docs: TypeScript examples in PUBLISHED package READMEs that do not compile — split the #18751 census into published-README vs internal-document blocks and correct the wrong examples in published READMEs only (#18715 ruling F)

[hotlong]

Filed by the director seat (objectstack #12708, summon #24, session_01Wj1HUjzyeiBQ8atRf1ZhaL) executing decision batch #156 item 2 — maintainer 「同意」 to F on #18715: ⛔ no gate is built; the user-facing half of the census is worked here.

The reading (PR #18751's census instrument, on main; re-run before acting)

1,002 TypeScript fenced blocks in package-root Markdown. Hand-written documents: 54 files / 282 blocks — 195 fail to compile; 76 under the tolerant reading (legal elisions allowed); 58 syntactically valid and wrong (copy-paste fails). CHANGELOGs (720 blocks) are out of scope: they correctly record deleted APIs.

Ask

1. Split the 76 / 58 by publication: a block is published when its document is inside its package's files[] (README.md and siblings that ship); everything else (PHASE2_IMPLEMENTATION.md, design notes, internal guides) is internal. Record both counts on this card with the instrument's output.
2. Correct the published-README examples that are syntactically valid and wrong, one PR per package or one PR overall — the claimant's call — so that every example a platform user's AI can copy compiles against the package it documents. A block that documents a removed API is rewritten or deleted, ⛔ never left as a fossil.
3. ⛔ Internal-document blocks are not worked here. If an internal document is dead, deleting it is the fix, on its own card.
4. ⛔ No gate, no ratchet, no CI wiring (ruling F). Clause-②: no; changeset skip-changeset (docs only) unless a package ships its README changes in a release note by convention.

Refs

#18715 (ruling F) · PR #18751 (the census instrument) · #15931 · #18620 / PR #18712

Dedup terms: published README examples compile, fenced TypeScript block typecheck, files[] published document split

---

Blocked-by: #18973

⚠️ 由 domain:devx 执行席在 PR #18968(Part of)落地后补上,⛔ 纯增量。本卡 44 个已发布错块修了 43;剩下的 1 块(packages/plugins/knowledge-ragflow/README.md)卡在 #18973 的裁决上 —— 那里在问 source.options 与 adapterConfig 哪个是契约,两种收法都是运行时改动,⛔ 不是文档卡能自答的。

⚠️ 释放机制是人工,⛔ 不会自动触发(H26 的形状):#18973 现读 needs-user-decision,而被回答⛔ 不产生关闭事件,所以纯 Blocked-by: 在这里是一个永不自动解锁的块。

• Release-fires-on: @objectstack/knowledge-ragflow must read source.adapterConfig.datasetId — the declared key (ruled: batch #160 item 2, letter A; README moves with it) #18973 被回答(A / B 任一被裁定)
• Release-mechanism: 人工 —— 承接席读到裁决后自行放回,⛔ 不等解锁扫描

---

Generated by Claude Code

[os-try-charles]

Claim: PM loop round 44
Session: session_017ef78bLdybu3AffehKkhfk
Branch: claude/issue-18915-published-readme-examples-compile
Worktree: objectstack-issue-18915
Domain: domain:devx
File surface: packages/**/README.md 及各包 files[] 实际发运的同级 Markdown(stop on breach; explain in the report)。⛔ 面外:packages/*/CHANGELOG.md(卡面自己划出)、任何内部文档(ruling F 第 3 条)、content/docs/**。
Container & model: M, mode:subagent, model: opus (当次 dispatch-gates.mjs --tier packages/spec/README.md packages/client/README.md 输出逐字:「no path-derived mandate: the surface hits none of the 3 declared glob(s)… floor sonnet · default opus · ceiling fable」⇒ 无路径强制,取默认档)
Clause-②: no
Thread-read: none
Serial constraints cleared: 34 个 open PR 的 files 逐个读过(共读到 413 个变更文件 ⇒ 读法有反应)。触碰 packages/**/*.md(不含 CHANGELOG)的只有 PR #18938(draft),面是 packages/spec/liveness/README.md 与 packages/spec/liveness/state-counts.md —— liveness/ 子目录,不是包根 README,与本卡的面不重叠;若施工席量到 liveness/README.md 落在 @objectstack/spec 的 files[] 里因而属于本卡总体,以它的实测为准,并把该文件让给 #18938。

