From b03ce8d778c9acf5f22dd292944841e221ce40bf Mon Sep 17 00:00:00 2001
From: Dash <125997726+dashhuang@users.noreply.github.com>
Date: Sun, 25 Oct 2026 12:58:36 +1300
Subject: [PATCH 1/2] =?UTF-8?q?fix(companion):=20=E4=BF=9D=E7=95=99?=
 =?UTF-8?q?=E5=90=8E=E5=8F=B0=E5=9B=9E=E6=8A=A5=E8=A7=A6=E5=8F=91=E7=9A=84?=
 =?UTF-8?q?=E8=BF=9B=E5=BA=A6=E6=96=87=E5=AD=97=E5=B9=B6=E6=98=8E=E7=A1=AE?=
 =?UTF-8?q?=E8=BD=AE=E6=AC=A1=E7=8A=B6=E6=80=81?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Signed-off-by: Dash <125997726+dashhuang@users.noreply.github.com>
---
 .../agent-island/__tests__/service.test.ts    |  2 +-
 .../__tests__/agent-input-coordinator.test.ts | 29 ++++++++++++++-
 .../__tests__/botTaskCoordination.test.ts     |  8 +----
 .../__tests__/botWorkbenchTranscripts.test.ts | 15 ++++----
 .../__tests__/sessionEventPipeline.test.ts    | 36 ++++++++++++-------
 .../main/maker-ipc/botWorkbenchTranscripts.ts | 14 +++++---
 .../src/main/maker-ipc/coordinationOutput.ts  | 12 -------
 .../main/maker-ipc/sessionEventPipeline.ts    |  6 +---
 .../main/maker-ipc/sessionEventPreparation.ts |  6 ++--
 .../desktop/src/shared/botTaskCoordination.ts |  8 +++--
 docs/product-rules/agent-user-parity.md       |  3 ++
 docs/product-rules/cindy-bots-runtime.md      |  7 ++--
 docs/product-rules/message-source.md          | 15 +++++---
 .../src/xdt-helper/get_task_digest.ts         |  5 +--
 .../src/xdt-helper/send_to_session.ts         |  2 +-
 15 files changed, 102 insertions(+), 66 deletions(-)
 delete mode 100644 apps/desktop/src/main/maker-ipc/coordinationOutput.ts

diff --git a/apps/desktop/src/main/agent-island/__tests__/service.test.ts b/apps/desktop/src/main/agent-island/__tests__/service.test.ts
index b31dfb4401f..94cfde33253 100644
--- a/apps/desktop/src/main/agent-island/__tests__/service.test.ts
+++ b/apps/desktop/src/main/agent-island/__tests__/service.test.ts
@@ -7013,7 +7013,7 @@ it('does not sound or mark attention for internal coordination completion, but s
   const meta = { sessionId: 'coordination', agentKind: 'codex' as const };
   service.handleAgentEvent(meta, { type: 'status', source: 'codex', data: { isRunning: true } });
   playSound.mockClear();
-  service.handleAgentEvent(meta, { ...doneEvent(), agentMeta: { botTaskCoordination: true, botPrivateReply: true } });
+  service.handleAgentEvent(meta, { ...doneEvent(), agentMeta: { botTaskCoordination: true, botPrivateReply: false } });
   expect(playSound).not.toHaveBeenCalled();
   expect(service.getSessionActivitySnapshot(meta.sessionId)?.attention).not.toBe(true);
   service.handleAgentEvent(meta, terminalErrorEvent('User action required'));
diff --git a/apps/desktop/src/main/maker-ipc/__tests__/agent-input-coordinator.test.ts b/apps/desktop/src/main/maker-ipc/__tests__/agent-input-coordinator.test.ts
index 7cf8fa84db4..459939145bb 100644
--- a/apps/desktop/src/main/maker-ipc/__tests__/agent-input-coordinator.test.ts
+++ b/apps/desktop/src/main/maker-ipc/__tests__/agent-input-coordinator.test.ts
@@ -2404,7 +2404,7 @@ describe('AgentInputCoordinator send transaction', () => {
   });
 
   it.each(['policy', 'preparation', 'provider'] as const)(
-    'keeps coordination quiet when human steering fails at %s', async (failure) => {
+    'keeps coordination ownership when human steering fails at %s', async (failure) => {
       const h = createHarness(), sid = `coordination-steer-${failure}`;
       h.sendToAgent.mockImplementationOnce(async () => { h.setRunning(true); return sendSuccess(); });
       h.coordinator.enqueue(sid, makeItem('internal', '[UI_ACTION_TRIGGER]File agreement', {
@@ -2436,6 +2436,33 @@ describe('AgentInputCoordinator send transaction', () => {
     },
   );
 
