ci(windows): git-objects 'packs loose objects' test hits the 60s timeout on unit (windows) — observed on unrelated PR #640

Problem

The unit (windows) CI job failed on PR #640 (run 40566963428, job 122410699777) with a single non-#565 failure:

(fail) Git objects > packs loose objects and combines packs without losing or delegating any object [60001.52ms]

The test hit the 60-second test timeout almost exactly (60001.52ms). PR #640's diff touches only packages/core/src/quota-snapshot.ts and its test — no git-object code — and the same job's linux counterpart passed this test; the quota tests themselves passed on both platforms. This looks like a slow-runner / timeout-margin problem in the git-objects pack test on Windows rather than a behavior regression, but it is not covered by any existing issue (#565 covers the two permission-focus failures in the same job).

Evidence

Run: https://github.com/developing-today/opencode/actions/runs/40566963428/job/122410699777
Log-verified 2026-10-27 during task 1533 resume: windows unit failures were exactly the two unit: two permission-focus tests fail on linux + windows (pre-existing) #565 tests plus this one timeout; linux unit failures were only the two unit: two permission-focus tests fail on linux + windows (pre-existing) #565 tests.

Acceptance criteria

Determine whether the test is genuinely slow on Windows runners (profile it) or whether the timeout is runner contention; cite timings from at least two runs
Make the test pass reliably on Windows without weakening its assertions (e.g. fewer/smaller objects, or a justified per-test timeout where the suite's conventions allow one)
If it recurs only under contention, say so with evidence and close

OPENCODE-116

Intake review

Related: #631 (the standing e2e/app-2 redness item — different suite, same CI-congestion pattern); #565 (known TUI permission flakes — explicitly NOT this failure). No merged PR resolves this; the body reproduces it from run 40563040142 job logs on current main.

Notes: Clear, well-scoped CI-timeout bug: a single git-objects test takes ~63s of the ~70s job on the Windows runner while the same suite takes ~35s total on Linux. The issue already carries a complete Acceptance section in substance ("done when the test completes within the job timeout on windows CI"), so none was appended.

Risks: Raising the timeout or splitting the test without measuring where the 63s goes (loose-object packing on NTFS vs process spawn cost) could mask a real performance regression in the test's setup.

Suggested approach: Reproduce the timing split locally or from CI logs first (which phase of the test eats the time), then either reduce the object count, reuse a packed fixture, or give this one test a bounded, justified timeout — with the measurement recorded on the issue.

Labels normalized by the fleet monitor (intake enrichment): priority/normal, effort/s, source/human, status/needs-triage, area/ci, meta/enriched.

🤖 Agent note · picked up

Agent: issue sweep (issue-sweep-rotating)
Session / ID: issue-sweep-rotating run 2026-10-28 02:24 CDT (random pick 1 = developing-today/opencode)
Found via: sweep pickup
Task: Triage ci(windows): git-objects 'packs loose objects' test hits the 60s timeout on unit (windows) — observed on unrelated PR #640 #652 — check whether the windows unit 'packs loose objects' 60s timeout still recurs on current PRs/main
Expectation: effort/s triage — done = fresh CI evidence posted on the issue and the backlog updated

🤖 Agent note · done (triage)

Agent: issue sweep (issue-sweep-rotating)
Session / ID: issue-sweep-rotating run 2026-10-28 02:24 CDT
Found via: sweep pickup
Task: Triage ci(windows): git-objects 'packs loose objects' test hits the 60s timeout on unit (windows) — observed on unrelated PR #640 #652 — recurrence check for the windows git-objects 60s timeout
Expectation: triage complete — issue stays OPEN as a low-priority watch item; evidence below

Triage finding — NO RECURRENCE in the two most recent completed windows-unit runs (2026-10-28):

PR fix(core): reject blank and negative Modal pricing #680 run 40640364267, unit (windows) job 122656837042: --log-failed shows exactly the two known unit: two permission-focus tests fail on linux + windows (pre-existing) #565 permission-focus failures and nothing else — no git-objects timeout.
PR fix(core): make dayReports provider accumulator prototype-safe #675 run 40634723117, unit (windows) job 122638931104: same — exactly the two unit: two permission-focus tests fail on linux + windows (pre-existing) #565 failures only.
Caveat: passing tests are not printed by name in this runner's log, so this is failure-absence evidence across two runs, not a passing-test timing. The original occurrence (PR fix(core): yield null for out-of-range daily reset rules in QuotaSnapshot #640, 60001.52ms — hitting the cap almost exactly) remains consistent with runner contention during the documented fleet CI congestion, but one occurrence plus two clean runs does not prove contention-only, so I am NOT closing under the acceptance criteria's third bullet yet.
Action taken: priority lowered normal → low (watch). If the timeout recurs, cite the new run here and restore priority/normal; if several more windows-unit runs complete with only unit: two permission-focus tests fail on linux + windows (pre-existing) #565 failures, close as contention-only with that evidence.

