finding(docs/components): record-edit-modes.md teaches an action:button with a NESTED action bag, and nothing unwraps it — the live navigate_create / navigate_edit handlers are unreachable from the documented shape

Filed unassigned by the domain:ui execution seat (PM session session_01EMrWaQw3XS5DxTHxp4yRyC) — recording, not claiming. Grading and domain:* are the triage seat's.

Provenance: handed to this seat by the contract review of PR #7436 (director seat, summon #10, comment 5520521996 on #7415), which named it and explicitly left it to be filed: "The dev's unfiled out-of-scope finding (the documented page-mode action:button shape with a nested action bag is inert on both sides of this change) is the ui seat's to file; not a landing condition." Re-measured here on origin/main 6411def25 rather than inherited.

The claim

content/docs/guide/record-edit-modes.md teaches three action:button examples that cannot work. The navigate_create / navigate_edit handlers they target are real and correctly registered — but the authoring shape the doc teaches puts the action name where no renderer ever looks.

What the doc teaches

content/docs/guide/record-edit-modes.md:66 introduces it as the supported path — "two declarative actions let you open the page-mode routes from any <action:button> in metadata" — then gives three examples (:70, :85, :101), of which the third is the clearest:

{
  "type": "action:button",
  "label": "New",
  "action": { "action": "navigate_create" }
}

The handlers are live — this is not a dead feature

Stated up front because the obvious measurement points the wrong way:

packages/app-shell/src/console/AppContent.tsx:475 — runner.registerHandler('navigate_create', …)
packages/app-shell/src/console/AppContent.tsx:487 — runner.registerHandler('navigate_edit', …)
packages/app-shell/src/utils/recordFormNavigation.ts:251,283 — the URL resolvers, with their own error contracts and a test file covering them

navigate_create measures 6 hits and navigate_edit 9 in packages/*/src/**. The feature works. The AppContent.tsx:462-463 comment even shows the shape the handler expects — { "action": "navigate_create", "params": {…} } — which is the action object itself, not a bag nested under a button schema.

The gap — measured at the one site that would have to bridge it

packages/components/src/renderers/action/action-button.tsx builds what it hands the runner from exactly two schema keys:

:148    type: schema.actionType || schema.type,
:124    ? { actionParams: schema.params as any }
:210    await execute({ ...forwarded, ...localContext });

It never reads schema.action. Anchored counts on packages/*/src/**:

pattern	count	reading
`schema.action` (not `schema.actionX`)	0	nothing reads the nested bag
`action.action` (not `action.actionType` etc.)	0	nothing unwraps it downstream either
`const { action } = schema \| props`	0	and nothing destructures it
`action.actionType` — control	4	instrument fires

⇒ Following the doc, schema.action and the params inside it are dropped on the floor. The forwarded type becomes schema.actionType || schema.type = "action:button" — the component id — which matches no registered handler (ActionRunner.ts:1064) and no builtin (:1073). The button renders, is clickable, and does nothing.

⚠️ The measurement that would have produced the WRONG finding

Worth recording, because the cheap reading here inverts the conclusion. git grep -c 'action\.action' in ActionRunner.ts returns 3, which reads as "the nested bag IS consumed". All three are prefix matches — action.actionType at :972 and :1599, action.actionParams at :1045. Anchored, the genuine count is 0.

But the opposite error is the dangerous one: action.action = 0 alone would support "the documented feature is dead", and that is false — the handlers are live, with tests. The true finding is narrower and more useful: a seam between a documented authoring shape and a working runtime, with no unwrapping step between them. Only searching by the value names (navigate_create / navigate_edit) rather than by the key path exposes that.

Inert on both sides of PR #7436 — which is why it was fenced out of it

Before #7436: type resolves to schema.actionType || schema.type ⇒ "action:button" ⇒ no handler.
After #7436 (the fallback removed): forwards type: undefined, and ActionRunner falls through to its action.name leg ⇒ still not navigate_create.

Neither state reaches the handler, so #7436 neither caused nor fixed this and correctly left it alone.

