Repository navigation
fix(convention): 持久性降级记 error 而非 warn —— 把 #4460 定成规则,并给规则装上机械检查 (#4632) - #4730
Conversation
…4632) #4420: the durable suspended-run store attached to a table that was never created. Every write failed into a warn nobody read, every restart dropped all in-flight approvals, and the process reported perfect health throughout. #4460 raised that ONE site to error. This makes it the rule, because the class is what recurs. The rule (AGENTS.md, new "Degradation log levels" section) is a judgment QUESTION rather than an adjective, so it can be applied while writing the catch: after the degradation, does the system still look "normal" from the outside while something it claims is persisted has not actually landed? Yes -> error. An error here owes the CONSEQUENCE and the FIX in its first line, once, per the #4460 reference text in service-automation's start(). Anchored in AGENTS.md rather than an ADR addendum: this is an authoring judgment every agent applies at every catch, not an architectural decision scoped to one subsystem, and it is broader than the plugin-ordering context ADR-0116 covers. It cross-references its two existing cousins -- Prime Directive #10 (declared != enforced) and "Absence must be loud". Sites raised to error, each judged individually: - objectql schema sync, per object and summary (the info line said "Schema sync complete" over a pass that lost DDL) - objectql reload-time schema sync (a Studio edit shows a field the column for which was never created) - ObjectQL.syncSchemas() -- an EMPTY catch; marketplace install and template seeding write into tables this failure means do not exist - service-automation suspended wait-timer re-arm, 4 paths Left at warn deliberately: the batch->sequential schema-sync fallback (it recovers) and "no job service registered" on the re-arm path (a declared absence, not a broken promise). Over-applying this rule trains everyone to skim error, which is what made the original #4420 warn unreadable. Teeth: pnpm check:durability-log-level walks the AST for catch blocks guarding a declared vocabulary of durability-critical operations and fails when one degrades below error without rethrowing. Follows same-file helpers so extracting a reporter cannot quietly defeat it; ships a 10-case --self-test; exceptions are hand-written and shrink-only with no --fix flag. Deliberately narrow -- a ratchet, not a proof; both limitations are stated in the script header. Findings in the two packages frozen this round are filed, not fixed: #4728 (metadata database-loader silent DDL swallow, baselined) and #4729 (metadata-protocol seed-loader: its own comment demands error). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015Br2xsJsczFsTR9bvbh2Ny
|
The latest updates on your projects. Learn more about Vercel for GitHub. 1 Skipped Deployment
|
📓 Docs Drift CheckThis PR changes 2 package(s): 15 hand-written doc(s) reference the affected code and may need an implementation-accuracy re-verification:
|
…context bag
CI caught what a filter-scoped typecheck could not:
src/engine.ts(5356,15): error TS2353: Object literal may only specify known
properties, and 'object' does not exist in type 'Error'.
`Logger` (packages/spec/src/contracts/logger.ts) is asymmetric on purpose:
warn (message: string, meta?: Record<string, any>): void
error(message: string, error?: Error, meta?: Record<string, any>): void
Every escalation in the previous commit was written as a mechanical
warn -> error swap, which silently slides the context bag from the `meta`
slot into the `Error` slot. Excess-property checking makes that a hard
error -- but ONLY where a declaration is emitted, so it surfaced in the
tsup DTS build and in the repo-wide TypeScript Type Check, while every
runtime test stayed green (a plain object in the Error slot logs fine).
Four call sites corrected to (message, Error, meta):
- engine.ts syncSchemas() -> e as Error, { object, tableName, driver }
- plugin.ts reload-time schema sync -> the caught error
- plugin.ts reportSyncFailure() -> the caught error, { object, tableName, driver }
- plugin.ts pass summary -> undefined (no single error owns a whole pass)
The tests now assert the ARGUMENT POSITIONS rather than a flattened `meta`,
so this exact slide fails a test run too instead of only a build:
expect(args[0]).toBeInstanceOf(Error)
expect(args[1]).toMatchObject({ object: 'invoice' })
Also fixed the overdue-resume test: AutomationEngine.resume() REPORTS
machine-state problems in its result rather than throwing, so an
unregistered flow never reached the catch under test. It now injects a
throwing resume stub, which is what that catch actually guards.
