@@ -0,0 +1,10 @@
+---
+'@object-ui/i18n': patch
+---
+
+feat(i18n): `chatbot.tool.create_record` / `update_record` / `remove_sample_data` labels and the record-write card's `console.ai.recordWrite*` labels in all ten packs; `chatbot.tool.draft_record` retired
+
+Clause-②: no
+Metadata-protocol: none
+
+The spec registry now lists the cloud chat agents' confirmed record write and the sample-data cleanup (cloud#3565) and drops `draft_record`, so the chatbot's tool timeline gets a sentence for each new step in every pack ("Create the record" / "新建记录", "Change the record" / "修改记录", "Remove the sample data" / "清理示例数据") and loses the retired one, keeping `toolLabels-locale-parity-7481` and `toolLabels-sentences-12724` green. The console's record-write card gets its eleven labels under `console.ai.recordWrite*`: the create and update titles (with the object's label), the three buttons 确认 / 打开完整表单 / 取消, the two stored reasons, the "not included" lead-in and the three states (writing, written, cancelled) — plus the 「清理示例数据」 card's four: `console.ai.sampleDataCleanupTitle` (its title over the seeded tables and counts), `console.ai.sampleDataCleanupIrreversible` (「删除后无法恢复」 — the removal cannot be undone, said on every card), and the two count families `console.ai.sampleDataCleanupAbsent` (the recorded rows the card does not list) and `console.ai.sampleDataCleanupEdited` (the rows kept because someone edited them), each with every plural slot its language selects (objectui#11445).
@@ -0,0 +1,16 @@
+---
+'@object-ui/plugin-chatbot': patch
+---
+
+feat(plugin-chatbot): the record-write card — ONE record the chat proposes and the user confirms, with 确认 / 打开完整表单 / 取消 (cloud#3565)
+
+Clause-②: no
+Metadata-protocol: none
+
+The cloud chat's `create_record` / `update_record` tools queue a HITL pending action in the user's name and answer the approval envelope plus the record: `{ status:'pending_approval', pendingActionId, recordWrite:{ operation, objectName, objectLabel, recordId?, label?, reason?, fields:[{ name, label, type, value, current? }], dropped } }`. The chat now draws that as a card instead of the generic Approve / Deny strip and the raw payload:
+
+- **New export** `detectRecordWrite(result)` (with the `RecordWrite` and `RecordWriteField` types), lifted onto `ChatbotEnhancedToolInvocation.recordWrite` by the live mapper beside `pendingActionId`. It reads the bare object, the persisted `{ type:'text', value }` wrapper and a JSON string; an envelope without the pending id or without a field to confirm is not a card.
+- **The card** (`data-testid="record-write"`): the create or update title with the object's label, the record's label and reason, every field as a labelled row — an update shows the stored value struck through, then →, then the proposed one — the fields the server left out and why, and three buttons. 确认 calls `onToolApprove(toolCallId, true)` (the pending action's dispatcher writes as the user); 取消 calls it with `false` and `recordWriteCancelReason`; 打开完整表单 calls it with `false` and `recordWriteOpenFormReason` FIRST, then the new `onOpenRecordForm(write)` prop, so a record saved in the form is never written a second time from the card (disabled where the host passes no handler). After a decision the card says written (with the record's link through `onOpenRecord`, read from the dispatcher's `{ status:'written', objectName, recordId, label }` result), cancelled, or the server's refusal.
+- **New props**: `onOpenRecordForm`, `recordWriteCreateTitle(objectLabel)`, `recordWriteUpdateTitle(objectLabel)`, `recordWriteConfirmLabel`, `recordWriteOpenFormLabel`, `recordWriteCancelLabel`, `recordWriteCancelReason`, `recordWriteOpenFormReason`, `recordWriteDroppedLabel`, `recordWriteSubmittingLabel`, `recordWriteWrittenLabel`, `recordWriteCancelledLabel`, all with English defaults.