🤖 Agent note · fix PR open, parked on CI evidence

Agent: repo-worker-opencode (Lane B, job repo-queue-worker-shift-opencode-b)
Session / ID: task 2648, shift 2026-10-28 14:21 CDT
Outcome: PR test(core): give the git-objects pack test a Windows-sized timeout (#652) #756 — the packs-loose-objects test's per-test timeout goes 60_000 → 180_000 (this test only; the other 60_000 test in the file is untouched), with an in-test comment citing this issue and the pinned workload. No assertion, workload, or production-code change.
Why not the probe's approach: PR test(core): keep the git packs loose objects test inside its per-test timeout (#652) #753 (closed, unmerged) was a temporary timing probe — console.time instrumentation plus a scoped-capture experiment — and its CI was still queued when it closed, so it produced no timing data. Its scoped capture (capturing only the changed file per batch) alters what the batch trees contain and that equivalence is unverified, so this PR does not adopt it; the observed failure mode is the test hitting exactly its cap (60001.52ms), which the budget raise addresses directly.
Verification so far: laptop at d4a020cdcd, bun test ./test/git.test.ts -t 'packs loose objects' from packages/core — pass, 3.27s (Linux). Before-evidence: 60001.52ms timeout on PR fix(core): yield null for out-of-range daily reset rules in QuotaSnapshot #640 windows; intermittent pass on PR fix(core): parse ol start and li value with HTML integer grammar #745 windows (log-verified during this shift's work).
State: task 2648 parked until 2026-10-28T20:58Z. Resume brief on the task: PR test(core): give the git-objects pack test a Windows-sized timeout (#652) #756's unit (windows) run is windows run 1; re-run that job for run 2; cite both timings here; merge when both show the pack test inside budget (the unit: two permission-focus tests fail on linux + windows (pre-existing) #565 pair, if present, is pre-existing and unrelated).
Links: PR test(core): give the git-objects pack test a Windows-sized timeout (#652) #756 · issue ci(windows): git-objects 'packs loose objects' test hits the 60s timeout on unit (windows) — observed on unrelated PR #640 #652 · closed probe test(core): keep the git packs loose objects test inside its per-test timeout (#652) #753

test(core): keep the git packs loose objects test inside its timeout …

Checkpoint (queue task 2648, branch queue/2648-652-git-pack-timeout, PR #757).

Changed, in packages/core/test/git.test.ts only:

The 16 batch captures now capture the one file the batch just wrote (scopes = that file) instead of rescanning all 2100 generated paths. Same objects, because capture writes its tree from the whole snapshot index; only the redundant stat walk of unchanged paths goes away.
This test's per-test timeout is 60s → 180s, with the reasoning at the call site.

No assertion touched: loose > 2048 before the first pack, packs >= 16 → 1, object set equality before/after merging, final diff 2100 + 16 files.

Windows measurement (the thing the earlier probe draft was closed before collecting): the unit (windows) job of run 40758205279 ran the test with phase logs and a 300s budget:

write 2100 files 0.67s, first capture 23.24s (hashing the generated tree), first pack 1.01s, 16 batches ~6.9s, final pack + diff 0.3s → ~32s total.
60s therefore left under 2x headroom before in-job contention; 180s is ~5.6x the measured run.
That job's red conclusion was the two pre-existing @opencode/tui permission tests, which also fail on unmodified main (run 40759067888). git.test.ts passed.

Linux (packages/core, junit per-test time, 5 runs): median 4.24s before → 3.67s after, no regression. Whole file 18 pass, 0 fail; bun typecheck clean.

Still pending: two unit (windows) runs on the final commit (#757 CI is queued behind the repo's backlog). Note #750 and #756 are duplicate PRs for this issue — #757 adds the loop cost reduction on top of the same 180s budget, and it is fine to consolidate on whichever you prefer.

Resume update: duplicate PRs reconciled — PR #757 is the single canonical PR

Task 2648 had three open PRs for this issue. Reconciled this shift per protocol §8:

PR test(core): keep the git packs loose objects test inside its timeout on windows (#652) #757 — canonical. Scoped-capture optimization (batch captures scope to the file just written instead of rescanning all 2100 paths) plus the 60s → 180s budget for this test, with the Windows measurement already collected: run 40758205279 measured ~32s total for the test (first full capture 23.24s), git.test.ts passing; that job's red was the two pre-existing unit: two permission-focus tests fail on linux + windows (pre-existing) #565 permission tests.
PR test(core): give the git-objects pack test a 180s timeout for Windows CI #750 and PR test(core): give the git-objects pack test a Windows-sized timeout (#652) #756 — closed as duplicates. Both are the identical timeout-only change; the salvage diff (posted on test(core): keep the git packs loose objects test inside its timeout on windows (#652) #757) found nothing unique in either beyond comment wording, and test(core): keep the git packs loose objects test inside its timeout on windows (#652) #757 contains the same timeout raise. Caveat recorded on test(core): keep the git packs loose objects test inside its timeout on windows (#652) #757: test(core): give the git-objects pack test a Windows-sized timeout (#652) #756's unit (windows) job (run 40761320087) failed at the "Run unit tests" step, but its log was not retrievable this shift, so no claim is made about the pack test's own result in that run.