Verified: turbo dependency-closure build, objectql + service-automation
builds WITH DTS, objectql typecheck, 7 + 25 tests, gate self-test 10/10.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015Br2xsJsczFsTR9bvbh2Ny
CI 红的根因已修(
|
| 位置 | 第二槽 |
|---|---|
engine.ts syncSchemas() |
e as Error + { object, tableName, driver } |
plugin.ts reload-time schema sync |
捕获到的 error |
plugin.ts reportSyncFailure() |
捕获到的 error + { object, tableName, driver } |
plugin.ts 汇总行 |
undefined —— 整趟 pass 没有哪一个 error 能代表它 |
教训:--filter … typecheck 覆盖不到 DTS 编译路径
我本地跑的是 pnpm --filter @objectstack/objectql typecheck,它是绿的;红的是 build 里的 DTS 步骤。两者读同一份类型却在不同时机报错,而这一类「对象字面量塞进了错误的参数槽」只有在产出声明时才会被超额属性检查逮到。
结论:改动涉及公开签名 / 日志调用形状时,推送前必须跑带 DTS 的 build,typecheck 不能替代它。 另外这个 worktree 是全新的,依赖包一个 dist 都没有 —— 第一次 tsc 直接报了一屏 Cannot find module '@objectstack/types',那不是我的改动,是 AGENTS.md §9 说的「新 worktree 先刷构建状态」。正确姿势是 pnpm exec turbo run build --filter=<pkg>,靠 ^build 把整个依赖闭包带起来。
测试也补了这个洞
原来的断言把日志尾参拍平成一个 meta 看,所以参数槽错位它照样通过。现在直接断言参数位置:
expect(args[0]).toBeInstanceOf(Error)
expect(args[1]).toMatchObject({ object: 'invoice' })
同一处滑动现在测试就会红,不必等构建。
顺带修了 overdue-resume 那条用例:AutomationEngine.resume() 对「flow 不存在」这类机器状态问题是在结果里报告而不是抛出的,所以我原来用「没注册 flow 的引擎」根本进不到被测的 catch;改成注入一个会抛的 resume stub —— 那才是那个 catch 真正守护的东西。
复验(全部本地跑过)
closure build (turbo ^build) exit 0
objectql build WITH DTS exit 0 ← CI 挂的就是这步
service-automation build WITH DTS exit 0
objectql typecheck exit 0
objectql tests 7 passed
service-automation tests 25 passed (新增 + wait-node + suspended-run-wiring)
check:durability-log-level self-test 10/10;8 缝隙全部响亮或重抛
(service-automation 没有 typecheck script —— 它在 ratchet 里带 DEBT 条目,其类型覆盖就是上面那条带 DTS 的 build。)
Generated by Claude Code
…rsession (objectstack-ai#18859) Fixes objectstack-ai#18828 Clause-②: no The claim protocol forbids a second `Claim:` — the rule is this file's own, in the objectstack-ai#17366 docblock at `scripts/pm/check-clause2-carriers.mjs:378` — and until this PR nothing READ it. A thread that carried the forbidden second line was RANKED, not refused: the governing-claim selector took the newest live claim that parses a branch and printed the loser as `rejected: 1 … a SUPERSEDED claim`, clean and green at exit 0, in the register reserved for a transition the protocol DESIGNED. A writer-side prohibition with no enforcing reader. `claimRepeats` and the C8 row are that reader.⚠️ **This is a RESUMED delivery.** Three commits were already on this branch from a run whose container was killed in its final-gates phase. Nothing it wrote was rewritten and nothing was redone — every item of the dispatch contract was re-verified against the tree, and all three readings were re-taken at the current tip and are dated below. The verified/fixed ledger is the first section. ## What was VERIFIED and what was FIXED Every contract item was found already correct and is left **byte-unchanged**. No code fix was needed; the only commit this run adds is a merge of `origin/main` (the derivation was answering about a stale tree — see Gates). | contract item | finding | |:---|:---| | the reading is a VERDICT at `EXIT_PAIR_ADVERSE` (4), never a NOTE at 0 | **verified** — `C8` is pushed in `pairRows` (:4430); `pairNotes` does not carry it, pinned | | `claimCarrierSelection` stays a pure function of the rows it is handed | **verified** — byte-identical to `origin/main` (sha256 of the whole function `40f59ba86082c358` at both revs) | | `CLAIM_COMMENT_MARKER` unwidened | **verified** — it is *imported* from `check-half-states.mjs` and read through `markerMatches`; that file is not in this diff at all | | `CLAIM_SELECTION_RULE` and the governing-claim choice unchanged | **verified** — byte-identical across revs; every hunk but two is a pure insertion | | SUPERSEDED / RETRACTED wordings byte-unchanged where they still apply | **verified** — no hunk touches them; the fixture control below prints the SUPERSEDED sentence identically at both revs | | every pin (a)–(i) present, each with a non-vacuity control | **verified** — 52 `t(...)