+- **「清理示例数据」** rides the same seam: `detectSampleDataCleanup(result)` (with the `SampleDataCleanup` type) lifts `{ status:'pending_approval', pendingActionId, sampleDataCleanup:{ tables:[{ objectName, label, count }], total, absent, edited } }` onto `sampleDataCleanup`, and the chat draws the tables and counts (`data-testid="sample-data-cleanup"`, title from the new `sampleDataCleanupTitleLabel` prop) above the generic Approve / Deny strip, with the raw payload hidden. The card always says the removal cannot be undone (`sampleDataCleanupIrreversibleLabel`, default "Removed rows cannot be recovered." — there is no recycle bin until objectstack#3146), says how many recorded rows it does not list when `absent` is above 0 (`sampleDataCleanupAbsentLabel(count)`) and how many stay because someone edited them when `edited` is above 0 (`sampleDataCleanupEditedLabel(count)`; an envelope without `edited` reads as 0).
+- **`ToolDecisionState`** (widened, additive): every state may carry `decision: 'approved' | 'rejected'`, and an approved `success` may carry `result`, the dispatcher's answer. `useHitlInChat` sets both; the generic strip keeps reading `message` alone.
@@ -0,0 +1,12 @@
+---
+'@objectstack/spec': patch
+---
+
+feat(spec): `PLATFORM_TOOLS_BY_PACKAGE['service-ai-studio']` lists `create_record`, `update_record` and `remove_sample_data`; `draft_record` is retired
+
+Clause-②: no
+Metadata-protocol: none
+
+The cloud chat agents now create or change ONE record of the user's data after the user confirms it on a card, and remove the sample rows a seed replay recorded after a card listing them (cloud#3565, the maintainer's ruling of 2026-11-04). The three tools replace `draft_record` (cloud#3383), whose "draft it and the user saves it in a form" ending was never built. The registry lists the three names in their alphabetical order and drops the retired one, so `PLATFORM_PROVIDED_TOOL_NAMES` carries them and a skill whose `tools[]` names them resolves under ADR-0109 reference integrity, while a skill still naming `draft_record` is reported as a fictional tool (`ai-skill-tool-unresolved`). That is the whole change: three accepted `skill.tools[]` values in, one out. No schema shape changes, no export is added or removed.
+
+**Reach of the retirement.** `draft_record` was a published constant, so removing it breaks any skill whose `tools[]` still names it. Measured before this landed: no skill in this repo names it (`validate-ai-tool-references.test.ts` names it only as the control proving the report fires; the two spec page fixtures called `draft_record` are page names, not tools); the one skill anywhere that names it is the cloud chat's `record_drafting` skill, which the cloud half of cloud#3565 (cloud#3574) moves to `create_record` / `update_record` in the same landing. A third-party skill that still names it reports `ai-skill-tool-unresolved` at validation and is not silently dropped. That is why this is graded `patch` and not left as a silent one.
@@ -226,18 +226,23 @@ describe('validate-ai-tool-references', () => {
   });
 
   /**
-   * The `ask` agent's new-record draft signal (cloud#3383, ui half #28104): a
-   * skill naming `draft_record` resolves; a near name nobody registers does not.
+   * The chat's confirmed record write (cloud#3565; it replaced cloud#3383's
+   * `draft_record`): a skill naming `create_record` / `update_record` resolves;
+   * a near name nobody registers does not, and neither does the retired name.
    */
-  it('resolves the service-ai-studio record-draft signal', () => {
-    expect(validateAiToolReferences({ skills: [{ name: 's', tools: ['draft_record'] }] })).toEqual([]);
+  it('resolves the service-ai-studio record-write pair', () => {
+    expect(
+      validateAiToolReferences({ skills: [{ name: 's', tools: ['create_record', 'update_record'] }] }),
+    ).toEqual([]);
 
     const control = validateAiToolReferences({
-      skills: [{ name: 's', tools: ['draft_record', 'draft_records'] }],
+      skills: [{ name: 's', tools: ['create_record', 'create_records', 'draft_record'] }],
     });
-    expect(control).toHaveLength(1);
+    expect(control).toHaveLength(2);
     expect(control[0].path).toBe('skills[0].tools[1]');
-    expect(control[0].message).toContain('draft_records');
+    expect(control[0].message).toContain('create_records');
+    expect(control[1].path).toBe('skills[0].tools[2]');
+    expect(control[1].message).toContain('draft_record');
   });
 
   it('reports stable paths and tolerates junk shapes', () => {
 
@@ -57,9 +57,12 @@ export const PLATFORM_TOOLS_BY_PACKAGE: Readonly<Record<string, readonly string[
   /**
    * `@objectstack/service-ai-studio` (cloud) — the shared schema-reader pair
    * (`describe_object`/`list_objects`), the `build` authoring product's
-   * metadata/blueprint/package/verify tools, and the `ask` agent's hand-off
-   * signals (`suggest_builder`, `open_record`, `draft_record`), which write
-   * nothing.
