ci(ui-merge): every merge-queue entry runs all four ui lanes because the ui filter output falls back to 'true' on merge_group — decide the ui lanes from the diff against merge_group.base_sha, as Test Core already does (parent #25677)

Filing gate ③ (maintainer-directed) + ④: aftercare of the ui/ merge (#25752), part of #25677. Touches .github/workflows/ci.yml = Tier H: GOVERNED_APPROVERS review, the PR stops at draft.

Maintainer, 2026-10-26 01:45Z, in the project thread that measured post-merge CI speed: 「创建issue」 (asked after reading the ranked fix list; this is item 1).

谁付代价（业务筛）: every PR that lands goes through the merge queue. Since the merge, a queue entry that changes only docs, scripts or a core package still runs Test UI (16 shards), Type Check UI, Build & E2E UI and Lint UI. Small entries used to finish in 4–10 min; now their floor is ~16–18 min, and each entry pays ~220 extra runner-minutes. With up to 8 card threads landing in parallel, this is the main post-merge slowdown the fleet feels.

Measured

Pre-merge merge_group CI runs (2026-10-25, list of 40): small entries 4.0–9.8 min (e.g. 40181079769 4.0, 40195098891 4.9, 40198037200 5.1), core entries 26–30 min.
Post-merge queue build 40251505048 (fix(ui-merge): two tests that #25745 broke on main read the merged tree — the Changesets config through changesetConfigPath(), the SDUI html tier by its tags (#25711) #25852): Test UI shards 8.4–17.7 min, Type Check UI 15.5 min, Build & E2E UI 5.0, Lint UI 4.5 — all scheduled although the job's queue base was only one commit behind.
Cause, ci.yml filter job: dorny/paths-filter runs only on pull_request / push (if: github.event_name == 'pull_request' || github.event_name == 'push'), and the outputs ui / ui_consumers fall back to 'true' (ui: ${{ steps.changes.outputs.ui || 'true' }}). Test Core does not have this problem: scripts/ci/select-shard-packages.sh already scopes merge_group runs with turbo ls --affected against merge_group.base_sha.

Wanted

On merge_group, compute the ui and ui_consumers outputs from git diff --name-only <merge_group.base_sha>...HEAD, using the same globs the pull_request filter uses. Fail open: if the base cannot be fetched or the diff fails, the output stays 'true'.
push / schedule / workflow_dispatch keep the full battery (maintainer ruling 2026-10-25 Q6-b: push-to-main runs stay).
The workflow's gate rosters follow (check-ci-filter-parity, check-expected-skips, check-step-collectors); each --self-test exits 0.

Acceptance

A queue entry that changes no ui/** file and no ui_consumers path shows the four ui lanes as skipped, and the required contexts (Lint UI, Build & E2E UI, and later Type Check UI) still report success. Skipped counts as passing for a required check, so name the run that shows this.
A queue entry touching ui/** still runs all four lanes.
On the card: queue CI wall time for one small entry before and after.

Expected saving: ~220 runner-min per non-ui queue entry; small entries back to 5–10 min. Size: S. Clause-②: no. Metadata-protocol: none.

Dedupe

search_issues 2026-10-26 01:50Z 「merge_group merge queue runs all ui jobs paths filter ui output true」: 5 hits, all closed and unrelated (#19036, #16574, #11621, #9283, #4897). New.

Measurement report: /mnt/project-files/d1-merge/out/ci-speed/post-merge-pr-ci-speed.md (project thread "Keep PR CI fast after the merge").