⚠️ 一条本席要施工席格外当心的判据倒转

本班多数文档卡的 skip-changeset 论证是「没有任何包的 files[] 够得着 content/docs/**」。本卡正好相反:README.md 通常就在 files[] 里,也就是说改动的正是会发运的东西。 卡面自己留了口子(「unless a package ships its README changes in a release note by convention」)。

⇒ ⛔ 不许照抄上一张卡的结论。changeset 要不要,由本卡自己的测量决定,并且要带阳性对照。

_{读数时刻 2026-09-18T08:00Z}

---

Generated by Claude Code

[os-try-charles]

os-dev-report

{
  "issue": 18915,
  "status": "rework",
  "branch": "claude/issue-18915-published-readme-examples-compile",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/18968",
  "premise_still_valid": true,
  "summary": "ACT 1 (split) delivered in full. PR #18751's instrument re-run on a fresh origin/main reproduces the card's numbers exactly: handwritten stratum 282 blocks / 54 files, raw 195, tolerant 76, well-formed-and-wrong 58. Split by publication (a block is published when its document is inside its package's files[]): TOLERANT 76 = 54 published + 22 internal; WELL-FORMED-AND-WRONG 58 = 44 published + 14 internal. The split is clean because every non-private package's files[] is [dist, README.md, CHANGELOG.md] (only @objectstack/spec lists more), so the published package-root Markdown is exactly README.md and every other package-root document is internal; confirmed with npm pack --dry-run against packages/core, where README.md packs True and ADVANCED_FEATURES.md / PHASE2_IMPLEMENTATION.md pack False (same-population control that must read the other way). No count in the split is zero, so no zero needed pairing. ACT 2 (correct the published ones) is 43 of 44, across 20 package READMEs; after-census on the final tree a8b75f978: published well-formed-and-wrong 44 -> 1, published tolerant 54 -> 11 (the other 10 are syntax-only fragments, outside act 2's mandate), internal 14/22 unchanged and untouched (ruling F item 3). All three instrument controls behave on every run cited. The ONE block not corrected is packages/plugins/knowledge-ragflow/README.md: it writes source.options.datasetId, which is what the shipped adapter reads (extractRagflowOptions casts to a shape with an optional options record and its error text names that path) and is NOT what KnowledgeSourceSchema declares (adapterConfig, on a plain z.object that strips unknown keys). Rewriting the doc to adapterConfig would make it compile and stop working; correcting the adapter is a runtime change and picks a winner between two live spellings. Contract-first says the defect is upstream, so the block is left and the conflict is reported. That is why the PR says Part of #18915 and not Fixes, and why this report is rework rather than done: act 2 is one block short pending that decision. CHANGESET, measured for this diff and not inherited: the house skip-changeset argument INVERTS here. README.md is in files[] for all 20 touched packages, so the changed bytes ship in the tarball (npm pack control above). AGENTS.md scopes skip-changeset to a diff that publishes nothing from any released package, so this PR carries a patch changeset naming all twenty and NO skip-changeset label. SERIAL CONSTRAINT: packages/spec/liveness/README.md measures as PUBLISHED (liveness is a files[] entry; npm pack puts liveness/README.md and liveness/state-counts.md in the spec tarball) but is NOT in this card's population - the census population is Markdown at a package ROOT and packages/spec/liveness carries no package.json, so it contributes none of the 76/58. No overlap with PR #18938 and nothing of theirs is touched. SCOPE NOTE: the one path outside the claim's declared file surface (packages/**/README.md) is .changeset/18915-published-readme-examples-compile.md, the companion artefact the measurement above obliges; named rather than slipped in. Beyond the measured set, three further client-react blocks (Master-Detail, Search with Debounce, the Type Safety comment) carrying the SAME defect were corrected - the census does not flag them only because they import nothing and type-check as any; declared in the PR body. assignee was already set by PM dispatch (os-try-charles) and the newest Claim: names this branch; I wrote neither.",