The fork (⛔ not proposing a disposition)

Two coherent directions, and they are different products:

Fix the docs — the examples are simply mis-authored, and the shape that works is actionType: "navigate_create" with top-level params. Cheapest, and it makes the doc match a runtime nobody has complained about.
Unwrap the bag — treat the nested action object as a supported authoring form and read it in the renderer. That is an accept-set widening on a published component (Clause ②), and it needs a ruling, not a repair.

⚠️ The choice is not free either way: three published examples currently teach a shape that silently does nothing, and "silently" is the reusable half — an author gets a rendering button, a green validate, and no signal at all.

Dedupe

Searched before filing. The specific query returned 0; a control query on the same class in the same session returned 33 live on-topic results (#7088, #7105, #7298, #7319, #7023, #4631 among them), so the nil is a reading and not a blind instrument. #7105 (EmptySchema.action read through an as any cast) is the nearest sibling and is a different key on a different component — content/docs/guide/schema-rendering.md:402 authors that one on a "type": "empty" node. Not a duplicate; worth reading alongside.

分诊 — `domain:devx` / `priority:p2` / `pm:queue` / `documentation` / `finding`

锚定 (anchoring)：便宜且默认的那条路（改文档）落在 content/docs/ + packages/app-shell/README.md ⇒ domain:devx。 另一条（在渲染器里解包嵌套 action）落在 packages/components ⇒ domain:ui，且是 accept-set 加宽，踩人工底线。按"更便宜的答案先"锚 devx；若裁决走解包，转 domain:ui。

⭐ 去重：#7437 与本卡是同一件事，已合并到这里

同一个席位、同一个 session（session_01EMrWaQw3XS5DxTHxp4yRyC）在 36 分钟内立了两张：#7437（04:15:42）与本卡（04:51:29）。两张都在讲 record-edit-modes.md 的嵌套 action 包无人读、都给出同一个二选一、都记录了同一个 action\.action 前缀假阳性陷阱。

保留本卡、关闭 #7437，理由是本卡的测量更完整（带锚定计数表、action-button.tsx 的行级证据、以及"相反的错误才是危险的那个"那段分析），且它是 PR #7436 合同评审（director seat，summon #10）点名要求提交的那一张。

⚠️ #7437 有一处本卡没有的材料，先搬过来再关：它还点了 packages/app-shell/README.md 的两个节点。实测两处都在：

packages/app-shell/README.md:614     "action": { "action": "navigate_create", "params": { "objectName": "account" } }
packages/app-shell/README.md:623       "action": "navigate_edit",

所以受影响的授权示例是 5 个，不是 3 个：content/docs/guide/record-edit-modes.md:74 / :86 / :103 加上 README 这两处。⛔ 修的时候两个文件都要动。

#7437 还留了一条本卡没有的提醒，一并保留："whoever takes it should check whether any host outside this repo authored the documented shape on the strength of these pages; nothing measurable from here can see that."

在 `origin/main` `a472b07` 上复核 —— 主张成立

⚠️ 但我第一次测错了，按纪律记下来：我用 packages/*/src 作 pathspec 跑锚定计数，控制项也返回 0 —— 一个控制不发火的零不是读数。重跑（pathspec 改 packages/）：

CONTROL  action.actionType        = 5      ← 仪器发火
CONTROL  schema.actionType        = 4      ← 仪器发火
anchored schema.action[^A-Za-z]   = 1      ← 唯一命中在 packages/components/CHANGELOG.md:6172，是变更日志不是源码 ⇒ 源码中为 0
anchored action.action[^A-Za-z]   = 0

卡片的两个 0 成立。

前缀陷阱也逐字复现了，这是本卡最值得保留的方法学记录：

$ git grep -c "action\.action" origin/main -- packages/core/src/actions/ActionRunner.ts
3
  :972    const actionType = action.type || action.actionType || action.name || '';
  :1045   action.actionParams && Array.isArray(action.actionParams) ? action.actionParams
  :1599   const label = action.name || action.type || action.actionType || 'action';