+   * metadata/blueprint/package/verify tools, the chat agents' hand-off
+   * signals (`suggest_builder`, `open_record`), which write nothing, and the
+   * two record-write proposals (`create_record`, `update_record`) and the
+   * sample-data cleanup (`remove_sample_data`), which write ONE record or the
+   * seeded rows only after the user confirms a card and as that user
+   * (cloud#3565; they replaced `draft_record`).
    */
   'service-ai-studio': [
     'add_field',
@@ -68,11 +71,11 @@ export const PLATFORM_TOOLS_BY_PACKAGE: Readonly<Record<string, readonly string[
     'create_metadata',
     'create_object',
     'create_package',
+    'create_record',
     'create_seed',
     'delete_field',
     'describe_metadata',
     'describe_object',
-    'draft_record',
     'get_active_package',
     'get_authoring_rules',
     'get_metadata_schema',
@@ -84,12 +87,14 @@ export const PLATFORM_TOOLS_BY_PACKAGE: Readonly<Record<string, readonly string[
     'modify_field',
     'open_record',
     'propose_blueprint',
+    'remove_sample_data',
     'set_active_package',
     'suggest_builder',
     'test_flow',
     'todo_write',
     'toggle_flow',
     'update_metadata',
+    'update_record',
     'validate_expression',
     'verify_build',
   ],
 
@@ -63,6 +63,8 @@ import {
   detectProposedPlan,
   detectBuilderHandoff,
   detectRecordHandoff,
+  detectRecordWrite,
+  detectSampleDataCleanup,
   detectProposedChanges,
   detectReplayOutcome,
   detectAuthoringVerdict,
@@ -143,6 +145,29 @@ import { isConversationZh } from './conversationLanguage.js';
 import { resolveOutboundAgentText, type OutboundAgentTextKey } from './outboundAgentText.js';
 import { resolveApiBase } from './resolveApiBase.js';
 import { AiUnavailable } from './AiUnavailable.js';
+import { recordFormPrefillQuery } from './recordFormPrefill.js';
+
+/**
+ * The app segment a record of `objectName` lives under: the object's owning
+ * package, read from `/api/v1/meta/object/<name>` (`_packageId`), with `setup`
+ * as the fallback shell for an object no package claims. Best-effort: a failed
+ * read still lands on the record. Shared by the record hand-off (cloud#1658)
+ * and the record-write card's form hand-off (cloud#3565).
+ */
+async function appOfObject(objectName: string): Promise<string> {
+  try {
+    const r = await fetch(`/api/v1/meta/object/${encodeURIComponent(objectName)}`, {
+      credentials: 'include',
+    });
+    if (r.ok) {
+      const j = (await r.json()) as { item?: { _packageId?: unknown } };
+      if (typeof j?.item?._packageId === 'string' && j.item._packageId) return j.item._packageId;
+    }
+  } catch {
+    /* resolution is best-effort; the caller's fallback still lands on the record */
+  }
+  return 'setup';
+}
 
 function partString(part: HydratedUIMessagePart, key: string): string | undefined {
   const value = part[key];
@@ -238,6 +263,10 @@ export function hydratedMessagesToChatMessages(messages: HydratedUIMessage[]): C
         // four; this is its hydration counterpart.
         const builderHandoff = detectBuilderHandoff(result);
         const recordHandoff = detectRecordHandoff(result);
+        // cloud#3565 — the record-write card rides the pending approval the
+        // id below is lifted from; it must survive a reload like the rest.
