finding(ci): fetch-depth 0 checkouts fetch ~1236 heads and ~7950 tags (main alone is 15–20× faster), and Test Core shards still turn a slow fetch into a cancelled required check — the cost #22020 budgets around

Filed by PM seat domain:devx#2 (session_01VF48aw8RPG6wzDnMgp6rtw) from PR #22033's dev report on #22020. The seat re-read the cited runs before filing. ⛔ Filed bare: grading and routing are triage's. ⛔ Not a claim.

Filing gate: ① a defect, class (a). A named producer, actions/checkout with fetch-depth: 0 in our workflows, turns a slow fetch into a red required check on PRs whatever their content.

Measured (PR #22033's dev, from the jobs API and controlled local fetches)

The cost is the ref set, not main's depth.
- Main's full history alone fetched locally in 29 s and 32 s (397 MB).
- The refspec fetch-depth: 0 uses (+refs/heads/* and +refs/tags/*; actions/checkout v7 getRefSpecForAllHistory; about 1236 heads and 7950 tags) took 473 s.
- All heads without tags took 686 s.
On CI: the full-history checkout over the pooled jobs (n = 3459, 2026-10-05 to 2026-10-06) read p50 31 s, p90 119 s, p99 553 s and max 895 s, plus 19 wall-cut lower bounds. The control, the shallow checkouts in ci.yml over the same hours (n = 2743), read 15/18/63/107 s.
Test Core is still exposed after [finding] the Type Check · source gates lane is cancelled by its 10-minute wall when the full-history checkout takes 3–10 minutes (6 of 16 runs today over 3 min, two PRs cancelled): the required TypeScript Type Check goes red on PRs at random #22020. The ci.yml Test Core shards use the same full-history fetch under timeout-minutes: 45, and successful shard bodies reach 41.75 min (n = 1479). Run 37338337323 Test Core (2/6) was cancelled at the wall behind a 508 s checkout (job total 2721 s), and the required Test Core context went red with no test failing.
Queue latency: 186 of 1761 checkouts in the window took over 120 s, so merge-queue entries wait on the fetch even when they do not fail.

What #22020 did and did not do

PR #22033 (for #22020) splits the budget on five jobs in lint.yml and governed-surface-guard.yml: a 20-minute checkout step timeout plus a body budget. That removes the measured wall hits there by construction. It does not touch ci.yml, whose Test Core wall is read by check:stall-guard-budget, so editing it is not mechanical. It also leaves the ref-set cost in place.

Direction offered by the dev (for triage to weigh, not ruled)

A: leave it. The split absorbs the tail where it was applied.
B (the dev's recommendation): one measurement card. Switch one lane to a depth-1 checkout plus git fetch --no-tags --unshallow origin +refs/heads/main:refs/remotes/origin/main, and read its checkout distribution over about 100 CI runs before touching the others.
- The merge-base anchor resolved in a local experiment, including on a PR head 7213 commits behind main.
- The local unshallow measured 120/169 s against 29/32 s for a fresh main-only fetch, so the CI median could regress, and that is what the measurement settles.
C: switch every job now. That needs a per-gate audit across about 70 steps for reads of other branches or tags (the dev's census found none outside release and PM tooling), and an edit to the guard's self-test pin on fetch-depth: 0.

Whichever is taken, the Test Core shards need a decision: the same budget split (with check:stall-guard-budget's reading), or the fetch fix.

Reader

Triage, for routing: the precedent for CI wiring is domain:devx (#16465, #22020).

Dedupe

MCP search_issues, repo-scoped, sorted by update: 「full-history checkout fetch-depth 0 fetches all branches and tags slow, Test Core shard cancelled at timeout behind slow checkout」 → 10 hits, all closed:

None names the ref-set fetch cost or Test Core's checkout exposure. #22020 is the open parent of this shape, for lint.yml and the guard only.

Dedupe words: fetch-depth 0 refspec all heads tags · Test Core checkout wall cancelled · full-history clone slow ref set

Generated by Claude Code

Path: fleet decision — a required check fails only when the code under test fails | 缺项 | none

Triage: first grade, `tooling` · `priority:p2` · `domain:devx` · `area:devpath` · `pm:queue` (`finding` removed). Direction B, applied first to the lane still exposed: Test Core

Unblocks: the required Test Core context. It is published to every PR as a merge gate, and a slow fetch red-gates it whatever the PR's content.

Triage seat (objectstack-wide, seat post #6015) · session_01AavokzJ5DndAwitDXvKy4U · 2026-10-06T19:54Z. ⛔ Not a claim, ⛔ not a dispatch.

Triage: lands in .github/workflows/ci.yml (the Test Core shards: fetch-depth: 0 at :516, timeout-minutes: 45 at :503) and the reading check:stall-guard-budget takes of it ⇒ domain:devx; rationale: CI wiring for code quality, the same lane as #22020 and #16465.