` cases in the battery block; the ablation shows all nine directions carried | | battery registered in the roster with its case count | **verified** — :792, pinned at 52, and the block declares exactly 52 | | `SELF_TEST_BATTERY_FLOOR` raised by exactly one | **verified** — 31 → 32 (:806) | | the live census extended for this shape, population named | **verified** — the five measured instances replayed from their real rows, with `objectstack-ai#18559` as the sanctioned-shape control | `rowAuthor`, `laterOnThread` and `claimRetractions` are byte-identical across the two revs as well. ## The before-reading — the fixture control through the exported reader The same two rows (one author, two claims each parsing a branch, no retraction between) through the same exported `claimCarrierSelection` / `pairInputRecord` / `pairRows`, at `origin/main` `88aa326deb` and at this branch: | | `origin/main` `88aa326deb` | this branch | |:---|:---|:---| | `claims` / `live` / `pool` / `rejected` | 2 / 2 / 1 / 1 | 2 / 2 / 1 / 1 — **unmoved** | | governing claim | `7100000002` | `7100000002` — **unmoved** | | `rejected[0].reason` | `a SUPERSEDED claim — it is not the newest LIVE claim that parses a branch …` | byte-identical | | `claimRepeats` exported | **no — the reader does not exist at that rev** | yes | | `claim.repeat` in the record | absent | `1 author(s) holding more than one LIVE claim comment …` | | `pairRows` codes | *(none)* | `C8` | | **verdict** | **exit 0 — nothing refused** | **exit 4 (`EXIT_PAIR_ADVERSE`)** | That is the defect and the repair in one table: the selector does not move, and the thread stops reading green. ## The live count — the five cards and the control, re-taken 2026-09-18T00:44:39Z Read per card through the gate's own `markerMatches` / `CLAIM_COMMENT_MARKER` and `claimRetractions`, not by eye.⚠️ Those threads may have left this state since. | card | `Claim:` comments (author) | `Clause-②-correction:` | `Release:` | the OLD reading | C8 today | card state | open delivering PR | |:---|:---|:---|:---|:---|:---|:---|:---| | objectstack-ai#18540 | 2 — `os-support-ai` (5719079496, 5720020876) | 0 | 0 | SUPERSEDED, exit 0 | **named, exit 4** | closed | **none** | | objectstack-ai#18677 | 2 — `os-support-ai` (5720104138, 5720190458) | 0 | 0 | SUPERSEDED, exit 0 | **named, exit 4** | closed | **none** | | objectstack-ai#18748 | 2 — `os-support-ai` (5720212595, 5720888122) | 0 | 0 | SUPERSEDED, exit 0 | **named, exit 4** | closed | **none** | | objectstack-ai#18651 | 2 — `os-support-ai` (5721424769, 5721997887) | 0 | 0 | SUPERSEDED, exit 0 | **named, exit 4** | closed | **none** | | objectstack-ai#18778 | 2 — `os-support-ai` (5721425530, 5722028692) | 0 | 0 | SUPERSEDED, exit 0 | **named, exit 4** | closed | **none** | | objectstack-ai#18559 *(control)* | 1 — `os-support-ai` (5721425131) | **1** (5721779100) | 0 | *(nothing rejected)* | silent | closed | **none** | The card's table reproduces exactly. All six cards are now **closed**, and the open-PR column is empty for every one of them: the only open PR cross-referenced from any of these threads is objectstack-ai#18857, whose body's closing keyword names `objectstack-ai#18780` instead — `prDeliversCard` answers `false` for all six and `true` for `objectstack-ai#18780` (the control leg), so it is not paired with any of them. ⛔ The repair of the five is `os-support-ai`'s; this PR only names them, and posts nothing on those cards. ## The escalation probe — the triage's p1 condition, MEASURED The triage marked this p2 because all five instances were one seat self-superseding, and named the escalation condition it had not run: a second `Claim:` from a DIFFERENT session on one card, which would be a silent ownership transfer printed green. I ran it. **Population, read 2026-09-18T00:47:37Z → 00:48:42Z:** every open card on both boards this gate reads — **529 open cards in `objectstack-ai/objectstack`, 413 in `objectstack-ai/objectui` (942 total)**, of which **880** carry at least one comment and were read; **163** carry at least one claim comment. 