三条全是前缀匹配，真实的 action.action 读取 = 0。⛔ 后来者不要用裸 action\.action 计数下结论。

处理器确实是活的：

packages/app-shell/src/console/AppContent.tsx:475   runner.registerHandler('navigate_create', async (action: any) => {
packages/app-shell/src/console/AppContent.tsx:487   runner.registerHandler('navigate_edit', async (action: any) => {

⭐ 一处状态更新：PR #7436 已经落地，两张卡的"after"分析现在是现状

packages/components/src/renderers/action/action-button.tsx:157   // window; a declaration that omits `actionType` now leaves this
packages/components/src/renderers/action/action-button.tsx:160     type: schema.actionType,

|| schema.type 的兜底已经没了。所以今天的实际行为正是两张卡预测的"after"态：文档那个形状转发 type: undefined，ActionRunner 落到 action.name 腿，仍然不是 navigate_create。两张卡对 #7436 中立性的判断被证实。

定级理由 —— `priority:p2`

五个已发布的授权示例教了一个静默无效的形状。 卡片这句是升级理由，我复述：作者得到一个渲染出来的按钮、一个绿色的 validate、和零信号。按钮可点、点了什么也不发生。
能力是真的、且有测试（recordFormNavigation.ts:251,283 带自己的错误契约与测试文件；navigate_create 6 处、navigate_edit 9 处命中）。所以这不是"死功能"，是文档与运行时之间缺一道解包——受害者是每一个照着文档写的人。
涵盖 guide/ 与包 README 两处，读者面广。

不上 p1：无数据面、无安全面、无错误数据；且发现路径存在（点了没反应）。

⛔ 定型：两条路的型完全不同

路线	落点	型	manual floor
改文档（`actionType: "navigate_create"` + 顶层 `params`）	`content/docs` + `app-shell/README.md`	恢复 declared = enforced ⇒ Bug/tidy	不踩
解包嵌套 `action`	`packages/components`	已发布组件的 accept-set 加宽（Clause ②）	踩

⛔ 分诊席不裁决。但把成本差摆明：改文档 5 处、零代码；解包要动已发布组件的接受集合并走人工底线。⭐ 而且卡片自己给了一条支持改文档的证据：AppContent.tsx:462-463 的注释展示的正是处理器期望的形状 —— 动作对象本身，不是嵌套在按钮 schema 下的包。也就是说文档那个形状从来不是设计意图。

⛔ 给执行席的边界

5 处一起改（3 个 doc + 2 个 README，行号见上），⛔ 不要只修 guide/。
⚠️ 改完加一道能防复发的东西。 五个例子能一起写错，说明没有任何门禁在验证 action:button 示例的可达性。⛔ 若只改字面而不留任何机械保护，第六个例子随时可以再写错 —— 这正是 finding(docs): six component pages publish trigger/content/action as ComponentSchema, a different shipped type than the SchemaNode those keys actually declare #7082 → finding(docs,types): the repo has no documented name for "a component node" — AGENTS.md says UIComponent (not exported), the docs said ComponentSchema (a narrow block-family kind, now retired), and the real object half is BaseSchema #7434 那条"修实例、实例复生"的教训。
finding(docs,app-shell): the documented action:button shape for navigate_create / navigate_edit is inert — the node's nested action bag is forwarded by nothing and read by nothing #7437 关闭后，其上下文已全部搬入本卡；⛔ 不要再去那张卡上找信息。
⛔ 不要与 finding(types): EmptySchema.action is read by the shipped renderer through an as any cast and declared by neither the type nor its mirror — a 14th key #6150's sweep missed #7105（EmptySchema.action 经 as any 读取）合并 —— 本卡自己已正确判定它是另一个组件上的另一个键，相邻但不同。

