diff --git a/CHANGELOG.md b/CHANGELOG.md
index 948072117..fc83d5b7e 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -19,12 +19,16 @@
 
 ### 新增
 
-- 知识库支持下载原文、按原排版预览，聊天内可直接预览引用
-- 技能可在技能包与专家工作区之间复制
-- 记忆树支持手动新建与修正；Token 统计支持日期筛选与 Excel 导出
-- 按用户限制存储根目录与 Token 配额；创建专家可带默认知识库与连接器
-- 聊天支持 @ 子专家、技能斜杠插入，以及 ask_agent 独立对话线程
-- 滴滴连接器；可选分段历史归档；便携运行时升级并支持 SQLite 备份
+- 知识库支持下载上传原文；PDF / DOCX / PPTX / XLSX 在预览弹窗中按原排版查看；文档区支持拖拽上传（#591）
+- 知识库 / 工作区 PDF 预览改用 PDF.js（react-pdf），替代浏览器内置 iframe 查看器
+- 技能包支持将选定技能一次性复制到专家工作区，工作区技能也可推送到当前用户有写权限的技能包（#618）。
+- 对话权限模式 Ask / Plan / Craft：turn 级 `conversation_mode`、`ConversationModeMiddleware`（工具过滤 + 系统提示 + denylist 硬拒；Ask 额外禁用 `write_todos`；Ask/Plan 拦截 `task` / `acp_runner` / `ask_agent`）、agent `default_conversation_mode` 与 `default_knowledge_base_ids`；Composer 模式菜单（默认 / 计划 / 问答）；IM `/mode`（及 `/ask` `/plan` `/craft`）线程粘性覆盖（#616）
+- 计划模式回合结束后展示「按计划执行 / 继续改计划」；确认后切到默认模式，计划 brief 经 `plan_brief` 由网关注入 system（不落对话气泡）（#616）
+- 记忆树支持手动新建主题/记忆，以及编辑已有记忆（replace/supersede，保留变更历史）。
+- Token 统计支持日期范围筛选（RangePicker 预设 + 自定义起止）并导出 Excel 用量明细（#154）
+- Token 用量 Excel 含专家名称、中英列表头，以及按天/专家/模型汇总 sheet 与图表
+- Token 用量 Excel 按报表规范打磨：冻结表头、千分位、合计公式行、筛选与仪表盘配色柱图
+- 知识库卡片「更多」菜单可一键添加到当前会话
 
 ### 修复
 
diff --git a/dashboard/src/context/AgentContext.tsx b/dashboard/src/context/AgentContext.tsx
index d1a28c476..fe4f60b24 100644
--- a/dashboard/src/context/AgentContext.tsx
+++ b/dashboard/src/context/AgentContext.tsx
@@ -37,6 +37,10 @@ export interface OctopAgent {
   description: string | null;
   persona_mbti: string | null;
   default_model: string | null;
+  /** Composer default: ask | plan | craft (#616). Missing → craft. */
+  default_conversation_mode?: "ask" | "plan" | "craft" | null;
+  /** Composer / turn default knowledge bases when the turn omits an explicit list. */
+  default_knowledge_base_ids?: string[] | null;
   system_prompt: string | null;
   template_name: string | null;
   state: "running" | "stopped" | "failed" | "starting" | "stopping" | string;
diff --git a/dashboard/src/locales/en.json b/dashboard/src/locales/en.json
index e68073df6..8a91fb61d 100644
--- a/dashboard/src/locales/en.json
+++ b/dashboard/src/locales/en.json
@@ -824,6 +824,11 @@
     "patchFailed": "Failed to save changes",
     "defaultModelLabel": "Default Model",
     "defaultModelPlaceholder": "Select a model",
+    "defaultConversationModeLabel": "Default conversation mode",
+    "defaultConversationModeHint": "Initial permission mode for new chats; change anytime in the composer.",
+    "defaultKnowledgeBasesLabel": "Default knowledge bases",
+    "defaultKnowledgeBasesHint": "Auto-attached when a new chat does not pick bases; still adjustable in the composer.",
+    "defaultKnowledgeBasesPlaceholder": "Select knowledge bases",
     "skillPackagesLabel": "Optional Skill Packages",
     "skillPackagesHint": "Mount packages at create time (local storage with root /).",
     "skillPackagesUnsupportedHint": "This sandbox cannot mount skill-package directories directly. You can still use Copy skills to create independent workspace copies.",
@@ -1128,6 +1133,17 @@
     "reasoningDisabled": "Off",
     "reasoningEffort": "Thinking effort",
     "reasoningAlways": "Always reasoning",
+    "conversationMode": "Mode",
+    "conversationModeAsk": "Ask",
+    "conversationModePlan": "Plan",
+    "conversationModeCraft": "Default",
+    "conversationModeStatusCraft": "Default mode — execute efficiently and finish the task.",
+    "conversationModeStatusPlan": "Plan mode — outline the approach first, then execute.",
+    "conversationModeStatusAsk": "Ask mode — answer and analyze without changing the environment.",
+    "planReadyTitle": "Ready to build?",
+    "planReadyHint": "Switch to Default mode and continue with this plan, or keep refining it.",
+    "planReadyExecute": "Execute plan",
+    "planReadyContinue": "Keep planning",
     "errorOccurred": "An error occurred",
     "imageLoadFailed": "Image failed to load",
     "mediaLoadFailed": "Unable to load media",
diff --git a/dashboard/src/locales/zh.json b/dashboard/src/locales/zh.json
index 13ab16228..ac77b8dc7 100644
--- a/dashboard/src/locales/zh.json
+++ b/dashboard/src/locales/zh.json
@@ -824,6 +824,11 @@
     "agentDisable": "停用",
     "defaultModelLabel": "默认模型",
     "defaultModelPlaceholder": "选择一个模型",
+    "defaultConversationModeLabel": "默认对话模式",
+    "defaultConversationModeHint": "新对话的初始权限模式；可在聊天输入栏随时切换。",
+    "defaultKnowledgeBasesLabel": "默认知识库",
+    "defaultKnowledgeBasesHint": "新对话未手动选择知识库时自动附带；仍可在输入栏增减。",
+    "defaultKnowledgeBasesPlaceholder": "选择知识库",
     "skillPackagesLabel": "可选技能包",
     "skillPackagesHint": "创建时挂载选中的技能包（需本地存储，且根目录为 /）。",
     "skillPackagesUnsupportedHint": "当前沙箱无法直接挂载技能包目录，但仍可使用“复制技能”在工作区创建独立副本。",
@@ -1128,6 +1133,17 @@
     "reasoningDisabled": "关闭",
     "reasoningEffort": "思考强度",
     "reasoningAlways": "始终推理",
+    "conversationMode": "模式",
+    "conversationModeAsk": "问答",
+    "conversationModePlan": "计划",
+    "conversationModeCraft": "默认",
+    "conversationModeStatusCraft": "当前为默认模式，可高效执行并完成任务。",
+    "conversationModeStatusPlan": "当前为计划模式，先梳理方案与步骤，再动手执行。",
+    "conversationModeStatusAsk": "当前为问答模式，只回答与分析，不改动环境。",
+    "planReadyTitle": "按计划执行？",
+    "planReadyHint": "确认后将切换到默认模式并带着这份计划继续执行；也可继续改计划。",
+    "planReadyExecute": "按计划执行",
+    "planReadyContinue": "继续改计划",
     "errorOccurred": "出现错误",
     "imageLoadFailed": "图片加载失败",
     "mediaLoadFailed": "无法加载媒体",
diff --git a/dashboard/src/pages/Chat/chatInputCore.partial.less b/dashboard/src/pages/Chat/chatInputCore.partial.less
index 17c7e2e35..2cdafe968 100644
--- a/dashboard/src/pages/Chat/chatInputCore.partial.less
+++ b/dashboard/src/pages/Chat/chatInputCore.partial.less
@@ -28,6 +28,76 @@
   }
 }
 
+.planReadyCard {
+  max-width: var(--chat-column-max, 960px);
+  width: calc(100% - 48px);
+  margin: 0 auto 10px;
+  padding: 12px 14px;
+  border: 1px solid var(--fn-border-secondary, rgba(0, 0, 0, 0.08));
+  border-radius: 12px;
+  background: var(--fn-bg-elevated, #fff);
+
+  @media (max-width: 767px) {
+    width: calc(100% - 24px);
+  }
+}
+
+.planReadyTitle {
+  font-size: 14px;
+  font-weight: 600;
+  color: var(--fn-text-primary);
+  margin-bottom: 4px;
+}
+
+.planReadyHint {
+  margin: 0 0 10px;
+  font-size: 12px;
+  line-height: 1.45;
+  color: var(--fn-text-secondary);
+}
+
+.planReadyBrief {
+  margin: 0 0 12px;
+  padding: 8px 10px;
+  max-height: 160px;
+  overflow: auto;
+  border-radius: 8px;
+  border: 1px solid var(--fn-border-secondary, rgba(0, 0, 0, 0.06));
+  background: var(--fn-bg-layout, #f5f5f5);
+  font-size: 12px;
+  line-height: 1.45;
+  white-space: pre-wrap;
+  word-break: break-word;
+  color: var(--fn-text-secondary);
+  font-family: inherit;
+}
+
+.planReadyActions {
+  display: flex;
+  flex-wrap: wrap;
+  gap: 8px;
+}
+
+.planReadyPrimary,
+.planReadySecondary {
+  border-radius: 8px;
+  padding: 6px 12px;
+  font-size: 13px;
+  cursor: pointer;
+}
+
+.planReadyPrimary {
+  border: none;
+  background: var(--fn-brand, #1677ff);
+  color: #fff;
+}
+
+.planReadySecondary {
+  border: 1px solid var(--fn-border-secondary, rgba(0, 0, 0, 0.12));
+  background: transparent;
+  color: var(--fn-text-primary);
+}
+
 .askQuestionDockInner {
   width: 100%;
   max-width: var(--chat-column-max, 960px);
diff --git a/dashboard/src/pages/Chat/chatInputPickers.partial.less b/dashboard/src/pages/Chat/chatInputPickers.partial.less
index 4721da58a..020a667b5 100644
--- a/dashboard/src/pages/Chat/chatInputPickers.partial.less
+++ b/dashboard/src/pages/Chat/chatInputPickers.partial.less
@@ -169,6 +169,35 @@
   padding-left: 6px;
 }
 