9 comment lists sit at the 100-comment cap and are UNJUDGED past it, exactly as objectstack-ai#18683 prescribes. 0 parse failures. **The answer is not 0 — it is 12.** Twelve open cards carry LIVE `Claim:` comments from two or more DIFFERENT authors with no retraction between them: | repo | card | authors holding live claims | |:---|:---|:---| | objectstack | `objectstack-ai#13503` | `claude[bot]` + `baozhoutao` | | objectstack | `#14026` | `hotlong` + `claude[bot]` | | objectstack | `objectstack-ai#15811` | `os-bill` + `os-litant` | | objectstack | `objectstack-ai#17852` | `os-warren` + `os-litant` | | objectui | `objectstack-ai#4730` | `yinlianghui` + `os-sales` | | objectui | `objectstack-ai#7070` | `os-warren` + `claude[bot]` | | objectui | `objectstack-ai#7696` | `os-justin` + `os-tesla` | | objectui | `objectstack-ai#7804` | `os-tesla` + `os-sam` + `os-justin` | | objectui | `objectstack-ai#7848` | `claude[bot]` + `baozhoutao` | | objectui | `objectstack-ai#7924` | `os-warren` + `os-sales` | | objectui | `objectstack-ai#8115` | `claude[bot]` + `yinlianghui` | | objectui | `objectstack-ai#9370` | `os-tesla` + `os-justin` |⚠️ **This PR is deliberately SILENT on all twelve** — and that silence is pinned, per direction (b). Refusing an ownership transfer between sessions is not this card's to do: it is a different state, it would need its own remedy sentence, and a row that answered both would make one sentence out of two states. This is reported here and in the dispatch report so the seat can file it; ⛔ it is not folded in. ## Who the new exit 4 meets before it lands The seat needs this before landing, so I swept it rather than assuming. Two facts: 1. **No CI job turns red.** `check:pm-clause2-carriers` — the only wiring, `.github/workflows/lint.yml:1140` — runs `--self-test` and nothing else. No workflow runs `--pair` or a sweep, so landing this changes no required context. The exit 4 appears only when a seat runs `--pair` or a sweep by hand. 2. **On the objectstack board: nobody.** Of 32 open PRs in objectstack, **none** delivers a card that would newly earn a C8. Twenty open cards across both boards would earn the row (report-only, listed in the dispatch report), but only one is reachable through an open PR, and it is in the sibling repo: **`objectstack-ai/objectui#9584`** (open, not draft; its closing keyword names `objectui#9499`), which delivers a card carrying two live claims by `os-try-charles` (5663366106 on 2026-09-14, 5690579598 on 2026-09-16). That pair answers exit 4 at its next `--pair`. `DEFAULT_SWEEP_REPO` is `objectstack-ai/objectstack`, so objectui is only ever read when passed explicitly. ⛔ Nothing was posted on objectstack-ai#9499, objectstack-ai#9584 or any of the twelve. ## The reading, and WHERE it is computed `claimRepeats` (:1898) is a **sibling pure reader beside `claimRetractions`, built on it** — the same map decides membership here and for governance, so the pool and this row cannot describe two different retractions. It names every author holding more than one LIVE claim comment, orders them by the file's one recency rule (`laterOnThread`, to ORDER the record, never to pick a winner), and resolves no state, no row and no exit code. `c8SecondClaimSameSeat` (:4380) renders the verdict; `pairRows` (:4430) pushes it as row `C8`; `pairInputRecord` adds `claim.repeat` (:5876, declared in `INPUT_RECORD_PAIR_FIELDS` at :5641) as the READING — one derivation feeding both, so the record and the verdict cannot disagree about how many claims a seat holds or which they are. **MEMBERSHIP first, and that is what makes the state repairable.