  "tests": "Tree a8b75f978 (origin/main merged in, pnpm install --frozen-lockfile after the lockfile moved, full rebuild through scripts/pm/os-verify-lock.sh). CENSUS: node scripts/measure-markdown-ts-blocks.mjs --json, exit 0, controls green=clean firing=fires unpublished-subpath=fires. handwritten blocks 284 files 54 raw 172 tolerant 33 wellFormedAndWrong 15; published tolerant 11 / WFW 1; internal tolerant 22 / WFW 14. (284 not 282 because the observability block that redeclared metrics four times is now three fences.) BEFORE-run on the same instrument: blocks 282 raw 195 tolerant 76 WFW 58. GATES: node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack derived 63 families on the merged tree with no stale-tree warning; all 63 run, all exit 0, each exit captured before any pipe; --ran reconciliation prints '63 derived familt(ies) accounted for - 63 run, 0 NOT-MEASURED (a DERIVED zero - all 63 recorded an exit code and none of them is 3)'. check:pm-dispatch-gates is NOT among the derived families for this change set, so its ~430s prescription does not apply. The tool's own caveat stands: this answers one link, not CI. TESTS: the diff changes no TypeScript, so no tsc program or vitest source set moves; the README-reading tests were found by grepping every test file under packages/ for README.md (19 hits) and triaging by the path each reads. Two read a README this PR edits, both run: pnpm --filter @objectstack/client exec vitest run --maxWorkers=2 src/readme-package-install-example.test.ts -> Test Files 1 passed (1), Tests 6 passed (6), CLIENT_README_EXIT=0 (this one parses packages/client/README.md with the TypeScript parser and validates the manifest against the install contract - the pin a README edit there could break); pnpm --filter @objectstack/spec exec vitest run --maxWorkers=2 src/api/package-api.test.ts src/kernel/plugin-structure.test.ts -> Test Files 2 passed (2), Tests 81 passed (81), SPEC_TARGETED_EXIT=0. LINT: zero files in this diff are in eslint's population, read from eslint's own config rather than assumed - ESLint#calculateConfigForFile gives packages/types/README.md rules 0 / ignored true, the changeset file rules 0 / ignored true, and the control packages/types/src/index.ts rules 6 / ignored false. eslint.config.mjs scopes every block to {ts,tsx,mts,cts,js,jsx,mjs,cjs}, so no config in this diff can move an untouched file's verdict; the narrowing is therefore a measurement, not a skip. turbo ls --affected is NOT cited: its set here is dominated by the origin/main merge (pnpm-workspace.yaml, pnpm-lock.yaml, spec sources), not by this diff. CONTROL BYTES: grep -naP over the C0 range on every changed path - no hits; a fixture carrying one such byte does hit, so the scan is live. pnpm check:nul-bytes is among the 63 green gates. CLAUSE-2: readClause2Line over the PR body prints {\"kind\":\"declared\",\"value\":\"no\",\"arm\":null,\"line\":\"Clause-\\u2461: no\"}; node scripts/pm/check-clause2-carriers.mjs --pair 18968 exits 0 with pair.1.claim.clause2-line: DECLARED `no` and pair.1.pr-body.clause2-line: DECLARED `no` (the tool notes the pr-body row is an INPUT only and that check-changeset-no-major.mjs is what reads that line). NO ablation and no reverse verification: this diff adds no guard and changes no runtime behaviour, so neither artefact has a subject here. PR body read back after both writes: create stored the session-URL footer byte-identical; the raw REST PATCH was sent with BOTH footer spellings stripped per the measured cell and stored exactly one bare footer, +58 bytes, body otherwise byte-identical.",
  "mcp_calls": "0 - no MCP GitHub tool was called; every GitHub read and write went through the REST proxy with curl.",
  "api_writes": "3 - POST /repos/objectstack-ai/objectstack/pulls (draft PR #18968); PATCH /repos/objectstack-ai/objectstack/pulls/18968 (body: added the Verification section); POST /repos/objectstack-ai/objectstack/issues/18915/comments (this report). Plus two git pushes (empty-branch routing probe, then the commit). NO label write was owed: a changeset ships, so no skip-changeset; Clause-2 is no, so no needs:contract-review. Budget was 4 writes including a label write; 3 used.",
  "open_questions": [
    {
      "question": "@objectstack/knowledge-ragflow reads source.options.datasetId at runtime while KnowledgeSourceSchema declares adapterConfig and, being a plain z.object, strips options on parse. The published README documents options, which therefore does not type-check. Which spelling is the contract?",
      "options": [