+  // The parent's reply ending releases its input lane even if the delegated
+  // child has not reported completion. No child terminal is needed to talk again.
+  it('dispatches owner input after a coordination turn ends and accepts another normal turn', async () => {
+    const h = createHarness(), sid = 'coordination-next-message';
+    h.coordinator.enqueue(sid, makeItem('report', '[UI_ACTION_TRIGGER]Build still running', {
+      botTaskCoordination: { delegationId: 'd', senderSessionId: 'child', runSequence: 1 },
+    }));
+    await flush();
+    h.coordinator.enqueue(sid, makeItem('owner', 'Check the other machine too'));
+    await flush();
+    expect(h.sendToAgent).toHaveBeenCalledOnce();
+    expect(h.coordinator.getProjection(sid).pendingQueue.map(item => item.clientId)).toEqual(['owner']);
+    h.setRunning(false);
+    h.coordinator.onTurnEvent(sid, 'done');
+    await flush();
+    expect(h.sendToAgent).toHaveBeenCalledTimes(2);
+    expect(h.sendToAgent.mock.calls[1]?.[1]).toEqual({ type: 'user', content: 'Check the other machine too' });
+    expect(h.coordinator.isActiveTaskCoordination(sid)).toBe(false);
+    h.setRunning(false);
+    h.coordinator.onTurnEvent(sid, 'done');
+    await flush();
+    h.coordinator.enqueue(sid, makeItem('next', 'What is the next step?'));
+    await flush();
+    expect(h.sendToAgent).toHaveBeenCalledTimes(3);
+    expect(h.coordinator.getProjection(sid).pendingQueue).toEqual([]);
+  });
+
   it.each(['projection', 'control', 'direct'] as const)(
     'keeps coordination queued instead of injecting it through %s steering', async (entry) => {
       const h = createHarness(), sid = `coordination-queue-steer-${entry}`;
diff --git a/apps/desktop/src/main/maker-ipc/__tests__/botTaskCoordination.test.ts b/apps/desktop/src/main/maker-ipc/__tests__/botTaskCoordination.test.ts
index c12b39afabf..b7053787829 100644
--- a/apps/desktop/src/main/maker-ipc/__tests__/botTaskCoordination.test.ts
+++ b/apps/desktop/src/main/maker-ipc/__tests__/botTaskCoordination.test.ts
@@ -8,8 +8,7 @@ import { assertBotTaskCoordination, classifySessionMessagePurpose, coordinationI
 import { buildMakerUserMessage, sanitizeQueuedMessageForPersistence, type AgentInputQueuedMessage } from '../../../shared/agentInputQueue';
 import { redactMessageOriginForSharedGuest, redactQueueItemForSharedGuest } from '../../device-link/sharedTaskMessageOrigin';
 import { rebuildSessionQueueItem } from '../sessionControlService';
-import { projectCoordinationOutput } from '../coordinationOutput';
-import { AUTO_REVIEW_DELEGATED_CONTINUATION, type AgentEvent } from '@cindy/maker-core';
+import { AUTO_REVIEW_DELEGATED_CONTINUATION } from '@cindy/maker-core';
 import { stripMainOnlySendOpts } from '../mobileClientPromptNote';
 import { captureAgentRequest, formatForwardedRequest } from '../forwardedRequest';
 
@@ -90,11 +89,6 @@ it('retains hidden persistence and routing when the sending task edits its queue
   expect(edited.botTaskCoordination).toEqual(receipt);
   expect(buildMakerUserMessage(edited)).toMatchObject({ content: expect.stringContaining('Updated agreement') });
 });
-it.each<AgentEvent['type']>(['interaction_request', 'interaction_dismissed', 'error', 'tool_use', 'tool_result', 'account_usage'])('does not suppress an attention/action event: %s', type => {
-  const event: AgentEvent = { type, source: 'codex', data: { requestId: 'interaction', kind: 'ask_user_question', questions: [] } };
-  expect(projectCoordinationOutput(event, true)).toBe(event);
-});
-
 it('uses the registered host entry for both quiet and explicitly visible delegated sends', async () => {
   const source = readFileSync(resolve(__dirname, '..', 'register.ts'), 'utf8').replace(/\r\n/g, '\n');
   const start = source.indexOf('    sendToSession: async (params) => {');
diff --git a/apps/desktop/src/main/maker-ipc/__tests__/botWorkbenchTranscripts.test.ts b/apps/desktop/src/main/maker-ipc/__tests__/botWorkbenchTranscripts.test.ts
index dfb46689623..736aa68ce14 100644
--- a/apps/desktop/src/main/maker-ipc/__tests__/botWorkbenchTranscripts.test.ts
+++ b/apps/desktop/src/main/maker-ipc/__tests__/botWorkbenchTranscripts.test.ts
@@ -1,9 +1,9 @@
 /**
  * 测试价值
- * - 生产入口与可观察结果:boundTranscript / lastExchange 的输出(伙伴读到的摘录与任务摘要的两句)。
- * - 规则依据:工作台摘录有界且去掉指令块;get_task_digest 取「主人最后的要求 / 助手最后一句」。
- * - 现有覆盖缺口:放弃「项目」后本机转录扫描已删除(其测试随之删除);lastExchange 为新增。
- * - 最小失败输入:末尾是空助手消息 / 指令块;多行助手回复。
+ * - 生产入口与可观察结果:boundTranscript / lastExchange 的摘录与任务摘要。
+ * - 规则依据:内部协调不是用户要求;最新助手回复保留重点,由工具层按预算截短。
+ * - 现有覆盖缺口:内部回报覆盖了用户要求;多行回复只保留最后一行而丢失结论。
+ * - 最小失败输入:用户要求之后出现隐藏协调输入;助手先说结论再说下一步。
  * - tier:unit(纯函数,数据库模块 mock 掉,不加载)。
  */
 import { describe, expect, it, vi } from 'vitest';
@@ -30,14 +30,17 @@ describe('boundTranscript', () => {
 });
 