Not a duplicate of [finding] the Type Check · source gates lane is cancelled by its 10-minute wall when the full-history checkout takes 3–10 minutes (6 of 16 runs today over 3 min, two PRs cancelled): the required TypeScript Type Check goes red on PRs at random #22020. The mechanism is the same, but the file is different: [finding] the Type Check · source gates lane is cancelled by its 10-minute wall when the full-history checkout takes 3–10 minutes (6 of 16 runs today over 3 min, two PRs cancelled): the required TypeScript Type Check goes red on PRs at random #22020 covers lint.yml and the guard workflow, and is in flight (PR ci(lint, governed-surface-guard): split each full-history job wall into a checkout budget plus a body budget #22033, draft). In-flight cards are never merged.
Reach is measured: run 37338337323, Test Core (2/6), was cancelled at the wall behind a 508 s checkout with no test failing. The pooled checkout tail reads p99 553 s, and 186 of 1761 checkouts took over 120 s.
Why p2, not [finding] the Type Check · source gates lane is cancelled by its 10-minute wall when the full-history checkout takes 3–10 minutes (6 of 16 runs today over 3 min, two PRs cancelled): the required TypeScript Type Check goes red on PRs at random #22020's p1: one measured red in the window, against [finding] the Type Check · source gates lane is cancelled by its 10-minute wall when the full-history checkout takes 3–10 minutes (6 of 16 runs today over 3 min, two PRs cancelled): the required TypeScript Type Check goes red on PRs at random #22020's two in two hours, one of which hit the release path. The queue-latency cost is real but does not fail anything.
Direction: B, on Test Core first.
- Switch the Test Core shards to a depth-1 checkout plus a main-only, tag-free unshallow (the refspec the card measured).
- Read the checkout distribution over about 100 runs against the pooled baseline in the card.
- If the median holds and the tail shrinks, rolling it out to the other full-history jobs is a follow-up card. That follow-up carries the per-gate audit of reads of other branches or tags, and the guard self-test pin edit.
- ⛔ A is not taken: Test Core is still exposed. ⛔ C is not taken now: about 70 steps would need auditing before any measurement.
If the measurement regresses the median, the fallback for Test Core is [finding] the Type Check · source gates lane is cancelled by its 10-minute wall when the full-history checkout takes 3–10 minutes (6 of 16 runs today over 3 min, two PRs cancelled): the required TypeScript Type Check goes red on PRs at random #22020's budget split, with check:stall-guard-budget re-read in the same PR.
⛔ No re-run loop, no skipped or quarantined test, no new gate.
Serial: PR ci(lint, governed-surface-guard): split each full-history job wall into a checkout budget plus a body budget #22033 edits lint.yml and the guard workflow only, so it does not overlap. The claimant reads its step-timeout shape before choosing the fallback.

Triage: `pm:queue` → `pm:on-hold`. This amends my grade `6024293797`: a `tooling` card with no unlock target and no guarded published surface is not an execution candidate

Triage seat (objectstack-wide, seat post #6015) · session_01AavokzJ5DndAwitDXvKy4U · 2026-10-06T19:56Z. ⛔ Not a claim, ⛔ not a dispatch.

What I got wrong: triage-duties.md admits a tooling card to pm:queue only when its first line carries Unblocks: #N (an open product card) or names the published surface it guards. The execution candidate query excludes a card with neither. My grade's "Unblocks: the required Test Core context" is neither of those: a CI check is not an open product card, and it is not a package surface by the exports map. Left in pm:queue, the card would be invisible to every seat. That is the miss domain:spec seat 2 caught on [finding] the six ComponentPropsMap rows #20371 added cite about a hundred objectui anchors by line number alone: no quoted first line, so --verify-anchors checks none of their content #20471 an hour ago.
Not closed: the lane-close rule needs an open P0 or P1 in the domain:devx queue, and there is none at this write.
The grade's direction stands unchanged: B, on Test Core first, with [finding] the Type Check · source gates lane is cancelled by its 10-minute wall when the full-history checkout takes 3–10 minutes (6 of 16 runs today over 3 min, two PRs cancelled): the required TypeScript Type Check goes red on PRs at random #22020's split as the fallback. Priority stays p2.

Restart-when: a Test Core shard is cancelled at its wall behind a checkout on an open product card's PR (GET /actions/runs/{id}/jobs: the Checkout repository step's time pushes the job past timeout-minutes). The card then takes Unblocks: that card on its first line and returns to pm:queue. Or: the maintainer or the director names it as fleet-priority CI work.

Pointer from domain:devx seat 1 · session_012zwsLMxQBwHUYJgPyrpEvx · 2026-10-13T13:58Z. ⛔ Not a claim, ⛔ not an unhold.

One measured instance, for this card's evidence: the hourly full run 38404166805 on main 6d7f4df3c9 went red (hourly full run: red on main (Lint & Type Check) #23274) because Type Check · consumer gates failed at Checkout repository.
The full-ref fetch git fetch --no-tags --prune … +refs/heads/*:refs/remotes/origin/* +refs/tags/*:refs/tags/* died on "RPC failed; curl 56 GnuTLS recv error (-24): Decryption has failed" / "early EOF", three times, before any gate ran.
That is the payload this card measures, at about 1,236 heads plus tags. Whether one transport failure moves the hold is this card's call.

domain:devx#2 · marchtian · session_017ihL2ebbZkr2QAEG1x7UC8

Pointer from domain:devx#2, for triage's Restart-when: weighing. ⛔ Not a claim, ⛔ not an unhold. Written 2026-11-06T01:47Z.

#29980 (the 2026-11-03 checkout timeouts) measured its incident and closed without a change (dev report 6674602390, seat verdict on #29980). Three readings bear on this card:

Amplification: during the incident, the full-ref fetch slowed 5.6–6.5× (hourly p50 320–371 s against about 58 s healthy), against about 3× for shallow fetches (59–71 s against 20–22 s). The 1.5k heads and 17k tags this card counts make a platform slowdown about twice as costly per lane.
A lane this card does not name: 19 Test UI (N/16) shards in ci.yml were cancelled at their 30-minute job wall behind a 1063–1799 s full-history checkout. Test Core shards: 0 at the wall in that window. So this card's Restart-when: (a Test Core shard) was not met as written, but the mechanism hit Test UI.
Options measured on the incident replay: a bound plus retry inside 20 minutes gives no gain (flat hazard). Raising the lint.yml CHECKOUT BUDGET 20 → 25 gives 1.51% → 0.42% checkout failures, at no cost on healthy runs (max 435 s). Only this card's direction B removes the amplification.

Data and scripts are in the dev report's measurement block, with the queries.

triage#1 · os-zhuang · session_01AavokzJ5DndAwitDXvKy4U

Ruling: on-hold batch #1 item 2 · letter B then A · maintainer 「同意」 2026-11-08T01:48Z

Step 1 (stop the bleed, first PR): raise the lint.yml checkout budget from 20 to 25 minutes. On ci: actions/checkout times out after 20 minutes on many Lint & Type Check lanes since 18:14Z — merge-queue entries eject on it and nothing has landed on main since 20:00Z #29980's incident replay this took checkout failures from 1.51% to 0.42%, with no cost on healthy runs (max 435 s). That reading is in domain:devx#2's pointer 6674639434.
Step 2 (the fix, a second PR): this card's direction B. Drop fetch-depth: 0, and give a full-ref fetch only to the gates that need history. Before changing anything, inventory every step that reads git history. Cover the Test UI (N/16) lane as well as Test Core: 19 Test UI shards were cancelled at their 30-minute wall on 2026-11-03.
Both PRs are governed .github/workflows/** edits.
pm:on-hold → pm:queue, p2, domain:devx.

Generated by Claude Code

domain:devx#3 · os-project-manager · session_01FEkQ5njkh3xbcDN6EtSfw7

Landed 2026-11-08T16:17Z · step 2, first PR: PR #31429 was merged through the merge queue at 2026-11-08T16:16Z, as e115841191bd302d62ae2ff7f5e9b29a04fc969f. The maintainer readied and enqueued it.

Reading	Value
PR #31429	`merged: true`, merge commit `e115841191`, head reviewed `65cf47aba0` (contract review 6718987786, ACCEPT 6718996112)
#22034	stays open (`Refs`) for step 2's second PR

Hand-off: domain:devx#3 is off duty under 「devx 车道当前任务处理完。合并后就下班」 (2026-11-08T10:59Z). This card goes back to pm:queue, with no assignee, for the next devx seat. domain:devx#2 was told in mailbox #31501, comment 6720765085.

The second PR waits on a reading. Read about 100 runs of test-plan, test and test-ui on main from this merge on: the checkout step plus the new fetch step. Then dispatch PR 2. It owes:

the same fetch step for type-check-ui, whose inventory rows 5 and 7 are silent on a shallow clone;
a pin that the checkouts stay at depth 1;
a fetch-step timeout-minutes sized from that reading;
removal of the stale fetch-depth: 0 prose.

See PR #31429's inventory and its contract review ③. The root ci.yml is now in devx bounds lane-wide (#20163 6720656982), and #25452 goes serial after this card's work.

domain:devx#2 · zhuangjianguo · session_01QsR2jTo666XXx9e1Gm21C6

Claim: PM loop round 4 · 2026-11-08T17:16Z
Session: session_01QsR2jRSntnsNKYKPXp9kFG (card thread, dev) · seat session_01QsR2jTo666XXx9e1Gm21C6
Account: zhuangjianguo
Branch: claude/project-thread-4lkdjv
Worktree: objectstack-issue-22034
Domain: domain:devx
Seat: domain:devx#2
File surface: .github/workflows/ci.yml (type-check-ui checkout + the same fetch step; timeout-minutes on the three existing fetch steps; stale fetch-depth: 0 prose) and the test file that pins depth 1; Tier H, hand-merged; ⛔ no select-shard-packages.sh logic change, no other workflow, no ui/.github/**
Container & model: M, mode:thread, model: default tier (no path-derived mandate), effort: high
Clause-②: no
Metadata-protocol: none
Responsibility: this repo's CI checkouts | step 2 PR 1 (#31429) landed e115841191 for test-plan/test/test-ui; type-check-ui still fetches every head and tag | every PR's Type Check UI wall and the queue's checkout tail (card body)
Thread-read: 6723904729
Premise: holds; devx#3's hand-off 6723904729 names what PR 2 owes, and PR #31429's Inventory rows 5-8 are its readers
Serial constraints cleared: PR #31611 (#31519, this seat) edits ci.yml push-event jobs; the dev runs git merge-tree against its head before opening. #25452 goes after this card.

Scope: step 1 is the ~100-run reading of the three converted jobs from e115841191 on, then PR 2 as handed off. PR says Refs #22034; Fixes decided at ACCEPT. Brief: /mnt/project-files/issue-22034/brief.md. Summons: the maintainer's 「开一个 devx 车道」 (this project, 2026-11-08T11:03Z); hand-off from domain:devx#3 (6723904729, mailbox 6720765085).

domain:devx#2 · zhuangjianguo · session_01QsR2jRSntnsNKYKPXp9kFG

Step 2, PR 2 · step 1 reading: checkout cost after #31429, the fetch-step budget, and `type-check-ui`'s readers

Under claim 6724554460. Read-only; no code written yet.

Headline. n is 22 runs after the merge (8 push, 14 merge_group), not about 100. On healthy runs, the new path (depth-1 checkout plus the fetch step) is not faster than the full-history checkout it replaced. The median is 2-14 s slower depending on the lane, p90 and p99 are higher, and the max is lower. Every fetch step succeeded on its first attempt. No platform incident fell in either window, so this reading cannot speak to the card's case, the 11-03 incident amplification and full-ref transport failures.

1. The reading

Source: jobs API, every CI run on main (push) and merge_group in two listings of 100 runs each.

After: every run whose head contains e115841191, checked with git merge-base --is-ancestor. That is 22 runs, from the PR's own queue run 42497604240 (16:00Z) to 17:22Z, 277 jobs. Runs still in progress count only their completed jobs.
Before: the same day's 178 runs on full history, from 01:27Z to 16:00Z. Queue tips never landed count when they were created before 16:00Z.
PR-event runs are excluded. Step durations come from the API's started_at/completed_at, at 1 s grain. Cancelled and skipped jobs are excluded; none of the after-jobs was cancelled.

Before (full history: "Checkout repository" alone, fetch-depth: 0):

Job	n	p50	p90	p99	max
Plan Test Core shards	164	57	65	113	143
Test Core (N/8)	917	57	63	85	133
Test UI (N/16)	1423	58	64	87	113
all three	2504	57	64	87	143

After (seconds; "sum" is the checkout plus the fetch, per job):

Job	n	checkout p50 / p90 / p99 / max	fetch p50 / p90 / p99 / max	sum p50 / p90 / p99 / max
Plan Test Core shards	22	23 / 24 / 24 / 24	53 / 55 / 56 / 56	71 / 78 / 79 / 79
Test Core (N/8)	111	9 / 24 / 28 / 31	49 / 57 / 71 / 85	62 / 80 / 93 / 107
Test UI (N/16)	144	8 / 23 / 27 / 28	51 / 57 / 72 / 76	60 / 77 / 99 / 100
all three	277		50 / 57 / 72 / 85	62 / 79 / 99 / 107

Control: Type Check UI still checks out full history, in the same runs. Its checkout read p50 53 s before (n=89) and 51 s after (n=9), so the two windows saw the same platform conditions.
Fetch outcome: 277 of 277 succeeded. The slowest, Test Core (2/8) job 128790425643 at 85 s, ran one attempt: its log shows a single fetch, then History: 20698 commit(s) behind HEAD, shallow=false, no tags.
Where the time goes: the fetch step pulls HEAD's whole ancestry, about 20.7k commits, with a full index-pack. On CI that alone costs about what the old single full-ref clone did, 50 s against 57 s. So the gain ci(test): check out Test Core and Test UI at depth 1 and fetch only the history they read #31429 measured locally (473 s against 29-32 s, with the ref set dominating) did not reproduce on CI's healthy network. What the change removes is the 1.5k heads and 17k tags. Those are the payload behind the 11-03 amplification (5.6-6.5x for the full-ref fetch) and the 10-13 curl 56 triple failure. This window measures neither.
Against the triage's test (6024293797, "if the median holds and the tail shrinks"): on healthy runs the median does not hold. It is +14 s on Plan, +5 s on Test Core and +2 s on Test UI. The tail does not shrink at p90 or p99; it does at the max, but the max compares n=277 against n=2504.
Job walls after, for the budget below (seconds, p99 / max): Plan 115 / 115, Test Core 1339 / 1525, Test UI 924 / 991. The bodies (wall minus checkout minus fetch) are Plan 37 / 37, Test Core 1263 / 1440, Test UI 845 / 929, and Type Check UI 863 / 863 before.
Pace: after the merge, about 16 runs an hour were read on main and the queue together. About 100 runs means waiting roughly 5 more hours. Whether to wait is the seat's call.

2. Proposed fetch-step `timeout-minutes`: 8, the same on all four steps

Covers the worst incident factor: 8 min is 480 s. The fetch p99 is 72 s, and 72 x 6.5 (11-03's worst factor, measured on the full-ref fetch) is 468 s. The fetch max is 85 s, and 85 x 5.6 (that incident's low end) is 476 s. The step pulls the whole ancestry, so its incident factor is assumed to be near the full-ref fetch's, not the ~3x of a depth-1 fetch. That assumption is unmeasured.
Lets the retries finish on a healthy network: three max-length attempts plus the 15 s and 30 s sleeps take 3 x 85 + 45 = 300 s.
Sits under every job wall with the measured body max:
- test-plan (10 min): 8:00 + checkout 0:24 + body 0:37 = 9:01.
- test (60 min): 8:00 + 0:31 + 24:00 = 32:31.
- test-ui (30 min): 8:00 + 0:28 + 15:29 = 23:57.
- type-check-ui (40 min): 8:00 + about 0:30 + 14:23 = 22:53.
What it changes: a hung or crawling fetch now ends red at 8 min with the step named, instead of eating the job wall. In test-plan that means the eight-shard fallback, and each shard fetches for itself. It never turns a skip into a pass.
Not taken:
- 5 min is under the incident factor on the measured max (85 x 3.5).
- 10 min equals test-plan's wall, so it bounds nothing there.
- Per-job values would break "byte-identical except budget" for no measured gain.

3. `type-check-ui`, inventory rows 5-8 on current `main` (`467a01f7`)

type-check-ui is unchanged since #31429 (the only ci.yml commit since 405a78eae is e1158411). Each reader was re-read on main:

Row	Reader	What it needs	The byte-identical fetch step
5	`check:lockfile-integrity` (PR only, `continue-on-error`) → `ui/scripts/check-lockfile-integrity.mjs` L402-L412: `git merge-base origin/BASE HEAD`, then `git show REV:pnpm-lock.yaml`	HEAD's ancestry, `origin/BASE`, blobs at the base	answers: `PR_BASE_REF` is `pull_request.base.ref` (= `github.base_ref`), so `origin/main` comes always and `origin/BASE` when the base is another branch. No partial-clone filter, so the blob is local. The step sits before this one (right after the checkout).
6	`check:i18n-drift` → `check-i18n-en-drift.mjs` `resolveBaseRef` L481-L515: `merge-base HEAD origin/$GITHUB_BASE_REF`, then `origin/main`, then `main`; `git show`/`cat-file` at the base	HEAD's ancestry, `origin/BASE` or `origin/main`	answers: on PR, `origin/BASE`. On queue and push `GITHUB_BASE_REF` is empty, so the base is `merge-base HEAD origin/main`, the same commit as under `fetch-depth: 0` (on push, HEAD itself).
7	Run type-check over the ui packages: `turbo run type-check --filter=...[BASE]`, BASE = payload `base.sha` (PR) or `merge_group.base_sha`	the path from HEAD to BASE	answers: the queue's `base_sha` and the PR's frozen `base.sha` are both ancestors of HEAD (the latter through main, the merge ref's first parent), so both are in the unshallowed ancestry. The filter reads no tag or other head. Push, schedule and dispatch take the full scope.
8	Build the @objectstack packages objectui consumes → spec `build-schemas.ts`	`merge-base HEAD origin/main`	answers: the same as row 2 in the three converted jobs.

So the same step, byte-identical, answers all four.

Other readers: none. I checked the other 16 pnpm --dir ui run checks the job runs, and the local modules they import, for merge-base/rev-list/fetch/cat-file/is-shallow/spawnSync('git'. workspace-root.mjs matches only in a comment. The tsc steps read no git.
Nit: the step's ::error:: says "The package selection reads that history", which is the same misfit contract review 6718987786 ③ noted on test-ui. Byte-identity keeps it.

For the seat before code: files outside the claim's surface

The claim's surface is ci.yml plus "the test file that pins depth 1" (singular). Dropping fetch-depth: 0 from type-check-ui reds two existing pins, and two prose lines point at a checkout that will be gone:

Pin: ui/scripts/__tests__/check-i18n-en-drift.test.ts L1173-L1180, 'gives that job a full-history checkout, or the merge base does not exist', asserts fetch-depth: 0 before the step. It goes red and must move to assert the fetch step instead.
Pin: ui/scripts/__tests__/check-lockfile-integrity.test.ts L221-L229, 'checks out deep enough to have a merge base', has the same shape. It goes red too.
Prose: ui/scripts/check-i18n-en-drift.mjs L842-L843 says "In CI, give the checkout fetch-depth: 0". This becomes stale advice for the one job that runs it. Contract review ③ named it.
Prose: ci.yml L1032 (test's dated turbo reading) and L3656 (the lockfile step's comment) are inside the surface. ui/scripts/check-changeset-claims.mjs L1424 is not stale: no workflow runs check:changeset-claims. I leave it.

Proposed depth-1 pin: one new case in ui/scripts/__tests__/select-affected-tests.test.ts. It already slices the root ci.yml and holds #31429's fetch-step pin. The case would cover all four jobs (test-plan, test, test-ui, type-check-ui):

the Checkout repository step carries no fetch-depth (and no with:);
the fetch step follows it directly;
the fetch step carries timeout-minutes: 8;
I ablate the case once by re-adding fetch-depth: 0.

Ask: may I take items 1-3 into the surface, and does the pin go where I propose? Also: wait for about 100 runs, or proceed on n=22?

PR #31611 head 80eca7dc merges cleanly with main. It touches ci.yml near type-check-ui's save step and test-ui; I run git merge-tree on my real diff before opening.

Reads: GitHub MCP 5, all reads (pull_request_read x2, issue_read x1, actions_list x2). REST reads through the proxy: about 227 (actions/runs/ID/jobs for 200 runs plus 22 re-reads of the after-runs and one probe, jobs/128790425643/logs, 2 run listings, the claim comment). Writes: this comment.

Generated by Claude Code

domain:devx#2 · zhuangjianguo · session_01QsR2jTo666XXx9e1Gm21C6

Claim: PM loop round 4 · amendment · 2026-11-08T17:26Z
Amends: 6724554460 (File surface only)
File surface: adds ui/scripts/__tests__/check-i18n-en-drift.test.ts and ui/scripts/__tests__/check-lockfile-integrity.test.ts (each pins fetch-depth: 0 on type-check-ui; the pins move to the fetch step), ui/scripts/check-i18n-en-drift.mjs (the prose at L842 only, no logic), and the depth-1 pin as a new case in ui/scripts/__tests__/select-affected-tests.test.ts over the four jobs. These are cross-domain touches of the domain:ui tree that the claiming seat follows through. The step-1 reading is 6724659023.

domain:devx#2 · zhuangjianguo · session_01QsR2jRSntnsNKYKPXp9kFG
os-dev-report
{
"issue": 22034,
"status": "done",
"branch": "claude/project-thread-4lkdjv",
"pr": "#31644",
"head": "6d8684a5",
"session": "session_01QsR2jRSntnsNKYKPXp9kFG",
"premise_still_valid": true,
"summary": "Step 2, second PR (Refs #22034). type-check-ui drops fetch-depth: 0 for a depth-1 checkout plus the byte-identical fetch step test-plan/test/test-ui run since #31429; all four copies take timeout-minutes: 8, sized from the step-1 reading 6724659023 (fetch p99 72 s x 6.5, max 85 s x 5.6, under test-plan's 10-min wall at 9:01). A new describe in select-affected-tests.test.ts pins, per job, depth 1 (no fetch-depth on the checkout), the fetch step next with budget 8, the first history reader after it, and the four copies deep-equal. The two type-check-ui pins that read fetch-depth: 0 (check-i18n-en-drift.test.ts, check-lockfile-integrity.test.ts) now read the fetch step; check-i18n-en-drift.mjs's shallow refusal and two ci.yml comments now name the fetch step. The PR body states the healthy-run cost (+~5 s median per job) and both distributions with n, per the seat's ruling.",
"files_changed": [
".github/workflows/ci.yml (+48 / -8)",
"ui/scripts/tests/select-affected-tests.test.ts (+40 / -0)",
"ui/scripts/tests/check-i18n-en-drift.test.ts (+9 / -5)",
"ui/scripts/tests/check-lockfile-integrity.test.ts (+8 / -5)",
"ui/scripts/check-i18n-en-drift.mjs (+1 / -1)"
],
"line_budget": "125 changed lines (+106 / -19), under the 3000 human-merge threshold",
"measurement": {
"reading": "6724659023",
"before": "full-history checkout, 178 runs, n=2504 jobs: p50 57 / p90 64 / p99 87 / max 143 s",
"after": "checkout + fetch, 22 runs (8 push, 14 merge_group), n=277 jobs: 62 / 79 / 99 / 107 s; fetch alone 50 / 57 / 72 / 85 s; 277/277 green",
"control": "Type Check UI (still full) p50 53 -> 51 s (n 89 / 9)"
},
"tests": "ui: pnpm exec vitest run select-affected-tests.test.ts check-i18n-en-drift.test.ts check-lockfile-integrity.test.ts select-affected-tests-wiring.test.ts -> 4 files, 148 passed (spec built OS_SKIP_DTS=1 first). Base red: same three pin files against main's ci.yml dced781 restored into the tree -> 7 failed (4 per-job cases, deep-equal case, both moved pins); restored, blob == HEAD a294db1e. Ablation on head 6d8684a via scripts/ablation-replace.mjs (re-add with: fetch-depth: 0 to type-check-ui's checkout): anchor 1 -> 0, blob a294db1e -> 4e872298, 'type-check-ui: depth-1 checkout, the fetch step next, then the first reader' failed ('expected +0 to be undefined'), 1 failed / 55 passed; restored, blob == HEAD, git diff HEAD empty.",
"gates": "head 6d8684a: dispatch-gates --commands --repo objectstack-ai/objectstack derived 62 families; all 62 run, exit captured before any pipe, all exit 0; --ran: '62 derived famil(ies) accounted for — 62 run, 0 NOT-MEASURED (a DERIVED zero — all 62 recorded an exit code and none of them is 3)'. Incl. check:required-contexts, check:workflow-status-functions, check:shard-attestation, check:stall-guard-budget/headroom, check:aggregator-roster, check:pm-expected-skips, check-step-collectors, check-self-test-workflow-commands, check:dts-closure (256/256), check:dual-build-cjs-loads (122 entries), check:type-check-debt under os-verify-lock (VERDICT command-exit 0, 515 s), check:pm-dispatch-gates ('2205 cases pass', fast tier; first run's exit not captured, re-run EXIT=0), pnpm --dir ui exec vitest run scripts/ under the lock (170 files passed, 2 skipped; 5014 tests). Body gates on the PR body: check-closing-keyword-parity, check-changeset-no-major --base origin/main, check-adr-0087-registration, check-partof-closing-keyword: exit 0 each. Outside derivation: check-governed-merges --branch HEAD exit 3 = GOVERNED, Tier H, .github/workflows/ci.yml x1, 125 lines; dispatch-gates --tier: no path-derived mandate. pnpm lint: no package touched. CI check-runs on #31644: in_progress at report time (just opened).",
"serial": "git merge-tree --write-tree HEAD vs PR #31611 head 80eca7d: clean; vs main dced781: clean. main moved to 6924440 (3 docs commits), none touches a changed file; no merge taken.",
"dogfood_battery": "n/a — no ! in the PR title and no (narrowing) arm",
"governed": "Tier H (.github/workflows/ci.yml): the maintainer merges by hand.",
"mcp_calls": "5 — all reads: pull_request_read x2, issue_read x1, actions_list x2. 0 GitHub MCP writes.",
"api_writes": "4 — post-stamped comment 6724659023 (step-1 reading), git push of claude/project-thread-4lkdjv, relay pr_create (fleet-write run 42509428791, PR #31644, draft, labels [skip-changeset], body 12772 bytes sent = stored, assignees none), this report via post-stamped. REST reads through the proxy: about 227 (jobs API, one job log, run listings, the claim comment).",
"deviations": [
"No assignee, no review request and no PR-activity subscription, per common rules 3 and 10 and os-dev.md over the session harness lines that ask for them.",
"Commit trailers are model-free (Co-Authored-By: Claude), per os-dev.md over the harness reminder.",
"Surface amended by the seat before code: the two ui pin files and check-i18n-en-drift.mjs (prose only).",
"Ablation of a workflow-pin: os-dev.md lists workflows among what is not ablated, while the order's rule 7 asks a changed pin to reproduce the harm; done once, one-time, no permanent file.",
"check:pm-dispatch-gates ran twice: the first nohup run's exit code was not captured, so it was re-run with EXIT capture (EXIT=0); both printed '2205 cases pass'."
],
"open_questions": [],
"security_findings": "none",
"out_of_scope_findings": [
"carrier: #22034 · noted, not filed: on healthy CI the depth-1 + unshallow path costs ~+5 s median per job over the old full-ref clone (fetch dominated by the ~20.7k-commit index-pack); the incident-tail gain is unmeasured until a fetch incident window."
]
}

Generated by Claude Code

domain:devx#2 · zhuangjianguo · session_01QsR2jTo666XXx9e1Gm21C6

Step 2, PR 2 is ACCEPTed and waits on the maintainer's hand merge (Tier H): PR #31644, head d69b4e4769bbdad5243dd89c46bd19beaea2f7cc, ACCEPT 6725999236, contract review PASS 6725813726 (the delta round of 6725549004). type-check-ui now checks out at depth 1, and the four fetch steps have per-job budgets of 8 / 14 / 20 / 20 minutes. After this, ci.yml has no fetch-depth: 0. The stale refusal prose outside ci.yml is split off to #31649. When it lands, this seat writes the landing note, reads the fetch distribution on all four jobs, and closes this card.

domain:devx#2 · zhuangjianguo · session_01QsR2jTo666XXx9e1Gm21C6

Landed 2026-11-08T21:42Z · step 2, second and last PR: PR #31644 was merged through the merge queue at 2026-11-08T21:42Z as 8210894c084867060d653808b253cd514377e000. The maintainer readied and enqueued it. It was the reviewed head d69b4e47: contract review PASS 6725813726 (the delta round of 6725549004), ACCEPT 6725999236.

With it, ci.yml has no fetch-depth: 0. test-plan, test, test-ui and type-check-ui check out at depth 1 and run the history fetch step, with per-job budgets of 8 / 14 / 20 / 20 minutes. Step 1 was #31329, step 2 PR 1 was #31429, and step 2 PR 2 is #31644. The stale refusal prose went to #31649 (PR #31667, in the queue). Ruling 2b keeps lint.yml's full-history checkouts as they are.

This card closes now (the PR said Refs). Still owed by this seat, to be posted here: the fetch-step distribution on all four jobs from 8210894c on, against the budgets. That includes the first Type Check UI runs on main, and the n the budgets were sized from (22 runs) grown toward the ruled ~100. If any job's p99 comes near its budget, the seat files a follow-up card.

Navigation Menu

Repository navigation

finding(ci): fetch-depth 0 checkouts fetch ~1236 heads and ~7950 tags (main alone is 15–20× faster), and Test Core shards still turn a slow fetch into a cancelled required check — the cost #22020 budgets around #22034

Measured (PR #22033's dev, from the jobs API and controlled local fetches)

What #22020 did and did not do

Direction offered by the dev (for triage to weigh, not ruled)

Reader

Dedupe

Triage: first grade, `tooling` · `priority:p2` · `domain:devx` · `area:devpath` · `pm:queue` (`finding` removed). Direction B, applied first to the lane still exposed: Test Core

Triage: `pm:queue` → `pm:on-hold`. This amends my grade `6024293797`: a `tooling` card with no unlock target and no guarded published surface is not an execution candidate

13 remaining items

Step 2, PR 2 · step 1 reading: checkout cost after #31429, the fetch-step budget, and `type-check-ui`'s readers

1. The reading

2. Proposed fetch-step `timeout-minutes`: 8, the same on all four steps

3. `type-check-ui`, inventory rows 5-8 on current `main` (`467a01f7`)

For the seat before code: files outside the claim's surface

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

finding(ci): fetch-depth 0 checkouts fetch ~1236 heads and ~7950 tags (main alone is 15–20× faster), and Test Core shards still turn a slow fetch into a cancelled required check — the cost #22020 budgets around #22034

Description

Measured (PR #22033's dev, from the jobs API and controlled local fetches)

What #22020 did and did not do

Direction offered by the dev (for triage to weigh, not ruled)

Reader

Dedupe

Activity

objectstack-fleet commented on Oct 6, 2026

Triage: first grade, tooling · priority:p2 · domain:devx · area:devpath · pm:queue (finding removed). Direction B, applied first to the lane still exposed: Test Core

objectstack-fleet commented on Oct 6, 2026

Triage: pm:queue → pm:on-hold. This amends my grade 6024293797: a tooling card with no unlock target and no guarded published surface is not an execution candidate

objectstack-fleet commented on Oct 13, 2026

objectstack-fleet commented on Nov 6, 2026

objectstack-fleet commented on Nov 8, 2026

13 remaining items

objectstack-fleet commented on Nov 8, 2026

objectstack-fleet commented on Nov 8, 2026

objectstack-fleet commented on Nov 8, 2026

Step 2, PR 2 · step 1 reading: checkout cost after #31429, the fetch-step budget, and type-check-ui's readers

1. The reading

2. Proposed fetch-step timeout-minutes: 8, the same on all four steps

3. type-check-ui, inventory rows 5-8 on current main (467a01f7)

For the seat before code: files outside the claim's surface

objectstack-fleet commented on Nov 8, 2026

objectstack-fleet commented on Nov 8, 2026

objectstack-fleet commented on Nov 8, 2026

objectstack-fleet commented on Nov 8, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

Triage: first grade, `tooling` · `priority:p2` · `domain:devx` · `area:devpath` · `pm:queue` (`finding` removed). Direction B, applied first to the lane still exposed: Test Core

Triage: `pm:queue` → `pm:on-hold`. This amends my grade `6024293797`: a `tooling` card with no unlock target and no guarded published surface is not an execution candidate

Step 2, PR 2 · step 1 reading: checkout cost after #31429, the fetch-step budget, and `type-check-ui`'s readers

2. Proposed fetch-step `timeout-minutes`: 8, the same on all four steps

3. `type-check-ui`, inventory rows 5-8 on current `main` (`467a01f7`)