        "A - the SPEC wins: change the adapter (and the README) to read source.adapterConfig. One runtime change in knowledge-ragflow, no spec change; any host already writing options breaks, and the schema would have silently stripped it anyway on any path that parses.",
        "B - the ADAPTER wins: declare options on KnowledgeSourceSchema (or an ADR-0087 conversion from options to adapterConfig) so the shipped spelling is a declared key. Spec change plus the generated artefacts it drags.",
        "C - leave both and correct nothing: the README keeps documenting a key the type rejects. Rejected here - that is the declared-not-enforced shape Prime Directive #10 forbids."
      ],
      "recommendation": "A, because adapterConfig is the DECLARED key, the cast in extractRagflowOptions is exactly the consumer-side leniency Prime Directive #12 forbids, and anything that routes a source through KnowledgeSourceSchema.parse already drops options - so B blesses a spelling that only works because nothing parses. Either way it is a runtime change outside a docs card, which is why this PR leaves the block rather than guessing."
    }
  ],
  "out_of_scope_findings": [
    "to file (class c, metadata-authoring trap; dedupe: knowledge source options adapterConfig, ragflow datasetId undeclared key, KnowledgeSourceSchema strips unknown keys, adapter config spelling drift): @objectstack/knowledge-ragflow's extractRagflowOptions casts a KnowledgeSource to a shape with an optional options record and errors with 'RAGFlow adapter requires source.options.datasetId', while KnowledgeSourceSchema declares adapterConfig and no options; the schema is a plain z.object, so a parse strips the key the adapter needs. Its published README documents options and does not compile. This is the one block act 2 leaves; see open_questions.",
    "to file (class a, reproducible defect; dedupe: driver-memory default export onEnable, kernel.use rejects default plugin, legacy onEnable plugin shape, drop-in plugin not a Plugin): packages/drivers/driver-memory/src/index.ts default-exports { id, version, onEnable }, which is not a kernel Plugin - it has no name and no init, so kernel.use(memoryPlugin) fails to type-check AND assertPluginContract refuses it at boot with PLUGIN_CONTRACT_VIOLATION. Nothing in the repo imports it (one git grep hit, the README). The README's quick start and its 'Key Exports' row are corrected in this PR; the export itself is a runtime decision (retire it, or make it a real Plugin) and is left alone.",
    "noted, not filed: 10 published blocks remain tolerant-failing and are ALL syntax-only - bare type-signature fragments in service-automation (5), trigger-record-change (2), trigger-schedule (2) and service-job (1). They are the instrument's own needs-explicit-partial-tag cost line, not the syntactically-valid-and-wrong set ruling F scopes. Successor: whoever answers the census's wiring letter on #18715, which is where the 'ts partial' info-string convention is proposed. Nothing files this today.",
    "noted, not filed: the internal half of the census is untouched by design (ruling F item 3) and still reads 22 tolerant / 14 well-formed-and-wrong, concentrated in packages/core/ADVANCED_FEATURES.md (6 WFW), packages/spec/V3_MIGRATION_GUIDE.md (4), packages/plugins/plugin-auth/ARCHITECTURE.md (2), packages/spec/DEVELOPMENT_PLAN.md (1), packages/spec/ZOD_SCHEMA_AUDIT_REPORT.md (1). I did NOT measure whether any of them is dead, so I file no deletion card. Successor: the internal-document card ruling F item 3 anticipates.",
    "noted, not filed: packages/client/README.md documents data.find()'s legacy vocabulary (select / filters / sort / top). Unlike the client-react case this COMPILES - QueryOptions still accepts it - so it is outside this card's set; but find itself carries @deprecated and data.query() is the canonical call. Successor: the PR that retires data.find(). In the PR's Acceptance notes.",
    "noted, not filed: createRestApiPlugin's config nests as { api: { api: { version, basePath } } } because RestApiPluginConfig.api holds a whole RestServerConfig whose own api member carries those keys. The README now states the real shape (and the repo's own tests use it), but the doubled key reads like a typo to every first reader. Ergonomics, not a defect; successor: none identified."
  ]
}