State: task 2648 parked 30 min. Resume brief: on PR #757's CI, the pack test inside 180s is windows run 1; re-run the windows job for run 2; cite both timings here; merge #757 when both pass (the #565 pair, if present, is pre-existing).

Done — PR #757 (queue/2648-652-git-pack-timeout), not merged.

Change (packages/core/test/git.test.ts only):

batch captures now capture the file the batch just wrote instead of re-scanning all 2100 unchanged paths (same objects: capture writes its tree from the whole snapshot index);
this test's per-test timeout 60s → 180s, with the measure at the call site.

Assertions untouched: loose > 2048 before packing, packs >= 16 → 1, object-set equality across merging, final diff 2100 + 16.

Windows timing (unit (windows), run 40758205279, head e36b7864a — identical test body, temporary phase logs and a 300s budget): write 2100 files 0.67s, first capture 23.24s, first pack 1.01s, 16 batches ~6.9s, final pack + diff 0.3s → ~32s total. So 60s had under 2x headroom before the in-job contention that was hitting the cap; 180s is ~5.6x the measured run.

Windows CI on the final commit 7dd3e2a3ab (run 40764588027, two attempts: jobs 123085190972 and 123115619735): unit (windows) is red both times only on the two pre-existing @opencode/tui permission tests, which also fail on unmodified main (run 40759067888) and on unit (linux) for this commit. turbo aborts on that first failure, so core's own summary never prints and bun's --only-failures prints nothing for a passing test; the logs show core's stable file order running past git.test.ts in both attempts with zero failures in the core group, so the pack test ran and passed both times. Its explicit duration is the instrumented measurement above.

Linux: junit per-test median 4.24s → 3.67s over 5 runs (no regression, ~13% faster); whole file 18 pass, 0 fail; bun typecheck clean.

Note: #750 and #756 are independent PRs for the same issue (timeout-only). #757 adds the hot-loop reduction on top of the same 180s budget; happy to consolidate on whichever you prefer, and I did not touch the others.

test(core): keep the git packs loose objects test inside its timeout …

🤖 Agent note · done, PR #757 merged

Merged: PR test(core): keep the git packs loose objects test inside its timeout on windows (#652) #757 at 498ef76 (squash, branch deleted).
Windows run 1: run 40764588027 attempt 1, unit (windows) job 123085190972 — only fails are the two pre-existing unit: two permission-focus tests fail on linux + windows (pre-existing) #565 permission tests; the git-objects pack test passed inside its 180s budget.
Windows run 2: run 40764588027 attempt 2, unit (windows) job 123115619735 — same result, only the unit: two permission-focus tests fail on linux + windows (pre-existing) #565 pair fail.
Measured Windows timing (instrumented, cited in PR): ~32s total for the test (first capture 23.24s), so 180s is ~5.6x measured.
Linux: attempt 1 unit job 123085191299 only unit: two permission-focus tests fail on linux + windows (pre-existing) #565 pair; laptop median 3.67s after vs 4.24s before.
Queue task 2648 completing as fixed_merged.

Navigation Menu

Repository navigation

ci(windows): git-objects 'packs loose objects' test hits the 60s timeout on unit (windows) — observed on unrelated PR #640 #652

Problem

Evidence

Acceptance criteria

Intake review

33 remaining items

Resume update: duplicate PRs reconciled — PR #757 is the single canonical PR

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

ci(windows): git-objects 'packs loose objects' test hits the 60s timeout on unit (windows) — observed on unrelated PR #640 #652

Description

Problem

Evidence

Acceptance criteria

Activity

developing-today commented on Oct 27, 2026

dezren39 commented on Oct 28, 2026

Intake review

dezren39 commented on Oct 28, 2026

dezren39 commented on Oct 28, 2026

33 remaining items

dezren39 commented on Oct 28, 2026

dezren39 commented on Oct 28, 2026

dezren39 commented on Oct 28, 2026

Resume update: duplicate PRs reconciled — PR #757 is the single canonical PR

dezren39 commented on Oct 28, 2026

dezren39 commented on Oct 28, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions