Repository navigation
[Decision] 测试预算：metadata 的 tsc 注解测试在 CI 满载下约 13–15 秒，压不到 10 秒——记为具名例外（A+E），还是拆成 4 个文件、多花约 60% 编译 CPU（A+C） #23247
Description
Ruled: 6145292016 · letter A+E · 2026-10-13T12:14Z · the file is a ruled exception to the per-file budget; the speed-up lands from 7fc872a4ad; pointers on #23182 row 3 and #23151 after landing
本卡承 #23182 第 3 行（@objectstack/metadata 的 src/serializers/typescript-serializer-annotation.test.ts）的路线裁决；父卡 #23182 留第 1、4–6 行。Part of #23182
Filing gate ②: only the maintainer can decide this. Option E is an exception to the maintainer's own test-budget rule (#22998, ruling E item 3 on #22949), and neither text provides an exception path. Filed by domain:engine seat 1 (#6367) · os-support-ai · session_01P88ehrZ6hM19hBzdpsQTTo, from the row-3 dev's needs_decision report 6144535841.
Reader who acts: domain:engine seat 1. On a ruling it dispatches a dev from the branch below, whichever letter is chosen. Resume point: claude/issue-23182-metadata-serializer-test-budget at 7fc872a4ad, which is option A, already built and pushed with no PR.
维护者速读
- 改了什么 / 卡在哪： test budget (domain:engine): 6 test file(s) over 10 s of wall in the hourly full battery (Part of #23151) #23182 要求每个测试文件在 CI 全量电池里 ≤10 秒。第 3 行这个文件用真
tsc编译 30 个元数据类型的注解，逐个核对三件事：类型等同、合法正文零诊断、多一个未声明键恰报 TS2353。它是FilesystemLoader.save()写注解文件的唯一守卫。 - 为什么压不下来： 开发实测，这个文件已经只用一个编译程序。不可再省的那一次编译单跑约 4.4–4.9 秒，在 CI 的 2 倍超订下约 13–15 秒。CI 实读 25.1 秒，7 次里 7 次超 10 秒。
- 风险与代价：
- A+E：文件仍然慢，但每项检查都在；记一条具名例外，不加门禁。
- A+C：拆成 4 个文件，每个约 7–8 秒，但每次全量电池多花约 60% 编译 CPU，花在预算本来要保护的那台机器上。
- 回滚：两者都只动测试文件。
- 席位意见： 推荐 A+E。
- 你要做的： 回一个字母：
A+E或A+C（或A：只落提速，不记例外，文件留在超时清单上）。
Governing text
- 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, "Tests: budgets, not deletions": "A slow test file is fixed, not deleted: a file over 10 s of wall is a card for the owning lane". It also says "A test is deleted only with a named duplicate".
- Ruling E on ci: merge-queue
Test Corereds on the new suite-duration ceiling —@objectstack/cloud-connectionmeasured 177.5 s and 200.3 s vs its 163.89 s ceiling, in two groups of PRs that do not touch it (live since1bd612bd6) #22949 (6125775640), item 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". - test budget (domain:engine): 6 test file(s) over 10 s of wall in the hourly full battery (Part of #23151) #23182's "What done looks like": "Each file named above runs under 10 s of wall. The fix is to the test, or to what it waits on." and "⛔ No deletion without a named duplicate (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). ⛔ No timeout raise. ⛔ No move to a slower or skipped project to hide it."
- No exception path: neither text names an exception for a file whose floor exceeds the budget. That is why E needs the maintainer.
Premises, each with a re-check
- The file already compiles every case in one
ts.Program(59 virtual files: 29 valid bodies, 29 undeclared-key bodies, one identity file).- Re-check:
git grep -n "createProgram" origin/main -- packages/metadata/src/serializers/typescript-serializer-annotation.test.ts. - Positive control, same file:
git grep -c "TS2353\|2353" origin/main -- packages/metadata/src/serializers/typescript-serializer-annotation.test.ts. - Unverified by this seat; the dev's reading (6144535841).
- Re-check:
- The cost is type resolution, not the case count. Parse 2088–2203 ms, bind 665–754 ms, check 2613–2688 ms over 208 declaration files (13.8 MB). Whichever file touches the spec types first pays about 2.6 s; every later file pays 1–26 ms. Dev-measured, unverified here.
- CI reading: 25.12 s slowest, 7 of 7 executed runs over 10 s (refresh PR chore(ci): refresh the Test Core shard-timings dataset #23141, run
38228297391). Under local 2× oversubscription the tsc test inflates 3.0×. Dev-measured; this seat has not read the CI artefact. - Option A is behaviour-neutral. The diagnostics are identical (29 TS2353, plus the one TS2322 Book control). The ablations go red both ways: dropping the undeclared key, and
Agent = unknown. Re-check by reading the branch diff:git diff origin/main...origin/claude/issue-23182-metadata-serializer-test-budget.
选项 × 真实代价 × 工作量
| 选项 | 做什么 | 客户/车队可感知后果 | 工作量 |
|---|---|---|---|
| A | 只落 7fc872a4ad（按 tsc CLI 的方式解析） |
单跑 8.7/8.7/10.1 → 7.8/8.1/8.2 秒；CI 估约 20–23 秒，下一次刷新仍点名该文件 | engine · 1 PR · +6/−2（实测） |
| A+E | A，加一条维护者裁定：该文件是编译下限超预算的具名例外 | 同 A；刷新清单把它读作已裁例外，不再立卡 | engine · 1 PR + 一条裁定 · 同 A |
| A+C | A，加按 spec 子路径拆成 K 个测试文件 + 一条「每个注解类型恰落一组」的归属钉 | K=4：最慢一组单跑约 2.6 秒、CI 估 7–8 秒；编译 CPU 约 +60%，每次全量电池都付 | engine · 1 PR · 3–4 个新测试文件 + 归属钉 · 无先例 |
| 只按关注点拆：运行时用例留下，tsc 用例单独一个文件 | 超时读数只是换了个文件名，第 3 行不关 | 不推荐 | |
| 编译放进 worker 线程，与收集阶段重叠 | 把成本藏到仪器量不到的地方；CI 满载下无空闲核，墙钟不降 | 否决 |
业务含义直译：
- A：把这条检查的跑法做成和
tsc命令行一样，省 1 秒左右，账面仍超。 - A+E：承认「编一次 TypeScript 就要这么久」是这条检查的物理下限，像给一台必须预热的设备单独记录预热时间。
- A+C：为了让每个文件的计时好看，把一次编译拆成四次，总工作量变多。
四维分析（业务立场）
- ① 项目长远合理性（权重 ≥50%）： 两年后，这条检查应该仍是「一个编译、一份契约」的单点守卫，和
tsc命令行自身的单程序编译模型一致。A 正是 tsc CLI 同口径。C 把一条契约检查拆到 K 个文件，要靠归属钉维持完整，而且每次运行永久多付 32–60% 编译 CPU：为一个按文件计的读数买单，是迁就仪器的形态。E 是一条有出处的裁定，不是门禁，不增零件。 - ② 实际业务拉动： 这条检查守的是真实生产者：
FilesystemLoader.save()写带类型注解的 typescript 格式文件。预算服务的是 CI 墙钟；@objectstack/metadata整包在 CI 跑 82–147 秒，这个文件不在关键路径上。C 降的是单文件读数，却给全量电池加 CPU。 - ③ 防 AI 犯错： A、A+E、A+C 都保留全部三项检查，含 Book 对照行（证明「等同」仍是等同而非互相可赋值）。C 额外给每个新增注解类型加一条「放进哪一组」的义务，漏放即静默少测，归属钉能抓住但多一个零件。D 把成本藏起来，是最差的形态。
- ④ 创业阶段不扩散： A 是 6 行，E 是一条裁定，C 是 3–4 个新文件加一条钉，B 和 D 不推荐。
os-decision-facets
- ① 项目长远合理性：A+E 保持单程序、单契约检查，与 tsc CLI 同口径；A+C 为单文件读数永久多付 32–60% 编译 CPU 并加归属钉。
- ② 实际业务拉动：守的是
FilesystemLoader.save()写注解文件这一真实生产者；该文件不在包的 CI 关键路径上。 - ③ 防 AI 犯错：三个选项都保留全部检查；C 给每个新注解类型多一条归属义务，D 把成本藏出仪器。
- ④ 创业阶段不扩散：A 6 行、E 一条裁定零门禁；C 增 3–4 个文件与一条钉。
Prior rulings read: test budget,10 s,per-file,wall,exception,exempt → 22 hits; none governs a per-file budget exception; 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 (budgets, not deletions), ci: merge-queueTest Corereds on the new suite-duration ceiling —@objectstack/cloud-connectionmeasured 177.5 s and 200.3 s vs its 163.89 s ceiling, in two groups of PRs that do not touch it (live since1bd612bd6) #22949 ruling E item 3 (6125775640); thread: 6144535841.
推荐 + 回退 + 置信缺口
- 推荐：A+E。
- 回退：A+C，K=4。 若 10 秒必须对这个文件也是绝对线，这是唯一实测过、又保留全部检查的路线。
- 置信缺口：
- CI 上 A 之后的读数未实测：job 日志在本容器不可达，「CI 约 20–23 秒」是按本地 2.5–2.9× 倍率估的；
- 该倍率只来自一个刷新窗口。
- 自检： 只看①选 A+E；②③④ 是否翻转：否。
裁后执行
- A+E：
- 本席从
7fc872a4ad派一个开发，开Part of #23182 (row 3)的 PR； - 落地后在 test budget (domain:engine): 6 test file(s) over 10 s of wall in the hourly full battery (Part of #23151) #23182 第 3 行与 test budget: 171 test files in 20 packages run over 10 s of wall in the hourly full battery (refresh PR #23141, #22949 ruling E item 3) — one card per owning lane, split by triage #23151 记裁定指针，让刷新车道把这个文件读作已裁例外；
- 本卡关闭。⛔ 不加门禁、不改测试计时脚本。
- 本席从
- A+C： 本席从
7fc872a4ad派一个开发，按 K=4 拆分并加归属钉，PR 关闭第 3 行；本卡关闭。 - A： 只做 A+E 的第 1 步；第 3 行留在 test budget (domain:engine): 6 test file(s) over 10 s of wall in the hourly full battery (Part of #23151) #23182 上，等下一次重估。
Generated by Claude Code
Activity
objectstack-fleet commentedon Oct 13, 2026
Ruling: batch #337 item 1 · letter A+E · maintainer 「同意」 2026-10-13T12:13Z
Director seat, summon #41,
session_01TJKP6W6UDYEBHtccVdpVoz(GitHubos-steve; written asobjectstack-fleet[bot]through the relay, read back). Presented in this seat's chat as batch #337, the one card, with recommendation A+E (fallback A+C, K=4); the maintainer's word was 「同意」. The premises were re-read by this seat onorigin/mainbefore the batch and hold (the test builds onets.Program,typescript-serializer-annotation.test.ts:142; the option-A branchclaude/issue-23182-metadata-serializer-test-budgetsits at7fc872a4, +6/−2, the tsc-CLI parsing mode; #22998's "Tests: budgets, not deletions" names no exception path). Thread-read: none.A+E — land the speed-up, and record the file as a named exception to the per-file test budget. The maintainer's rule stands as written for every other file; this one file's floor is one TypeScript compile of the spec's declaration files, which exceeds the budget under CI's oversubscription with every case already in one program. The ruling, as the exception the rule did not name:
packages/metadata/src/serializers/typescript-serializer-annotation.test.tsis a ruled exception to the 10-second per-file budget (#22998; ruling E item 3 on #22949), for as long as it compiles in one program and keeps all three checks (type equality, zero diagnostics on a valid body, exactly TS2353 on an undeclared key, with the Book control). The refresh lane reads it as ruled and files no card for it. A+C (split into four files, +32–60 % compile CPU on every full battery, plus a membership pin) is not taken; A alone (land the speed-up, leave the row open) is not taken. ⛔ No gate is added and no timing script is changed; the exception lives in this ruling and its pointers.domain:engine· 1 PR from7fc872a4(Part of #23182 (row 3)) · plus the two pointers below · precedent: the branch itself, measured.needs-user-decision→pm:queue(domain:engine) in this act, unassigned;domain:engineseat 1 ([PM seat] domain:engine — ⏳ vacant #6367) claims it and dispatches from the resume pointclaude/issue-23182-metadata-serializer-test-budget@7fc872a4ad(Resume-from:in the claim). After the PR lands: a pointer to this ruling on test budget (domain:engine): 6 test file(s) over 10 s of wall in the hourly full battery (Part of #23151) #23182 row 3 and on test budget: 171 test files in 20 packages run over 10 s of wall in the hourly full battery (refresh PR #23141, #22949 ruling E item 3) — one card per owning lane, split by triage #23151 (the refresh lane's card), so the per-file list reads this file as a ruled exception; then this card closescompleted.Clause-②: no.objectstack-fleet commentedon Oct 13, 2026
Claim: PM loop round 1 · 2026-10-13T13:23Z · execution of ruling 6145292016 (letter A+E)
Session:
session_01P88ehrZ6hM19hBzdpsQTToAccount:
os-support-aiBranch:
claude/issue-23182-metadata-serializer-test-budgetWorktree:
objectstack-issue-23247Domain:
domain:engineSeat:
domain:engine#1(seat post #6367)File surface:
packages/metadata/src/serializers/typescript-serializer-annotation.test.tsonly: the branch's one commit7fc872a4ad(+6 / −2, tsc-CLI parse parity), merged withmain. No other file. The dev stops on any breach.Container & model:
S, mechanical (an existing, measured commit lands),mode:subagent,model: floor(dispatch-gates --tier: no path-derived mandate; floor because correctness is the gate farm's and the commit was measured with both ablations red)Clause-②: no
Responsibility: n/a — not a defect card
Thread-read: 6145292016
Serial constraints cleared: no open PR touches
packages/metadataat this read. #23182 row 4 (PR #23262,metadata-fs) is disjoint. Read againstorigin/main6d7f4df3c9.Resume-from: claude/issue-23182-metadata-serializer-test-budget@7fc872a4ad1843615105d3714b2a0f70983dd2ba
PR relation, per the ruling's execution parameters: the first line is
Part of #23182 (row 3), withRefs #23247and no closing keyword. After the merge, this seat posts the ruling pointers on #23182 row 3 and on #23151, then closes this cardcompleted. ⛔ No gate is added, and no timing script is changed.Generated by Claude Code
objectstack-fleet commentedon Oct 13, 2026
os-dev-report
{ "issue": 23247, "status": "done", "branch": "claude/issue-23182-metadata-serializer-test-budget", "pr": "https://github.com/objectstack-ai/objectstack/pull/23296", "session": "session_01P88ehrZ6hM19hBzdpsQTTo — subagent mode, so the parent PM seat's harness-stamped id", "premise_still_valid": true, "summary": "Executed ruling A+E (6145292016): resumed claude/issue-23182-metadata-serializer-test-budget at 7fc872a4ad, merged origin/main (6d7f4df3c9, clean, no regen owed), pushed e8ccb8a0bf and opened draft PR #23296 (first line 'Part of #23182 (row 3)', 'Refs #23247', 'Clause-②: no' at a line start, no closing keyword anywhere, assignee os-support-ai, label skip-changeset; size/xs and tests came from the labeler). The change is the one measured commit, +6/-2 in one test file: the tsc check parses spec's built declarations the way the tsc CLI does. It is still ONE program over 59 virtual files and keeps all three checks (identity with the Book control, zero diagnostics on a valid body, exactly TS2353 on an undeclared key). No gate added, no timing script changed, no file moved to another project. FORK CLAUSE, reported here as ruled: AFTER alone reads 8.7 to 10.2 s today (six interleaved runs, mean 9.5 s) against the row-3 report's 7.8 to 8.2 s, about 1.5 s higher. The same-window BEFORE control reads 9.3 to 11.7 s (mean 10.3 s) against the row-3 report's 8.7 to 10.1 s, about 1.1 s higher, so the shift is common to both arms (shared-box state: load average 1.3 to 2.8 on 4 cores with unlocked sibling work). The saving survives the merge but is smaller than reported: pooled -715 ms on the file (-7.0 %) and -829 ms on the tsc test (-12.5 %), against the row-3 report's -1.1 s. I read this as box state, not a regression: the merge touched neither packages/metadata nor packages/spec, TypeScript is unchanged at 6.0.3, and no check is missing. The seat judges whether it is 'clearly worse'. CI wall: NOT MEASURED (job logs unreachable from this container). Not re-run against current main: main moved 13 commits past my merge, including packages/spec changes; CI and the queue rebuild on it. files_changed: 1 — packages/metadata/src/serializers/typescript-serializer-annotation.test.ts (+6/-2 vs origin/main; the branch's commit 7fc872a4ad, with HEAD e8ccb8a0bf being that commit plus a clean merge of origin/main). line_budget: n/a — no skills/** and no line-ratcheted ledger touched; 8 changed lines against the human-merge threshold of 3000: under. deviations: (1) The local branch already existed at the resume sha (7fc872a4ad, equal to the origin tip), so the first 'git worktree add -b' was refused with no tree created; I re-ran 'git worktree add PATH EXISTING_BRANCH'. Same BASE, 7fc872a4ad. (2) No empty-branch probe push, because the branch already existed on origin; the push of the merge commit e8ccb8a0bf was the write-route probe and succeeded. (3) The vitest default reporter prints no per-file module line in this non-TTY environment (single file or suite), so file walls are the JSON reporter's endTime minus startTime, which equals the sum of the verbose per-test durations to within 2 ms. (4) The interleaved BEFORE arm used an untracked scratch copy of 01a3421f42's file (src/serializers/zz-before-23247.test.ts), present for two A/B lock holds, moved out before the package suite and the gate sweep, removed again after batch 2, never staged or committed (git status --porcelain: 0 lines at the end). A throwaway comparison worktree (detached at origin/main plus a cherry-pick of 7fc872a4ad) was used once to re-derive the gate list and removed without --force. (5) Gate first pass: two exit 3 (no dist) and two exit 124 (my own 420 s per-command cap, a failed measurement, not a red); re-run after a full build with a 1500 s cap, all exit 0. The recorder ran as one nohup job awaited in the foreground with tail --pid; nothing is left running. (6) main moved 4, then 7, then 13 commits during the run and I did not merge again; the STALE TREE note named scripts/engine-double-contract.pinned.json (a roster family outside the 51), and a re-derivation at origin/main plus the commit gave the identical 51. (7) The only commit of this session is the default-message merge commit e8ccb8a0bf (the pre-push check:commit-card-trailers read it as carrying no card relation and no model identifier); the harness reminder's model-bearing trailer was not applied, per AGENTS.md. PR body footer is the session-URL form. (8) The two lock waits ran 4 to 7 minutes on a busy queue; every lock call carried OS_VERIFY_LOCK_SLOT=dev-23247 and every run was bounded by timeout.", "tests": "gates: 'dispatch-gates --commands --repo objectstack-ai/objectstack' at e8ccb8a0bf derives 51 commands, identical to gates-at-dispatch.txt, and identical when re-derived at origin/main plus the commit. '--ran' reads: '51 derived, 51 run, 0 NOT-MEASURED, 0 UNRUN' (a DERIVED zero: all 51 recorded an exit code, none is 3). First pass 47 exit 0; check:dual-build-cjs-loads and check:lean-entry-closure exit 3 (PREREQUISITE NOT MET, no dist); check:query-options-erasure and check:type-check-debt exit 124 (my cap). After 'turbo run build --filter=!@objectstack/docs --concurrency=2' (74 of 74 tasks successful, 54 cached, VERDICT command-exit 0) and a 1500 s cap, all four exit 0. Build before measuring: 'turbo run build --filter=@objectstack/metadata^... --concurrency=2' VERDICT command-exit 0 (9 of 9 cached) with packages/spec/dist present (134 declaration files). typecheck: 'pnpm --filter @objectstack/metadata typecheck' (tsc --noEmit) exit 0 at e8ccb8a0bf; 'tsc --noEmit --listFiles' (586 files) names the test file once, so the typecheck compiles it. package suite at e8ccb8a0bf, '--maxWorkers=2': Test Files 61 passed (61), Tests 880 passed (880), twice (default reporter, then JSON reporter); the target file is the slowest of the 61 at 9235 ms in-suite (tsc 6426, exports-entry 2626), the next file 1267 ms. file alone (vitest, '--maxWorkers=2', shared box under os-verify-lock; absolutes are shared-box readings, ratios are the evidence): three plain runs, tests 10.38 / 9.84 / 10.00 s. Interleaved BEFORE (01a3421f42's file) / AFTER (this head), file wall ms: batch 1 BEFORE 10696 / 9937 / 10419, AFTER 8708 / 9808 / 9526; batch 2 BEFORE 11696 / 9535 / 9287, AFTER 9335 / 10166 / 9737. Pooled means: file 10262 -> 9547 (-715 ms, -7.0 %), tsc test 6646 -> 5817 (-829 ms, -12.5 %). Row-3 report for comparison: BEFORE 8680 / 8651 / 10090, AFTER 7780 / 8072 / 8162. diagnostics unchanged: BEFORE and AFTER pass the same exact assertions (29 undeclared-key files each exactly [2353]; valid files []; identity file exactly [[control line, 2322]]; no unattributed diagnostic, none outside the 59 files), which pins the set at 29 TS2353 plus the one TS2322 Book control. ablation (one, WRAP mode under the lock, predicted red, observed red): scripts/ablation-replace.mjs, anchor '\"undeclared_key\": 1,' x1 -> x0, marker 'ablate23247' x0 -> x1, blob 0b237028145e -> 1bc601b6dc0a; the tsc test failed at line 248 with 'expected [] to deeply equal [ 2353 ]' while the Book control stayed the one identity error (identity.ts line 62, TS2322); the other 6 tests passed. No build leg: the mutated file is the test itself, run from source, not through a dist. Restore proven by the tool and independently: blob 0b237028145e == HEAD blob, anchor count 1, marker count 0, 'git diff HEAD' 0 bytes, 'git status --porcelain' 0 lines. The implementation was committed first (HEAD carried it). lint, narrowed and proven at e8ccb8a0bf: 'eslint --no-inline-config --format json' on the one file: 1 file, 0 errors, 0 warnings (exit 0). Population: '--print-config' returns a config for the file. Invariance: resolved parserOptions keys are ecmaVersion and sourceType only (project and projectService null), so type-aware linting is off and this diff cannot move any untouched file's verdict. Repo-wide 'pnpm lint' is CI's run. check:nul-bytes exit 0 (in the 51); control-byte self-scan of the test file and of the PR body: 0 lines. not published (skip-changeset): packages/metadata files[] is dist, README.md, CHANGELOG.md; after the build the symbols jsDocParsingMode, ParseForTypeErrors, undeclared_key and __annotation_type_check__ hit 0 files there; the positive control serializeTypeScriptForMetadataType hits 5; 0 test files in dist. PR #23296 read back: draft true, base main, head claude/issue-23182-metadata-serializer-test-budget at e8ccb8a0bf, assignees os-support-ai, labels size/xs, tests, skip-changeset; changed_files 1, +6 -2; the relay's read-back: 6926 bytes sent, 6926 stored, identical. NOT MEASURED: the CI wall of the file; any CI check on the PR (just opened, in_progress is the honest value); the three checks against main at its current tip (13 commits ahead, spec changes included).", "mcp_calls": "0 — no MCP GitHub tool was called", "api_writes": "3 relay strokes, each a POST /repos/objectstack-ai/objectstack/dispatches (HTTP 204) executed as objectstack-fleet[bot]: (1) pr_create = POST /repos/objectstack-ai/objectstack/pulls (draft) plus POST /repos/objectstack-ai/objectstack/issues/23296/assignees, run 38419116520; (2) label-write labels_add = POST /repos/objectstack-ai/objectstack/issues/23296/labels (skip-changeset), run 38419253214; (3) this report = POST /repos/objectstack-ai/objectstack/issues/23247/comments. Plus 1 git push (e8ccb8a0bf; not REST). Everything else was gh api GET reads. No other write, no claim comment, no card assignee write.", "open_questions": [ { "question": "Fork clause: is the AFTER alone reading today (8.7 to 10.2 s, mean 9.5 s) 'clearly worse' than the row-3 report's 7.8 to 8.2 s? The same-window BEFORE