 describe('lastExchange', () => {
-  it('takes the latest owner ask and the last line of the latest non-empty assistant reply', () => {
+  it('keeps the latest user ask and the full latest non-empty assistant reply', () => {
     expect(lastExchange([
       { role: 'user', text: '先修登录' },
       { role: 'assistant', text: '好的' },
       { role: 'user', text: '<system-reminder>internal</system-reminder>再跑一遍测试' },
       { role: 'assistant', text: '跑完了。\n还有 2 个失败' },
+      { role: 'user', text: '[UI_ACTION_TRIGGER]后台回报：正在修复失败项' },
       { role: 'assistant', text: '<system-reminder>only a hint</system-reminder>' },
-    ])).toEqual({ lastUserAsk: '再跑一遍测试', lastAssistantLine: '还有 2 个失败' });
+    ])).toEqual({ lastUserAsk: '再跑一遍测试', lastAssistantLine: '跑完了。\n还有 2 个失败' });
     expect(lastExchange([])).toEqual({ lastUserAsk: null, lastAssistantLine: null });
+    expect(lastExchange([{ role: 'user', text: '[UI_ACTION_TRIGGER]仅内部协调' }]))
+      .toEqual({ lastUserAsk: null, lastAssistantLine: null });
   });
 });
diff --git a/apps/desktop/src/main/maker-ipc/__tests__/sessionEventPipeline.test.ts b/apps/desktop/src/main/maker-ipc/__tests__/sessionEventPipeline.test.ts
index c7c4a24eef9..fe16981efe0 100644
--- a/apps/desktop/src/main/maker-ipc/__tests__/sessionEventPipeline.test.ts
+++ b/apps/desktop/src/main/maker-ipc/__tests__/sessionEventPipeline.test.ts
@@ -425,31 +425,43 @@ function ordered(...names: string[]) {
 }
 
 describe('production Session event pipeline', () => {
-  it('delivers coordination actions without public prose, then preserves ordinary and completion replies', async () => {
+  // Regression: a delegated report used to leave only tool rows in the chat.
+  // The real pipeline must persist and broadcast its reply, then settle normally.
+  it('delivers coordination progress and final text through the ordinary message pipeline', async () => {
     const h = harness();
+    h.deps.redactEventForRenderer.mockImplementation(value => value);
     h.deps.agentInputCoordinatorHolder.isActiveTaskCoordination.mockReturnValue(true);
     h.deps.agentInputCoordinatorHolder.getActiveInputClientId.mockReturnValue('coordination');
     try {
-      h.emit(event('text', { text: 'Internal file ownership agreement', isFinal: true }));
-      h.emit(event('text', { text: 'Internal standalone' }, { standaloneText: true }));
-      h.emit(event('thinking', { text: 'Internal reasoning' }));
-      expect(effects.fn('onAssistantTextEvent')).not.toHaveBeenCalled();
-      expect(effects.fn('onStandaloneTextEvent')).not.toHaveBeenCalled();
-      expect(h.deps.broadcastToAllWindows).not.toHaveBeenCalled();
+      h.emit(event('text', { text: 'The new machines are running; checking the test results.', isFinal: false },
+        { agentMeta: { assistantPhase: 'commentary' } }));
+      expect(effects.fn('onAssistantTextEvent')).toHaveBeenCalledWith('task',
+        expect.objectContaining({ text: 'The new machines are running; checking the test results.' }),
+        expect.objectContaining({ botPrivateReply: false, assistantPhase: 'commentary' }), undefined);
+      expect(h.deps.broadcastToAllWindows).toHaveBeenCalledWith('maker:event', expect.objectContaining({
+        event: expect.objectContaining({ type: 'text', data: expect.objectContaining({
+          text: 'The new machines are running; checking the test results.',
+        }) }),
+      }));
+      h.emit(event('text', { text: 'Progress notice' }, { standaloneText: true }));
+      expect(effects.fn('onStandaloneTextEvent')).toHaveBeenCalledWith('task', 'Progress notice', { botPrivateReply: false });
+      h.emit(event('thinking', { text: 'Check the test results' }));
+      expect(effects.fn('onThinkingEvent')).toHaveBeenCalledOnce();
       h.emit(event('tool_use', { id: 'tool', name: 'read', input: {} }));
       expect(effects.fn('onToolUseEvent')).toHaveBeenCalledOnce();
       h.deps.agentInputCoordinatorHolder.onTurnEvent.mockImplementation(() => {
         h.deps.agentInputCoordinatorHolder.isActiveTaskCoordination.mockReturnValue(false);
       });
-      h.emit(event('done', { result: 'Internal acknowledgement', finalText: 'Internal acknowledgement' }));
+      h.emit(event('done', { result: 'The tests passed.', finalText: 'The tests passed.' }));
       expect(h.deps.broadcastToAllWindows).toHaveBeenCalledWith('maker:event', expect.objectContaining({
-        event: expect.objectContaining({ type: 'done', data: expect.objectContaining({ result: '', finalText: '' }),
-          agentMeta: expect.objectContaining({ botPrivateReply: true, botTaskCoordination: true }) }),
+        event: expect.objectContaining({ type: 'done', data: expect.objectContaining({ result: 'The tests passed.', finalText: 'The tests passed.' }),
+          agentMeta: expect.objectContaining({ botPrivateReply: false, botTaskCoordination: true }) }),
       }));
-      expect(JSON.stringify(h.deps.broadcastToAllWindows.mock.calls)).not.toContain('Internal acknowledgement');
+      expect(h.deps.agentInputCoordinatorHolder.onTurnEvent).toHaveBeenCalled();
+      const persisted = effects.fn('onAssistantTextEvent').mock.calls.length;
       h.deps.agentInputCoordinatorHolder.getActiveInputClientId.mockReturnValue('bot-delegation-completion:result');
       h.emit(event('text', { text: 'User requested final result', isFinal: true }));
-      expect(effects.fn('onAssistantTextEvent')).toHaveBeenCalledOnce();
+      expect(effects.fn('onAssistantTextEvent')).toHaveBeenCalledTimes(persisted + 1);
     } finally { await h.dispose(); }
   });
 
