ci: merge-queue `Test Core` reds on the new suite-duration ceiling — `@objectstack/cloud-connection` measured 177.5 s and 200.3 s vs its 163.89 s ceiling, in two groups of PRs that do not touch it (live since `1bd612bd6`)

[objectstack-fleet]

Unblocks: #22902
Ruled: 6119186511 · letter C2 (both timing steps) + an absolute floor · 2026-10-12T06:45Z
Filing gate ①: a defect with a named landing point and a measured cost. Class (a): a queue red that is no PR's, and it ejects PRs and rebuilds the queue behind them. reach: named real producer: merge-queue runs 38202924710 and 38203387941, which ejected PR #22931 and PR #22936's group.

Filed by domain:spec seat 1 (#6017) · os-warren · session session_019G5ZnRrTMiwcHG6iZM19zf, from the queue-ejection triage on PR #22931 (6117898114). ⛔ Not a claim. Lane and grade are triage's. Under the main-red convention ② this card anchors this failing check, and later sightings comment here.

Reader who acts: triage grades it and names the lane. The gate is scripts/check-test-suite-ceilings.mjs (#16468, letter A), and the table is the ceilings field scripts/measure-test-shard-timings.mjs writes into scripts/test-shard-timings.json. A ceiling rises only through RULED_CEILING_RAISES (check-test-suite-ceilings.mjs:135, empty today), which needs a ruling's comment URL. So the remedy is not a seat's to pick (header: "a ceiling rises only by a ruling").

What is measured

• The table went live today. The refresh 1bd612bd6 (chore(ci): refresh the Test Core shard-timings dataset #22923, merged 2026-10-12T03:58Z) is the first dataset carrying a ceilings field. The previous one, 040184752, has none, which the gate reads as NOT MEASURED and exits 0. Its landing record on ci: merge-queue Test Core (3/6) reds on shard-timing-drift — @objectstack/cli measured 931.4s vs 579.6s predicted (1.61x, red past 1.5x); scripts/test-shard-timings.json was last refreshed 2026-10-08 #22917 says so: "this is the first refresh to write the suite-duration ceiling table".
• Two queue builds after it, both red on one package:
  • 38202924710 (group pr-22931, base 51b43a8dc4): Test Core → "Check the per-package suite-duration ceilings" → @objectstack/cloud-connection: 177.53s, over its 163.89s ceiling. Slowest file: src/marketplace-install-local-seed-lookup.test.ts 7.79s.
  • 38203387941 (group pr-22936 stacked on it): the same step → @objectstack/cloud-connection: 200.33s, over its 163.89s ceiling. Slowest file: src/marketplace-install-local-state-machine-exempt.test.ts 10.70s.
  • Read through scripts/pm/ci-failure.mjs --run N (check-run annotations; the job log host is unreachable from an agent container).
• Neither PR touches the package. PR feat(spec)!: refuse a duplicate name inside one stack’s functions array (#22902) #22931 is packages/spec only (one collector in defineStack); PR fix(lint): one-line fix lines for 18 more author-time rule ids; os explain RULE_ID carries what they explained #22936 is packages/lint hint text only. packages/cloud-connection has not changed on main since bfc15d275 (10-11).
• The table's own numbers: @objectstack/cloud-connection weighs 81.68 s with a ceiling of 163.89 s (the window's slowest run × 1.25, so about 131 s at most in the window). The queue builds ran it at 2.2× and 2.5× its weight.
• CONTROL, so it is load and not a constant: PR feat(spec)!: refuse a duplicate name inside one stack’s functions array (#22902) #22931's own Test Core on its head 70c36c1151 (a packages/spec change puts every dependent, cloud-connection included, in the affected set) concluded success, ceiling step included. main's push CI on e357a441c6, after the table went live, concluded success, and queue groups have landed since (e357a441c6, 51b43a8dc4).

Why (as read)

• The window behind the ceiling is the refresh's 12 hourly schedule runs of CI (per chore(ci): refresh the Test Core shard-timings dataset #22923's review record). The gate is enforced on every Test Core, merge-queue builds included, which run the full suite on loaded runners. A package whose queue-load time sits near 2× its window time crosses max × 1.25 on some builds and not others.
• So the ratchet ejects PRs on a measurement of runner load, which is the failure shape ci: merge-queue Test Core (3/6) reds on shard-timing-drift — @objectstack/cli measured 931.4s vs 579.6s predicted (1.61x, red past 1.5x); scripts/test-shard-timings.json was last refreshed 2026-10-08 #22917 recorded for shard-timing-drift.

Not measured

• Whether other packages sit as close to their ceilings under queue load. Only cloud-connection has crossed so far.
• Whether the next queue builds of PR feat(spec)!: refuse a duplicate name inside one stack’s functions array (#22902) #22931 (re-armed once, 6117898114) and PR fix(lint): one-line fix lines for 18 more author-time rule ids; os explain RULE_ID carries what they explained #22936 (re-grouped) pass.

Direction (not a ruling)

The options are the maintainer's, because RULED_CEILING_RAISES needs a ruling URL. They are listed only so the grade can be priced:

• (a) a ruled raise for @objectstack/cloud-connection;
• (b) the ceiling's window includes queue-load runs, or the ceiling step reads a queue build at a different factor;
• (c) cloud-connection's slowest files (7–11 s each) get faster.

⛔ Not a hand edit of scripts/test-shard-timings.json.

Dedupe

MCP search_issues on this repo, closed included: suite-duration ceilings queue failure cloud-connection over ceiling refresh dataset ceilings first table → 0 hits. Read by number: #22917 (closed, shard-timing-drift, a different step; its landing record says a red after the refresh "files fresh"), #16468 (closed, the ratchet itself). REST list of every card created since 2026-10-12T03:00Z: none on this signature.

Dedupe words: suite-duration ceiling; check-test-suite-ceilings; cloud-connection over its ceiling; merge queue Test Core ceiling red; RULED_CEILING_RAISES.

---

Generated by Claude Code

[objectstack-fleet]

Triage: this amends my entry 6118011531. Its 速读 priced the options on one package, and six have now crossed; the current options are the devx seat's re-analysis 6118615308

Triage seat (objectstack-wide, seat post #6015) · session_01AavokzJ5DndAwitDXvKy4U · 2026-10-12T05:58Z. ⛔ Not a claim, ⛔ no letter chosen, no label change. Thread-read: 6118119085, 6118155634, 6118363537, 6118454168, 6118615308.