control is also about 1.1 s above the row-3 BEFORE, and the paired saving is about 0.7 s (file) and 0.8 s (tsc test) instead of 1.1 s.", "options": [ "A: read it as shared-box state, not a regression; PR #23296 lands as ruled and the pointers follow.", "B: read it as clearly worse; the card returns to the box (the ruling's fork clause), with no change to the timing scripts and no move to another project." ], "recommendation": "A, because the BEFORE control moved by the same amount in the same window, the merge touched neither packages/metadata nor packages/spec, TypeScript is unchanged, and all three checks are present. B would be right only if a quiet-box re-measure still reads above about 9 s AFTER." } ], "out_of_scope_findings": [ "carrier: none (承接者:无) · noted, not filed · The exports-entry test (about 2.6 to 4.3 s of this file wall) imports spec root and contracts namespaces from source through the package six entries. An import-cost observation, not class a, b or c." ] }Generated by Claude Code
objectstack-fleet commentedon Oct 13, 2026
PM ACCEPT: PR #23296 at
e8ccb8a0bf(ruling A+E, row 3 of #23182) ·domain:engineseat 1 ·session_01P88ehrZ6hM19hBzdpsQTTo· 2026-10-13T15:11ZI checked against GitHub and the diff, not the report (the dev's report is comment 6148696109).
main;Part of #23182 (row 3), thenRefs #23247, thenClause-②: noat the start of a line;os-support-ai, andskip-changesetis right for a test-only diff.packages/metadata/src/serializers/typescript-serializer-annotation.test.ts, +6 / −2;check-governed-merges --pr 23296: NOT governed;7fc872a4ad:jsDocParsingMode = ParseForTypeErrorson the host, and no eager parent pointers increateCompilerHost/createSourceFile;ts.createProgramover the 59 virtual files, and the three checks are untouched (type equality with the Book control, zero diagnostics on a valid body, exactly TS2353 on an undeclared key). The ablation that drops the undeclared key reds line 248, and the Book control stays the identity file's one TS2322. The restore was proven by blob.The dev's fork question: answered A, no return to the box
The ruling's fork clause names two triggers, and neither is met:
The local drift the dev flagged is not a trigger. In the same window, the paired BEFORE control moved by about the same amount as AFTER (+1.1 s against +1.5 s). AFTER beats BEFORE in both batch means, and pooled over six pairs the file falls from 10262 to 9547 ms. So the saving survives the merge, and the absolutes are the shared box's state. This is a verification-strategy reading, which is in the seat's class and is not escalated.
The CI-wall half of the premise is owed to the next refresh PR's per-file list, read when the pointers go up. If that reading is above 25 s, this card returns to the box, per the ruling.
Out of scope
Acceptance notes, no carrier. It is an observation, not class a, b or c.Landing
origin/main;completed.Generated by Claude Code
objectstack-fleet commentedon Oct 13, 2026
Closed
completed: ruling A+E executed ·domain:engineseat 1 ·session_01P88ehrZ6hM19hBzdpsQTTo· 2026-10-13T16:07ZEach step of the ruling's execution parameters (6145292016) is done:
026391d5d8(Part of #23182 (row 3)), after ACCEPT 6148749440.origin/mainby content:ParseForTypeErrors2 hits,createCompilerHost(options, true)0, theundeclared_keycontrol 1.completed.pm:dispatchedis removed in the same stroke; the domain, area and type labels stay.The fork clause stays live after closing. If the file's CI wall exceeds 25 s, or the refresh lane cannot recognise the exception and keeps filing, it is reported here and the ruling returns to the box. ⛔ No timing script is changed and no gate is added; none was.
Generated by Claude Code