+        const recordWrite = detectRecordWrite(result);
+        const sampleDataCleanup = detectSampleDataCleanup(result);
         const proposedChanges = detectProposedChanges(result);
         // objectui#5695 — a confirm-replay result becomes a terminal state on
         // the original 确认修改 card; it must never rehydrate as an ordinary
@@ -285,6 +314,8 @@ export function hydratedMessagesToChatMessages(messages: HydratedUIMessage[]): C
           ...(proposedPlan ? { proposedPlan } : {}),
           ...(builderHandoff ? { builderHandoff } : {}),
           ...(recordHandoff ? { recordHandoff } : {}),
+          ...(recordWrite ? { recordWrite } : {}),
+          ...(sampleDataCleanup ? { sampleDataCleanup } : {}),
           ...(proposedChanges ? { proposedChanges } : {}),
           ...(replayOutcome ? { replayOutcome } : {}),
           ...(part.errorText ? { errorText: String(part.errorText) } : {}),
@@ -1706,25 +1737,40 @@ export function ChatPane({
   // shell for an object no package claims.
   const openRecord = useCallback(
     async (handoff: { objectName: string; recordId: string }) => {
-      let app = 'setup';
-      try {
-        const r = await fetch(`/api/v1/meta/object/${encodeURIComponent(handoff.objectName)}`, {
-          credentials: 'include',
-        });
-        if (r.ok) {
-          const j = (await r.json()) as { item?: { _packageId?: unknown } };
-          if (typeof j?.item?._packageId === 'string' && j.item._packageId) app = j.item._packageId;
-        }
-      } catch {
-        /* resolution is best-effort; the fallback below still lands on the record */
-      }
+      const app = await appOfObject(handoff.objectName);
       navigate(
         `/apps/${encodeURIComponent(app)}/${encodeURIComponent(handoff.objectName)}/record/${encodeURIComponent(handoff.recordId)}`,
       );
     },
     [navigate],
   );
 
+  // cloud#3565 — 打开完整表单 on a record-write card. A create opens the New
+  // form with the card's values through the query-string prefill
+  // `RecordFormPage` already reads for related-list "+ New" (one `field=value`
+  // pair per card field; the form's own defaults fill the rest). An update
+  // opens the record's edit form: the card has shown old → new, and the form
+  // opens on the stored record. The card cancelled the pending write before
+  // calling this, so a record saved here is never written a second time.
+  const openRecordForm = useCallback(
+    async (write: {
+      operation: 'create' | 'update';
+      objectName: string;
+      recordId?: string;
+      fields: Array<{ name: string; value: unknown }>;
+    }) => {
+      const app = await appOfObject(write.objectName);
+      const base = `/apps/${encodeURIComponent(app)}/${encodeURIComponent(write.objectName)}`;
+      if (write.operation === 'update' && write.recordId) {
+        navigate(`${base}/record/${encodeURIComponent(write.recordId)}/edit`);
+        return;
+      }
+      const query = recordFormPrefillQuery(write.fields);
+      navigate(`${base}/new${query ? `?${query}` : ''}`);
+    },
+    [navigate],
+  );
+
   // objectui#12894 — 查看变更并发布 on the applied card of a change the copilot
   // saved as a draft (cloud#2924): open Studio's 变更 drawer for the package
   // through its one declared way in, `?changes=review` (`studioChangesLink.ts`).
@@ -2612,6 +2658,23 @@ export function ChatPane({
         builderHandoffOpenLabel={t('console.ai.builderHandoffOpen')}
         builderHandoffSupersededTitle={t('console.ai.builderHandoffSuperseded')}
         onOpenRecord={openRecord}
+        // cloud#3565 — the record-write card's three buttons and its states.