diff --git a/apps/desktop/src/main/maker-ipc/botWorkbenchTranscripts.ts b/apps/desktop/src/main/maker-ipc/botWorkbenchTranscripts.ts
index 1a5e18d0fb7..23569720156 100644
--- a/apps/desktop/src/main/maker-ipc/botWorkbenchTranscripts.ts
+++ b/apps/desktop/src/main/maker-ipc/botWorkbenchTranscripts.ts
@@ -1,11 +1,12 @@
 /**
- * 伙伴工作台与任务摘要的只读会话读取:有界摘录与「最后一问 / 最后一句」。
+ * 伙伴工作台与任务摘要的只读会话读取:有界摘录与最近用户消息 / 助手回复。
  *
  * 2026-10-17 放弃「项目」后,不再扫描本机 Claude Code / Codex / Pi 转录目录,也不再为项目
  * 现算摘要;只读 Cindy 自己库里的任务,两条有界查询,输出统一去掉 `<system-reminder>` 之类的
  * 指令块。不写库、不导入。
  */
 import { and, desc, eq, gt, inArray, isNull } from 'drizzle-orm';
+import { isSyntheticTriggerText } from '@cindy/maker-shared/synthetic-trigger';
 
 import { getDbClient } from '../localDb/client/current.js';
 import { messages, sessions } from '../localDb/schema.js';
@@ -18,7 +19,7 @@ import {
 export const WORKBENCH_TRANSCRIPT_MAX_CHARS = 4_000;
 const PER_MESSAGE_MAX_CHARS = 1_200;
 const SESSION_MESSAGE_ROWS = 60;
-/** 找「最后一问 / 最后一句」时最多往回看的消息条数。 */
+/** 找最近用户消息 / 助手回复时最多往回看的消息条数。 */
 const LAST_EXCHANGE_ROWS = 20;
 
 function cleanText(raw: string): string {
@@ -112,13 +113,16 @@ export function lastExchange(items: readonly Pick<WorkbenchTranscriptItem, 'role
   for (let index = items.length - 1; index >= 0 && (!lastUserAsk || !lastAssistantLine); index -= 1) {
     const text = cleanText(items[index].text);
     if (!text) continue;
-    if (items[index].role === 'user') lastUserAsk ??= text;
-    else lastAssistantLine ??= text.split('\n').filter(Boolean).pop() ?? text;
+    if (items[index].role === 'user') {
+      // Shared with remote history: coordination and recovery inputs carry the
+      // same durable marker and must not replace the user's actual request.
+      if (!isSyntheticTriggerText(text)) lastUserAsk ??= text;
+    } else lastAssistantLine ??= text;
   }
   return { lastUserAsk, lastAssistantLine };
 }
 