+.conversationModeMenuPanel {
+  width: 260px;
+  min-width: 260px;
+  max-height: none;
+}
+
+.conversationModeStatus {
+  padding: 8px 10px 6px;
+  color: var(--fn-text-tertiary);
+  font-size: 12px;
+  line-height: 1.45;
+}
+
+.conversationModeChoiceLabel {
+  display: inline-flex;
+  align-items: center;
+  gap: 8px;
+  min-width: 0;
+
+  svg {
+    display: block;
+    flex-shrink: 0;
+  }
+
+  span {
+    line-height: 1;
+  }
+}
+
 .reasoningMenuHeader {
   display: flex;
   align-items: center;
diff --git a/dashboard/src/pages/Chat/components/ChatInput.tsx b/dashboard/src/pages/Chat/components/ChatInput.tsx
index fcd03143c..479fee88a 100644
--- a/dashboard/src/pages/Chat/components/ChatInput.tsx
+++ b/dashboard/src/pages/Chat/components/ChatInput.tsx
@@ -19,6 +19,7 @@ import type { KnowledgeBase } from "../../../api/modules/knowledgeBases";
 import type { SkillSpec } from "../../Agent/Skills/useSkills";
 import type { ChatAgentOption } from "./ExpertAgentAvatar";
 import type { AgentSubagentSummary } from "../../../api/modules/subagents";
+import type { ConversationMode } from "../utils/conversationMode";
 import MentionPickerMenu from "./MentionPickerMenu";
 import ChatInputPreviewBar from "./ChatInputPreviewBar";
 import ChatInputActionsRow from "./ChatInputActionsRow";
@@ -86,6 +87,8 @@ interface ChatInputProps {
     mode: "auto" | "enabled" | "disabled",
     effort: string | null,
   ) => void;
+  conversationMode?: ConversationMode;
+  onConversationModeChange?: (mode: ConversationMode) => void;
   availableConnectors?: {
     mcp_server_name: string;
     label: string;
@@ -140,6 +143,8 @@ const ChatInput = forwardRef<ChatInputHandle, ChatInputProps>(
       reasoningMode = "auto",
       reasoningEffort = null,
       onReasoningChange,
+      conversationMode = "craft",
+      onConversationModeChange,
       availableConnectors,
       selectedConnectors = [],
       onConnectorsChange,
@@ -809,6 +814,8 @@ const ChatInput = forwardRef<ChatInputHandle, ChatInputProps>(
             reasoningMode={reasoningMode}
             reasoningEffort={reasoningEffort}
             onReasoningChange={onReasoningChange}
+            conversationMode={conversationMode}
+            onConversationModeChange={onConversationModeChange}
             defaultModel={defaultModel}
             availableConnectors={availableConnectors}
             selectedConnectors={selectedConnectors}
diff --git a/dashboard/src/pages/Chat/components/ChatInputActionsRow.test.tsx b/dashboard/src/pages/Chat/components/ChatInputActionsRow.test.tsx
index ed684f9cf..3f94aa1d9 100644
--- a/dashboard/src/pages/Chat/components/ChatInputActionsRow.test.tsx
+++ b/dashboard/src/pages/Chat/components/ChatInputActionsRow.test.tsx
@@ -1,4 +1,5 @@
-import { fireEvent, render, waitFor } from "@testing-library/react";
+import { fireEvent, render, waitFor, screen } from "@testing-library/react";
+import type { ReactElement } from "react";
 import { MemoryRouter } from "react-router-dom";
 import { describe, expect, it, vi } from "vitest";
 import type { ResolvedModel } from "../../../api/types";
@@ -19,32 +20,38 @@ const models: ResolvedModel[] = [
   },
 ];
 
+const baseProps = {
+  isMobile: false,
+  isStreaming: false,
+  canSend: false,
+  text: "",
+  polishing: false,
+  uploading: false,
+  recording: false,
+  transcribing: false,
+  slashPickerGroups: null as null,
+  slashMenuItems: [] as [],
+  onSlashShortcutSelect: vi.fn(),
+  onFileSelect: vi.fn(),
+  onNewChat: vi.fn(),
+  onPolish: vi.fn(),
+  onToggleVoice: vi.fn(),
+  onCancel: vi.fn(),
+  onSubmit: vi.fn(),
+};
+
+function renderRow(ui: ReactElement) {
+  return render(<MemoryRouter>{ui}</MemoryRouter>);
+}
+
 describe("ChatInputActionsRow compact pickers", () => {
   it("uses a popover instead of a full-width drawer on narrow desktop", async () => {
-    const { container } = render(
-      <MemoryRouter>
-        <ChatInputActionsRow
-          isMobile={false}
-          isStreaming={false}
-          canSend={false}
-          text=""
-          polishing={false}
-          uploading={false}
-          recording={false}
-          transcribing={false}
-          availableModels={models}
-          onModelChange={vi.fn()}
-          slashPickerGroups={null}
-          slashMenuItems={[]}
-          onSlashShortcutSelect={vi.fn()}
-          onFileSelect={vi.fn()}
-          onNewChat={vi.fn()}
-          onPolish={vi.fn()}
-          onToggleVoice={vi.fn()}
-          onCancel={vi.fn()}
-          onSubmit={vi.fn()}
-        />
-      </MemoryRouter>,
+    const { container } = renderRow(
+      <ChatInputActionsRow
+        {...baseProps}
+        availableModels={models}
+        onModelChange={vi.fn()}
+      />,
     );
 
     const modelButton = container
@@ -60,3 +67,32 @@ describe("ChatInputActionsRow compact pickers", () => {
     expect(document.querySelector(".ant-drawer-content")).toBeNull();
   });
 });
+
+describe("ChatInputActionsRow conversation mode (#616 M13)", () => {
+  it("shows status + Default/Plan/Ask labels and notifies on Ask select", async () => {
+    const onConversationModeChange = vi.fn();
+    renderRow(
+      <ChatInputActionsRow
+        {...baseProps}
+        conversationMode="craft"
+        onConversationModeChange={onConversationModeChange}
+      />,
+    );
+
+    fireEvent.click(screen.getByTestId("conversation-mode-trigger"));
+
+    await waitFor(() => {
+      expect(
+        screen.getAllByTestId("conversation-mode-menu").length,
+      ).toBeGreaterThan(0);
+    });
+    const menu = screen.getAllByTestId("conversation-mode-menu")[0]!;
+    expect(menu).toHaveTextContent("当前为默认模式");
+    expect(menu).toHaveTextContent("默认");
+    expect(menu).toHaveTextContent("计划");
+    expect(menu).toHaveTextContent("问答");
+
+    fireEvent.click(screen.getAllByTestId("conversation-mode-option-ask")[0]!);
+    expect(onConversationModeChange).toHaveBeenCalledWith("ask");
+  });
+});
diff --git a/dashboard/src/pages/Chat/components/ChatInputActionsRow.tsx b/dashboard/src/pages/Chat/components/ChatInputActionsRow.tsx
index 3e2834a91..9bf847b55 100644
--- a/dashboard/src/pages/Chat/components/ChatInputActionsRow.tsx
+++ b/dashboard/src/pages/Chat/components/ChatInputActionsRow.tsx
@@ -10,6 +10,7 @@ import {
   Link2,
   Sparkles,
   Wand2,
+  ListTodo,
   Mic,
   CircleDot,
   Play,
@@ -47,6 +48,10 @@ import type { SlashMenuItem } from "../hooks/useSlashMentionInput";
 import { SHORTCUT_ICON_TONE_CLASS } from "../utils/slashShortcutStyles";
 import { isSttAvailable } from "../../../hooks/useVoiceInput";
 import { resolveTurnModelOverride } from "../utils/chatMessages";
+import {
+  COMPOSER_CONVERSATION_MODES,
+  type ConversationMode,
+} from "../utils/conversationMode";
 import {
   mentionedExpertIds,
   mentionedSubagentSlugs,
@@ -96,6 +101,8 @@ interface ChatInputActionsRowProps {
     mode: "auto" | "enabled" | "disabled",
     effort: string | null,
   ) => void;
+  conversationMode?: ConversationMode;
+  onConversationModeChange?: (mode: ConversationMode) => void;
   availableConnectors?: {
     mcp_server_name: string;
     label: string;
@@ -150,6 +157,8 @@ export default function ChatInputActionsRow({
   reasoningMode = "auto",
   reasoningEffort = null,
   onReasoningChange,
+  conversationMode = "craft",
+  onConversationModeChange,
   availableConnectors,
   selectedConnectors = [],
   onConnectorsChange,
@@ -181,6 +190,7 @@ export default function ChatInputActionsRow({
   const [subagentPickerOpen, setSubagentPickerOpen] = useState(false);
   const [connectorPickerOpen, setConnectorPickerOpen] = useState(false);
   const [knowledgePickerOpen, setKnowledgePickerOpen] = useState(false);
+  const [conversationModeOpen, setConversationModeOpen] = useState(false);
   const [shortcutOpen, setShortcutOpen] = useState(false);
   const [modelPickerOpen, setModelPickerOpen] = useState(false);
   const [reasoningModelRef, setReasoningModelRef] = useState<string | null>(
@@ -322,6 +332,72 @@ export default function ChatInputActionsRow({
     setReasoningModelRef(modelRef);
   };
 