Claim: epic #25677 round 8 (post-cutover CI speed cards; Tesla 2026-10-26 01:45Z 「创建issue」) 2026-10-26T01:55Z
Session: session_01RSDL3SFtJaKzcymvCD8RsU
Account: os-tesla
Branch: claude/project-thread-8dceid (the card thread's designated branch in objectstack-ai/objectstack — the thread can push only this branch, a measured platform exception to claude/issue-<n>-<slug>)
Worktree: objectstack-issue-25867
Domain: domain:devx
Seat: epic:#25677
Scope: the whole card — ci.yml filter job computes ui / ui_consumers on merge_group from git diff --name-only <merge_group.base_sha>...HEAD with the pull_request globs, fail-open to 'true'; push/schedule/workflow_dispatch keep the full battery (Q6-b); the gate rosters follow with --self-test 0; Tier H (draft, os-zhuang + hotlong), skip-changeset
Why now: every queue entry pays ~220 runner-min and a 16–18 min floor for four ui lanes it does not need; with 6 card threads landing tonight this is the queue's pace
Container & model: S, mode:thread (thread session session_01Suqt7vsGh5kE2NhDZyqeh6; order: this card's body, FRAME.md applies), model: opus, effort: xhigh
Clause-②: no
Metadata-protocol: none
Responsibility: n/a — epic step
Thread-read: none
Serial constraints cleared: #25868 edits only the test-ui job of the same file (branch claude/project-thread-n0iri6); the #25853 seat's Type Check UI row PR edits scripts/check-required-contexts.mjs and the two statesTheSet surfaces, not ci.yml; #25869 is held until this and #25868 land. Rebase over whichever lands first; ⛔ do not change any lane's if: beyond the filter outputs this card names

os-dev-report

{
 "issue": 25867,
 "status": "done",
 "branch": "claude/project-thread-8dceid",
 "pr": "https://github.com/objectstack-ai/objectstack/pull/25881",
 "session": "session_01Suqt7vsGh5kE2NhDZyqeh6",
 "premise_still_valid": true,
 "summary": "ci.yml filter job: the dorny/paths-filter step's allow-list gains merge_group (the pinned v4.0.3 reads merge_group.base_sha/head_sha from the payload and diffs base..head), continue-on-error on that event so a failed diff fails open to 'true'. ui and ui_consumers read the queue diff; docs/core/crosspkg are guarded with github.event_name != 'merge_group' and stay 'true' on queue builds; schedule/dispatch unchanged. check-expected-skips header + ui roster reasons and the dispatch-gates marker census (133 -> 137) follow.",
 "tests": "Replayed the pinned action dist/index.js with synthetic merge_group payloads and this branch's filter block: non-ui entry abebc4e1 -> ui=false ui_consumers=false; spec+lint b00fc55c -> ui=false ui_consumers=true; ui/ entry 12105449 -> ui=true; 3-commit group 3fca5996..c475b45b -> ui=true; depth-1 checkout with base absent -> action fetched base by sha and diffed; unfetchable base -> action exit 1 with no outputs (fail-open via continue-on-error). Live queue evidence owed after landing (this PR's own entry changes ci.yml, which is in the ui filter, so it runs all four ui lanes).",
 "gates": "be2cff5e: dispatch-gates --commands derives 67; all 67 exit 0 (incl. check:pm-dispatch-gates fast tier 1917/1917, check-ci-filter-parity + --self-test, check:pm-expected-skips, check-step-collectors + --self-test, check:type-check-debt); --ran: 67 derived, 67 run, 0 NOT-MEASURED (derived zero). Earlier head 5aa9fcb3: check:pm-dispatch-gates 1 of 1917 red (marker census pinned check-expected-skips.mjs:133), fixed in be2cff5e. PR check-runs: in_progress at report time.",
 "dogfood_battery": "n/a — no `!` in the PR title and no (narrowing) arm",
 "files_changed": [
  ".github/workflows/ci.yml",
  "scripts/pm/check-expected-skips.mjs",
  "scripts/pm/dispatch-gates.self-test.mjs"
 ],
 "line_budget": "+70 / -37",
 "deviations": [
  "Route: the card says git diff --name-only with the pull_request globs; the step runs dorny/paths-filter itself on merge_group instead, so the globs are the same block by construction and check-ci-filter-parity's one-step rule holds.",
  "Third file outside the card's named rosters: dispatch-gates.self-test.mjs marker-census row moved 133 -> 137 because the check-expected-skips header grew four lines."
 ],
 "mcp_calls": "4 — issue_read get, issue_read get_comments, list_pull_requests, pull_request_read get (all reads, no write tool)",
 "api_writes": "2 — relay pr_create (fleet-write run 40259928648; relay's governed-review follow-up requested os-zhuang + hotlong), this report comment via post-stamped",
 "open_questions": [],
 "out_of_scope_findings": [
  "carrier: none · noted, not filed — docs/core/crosspkg could read the queue diff the same way (Build Core / Build Docs would skip on queue entries outside their paths); left at 'true' per the filter's 'queue build is the last validation before main' reasoning"
 ]
}