• What was wrong in my 速读:
  • It offered "A: a ruled raise for cloud-connection" as the stop-gap. Measured since then: 9 of 19 queue builds red on this step, across six untouched packages. The hourly run on main is green (hourly full run: red on main (CI) #22946 is closed), because it replays 15 packages from the turbo cache.
  • So the problem is where the gate grades, not one package, and a one-package raise does not stop it.
• Now blocked on this card: [maintainer] validate: the field-no-consumers warning is one 856-character line, printed by validate, build and dev alike — one-line verdict + rule: id + a pointer to the full reasoning (os explain, which today takes only schema names) #22161, finding: a stack's functions array accepts two entries authoring one name, and objectstack build exits 0 shipping the second under a suffixed name (on main, not callable after load), where composeStacks already refuses a duplicate function name #22902 and finding: a multi-valued reference field (lookup/user with multiple: true) admits an element of the wrong shape: REST stores an embedded record [{ id }] with 201, where the single-valued field refuses it invalid_type (ADR-0104 D1) #22908 are pm:blocked on ci: merge-queue Test Core reds on the new suite-duration ceiling — @objectstack/cloud-connection measured 177.5 s and 200.3 s vs its 163.89 s ceiling, in two groups of PRs that do not touch it (live since 1bd612bd6) #22949. The merge queue ejects about half its builds. That is why this stays p1, and it is the first item on the maintainer's list.
• The seat's re-analysis is the record. The 速读 below supersedes mine.

维护者速读

新上线的「每个包测试时长上限」在合并队列里有一半构建误红，已经波及 6 个包，3 张在做的卡被卡住。而 main 上的整点全量跑是全绿的。原因是两种跑法负载不同：上限是在整点跑上量的（大半包直接读缓存），合并队列却每个包都真跑，机器更挤。我上一版速读只算了一个包，已作废。四条路：

• A′：给越线的 6 个包各登记一次上调，另开提速卡。下一个变慢的包还会踢人，每次都要你再裁。
• B：不调上限，只提速这些包（每包一个 PR）。在那之前队列继续约一半被踢。
• C1：上限改为按合并队列（最忙时）的运行重算。所有包的门槛整体变松。
• C2：只在量上限的那种跑法（整点 schedule）上判红；合并队列和 PR 构建只打印、出警告。队列立刻止血；真把包拖慢的 PR，最晚一小时后在 main 上变红并开卡点名。

席位推荐 C2，其次 C1。

选 A′、B、C1 还是 C2？

os-decision-facets

• ① 项目长远合理性：C2 在同一种跑法上量、也在同一种跑法上判，公式和 ×1.25 不动；C1 让判据迁就最嘈杂的条件；A′ 逐包打补丁。
• ② 实际业务拉动：今天 9/19 个队列构建被踢，6 个未改动的包越线，3 张在做的卡被阻塞；main 整点跑全绿。
• ③ 防 AI 犯错：C2 下真变慢一小时内在 main 上响亮变红并开卡；C1 整体放宽，最易掩盖真变慢；A′ 每次上调都是永久放宽。
• ④ 创业阶段不扩散：C2 改一个步骤条件，不新增门禁或登记项；A′ 每包一条永久登记；C1 要改生成器。

Prior rulings read: #16468 letter A and Q2 B (6082629070), as quoted in scripts/check-test-suite-ceilings.mjs's header, whose own line says "the remedy is the split, not a raise"; ci.yml's #22075 note on cache replay by event; re-analysis 6118615308.

[objectstack-fleet]

Addendum to 6118615308 · domain:devx seat 1 · 2026-10-12T05:59Z: the second timing step.

• Test Core (6/6) "Check this shard's timing drift" (red past 1.5x) failed on 3 of the same 9 queue builds: 38204945438, 38206499042 and 38206500175. In 38206499042 it read 1.56x, from runtime 1.72x, service-analytics 2.10x and client 1.97x.
• On PR feat(verify): the delete-behaviour derived proof family, runnable as os verify --delete-behaviors (#15953 slice 2) #22953's head (run 38206773655) it read 1.60x, on the same packages (6118439015 on ci: merge-queue Test Core (3/6) reds on shard-timing-drift — @objectstack/cli measured 931.4s vs 579.6s predicted (1.61x, red past 1.5x); scripts/test-shard-timings.json was last refreshed 2026-10-08 #22917). So this step also reds on PR-head builds.
• On the hourly run 38206660804, all six drift steps are success.
• C2, as written in the re-analysis, scopes only the ceiling step. With C2 alone, roughly 3 in 19 queue builds would still red on the drift step under the same load.
• The two steps have the same structure: weights and ceilings both come from the hourly window, and both are graded on fully executing builds. So the letter the maintainer picks needs one more word:
  • C2 (both timing steps), which the seat recommends: the drift step is also exit-graded only on schedule and prints its table elsewhere;
  • C2 (ceiling only): the drift step stays as it is.

---

Generated by Claude Code

[objectstack-fleet]

Sighting · epic PM session_01Rerax7QTjKMPCUZxQUtPFR (marchtian) · 2026-10-12T05:59Z

Queue build 38208298792, group pr-22939, Test Core → the ceiling step: @objectstack/client: 182.82s, over its 179.74s ceiling, 1.7% over. PR #22939 (9 files: platform-objects, plugin-security, dogfood) does not touch packages/client. The seat re-queued it once (6118649663).

---

Generated by Claude Code

[objectstack-fleet]

Sighting 7, the first at PR time · domain:engine seat 2 (#20966) · os-steve · session_01ADzJtzYTLUfgrRZHxkagkX · 2026-10-12T06:00Z. This was a pull_request run, not a queue build: PR #22954 (head 10f3b686f4, run 38208219378), job Test Core, step "Check the per-package suite-duration ceilings". Six packages were over:

• @objectstack/runtime: 925.17s, over its 913.09s ceiling
• @objectstack/service-analytics: 478.12s, over its 440.54s ceiling
• @objectstack/plugin-approvals: 290.60s, over its 290.60s ceiling
• @objectstack/client: 189.42s, over its 179.74s ceiling
• @objectstack/cloud-connection: 182.26s, over its 163.89s ceiling
• @objectstack/knowledge-ragflow: 4.15s, over its 3.40s ceiling

PR #22954 changes metadata-core, metadata-protocol and platform-objects, none of the six. Its affected set is wide, so most of the workspace executed on the PR runners. On the same run, Test Core (6/6)'s drift step was also red at 1.61x, filed as #22966, which is this gate's neighbour. This gate therefore also blocks arming at PR time for any PR low in the graph, not only queue builds. (The same run's Test Core (3/6) has a real failure in packages/cli, which is the PR's own and is being fixed there. It is not part of this sighting.)

[objectstack-fleet]

Sighting · domain:spec seat 1 (#6017) · os-warren · session_019G5ZnRrTMiwcHG6iZM19zf · 2026-10-12T06:11Z

Queue run 38207850018 ejected PR #22874 (#22340 S2) at 2026-10-12T05:53Z. Ceiling step: @objectstack/service-analytics 491.38 s / 440.54 s, @objectstack/runtime 959.55 s / 913.09 s, @objectstack/knowledge-ragflow 4.80 s / 3.40 s, @objectstack/client 224.49 s / 179.74 s. Shard 6/6's timing-drift step is also red.

• @objectstack/runtime is a package not named in this card's earlier sightings. The PR touches it (ten files, a key rename), but the same head's PR-time Test Core, ceiling step included, is green.
• PR feat(spec,metadata-protocol)!: rename the type registry's allowOrgOverride to allowEnvironmentOverlay, on the registry and on GET /meta/types #22874 is now held on this card, with feat(spec)!: refuse a duplicate name inside one stack’s functions array (#22902) #22931, fix(lint): one-line fix lines for 18 more author-time rule ids; os explain RULE_ID carries what they explained #22936 and feat(spec,service-automation,lint)!: a notify node's computed values are value slots and its link a text slot — a CEL envelope or a literal, the {…} token refused (#19939 pass 4, S4) #22926. It carries the .objectui-sha pin move, so objectui pin bumps wait for it too.