+  const conversationModeLabel = (mode: ConversationMode) => {
+    if (mode === "ask") return t("chat.conversationModeAsk", "问答");
+    if (mode === "plan") return t("chat.conversationModePlan", "计划");
+    return t("chat.conversationModeCraft", "默认");
+  };
+
+  const conversationModeStatus = (mode: ConversationMode) => {
+    if (mode === "ask") {
+      return t(
+        "chat.conversationModeStatusAsk",
+        "当前为问答模式，只回答与分析，不改动环境。",
+      );
+    }
+    if (mode === "plan") {
+      return t(
+        "chat.conversationModeStatusPlan",
+        "当前为计划模式，先梳理方案与步骤，再动手执行。",
+      );
+    }
+    return t(
+      "chat.conversationModeStatusCraft",
+      "当前为默认模式，可高效执行并完成任务。",
+    );
+  };
+
+  const conversationModeIcon = (mode: ConversationMode) => {
+    if (mode === "ask") return <MessageSquarePlus size={16} />;
+    if (mode === "plan") return <ListTodo size={16} />;
+    return <Zap size={16} />;
+  };
+
+  const conversationModeMenu = (
+    <div
+      className={`${styles.reasoningMenuPanel} ${styles.conversationModeMenuPanel}`}
+      data-testid="conversation-mode-menu"
+    >
+      <div
+        className={styles.conversationModeStatus}
+        data-testid="conversation-mode-status"
+      >
+        {conversationModeStatus(conversationMode)}
+      </div>
+      <div className={styles.reasoningMenuDivider} />
+      {COMPOSER_CONVERSATION_MODES.map((mode) => (
+        <button
+          key={mode}
+          type="button"
+          data-testid={`conversation-mode-option-${mode}`}
+          className={`${styles.reasoningMenuChoice} ${
+            conversationMode === mode ? styles.reasoningMenuChoiceActive : ""
+          }`}
+          onClick={() => {
+            onConversationModeChange?.(mode);
+            setConversationModeOpen(false);
+          }}
+        >
+          <span className={styles.conversationModeChoiceLabel}>
+            {conversationModeIcon(mode)}
+            <span>{conversationModeLabel(mode)}</span>
+          </span>
+          {conversationMode === mode && <Check size={16} />}
+        </button>
+      ))}
+    </div>
+  );
+
   const reasoningMenu = reasoningModelCapability ? (
     <div className={styles.reasoningMenuPanel}>
       <div className={styles.reasoningMenuHeader}>
@@ -1228,6 +1304,35 @@ export default function ChatInputActionsRow({
             </button>
           </Tooltip>
         )}
+        {onConversationModeChange && (
+          <Popover
+            trigger="click"
+            placement="topLeft"
+            open={conversationModeOpen}
+            onOpenChange={setConversationModeOpen}
+            overlayClassName={styles.skillPickerPopover}
+            content={conversationModeMenu}
+          >
+            <Tooltip
+              title={`${t(
+                "chat.conversationMode",
+                "模式",
+              )}: ${conversationModeLabel(conversationMode)}`}
+              mouseEnterDelay={0.4}
+            >
+              <button
+                className={`${styles.secondaryBtn} ${
+                  conversationMode !== "craft" ? styles.secondaryBtnActive : ""
+                }`}
+                type="button"
+                data-testid="conversation-mode-trigger"
+                aria-label={conversationModeLabel(conversationMode)}
+              >
+                {conversationModeIcon(conversationMode)}
+              </button>
+            </Tooltip>
+          </Popover>
+        )}
         {isStreaming ? (
           canSend ? (
             <Tooltip title={t("chat.queue.action")} mouseEnterDelay={0.4}>
diff --git a/dashboard/src/pages/Chat/components/MessageBubble.tsx b/dashboard/src/pages/Chat/components/MessageBubble.tsx
index 99b1130f6..42f9a0b3d 100644
--- a/dashboard/src/pages/Chat/components/MessageBubble.tsx
+++ b/dashboard/src/pages/Chat/components/MessageBubble.tsx
@@ -21,6 +21,7 @@ import UserMessageComposerTags, {
   type ComposerTagLookups,
 } from "./UserMessageComposerTags";
 import { deriveMessageContent } from "../utils/messageContent";
+import { stripConversationModeUiInstructions } from "../utils/planArtifact";
 import { inferKindFromNameAndMime } from "../utils/chatAttachments";
 import { ChatMediaPlayer } from "./ChatMediaPlayer";
 import { useAuthImageSrc } from "../../../hooks/useAuthImageSrc";
@@ -594,10 +595,16 @@ function MessageBubble({
     [message.errorInfo],
   );
 
-  const { textContent } = useMemo(
+  const { textContent: rawTextContent } = useMemo(
     () => deriveMessageContent(message),
     [message],
   );
+  const textContent = useMemo(() => {
+    if (message.role === "user" || message.toolData || message.hitlData) {
+      return rawTextContent;
+    }
+    return stripConversationModeUiInstructions(rawTextContent);
+  }, [message.role, message.toolData, message.hitlData, rawTextContent]);
   const speechText = useMemo(
     () => prepareSpeechText(textContent),
     [textContent],
diff --git a/dashboard/src/pages/Chat/components/PlanReadyCard.test.tsx b/dashboard/src/pages/Chat/components/PlanReadyCard.test.tsx
new file mode 100644
index 000000000..87bb7d9f8
--- /dev/null
+++ b/dashboard/src/pages/Chat/components/PlanReadyCard.test.tsx
@@ -0,0 +1,28 @@
+import { fireEvent, render, screen } from "@testing-library/react";
+import { describe, expect, it, vi } from "vitest";
+import PlanReadyCard from "./PlanReadyCard";
+
+describe("PlanReadyCard (#616 P6–P8)", () => {
+  it("shows execute and continue actions plus brief preview", () => {
+    const onExecute = vi.fn();
+    const onContinue = vi.fn();
+    render(
+      <PlanReadyCard
+        brief={
+          "## Approved plan\n\nDo it.\n\n1. Write file\n\nExecute this plan now.\n"
+        }
+        onExecute={onExecute}
+        onContinue={onContinue}
+      />,
+    );
+    expect(screen.getByTestId("plan-ready-card")).toBeInTheDocument();
+    expect(screen.getByTestId("plan-ready-brief")).toHaveTextContent("Do it.");
+    expect(screen.getByTestId("plan-ready-brief")).toHaveTextContent(
+      "Write file",
+    );
+    fireEvent.click(screen.getByTestId("plan-ready-execute"));
+    expect(onExecute).toHaveBeenCalledTimes(1);
+    fireEvent.click(screen.getByTestId("plan-ready-continue"));
+    expect(onContinue).toHaveBeenCalledTimes(1);
+  });
+});
diff --git a/dashboard/src/pages/Chat/components/PlanReadyCard.tsx b/dashboard/src/pages/Chat/components/PlanReadyCard.tsx
new file mode 100644
index 000000000..a340b8ca4
--- /dev/null
+++ b/dashboard/src/pages/Chat/components/PlanReadyCard.tsx
@@ -0,0 +1,55 @@
+import { useTranslation } from "react-i18next";
+import styles from "../index.module.less";
+import { planBriefPreview } from "../utils/planArtifact";
+
+interface PlanReadyCardProps {
+  brief: string;
+  onExecute: () => void;
+  onContinue: () => void;
+}
+
+/** Cursor/WorkBuddy-style Plan → Craft checkpoint (#616 P6). */
+export default function PlanReadyCard({
+  brief,
+  onExecute,
+  onContinue,
+}: PlanReadyCardProps) {
+  const { t } = useTranslation();
+  const preview = planBriefPreview(brief);
+  return (
+    <div className={styles.planReadyCard} data-testid="plan-ready-card">
+      <div className={styles.planReadyTitle}>
+        {t("chat.planReadyTitle", "按计划执行？")}
+      </div>
+      <p className={styles.planReadyHint}>
+        {t(
+          "chat.planReadyHint",
+          "确认后将切换到默认模式并带着这份计划继续执行；也可继续改计划。",
+        )}
+      </p>
+      {preview ? (
+        <pre className={styles.planReadyBrief} data-testid="plan-ready-brief">
+          {preview}
+        </pre>
+      ) : null}
+      <div className={styles.planReadyActions}>
+        <button
+          type="button"
+          className={styles.planReadyPrimary}
+          data-testid="plan-ready-execute"
+          onClick={onExecute}
+        >
+          {t("chat.planReadyExecute", "按计划执行")}
+        </button>
+        <button
+          type="button"
+          className={styles.planReadySecondary}
+          data-testid="plan-ready-continue"
+          onClick={onContinue}
+        >
+          {t("chat.planReadyContinue", "继续改计划")}
+        </button>
+      </div>
+    </div>
+  );
+}
diff --git a/dashboard/src/pages/Chat/hooks/chatStore.ts b/dashboard/src/pages/Chat/hooks/chatStore.ts
index 48cea7db1..e1d250500 100644
--- a/dashboard/src/pages/Chat/hooks/chatStore.ts
+++ b/dashboard/src/pages/Chat/hooks/chatStore.ts
@@ -18,6 +18,7 @@ import type {
   SessionSnapshot,
   SessionStreamState,
 } from "./sseHelpers";
+import type { ConversationMode } from "../utils/conversationMode";
 import {
   parseHarnessChunk,
   type HarnessChunk,
@@ -1851,6 +1852,8 @@ async function sendTurnWebSocket(
   onStreamEnd?: () => void,
   reasoningMode?: "auto" | "enabled" | "disabled",
   reasoningEffort?: string | null,
+  conversationMode?: ConversationMode,
+  planBrief?: string,
 ): Promise<boolean> {
   const state = getOrCreate(sessionId);
   const resolvedThreadId = (threadId || sessionId).trim();
@@ -1928,6 +1931,8 @@ async function sendTurnWebSocket(
       }
       if (reasoningMode) payload.reasoning_mode = reasoningMode;
       if (reasoningEffort) payload.reasoning_effort = reasoningEffort;
+      if (conversationMode) payload.conversation_mode = conversationMode;
+      if (planBrief?.trim()) payload.plan_brief = planBrief.trim();
       ws.send(JSON.stringify(payload));
     };
 
@@ -2052,6 +2057,8 @@ export async function sendTurn(
   targetAgentIds?: string[] | null,
   reasoningMode?: "auto" | "enabled" | "disabled",
   reasoningEffort?: string | null,
+  conversationMode?: ConversationMode,
+  planBrief?: string,
 ): Promise<void> {
   const state = getOrCreate(sessionId);
 
@@ -2123,6 +2130,8 @@ export async function sendTurn(
     onStreamEnd,
     reasoningMode,
     reasoningEffort,
+    conversationMode,
+    planBrief,
   );
   if (!wsOk) {
     state.messages = [
diff --git a/dashboard/src/pages/Chat/hooks/sseHelpers.ts b/dashboard/src/pages/Chat/hooks/sseHelpers.ts
index 9933981a9..799d8fe63 100644
--- a/dashboard/src/pages/Chat/hooks/sseHelpers.ts
+++ b/dashboard/src/pages/Chat/hooks/sseHelpers.ts
@@ -11,6 +11,7 @@ import type {
   TokenUsage,
 } from "../../../api/types";
 import type { ContentBlockItem } from "../../../utils/messageParser";
+import type { ConversationMode } from "../utils/conversationMode";
 
 export interface ToolCallData {
   name?: string;
@@ -53,6 +54,7 @@ export interface UserComposerContext {
   model?: string;
   reasoningMode?: "auto" | "enabled" | "disabled";
   reasoningEffort?: string | null;
+  conversationMode?: ConversationMode;
 }
 
 export interface ChatMessage {
@@ -71,6 +73,8 @@ export interface ChatMessage {
   errorInfo?: ProcessErrorInfo;
   status?: "streaming" | "done" | "error";
   timestamp: number;
+  /** Plan→Craft silent handoff — omit from chat transcript UI. */
+  uiHidden?: boolean;
 }
 