Session: session_017ef78bLdybu3AffehKkhfk —— 本 os-dev-report 由本席在本会话内派发的 os-dev 子代理产出并由本席原样转贴,⛔ 非本席自述。补记于 H64 的未署名写入行(2026-09-19 巡逻捞出)。

---

Generated by Claude Code

[os-try-charles]

ACCEPT(部分) —— PR #18968 复核通过,⛔ 但本卡不关:它是 Part of,余一块待裁。

domain:devx 执行席(座位贴 #6023)。以下读数都是本席自己在 PR head a8b75f978(取数 2026-09-18T08:57Z)的独立工作树上量的,⛔ 不是转录报告的自述。

⭐ 那处「判据倒转」,施工席量了,没有照抄本班的房屋论证

本班多数文档卡用「没有任何包的 files[] 够得着 content/docs/**」来论证 skip-changeset。本席在派发令里明写这条在本卡上反向。复核结果:

本 PR 改动的 README : 20 个
其中 README.md 被列进所属包 files[] 的 : 20 / 20
未被列进的 : 无

⇒ 改的字节就是发运的字节,skip-changeset 在这里是错的。PR 携带一份 patch changeset 并点名全部 20 个包,⛔ 没有 skip-changeset 标签 —— 与测量一致。⭐ 这正是「顺手的论证要先问它量的是什么」的一次正面用例。

冲突本身,本席独立复核过

读到	位置
适配器读 source.options.datasetId	knowledge-ragflow/src/index.ts:60,:66,:108
schema 声明的却是 adapterConfig	packages/spec/src/ai/knowledge-source.zod.ts:137
KnowledgeSourceSchema = lazySchema(() => z.object({ —— 没有 .passthrough()	同文件 :124

⇒ 任何真的 parse 的路径都会把适配器赖以工作的键剥掉。把 README 改成 adapterConfig 会让它编译通过但停止工作 —— 所以不修那一块是对的,⛔ 硬修才是错的。

H52:这一问⛔ 不重新挂回本卡

那是执行过程中冒出来的余项,不是本卡的派发问题。按 H52 的处置,已另立 #18973(决策卡,四棱块与 Prior rulings read: 已在正文里;读过的八条 ADR 中最近的是 ADR-0087,它提供 B 所需的转换机制但⛔ 没裁过哪个拼写是契约)。

另接出 #18974 —— driver-memory 的默认导出不是 kernel Plugin(给 id 不给 name、用 onEnable 不用 init),本席复核:全仓按默认导入的命中为 0,真实使用点都走具名的 InMemoryDriver。⇒ 它是一个只对文档读者存在的导出,而退役或改造是运行时决定,⛔ 超出本卡的面。

本卡余量与去向

• 已落:44 个已发布错块修了 43,跨 20 个包 README。
• 未落:1 块(knowledge-ragflow/README.md),卡在 @objectstack/knowledge-ragflow must read source.adapterConfig.datasetId — the declared key (ruled: batch #160 item 2, letter A; README moves with it) #18973 的裁决上。
• ⛔ 不算余量:另外 10 个已发布 tolerant 块是纯语法片段,ruling F 的第 2 条不含它们;内部文档 22/14 按 ruling F 第 3 条刻意不动。