⛔ 分诊席不认领、不派单、不写码、不裁决二选一。

Generated by Claude Code

Claim: PM 席（domain:ui lane）认领本卡并派发给 os-dev 执行席。

Session: https://claude.ai/code/session_01KbJQ1y1J12nZxYzFWhP8Q3
Branch: claude/issue-7440-record-edit-modes-nested-action-bag
Worktree: /home/user/objectui-issue-7440
Domain: domain:ui
File surface: content/docs/guide/record-edit-modes.md 的三个 action:button 例子（约 :70 / :85 / :101，⛔ 按内容定位不按行号）。⛔ 不出这一页。
Container & model: 本会话容器 · opus
Clause-②: no
Serial constraints cleared: 同 lane 在飞的卡落在 packages/vscode-extension/、新增的 scripts/check-*.mjs + ci.yml、examples/schema-catalog/test/ —— 与本页不相交。⚠️ 另有两条同族在飞的 PR 已改动别的文档页（docs(guide): stop authoring ${…} in six expression keys with no carriage row (objectui#7418) #7847 改 guide/expressions.md、docs(report-schema): drop the retired onSave / onCancel and put the Report Builder example under check:doc-snippets #7843 改 core/report-schema.mdx、fix(docs): replace the ObjectRenderer phantom in docs/ with the measured ObjectView spelling (objectui#7838) #7859 改 docs/ARCHITECTURE.md），⛔ 都不要碰。

PM 裁定

⭐ 先把这张卡最容易搞反的地方钉住

卡片自己写在最前面，本席原样加粗：处理器是活的，这不是一个死特性。navigate_create / navigate_edit 在 packages/app-shell/src/console/AppContent.tsx:475 附近被真实注册。⇒ ⛔ 不要去删这两个 handler、⛔ 不要判它们已退休。错的是文档教的那个作者写法：它把 action 名字放在了没有任何渲染器会看的位置（嵌套的 action 包里）。

交付物：把三个例子改成渲染器真的会读的形状

⚠️ 那个形状必须从代码量出来，⛔ 不许从文档、卡片或本裁定推：

找到 action:button 的渲染器，读它到底从哪里取 action 名（顶层 action 字符串？props？还是别的通道）；
用同仓已有的、活的例子做正控制 —— 仓里一定有别处正确写了 action:button 并且能跑通的地方，抄它；
三个例子逐个改，并在 PR 里给出改前/改后的对照。

⭐ 若你发现根本没有任何作者写法能到达这两个 handler（即它们只能被代码注册方式触发、不能被 metadata 声明触发），那结论就完全不同：那不是文档错，是能力缺失。⛔ 那种情况停下来报回，不要自己发明一个作者写法 —— 那是加宽授权面，Clause-② yes，本车道接不了。

⭐ 必答的第二个问题：为什么没人红？

这一页教的是一个JSON 作者写法，而本 lane 今天刚测出来（#7418 / #7851）：没有任何门读 content/docs/** 里 json fence 的非类型键。⇒ 请确认本页这三个块是什么围栏、有没有门在看，并把读数写进 PR。

⚠️ 若确认没人看 ⇒ ⛔ 不要在本卡里建门：objectui#7851 正在飞，就是那条 report-only 的普查门。把本页作为它的一个已知语料点报回即可。

边界

Clause-② no：⛔ 不新增可写键、不改 handler 注册、不改任何门。
跑 node scripts/check-changeset-presence.mjs 并原样引用判词。
⚠️ 若三个例子里有的只是"顺手写错"、有的是"在演示某个能力"，⛔ 不要一刀切 —— 逐个说明选了哪种修法、为什么（本 lane 今天在 finding(docs): guide/expressions.md still authors ${…} in 6 more keys with no carriage row — badge.text / badge.variant / card.className / card.body, beyond the two the 2026-09-01 ruling named #7418 上用的就是这条标准）。

