Repository navigation
[devx] Live E2E UI skips every server-only PR: widen its gate to the server packages the Console runs against, and pin the spec-time budget #27550
Description
domain:devx#2 · os-justin · session_01QSeKYmTYfaPbLQD4BJfomW
Filing gate ③ (maintainer-directed). Maintainer, this session's chat, 2026-10-29, verbatim: 「同意，立卡」, answering the seat's Live E2E UI assessment (B: widen the gate to the server packages the Console depends on; C: the scenarios below), after the seat measured that neither adds CI wall-clock time. This is B. Ruling-ref: this card's quote, which is the ci.yml ruling the standing constraint asks for. It is later than, and for this lane supersedes, #16455's "e2e and live tiers to nightly" (#25914 ruling 6435339772 Q1 A already flipped it for this lane).
维护者速读
- 改了什么：
Live E2E UI现在只在改到ui/**或 7 个 ui 依赖包时才跑；改为同时覆盖 Console 实际依赖的服务端包（objectql、runtime、rest、metadata-protocol、各业务插件、examples/app-showcase等），并在 job 注释里写死「浏览器测试部分 ≤ 2 分钟」的预算。 - 为什么改：近期大量「后端改了、Console 跟着坏」的 bug 都是合并后人工 QA 才发现的；纯后端 PR 和合并队列都会跳过这条检查，正好漏掉这一类。
- 风险与回滚：不增加 PR 墙钟（实测纯后端 PR 的墙钟由
Test Core决定，最快也约 11 分钟，而它最晚第 9 分钟就跑完）；代价是约三分之一的 PR 每次多约 6 分钟 runner 时间。回滚 = 还原if:。 - 先看：
ci.yml的live-e2e-uijob 的if:和 filter 输出。 - 两分钟验证：开一个只改
packages/objectql的 PR，Live E2E UI应该运行而不是跳过。 - 需要你做的：Tier H，批准 PR。
Measured (seat, 2026-10-28/29)
- Gate today:
ui_tree != 'false' && (ui != 'false' || ui_consumers != 'false'), whereui_consumers= spec, client, core, formula, lint, sdui-parser, types. Themerge_groupbuild reads the same filter against the queue's base, so it skips there too. - Of 60 successful PR CI runs since 2026-10-28T12:00Z, 21 skipped it. Their wall clock was min ≈ 651 s (one 50 s run touched no package), median 894 s, max 1229 s. The slowest job was a
Test Core (k/8)shard every time. - Live E2E UI on 37 runs: it starts ≈ 25 s in and runs 77–509 s (median 391 s), so it ends by ≈ 535 s. It finished ≥ 176 s before the run's last job on every run, and failed 0 times.
- Escaped defects of the class (all found after merge, by QA/dogfood in a browser): Record edit dialog: on a row the caller may read but not write (sharing / RLS / ownership), the fields are offered as editable and the save is refused whole — the dialog never asks whether the row is writable #27138, Record detail page inline edit: a row-conditioned field lock (
when) still disables the field on every row — the detail page does not ask the record projection the edit dialog now asks #27146 and Record form field-access scope: the record projection is asked once and not again when the form re-reads the record, so a field a save just locked is drawn enabled until remount #27147 (access), console: a full-page record form opened by direct URL saves silently and then leaves the app for /_console/home #27454 (record save by URL), Studio: "Add item" on a dashboard list property autosaves an empty item (422), and every later edit is refused and lost until reload #27459 and studio: the Automations and App creates can write a blank draft over a published flow or app of the same identifier (first-write pin is draft-only) #26917 (Studio save), A screen-flow result screen (onlyrecord_linkfields) still shows Cancel, and closing it neither completes the run nor refreshes the record — the converted lead still reads Qualified #26902 (screen flow), console: a bare-string CEL predicate (navvisible, pagevisibleWhen) runs on the legacy evaluator and fails, though the spec defines a bare string as CEL shorthand #27457 (CEL).
Task
- Widen the condition. Recommended: add
|| needs.filter.outputs.core != 'false'(thecorefilter already coverspackages/**,examples/**andapps/!(docs)/**). The alternative is a narrower new filter output naming only the server packages the Console calls. The reader measures how many PRs each option triggers on (from the last 60 runs), picks one, and records why. ⛔ No retry wrapper, nocontinue-on-error. - In the job comment, state the budget: the
Run the live e2e specsstep stays under 2 minutes. A spec that would push it over stays intest:e2e:live:nightly. A gate or test that reads the step duration is welcome but optional. - Re-run the gates the change derives:
check:required-contexts,check:workflow-status-functions(it now carries the context-availability rule),check:ci-filter-parity, andcheck-expected-skipson the PR. The derived expected-skip roster changes with theif:.
Reader who acts. The devx seat. One Tier H PR on .github/workflows/ci.yml, with a contract review at CONTRACT_REVIEW_TIER.
Dedupe. search_issues (semantic) for "Live E2E UI live e2e spec gate backend packages persisted batch access member Studio promote nightly": #27228 (the ceiling pin, a different item), #27356, #16455 (closed, superseded for this lane). New.
Generated by Claude Code
Activity
objectstack-fleet commentedon Oct 29, 2026
domain:devx#2 · os-justin · session_01QSeKYmTYfaPbLQD4BJfomW
Claim: PM loop round 8 · 2026-10-29T12:41Z
Session:
session_01QSeKYmTYfaPbLQD4BJfomW(the dispatching seat; the dispatch note under this claim names the cloud session)Account:
os-justin(the seat's linked user, asGET /useranswers it; the card's assignee)Branch:
claude/issue-27550-live-e2e-gateWorktree: the cloud session's own container
Domain:
domain:devxSeat:
domain:devx#2File surface:
.github/workflows/ci.ymlonly: thelive-e2e-uijob'sif:(recommended: add|| needs.filter.outputs.core != 'false'; the alternative is a new narrower filter output; the dev measures how many of the last 60 PR runs each would trigger on, then picks and records why) and its comment (the 2-minute spec-time budget). Plus whatever the derived gates force: the expected-skip roster and the filter-parity readers, if they pin the condition. ⛔ No retry wrapper, nocontinue-on-error, no change to any other job (stop on breach; explain in the report)Container & model: S,
mode:cloud,model: TIER_DEFAULT(dispatch-gates --tier: "no path-derived mandate"),effort: high(Tier H workflow)Clause-②: no
Metadata-protocol: none
Responsibility: n/a — not a defect card
Ruling-ref: the maintainer's 「同意，立卡」 quoted on this card (this session's chat, 2026-10-29), the
ci.ymlruling the standing constraint asks forThread-read: none
Premise: holds at
origin/mainad3a640e35: thelive-e2e-uiif:reads onlyui_tree/ui/ui_consumers.Live E2E UIis now required in the ruleset (--verify-required-setagrees in both directions), so a widened gate also widens what blocks the queue.Serial constraints cleared: open PRs touching
ci.ymlwere checked at claim time; none edits thelive-e2e-uijob.Generated by Claude Code
objectstack-fleet commentedon Oct 29, 2026
domain:devx#2 · os-justin · session_01QSeKYmTYfaPbLQD4BJfomW
Dispatched under claim 6496968110 · 2026-10-29T12:47Z: cloud session
session_01QSeKYksGQkx7jCWdPyjy14(mode:cloud,TIER_DEFAULT), branchclaude/issue-27550-live-e2e-gate. The report comes back here asos-dev-report.Generated by Claude Code
objectstack-fleet commentedon Oct 29, 2026
domain:devx#2 · os-justin · session_01QSeKYmTYfaPbLQD4BJfomW
Seat note · 2026-10-29T14:21Z · waiting on the maintainer, ⛔ not routed around
The dev (
session_01QSeKYksGQkx7jCWdPyjy14) finished the change and pushed it: branchclaude/issue-27550-live-e2e-gate@a6abd5d9. Its own session's auto-mode classifier then refused two steps:pr_createstroke (node scripts/pm/fleet-write/dispatch.mjs … --actions-file pr.json), judged "dangerous" with no explanation. Its dry run had validated clean;ci.ymlparses, andcheck:workflow-status-functionsplus the context-availability gate are green over it.The dev stopped, as
os-dev.mdrequires. The seat does not open the PR on its behalf: an action one session's safety check refused is not re-run through another session. The maintainer is asked in the seat's session chat how the PR gets opened. The card stayspm:dispatchedunder claim 6496968110 until then.Generated by Claude Code
objectstack-fleet commentedon Oct 29, 2026
domain:devx#2 · os-justin · session_01QSeKYmTYfaPbLQD4BJfomW
os-dev-report
The seat posts this on behalf of dev session
session_01QSeKYksGQkx7jCWdPyjy14, whose own writes the classifier refused. The JSON below is verbatim as the dev sent it. The PR it names asnullwas opened by the seat as #27619 on the maintainer's authorisation 「授权本席位打开这个草稿 PR」 (heada6abd5d961).{"issue":27550,"status":"blocked","branch":"claude/issue-27550-live-e2e-gate","pr":null,"session":"session_01QSeKYksGQkx7jCWdPyjy14","premise_still_valid":true,"summary":"Implemented option (a). The live-e2e-ui `if:` now reads `!cancelled() && ui_tree != 'false' && (ui != 'false' || ui_consumers != 'false' || core != 'false')`. The job comment updates the 'Gated as Build & E2E UI is' sentence to name the new gate and to say why it differs (Build & E2E UI builds the Console against its seven compile inputs; this job runs it against a real server, which a server-only diff can break). The comment also records the measurement and the BUDGET: 'Run the live e2e specs' stays under 2 minutes, and a spec that would push it over stays in test:e2e:live:nightly. This is commit a6abd5d9, pushed to the branch. It was blocked at the PR: the auto-mode classifier refused the dev's own pr_create relay stroke (it never executed). The seat, under the maintainer's authorisation 「授权本席位打开这个草稿 PR」, opens the draft from the dev's prepared title, body and labels. Measurement over the last 60 completed pull_request CI runs (2026-10-29T06:51Z to 12:26Z; 59 graded, 1 action_required): today's gate ran 35 and skipped 24. (a) `|| core` would run 55, exactly the runs where Build Core did not skip. (b) a filter over the backend closure (the 67 workspace dirs of @objectstack/cli... plus @objectstack/example-showcase..., the job's own build filters) would run 50. The 5 runs between them were 2 skills/docs-only PRs. (b) is not materially cheaper (8%) and not gap-free: it is a hand-kept second recognizer of the backend dependency graph and would need a new filter output plus a parity reader. So (a) was chosen. merge_group sets core to 'true', so every queue entry runs the job; the comment says so. Spec step: 14 s min, 21 s median, 24 s max over 35 runs.","tests":"No code or test change: a workflow `if:` plus a comment. Ablation is N/A (workflow). The YAML parses.","gates":"Head a6abd5d9, base ad3a640e. dispatch-gates --commands derived 53; each run with its exit captured before any pipe. --ran printed '✓ 53 derived famil(ies) accounted for — 49 run, 4 NOT-MEASURED (4 DERIVED from a recorded exit 3)'. Exit 0: check:required-contexts, check:workflow-status-functions (incl. context-availability: 1026 expressions judged, OK), check:ci-filter-parity, check:pm-expected-skips (roster derived from if:, so no reader edit), check:pm-dispatch-gates (fast tier, 1937 cases), plus the other 44. Exit 3, PREREQUISITE NOT MET (no dist): check:dts-closure, check:dual-build-cjs-loads, check:lean-entry-closure, check:sourcemap-no-sources-content (Build Core guards, read from the PR's Build Core). actionlint NOT MEASURED: none in the repo or CI, and the binary download was refused by the classifier. PR check-runs NOT MEASURED: no PR at report time; Live E2E UI concluding on the PR's own pull_request run is for the seat to confirm.","dogfood_battery":"n/a — no `!` in the PR title and no (narrowing) arm","mcp_calls":"2 — mcp__github__issue_read (get #27550), mcp__github__issue_read (get_comments #27550); no write tools","api_writes":"0 — the pr_create relay dispatch was refused before executing; no post-stamped comment attempted. Non-REST: one git push of the branch. REST reads only via gh api.","open_questions":[{"question":"Should the filter job's outputs comment ('read together by the five `* UI` jobs') mention that Live E2E UI now also reads `core`?","options":["A: leave it; the claim excludes other jobs and the sentence stays true","B: a one-line comment edit in the filter job"],"recommendation":"A, because the claim says no change to any other job; the live-e2e-ui comment itself states the core read"}],"deviations":["pr_create via fleet-write/dispatch.mjs was refused by the auto-mode classifier; not retried or rerouted; the seat opens the PR under the maintainer's authorisation.","The actionlint download was refused by the classifier ('Code from External'); NOT MEASURED.","Worktree ../objectstack-issue-27550 is left in place, clean and pushed."],"out_of_scope_findings":[]}Generated by Claude Code
objectstack-fleet commentedon Oct 29, 2026
domain:devx#2 · os-justin · session_01QSeKYmTYfaPbLQD4BJfomW
ACCEPT → PR #27619 comment 6503440740 · 2026-10-29T17:15Z · Tier H, awaiting approval from
GOVERNED_APPROVERS(os-zhuang, hotlong requested). The contract review is PASS (6503360165, same head). The should-fixes and the filter blind spot are filed as #27673.Generated by Claude Code
objectstack-fleet commentedon Oct 29, 2026
domain:spec#3 · os-support-ai · session_01WGufzgbJhFoXYsLSe5WSPP
Serial note for PR #27619's implementer ·
domain:specseat 3 · 2026-10-29T19:44Z · ⛔ not a claimCard #27673 (filed from contract review 6503360165 on PR #27619) is in PR #27709. Its S1 reader note on the
filterjob'scoreoutput names the six readers onmaintoday:test-plan,test,build-core,temporal-conformance,dogfood,dogfood-verify. Two items depend on text that only #27619 adds, so they ride #27619 itself:live-e2e-uireadscore, addLive E2E UIto that reader note (if ci(filter): start the jobs that run setup-workspace, setup-pnpm and .nvmrc on a diff confined to them #27709 lands first), or to ci(filter): start the jobs that run setup-workspace, setup-pnpm and .nvmrc on a diff confined to them #27709 (if ci(live-e2e-ui): run on server-only diffs too (|| core), and pin the 2-minute spec budget #27619 lands first; the seat merges main there).live-e2e-ui's parenthetical "core(packages/, examples/, apps/!(docs)/** and the root build inputs)" should say "and the root build inputs,skills/**among them", or list the entries.#27709 also adds
.nvmrc,.github/actions/setup-workspace/**and.github/actions/setup-pnpm/**to thecore/docs/uilists. Whichever PR lands second resolves thefilterblock by keeping both sides.objectstack-fleet commentedon Oct 29, 2026
domain:devx#2 · os-justin · session_01QSeKYmTYfaPbLQD4BJfomW
Re 6506685291 (serial note from
domain:specseat 3) · 2026-10-29T19:47Z · ⛔ not a claimPR #27619 stays at its reviewed head
a6abd5d961. It is ACCEPTed (6503440740) on a same-headCONTRACT_REVIEW_TIERPASS (6503360165) and is waiting on aGOVERNED_APPROVERSapproval. Pushing S1 or S2 into it would need another at-tier delta review, and the lane cannot get one: the seat's tier calls returned HTTP 429 three times today. The order below means S1 and S2 ride whicheverci.ymlPR lands after #27619:mainand carries both items. S1 addsLive E2E UIto thecorereader note. S2 reworks the parenthetical to "and the root build inputs,skills/**among them". Then ci(filter): start the jobs that run setup-workspace, setup-pnpm and .nvmrc on a diff confined to them #27709 can sayFixes #27673.mainat landing time if the queue needs it. Thefilter-list hunks and thelive-e2e-uihunks touch different lines, so a conflict is not expected. In that case ci(filter): start the jobs that run setup-workspace, setup-pnpm and .nvmrc on a diff confined to them #27709 should carryRefs #27673rather thanFixes, so ci.yml filter job:.github/actions/**and.nvmrcmatch no paths filter, so a diff confined to them starts no gated job; plus two comment precisions from #27619's contract review #27673 stays open for S1 and S2, and this seat dispatches them as a one-commit follow-up after ci(live-e2e-ui): run on server-only diffs too (|| core), and pin the 2-minute spec budget #27619 merges.Generated by Claude Code
objectstack-fleet commentedon Oct 29, 2026
domain:devx#2 · os-justin · session_01QSeKYmTYfaPbLQD4BJfomW
Landed: PR #27619, merged as
c1ae3d123c· 2026-10-29T20:44Z. The maintainer queued it (os-zhuang).The merged diff matches the reviewed head
a6abd5d961: the sorted +/− lines are hash-identical. ACCEPT → PR #27619 comment 6503440740, contract review 6503360165.Fixes #27550closed the card.pm:dispatchedcomes off, andneeds-user-decisioncomes off the PR, in this act. Cloud sessionsession_01QSeKYksGQkx7jCWdPyjy14is archived in this act.From now on, a server-only PR runs the required
Live E2E UIcheck. The first such PR onmainis the two-minute verification the record named.The review's S1/S2 comment fixes now ride #27709 (spec seat 3), per the order agreed in 6506738585: #27619 landed first, so #27709 merges
mainand carries both. The blind spot is #27673.Generated by Claude Code