 /** Per-session state held in the chat store's module-scoped Map. */
diff --git a/dashboard/src/pages/Chat/hooks/useChat.ts b/dashboard/src/pages/Chat/hooks/useChat.ts
index 259682744..30e40a44b 100644
--- a/dashboard/src/pages/Chat/hooks/useChat.ts
+++ b/dashboard/src/pages/Chat/hooks/useChat.ts
@@ -24,6 +24,8 @@ import {
   type ContentBlock,
 } from "../../../utils/messageParser";
 import { normalizeComposerContext } from "../utils/chatMessages";
+import { PLAN_EXECUTE_USER_TRIGGER } from "../utils/planArtifact";
+import type { ConversationMode } from "../utils/conversationMode";
 import { resolveMessageTimestampMs } from "../../../utils/formatMessageTime";
 import { inferKindFromNameAndMime } from "../utils/chatAttachments";
 import {
@@ -442,6 +444,12 @@ function convertCallEntries(entries: CallEntry[]): ChatMessage[] {
                 ?.composer_context,
             )
           : undefined,
+      uiHidden:
+        entry.role === "user" &&
+        Boolean(
+          (entry.metadata as Record<string, unknown> | null | undefined)
+            ?.ui_hidden,
+        ),
       toolData: tool?.data,
       usage: normalizeTokenUsage(entry.usage ?? undefined) ?? undefined,
       metadata: normalizeMessageMetadata(entry.metadata ?? undefined),
@@ -637,6 +645,15 @@ function toHistoryContentBlocks(content: unknown): unknown[] {
 }
 
 function isDisplayableHistoryMessage(message: ChatMessage): boolean {
+  if (message.uiHidden) return false;
+  // Silent Plan→Craft trigger (in case older servers omitted ui_hidden).
+  if (
+    message.role === "user" &&
+    message.content.trim() === PLAN_EXECUTE_USER_TRIGGER &&
+    !(message.attachments && message.attachments.length > 0)
+  ) {
+    return false;
+  }
   if (message.hitlData) return true;
   if (message.toolData) return true;
   if (message.attachments && message.attachments.length > 0) return true;
@@ -653,6 +670,7 @@ export function convertHistoryMessages(
     timestamp?: number;
     composer_context?: unknown;
     inbound_attachments?: unknown;
+    ui_hidden?: unknown;
     status?: string;
     error_code?: string;
   }>,
@@ -666,6 +684,9 @@ export function convertHistoryMessages(
     if (message.inbound_attachments) {
       meta.inbound_attachments = message.inbound_attachments;
     }
+    if (message.ui_hidden) {
+      meta.ui_hidden = true;
+    }
     return {
       message_id: message.id,
       entry_index: `hist-${index}`,
@@ -831,20 +852,25 @@ export function useChat(
       composerContext?: UserComposerContext,
       reasoningMode?: "auto" | "enabled" | "disabled",
       reasoningEffort?: string | null,
+      conversationMode?: ConversationMode,
+      planBrief?: string,
+      hideUserMessage?: boolean,
     ) => {
       const key = storeKey || stableSessionId;
 
-      const userMsg: ChatMessage = {
-        id: generateId(),
-        role: "user",
-        content: text,
-        attachments:
-          attachments && attachments.length > 0 ? attachments : undefined,
-        composerContext,
-        status: "done",
-        timestamp: Date.now(),
-      };
-      chatStore.appendUserMessage(key, userMsg);
+      if (!hideUserMessage) {
+        const userMsg: ChatMessage = {
+          id: generateId(),
+          role: "user",
+          content: text,
+          attachments:
+            attachments && attachments.length > 0 ? attachments : undefined,
+          composerContext,
+          status: "done",
+          timestamp: Date.now(),
+        };
+        chatStore.appendUserMessage(key, userMsg);
+      }
 
       const threadIdForApi =
         storeKey ||
@@ -863,6 +889,8 @@ export function useChat(
         targetAgentIds,
         reasoningMode,
         reasoningEffort,
+        conversationMode,
+        planBrief,
       );
     },
     [stableSessionId],
diff --git a/dashboard/src/pages/Chat/hooks/useChatComposerResources.ts b/dashboard/src/pages/Chat/hooks/useChatComposerResources.ts
index 6d172ca71..0ae4c8cfe 100644
--- a/dashboard/src/pages/Chat/hooks/useChatComposerResources.ts
+++ b/dashboard/src/pages/Chat/hooks/useChatComposerResources.ts
@@ -24,7 +24,13 @@ import {
   peekPendingAttachKnowledgeBaseId,
 } from "../utils/pendingAttachKnowledgeBase";
 import { withDefaultOpenKnowledgeBases } from "../utils/withDefaultOpenKnowledgeBases";
+import { PENDING_THREAD_ID } from "../constants";
 import { isPendingThread } from "./useSessions";
+import {
+  parseConversationMode,
+  type ConversationMode,
+} from "../utils/conversationMode";
+import { resolveSeedConversationMode } from "../utils/seedConversationMode";
 
 export function useChatComposerResources(
   resolvedAgentId: string | null | undefined,
@@ -32,15 +38,15 @@ export function useChatComposerResources(
   stickyModel?: string | null,
   stickyReasoningMode?: "auto" | "enabled" | "disabled" | null,
   stickyReasoningEffort?: string | null,
+  defaultConversationMode?: ConversationMode | string | null,
+  defaultKnowledgeBaseIds?: string[] | null,
 ) {
   const user = useCurrentUser();
   const currentUserId = user?.id ?? null;
   const { agents } = useAgent();
   const expert = agents.find((item) => item.agent_id === resolvedAgentId);
   const expertMcpServers = expert?.mcp_servers;
-  const expertKnowledgeBaseIds = expert?.knowledge_base_ids;
   const expertMcpKey = (expertMcpServers ?? []).join("\0");
-  const expertKbKey = (expertKnowledgeBaseIds ?? []).join("\0");
   const isNewSession = !activeThreadId || isPendingThread(activeThreadId);
   const composerTouchedRef = useRef(false);
   const [selectedConnectors, setSelectedConnectors] = useState<string[]>([]);
@@ -72,6 +78,9 @@ export function useChatComposerResources(
     "auto" | "enabled" | "disabled"
   >("auto");
   const [reasoningEffort, setReasoningEffort] = useState<string | null>(null);
+  const [conversationMode, setConversationMode] = useState<ConversationMode>(
+    () => parseConversationMode(defaultConversationMode),
+  );
   const [conversationOverrides, setConversationOverrides] = useState<
     Record<
       string,
@@ -79,17 +88,80 @@ export function useChatComposerResources(
         model: string | null;
         mode: "auto" | "enabled" | "disabled";
         effort: string | null;
+        conversationMode?: ConversationMode;
       }
     >
   >({});
 
-  useEffect(() => {
-    composerTouchedRef.current = false;
-  }, [resolvedAgentId]);
+  // Keep refs so thread-id transitions can stamp the composer mode the user
+  // already picked on an empty / pending chat (PlanReady P0).
+  const conversationModeRef = useRef(conversationMode);
+  conversationModeRef.current = conversationMode;
+  const selectedModelRef = useRef(selectedModel);
+  selectedModelRef.current = selectedModel;
+  const reasoningModeRef = useRef(reasoningMode);
+  reasoningModeRef.current = reasoningMode;
+  const reasoningEffortRef = useRef(reasoningEffort);
+  reasoningEffortRef.current = reasoningEffort;
+  const prevThreadIdRef = useRef(activeThreadId);
 
+  // Seed from agent default, then restore any per-thread override (#616 OCR).
+  // When a new chat first gets a thread id (empty → pending → real), preserve
+  // the composer selection instead of wiping it back to the agent default —
+  // otherwise Plan turns never show PlanReady on the first send.
   useEffect(() => {
-    if (isNewSession) composerTouchedRef.current = false;
-  }, [isNewSession]);
+    const prevThreadId = prevThreadIdRef.current;
+    prevThreadIdRef.current = activeThreadId;
+    const decision = resolveSeedConversationMode({
+      activeThreadId,
+      previousThreadId: prevThreadId,
+      override: activeThreadId
+        ? conversationOverrides[activeThreadId]?.conversationMode
+        : undefined,
+      agentDefault: defaultConversationMode,
+      currentComposerMode: conversationModeRef.current,
+    });
+
+    if (decision.action === "set") {
+      setConversationMode(decision.mode);
+      // Drop the reusable pending bucket so the next new-chat first-send does
+      // not revive a previous Plan/Ask selection.
+      if (!activeThreadId) {
+        setConversationOverrides((current) => {
+          if (!(PENDING_THREAD_ID in current)) return current;
+          const next = { ...current };
+          delete next[PENDING_THREAD_ID];
+          return next;
+        });
+      }
+      return;
+    }
+
+    if (decision.action === "stamp-override" && activeThreadId) {
+      const mode = decision.mode;
+      setConversationMode(mode);
+      setConversationOverrides((current) => {
+        const prev = current[activeThreadId];
+        if (prev?.conversationMode === mode) {
+          return current;
+        }
+        return {
+          ...current,
+          [activeThreadId]: {
+            model: prev?.model ?? selectedModelRef.current,
+            mode: prev?.mode ?? reasoningModeRef.current,
+            effort: prev?.effort ?? reasoningEffortRef.current,
+            conversationMode: mode,
+          },
+        };
+      });
+    }
+  }, [
+    resolvedAgentId,
+    activeThreadId,
+    defaultConversationMode,
+    conversationOverrides,
+  ]);
 
   // Auto = omit turn model; backend applies the expert default.
   useEffect(() => {
@@ -211,6 +283,9 @@ export function useChatComposerResources(
         return knowledgeBasesApi.list().then((bases) => {
           if (cancelled) return;
           setChatKnowledgeBases(bases);
+          const agentDefaults = (defaultKnowledgeBaseIds ?? []).filter((id) =>
+            bases.some((base) => base.id === id),
+          );
           const ownedDefaults = bases
             .filter(
               (base) =>
@@ -219,34 +294,16 @@ export function useChatComposerResources(
                 base.owner_user_id === currentUserId,
             )
             .map((base) => base.id);
-          const allowed = new Set(bases.map((base) => base.id));
-          const defaults = withDefaultOpenKnowledgeBases(
-            ownedDefaults,
-            (expertKnowledgeBaseIds ?? []).filter((id) => allowed.has(id)),
-          );
-          setSelectedKnowledgeBaseIds((previous) => {
-            if (composerTouchedRef.current) {
-              return pendingId &&
-                allowed.has(pendingId) &&
-                !previous.includes(pendingId)
-                ? [...previous, pendingId]
-                : previous;
-            }
-            if (isNewSession) {
-              return withDefaultOpenKnowledgeBases(
-                pendingId && allowed.has(pendingId) ? [pendingId] : [],
-                defaults,
-              );
-            }
-            return withDefaultOpenKnowledgeBases(
-              pendingId &&
-                allowed.has(pendingId) &&
-                !previous.includes(pendingId)
+          const seedDefaults =
+            agentDefaults.length > 0 ? agentDefaults : ownedDefaults;
+          setSelectedKnowledgeBaseIds((previous) =>
+            withDefaultOpenKnowledgeBases(
+              pendingId && !previous.includes(pendingId)
                 ? [...previous, pendingId]
                 : previous,
-              ownedDefaults,
-            );
-          });
+              seedDefaults,
+            ),
+          );
           if (pendingId) consumePendingAttachKnowledgeBaseId();
         });
       })
@@ -256,7 +313,7 @@ export function useChatComposerResources(
     return () => {
       cancelled = true;
     };
-  }, [resolvedAgentId, currentUserId, isNewSession, expertKbKey]);
+  }, [resolvedAgentId, currentUserId, defaultKnowledgeBaseIds]);
 
   useEffect(() => {
     let cancelled = false;
@@ -330,6 +387,26 @@ export function useChatComposerResources(
     setSelectedKnowledgeBaseIds(ids);
   }, []);
 
+  const handleConversationModeChange = useCallback(
+    (mode: ConversationMode) => {
+      setConversationMode(mode);
+      if (!activeThreadId) return;
+      setConversationOverrides((current) => {
+        const prev = current[activeThreadId];
+        return {
+          ...current,
+          [activeThreadId]: {
+            model: prev?.model ?? selectedModel,
+            mode: prev?.mode ?? reasoningMode,
+            effort: prev?.effort ?? reasoningEffort,
+            conversationMode: mode,
+          },
+        };
+      });
+    },
+    [activeThreadId, selectedModel, reasoningMode, reasoningEffort],
+  );
+
   const handleModelChange = useCallback(
     (model: string | null) => {
       setSelectedModel(model);
@@ -342,14 +419,18 @@ export function useChatComposerResources(
       setReasoningMode(nextMode);
       setReasoningEffort(nextEffort);
       if (activeThreadId) {
-        setConversationOverrides((current) => ({
-          ...current,
-          [activeThreadId]: {
-            model,
-            mode: nextMode,
-            effort: nextEffort,
-          },
-        }));
+        setConversationOverrides((current) => {
+          const prev = current[activeThreadId];
+          return {
+            ...current,
+            [activeThreadId]: {
+              model,
+              mode: nextMode,
+              effort: nextEffort,
+              conversationMode: prev?.conversationMode,
+            },
+          };
+        });
       }
       if (
         resolvedAgentId &&
@@ -371,14 +452,18 @@ export function useChatComposerResources(
       setReasoningMode(mode);
       setReasoningEffort(effort);
       if (activeThreadId) {
-        setConversationOverrides((current) => ({
-          ...current,
-          [activeThreadId]: {
-            model: selectedModel,
-            mode,
-            effort,
-          },
-        }));
+        setConversationOverrides((current) => {
+          const prev = current[activeThreadId];
+          return {
+            ...current,
+            [activeThreadId]: {
+              model: selectedModel,
+              mode,
+              effort,
+              conversationMode: prev?.conversationMode,
+            },
+          };
+        });
       }
       if (
         resolvedAgentId &&
@@ -400,6 +485,8 @@ export function useChatComposerResources(
     reasoningMode,
     reasoningEffort,
     handleReasoningChange,
+    conversationMode,
+    handleConversationModeChange,
     selectedConnectors,
     selectedKnowledgeBaseIds,
     chatConnectors,
diff --git a/dashboard/src/pages/Chat/hooks/useChatSend.ts b/dashboard/src/pages/Chat/hooks/useChatSend.ts
index 9273e6b1a..d3a8ece25 100644
--- a/dashboard/src/pages/Chat/hooks/useChatSend.ts
+++ b/dashboard/src/pages/Chat/hooks/useChatSend.ts
@@ -13,6 +13,7 @@ import {
   buildUserMessage,
   resolveTurnModelRef,
 } from "../utils/chatMessages";
+import type { ConversationMode } from "../utils/conversationMode";
 
 interface UseChatSendParams {
   resolvedAgentId: string | null | undefined;
@@ -25,6 +26,7 @@ interface UseChatSendParams {
   selectedTargetAgents?: string[];
   reasoningMode: "auto" | "enabled" | "disabled";
   reasoningEffort: string | null;
+  conversationMode: ConversationMode;
   defaultModel?: string | null;
   sendMessage: (
     text: string,
@@ -39,6 +41,9 @@ interface UseChatSendParams {
     composerContext?: UserComposerContext,
     reasoningMode?: "auto" | "enabled" | "disabled",
     reasoningEffort?: string | null,
+    conversationMode?: ConversationMode,
+    planBrief?: string,
+    hideUserMessage?: boolean,
   ) => void;
   createSession: () => { session: Session; resolvedId: Promise<string> };
   renameSession: (id: string, name: string) => void;
@@ -59,6 +64,11 @@ export type ChatSendOverrides = {
   selectedTargetAgents?: string[];
   composerContext?: UserComposerContext;
   modelRef?: string | null;
+  conversationMode?: ConversationMode;
+  /** Internal plan brief for Craft handoff — not shown as chat text. */
+  planBrief?: string;
+  /** Do not append a local user bubble (PlanReady execute). */
+  hideUserMessage?: boolean;
   /** Send to this thread instead of the active one (queued flush). */
   threadId?: string | null;
   /** Send as this agent instead of the active one (queued flush). */
@@ -76,6 +86,7 @@ export function useChatSend({
   selectedTargetAgents = [],
   reasoningMode,
   reasoningEffort,
+  conversationMode,
   defaultModel,
   sendMessage,
   createSession,
@@ -99,9 +110,14 @@ export function useChatSend({
       }
 
       const trimmed = text.trim();
-      if (!trimmed && !(attachments && attachments.length > 0)) return false;
+      const planBrief = overrides?.planBrief?.trim() || "";
+      const hideUserMessage = Boolean(overrides?.hideUserMessage);
+      if (!trimmed && !(attachments && attachments.length > 0) && !planBrief) {
+        return false;
+      }
 
       const maybeRenameNewThread = (tid: string, hadMessages: boolean) => {
+        if (!trimmed) return;
         const current = sessions.find((s) => s.id === tid);
         if (current?.name === "New Chat" && !hadMessages) {
           renameSession(tid, deriveThreadTitle(trimmed));
@@ -129,6 +145,10 @@ export function useChatSend({
             overrides?.composerContext?.reasoningMode ?? reasoningMode,
           reasoningEffort:
             overrides?.composerContext?.reasoningEffort ?? reasoningEffort,
+          conversationMode:
+            overrides?.composerContext?.conversationMode ??
+            overrides?.conversationMode ??
+            conversationMode,
         });
 
       const modelOverride =
@@ -151,6 +171,9 @@ export function useChatSend({
           composerContext,
           composerContext?.reasoningMode ?? reasoningMode,
           composerContext?.reasoningEffort ?? reasoningEffort,
+          composerContext?.conversationMode ?? conversationMode,
+          overrides?.planBrief,
+          hideUserMessage,
         );
       };
 
@@ -170,6 +193,11 @@ export function useChatSend({
         return true;
       }
 
+      // Silent plan execute requires an existing thread (no empty-bubble draft).
+      if (hideUserMessage || planBrief) {
+        return false;
+      }
+
       const userMsg = buildUserMessage(trimmed, attachments, composerContext);
       chatStore.appendUserMessage(EMPTY_CHAT_SESSION_KEY, userMsg);
 
@@ -210,6 +238,8 @@ export function useChatSend({
           targetAgents,
           composerContext?.reasoningMode ?? reasoningMode,
           composerContext?.reasoningEffort ?? reasoningEffort,
+          composerContext?.conversationMode ?? conversationMode,
+          overrides?.planBrief,
         );
         navigate(`/chat/${agent}/${tid}`, { replace: true });
       });
@@ -230,6 +260,7 @@ export function useChatSend({
       selectedTargetAgents,
       reasoningMode,
       reasoningEffort,
+      conversationMode,
       defaultModel,
       t,
     ],
diff --git a/dashboard/src/pages/Chat/index.tsx b/dashboard/src/pages/Chat/index.tsx
index 836d419fb..758d6d1c5 100644
--- a/dashboard/src/pages/Chat/index.tsx
+++ b/dashboard/src/pages/Chat/index.tsx
@@ -81,6 +81,12 @@ import ChatSidebarPanel from "./components/ChatSidebarPanel";
 import ChatTitleBar from "./components/ChatTitleBar";
 import ChatComposerChrome from "./components/ChatComposerChrome";
 import AskQuestionCard from "./components/AskQuestionCard";
+import PlanReadyCard from "./components/PlanReadyCard";
+import {
+  buildPlanBriefFromMessages,
+  planContinueHandoff,
+  planExecuteHandoff,
+} from "./utils/planArtifact";
 import { extractAskQuestions, isAskHitl } from "../../api/types/hitl";
 import { isAgentChatReady } from "../../utils/agentError";
 import { useMemoryMaintenance } from "./hooks/useMemoryMaintenance";
@@ -441,6 +447,8 @@ function ChatPageInner() {
     reasoningMode,
     reasoningEffort,
     handleReasoningChange,
+    conversationMode,
+    handleConversationModeChange,
     handleConnectorsChange,
     handleKnowledgeBaseIdsChange,
   } = useChatComposerResources(
@@ -449,6 +457,8 @@ function ChatPageInner() {
     composerSession?.modelRef,
     composerSession?.reasoningMode,
     composerSession?.reasoningEffort,
+    activeAgent?.default_conversation_mode,
+    activeAgent?.default_knowledge_base_ids,
   );
 
   const { contextMaxTokens, contextUsedTokens } = useChatContextWindow(
@@ -546,6 +556,7 @@ function ChatPageInner() {
     selectedKnowledgeBaseIds,
     reasoningMode,
     reasoningEffort,
+    conversationMode,
     defaultModel: activeAgent?.default_model ?? null,
     sendMessage,
     createSession,
@@ -569,7 +580,13 @@ function ChatPageInner() {
     setBrowserLastRecordingId,
   });
 