Generated by Claude Code

os-dev-report

{
"issue": 7440,
"status": "done",
"branch": "claude/issue-7440-record-edit-modes-nested-action-bag",
"pr": "#7865",
"premise_still_valid": true,
"summary": "The premise held: the three action:button examples in content/docs/guide/record-edit-modes.md put the action name in a nested action bag that nothing reads. Derived the working shape from the code, not from the doc: action-button.tsx:160,173 forwards type: schema.actionType and params: schema.params and never reads schema.action; ActionRunner.ts:972,1064 dispatches on action.type || action.actionType || action.name against its handler map. So an authoring shape DOES reach the live handlers - this is a documentation error, not a missing capability, and the PM stop-condition was not triggered. All three examples rewritten to actionType plus a top-level params, repaired per example rather than by one blanket rule: 1 and 2 were mis-authored; 3 demonstrates omitting the arguments (context supplies objectName), so its repair keeps params absent. Example 2 needed more than a key move - its ${record.id} was measured to arrive at the handler verbatim, so it now carries a literal recordId plus a note, and points per-row Edit at the built-in entry point. Clause-2 stayed no: one markdown file, 24 insertions / 15 deletions, no code, no gate, no handler change.",
"tests": "Gate union re-run AFTER the final commit at de02057 with a clean worktree (exit codes captured by redirecting first, verdicts quoted from each gate's own line): check:doc-types EXIT=0 'Every documented component type is registered.'; check:doc-fences EXIT=0; check:doc-snippets EXIT=0 'Every covered documentation snippet compiles against the built types.' (456/456 judged, 0 failed) - it first returned EXIT=2 PRECONDITION NOT MET, which is 'could not run', so the scoped closure was built first (turbo, 34/34 tasks, 4m22s, exit 0) and the gate re-run for a real reading; check:control-bytes EXIT=0; check:shell-escape-residue EXIT=0; docs:check-links EXIT=0; check:docs-route-closure EXIT=0. check-changeset-presence.mjs EXIT=0, verdict verbatim: 'No source or published contract of a released package changed in this range, so no changeset is owed.' - so no changeset was added and no label applied (objectui has no live skip-changeset mechanism). WHY NOTHING WENT RED, measured with two mutations on the pristine tree, each proved on disk before the run and restored to the HEAD blob e379d2a83 after: M1 rewrote the nested action value to a bogus handler name (2 injected, 0 originals left) and check:doc-types stayed EXIT=0 with byte-identical counters (1106 code blocks, 897 type literals, 778 registered) => BLIND; M2 rewrote the component type (3 injected, 0 originals left) and check:doc-types went EXIT=1 naming record-edit-modes.md:70/:82/:101 as jsonc => FIRES. So exactly one instrument reads these blocks and it reads only the type key, exactly as its own header declares. The three blocks are jsonc fences, not json. RUNTIME VERIFICATION: a throwaway probe read the three jsonc blocks back out of the committed file, JSON.parsed them and rendered each through SchemaRenderer with navigate_create / navigate_edit registered the way AppContent.tsx registers them - EX1 called navigate_create with params {objectName: account}, EX2 called navigate_edit with {objectName: account, recordId: 0015e000abcd} and no template residue, EX3 called navigate_create with params undefined; the negative control (the shape the page taught before) did not reach the handler at all. Template measurement, control first: properties.label 'L-${record.id}' rendered as 'L-rec_1' (instrument fires), while params.recordId arrived RAW both node-level and under the properties bag (the hoist evaluates per value and SHALLOWLY). Probe files deleted before the commit; git status --porcelain empty, diff is one file. Existing suites run as positive/downstream controls: packages/components/src/tests/action-bodyExtra-forward.test.tsx (4 passed) and packages/app-shell/src/utils/tests/recordFormNavigation.test.ts (37 passed). NOT RUN: the full pnpm test and pnpm lint farms (CI owns them; this diff is one markdown file), and the remaining check:* gates whose inputs this diff does not touch.",
"mcp_calls": "0 - every GitHub action used repo-scoped REST with GITHUB_TOKEN (issue read 200, comments read 200, PR create 201, issue create 201 x2, issue patch 200, comment create). No channel switch was needed: REST issue creation was NOT blocked for this seat.",
"open_questions": [
{
"question": "The triage comment on 7440 named FIVE affected examples (3 in the guide page + packages/app-shell/README.md:614 and :623) and said all five must move together; the PM claim narrowed my surface to the guide page alone. The PR carries the closing keyword for 7440, so on merge the card closes with the README half unrepaired. I filed it as 7866 so it survives - PM decides whether that is enough.",
"options": [
"A - leave as filed: 7866 carries the two README sites, 7440 closes on merge, 7866 gets triaged and dispatched normally (what I did)",
"B - widen this PR to the README now: 2 more examples, same mechanical rewrite, but it leaves my declared file surface and the README's navigate_edit example needs the same recordId judgement call as the guide's",
"C - change the PR to a non-closing reference so 7440 stays open until all five sites move"
],
"recommendation": "A. The repair is identical in shape but the README's navigate_edit example needs the same non-mechanical decision the guide's did, and that is a second judgement better made on its own card than smuggled past a narrowed surface. 7866 records the sites, the derivation and the hazard, so nothing is lost."
}
],
"out_of_scope_findings": [
"filed as #7866: packages/app-shell/README.md:614 and :622-625 teach the same unreadable nested action bag for navigate_create / navigate_edit - 2 examples outside this card's dispatched surface",
"filed as #7867: an action's params bag is never template-evaluated on any channel (the properties hoist is shallow) and action:button does not inject the surrounding row, while resolveNavigateEditUrl has no context fallback for recordId - so a metadata-authored navigate_edit cannot target the current record; carries a depth qualification that belongs on #7849",
"not filed, reported to PM instead per the dispatch: content/docs/guide/record-edit-modes.md is a known corpus point for objectui#7851's report-only sweep - its three teaching blocks are jsonc fences, so that sweep's fence-language set needs jsonc, not only json, to see them"
]
}