+        onOpenRecordForm={openRecordForm}
+        recordWriteCreateTitle={(objectLabel) => t('console.ai.recordWriteCreateTitle', { object: objectLabel })}
+        recordWriteUpdateTitle={(objectLabel) => t('console.ai.recordWriteUpdateTitle', { object: objectLabel })}
+        recordWriteConfirmLabel={t('console.ai.recordWriteConfirm')}
+        recordWriteOpenFormLabel={t('console.ai.recordWriteOpenForm')}
+        recordWriteCancelLabel={t('console.ai.recordWriteCancel')}
+        recordWriteCancelReason={t('console.ai.recordWriteCancelReason')}
+        recordWriteOpenFormReason={t('console.ai.recordWriteOpenFormReason')}
+        recordWriteDroppedLabel={t('console.ai.recordWriteDropped')}
+        recordWriteSubmittingLabel={t('console.ai.recordWriteSubmitting')}
+        recordWriteWrittenLabel={t('console.ai.recordWriteWritten')}
+        recordWriteCancelledLabel={t('console.ai.recordWriteCancelled')}
+        sampleDataCleanupTitleLabel={t('console.ai.sampleDataCleanupTitle')}
+        sampleDataCleanupIrreversibleLabel={t('console.ai.sampleDataCleanupIrreversible')}
+        sampleDataCleanupAbsentLabel={(count) => t('console.ai.sampleDataCleanupAbsent', { count })}
+        sampleDataCleanupEditedLabel={(count) => t('console.ai.sampleDataCleanupEdited', { count })}
         surface="plain"
         maxHeight="100%"
         headerSlot={headerSlot}
 
@@ -0,0 +1,88 @@
+// Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
+
+/**
+ * cloud#3565 — a record-write card survives a reload with everything the
+ * three buttons need: the pending id (for 确认 / 取消 through `useHitlInChat`),
+ * the `approval-requested` state (the render gate) and the record itself (what
+ * the card draws, and what 打开完整表单 hands the form). The server persists
+ * the proposal split across an assistant row and a tool row; this drives the
+ * real hydration path over those rows.
+ */
+
+import { describe, expect, it } from 'vitest';
+import { hydratedMessagesToChatMessages } from '../AiChatPage';
+import {
+  aiMessageRowsToServerMessages,
+  toUIMessages,
+} from '../../../hooks/useChatConversation';
+
+const PROPOSED = {
+  success: true,
+  status: 'pending_approval',
+  pendingActionId: 'pa_91',
+  recordWrite: {
+    operation: 'update',
+    objectName: 'crm_lead',
+    objectLabel: '线索',
+    recordId: 'lead_zhangwei',
+    label: '张伟',
+    fields: [{ name: 'status', label: '状态', type: 'select', value: 'contacted', current: 'new' }],
+    dropped: [],
+    fieldSecurity: 'record_projection',
+  },
+};
+
+function hydrateServerRows() {
+  return hydratedMessagesToChatMessages(
+    toUIMessages(
+      aiMessageRowsToServerMessages([
+        { id: 'u1', role: 'user', content: '把张伟的状态改成已联系' },
+        {
+          id: 'a1',
+          role: 'assistant',
+          content: '请在上方卡片确认。',
+          tool_calls: JSON.stringify([
+            {
+              type: 'tool-call',
+              toolCallId: 'c1',
+              toolName: 'update_record',
+              input: { objectName: 'crm_lead', recordId: 'lead_zhangwei', values: { status: '已联系' } },
+            },
+          ]),
+        },
+        {
+          id: 't1',
+          role: 'tool',
+          tool_call_id: 'c1',
+          content: JSON.stringify([
+            {
+              type: 'tool-result',
+              toolCallId: 'c1',
+              toolName: 'update_record',
+              output: { type: 'text', value: JSON.stringify(PROPOSED) },
+            },
+          ]),
+        },
+      ]),
+    ),
+  );
+}
+
+describe('a rehydrated record-write card is complete (cloud#3565)', () => {
+  it('hydrates to approval-requested with the pending id and the record', () => {
+    const messages = hydrateServerRows();
+    const tool = messages[1]?.toolInvocations?.[0];
+    expect(tool?.toolName).toBe('update_record');
+    expect(tool?.state).toBe('approval-requested');
+    expect(tool?.pendingActionId).toBe('pa_91');
+    expect(tool?.recordWrite).toEqual({
+      operation: 'update',
+      objectName: 'crm_lead',
+      objectLabel: '线索',
+      recordId: 'lead_zhangwei',
+      label: '张伟',
+      fields: [{ name: 'status', label: '状态', type: 'select', value: 'contacted', current: 'new' }],
+      dropped: [],
+    });
+  });
+});