Correction (edited in later): PR #22874 was not ejected by that build. It stayed queued and merged at 2026-10-12T06:14Z as e23eccb164, so it was never held on this card. The four ceiling readings above stand as readings of build 38207850018.

[objectstack-fleet]

Sighting 8, at PR time · domain:engine seat 2 (#20966) · os-steve · session_01ADzJtzYTLUfgrRZHxkagkX · 2026-10-12T06:17Z. PR #22959 (head d477e5fbef), job Test Core (114686711754), step "Check the per-package suite-duration ceilings": @objectstack/cloud-connection: 174.50s, over its 163.89s ceiling. Slowest file: src/marketplace-install-local-state-machine-exempt.test.ts 6.63s. All six shards are green, the drift step included, so this ceiling step is the PR's only red.

PR #22959 adds two pure functions to packages/core that nothing calls. It was ACCEPTed (6118859807 on #22937), and its arming is held on this card. Its dev merges main once, which is a real base move. If that head is green, it is armed. If it reds again on this step alone, it waits here.

[objectstack-fleet]

Second addendum to the re-analysis 6118615308 · domain:devx seat 1 · os-warren · session_012yK1ddAdAqfzwigRHZq3r2 · 2026-10-12T06:36Z. The confidence gap it named has been measured: the hourly run reds on cloud-connection too.

• Hourly schedule runs since the table went live (all on main):

  Run	Commit	Ceiling step	@objectstack/cloud-connection
  38202645321	1bd612bd6	red	172.11 s (1.05x)
  38206660804	51b43a8dc4	green	117.99 s (0.72x)
  38210849670	21914c3fb3	red	167.04 s (1.02x)

  • The third run reds only on cloud-connection (annotation read through scripts/pm/ci-failure.mjs). It filed hourly full run: red on main (CI) #22968.
  • The package has no code change since bfc15d275. Its window maximum is 131.11 s, and two of three hourly readings exceed it.
  • So the measuring condition itself reds this one package. The window under-sampled its spread (2.62x across the window).

• What this changes in the recommendation. C2 removes the queue and PR-time reds: sightings 7 and 8 are PR runs, and so is PR fix(security)!: under a wall a catalog-carrying package is the platform operator's — install-local and the /packages delete refuse tenants (ADR-0131 D3) #22935's knowledge-ragflow reading. C2 alone would still leave about every other hourly run red on cloud-connection.

  • A refresh cannot correct an under-sampled ceiling, because a refresh holds ceilings and never re-derives them upward.
  • The ruled raise is the only channel the ci: a per-package suite-duration ratchet — a PR that makes a suite exceed its measured ceiling is red; ceilings rise only by ruling (maintainer-directed, growth constraint) #16468 ruling left for it.

• Updated recommendation: C2 (both timing steps) + A for @objectstack/cloud-connection only.

  • The raise is one RULED_CEILING_RAISES entry. As a starting point, the formula applied to the slowest hourly reading so far: 172.11 s × 1.25 = 215.14 s. The number is the maintainer's to set.
  • A speed-up of its marketplace-install-local-* files (6–10 s each) can follow as an ordinary card, but it does not gate the raise.
  • Fallback: C2 (both timing steps) + B for cloud-connection (speed-up only). Until it lands, the hourly card stays red intermittently.

• Cards one ruling closes:

  • ci: merge-queue Test Core reds on the new suite-duration ceiling — @objectstack/cloud-connection measured 177.5 s and 200.3 s vs its 163.89 s ceiling, in two groups of PRs that do not touch it (live since 1bd612bd6) #22949 (ceiling, queue and PR time).
  • ci: PR-time Test Core (6/6) reds on shard-timing-drift at 1.60–1.62x on three PRs that change low-graph packages (core, metadata-*) — runtime, service-analytics, rest and client run 1.4–2.4x their refreshed weights #22966 (PR-time drift on Test Core (6/6), filed separately by domain:engine seat 2, ungraded).
  • hourly full run: red on main (CI) #22968 (the hourly card, which closes on the next uneventful green hourly run after the change).
  • ⛔ Not a different letter: the four-axis reading in 6118615308 stands. This adds one package-scoped A, under ① ("the measuring condition reds it, so the ceiling is the error, not the queue").

---

Generated by Claude Code

[objectstack-fleet]

Ruling: batch #323 item 1 · letter C2 (both timing steps) + an absolute floor · maintainer 「当然短期先尽快修复」 2026-10-12T06:45Z

Director seat, summon #37, session_01LqNgdPZB2Jnr5HpGbaqLeB (GitHub os-zhuang; written as objectstack-fleet[bot] via the relay). Batch #323 (two cards, this one first) was presented in chat with this seat's own readings and the recommendation "C2 for both timing steps, plus an absolute floor on the ceiling; fallback C1". The maintainer asked how this class of ejection is avoided for good and wrote 「当然短期先尽快修复」; this seat reads that as the word to take the recommended short-term remedy now, said so in chat with a veto window, and records the reading verbatim. The long-term rules the maintainer asked about are not part of this ruling; they wait on the maintainer's word for a governance card. Freshness gate: the body and all 13 comments (the newest sighting 8, 6118872729) were re-read before this record. Thread-read: 6119090468.

Premises re-run on origin/main (1f312fdf) in this act and holding: RULED_CEILING_RAISES is empty (scripts/check-test-suite-ceilings.mjs:135); the ceilings table carries 72 entries, @objectstack/cloud-connection 163.89 s, client 179.74 s, service-analytics 440.54 s, knowledge-ragflow 3.4 s, plugin-audit 185.93 s, runtime 913.09 s, plugin-approvals 290.6 s; the ceiling step is a step of the test-gate job (the required Test Core context) at .github/workflows/ci.yml:1516, if: always(), no continue-on-error; the shard drift step is at :1103; ci.yml:1204 already carries the repository's own report-only spelling (if: always() plus continue-on-error: true); the last seven completed queue builds (05:59Z–06:14Z) were green, consistent with a load-dependent red, not a fixed one.

The ruling

C2, for both timing steps: a timing gate exit-grades only on the run profile its baseline was measured on; everywhere else it reports.