** The state is read over LIVE claims only. A re-claim after a `Release:` is the protocol working and reads exactly as it did before. And a seat that already wrote a second claim has an act that clears the row: `Release:` what it holds, then one fresh `Claim:`. A rule written over the writing *moment* instead ("no retraction strictly BETWEEN the two lines") would have been unrepairable by construction — nothing un-writes a comment — so the row would have been a permanent red with a remedy nobody could execute. ### The four axes - **实际业务需求** — measured, not assumed. Five live instances on the objectstack board at filing, re-confirmed today, every one of them read green before this row; plus 20 open cards across both boards that carry the state now. The first signal in five occurrences came from a dev reading a docblock, not from any instrument. This is a real shape occurring repeatedly, not a speculative surface. - **项目长远合理性** — contract-first, and no workaround. The rule already existed in writing at :378; this adds the reader that enforces it, in the same file, over the same thread, through the same membership derivation governance uses. No new exit code was minted, no second selector, no second reader of the marker. The prohibition and its reader now live one screen apart. - **防 AI 写代码犯错** — this is the axis that decides the exit. A second `Claim:` re-enters the pool as the newest claim and becomes what every downstream reader is handed — the property the correction key was deliberately designed NOT to have. Rendering that as a NOTE at exit 0 is precisely the tolerant-consumer shape this repo refuses: an adverse fact printed green is how a batch of identical mistakes stays invisible. Declaring the prohibition and not enforcing it is the "声明而未兑现" gap; the repair is to enforce it loudly, at `EXIT_PAIR_ADVERSE`. The row also ⛔ never prescribes WHICH repair — choosing between a correction and a release would be choosing whether the card is being re-taken, which is the seat's judgement, so it names both and writes nothing. - **创业阶段不扩散需求** — the surface added is one file, one pure reader, one row, one record field. It refuses exactly one shape the protocol already forbade in writing and re-blocks no legal workflow: a card claimed once reads as it always did, a re-claim after a `Release:` reads as it always did, a `Clause-②-correction:` is not a claim and never was. The cross-seat question, which is a genuine second capability, is explicitly NOT taken here. ## The pins — per direction, each with a non-vacuity control | | direction | reading | |:---|:---|:---| | (a) | same author, two claims, no retraction | **named, exit 4**; the row carries both ids, the author and both repairs | | (b) | DIFFERENT authors | supersession as today, exit 0 — ⛔ not this state | | (c) | same author after a `Release:` **or** the id-naming retraction | RETRACTED as today, exit 0; the `⛔ NOT superseded` wording byte-unchanged | | (d) | a `Clause-②-correction:` as the later row | silent; the objectstack-ai#17366 exit still reads the declaration off it | | (e) | a DECORATED second claim (bold, backticked) | counted through `markerMatches` exactly as a bare one; the raw constant refuses both, so the counting is the sibling's ONE reading | | (f) | a second claim whose `Branch:` parses to zero branches | still named — the prohibition is on the WRITING, not the parse | | (g) | three claims by one seat | **ONE** refusal naming all three, not two | | (h) | an unattributable row (`rowAuthor` null) | fail closed, as `claimRetractions` does — and `null` never groups with `null` | | (i) | a later same-author comment that QUOTES or DISCUSSES the word | silent — the marker is read at line start | Direction (i) has a control in the wild on this very card: the triage comment 5722477144 contains the word `Claim:` mid-line, and `markerMatches` refuses it — card objectstack-ai#18828 reads one claim comment, so `--pair` on this PR is silent. **Roster line** (:792): `'objectstack-ai#18828: a SECOND \`Claim:\` by ONE seat — the writer-side prohibition, finally READ': 52` — and the battery block declares exactly 52 `t(...)