-/** 本机任务的「最后一问 / 最后一句」:一条有界查询。 */
+/** 本机任务的最近用户消息 / 助手回复:一条有界查询。 */
 export async function readSessionLastExchange(sessionId: string) {
   const { db, where } = await sessionMessageWhere(sessionId);
   const rows = await db
diff --git a/apps/desktop/src/main/maker-ipc/coordinationOutput.ts b/apps/desktop/src/main/maker-ipc/coordinationOutput.ts
deleted file mode 100644
index a7ff4613df2..00000000000
--- a/apps/desktop/src/main/maker-ipc/coordinationOutput.ts
+++ /dev/null
@@ -1,12 +0,0 @@
-import type { AgentEvent } from '@cindy/maker-core';
-
-/** Preserve tool actions, errors, interactions and usage; only internal prose is omitted. */
-export function projectCoordinationOutput(event: AgentEvent, coordination: boolean): AgentEvent | null {
-  if (!coordination || event.runtimeRecovery) return event;
-  if (event.type === 'text' || event.type === 'thinking') return null;
-  if (event.type === 'done') {
-    const data = event.data && typeof event.data === 'object' ? event.data : {};
-    return { ...event, data: { ...data, result: '', finalText: '' } };
-  }
-  return event;
-}
diff --git a/apps/desktop/src/main/maker-ipc/sessionEventPipeline.ts b/apps/desktop/src/main/maker-ipc/sessionEventPipeline.ts
index 55850d47f53..cc2de0e14dc 100644
--- a/apps/desktop/src/main/maker-ipc/sessionEventPipeline.ts
+++ b/apps/desktop/src/main/maker-ipc/sessionEventPipeline.ts
@@ -1,4 +1,3 @@
-import { projectCoordinationOutput } from './coordinationOutput.js';
 import { observeBotLearningTurn } from './botLearningFeedback.js';
 import { projectQuietScheduledOutput } from '../scheduler-host/silent-output.js';
 import type { AgentEvent, Session } from '@cindy/maker-core';
@@ -45,10 +44,7 @@ export function handleSessionEvent(
 ): void {
   const prepared = prepareSessionEvent(deps, session, event);
   if (!prepared) return;
-  const project = (input: AgentEvent) => {
-    const scheduled = projectQuietScheduledOutput(input);
-    return scheduled && projectCoordinationOutput(scheduled, prepared.taskCoordination);
-  };
+  const project = projectQuietScheduledOutput;
   const projected = project(prepared.event);
   if (!projected) return;
   if (projected !== prepared.event) {
diff --git a/apps/desktop/src/main/maker-ipc/sessionEventPreparation.ts b/apps/desktop/src/main/maker-ipc/sessionEventPreparation.ts
index 80bdc56f328..2a77e4a2441 100644
--- a/apps/desktop/src/main/maker-ipc/sessionEventPreparation.ts
+++ b/apps/desktop/src/main/maker-ipc/sessionEventPreparation.ts
@@ -185,7 +185,6 @@ export function prepareSessionEvent(
   const taskCoordination = event.turnScope !== 'background'
     && (deps.agentInputCoordinatorHolder?.isActiveTaskCoordination?.(session.id, event.sessionTurnGeneration) ?? false);
   if (event.type === 'text' && event.standaloneText === true) {
-    if (taskCoordination && !event.runtimeRecovery) return;
     if (isQuietScheduledOutput(event) && !event.runtimeRecovery) return;
     // Deliver through the existing persisted-row channel only. Sending a text
     // event as well would let older renderers adopt the notice as their active
@@ -226,7 +225,9 @@ export function prepareSessionEvent(
         ...event,
         agentMeta: {
           ...event.agentMeta,
-          botPrivateReply: isBotPrivateInput(activeInputId) || (taskCoordination && !event.runtimeRecovery),
+          // Coordination hides the incoming report, not the teammate's reply.
+          botPrivateReply: isBotPrivateInput(activeInputId),
+          // Retain provenance for the existing quiet completion notification policy.
           botTaskCoordination: taskCoordination && !event.runtimeRecovery,
           // The user input this turn answers; companion chats can show it above the reply
           // (cindy-bots-runtime.md「回复对应的消息」). Hidden trigger inputs are skipped on display.
@@ -603,7 +604,6 @@ export function prepareSessionEvent(
     attributedEvent,
     botTaskResultInputIds,
     groupLaneInputClientId,
-    taskCoordination,
     broadcastEvent,
     eventAgentMeta,
     pendingContextSnapshot,
diff --git a/apps/desktop/src/shared/botTaskCoordination.ts b/apps/desktop/src/shared/botTaskCoordination.ts
index 2e6cb76c4ac..8fbe1fae0da 100644
--- a/apps/desktop/src/shared/botTaskCoordination.ts
+++ b/apps/desktop/src/shared/botTaskCoordination.ts
@@ -10,7 +10,9 @@ export type SessionMessagePurpose = 'coordination' | 'user-visible';
 /** Per-input model instruction; never included in visible queue/history text. */
 export function coordinationModelPrefix(receipt: BotTaskCoordination): string {
   return `[Internal task coordination from session ${receipt.senderSessionId}; delegation ${receipt.delegationId}; run ${receipt.runSequence}]\n`
-    + 'Process this coordination and use tools as needed. Prose in this turn is internal. '
-    + 'Use question/permission tools for necessary user decisions. For a user-facing result or a failure requiring user intervention, '
-    + 'use send_to_session with message_purpose="user-visible". Normal task completion is delivered separately.\n\n';
+    + 'This incoming report is internal context, not a new request from the owner. Process it and use tools as needed. '
+    + 'Your replies appear in the owner conversation: directly explain meaningful progress, results or blockers; '
+    + 'if there is nothing useful to report, you may stay silent. Use question/permission tools for necessary user decisions. '
+    + 'Ending your reply ends only this turn; delegated work may continue. State what remains in progress and report completion only when verified. '
+    + 'Normal task completion is delivered separately.\n\n';
 }
diff --git a/docs/product-rules/agent-user-parity.md b/docs/product-rules/agent-user-parity.md
index a7366151065..99af1d68bd7 100644
--- a/docs/product-rules/agent-user-parity.md
+++ b/docs/product-rules/agent-user-parity.md
@@ -91,6 +91,9 @@ Worker 工具省略 working_dir 时保持继承 Lead 任务记录目录，不跟
 - 任务摘要（`get_task_digest`）：`packages/lizi-mcps/src/xdt-helper/get_task_digest.ts`（输出合同 ≤2KB）、
   `apps/desktop/src/main/mcp-integrations/taskDigest.ts` 与 `taskDigestHost.ts`（跨电脑只读，复用
   `local-db:sessions:get`、`maker:list-active`、历史消息与 PR 状态通道）。
+  本机与跨电脑均排除带隐藏触发标记的协调／续跑输入，保留最新助手回复全文供工具按预算截短；
+  不用内部回报替代用户要求，也不只取助手回复最后一行。`finished` 只表示所查任务本轮结束，
+  不证明委派工作或用户目标完成；不参与输入放行判定，下一条消息仍按实时运行状态和队列派发。
 - 伙伴跟进任务：`apps/desktop/src/main/maker-ipc/botWorkbenchAccess.ts`（授权：自动跟进所有任务的伙伴
   可读取既有范围内的任务，含在线电脑；继续或接手还需行动授权，其它伙伴只读自己开的或已接手的任务）。
 - 协议与兼容见 [`../dev-rules/protocol-compatibility.md`](../dev-rules/protocol-compatibility.md)
diff --git a/docs/product-rules/cindy-bots-runtime.md b/docs/product-rules/cindy-bots-runtime.md
index 67096e9ecaa..a783ede6185 100644
--- a/docs/product-rules/cindy-bots-runtime.md
+++ b/docs/product-rules/cindy-bots-runtime.md
@@ -713,9 +713,10 @@ Session 任务遵守同一套机制与呈现契约：
   超时终态也必须检查旧引擎是否已停，仍活跃时不得通过续接启动第二份执行。
 
