You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[Bug]: Workspace deletion leaves an undeletable record when git worktree remove partially fails (metadata_purge skipped, retries refused as unregistered) #21763
Short summary: When git worktree remove --force partially fails during workspace deletion, Orca keeps the workspace record but git has already dropped its registration. Every retry is then refused with Refusing to delete unregistered worktree path, so the card can never be removed from the UI. The card is labelled Detached HEAD @ <sha> because no branch resolves for it.
This is the same error string as #5864 (closed as fixed in v1.4.134 by #5888), but the cause is different: in #5864 the guard fired spuriously on a registered path; here the guard fires correctly — the path really is unregistered — and that is exactly the state Orca needs to be able to clean up. Related: #10629 (registration dropped, files left behind, Windows), #10951 (ghost worktreeMeta rows), #17654 (a different "cannot delete" dead end).
What the trace shows
worktree.remove runs watcher_gate → pty_sweep → git_remove (trash_rename, branch_delete) → metadata_purge → cache_invalidation. When git_remove fails, the parent span fails and metadata_purge never runs. From main.trace.ndjson (SSH execution host, the worktree lives on a 9p bind mount of a Windows folder inside a dev container):
time
path
error
11:20:48
/workspaces/<repo>/244-2
Request "git.removeWorktree" timed out after 30000ms
19:32:19
/workspaces/<repo>/seacucumber
Command failed: git worktree remove --force … → error: failed to delete '…': Directory not empty
19:32:55
same (retry)
Refusing to delete unregistered worktree path: …
19:33:24
same (retry)
Refusing to delete unregistered worktree path: …
20:53:12
/workspaces/<repo>/npm
Request "git.removeWorktree" timed out after 30000ms
After the first seacucumber attempt, git worktree list no longer showed the path (registration gone), but 29,335 files remained on disk. Orca's record survived because metadata_purge was skipped, and from then on the guard blocked every retry.
How to reproduce
Add a repo on an SSH execution host where the checkout is a bind mount (Docker Desktop 9p/drvfs mount of a Windows folder works). Create a workspace inside that mount (set the host's workspace directory to the mount or its parent with nested workspaces on).
Make the directory undeletable by git worktree remove --force from inside the container: e.g. a node_modules tree with files held open on the Windows side, or a root-owned file.
Delete the workspace from the UI → fails with Directory not empty. git worktree list no longer lists it.
Delete again → Refusing to delete unregistered worktree path. Repeat forever.
A 30 s git.removeWorktree RPC timeout on a large tree (steps 1–3 with a big node_modules over 9p) ends in the same state.
Expected
Treat partial success as partial success. Once git's registration is gone, run metadata_purge and report the leftover directory as a separate warning instead of keeping a record that points at nothing.
"Not registered in git" should not block deletion. It is the normal end state after a half-failed remove; the right action is to drop Orca's record (and optionally offer to delete the leftover directory).
Always offer an escape hatch — "Remove from Orca (keep files)" — so a record can never become permanent.
The fixed 30 s timeout for git.removeWorktree is too short for large worktrees over SSH/9p; scale it or make it long-running with progress.
Evidence (trace spans, config snapshot, container git worktree list --porcelain) is archived and can be attached on request.
Confirmed still reproducible on Linux / Orca 1.4.239, with a different trigger than the Windows/9p case reported here: a root-owned file inside the worktree makes git worktree remove fail with Permission denied instead of Directory not empty.
Trace (local execution host, plain local checkout — no bind mount)
2026-10-28T14:48:41 worktree.remove FAIL
worktree.remove.git_remove FAIL
git worktree remove <workspaces>/<repo>/<worktree>
error: failed to delete '<worktree-path>': Permission denied
2026-10-28T14:48:53 worktree.remove FAIL "<worktree-path> is no longer a Git worktree, so its folder was left on disk. Delete the folder yourself to finish removing this workspace."
2026-10-28T14:49:14 worktree.remove FAIL (same)
2026-10-28T14:49:23 worktree.remove FAIL (same)
2026-10-28T21:32:59 worktree.remove FAIL (same)
2026-10-29T09:33:17 worktree.remove FAIL (same)
A second workspace reproduced it independently at 2026-10-30T08:51:47 with the same Permission denied, after a root-owned dist/ directory was produced inside it. In both cases git had already dropped the worktree's admin entry under .git/worktrees/<id> before failing, so the checkout was no longer a git worktree while its directory remained on disk.
Differences from the original report
The error string changed in 1.4.239. The guard now says "is no longer a Git worktree, so its folder was left on disk. Delete the folder yourself to finish removing this workspace." (emitted from out/main/chunks/orcad-migration-source-retention-*.js), rather than "Refusing to delete unregistered worktree path".
--force does not help.orca worktree rm --worktree <path> --force fails with the identical error.
The message is a dead end. It instructs the user to delete the folder manually, but there is no supported way to finish the removal inside Orca afterwards:
orca worktree rm against the same selector then returns selector_not_found (an error) rather than success.
The workspace record disappears only via a background reconcile (a git worktree list pass) a few seconds later, so scripted and CLI flows get no success signal at all.
The trigger is common on Linux: any root-owned file inside the worktree is enough — Docker or sudo builds and CI helpers produce these routinely. In both of my cases the only offending path was a dist/ directory owned by root.
Expected (same as the original report)
Treat partial success as partial success: once git's registration is gone, run metadata_purge and report the leftover directory as a separate warning instead of keeping a record that points at nothing.
"Not registered in git" must not block deletion — it is the normal end state after a half-failed remove.
Always offer an escape hatch — "Remove from Orca (keep files)" — so a record can never become permanent.
Activity
AmethystLiang commentedon Sep 20, 2026
thanks for reporting. looking into it
fix(worktrees): finish SSH worktree deletes that git left half done
MatiasOlivera commentedon Oct 30, 2026
Confirmed still reproducible on Linux / Orca 1.4.239, with a different trigger than the Windows/9p case reported here: a root-owned file inside the worktree makes
git worktree removefail withPermission deniedinstead ofDirectory not empty.Trace (local execution host, plain local checkout — no bind mount)
A second workspace reproduced it independently at
2026-10-30T08:51:47with the samePermission denied, after aroot-owneddist/directory was produced inside it. In both cases git had already dropped the worktree's admin entry under.git/worktrees/<id>before failing, so the checkout was no longer a git worktree while its directory remained on disk.Differences from the original report
out/main/chunks/orcad-migration-source-retention-*.js), rather than "Refusing to delete unregistered worktree path".--forcedoes not help.orca worktree rm --worktree <path> --forcefails with the identical error.orca worktree rmagainst the same selector then returnsselector_not_found(an error) rather than success.git worktree listpass) a few seconds later, so scripted and CLI flows get no success signal at all.root-owned file inside the worktree is enough — Docker or sudo builds and CI helpers produce these routinely. In both of my cases the only offending path was adist/directory owned byroot.Expected (same as the original report)
metadata_purgeand report the leftover directory as a separate warning instead of keeping a record that points at nothing.Version matrix
Requesting the
os:linuxlabel: this reproduces on Linux with a plain local checkout and no bind mount.