` cases. **Floor** (:806): `SELF_TEST_BATTERY_FLOOR` **31 → 32**, raised by exactly one. ## The ablation Run from the **committed** fix, twice, each leg proving its mutation landed on disk before the reading was taken and proving its restore by an empty `git diff HEAD` and by blob hash — never by an editing command's exit code. `HEAD` blob `3a270ef2eb5f33780e04e4732714f8e88d74a017`. | leg | mutation | mutated blob | result | |:---|:---|:---|:---| | baseline | none | `3a270ef2eb…` | **941 cases pass, exit 0** | | **A** — the repeat detection neutered (a group is never reported) | `72115d2796ead200f93aa855c8ba5820a83f5f8f` | | **23 of 941 failed**, exit 1 | | **B** — the SAME-AUTHOR check removed (the author no longer decides the grouping) | `40cfadd4f5740f34210675ceb998fb2977823769` | | **3 of 941 failed**, exit 1 | Both legs restored: `git diff HEAD` empty, blob back to `3a270ef2eb…`. **Total case count is 941 in all three runs** — the rest of the self-test is byte-identical in its case count, and in both legs **0 of the failures fall outside the objectstack-ai#18828 battery**. Leg A is the interesting one, because it shows the per-direction controls doing their job. Five of the nine directions assert SILENCE and therefore *cannot* go red when the detection is removed — their non-vacuity controls go red instead. All nine directions are carried: - pin itself red: **(a) (e) (f) (g)** - carried by its control: **(b) (c) (d) (h) (i)** — "make those two authors ONE", "drop the retraction", "write that same correction as a SECOND `Claim:`", "give that same row a login", "move that same word to the OPENING of a line" Leg B is the narrower, sharper one: removing only the author test reds **3** cases, and pin (b) is among them. That is the pin which distinguishes this card from the cross-seat question — proof the author test is load-bearing and that (b) is not vacuous. **Self-test count: 889 before → 941 after** (+52, exactly the registered battery). The 889 was measured by running `--self-test` in a detached worktree at `origin/main` `88aa326deb`. ## Gates Derived from the worktree with `node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack` (no hand-fed path list).⚠️ The first derivation printed **STALE TREE** — the branch was 2 commits behind `origin/main` and 8 files the derivation reads had changed — so `origin/main` was merged in first and the list re-derived on the merged tree at `7424cf3f44`; the `--repo` assertion holds against this checkout's `origin`. **34 derived, 34 run, all exit 0.** Each exit code captured redirect-then-`$?`, never across a pipe. ``` node scripts/check-adr-0087-registration.mjs --base origin/main :: exit 0 node scripts/check-adr-0087-registration.mjs --self-test :: exit 0 node scripts/check-changeset-no-major.mjs --base origin/main :: exit 0 node scripts/check-changeset-no-major.mjs --self-test :: exit 0 node scripts/check-ci-filter-parity.mjs :: exit 0 node scripts/check-closing-keyword-parity.mjs :: exit 0 node scripts/check-closing-keyword-parity.mjs --self-test :: exit 0 node scripts/check-comment-mask-corpus.mjs :: exit 0 node scripts/check-declaration-mirrors.mjs :: exit 0 node scripts/check-declaration-mirrors.mjs --self-test :: exit 0 node scripts/check-scripts-symbol-anchors.mjs :: exit 0 node scripts/check-scripts-symbol-anchors.mjs --self-test :: exit 0 node scripts/check-self-test-wired.mjs :: exit 0 node scripts/check-self-test-wired.mjs --self-test :: exit 0 node scripts/check-self-test-workflow-commands.mjs :: exit 0 node scripts/check-self-test-workflow-commands.mjs --self-test :: exit 0 node scripts/check-whole-set-label-write.mjs :: exit 0 node scripts/check-whole-set-label-write.mjs --self-test :: exit 0 node scripts/pm/bare-root-worklist.mjs --self-test :: exit 0 pnpm check:agent-test-spelling :: exit 0 pnpm check:bash32-floor :: exit 0 pnpm check:changeset-gate-self-tests :: exit 0 pnpm check:cli-command-ids :: exit 0 pnpm check:cross-package-test-inputs :: exit 0 pnpm check:driver-memory-census :: exit 0 pnpm check:entry-guard :: exit 0 pnpm check:nul-bytes :: exit 0 pnpm check:parse-guard :: exit 0 pnpm check:pm-clause2-carriers :: exit 0 pnpm check:pm-dispatch-gates :: exit 0 pnpm check:pnpm-filter-targets :: exit 0 pnpm check:ratchet-remedy-authority :: exit 0 pnpm check:refd-timer-probe :: exit 0 pnpm check:watch-hint-literal :: exit 0 ``` Reconciled with `--ran`, exit codes included: **34 derived, 34 run, 0 NOT-MEASURED, 0 UNRUN** — "a DERIVED zero — all 34 recorded an exit code and none of them is 3". Repo-wide `pnpm lint` (`eslint . --no-inline-config`): **exit 0**. The heavy run took a ticket through `scripts/pm/os-verify-lock.sh` (slot `issue-18828-dev`), queued behind the seat's own `dispatch-gates.mjs --self-test`. `node scripts/pm/check-clause2-carriers.mjs --pair` on this PR is reported in the dispatch report — this card carries one claim comment, so the row is silent on it. `skip-changeset`: `scripts/pm/**` publishes nothing from any released package — the whole diff is one non-published script. --- _Generated by [Claude Code](https://claude.ai/code/session_01Gqi43smmqjJ5sUrhfoPeKu)_ --------- Co-authored-by: Claude <noreply@anthropic.com>
Fixes #4632
把 #4460 的点状修复定成规则,并且给规则装上牙齿。
一、规则落在哪,为什么
锚点选了
AGENTS.md,新增一节「Degradation log levels —warnvserror」,放在 “Route & surface ownership” 之后。理由:catch时当场适用的判据,不是某个子系统的架构决策。ADR 回答「我们决定采用什么架构」,而这条回答「你现在敲的这一行该用哪个级别」—— 后者属于每个 agent 编辑前必读的 AGENTS.md。catch,不限于插件 init 期。挂在 ADR-0116(plugin ordering declared contract)的 addendum 上会把它误缩到「插件排序」这一个语境里 —— 而 [automation/approvals] 进程重启后审批决策静默失效:挂起 flow run 仍只存内存(#1518 标记 COMPLETED 但 17.0.0-rc.1 未生效),approve 落库却永不推进且零报错 #4420 只是碰巧发生在那里。warn,正是「对外宣称持久化、运行时并不交付」的一个实例。新节显式交叉引用了这两处。文案是判定问句,不是形容词(按验收要求):
并规定提到
error的每一处必须在第一行同时给出后果与修复动作(照service-automation/src/plugin.tsstart()的 #4460 范本),且只说一次,不是每次写失败都刷一遍。同时明确写了反向失败:把功能性降级误提到
error会训练所有人跳读error,而那正是当初 #4420 那条warn没人读的原因。规则特别点出if (!service)的组合分支通常是功能性的、该留在warn;真正咬人的是包住写入 / DDL / store 初始化的catch。二、机械检查:做了
「一条只写在 AGENTS.md 里、没有任何东西检查的约定,就是本仓反复修的那个『声明 ≠ 强制』形状」—— 这个批评是对的,所以做了闸门。
pnpm check:durability-log-level(scripts/check-durability-degradation-log-level.mjs,已接入lint.yml的必需 job):AST 遍历所有catch,若其try块调用了词表内的持久化关键操作,则该catch必须error(或fatal)或重新抛出;记warn/info/debug或静默吞掉即失败。几个刻意的设计:
/save|write|persist/之类的正则去猜,误报率会高到让人把闸门关掉 —— 而一个被关掉的闸门比没有闸门更糟,因为它还会报成功。词表(DURABILITY_CRITICAL_CALLEES)很小、每条都写明「为什么它的失败意味着字节没落盘」,加一条是需要评审的动作。await driver.syncSchema(...)和它的日志隔着两层嵌套循环;[automation/approvals] 进程重启后审批决策静默失效:挂起 flow run 仍只存内存(#1518 标记 COMPLETED 但 17.0.0-rc.1 未生效),approve 落库却永不推进且零报错 #4420 那次干脆在 init() 调用的私有方法里)。判据是块结构,不是行距离。try块内不下沉进嵌套函数体(ctx.hook('kernel:ready', async () => …)里的调用晚于这个 catch 执行,不受它守护),这和check-init-service-contract.mjs的取舍一致。catch侧反过来会跨同文件 helper 追踪。这一条是实测逼出来的:我先把两处重复的错误文案抽成reportSyncFailure()helper,闸门立刻把它们报成「静默吞掉」。如果不追踪 helper,「把失败藏到一层间接后面」就成了让闸门闭嘴的最省事做法 —— 那正是 [automation/approvals] 进程重启后审批决策静默失效:挂起 flow run 仍只存内存(#1518 标记 COMPLETED 但 17.0.0-rc.1 未生效),approve 落库却永不推进且零报错 #4420 形状本身。console.error与logger.error同等响亮。也是实测发现的:metadata/src/loaders/database-loader.ts的 history-schema 分支用console.error老实上报,把它判成违规会是纯误报。--self-test(10 例,双向都钉:必须抓到 [automation/approvals] 进程重启后审批决策静默失效:挂起 flow run 仍只存内存(#1518 标记 COMPLETED 但 17.0.0-rc.1 未生效),approve 落库却永不推进且零报错 #4420 形状,也必须不抓功能性降级),--list打印全部缝隙及判定。scripts/durability-degradation.baseline.json,手写、带理由、shrink-only,且故意不提供--fix/--update—— 有了那个 flag,新违规就能靠「跑一下更新命令」混进来,闸门也就不再意味着什么。条目修好后不删会报 stale 而变红。闸门的局限,写在脚本头部而不是等人踩:它发现不了词表之外的新持久化缝隙,只能保证已经付过学费的那些不再退回