 委派子任务向发起伙伴发送文件分工等内部协调时，使用 `send_to_session` 的协调用途，
-其默认识别、来源核验、静默范围和历史兼容遵守 [message-source.md](message-source.md#委派任务的内部协调)。
-这条路径仍唤醒伙伴消费输入，不创建公开进度答复；正常完成结果、必要交互与错误继续保留。
-普通任务来信和明确 `user-visible` 的发送不受此静默规则影响。
+其默认识别、来源核验、通知静默和历史兼容遵守 [message-source.md](message-source.md#委派任务的内部协调)。
+原始输入隐藏，但伙伴处理后的进度与结果按普通助手消息显示、保存，最终答复后折叠过程。
+本轮结束不等待后台任务完成，也不阻止用户继续发消息；运行中排队／插话、结束后开始新轮的规则不变。
+正常完成结果、必要交互与错误继续保留；普通任务来信和明确 `user-visible` 的输入显示规则不变。
 
 `send_to_agent` 是另一条更轻的伙伴间通讯通道，不得伪装成任务：
 
diff --git a/docs/product-rules/message-source.md b/docs/product-rules/message-source.md
index 3909fe7c8a3..11e6e788f70 100644
--- a/docs/product-rules/message-source.md
+++ b/docs/product-rules/message-source.md
@@ -108,13 +108,18 @@ Agent 用 `send_to_session` 把消息发给另一台电脑上的既有任务时
 校验失败沿正常错误/重试路径处理，不能丢弃后假报送达。用途回执不接受 renderer/device-link 自报，
 也不为工具执行提供新授权；共享访客投影移除回执里的私有标识。
 
-宿主只对该协调输入所启动的文字答复与思考做静默投影，成功结束不产生通知/未读；
-工具审计、交互、权限请求、错误和宿主恢复说明保留。需要用户决定时使用交互工具；
-给用户的正文结果通过明确 `user-visible` 的发送提供。用户插话接管后的答复恢复可见。
+**2026-10-25：内部输入与对用户的回复分开。** 原始协调输入继续隐藏，伙伴处理后的进度与结果
+直接走普通助手消息的落库、广播和历史读取；不再丢弃整轮文字、思考或清空结束事件的答复。
+有重要进展就直接说明，没有值得汇报的信息可以不发言；需要用户决定时使用交互工具。
+不要求伙伴通过 `send_to_session` 给自己发消息来汇报，也不扩大任务写工具的权限。
+协调轮次成功结束的通知静默规则保持原样；通知静默不等于隐藏聊天记录。
+进度运行中可见，最终答复后沿用普通任务折叠规则；工具审计、交互、权限请求、错误和恢复说明保留。
 委派正常完成仍走原本独立的结果记录和隐藏 completion 唤醒，不改成内部协调。
 
-无需数据库迁移。旧消息不重新分类；新版宿主产生的隐藏输入和静默输出可由现有桌面/手机读取。
-旧宿主不会因新客户端自行获得协调用途识别能力；降级不承诺继续执行新队列的静默语义。
+本轮结束仅释放当前任务的输入通道，不代表委派工作或用户目标已完成；后台任务继续运行时，
+用户仍可给主任务发送下一条消息。主任务运行中沿用原有排队／插话，结束后正常开始下一轮。
+无需数据库迁移或新协议。旧消息不重新分类，也不补造此前已被过滤的正文；新消息使用已有
+桌面／手机消息通道。旧宿主仍会过滤协调回复，需更新宿主才能恢复该路径的进度显示。
 
 ## 对话使用的模型
 
diff --git a/packages/lizi-mcps/src/xdt-helper/get_task_digest.ts b/packages/lizi-mcps/src/xdt-helper/get_task_digest.ts
index 80f1ce05ac7..7c50b957487 100644
--- a/packages/lizi-mcps/src/xdt-helper/get_task_digest.ts
+++ b/packages/lizi-mcps/src/xdt-helper/get_task_digest.ts
@@ -108,12 +108,13 @@ export function boundTaskDigest(raw: TaskDigestWire): TaskDigestWire {
 }
 
 const DESCRIPTION = [
-  '读一件任务的短摘要(≤2KB):状态 phase、已持续分钟数、主人最后的要求、助手最后一句、待回答的问题、错误、当前模型路线、',
+  '读一件任务的短摘要(≤2KB):状态 phase、已持续分钟数、最近用户消息(排除隐藏协调/续跑输入)、最新助手回复片段、待回答的问题、错误、当前模型路线、',
   '建议换到的下一条路线(只在出错或撞用量上限、换模型确实有用且装得下当前上下文时,给你候选链里还没试过的那条;否则 null,此时提醒主人)、PR 状态、排队消息数、谁开的 / 谁在跟。',
   'idle_minutes 距最后一次动静的分钟数;tool_in_flight 为 true 表示有工具调用(如长时间构建、测试)正在执行,宿主不知道时不出现。',
   '被程序叫醒跟进任务时先读这一份再决定;要看原文再用 get_chat_history。',
   '作用范围:本机任务直接传 id;另一台电脑上的任务传 deviceId::sessionId(只读,对方须在线)。',
-  'phase:running 运行中 / waiting-permission 等权限 / waiting-answer 等回答 / errored 出错 / usage-limit 撞用量上限 / finished 刚做完 / idle 空闲。',
+  'phase:running 运行中 / waiting-permission 等权限 / waiting-answer 等回答 / errored 出错 / usage-limit 撞用量上限 / finished 本轮已结束 / idle 空闲。',
+  'finished 只描述所查任务的本轮，不代表它委派的后台任务或用户目标已完成。下一条消息的派发仍由该任务的实时运行状态和输入队列决定。',
   'suggested_route 只供判断，不是换模型授权。获得目标和变更的明确授权后才用 set_session_runtime(expected_generation 填 runtime_generation;只支持本机任务)。',
 ].join('\n');
 
