Findings: branch lifecycle (oak 0.96, 2026-06-11)
Observed while reviewing/merging the accuracy-push branches on oak/benchmarks. All reproduced from fresh clones; verification commands inline.
1. oak close never propagates to the server (P0)
oak close <name> and oak close (current branch) both print
✓ Closed branch '<name>', and oak branch list in that checkout shows
[closed] — but a fresh clone (and the web UI) still reports the branch
open. Tried from three separate clones, by name and as current branch;
no form propagates. There is no --remote flag and no error.
Verify: oak clone oak/benchmarks /tmp/a && cd /tmp/a && oak close oak clone oak/benchmarks /tmp/b && cd /tmp/b && oak branch list # still open
Consequence: server branch lists accumulate dead branches forever; an agent told to "clean up branches" reports success while changing nothing visible to anyone else. The only working path today appears to be the web UI.
2. Every clone leaves a transient open branch in server listings
Each oak clone immediately creates its auto-named working branch
(<user>-<hex>), and that branch shows open in branch list --json
from other checkouts even when it carries zero commits. Review/CI/probe
clones therefore pollute the branch UI. Zero-commit branches appear to be
garbage-collected eventually (some probe branches vanished from listings),
but the lifecycle is undocumented and listings are noisy meanwhile.
1b. Strengthened evidence for finding 1
oak --verbose close <name> shows zero server interaction (no request is
even attempted), and a subsequent oak push + oak desc (both of which DO
sync) still leave the server reporting open. Close is pure local
bookkeeping in 0.96.
1c. Working CLI removal: equalize-then-merge
Branches whose content predates main conflict on merge ("N file(s) modified
on both branches since fork"). The working pattern: switch to the branch,
overwrite its tree with main's content (rsync, excluding .oak), commit, then
oak merge — the no-op-content merge succeeds and removes the branch
server-side. Used to remove bench-accuracy-push, mrmrs-baf5cf, and
mrmrs-bacba9 on 2026-06-11.
1d. Self-inconsistent poisoned branch (P0)
bench-phase0-fleet-sync (the rename-target of a branch that hit the
false-divergence trap) cannot be removed at all from the CLI: after
oak pull --force syncs the checkout to the server's OWN tip (local ==
remote byte-for-byte), oak merge/push still report "remote has diverged".
The branch's server-side state disagrees with itself; rename appears to have
carried the corruption. Only server-side intervention / web UI can remove it.
3. merge is the only verified server-side branch remover
Merging a branch into its parent (oak switch <branch> && oak merge)
removes it server-side and rotates the checkout onto a fresh branch — used
successfully for the accuracy-push merge (2c45d63d7256). For abandoned
(never-to-merge) branches there is no CLI-visible equivalent; see finding 1.
Related, already filed
False-divergence after concurrent desc/rename/push on a shared branch and
the destructive-only pull --force recovery path are covered in
prompts/oak-fix-handoff.md (P0 section) and were re-hit live during this
work.