Repository navigation
Git hygiene reconciliation: push finished work, PR orphaned branches, classify dirty trees — and make it a standing sweep #1371
Description
Problem (audit 2026-10-27, 188 working trees / 26 repos)
The fleet's publish step evaporates the same way its claims used to: work finishes in a worktree and the push/PR step never happens.
- 62 commits on no origin branch, across 37 trees — almost all finished queue/sweep work (worst:
wt-1058, 10 commits). Includes task 1149's completed lab-ops-hub: judge container crawl hosts on the full UTS-46 mapped form (#516, #526) #530 branch commit. - 13 trees with uncommitted work (139 entries) — worst: the home repo (
~): 98 dirty files and the fleet's own Supabase migration commits 0009–0012 unpushed. - 143 GitHub branches ahead of main with no open PR (code 66, opencode 44, openchamber 19, nix 14) — a mix of Jules debris, bot branches, and real feature work (opencode
session-messaging+44,github-tools+33). - 289 fully-merged branches still on the remotes as cleanup candidates.
- Hazard: task 1157's cursor-quota work sits dirty in TWO worktrees with divergent heads (
~/git/devtoday-openchamber-wt-1157,~/work/oc-wt-1157). - Primary clones are healthy (code/openchamber/nix clean+pushed; opencode 2 behind). One fleet submodule fetch timed out during the audit — one tree's status unverified.
This effort
- Push every already-committed local-only branch (no force-push; rejections recorded, not fought).
- Open PRs for finished queue-task branches whose work is complete and whose issue is still open (body notes the reconciliation origin); skip branches whose issue is closed or whose PR already merged.
- Classify — do not commit — every dirty tree: in-flight lane work vs abandoned vs the divergent 1157 pair (flagged for a decision, untouched).
- Classify the 143 stray GitHub branches (fleet work / Drew's features / bot debris) on this issue; no mass PRs, no deletions in this pass.
- Design the standing version: a git-hygiene sweep job (daily) that reconciles dirty/unpushed trees, pushes finished queue branches, opens their PRs, and flags divergent duplicates — so this does not decay again. File its build as a sub-issue when designed.
Acceptance
- Zero local-only commits remain outside an explicit, recorded exception list.
- Every finished queue branch has a PR or a recorded reason it should not.
- The home repo's migration commits are pushed.
- Dirty-tree + stray-branch classifications recorded here; the 1157 pair has a named resolution path.
- Standing hygiene sweep designed (sub-issue filed).
Filed by Muse (operator) 2026-10-27 at Drew's direction. Parent: fleet reliability epic #1233. Related: #1130, #1131.
Activity
developing-today commentedon Oct 27, 2026
CODE-467
dezren39 commentedon Oct 27, 2026
Standing job built (Drew, 2026-10-27):
git-hygiene-sweep, every 4h — instruction at hidden_files/git-hygiene-sweep-instruction.md in the fleet goal workspace. Two duties: (A) publish reconciliation — push committed fleet branches (queue//sweep//creator/ + home-repo mains), PR finished queue branches whose issue is open (≤10/run), classify dirty trees without ever committing them, flag divergent duplicates; (B) idle build-artifact cleanup — target/build dirs in worktrees idle ≥4h (no mtimes, no lane lock, no live build process), realpath-verified inside their tree, moved to recoverable trash (never rm -rf, never wildcards, symlink escapes skipped+reported), ≤20 dirs/run, primary clones report-only. The one-time reconciliation of the audit backlog is running now as the first pass under this issue.dezren39 commentedon Oct 27, 2026
Reconciliation — Phase 1 complete: pushes (resumed pass, 2026-10-27)
A previous worker died mid-job; this pass resumed from verified state. Method note: the audit's "local-only" metric compared against stale remote-tracking refs — verified against live
ls-remote/GitHub API, most branches were already on origin, and many "missing" branches turned out to be post-merge leftovers (branch deleted upstream when its PR merged). Pushing those would resurrect deleted branches, so they were verified and left alone. The standinggit-hygiene-sweepjob's first run overlapped this pass; its pushes are included below.Pushed / verified on origin:
debug/fod-ci-capture(5 commits, no PR existed), openchamberfix/fod-node-modules-determinism(PR wave-2 workflow: PR #486 templates and PR #490 tool disagree on the generated-view contract — a template-built design.md can never pass check (run-table vs run-status) #491 was closed unmerged).queue/*branches, nixcreator/584-trailing-dot-host, nixmcpx-buildvcs-off, nix-cachemain, fleet code-writersfeat/usage-snapshot-writers, opencodequeue/1149-530-governance-expansion,queue/1334-561-identifier-prefix,sweep/546-env-file-chain, and the bulk of opencode sweep/creator branches (64 opencode local branches match origin exactly).Post-merge leftovers — deliberately NOT pushed (remote branch deleted at merge; local head equals the merged PR head or is patch-equivalent in main):
Exceptions (recorded, not fought):
check/typecheckover the pushing worktree and it fails in every worktree tried (main clone, wt-1128, wt-535/538/540 via the sweep). Verbatim:error: script "typecheck" exited with code 1 · error: script "check" exited with code 1 · husky - pre-push script failed (code 1). Blocked by this:queue/1128-526-quota-reset-decimal-compound(genuinely unpublished finished work for OPEN issue opencode#526, no PR — highest-value exception),creator/235-roadmap-done-229-232(docs commit, no PR),ci-baseline-check(PR 469 closed unmerged).bolt-cache-prosemirror-coords-at-pos-5260428214093938082(Jules branch; origin moved on), opencodecreator/120-wrangler-state,creator/141-service-turbo-tests,creator/153-lint-rule-tests.Home repo migrations (0009–0012) and hatch-personal were already pushed before this pass (verified by parent) — not redone.
dezren39 commentedon Oct 27, 2026
Reconciliation — Phase 2 complete: PRs for finished queue branches
Opened (1):
queue/1308-539-bench-harness→ Refs id(ci): F11.8 benchmark harness with committed baselines #539 (issue OPEN, no PR existed). Head commit self-describes as "wip slice 1", flagged in the PR body for the reviewer.Skipped — PR already exists (the common case; the work was never as orphaned as stale refs suggested):
Could not PR (1): opencode
queue/1128-526-quota-reset-decimal-compound— finished work for OPEN opencode#526, but the branch cannot be pushed (pre-push hook failure, Phase 1 exceptions), so no PR can be opened from it yet.Cap of 25 PRs not approached (1 opened). No overflow.
dezren39 commentedon Oct 27, 2026
Reconciliation — Phase 3 complete: classifications (nothing touched)
Dirty trees (fresh enumeration; audit found 13 trees / 139 entries)
/home/hatch[main] — 98 dirty — stale/operator: accumulated tooling + docs edits in the home repo working tree. Not lane work; needs an operator commit pass. Untouched.~/git/wt-quota-visibility[feat/quota-current-visibility] — 6 dirty — finished-uncommitted: the quota-visibility feature (aggregator config/model/render/sources + test), complete-looking but never committed.~/git/fleet[main] — 9 dirty — stale: submodule pointer drift + untracked submodule dirs in the superproject.~/git/fleet/wt-1160-500[queue/1160-500-ipv6-host-parsing] — 2 dirty — finished-uncommitted queue work (request-security.js + test).~/git/wt-1368[creator-b/1368-collab-reconnect-cap] — 2 dirty — in-flight (today's task; collab.ts + new collab.test.ts).~/git/devtoday-code[main] — 1 dirty — in-flight: collab.ts edit on the primary clone, pairs with wt-1368's work.~/git/wt-cold-store-step-isolation— 1 dirty — in-flight (export-turso.sh).~/git/wt-1116[queue/1116-968-nested-marks-triage] — 1 dirty — finished-uncommitted (markdown.rs from a completed triage task).~/git/wt-aa-explorer[pm/aa-explorer] — 1 dirty — generated churn (fleet-feed.json).~/git/pm-feed[main] — 1 dirty — generated churn (fleet-feed.json).~/git/nix-cache[main] — 2 dirty — stale (bench flake.locks).~/hatch-personal[main] — 0 dirty — clean (was already pushed; verified).⚠ Task 1157 pair — divergent copies, decision needed, contents untouched
~/git/devtoday-openchamber-wt-1157[queue/1157-495-cursor-windowseconds] — head93f7a18e6(2026-10-27 03:46, "fix(quota): Cursor windowSeconds is the billing-cycle duration, not time until reset (fix(bin): validate ANTIGRAVITY_LS_ADDRESS port strictly (#494) #495)"), tree clean. This is the live copy — it backs open PR openchamber#537.~/work/oc-wt-1157[same branch name] — head501f171(2026-10-27 01:02, an older main merge), with uncommitted modifications to the same two files (packages/web/server/lib/quota/providers/cursor.js, cursor.test.js).GitHub branches ahead of main with no open PR (fresh count: 168)
Branch totals grew today (code 129, opencode 102, openchamber 159, nix 112) partly because Phases 1 pushed ~25 branches. Caveat: GitHub's ahead-by counts commits unreachable from main, so squash/rebase-merged fleet branches register as "ahead" even though their content is in main — the fleet bucket below is mostly that residue, not lost work.
full-platform+12,feat/openclaw-summarize+4.session-messaging+8 (audit said +44; main has since absorbed most of it),tui-nested-asks+13,github-toolsnow +0 (fully landed),fix/plugin-host-bun-resolve-messagenow +1 (was +37).nested-asks+7,feat/session-messaging+6,debug/fod-ci-capture+5 (pushed in Phase 1; no PR — fod investigation branch). One compare error:pr-evidence/session-messagingshares no ancestor with main (orphan evidence branch).No mass PRs opened and no branches deleted, per the effort's scope.
dezren39 commentedon Oct 27, 2026
Reconciliation — final totals (2026-10-27)
Phase 1 (pushes): 23 branches pushed across the two passes (21 by the standing sweep's first run, 2 by this pass: openchamber
debug/fod-ci-capture,fix/fod-node-modules-determinism); dozens more verified already-on-origin against live refs. All remaining local-only commits fall in three explicitly recorded categories (Phase 1 comment): post-merge leftovers whose content is in main (59 branches, named), diverged branches left unforced (4, named), and opencode branches blocked by the repo's failing husky pre-pushcheck(3, named, verbatim error recorded) — of those, onlyqueue/1128-526…carries genuinely unpublished feature work (opencode#526).Phase 2 (PRs): 1 opened — code#1387 for queue/1308-539 (Refs #539). Every other finished queue branch already had a PR (mostly MERGED) or a recorded skip reason; queue/1128 could not be PR'd because its push is hook-blocked.
Phase 3 (classification): 12 dirty trees classified line-by-line; the task-1157 divergent pair flagged with a resolution path (the clean, PR-#537-backed ~/git copy is live; the ~/work copy's uncommitted cursor edits need a keep/discard decision). 168 GitHub branches ahead-with-no-open-PR classified by repo and kind, with the squash-merge measurement caveat recorded.
Acceptance: boxes 1–4 checked. Box 5 (standing sweep sub-issue) left unchecked: the
git-hygiene-sweepjob itself is built and running (see the earlier comment on this issue; its runs overlapped and assisted this pass), but I found no sub-issue filed for it — filing/confirming that is the one remaining clerical step.dezren39 commentedon Nov 2, 2026
agent-note:v1
event: done
task: 1352 (repo-health-code-pilot, repo-worker-code, lane code-d)
state: triaged — routing decided, closing #1371 on its merits
verified (fresh, 2026-11-02): the one-time reconciliation this issue asked for was completed 2026-10-27 — Phase 1 pushes, Phase 2 PRs (PR #1387 MERGED 2026-10-27T19:49Z), Phase 3 classifications, final totals; acceptance boxes 1–4 checked on this issue. The standing mechanism exists and is live: cron git-hygiene-sweep enabled, every 4h, last run today 15:27 CST succeeded, next 19:05 CST; instruction at hidden_files/git-hygiene-sweep-instruction.md (fleet goal workspace).
routing: the ONLY open acceptance item was box 5 (standing-sweep sub-issue) — now filed as #4604, the standing tracking issue for the sweep. Remaining classified items from Phase 3 (operator commit pass for /home/hatch dirt, the task-1157 divergent-pair decision, the husky-blocked opencode queue/1128 push for opencode#526) carry their resolution paths in the Phase 1/3 comments and belong to the standing sweep classification duty / their owners — no new queue packet is needed from this triage.
next: ongoing hygiene = git-hygiene-sweep (tracking #4604). Closing #1371: work complete, standing owner named.