diff --git a/packages/lizi-mcps/src/xdt-helper/send_to_session.ts b/packages/lizi-mcps/src/xdt-helper/send_to_session.ts
index 310946f1fac..35e6e6e82c6 100644
--- a/packages/lizi-mcps/src/xdt-helper/send_to_session.ts
+++ b/packages/lizi-mcps/src/xdt-helper/send_to_session.ts
@@ -191,7 +191,7 @@ export function registerSendToSessionTool(
           "要发送的消息。伙伴代用户下达任务要求时，宿主另附可核实的用户原话；只补充必要上下文，保留原意和范围，不复述授权、不要求对方回传。普通任务与自动化消息按正文发送。",
         ),
       message_purpose: z.enum(["coordination", "user-visible"]).optional().describe(
-        "委派子任务发给发起伙伴默认 coordination：送达模型但内部文字不进正文或通知。user-visible 用于用户要求的结果、需要用户介入的消息。其他普通任务默认保持可见；coordination 仅限宿主核实的委派关系。正常完成自动回传，无需重发。",
+        "委派子任务发给发起伙伴默认 coordination：原始来信隐藏，伙伴处理后的进度和结果可直接正常回复。user-visible 让来信本身可见。其他普通任务默认保持可见；coordination 仅限宿主核实的委派关系，不改变操作权限。正常完成自动回传，无需重发。",
       ),
       title: z
         .string()

From 4a852f3591396986271a5979b6c24b4bdd46e7f4 Mon Sep 17 00:00:00 2001
From: Dash <125997726+dashhuang@users.noreply.github.com>
Date: Sun, 25 Oct 2026 13:13:52 +1300
Subject: [PATCH 2/2] =?UTF-8?q?fix(companion):=20=E4=B8=BA=E4=BB=BB?=
 =?UTF-8?q?=E5=8A=A1=E6=91=98=E8=A6=81=E4=BF=9D=E7=95=99=E9=95=BF=E5=9B=9E?=
 =?UTF-8?q?=E5=A4=8D=E9=A6=96=E5=B0=BE=E7=BB=93=E8=AE=BA?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Signed-off-by: Dash <125997726+dashhuang@users.noreply.github.com>
---
 docs/product-rules/agent-user-parity.md       |  1 +
 .../src/__tests__/getTaskDigestTool.test.ts   | 25 +++++++++++++++++--
 .../src/xdt-helper/get_task_digest.ts         | 13 ++++++----
 3 files changed, 32 insertions(+), 7 deletions(-)

diff --git a/docs/product-rules/agent-user-parity.md b/docs/product-rules/agent-user-parity.md
index 99af1d68bd7..cfd55cfc6c9 100644
--- a/docs/product-rules/agent-user-parity.md
+++ b/docs/product-rules/agent-user-parity.md
@@ -92,6 +92,7 @@ Worker 工具省略 working_dir 时保持继承 Lead 任务记录目录，不跟
   `apps/desktop/src/main/mcp-integrations/taskDigest.ts` 与 `taskDigestHost.ts`（跨电脑只读，复用
   `local-db:sessions:get`、`maker:list-active`、历史消息与 PR 状态通道）。
   本机与跨电脑均排除带隐藏触发标记的协调／续跑输入，保留最新助手回复全文供工具按预算截短；
+  长回复保留首尾片段，总预算进一步收紧时仍保留首尾，避免丢掉末尾结果或阻塞说明。
   不用内部回报替代用户要求，也不只取助手回复最后一行。`finished` 只表示所查任务本轮结束，
   不证明委派工作或用户目标完成；不参与输入放行判定，下一条消息仍按实时运行状态和队列派发。
 - 伙伴跟进任务：`apps/desktop/src/main/maker-ipc/botWorkbenchAccess.ts`（授权：自动跟进所有任务的伙伴
diff --git a/packages/lizi-mcps/src/__tests__/getTaskDigestTool.test.ts b/packages/lizi-mcps/src/__tests__/getTaskDigestTool.test.ts
index 3c237c5ccd8..73d5fee30c2 100644
--- a/packages/lizi-mcps/src/__tests__/getTaskDigestTool.test.ts
+++ b/packages/lizi-mcps/src/__tests__/getTaskDigestTool.test.ts
@@ -2,8 +2,8 @@
  * 测试价值
  * - 生产入口与可观察结果:get_task_digest 经 registry.call 返回给模型的 JSON(字段、大小、错误码)。
  * - 规则依据:被程序叫醒的伙伴先读 ≤2KB 摘要再决定;每段文字 ≤300 字;只读调用方可读范围内的任务。
- * - 现有覆盖缺口:新工具。
- * - 最小失败输入:每段文字都是长中文(UTF-8 三字节)的摘要;受限历史范围里请求范围外的任务。
+ * - 现有覆盖缺口:长助手回复只截开头,丢掉末尾结果;总预算收紧时也须保留结论。
+ * - 最小失败输入:超过 300 字、末尾才写失败结论的回复,分别带有/不带其他长字段;受限范围外任务。
  * - tier:unit(纯工具层,宿主回调 mock)。
  */
 import { describe, expect, it, vi } from 'vitest';
@@ -73,6 +73,27 @@ describe('get_task_digest', () => {
     expect(boundTaskDigest(short)).toEqual(short);
   });
 