-  // Wrap handleSend to intercept skill recording workflow keywords
+  const [planReadyBrief, setPlanReadyBrief] = useState<string | null>(null);
+  const wasStreamingRef = useRef(false);
+  const planModeTurnRef = useRef(false);
+
+  // Wrap handleSend to intercept skill recording workflow keywords.
+  // Also latch plan-mode at send time so PlanReady still appears if the
+  // composer reseeds while the new thread id is assigned mid-turn.
   const wrappedHandleSend = useCallback(
     (
       text: string,
@@ -580,11 +597,56 @@ function ChatPageInner() {
         // The workflow intercepted the message — don't send it to the agent
         return;
       }
+      const modeForTurn =
+        overrides?.conversationMode ??
+        overrides?.composerContext?.conversationMode ??
+        conversationMode;
+      if (modeForTurn === "plan") {
+        planModeTurnRef.current = true;
+      }
       handleSend(text, attachments, overrides);
     },
-    [interceptUserMessage, handleSend],
+    [interceptUserMessage, handleSend, conversationMode],
   );
 
+  useEffect(() => {
+    if (conversationMode === "plan" && isStreaming) {
+      planModeTurnRef.current = true;
+    }
+  }, [conversationMode, isStreaming]);
+
+  useEffect(() => {
+    const wasStreaming = wasStreamingRef.current;
+    wasStreamingRef.current = isStreaming;
+    if (!wasStreaming || isStreaming) return;
+    if (!planModeTurnRef.current) return;
+    planModeTurnRef.current = false;
+    const brief = buildPlanBriefFromMessages(messages);
+    if (brief) setPlanReadyBrief(brief);
+  }, [isStreaming, messages]);
+
+  useEffect(() => {
+    setPlanReadyBrief(null);
+  }, [activeThreadId, resolvedAgentId]);
+
+  const handlePlanExecute = useCallback(() => {
+    if (!planReadyBrief) return;
+    const brief = planReadyBrief;
+    setPlanReadyBrief(null);
+    const handoff = planExecuteHandoff(brief);
+    handleConversationModeChange(handoff.conversationMode);
+    wrappedHandleSend(handoff.text, undefined, {
+      conversationMode: handoff.conversationMode,
+      planBrief: handoff.planBrief,
+      hideUserMessage: handoff.hideUserMessage,
+    });
+  }, [planReadyBrief, handleConversationModeChange, wrappedHandleSend]);
+
+  const handlePlanContinue = useCallback(() => {
+    setPlanReadyBrief(null);
+    handleConversationModeChange(planContinueHandoff().conversationMode);
+  }, [handleConversationModeChange]);
+
   const flushQueuedItem = useCallback(
     (item: QueuedChatItem, ctx: ChatQueueFlushContext): boolean => {
       if (!ctx.threadId) {
@@ -1351,6 +1413,13 @@ function ChatPageInner() {
                 </div>
               </div>
             ) : null}
+            {planReadyBrief && !pendingAsk && !isStreaming ? (
+              <PlanReadyCard
+                brief={planReadyBrief}
+                onExecute={handlePlanExecute}
+                onContinue={handlePlanContinue}
+              />
+            ) : null}
             <ChatInput
               ref={chatInputRef}
               onSend={wrappedHandleSend}
@@ -1372,6 +1441,8 @@ function ChatPageInner() {
               reasoningMode={reasoningMode}
               reasoningEffort={reasoningEffort}
               onReasoningChange={handleReasoningChange}
+              conversationMode={conversationMode}
+              onConversationModeChange={handleConversationModeChange}
               availableConnectors={chatConnectors}
               selectedConnectors={selectedConnectors}
               onConnectorsChange={handleConnectorsChange}
diff --git a/dashboard/src/pages/Chat/utils/chatMessages.conversationMode.test.ts b/dashboard/src/pages/Chat/utils/chatMessages.conversationMode.test.ts
new file mode 100644
index 000000000..6edc31bd4
--- /dev/null
+++ b/dashboard/src/pages/Chat/utils/chatMessages.conversationMode.test.ts
@@ -0,0 +1,26 @@
+import { describe, expect, it } from "vitest";
+import { buildComposerContext, normalizeComposerContext } from "./chatMessages";
+
+describe("buildComposerContext conversationMode", () => {
+  it("includes conversationMode in composer context (M15 payload seed)", () => {
+    const ctx = buildComposerContext({
+      conversationMode: "ask",
+    });
+    expect(ctx).toEqual({ conversationMode: "ask" });
+  });
+
+  it("omits conversationMode when unset", () => {
+    const ctx = buildComposerContext({ skills: ["a"] });
+    expect(ctx?.conversationMode).toBeUndefined();
+  });
+});
+
+describe("normalizeComposerContext conversationMode", () => {
+  it("keeps conversationMode from history payloads", () => {
+    const ctx = normalizeComposerContext({
+      skills: ["a"],
+      conversationMode: "plan",
+    });
+    expect(ctx).toEqual({ skills: ["a"], conversationMode: "plan" });
+  });
+});
diff --git a/dashboard/src/pages/Chat/utils/chatMessages.ts b/dashboard/src/pages/Chat/utils/chatMessages.ts
index 552362cbd..2f5d638f9 100644
--- a/dashboard/src/pages/Chat/utils/chatMessages.ts
+++ b/dashboard/src/pages/Chat/utils/chatMessages.ts
@@ -5,6 +5,7 @@ import type {
   ChatMessage,
   UserComposerContext,
 } from "../hooks/sseHelpers";
+import type { ConversationMode } from "./conversationMode";
 
 export function normalizeComposerContext(
   value: unknown,
@@ -48,6 +49,14 @@ export function normalizeComposerContext(
     ctx.reasoningEffort = raw.reasoningEffort.trim();
     has = true;
   }
+  if (
+    raw.conversationMode === "ask" ||
+    raw.conversationMode === "plan" ||
+    raw.conversationMode === "craft"
+  ) {
+    ctx.conversationMode = raw.conversationMode;
+    has = true;
+  }
 
   return has ? ctx : undefined;
 }
@@ -88,6 +97,7 @@ export function buildComposerContext(params: {
   selectedModel?: string | null;
   reasoningMode?: "auto" | "enabled" | "disabled";
   reasoningEffort?: string | null;
+  conversationMode?: ConversationMode;
 }): UserComposerContext | undefined {
   const ctx: UserComposerContext = {};
   let has = false;
@@ -122,6 +132,10 @@ export function buildComposerContext(params: {
     ctx.reasoningEffort = params.reasoningEffort;
     has = true;
   }
+  if (params.conversationMode) {
+    ctx.conversationMode = params.conversationMode;
+    has = true;
+  }
 