Generated by Claude Code

Live queue evidence for #25881 (merged 03:27Z), read 2026-10-26T05:25Z.

Non-ui entry: #25929 (fix(service-automation), 3 files: packages/services/service-automation/src/** + a changeset). Merge-group run 40271496679, base dd5e3554f0:

job	conclusion	window
filter	success	04:55:07–04:55:34
Test UI (pre-expansion name, all 16 shards)	skipped	04:55:34
Type Check UI	skipped	04:55:34
Build & E2E UI	skipped	04:55:34
Lint UI	skipped	04:55:34
Build Core / Build Docs	success	until 05:01:26 / 05:00:41
Test Core (8 shards + aggregate)	success	until 05:15:39 (long pole: 4/8, 19.5 min)

The entry landed on main as de2b7e02 with Lint UI and Build & E2E UI (required) reported as skipped, so the queue accepted skipped as passing.

Spec entry: #25830 (spec,plugin-auth,runtime,cli). Run 40264874688 took the ui_consumers path, as designed:

Test UI and Lint UI were skipped.
Type Check UI (03:36:39–03:48:37) and Build & E2E UI (03:36:39–03:43:43) ran.
The run concluded success and landed as f00e8d5c.

Before vs after (ui lanes per queue entry):

Before, on fix(ui-merge): two tests that #25745 broke on main read the merged tree — the Changesets config through changesetConfigPath(), the SDUI html tier by its tags (#25711) #25852 (run 40251505048, one commit behind its base), every ui lane was scheduled: Test UI shards ran 8.4–17.7 min, Type Check UI 15.5, Build & E2E UI 5.0, Lint UI 4.5. That is about 220 runner-min per entry.
After, on 40271496679, the ui lanes took 0 runner-min and occupied 0 runner slots.

Wall clock: 40271496679 ran 20.6 min (04:55:05 to 05:15:39). For an entry that touches a core package, Test Core 4/8 is the critical path either way, so the ui skip does not shorten it; the saving there is runner-minutes and slots.

Pending: a docs- or scripts-only entry, where the ui lanes used to set the ~16–18 min floor, has not entered the queue since the merge. When one does, its wall clock is the remaining before/after number on this card.

Generated by Claude Code

Navigation Menu

Repository navigation

ci(ui-merge): every merge-queue entry runs all four ui lanes because the `ui` filter output falls back to 'true' on merge_group — decide the ui lanes from the diff against `merge_group.base_sha`, as Test Core already does (parent #25677) #25867

Measured

Wanted

Acceptance

Dedupe

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

ci(ui-merge): every merge-queue entry runs all four ui lanes because the ui filter output falls back to 'true' on merge_group — decide the ui lanes from the diff against merge_group.base_sha, as Test Core already does (parent #25677) #25867

Description

Measured

Wanted

Acceptance

Dedupe

Activity

objectstack-fleet commented on Oct 26, 2026

objectstack-fleet commented on Oct 26, 2026

objectstack-fleet commented on Oct 26, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

ci(ui-merge): every merge-queue entry runs all four ui lanes because the `ui` filter output falls back to 'true' on merge_group — decide the ui lanes from the diff against `merge_group.base_sha`, as Test Core already does (parent #25677) #25867