+  it.each([false, true])('retains the opening and final blocker of a long reply (budget pressure=%s)', async (pressure) => {
+    const opening = '构建进展：';
+    const conclusion = '最终结果：部署失败，需要回滚。';
+    const raw = {
+      ...digest,
+      title: '部署结果',
+      last_user_ask: pressure ? longText : '检查部署结果',
+      last_assistant_line: `${opening}${'正在检查日志与测试结果。'.repeat(80)}\n${conclusion}`,
+      pending_question: pressure ? longText : null,
+      error: pressure ? longText : null,
+      pr_status: pressure ? longText : null,
+    };
+    const reg = setup({ taskDigest: async () => ({ ok: true as const, digest: raw }) });
+    const { text, json } = parse(await reg.call('get_task_digest', { session_id: raw.session_id }));
+    expect(new TextEncoder().encode(text).length).toBeLessThanOrEqual(TASK_DIGEST_MAX_BYTES);
+    const reply = String(json.last_assistant_line);
+    expect(reply.startsWith(opening)).toBe(true);
+    expect(reply.endsWith(conclusion)).toBe(true);
+    expect(Array.from(reply).length).toBeLessThanOrEqual(pressure ? 299 : 300);
+  });
+
   it('refuses a task outside a scoped caller\'s readable history and passes host errors through', async () => {
     const taskDigest = vi.fn(async () => ({ ok: false as const, errorCode: 'NOT_FOUND' as const, message: 'gone' }));
     const scoped = setup({
diff --git a/packages/lizi-mcps/src/xdt-helper/get_task_digest.ts b/packages/lizi-mcps/src/xdt-helper/get_task_digest.ts
index 7c50b957487..bfb09c75e08 100644
--- a/packages/lizi-mcps/src/xdt-helper/get_task_digest.ts
+++ b/packages/lizi-mcps/src/xdt-helper/get_task_digest.ts
@@ -68,12 +68,15 @@ export const TASK_DIGEST_TEXT_MAX = 300;
 const TITLE_MAX = 120;
 const SHRINKABLE = ['last_assistant_line', 'last_user_ask', 'pending_question', 'error', 'pr_status'] as const;
 
-function clip(value: string | null | undefined, max: number): string | null {
+function clip(value: string | null | undefined, max: number, keepEnd = false): string | null {
   if (typeof value !== 'string') return null;
   const text = value.replace(/\s+/g, ' ').trim();
   if (!text) return null;
   const chars = Array.from(text);
-  return chars.length <= max ? text : `${chars.slice(0, max - 1).join('')}…`;
+  if (chars.length <= max) return text;
+  if (!keepEnd) return `${chars.slice(0, max - 1).join('')}…`;
+  const head = Math.floor((max - 1) / 2);
+  return `${chars.slice(0, head).join('')}…${chars.slice(chars.length - (max - 1 - head)).join('')}`;
 }
 
 const encoder = new TextEncoder();
@@ -86,13 +89,13 @@ export function boundTaskDigest(raw: TaskDigestWire): TaskDigestWire {
     ...raw,
     title: clip(raw.title, TITLE_MAX) ?? '',
     last_user_ask: clip(raw.last_user_ask, TASK_DIGEST_TEXT_MAX),
-    last_assistant_line: clip(raw.last_assistant_line, TASK_DIGEST_TEXT_MAX),
+    last_assistant_line: clip(raw.last_assistant_line, TASK_DIGEST_TEXT_MAX, true),
     pending_question: clip(raw.pending_question, TASK_DIGEST_TEXT_MAX),
     error: clip(raw.error, TASK_DIGEST_TEXT_MAX),
     pr_status: clip(raw.pr_status, TASK_DIGEST_TEXT_MAX),
   };
   for (let max = TASK_DIGEST_TEXT_MAX; payloadBytes(digest) > TASK_DIGEST_MAX_BYTES && max >= 20; max = Math.floor(max / 2)) {
-    for (const key of SHRINKABLE) digest[key] = clip(digest[key], max);
+    for (const key of SHRINKABLE) digest[key] = clip(digest[key], max, key === 'last_assistant_line');
     digest.title = clip(digest.title, Math.min(TITLE_MAX, max)) ?? '';
   }
   // 极端情况(超长 id / 模型名):去掉可选文字,保证合同成立。
@@ -108,7 +111,7 @@ export function boundTaskDigest(raw: TaskDigestWire): TaskDigestWire {
 }
 
 const DESCRIPTION = [
-  '读一件任务的短摘要(≤2KB):状态 phase、已持续分钟数、最近用户消息(排除隐藏协调/续跑输入)、最新助手回复片段、待回答的问题、错误、当前模型路线、',
+  '读一件任务的短摘要(≤2KB):状态 phase、已持续分钟数、最近用户消息(排除隐藏协调/续跑输入)、最新助手回复首尾片段、待回答的问题、错误、当前模型路线、',
   '建议换到的下一条路线(只在出错或撞用量上限、换模型确实有用且装得下当前上下文时,给你候选链里还没试过的那条;否则 null,此时提醒主人)、PR 状态、排队消息数、谁开的 / 谁在跟。',
   'idle_minutes 距最后一次动静的分钟数;tool_in_flight 为 true 表示有工具调用(如长时间构建、测试)正在执行,宿主不知道时不出现。',
   '被程序叫醒跟进任务时先读这一份再决定;要看原文再用 get_chat_history。',