⇒ PR 落地后本卡将转 pm:blocked,Blocked-by: #18973,并同笔清 assignee、留交接行。⛔ 不转 pm:queue:余下那一块此刻没有人可以合法地把它做完,状态标签说的是「现在能不能被拿起来做」,⛔ 不是「还剩多少活」。

CI 仍在跑(16 项未完),绿了再按三步装弹。

_{读数时刻 2026-09-18T08:59Z}

---

Generated by Claude Code

[os-try-charles]

Release: session_017ef78bLdybu3AffehKkhfk · 因:PR #18968 以 Part of 落地 44 个已发布错块中的 43,卡按规矩⛔不关 · 去向:转 pm:blocked,Blocked-by: #18973(正文已补),⛔ 不转 pm:queue —— 余下那一块此刻没有人可以合法做完

⭐ 这一条是 Release: 指令的现在时正确用法:卡是真的离手。(⚠️ 对照:本席上一轮在 #18143 上拿它写补记,结果撤销了自己当时的活认领 —— 该指令⛔ 不读日期,只读作者。)

---

落地两读(卡 #18915 · PR #18968)

读① 队列侧 —— ready_for_review 2026-09-18T09:17Z → added_to_merge_queue 2026-09-18T09:19Z → merged 2026-09-18T09:41Z。对照:同一次 git ls-remote --heads origin 'gh-readonly-queue/*' 读到其他队列分支 1 条 ⇒ 列表是活的,这条腿成立。

读② 内容侧 —— 重取 origin/main(a48496640,取数 2026-09-18T09:42Z):

探什么	读到	发火对照(取自被探文件自身)
changeset .changeset/18915-published-readme-examples-compile.md	在树上,点名 20 个包	—
packages/client-react/README.md 的旧词 select: / sort: / top: / filters:	各 0	npm install — diff 内 0 / 文件内 1;React — 0 / 2

⚠️ 候选 Peer 被否:它 in-file 0,⛔ 不能当对照。⭐ 预挑的三个候选里有一个当场作废 —— 与 #18946 那次(四个候选全废)同一课:预挑不等于可用,落地当刻必须重验。

⛔ 两读都不取 merged 布尔,也不取 head-sha 祖先关系。

⚠️ 一条本席没有测成的读数,如实记为 NOT MEASURED

本想用更强的后置条件收口 —— 在合并后的树上重跑普查仪器,确认已发布 WFW 44 → 1。仪器拒绝出数,并给了理由:

✗ measure-markdown-ts-blocks REFUSES to report a census:
  GREEN_CONTROL did NOT compile (TS2307). A block that must compile did not,
  so nothing else this run would report is a reading.
  The usual cause is an unbuilt workspace: run `pnpm build` …

⭐ 这是仪器做对了:它宁可不报,也不报一个错的普查。

本席随后在带锁的工作树上跑了 pnpm build 再取一次,仍然没取到,而原因是本席自己的探针装置,⛔ 不是这次落地、也⛔ 不是仓库:

@objectstack/spec:build: src/ai/agent.zod.ts(3,19): error TS2307: Cannot find module 'zod'
@objectstack/spec:build: WARN Local package.json exists, but node_modules missing, did you mean to install?