warn。这是棘轮,不是证明。发现新缝隙时,请在修它的同一个 PR 里把它加进词表。第二条局限:catch的响亮判定扫的是整棵子树,所以一个通过嵌套 try 恢复的 catch(batch→sequential 那个)会因恢复路径的error而通过 —— 对那个形状这是正确判决,但确实意味着一个无关的嵌套logger.error也能满足闸门。收紧它会误伤合法的恢复型 catch,那是更糟的交易。当前实测结果:8 处持久化关键缝隙,7 处已响亮,1 处 baseline(冻结包,见下)。
三、盘点结果(逐条,含判为第一类的)
判定问句 = 「降级后系统对外表现是否仍然『正常』,而声称已持久化的东西实际没有落盘?」
改为
error(第二类)objectql/src/plugin.ts~405 reload-time schema syncobjectql/src/plugin.ts~1002 / ~1018 per-object syncobjectql/src/plugin.ts汇总行info: Schema sync complete—— 这就是「看起来正常」的那一半。现在failed > 0时是error并报出计数objectql/src/engine.ts~5340syncSchemas()catch {},注释写「log suppressed to avoid noise on already-synced tables」。但syncSchema契约要求幂等,能走到这个 catch 的驱动就是没同步成功;唯二调用方是 marketplace 插件安装与模板 seeding,它们随后就往这个失败意味着不存在的表里 INSERT,然后报告安装成功。这一处是闸门发现的,不在 PM 名单里service-automation/src/plugin.ts~882 wait-timer 外层service-automation/src/builtin/wait-node.ts~160 list 失败return 0,一个 timer 都没武装,整个子系统级别的失效service-automation/src/builtin/wait-node.ts~176 overdue resume 失败service-automation/src/builtin/wait-node.ts~205 re-schedule 失败关于 wait-node 这三处:数据确实落盘了,坏掉的是恢复。之所以仍判第二类 —— 持久化状态与运行时状态互相矛盾,而看起来没事的是运行时;对那个等审批的人来说,结果与 #4420 的丢失不可区分。这也是为了和同一子系统的外层 catch 保持一致:在一个文件里对同一后果给出两种级别,比任何一种选择都糟。
判为第一类,不改(逐条给理由)
warn/infomcp/src/plugin.ts164objectql/src/plugin.ts215 / 224 authored hook/action rebindobjectql/src/plugin.ts431app.*弃用提示objectql/src/plugin.ts611 / 682 registry ingest / event handlerobjectql/src/plugin.ts938 external object 注册objectql/src/plugin.ts992 batch→sequentialerrorobjectql/src/plugin.ts1279objectql/src/plugin.ts1692 / 1706 authored action runnerservice-analytics/src/plugin.ts213data服务,每次查询重试;真的一直缺就在调用点 throw,响亮service-analytics/src/plugin.ts585 getReadScope 缺失error会在每个单租户 host 每次启动制造噪音 —— 正是规则的反面失败。规则按维护者原文保持在「持久性 / 数据一致性」,我没有擅自扩到安全域service-automation/src/plugin.ts452error兜底service-automation/src/plugin.ts491 / 758 / 766 / 1355 / 1412 / 1456service-automation/src/plugin.ts847 / 910warn(触发器未绑定 / 声明式 connector 未注册),都已带修复动作service-automation/src/plugin.ts1180errorservice-automation/src/builtin/wait-node.ts184!joberror会在每个无 job host 的每次启动、每个挂起 run 各刷一次trigger-api64 / 68 / 80trigger-record-change49 / 57trigger-schedule53 / 62四、排除的两个包:有发现,已单开 issue
按本轮硬约束,
packages/metadata/**与packages/metadata-protocol/**零改动,发现只记录:packages/metadata/src/loaders/database-loader.tsensureSchema():catch完全静默且把schemaReady置为true,与成功路径不可区分。注释的免责理由("e.g. table already exists")只覆盖最良性的一种失败原因,却为全部失败原因开脱。闸门能自动发现这一处,已挂在 baseline 并指向 [metadata] database-loader 吞掉 sys_metadata 的 DDL 失败后仍置 schemaReady=true —— 第二类降级(#4632 规则),本轮因包冻结未修 #4728;修复时请一并删掉 baseline 条目(shrink-only,残留会变红)。packages/metadata-protocol/src/seed-loader.ts~961:注释与代码直接矛盾 —— 注释明写 "this must be a reported, counted error, never a silent warning",紧跟着的调用是logger.warn。后果按注释自己的说法:引用停在NULL,循环关系半写入。(公允地说,它确实通过recordDeferredError计入了allErrors,不是完全静默;但日志行本身是这次 seed 唯一可见的痕迹。)闸门发现不了这一处 —— 写路径没有可识别的调用名,这是上面那条局限的真实例子,issue 里也写明了。另有
packages/metadata/src/plugin.ts388(Failed to attach FileSystemRepository)存疑:后果取决于 manager 无 repository 时的行为,冻结包内未深挖,未作断言。若确证属第二类,建议并入 #4728。packages/spec/**零改动;content/docs/releases/未触碰。五、验证
pnpm check:durability-log-level—— self-test 10 例通过;实测 8 缝隙全部响亮或重抛(1 条 baseline)pnpm --filter @objectstack/objectql test/pnpm --filter @objectstack/service-automation testpnpm --filter @objectstack/objectql typecheck/pnpm --filter @objectstack/service-automation typecheckerror且文案含后果与修复动作;功能性降级(batch 回退、无 job 服务)必须留在warn;全部成功时一句不吭已加 changeset(
.changeset/durability-degradation-log-level.md)。无 API / schema / 行为变更 —— 只改了已经在失败的路径报告的级别和文案。Generated by Claude Code