   return has ? ctx : undefined;
 }
diff --git a/dashboard/src/pages/Chat/utils/conversationMode.test.ts b/dashboard/src/pages/Chat/utils/conversationMode.test.ts
new file mode 100644
index 000000000..f3619e02f
--- /dev/null
+++ b/dashboard/src/pages/Chat/utils/conversationMode.test.ts
@@ -0,0 +1,19 @@
+import { describe, expect, it } from "vitest";
+import {
+  DEFAULT_CONVERSATION_MODE,
+  parseConversationMode,
+} from "./conversationMode";
+
+describe("parseConversationMode", () => {
+  it("accepts ask / plan / craft", () => {
+    expect(parseConversationMode("ask")).toBe("ask");
+    expect(parseConversationMode("plan")).toBe("plan");
+    expect(parseConversationMode("craft")).toBe("craft");
+  });
+
+  it("defaults unknown / null to craft", () => {
+    expect(parseConversationMode(null)).toBe(DEFAULT_CONVERSATION_MODE);
+    expect(parseConversationMode(undefined)).toBe("craft");
+    expect(parseConversationMode("agent")).toBe("craft");
+  });
+});
diff --git a/dashboard/src/pages/Chat/utils/conversationMode.ts b/dashboard/src/pages/Chat/utils/conversationMode.ts
new file mode 100644
index 000000000..54cd3bb77
--- /dev/null
+++ b/dashboard/src/pages/Chat/utils/conversationMode.ts
@@ -0,0 +1,46 @@
+/** Conversation permission modes (#616): Ask / Plan / Craft. */
+
+import type { TFunction } from "i18next";
+
+export type ConversationMode = "ask" | "plan" | "craft";
+
+export const DEFAULT_CONVERSATION_MODE: ConversationMode = "craft";
+
+const VALID: ReadonlySet<string> = new Set(["ask", "plan", "craft"]);
+
+/** Composer options — default-first (WorkBuddy-style menu). */
+export const COMPOSER_CONVERSATION_MODES: readonly ConversationMode[] = [
+  "craft",
+  "plan",
+  "ask",
+] as const;
+
+export function parseConversationMode(
+  value: unknown,
+  fallback: ConversationMode = DEFAULT_CONVERSATION_MODE,
+): ConversationMode {
+  if (typeof value === "string" && VALID.has(value)) {
+    return value as ConversationMode;
+  }
+  return fallback;
+}
+
+/** Ant Design Select options shared by Experts create/edit drawers. */
+export function conversationModeSelectOptions(
+  t: TFunction,
+): Array<{ value: ConversationMode; label: string }> {
+  return [
+    {
+      value: "craft",
+      label: t("chat.conversationModeCraft", "默认"),
+    },
+    {
+      value: "plan",
+      label: t("chat.conversationModePlan", "计划"),
+    },
+    {
+      value: "ask",
+      label: t("chat.conversationModeAsk", "问答"),
+    },
+  ];
+}
diff --git a/dashboard/src/pages/Chat/utils/planArtifact.test.ts b/dashboard/src/pages/Chat/utils/planArtifact.test.ts
new file mode 100644
index 000000000..c81bb7ef1
--- /dev/null
+++ b/dashboard/src/pages/Chat/utils/planArtifact.test.ts
@@ -0,0 +1,186 @@
+import { describe, expect, it } from "vitest";
+import type { ChatMessage } from "../hooks/useChat";
+import {
+  buildPlanBriefFromMessages,
+  formatPlanBrief,
+  planBriefPreview,
+  planContinueHandoff,
+  planExecuteHandoff,
+  stripConversationModeUiInstructions,
+} from "./planArtifact";
+
+describe("planArtifact (#616 P5/P7)", () => {
+  it("formats brief with summary and todos", () => {
+    const brief = formatPlanBrief({
+      summary: "Ship modes.",
+      todos: [{ id: "1", content: "Ask overlay", status: "completed" }],
+    });
+    expect(brief).toContain("Approved plan");
+    expect(brief).toContain("Ship modes.");
+    expect(brief).toContain("1. Ask overlay (completed)");
+    expect(brief).toContain("Execute this plan now.");
+  });
+
+  it("planBriefPreview drops headers and truncates", () => {
+    const brief = formatPlanBrief({
+      summary: "Do the thing.",
+      todos: [{ id: "1", content: "Step one", status: "pending" }],
+    });
+    const preview = planBriefPreview(brief);
+    expect(preview).toContain("Do the thing.");
+    expect(preview).toContain("Step one");
+    expect(preview).not.toContain("Approved plan");
+    expect(preview).not.toMatch(/Execute this plan now/i);
+  });
+
+  it("planExecuteHandoff switches to craft with silent planBrief metadata", () => {
+    const handoff = planExecuteHandoff(formatPlanBrief({ summary: "Go." }));
+    expect(handoff.conversationMode).toBe("craft");
+    expect(handoff.hideUserMessage).toBe(true);
+    expect(handoff.text).toBe("Execute the approved plan now.");
+    expect(handoff.planBrief).toContain("Go.");
+    expect(handoff.planBrief).toContain("Approved plan");
+  });
+
+  it("planContinueHandoff keeps plan mode (P8)", () => {
+    expect(planContinueHandoff()).toEqual({ conversationMode: "plan" });
+  });
+
+  it("strips mode-switch UI instructions from assistant copy", () => {
+    const raw = [
+      "当前仍在计划模式，edit_file 等修改类工具不可用。",
+      "",
+      "需要你手动切换到默认（Craft / 做一做）模式后，我才能直接修 bug 和跑测试。",
+      "",
+      "切换入口：输入框旁边的模式选择按钮，从「计划」切回默认即可。",
+      "",
+      "脚本将用 awk 采集 CPU。",
+    ].join("\n");
+    const cleaned = stripConversationModeUiInstructions(raw);
+    expect(cleaned).toContain("awk");
+    expect(cleaned).not.toMatch(/手动切换/);
+    expect(cleaned).not.toMatch(/模式选择按钮/);
+    expect(
+      stripConversationModeUiInstructions(
+        [
+          "当前仍在计划模式，edit_file 等修改类工具不可用。",
+          "",
+          "需要你手动切换到默认（Craft / 做一做）模式。",
+          "",
+          "切换入口：输入框旁边的模式选择按钮。",
+        ].join("\n"),
+      ),
+    ).toBe("");
+  });
+
+  it("keeps ordinary mentions of chat input", () => {
+    expect(
+      stripConversationModeUiInstructions(
+        "You can paste content into the chat input area.",
+      ),
+    ).toContain("chat input");
+  });
+
+  it("buildPlanBriefFromMessages merges assistant text and write_todos", () => {
+    const messages: ChatMessage[] = [
+      {
+        id: "1",
+        role: "user",
+        content: "plan",
+        status: "done",
+        timestamp: 1,
+      },
+      {
+        id: "2",
+        role: "assistant",
+        content: "",
+        toolData: {
+          name: "write_todos",
+          arguments: JSON.stringify({
+            todos: [{ id: "1", content: "Write tests", status: "pending" }],
+          }),
+        },
+        status: "done",
+        timestamp: 2,
+      },
+      {
+        id: "3",
+        role: "assistant",
+        content: "Here is the plan.",
+        status: "done",
+        timestamp: 3,
+      },
+    ];
+    const brief = buildPlanBriefFromMessages(messages);
+    expect(brief).toContain("Here is the plan.");
+    expect(brief).toContain("Write tests");
+  });
+
+  it("buildPlanBriefFromMessages ignores short replies without todos", () => {
+    const messages: ChatMessage[] = [
+      {
+        id: "1",
+        role: "user",
+        content: "plan",
+        status: "done",
+        timestamp: 1,
+      },
+      {
+        id: "2",
+        role: "assistant",
+        content: "好的，我先想想。",
+        status: "done",
+        timestamp: 2,
+      },
+    ];
+    expect(buildPlanBriefFromMessages(messages)).toBeNull();
+  });
+
+  it("buildPlanBriefFromMessages ignores stale write_todos from earlier turns", () => {
+    const messages: ChatMessage[] = [
+      {
+        id: "1",
+        role: "user",
+        content: "plan a file",
+        status: "done",
+        timestamp: 1,
+      },
+      {
+        id: "2",
+        role: "assistant",
+        content: "",
+        toolData: {
+          name: "write_todos",
+          arguments: JSON.stringify({
+            todos: [{ id: "1", content: "Write tests", status: "pending" }],
+          }),
+        },
+        status: "done",
+        timestamp: 2,
+      },
+      {
+        id: "3",
+        role: "assistant",
+        content:
+          "Here is a detailed plan with plenty of substance for the card.",
+        status: "done",
+        timestamp: 3,
+      },
+      {
+        id: "4",
+        role: "user",
+        content: "嗯",
+        status: "done",
+        timestamp: 4,
+      },
+      {
+        id: "5",
+        role: "assistant",
+        content: "好的",
+        status: "done",
+        timestamp: 5,
+      },
+    ];
+    expect(buildPlanBriefFromMessages(messages)).toBeNull();
+  });
+});
diff --git a/dashboard/src/pages/Chat/utils/planArtifact.ts b/dashboard/src/pages/Chat/utils/planArtifact.ts
new file mode 100644
index 000000000..1ec8d4422
--- /dev/null
+++ b/dashboard/src/pages/Chat/utils/planArtifact.ts
@@ -0,0 +1,157 @@
+import type { TodoListItem } from "../../../utils/parseWriteTodos";
+import {
+  collectWriteTodosFromMessages,
+  type WriteTodosMessageSource,
+} from "../../../utils/parseWriteTodos";
+import type { ChatMessage } from "../hooks/useChat";
+import { deriveMessageContent } from "./messageContent";
+
+/**
+ * Lines that tell the user how to operate Plan/Craft UI — never show in chat.
+ * Keep patterns mode-specific so ordinary assistant prose is not stripped.
+ */
+const MODE_UI_INSTRUCTION_RE =
+  /手动切换|模式选择按钮|输入框旁边|切换入口|切换到\s*默认\s*模式|切换到\s*做一做|切回默认|请.{0,24}切换.{0,24}模式|switch to (?:default|craft)\s+mode|mode (?:selection )?button|(?:near|beside|next to) (?:the )?chat input|做一做[）)]?\s*模式|Craft\s*[/／]\s*做一做|修改类工具不可用|仍在计划模式/i;
+
+/**
+ * Remove assistant copy that instructs the user to flip the composer mode
+ * control. Product UI (PlanReadyCard) owns that handoff.
+ */
+export function stripConversationModeUiInstructions(text: string): string {
+  const raw = text.replace(/\r\n/g, "\n").trim();
+  if (!raw) return "";
+
+  const blocks = raw.split(/\n{2,}/);
+  const kept = blocks.filter((block) => {
+    const compact = block.replace(/\s+/g, "");
+    if (!compact) return false;
+    if (MODE_UI_INSTRUCTION_RE.test(block)) return false;
+    // Whole block is only a mode-name reminder.
+    if (
+      /^(当前)?(仍?在)?计划|仅问答|问答|Ask|Plan|Craft/.test(compact) &&
+      compact.length < 40
+    ) {
+      return !/模式|mode|不可用|unavailable/i.test(block);
+    }
+    return true;
+  });
+
+  // Also drop leftover single lines that match.
+  const lines = kept.join("\n\n").split("\n");
+  const cleaned = lines
+    .filter((line) => !MODE_UI_INSTRUCTION_RE.test(line))
+    .join("\n")
+    .replace(/\n{3,}/g, "\n\n")
+    .trim();
+  return cleaned;
+}
+
+export function formatPlanBrief(params: {
+  summary?: string;
+  todos?: TodoListItem[];
+}): string {
+  const lines: string[] = ["## Approved plan", ""];
+  const summary = stripConversationModeUiInstructions(
+    params.summary || "",
+  ).trim();
+  if (summary) {
+    lines.push(summary, "");
+  }
+  const todos = params.todos || [];
+  if (todos.length > 0) {
+    lines.push("### Steps");
+    todos.forEach((todo, index) => {
+      const status = todo.status ? ` (${todo.status})` : "";
+      lines.push(`${index + 1}. ${todo.content}${status}`);
+    });
+    lines.push("");
+  }
+  lines.push("Execute this plan now.");
+  return `${lines.join("\n").trim()}\n`;
+}
+
+/** Short preview of an approved brief for PlanReadyCard (drop headers / footer). */
+export function planBriefPreview(brief: string, maxLen = 320): string {
+  const lines = brief
+    .replace(/\r\n/g, "\n")
+    .split("\n")
+    .map((line) => line.trim())
+    .filter(
+      (line) =>
+        Boolean(line) &&
+        !/^#{1,6}\s/.test(line) &&
+        !/^Execute this plan now\.?$/i.test(line),
+    );
+  const text = lines.join("\n").trim();
+  if (!text) return "";
+  if (text.length <= maxLen) return text;
+  return `${text.slice(0, maxLen).trimEnd()}…`;
+}
+
+/** Collect plan brief from the latest assistant turn (text + write_todos). */
+export const PLAN_READY_MIN_SUMMARY_CHARS = 80;
+
+/** True when a built brief is substantial enough to show PlanReadyCard. */
+export function isPlanReadyBrief(brief: string | null | undefined): boolean {
+  if (!brief?.trim()) return false;
+  const hasSteps = /^### Steps$/m.test(brief) && /^\d+\. /m.test(brief);
+  if (hasSteps) return true;
+  const preview = planBriefPreview(brief, Number.MAX_SAFE_INTEGER);
+  return preview.trim().length >= PLAN_READY_MIN_SUMMARY_CHARS;
+}
+
+/** Collect plan brief from the latest assistant turn (text + write_todos). */
+export function buildPlanBriefFromMessages(
+  messages: readonly ChatMessage[],
+): string | null {
+  // Only the latest user→assistant turn. Older write_todos must not resurrect
+  // PlanReady when the model later replies with a short ack ("好的").
+  let lastUserIdx = -1;
+  for (let i = messages.length - 1; i >= 0; i -= 1) {
+    if (messages[i]?.role === "user") {
+      lastUserIdx = i;
+      break;
+    }
+  }
+  const turn = messages.slice(lastUserIdx + 1);
+  const todos = collectWriteTodosFromMessages(
+    turn as unknown as WriteTodosMessageSource[],
+  );
+  let summary = "";
+  for (let i = turn.length - 1; i >= 0; i -= 1) {
+    const msg = turn[i];
+    if (!msg || msg.role !== "assistant" || msg.toolData) continue;
+    summary = stripConversationModeUiInstructions(
+      deriveMessageContent(msg).textContent,
+    ).trim();
+    if (summary) break;
+  }
+  if (!summary && todos.length === 0) return null;
+  const brief = formatPlanBrief({ summary, todos });
+  return isPlanReadyBrief(brief) ? brief : null;
+}
+
+/** Minimal user-turn text for silent Plan→Craft (matched on history reload). */
+export const PLAN_EXECUTE_USER_TRIGGER = "Execute the approved plan now.";
+
+/** P7: Execute plan → Craft mode; brief goes via metadata (not chat text). */
+export function planExecuteHandoff(brief: string): {
+  conversationMode: "craft";
+  /** Silent send — no visible user bubble; planBrief drives the turn. */
+  hideUserMessage: true;
+  /** Non-empty WS text so servers that only check text still accept the turn. */
+  text: string;
+  planBrief: string;
+} {
+  return {
+    conversationMode: "craft",
+    hideUserMessage: true,
+    text: PLAN_EXECUTE_USER_TRIGGER,
+    planBrief: brief.trim(),
+  };
+}
+
+/** P8: Keep planning → stay in Plan mode. */
+export function planContinueHandoff(): { conversationMode: "plan" } {
+  return { conversationMode: "plan" };
+}
diff --git a/dashboard/src/pages/Chat/utils/seedConversationMode.test.ts b/dashboard/src/pages/Chat/utils/seedConversationMode.test.ts
new file mode 100644
index 000000000..5a6c3924f
--- /dev/null
+++ b/dashboard/src/pages/Chat/utils/seedConversationMode.test.ts
@@ -0,0 +1,73 @@
+import { describe, expect, it } from "vitest";
+import { PENDING_THREAD_ID } from "../constants";
+import { resolveSeedConversationMode } from "./seedConversationMode";
+
+describe("resolveSeedConversationMode", () => {
+  it("reseeds agent default on empty new-chat composer", () => {
+    expect(
+      resolveSeedConversationMode({
+        activeThreadId: null,
+        previousThreadId: "thr_old",
+        agentDefault: "craft",
+        currentComposerMode: "plan",
+      }),
+    ).toEqual({ action: "set", mode: "craft" });
+  });
+
+  it("restores per-thread override when present", () => {
+    expect(
+      resolveSeedConversationMode({
+        activeThreadId: "thr_a",
+        previousThreadId: "thr_b",
+        override: "ask",
+        agentDefault: "craft",
+        currentComposerMode: "plan",
+      }),
+    ).toEqual({ action: "set", mode: "ask" });
+  });
+
+  it("stamps composer mode when empty chat gets a real thread id", () => {
+    expect(
+      resolveSeedConversationMode({
+        activeThreadId: "thr_new",
+        previousThreadId: null,
+        agentDefault: "ask",
+        currentComposerMode: "plan",
+      }),
+    ).toEqual({ action: "stamp-override", mode: "plan" });
+  });
+
+  it("stamps composer mode when pending chat resolves to a real thread id", () => {
+    expect(
+      resolveSeedConversationMode({
+        activeThreadId: "thr_new",
+        previousThreadId: PENDING_THREAD_ID,
+        agentDefault: "craft",
+        currentComposerMode: "plan",
+      }),
+    ).toEqual({ action: "stamp-override", mode: "plan" });
+  });
+
+  it("ignores stale pending override and stamps current composer mode", () => {
+    expect(
+      resolveSeedConversationMode({
+        activeThreadId: PENDING_THREAD_ID,
+        previousThreadId: null,
+        override: "plan",
+        agentDefault: "craft",
+        currentComposerMode: "ask",
+      }),
+    ).toEqual({ action: "stamp-override", mode: "ask" });
+  });
+
+  it("reseeds agent default when switching between existing threads", () => {
+    expect(
+      resolveSeedConversationMode({
+        activeThreadId: "thr_b",
+        previousThreadId: "thr_a",
+        agentDefault: "craft",
+        currentComposerMode: "plan",
+      }),
+    ).toEqual({ action: "set", mode: "craft" });
+  });
+});
diff --git a/dashboard/src/pages/Chat/utils/seedConversationMode.ts b/dashboard/src/pages/Chat/utils/seedConversationMode.ts
new file mode 100644
index 000000000..070bc6e88
--- /dev/null
+++ b/dashboard/src/pages/Chat/utils/seedConversationMode.ts
@@ -0,0 +1,56 @@
+import { isPendingThread } from "../hooks/useSessions";
+import {
+  parseConversationMode,
+  type ConversationMode,
+} from "./conversationMode";
+
+export type SeedConversationModeResult =
+  | { action: "set"; mode: ConversationMode }
+  | { action: "stamp-override"; mode: ConversationMode }
+  | { action: "keep" };
+
+/**
+ * Decide how the composer conversation mode should react when the active
+ * thread id (or agent default) changes.
+ *
+ * Critical path: empty / pending chat → first real thread id must **stamp**
+ * the mode the user already picked, not wipe back to the agent default.
+ * Otherwise a Plan first-send never shows PlanReady.
+ */
+export function resolveSeedConversationMode(params: {
+  activeThreadId?: string | null;
+  previousThreadId?: string | null;
+  override?: ConversationMode;
+  agentDefault?: ConversationMode | string | null;
+  currentComposerMode: ConversationMode;
+}): SeedConversationModeResult {
+  const {
+    activeThreadId,
+    previousThreadId,
+    override,
+    agentDefault,
+    currentComposerMode,
+  } = params;
+
+  if (!activeThreadId) {
+    return { action: "set", mode: parseConversationMode(agentDefault) };
+  }
+
+  // `__pending__` is reused for every new chat. Never restore a stale override
+  // from a previous first-send — always stamp the composer mode the user has
+  // now. Same when empty composer first receives a real thread id.
+  const fromEmptyComposer =
+    previousThreadId == null ||
+    previousThreadId === "" ||
+    isPendingThread(previousThreadId) ||
+    isPendingThread(activeThreadId);
+  if (fromEmptyComposer) {
+    return { action: "stamp-override", mode: currentComposerMode };
+  }
+
+  if (override !== undefined) {
+    return { action: "set", mode: override };
+  }
+
+  return { action: "set", mode: parseConversationMode(agentDefault) };
+}
diff --git a/dashboard/src/pages/Experts/components/CreateFromExpertDrawer.tsx b/dashboard/src/pages/Experts/components/CreateFromExpertDrawer.tsx
index fae7d0fb3..bc0cca170 100644
--- a/dashboard/src/pages/Experts/components/CreateFromExpertDrawer.tsx
+++ b/dashboard/src/pages/Experts/components/CreateFromExpertDrawer.tsx
@@ -59,6 +59,8 @@ import AgentBackendFields from "./AgentBackendFields";
 import ExpertAvatarPicker from "./ExpertAvatarPicker";
 import ExpertComposerDefaultsFields from "./ExpertComposerDefaultsFields";
 import styles from "../index.module.less";
+import { conversationModeSelectOptions } from "../../Chat/utils/conversationMode";
+import { knowledgeBasesApi } from "../../../api/modules/knowledgeBases";
 