本席的复核工作树把 node_modules 符号链接到另一个工作树,因此各包自己的 node_modules 不存在,zod 解析不到 ⇒ @objectstack/spec#build 失败 ⇒ 没有 dist/*.d.ts ⇒ 普查仪器继续拒绝出数。

⇒ 这条读数记为 NOT MEASURED,原因是本席的装置。⛔ 本席不转述施工席报的「已发布 WFW 44 → 1」 —— 那是它的读数,不是本席的。落地本身由上面两读独立成立,⛔ 不依赖这一条。

本卡余量

• 已落:43 块,跨 20 个包 README。
• 未落:1 块(knowledge-ragflow/README.md),等 @objectstack/knowledge-ragflow must read source.adapterConfig.datasetId — the declared key (ruled: batch #160 item 2, letter A; README moves with it) #18973。
• ⛔ 不算余量:另 10 个已发布 tolerant 块是纯语法片段(ruling F 第 2 条不含);内部文档 22/14 按 ruling F 第 3 条刻意不动。

changeset:判据倒转,本席复核过

本班多数文档卡用「没有包的 files[] 够得着」论证 skip-changeset。本卡反向,本席实测:改动的 20 个 README,20 个都在所属包的 files[] 里 ⇒ 改的字节就是发运的字节。PR 因此带 patch changeset 点名全部 20 包、⛔ 无 skip-changeset 标签 —— 与测量一致。

席位状态随此清理:pm:dispatched → pm:blocked,解除指派。

_{读数时刻 2026-09-18T09:45Z}

---

Generated by Claude Code

[os-try-charles]

⏱️ 本卡的释放条件已触发,且它的性质变了 —— 从「人工释放」变成「机械可释放」。

domain:devx 执行席(座位贴 #6023)。⛔ 本条不改状态:释放动作留给解锁扫描,因为条件现在够得着机械判据了。以下是理由。

裁决已下:批 #160 第 2 项 · 字母 A

维护者 2026-09-18T11:58Z 裁定 A —— 规格为准:extractRagflowOptions 改读 source.adapterConfig(被声明的那个键),读 options 的那个 cast 撤掉(Prime Directive #12:消费侧不得宽容)。⛔ B 被否(册封一个只因无人 parse 才能用的拼法)· ⛔ C 被否(留着 declared ≠ enforced)。

⭐ 裁决把本卡的余项吸收进了 #18973 的执行面

裁决逐字:「The README example moves to adapterConfig and compiles against the package.」

⇒ 本卡唯一未落的那一块 —— packages/plugins/knowledge-ragflow/README.md —— 由 #18973 的 PR 交付,⛔ 不再需要 domain:devx 另开一个 PR。⚠️ 裁决还写明车道会重新落到 domain:services(diff 落在 packages/plugins/knowledge-ragflow)。

⇒ ⭐ 本卡的余项不再由本席交付。 下一个读到本卡的人:⛔ 不要为那一块另立 PR,它已经有主。

⚠️ H26 的形状解除了,这一点值得单独记

本席落地时按 H26 记过:Blocked-by: #18973 当时指向一张 needs-user-decision 卡,而那类卡被回答⛔不产生关闭事件 ⇒ 纯 Blocked-by: 是个永不自动解锁的块,所以写了 Release-mechanism: 人工。

现在 #18973 已转 pm:queue —— 它是一张会被执行、会 CLOSE 的卡 ⇒ 解锁谓词(「Blocked-by: 目标已关闭」)从此够得着它。

• Release-fires-on: @objectstack/knowledge-ragflow must read source.adapterConfig.datasetId — the declared key (ruled: batch #160 item 2, letter A; README moves with it) #18973 关闭(其 PR 落地,同一笔带走本卡余下那一块 README)
• Release-mechanism: 解锁扫描(⛔ 不再是人工)

⇒ 本席⛔ 现在不放行:#18973 尚未关闭,余项尚未交付。提前放行会让本卡回到队列而实际无活可做 —— 那正是 pm:queue 语义要避免的。

顺带:本席已把 #18973 的标题前缀 [decision] 摘掉

H54 在本轮巡查里点名它:标题写着 [Decision] 而标签已是 pm:queue,两个载体互相矛盾,而 H54 测量过那个代价 ——「每个席位都要读完正文和评论才能确认它能不能被拿起来,而下一个席位再付一次,因为没人把读数写回去」。标题是本席立卡时写的,所以由本席改,⛔ 不代 domain:spec 动任何标签。

_{读数时刻 2026-09-18T12:27Z}

---

Generated by Claude Code