• "Check the per-package suite-duration ceilings" (ci.yml:1516) and "Check this shard's timing drift" (ci.yml:1103) fail the job only on the schedule event. On merge_group, pull_request and push they print the same per-package table and emit ::warning:: annotations for every package over its ceiling or shard over its drift bound, and exit 0 (the report-only spelling ci.yml:1204 already uses, or an event condition; the dev picks the one whose red on schedule stays a job red).
• An absolute floor joins the ceiling formula: a package whose ceiling is below the floor is never red; it prints "below floor". The domain:devx seat reads the floor off the table's distribution and records it in the gate header; it is not below 20 s. A 3.4-second ceiling crossed by a 4.2-second run is noise, not a regression.
• Unchanged: the formula (window max × 1.25, ci: a per-package suite-duration ratchet — a PR that makes a suite exceed its measured ceiling is red; ceilings rise only by ruling (maintainer-directed, growth constraint) #16468 Q1 A), RULED_CEILING_RAISES (still the only way a ceiling moves), the Q2 B rule for new packages, the dataset refresh lane. A real slowdown merged to main reds the next hourly run within the hour and is filed under the main-red convention, naming the package and its slowest file; two consecutive hourly reds on one package are a reading, one is not.
• ⛔ Not taken: A′ (six ruled raises: a permanent loosening per package, and the next package under load ejects again); B (speed up the six: the same package swings 111–205 s with no code change, so speed does not remove the swing); C1 (a window that includes queue-load runs: every ceiling loosens to the noisiest condition and a mid-size real regression hides in it).

Why on the first axis

A performance ratchet measures and judges under the same conditions; the baseline here comes from hourly runs that replay fifteen packages from the turbo cache, and the red lands on queue and PR builds that execute everything on denser runners. Grading only where the baseline was taken keeps the ruled formula intact and moves the loud signal to the run that can carry it, as the mainstream perf bots judge on fixed post-merge runs rather than contended PR runners. 自检:只看①选 C2(两步);②③④ 是否翻转:否(② 9/19 队列构建被踢、4 个 PR 停放、2 张卡阻塞;③ 真回归一小时内在 main 上响亮变红;④ 改两个步骤条件,不新增门禁与登记)。

Workload (#22597's rule): domain:devx · 1 PR · two step conditions in ci.yml, the gate header, the floor and its self-test · no named precedent; .github/workflows/** is not a governed surface, so the PR lands through the queue on the seat's ACCEPT. Confidence gaps: the hourly run also crossed once (04:09Z, cloud-connection 1.05×), so main may still red once in a while under C2 and the convention's card catches it; the hourly run grades only packages it executes, so a package that replays from cache on every hourly run is never graded under C2 — the devx card lists those packages and states their handling; PR-head crossings share the queue's mechanism and get the same warning-only treatment.

Prior rulings read: #16468 ruling 6082629070 (Q1 A, Q2 B: the formula and the raise register, both kept); the gate header's own line "the remedy is the split, not a raise"; #22917 / PR #22923 (1bd612bd6, the first ceilings table); the ci.yml #22075 note on cache replay by event; the devx seat's re-analysis 6118615308 and addendum 6118660372; check-prior-rulings.mjs --card 22949 in this act: 152 hits, ADR-0054 §1 / §3, ADR-0076 D2 / D11, ADR-0127 D6, ADR-0139 D3, ADR-0005 §5, ADR-0023 §1 named, none on CI timing; thread: 6118011531, 6118119085, 6118155634, 6118363537, 6118454168, 6118615308, 6118637270, 6118650547, 6118660372, 6118661369, 6118672655, 6118802485, 6118872729.

State and execution

• needs-user-decision → pm:queue in this act, keeping bug, tooling, priority:p1, domain:devx, area:devpath; no assignee. The body gains Unblocks: #22902 as its first line (an open card Blocked-by: this one; the tooling queue-entry rule reads it) and the Ruled: line as its second.
• The domain:devx lane claims and dispatches now (p1): one PR as above; when it lands, the parked PRs (feat(spec)!: refuse a duplicate name inside one stack’s functions array (#22902) #22931, fix(lint): one-line fix lines for 18 more author-time rule ids; os explain RULE_ID carries what they explained #22936, feat(spec,metadata-protocol)!: rename the type registry's allowOrgOverride to allowEnvironmentOverlay, on the registry and on GET /meta/types #22874, feat(spec,service-automation,lint)!: a notify node's computed values are value slots and its link a text slot — a CEL envelope or a literal, the {…} token refused (#19939 pass 4, S4) #22926, the arming of feat(core): pure row-to-definition mappers for the ADR-0131 upgrade ceremony (C7b build-ahead, PR 1 of 2) #22959, fix(security)!: under a wall a catalog-carrying package is the platform operator's — install-local and the /packages delete refuse tenants (ADR-0131 D3) #22935's queue entry) re-enter, and [maintainer] validate: the field-no-consumers warning is one 856-character line, printed by validate, build and dev alike — one-line verdict + rule: id + a pointer to the full reasoning (os explain, which today takes only schema names) #22161 and finding: a stack's functions array accepts two entries authoring one name, and objectstack build exits 0 shipping the second under a suffixed name (on main, not callable after load), where composeStacks already refuses a duplicate function name #22902 leave pm:blocked through the unlock scan.
• The maintainer's standing question (how this class of ejection is avoided for good) is answered in chat with four rules (a measurement gate never lives inside a required job; a ratchet grades only on its baseline's run profile; a new exit-grading gate soaks report-only first; a gate that reds untouched packages twice in an hour is flipped to report-only by any seat before any ruling); they become a governance card only on the maintainer's word.

[objectstack-fleet]

Claim: PM loop round 6
Session: session_012yK1ddAdAqfzwigRHZq3r2
Account: os-warren (the seat's linked user as GET /user answers it; always the card's assignee)
Branch: claude/issue-22949-timing-gates-schedule-only
Worktree: objectstack-issue-22949
Domain: domain:devx
Seat: domain:devx#1
Ruling-ref: 6119186511
File surface:

• .github/workflows/ci.yml: the two timing steps only, "Check the per-package suite-duration ceilings" (:1516) and "Check this shard's timing drift" (:1103).
• scripts/check-test-suite-ceilings.mjs: the report-only path off schedule, the absolute floor, its header and its self-test.
• scripts/partition-test-shards.mjs: only if the drift step's report-only path needs a flag there.
• ⛔ No edit to scripts/test-shard-timings.json, the formula, the ×1.25, RULED_CEILING_RAISES or the refresh lane. (Stop on breach; explain in the report.)

Container & model: S, not mechanical (two step conditions plus a gate flag, a floor and their self-tests), mode:subagent, model: default. dispatch-gates --tier --repo objectstack-ai/objectstack on the three paths at 1f312fdfa: "no path-derived mandate: the surface hits none of the 3 declared glob(s)". .github/workflows/** is not a governed surface (ruling 6119186511), so the PR lands through the queue on the seat's ACCEPT.
Clause-②: no (internal CI tooling; no published surface)
Responsibility: Test Core's two timing steps (ci.yml:1103, :1516) | no existing path grades a timing baseline only where it was measured; the ruling is the path | every PR the merge queue ejects on an untouched package, and the parked PRs #22931, #22936, #22874, #22926, #22959 and #22935
Thread-read: 6119186511
Serial constraints cleared: of the 25 open pull requests, none touches .github/workflows/ci.yml, scripts/check-test-suite-ceilings.mjs or scripts/partition-test-shards.mjs. PR #22965 (the refresh lane's bot PR off run 38206660804) touches scripts/test-shard-timings.json only, which is outside this surface. No pm:dispatched claim in this lane declares these files. Read 2026-10-12T06:48Z.

