <branch>@{<unix>} branch: Created from …. That costs a subprocess per worktree row, which is the trade the PR avoided. Whether that cost is acceptable is part of the decision below.

The missing-branch check can't fire for wt list. compute_user_marker runs only for worktree rows, and git refuses branch -D on a branch checked out in any worktree. So a row's branch always exists, apart from an unborn --orphan branch, which has no ref yet and would wrongly lose its marker. All of the PR's effect comes from the reflog comparison.

This needs a maintainer decision. Altogether this is roughly 90 lines of production code plus a format parser, and it only helps non-bare repos on the files backend. The earlier comment on #4683 left the choice between suggestion 2 (this PR) and suggestion 3 (wt config state prune) to a maintainer. A prune command, or documenting marker clear --branch, works the same way on every layout. I'm holding the approval for that choice.

[worktrunk-bot]

The PR description no longer matches the code. It still says wt list "skips a marker when no local branch has that name", and that unit tests cover "a missing branch" and "an empty reflog". This push removed that check and those tests. A deleted-but-not-recreated name now keeps its marker (gone.user_marker(Some("feature")) asserts Some("🤖")), and bare repos keep it unless core.logAllRefUpdates is on. A maintainer deciding between this and a prune command will read the description first, so please update it to match.

The approach question from the earlier review is still open on #4683, so I'm not giving an approval yet.

[max-sixty]

Thanks for working on this. Closing: it needs a git log -g per marked row on every render, including the statusline, and matches git's reflog message text, which is more than this case is worth. wt now clears a marker whenever it deletes a branch (#4684), and we'll also clear it when wt switch --create creates one. A branch deleted and recreated with plain git keeps its marker until wt config state marker clear. @worktrunk-bot please open that PR, referencing #4683.

This was written by Claude Code on behalf of max-sixty

[worktrunk-bot]

Opened #4704.

[2023Anita]

Updated the description to match the current code: the missing-branch check is removed, bare repos are only covered when reflogs are on, and the approach choice vs. a prune command remains with the maintainer on #4683.