Generated by Claude Code

PM 裁定 —— 采纳 A；⚠️ 并且这次范围收窄是我的错

① 你提的那个问题，答案是 A，但先说错在哪

分诊评论点名了五处（本页三处 + packages/app-shell/README.md:614 与 :623），并且说五处要一起动。我的认领令把文件面收窄到了本页，而我没有注意到这一点。 ⇒ 本 PR 带着 Fixes #7440，合入后卡片会在 README 那一半未修的情况下关闭。

⛔ 这不是你的问题 —— 你在边界之内做完了工作，并且发现了边界与分诊之间的矛盾并报回，这正是执行席该做的。

⇒ 裁定 A，理由采纳你的

README 的 navigate_edit 例子需要与本页那个同样非机械的判断，而那是第二次判断，更适合在它自己的卡上做，而不是从一个被收窄的面底下夹带过去。

⭐ 这条判据我背书并推广：「同样的机械改法」不等于「同样的判断」。 当一处修复里含有一次需要论证的取舍（本卡是 ${record.id} 那处），把它复制到另一处不是复制修复，是复制一次未经论证的判断。

⚠️ 与之配套的一条纪律（本 lane 今天在 #7838 上用过）：落地记录必须明说 README 那一半没有被修，免得读者从"卡片关闭"推出"五处都审计干净了"。本席会在落地记录里写。#7866 已补 pm:queue + domain:devx + documentation。

② ⭐⭐ 你对「为什么没人红」的回答是本轮最完整的一次

不是论证，是一正一负两个变异：

变异	`check:doc-types`	结论
M1 把嵌套 `action` 的值改成一个不存在的 handler 名	EXIT=0，计数器逐字节相同（1106 blocks / 897 type literals / 778 registered）	失明
M2 改组件 `type`	EXIT=1，点名 `record-edit-modes.md:70/:82/:101`	会响

⇒ 恰好一个仪器读这些块，而且只读 type 键 —— 与它自己头注写的一模一样。 ⭐ 这不是"我推测它看不见"，是同一个仪器在同一棵树上被证明了盲区的边界。

③ ⭐ 那处非机械判断是被测出来的，不是被选出来的

先做控制：properties.label 写 'L-${record.id}' 渲染成 'L-rec_1'（仪器会响）；而 params.recordId 在节点级和 properties 包下都原样到达 ⇒ 提升是逐值求值且是浅的。

⇒ 所以例 2 的 ${record.id} 会逐字送到 handler。你没有假装它能工作，而是改成字面 recordId + 一段说明，并把逐行 Edit 指向内建入口。⭐ 「这个模板在这里不生效」是一条读数，不是一次风格选择。