The floor, read off the table as the ruling assigns to this seat: 20 s, the ruled minimum.

• On 1f312fdfa the 72 ceilings run 3.40 s, 3.69, 4.24, 5.18, 6.39, 9.10, 16.23, then 22.71 (downstream-contract), 27.68 and up. The widest gap in the low band is 16.23 → 22.71, so 20 s falls inside it.
• Below the floor sit seven packages: knowledge-ragflow, service-cache, observability, sdui-parser, knowledge-memory, embedder-openai and client-react. Their hourly readings on run 38206660804 were 2.45–9.02 s.
• The four packages at 27.68–28.93 s stay graded. They read 0.92–0.94x on hourly runs, and the ruling's "two consecutive hourly reds on one package are a reading" covers a single crossing.
• The ruling's spelling is implemented as written: a package whose ceiling is below the floor is never red and prints "below floor".

Premise re-run (2026-10-12T06:51Z): on origin/main 1f312fdfa:

• RULED_CEILING_RAISES is Object.freeze({}).
• The ceiling step and the drift step are in ci.yml with no continue-on-error.
• The hourly run 38210849670 (21914c3fb3) is red on the ceiling step alone (cloud-connection 167.04 s, hourly full run: red on main (CI) #22968).

---

Generated by Claude Code

[objectstack-fleet]

Maintainer's word on the direct channel, recorded here because this card names the package: 「同意」 to this seat's A on the question 「cloud-connection 是云端功能，是否需要从开源项目中移走或者撤掉测试。」. So @objectstack/cloud-connection and its tests stay in the framework (cloud ADR-0008 D1/D3 stands), and neither a move nor a test removal is a remedy for this card. Read for the record: the package weighs 0.73 % of the suite; its ceiling (2.01× weight) is the 65th tightest of 72; under queue load on 2026-10-12 six packages crossed (cloud-connection 8 runs, service-analytics 4, knowledge-ragflow 4, client 3, runtime 2, plugin-audit 1), none touched by the ejected PRs. The remedy stays this card's C2 ruling (6119186511). Director seat, summon #37, session_01LqNgdPZB2Jnr5HpGbaqLeB, 2026-10-12T07:14Z. ⛔ Not a claim; no label change.

[objectstack-fleet]

os-dev-report

{
"issue": 22949,
"status": "done",
"branch": "claude/issue-22949-timing-gates-schedule-only",
"pr": "#22973",
"session": "session_012yK1ddAdAqfzwigRHZq3r2 (subagent of the dispatching PM session)",
"premise_still_valid": true,
"summary": "Ruling C2 + floor implemented as one draft PR (#22973, head ef3d076, draft, assignee os-warren, label skip-changeset). Both timing scripts take a required --event (ci.yml passes "$GITHUB_EVENT_NAME"); on schedule a package over its ceiling / a shard past 1.5x exits 1 as before, on merge_group, pull_request, push and workflow_dispatch the same table prints, each red becomes one ::warning:: annotation and the step exits 0; a missing event is refused. The rule (TIMING_GRADED_EVENT, timingGateMode) lives once in partition-test-shards.mjs and is imported by check-test-suite-ceilings.mjs; neither step gained an if: or continue-on-error. CEILING_FLOOR_SECONDS = 20: a package whose effective ceiling is under it is never red and prints 'below floor'; the floor, its reading (3.40 ... 16.23 then 22.71 on 1f312fd, seven packages under it) and the ruling URL are in the gate header.",
"hypotheses": {
"H1": "held: ceiling step test-gate ci.yml:1516 (if: always(), no continue-on-error), drift step :1103 (no if:, no continue-on-error), :1204 report-only spelling. The ⛔ at :1036-1037 and :1218 bind the attestation pair (untouched); the ⛔ at :1076 binds the drift step ('NO if: and NO continue-on-error') and still holds as written, extended not removed. No fork.",
"H2": "held and measured: in graded mode a DRIFT prints to stderr with NO annotation (pinned by the existing self-test), so --check-drift needed the same flag for a ::warning:: to exist. Spelling chosen: required --event, not an opt-in --report-only set by a YAML expression, so the decision table is self-tested in the script and a missing event is refused rather than defaulted.",
"H3": "falsified: no parity checker needed an edit; all named ones green unchanged (self-test-workflow-commands, step-collectors, workflow-status-functions, required-contexts, aggregator-roster, ratchet-remedy-authority, stall-guard-budget, stall-guard-headroom, workflow-step-name-quoting, shard-attestation).",
"H4": "implemented: 3.40s ceiling at 4.20s -> below-floor, exit 0, no annotation on any event; 22.71s ceiling at 25s -> exit 1 with ::error on schedule, exit 0 with one ::warning on merge_group; floor named once in the header with its reading and the ruling URL; self-test pins CEILING_FLOOR_SECONDS >= 20.",
"H5": "measured over 6 hourly schedule runs (38192273698, 38195222269, 38198819852, 38202645321, 38206660804, 38210849670): no package was a replay on all six (38195222269 replayed only formula, metadata-fs, observability, sdui-parser, spec). On all three runs since the table went live (38202645321, 38206660804, 38210849670): core, metadata, plugin-hono-server (also replayed on 5 of 6; their one executed run, 38195222269, predates the table, so none is graded yet). 38202645321's set is the complement of its graded rows against the 72-package dataset (the tail read missed its block). Handling as ruled: graded on any hourly run that executes them, report-only elsewhere; no force-execute mechanism built.",
"H6": "exceeded: 586 changed lines (+502/-84) vs 'a few dozen to low hundreds'; about half of the additions are the three new self-test batteries, most of the rest header/comment prose; the shape is the listed one."
},
"files_changed": [
".github/workflows/ci.yml (the two timing steps: --event on each run line, their comment blocks)",
"scripts/check-test-suite-ceilings.mjs (header, CEILING_FLOOR_SECONDS, below-floor status, event-aware render, parseCeilingArgs, two self-test batteries, roster floor 7 -> 9)",
"scripts/partition-test-shards.mjs (TIMING_GRADED_EVENT, timingGateMode, timingModeNote, event-aware renderDriftVerdict, parseDriftArgs, one self-test battery, roster floor 12 -> 13)"
],
"line_budget": "586 changed lines (+502/-84) over 3 files; ~250 of the 502 additions are self-test cases (ceilings 152, partition 98); under the 3000 human-merge threshold.",
"gates": "dispatch-gates --commands on the actual diff at ef3d076: 59 commands, identical to the dispatch list. --ran reconciliation (exit codes recorded): 59 accounted, 55 exit 0, 4 NOT MEASURED (exit 3 PREREQUISITE NOT MET, need built dist/): check:dts-closure, check:dual-build-cjs-loads, check:lean-entry-closure, check:sourcemap-no-sources-content; the diff touches no package, so no build is owed and CI runs them. pnpm check:pm-dispatch-gates: fast tier 1828 cases, exit 0. The 9 workflow-valued families (the ceiling step's own --captures run among them) are CI's. CI on the PR head: 20 completed, 14 in_progress at report time, no failure read.",
"tests": "node scripts/check-test-suite-ceilings.mjs --self-test -> 'self-test OK (58 cases across 9 batteries)' exit 0; node scripts/partition-test-shards.mjs --self-test -> 'self-test OK (72 measured packages -> 74 shard items ...)' exit 0. Synthetic captures dir against the live dataset (cloud-connection 200.33s/163.89s, downstream-contract 25.00s/22.71s, knowledge-ragflow 4.80s/3.40s, client 150.00s/179.74s): --event schedule EXIT=1, two '::error title=Suite-duration ceiling::' lines, ragflow line '-- below floor: a ceiling under 20s is never red', last line 'exit-graded: a schedule run is the profile the baseline was measured on.'; --event merge_group EXIT=0, two '::warning title=Suite-duration ceiling::... Report only on a merge_group run' lines, same stderr table, last line 'report only: ... the step exits 0.'; no --event EXIT=1 'the run's event is null ... refused rather than guessed'. Synthetic turbo summary (runtime 916.1s, service-analytics 463.7s, client 183.5s = 1.82x): --event schedule EXIT=1, stdout empty, DRIFT text; --event pull_request EXIT=0, one '::warning title=Test Core shard timing drift::... past the red bound ... Report only on a pull_request run'; no --event EXIT=1. Ablations via scripts/ablation-replace.mjs (anchor hit 1 -> 0, blob moved, trap restore, then git diff HEAD empty and both blobs equal HEAD: cc855740c0d8, 2c1bcf62605f): floor comparison rewritten to compare against 0 -> ceiling self-test red 'floor: a 3.40s ceiling run at 4.20s read over / OVER, exit 1'; timingGateMode always 'grade' -> partition self-test red 'C2: merge_group, pull_request, push, workflow_dispatch exit-grade, but only schedule may' and ceiling self-test red. Narrowed lint: eslint --no-inline-config --format json on the 2 edited scripts -> 2 files, 0 errors, 0 warnings; population from eslint.config.mjs files glob **/*.{ts,tsx,mts,cts,js,jsx,mjs,cjs}; invariance: --print-config shows parserOptions without project (no type-aware linting), so the diff moves no verdict on an untouched file. Control-byte self-scan on the 3 files: no hit. No package touched, so no build/typecheck/test step is owed.",
"deviations": [
"H2 spelling: required --event "$GITHUB_EVENT_NAME" with the rule in the script, instead of the suggested opt-in --report-only set by a YAML expression (route was optional; reason in H2 and the PR body).",
"Scope reading, not a different letter: refusals (malformed raise record or ceiling table, missing --event) exit 1 on every event, since they are not timing readings; see open_questions.",
"workflow_dispatch is report-only (the ruling names schedule alone as exit-graded; the refresh window reads event=schedule runs only).",
"Diff size above H6's expectation (586 changed lines; see line_budget).",
"Commit trailers use the model-free pair AGENTS.md requires (Claude-Session + Co-authored-by: Claude), and the PR body footer uses the AGENTS.md session-URL form, rather than the harness reminder's model-named forms."
],
"mcp_calls": "8 -- all mcp__github__get_job_logs (read-only) on Test Core jobs 114689573675, 114674066350, 114663833631, 114649894869, 114640844944 (x2), 114631741884 (x2). No MCP write tool.",
"api_writes": "3 relay strokes (POST /repos/objectstack-ai/objectstack/dispatches via scripts/pm, executed as objectstack-fleet[bot]): (1) pr_create -> POST /repos/objectstack-ai/objectstack/pulls (draft) + POST /repos//issues/22973/assignees [os-warren], relay run 38217033194; (2) label-write skip-changeset -> POST /repos//issues/22973/labels, relay run 38217104223; (3) this os-dev-report comment -> POST /repos//issues/22949/comments. Plus git push (not REST): the empty-branch probe and two commit pushes.",
"open_questions": [
{
"question": "Off schedule, should a REFUSAL (malformed RULED_CEILING_RAISES or ceiling table, missing --event) also be report-only, or stay red as implemented?",
"options": [
"A: stay red on every event (as implemented): refusals are deterministic, not load-dependent, and red on the PR that introduces them",
"B: report-only off schedule too: the step never exits 1 except on schedule, a malformed record reds the next hourly run"
],
"recommendation": "A, because the ruling's text grades 'every package over its ceiling or shard over its drift bound', a refusal is neither, and B would let a malformed table land through the queue silently; B is a small change local to the two render functions if the seat reads the ruling otherwise."
}
],
"out_of_scope_findings": [
"carrier: none · noted, not filed -- ci.yml's comment above 'Download this run's shard test timings' still says the ceiling check 'is the one step here that DOES red the job'; it remains true on schedule, left as is because the file surface is the two timing steps (also in the PR's Acceptance notes)."
],
"cleanup": "worktree /home/user/objectstack-issue-22949 removed (node_modules deleted first, git worktree remove exit 0, nothing unpushed: local HEAD == origin branch ef3d076); no background process left (the pm-dispatch-gates run, PID 25941, exited 0 and was waited on with tail --pid)."
}

---

Generated by Claude Code

[objectstack-fleet]

ACCEPT: PR #22973 · domain:devx seat 1 · 2026-10-12T07:23Z

session_012yK1ddAdAqfzwigRHZq3r2 · os-warren. Reviewed against GitHub on head ef3d0764b (merge base 246edf5b3), dev report 6119644687 on this card.

• Shape: draft, base main, first line Fixes #22949, then Refs #22966 and Refs #22968 with no closing keyword. skip-changeset. Assignee os-warren. check-governed-merges --pr 22973: NOT governed, 586 changed lines (+502 / −84), under 3000.
• Diff, against ruling 6119186511:
  • ci.yml changes only the two timing steps (:1103 drift, :1516 ceilings). Each passes --event "$GITHUB_EVENT_NAME". Neither gains an if: or continue-on-error, so the ⛔ at :1076 ("NO if: and NO continue-on-error") still holds as written.
  • The rule lives once, TIMING_GRADED_EVENT = 'schedule' / timingGateMode() in partition-test-shards.mjs, and check-test-suite-ceilings.mjs imports it. On schedule an over-ceiling package or a shard past 1.5x exits 1, as before. Every other event prints the same table, raises each red as ::warning:: and exits 0.
  • CEILING_FLOOR_SECONDS = 20: an effective ceiling under it prints "below floor" and is never red, the ruling's spelling. The header carries the floor, the seat's reading (3.40 … 16.23, then 22.71 on 1f312fdfa, seven packages under it) and the ruling URL.
  • Unchanged, as ruled: the formula, ×1.25, RULED_CEILING_RAISES (still empty), the dataset and the refresh lane.
• Re-run by this seat in a detached worktree at ef3d0764b:
  • check-test-suite-ceilings.mjs --self-test → "self-test OK (58 cases across 9 batteries)", exit 0.
  • partition-test-shards.mjs --self-test → OK, exit 0.
  • No --event → exit 1, refused with "A missing event is refused rather than guessed".
  • --event schedule with an unreadable captures dir → exit 0, ungraded and named.
• The dev's open question, decided by this seat: a refusal (a malformed table or raise record, a missing --event) stays red on every event, as implemented, which is the dev's A. The ruling moves timing readings to report-only, and a refusal is not a timing reading. Reporting it only off schedule would let a malformed table pass the queue silently.
• Deviations accepted:
  • --event is a required argument and not an opt-in --report-only, so the decision table is self-tested in the script.
  • workflow_dispatch is report-only, because the ruling names schedule alone and the refresh window reads schedule runs.
  • The size is above the dispatch's guess. About half of the additions are the three new self-test batteries.
• H5, recorded per the ruling's confidence gap: on all three hourly runs since the table went live (38202645321, 38206660804, 38210849670), core, metadata and plugin-hono-server replayed from cache, so none has been graded yet. As ruled, each is graded on any hourly run that executes it and reported elsewhere. No force-execute mechanism was built.
• Non-blocking: a missing --event surfaces as an uncaught throw with a stack trace rather than a formatted refusal. CI always passes the event, so this is a local-use cosmetic only.
• CI: success and expected skips so far, with Test Core (1/6), Lint & Repo Gates, Console Pin Gate and Temporal Conformance still running at this comment. It lands through the queue once every check on the head is green.
• After landing: the parked PRs re-enter (fix(security)!: under a wall a catalog-carrying package is the platform operator's — install-local and the /packages delete refuse tenants (ADR-0131 D3) #22935, feat(spec,metadata-protocol)!: rename the type registry's allowOrgOverride to allowEnvironmentOverlay, on the registry and on GET /meta/types #22874, feat(spec)!: refuse a duplicate name inside one stack’s functions array (#22902) #22931, fix(lint): one-line fix lines for 18 more author-time rule ids; os explain RULE_ID carries what they explained #22936, feat(spec,service-automation,lint)!: a notify node's computed values are value slots and its link a text slot — a CEL envelope or a literal, the {…} token refused (#19939 pass 4, S4) #22926, and the arming of feat(core): pure row-to-definition mappers for the ADR-0131 upgrade ceremony (C7b build-ahead, PR 1 of 2) #22959). [maintainer] validate: the field-no-consumers warning is one 856-character line, printed by validate, build and dev alike — one-line verdict + rule: id + a pointer to the full reasoning (os explain, which today takes only schema names) #22161 and finding: a stack's functions array accepts two entries authoring one name, and objectstack build exits 0 shipping the second under a suffixed name (on main, not callable after load), where composeStacks already refuses a duplicate function name #22902 leave pm:blocked through the unlock scan. hourly full run: red on main (CI) #22968 closes on the next uneventful green hourly run.

---

Generated by Claude Code

[objectstack-fleet]

Landed: PR #22973 merged as eede467f3 · 2026-10-12T08:34Z

domain:devx seat 1 · os-warren · session_012yK1ddAdAqfzwigRHZq3r2. ACCEPT 6119677604 on this card is the review record.

• Route: not governed. The PR went ready with auto-merge at 07:56Z, entered the merge queue the same minute, and its group's builds read success. It merged at 08:33:50Z. Its Fixes #22949 closed this card, and the seat removes pm:dispatched in this act.
• On main (eede467f3), as ruled in 6119186511:
  • Both timing steps pass --event "$GITHUB_EVENT_NAME".
  • "Check the per-package suite-duration ceilings" and "Check this shard's timing drift" exit-grade on schedule only. Every other event prints the same table, raises reds as ::warning:: and exits 0.
  • CEILING_FLOOR_SECONDS = 20 (scripts/check-test-suite-ceilings.mjs:176).
• For the seats holding PRs on this card: a merge-group or PR build that starts from eede467f3 or later no longer fails on either timing step. The parked PRs can re-enter the queue: fix(security)!: under a wall a catalog-carrying package is the platform operator's — install-local and the /packages delete refuse tenants (ADR-0131 D3) #22935, feat(spec,metadata-protocol)!: rename the type registry's allowOrgOverride to allowEnvironmentOverlay, on the registry and on GET /meta/types #22874, feat(spec)!: refuse a duplicate name inside one stack’s functions array (#22902) #22931, fix(lint): one-line fix lines for 18 more author-time rule ids; os explain RULE_ID carries what they explained #22936, feat(spec,service-automation,lint)!: a notify node's computed values are value slots and its link a text slot — a CEL envelope or a literal, the {…} token refused (#19939 pass 4, S4) #22926, and the arming of feat(core): pure row-to-definition mappers for the ADR-0131 upgrade ceremony (C7b build-ahead, PR 1 of 2) #22959. [maintainer] validate: the field-no-consumers warning is one 856-character line, printed by validate, build and dev alike — one-line verdict + rule: id + a pointer to the full reasoning (os explain, which today takes only schema names) #22161 and finding: a stack's functions array accepts two entries authoring one name, and objectstack build exits 0 shipping the second under a suffixed name (on main, not callable after load), where composeStacks already refuses a duplicate function name #22902 (Blocked-by: #22949) are the unlock scan's.
• Still open, as ruled:
  • hourly full run: red on main (CI) #22968 (the hourly card) closes on the first uneventful green hourly run.
  • A second consecutive hourly red on one package is a reading, and is posted here.
  • ci: PR-time Test Core (6/6) reds on shard-timing-drift at 1.60–1.62x on three PRs that change low-graph packages (core, metadata-*) — runtime, service-analytics, rest and client run 1.4–2.4x their refreshed weights #22966 was already closed not_planned.

---

Generated by Claude Code

[objectstack-fleet]

Reading under the ruling's rule ("two consecutive hourly reds on one package are a reading") · domain:devx seat 1 · os-warren · session_012yK1ddAdAqfzwigRHZq3r2 · 2026-10-12T10:47Z. ⛔ Not a claim, ⛔ no label change, ⛔ no letter chosen.

• The two consecutive hourly reds on @objectstack/cloud-connection, both on trees carrying eede467f3, where a schedule red is by design:
  • 38228297391 (74c4b3435e, 09:11Z): 165.64 s against 163.89 s (1.01x), with service-cluster-redis 28.61 s / 28.00 s;
  • 38234348113 (111ffff556, 10:07Z): 209.04 s (1.28x). In the same run, service-analytics 491.37 s (1.12x), plugin-approvals 305.06 s (1.05x) and client 182.21 s (1.01x) are also over, and Test Core (6/6) "Check this shard's timing drift" is red.
  • Read through scripts/pm/ci-failure.mjs and MCP get_job_logs (job 114769678847).
• What is new: the 10:07Z hourly run executed the whole workspace.
  • Its table grades cli (2389.93 s, all 3 slices), spec (1335.34 s), core, lint, metadata, metadata-protocol and rest. Those are the packages the 05:08Z hourly run (38206660804) excluded as cache replays.
  • Between the two hourly runs, ten PRs landed on main (74c4b3435e..111ffff556), several of them spec or core changes: feat(spec)!: refuse a duplicate name inside one stack’s functions array (#22902) #22931, feat(spec,service-automation,lint)!: a notify node's computed values are value slots and its link a text slot — a CEL envelope or a literal, the {…} token refused (#19939 pass 4, S4) #22926, feat(core): pure row-to-definition mappers for the ADR-0131 upgrade ceremony (C7b build-ahead, PR 1 of 2) #22959, fix(metadata-protocol): two locales of one email template survive a restart — sys_metadata overlay rows keyed by bundle member (#22877, stage 1) #22954 and others. The hourly run then had no cache to replay.
  • So the hourly profile is not one density. On a busy hour, it runs at the queue-build density the re-analysis 6118615308 measured, and it reds the same packages.
• What this means for the ruling's premise:
  • C2 grades only on schedule because the ceilings were measured there. The ceilings came from a refresh window in which most packages were replayed most hours, so they fit the quiet-hour density, not the busy-hour one.
  • The ruled channel for a reading is a ruled raise or a faster suite. A raise for cloud-connection alone leaves the other three packages crossing on the next busy hour.
• The maintainer's call is presented in chat by this seat, with options. hourly full run: red on main (CI) #23007 (the generated hourly card) stays open until then.

---

Generated by Claude Code

[objectstack-fleet]

Sighting after the reading 6122739059 · domain:devx seat 1 · 2026-10-12T12:45Z. Hourly schedule runs on trees carrying eede467f3:

Run	Commit	Ceiling step	Over
38228297391 (09:11Z)	74c4b3435e	red	cloud-connection 1.01x, service-cluster-redis 1.02x
38234348113 (10:07Z)	111ffff556	red	cloud-connection 1.28x, service-analytics 1.12x, plugin-approvals 1.05x, client 1.01x; shard 6/6 drift red
38240128220 (11:02Z)	9432bac077	green	none
38246590445 (12:03Z)	0ede0f8ab3	red	client 226.02 s (1.26x), cloud-connection 187.34 s (1.14x), service-analytics 479.68 s (1.09x), runtime 938.28 s (1.03x); shard 6/6 drift red

Three of the four hourly runs since C2 landed are red. They cross the same packages the queue builds crossed, and on busy hours they carry shard 6/6 drift too. The remedy stays the maintainer's call, which this seat has put to the maintainer in chat. #23007 stays open.

---

Generated by Claude Code

[objectstack-fleet]

Ruling: batch #330 item 1 · letter E · maintainer 「22949 同意」 2026-10-12T13:57Z

Director seat, summon #37, session_01LqNgdPZB2Jnr5HpGbaqLeB (GitHub os-zhuang; written as objectstack-fleet[bot] via the relay). The maintainer asked for an analysis of this card's ceiling check on the direct channel (「帮我分析 #22949 上限检查」) with the readings 6122739059 and 6124529838 (three of the four hourly runs since C2 red, on client, cloud-connection, service-analytics, runtime); this seat presented the diagnosis and the letters A (ruled raises per package), B (grade only on a daily forced full-density run that also feeds the window) and E (retire the wall-clock ceiling as a red; regression detection moves to the refresh dataset's diff and the per-file budget), with its own four-axis reading and the recommendation E, and the maintainer answered 「22949 同意」.

The diagnosis the ruling rests on

A per-package suite wall on a shared runner measures the run's load, not the package: the same package reads 82 s on a quiet hour (most packages replayed) and 209 s on a busy hour (the whole workspace executing after spec / core landed), 2.56×, while the ceiling formula allows 2.0× of weight (max × 1.25). The refresh window (12 hourly runs) was mostly quiet-density, the grading on schedule meets both densities, so busy hours red the same near-ceiling packages whatever their code does. No factor short of 3× covers it, and 3× hides a real regression. C2 (6119186511) was right to take the red off the queue; it left the red on schedule, where the same density problem lives.

The ruling (E)

1. Both timing steps report-only on every event, schedule included: "Check the per-package suite-duration ceilings" and "Check this shard's timing drift" print their tables and raise ::warning:: annotations everywhere and exit 0; TIMING_GRADED_EVENT goes (the spelling is the dev's), the self-tests move with it, the formula, CEILING_FLOOR_SECONDS and RULED_CEILING_RAISES are untouched (the table stays a reading). Not taken: A (a raise per package is a reading of load, and the next busy hour crosses the next package); B (a daily forced run would grade at one density, at about 60–90 runner-minutes a day, and the window would still need rebasing).
2. Regression detection moves to the refresh lane: each dataset refresh (scripts/measure-test-shard-timings.mjs, the bot PR) compares the new window's per-package weight with the previous dataset's; a package up by 25 % or more (the ceiling's own factor) is named in the refresh PR's body and filed as one card for the owning lane by the refresh lane's seat, under objectstack#22998's test-budget rule, never a red build. The first diff after this ruling baselines on a window of comparable density (the dev states which runs it took and why); a package whose only change is density is not a finding.
3. The per-file budget (a test file over 10 s of wall is a card for the owning lane, skills(ci): two-speed CI — fast set on PR and merge-queue runs, full battery hourly on main and before a release, main-red stops the lane, four required contexts plus the guard, timing gates report-only, test budgets not deletions #22998) reads the per-file durations the dataset already carries (Slowest file: is in the gate's output today); the refresh lane lists the files over budget in the same PR body.
4. hourly full run: red on main (CI) #23007 (the generated hourly-red card) closes on this ruling: its reds are the gate's, not a package's; the hourly run goes green at the next schedule after item 1 lands.

After this ruling

• Two execution cards filed in this act for the domain:devx lane (the seat dispatches CI cards on the maintainer's standing order): E1 the two steps report-only everywhere; E2 the refresh diff and the per-file list with card filing. Both queue behind the devs in flight (the maintainer's word this hour on the number of parallel devs). ⛔ No code by this seat.

os-decision-facets

• ① 项目长远合理性：共享 runner 上的墙钟是负载的函数，数据集同口径差分才是回归；E 取消 schedule 这个特例而不是加裁定条目。
• ② 实际业务拉动：忙碌小时 3/4 红，每次都让 devx 线按 main-red 规则去修一个不存在的回归；真正的测试增长反而没有门在量。
• ③ 防 AI 犯错：门自己的头注释写着"红灯常态化本身有毒"；差分立卡是封闭、可核、点名到包的读数。
• ④ 创业阶段不扩散：不加 runner、不加裁定表条目、不改公式；复用刷新车道与每日强制跑的报告器（cloud#2806 同形）。
  Prior rulings read: ci: merge-queue Test Core reds on the new suite-duration ceiling — @objectstack/cloud-connection measured 177.5 s and 200.3 s vs its 163.89 s ceiling, in two groups of PRs that do not touch it (live since 1bd612bd6) #22949 C2 (6119186511), ci: a per-package suite-duration ratchet — a PR that makes a suite exceed its measured ceiling is red; ceilings rise only by ruling (maintainer-directed, growth constraint) #16468 A (the ratchet), ci: merge-queue Test Core (3/6) reds on shard-timing-drift — @objectstack/cli measured 931.4s vs 579.6s predicted (1.61x, red past 1.5x); scripts/test-shard-timings.json was last refreshed 2026-10-08 #22917 (shard-timing-drift), objectstack#22998 (timing gates report-only; test budgets); thread: 6122739059, 6124529838.