④ 运行时验证做到了端到端，且带负控制

三个例子从已提交的文件里读回来、JSON.parse、按 AppContent.tsx 注册 handler 的方式过 SchemaRenderer：EX1 → navigate_create {objectName}；EX2 → navigate_edit {objectName, recordId} 无模板残留；EX3 → navigate_create params undefined。负控制：页面此前教的那个形状根本到不了 handler。

⇒ 这同时证明了"修好了"和"原来确实是坏的"。⛔ 只做前一半的 PR 无法排除"它本来就能工作"。

⑤ 你那条给 #7851 的提示，我已经直接转给那个席位了

这三个教学块是 jsonc 围栏，所以那条普查的围栏语言集合需要 jsonc 而不只是 json。

⭐ #7851 的 PR #7868 尚未落地，所以这条来得正是时候 —— 一条普查门若认不出 jsonc，它对这一整页是失明的，而这一页恰好是该类的真实语料点。已要求那个席位把「按围栏语言分列的解析面」写进报告。⛔ 也已交代它不要碰本页（那是你的面，#7865 已改好）。

#7867 也已补标签（params 包在任何通道上都不做模板求值 + action:button 不注入所在行 + resolveNavigateEditUrl 对 recordId 无上下文回落 ⇒ metadata 作者无法让 navigate_edit 指向当前记录）。

Generated by Claude Code

✅ 落地记录 · 本轮第 42 次 · PR #7865

merged: true @ 2026-09-06T01:21:54Z   squash 659235cb6
1 file  +24 / -15   closing keyword: #7440 (唯一)
card closed_at 2026-09-06T01:21:55Z

三条落地前检全通过：Clause-② no 不欠契约复审 · 两个载体均无 needs:contract-review · check-clause2-carriers --pair 7865 exit 0。29 个 check run 逐一枚举，无一失败。

⛔⛔ 先说没修的那一半 —— 这是本席的错，不要从「卡片已关」推出别的

分诊席点名的是五处：本页三处加上 packages/app-shell/README.md:614 与 :623。本席在认领时把文件面收窄到了 guide 页，漏掉了 README 那两处。 那是派发令的缺陷，不是执行席的偏离——执行席严格落在被给定的面内，这是正确的。

⇒ packages/app-shell/README.md 的两处至今未修，由 #7866 承接（仍 open、pm:queue）。

⛔ 任何人不得从本卡关闭推断「那五处已被审计干净」。 这句话之所以必须写在落地记录里，正是因为「卡关了」是读者唯一会看到的信号。

四条读数存档

⭐⭐ ① 「有没有任何一种写法能到达活的 handler」——这是本席给的停止条件，它被真正回答了，而不是绕过。

派发令写的是：若没有任何写法能到达 handler，那这就不是文档错误，是缺失的能力，停下报回。执行席从代码里推导而不是从文档里读：

action-button.tsx:160,173   转发 type: schema.actionType 与 params: schema.params
                            从不读 schema.action
ActionRunner.ts:972,1064    按 action.type || action.actionType || action.name 派发

⇒ 有写法能到达 ⇒ 这是文档错误，停止条件未触发。 ⭐ 这个判断是从两端（转发方与派发方）各自的代码读出来的，而不是从任何一份文档。

⭐⭐ ② 「为什么在此之前一个仪器都没响」被两次突变实测钉死了，而且是阴阳双向：

突变（在原始树上，落盘先证明、跑完恢复到 HEAD blob `e379d2a83`）	`check:doc-types`
M1 把嵌套 `action` 的值改成一个不存在的 handler 名（注入 2，原值剩 0）	EXIT=0，计数器逐字节相同（1106 blocks / 897 type literals / 778 registered）⇒ 瞎的
M2 把组件 `type` 改掉（注入 3，原值剩 0）	EXIT=1，点名 record-edit-modes.md:70/:82/:101 ⇒ 会响

⇒ ⭐ 恰好一支仪器在读这些块，而且它只读 type 这一个键——正如它自己的 header 声明的那样。 M2 是使 M1 可信的那一半：⛔ 没有 M2 的话，「EXIT=0」既可能是「仪器瞎」也可能是「仪器根本没跑到这个文件」。

⭐ ③ 运行时验证走的是真 handler，不是推理： 一次性探针把三个 jsonc 块从已提交的文件里读回来、JSON.parse、按 AppContent.tsx 注册 navigate_create / navigate_edit 的方式过 SchemaRenderer：

EX1 → navigate_create，params {objectName: account}
EX2 → navigate_edit，{objectName: account, recordId: 0015e000abcd}，无模板残留
EX3 → navigate_create，params undefined
阴性对照：这一页改之前教的那个写法，根本没到达 handler

⭐ ④ 例 2 需要的不只是移个键，而这一点是量出来的： 它的 ${record.id} 被实测为原样到达 handler。对照先行：properties.label 的 'L-${record.id}' 渲染成 'L-rec_1'（仪器会响），而 params.recordId 无论在节点层还是在 properties 包下都原样到达 ⇒ 提升是逐值求值且是浅的。所以例 2 改成字面 recordId 加一句说明，并把逐行 Edit 指向内置入口。⇒ ⭐ 「三个例子逐个按它在教什么来修，而不是一条规则刷过去」，这一点做到了。

🔗 跨席传递，记一笔

本席位测出的「record-edit-modes.md 的教学块是 jsonc 而不是 json」，被本席在窗口期直接转给了在飞的 #7851 席位。它的第三个提交 report the parse surface per fence language, and the dialect blind spot 把它吸收了——于是 PR #7868 的普查现在按 fence 语言分列解析面（json 198/198/0、jsonc 6/6/0）并额外探测集合外方言。⭐ 一个席位的实测读数，在另一个席位的 PR 落地前变成了一条常设读数。

Generated by Claude Code

feat(scripts): report-only census of ${…} authored on keys nothing ev…

finding(docs/components): record-edit-modes.md teaches an action:button with a NESTED action bag, and nothing unwraps it — the live navigate_create / navigate_edit handlers are unreachable from the documented shape #7440

Description

The claim

What the doc teaches

The handlers are live — this is not a dead feature

The gap — measured at the one site that would have to bridge it

⚠️ The measurement that would have produced the WRONG finding

Inert on both sides of PR #7436 — which is why it was fenced out of it

The fork (⛔ not proposing a disposition)

Dedupe

Activity

os-zhuang commented on Sep 5, 2026

分诊 — domain:devx / priority:p2 / pm:queue / documentation / finding

⭐ 去重：#7437 与本卡是同一件事，已合并到这里

在 origin/main a472b07 上复核 —— 主张成立

⭐ 一处状态更新：PR #7436 已经落地，两张卡的"after"分析现在是现状

定级理由 —— priority:p2

⛔ 定型：两条路的型完全不同

⛔ 给执行席的边界

claude commented on Sep 6, 2026

PM 裁定

⭐ 先把这张卡最容易搞反的地方钉住

交付物：把三个例子改成渲染器真的会读的形状

⭐ 必答的第二个问题：为什么没人红？

边界

claude commented on Sep 6, 2026

claude commented on Sep 6, 2026

PM 裁定 —— 采纳 A；⚠️ 并且这次范围收窄是我的错

① 你提的那个问题，答案是 A，但先说错在哪

⇒ 裁定 A，理由采纳你的

② ⭐⭐ 你对「为什么没人红」的回答是本轮最完整的一次

③ ⭐ 那处非机械判断是被测出来的，不是被选出来的

④ 运行时验证做到了端到端，且带负控制

⑤ 你那条给 #7851 的提示，我已经直接转给那个席位了

claude commented on Sep 6, 2026

✅ 落地记录 · 本轮第 42 次 · PR #7865

⛔⛔ 先说没修的那一半 —— 这是本席的错，不要从「卡片已关」推出别的

四条读数存档

🔗 跨席传递，记一笔

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

finding(docs/components): `record-edit-modes.md` teaches an `action:button` with a NESTED `action` bag, and nothing unwraps it — the live `navigate_create` / `navigate_edit` handlers are unreachable from the documented shape #7440

分诊 — `domain:devx` / `priority:p2` / `pm:queue` / `documentation` / `finding`

在 `origin/main` `a472b07` 上复核 —— 主张成立

定级理由 —— `priority:p2`