 type FileContent = NamedFileContent;
 
@@ -158,6 +160,8 @@ export default function CreateFromExpertDrawer({
       agent_id?: string;
       welcome_message?: string;
       default_model: string;
+      default_conversation_mode: "ask" | "plan" | "craft";
+      default_knowledge_base_ids?: string[];
       backend_choice: string;
       composite_default: string;
       root_dir?: string;
@@ -181,6 +185,9 @@ export default function CreateFromExpertDrawer({
   const [colorPalette, setColorPalette] = useState<string>("rose");
   const [avatarFile, setAvatarFile] = useState<File | null>(null);
   const [avatarPreview, setAvatarPreview] = useState<string | null>(null);
+  const [knowledgeBaseOptions, setKnowledgeBaseOptions] = useState<
+    Array<{ value: string; label: string }>
+  >([]);
 
   const backendChoice =
     Form.useWatch("backend_choice", form) ?? DEFAULT_BACKEND;
@@ -190,6 +197,28 @@ export default function CreateFromExpertDrawer({
     rootDir: watchedRootDir,
   });
 
+  useEffect(() => {
+    if (!open) return;
+    let cancelled = false;
+    void knowledgeBasesApi
+      .list()
+      .then((bases) => {
+        if (cancelled) return;
+        setKnowledgeBaseOptions(
+          bases.map((base) => ({
+            value: base.id,
+            label: base.name?.trim() || base.id,
+          })),
+        );
+      })
+      .catch(() => {
+        if (!cancelled) setKnowledgeBaseOptions([]);
+      });
+    return () => {
+      cancelled = true;
+    };
+  }, [open]);
+
   useEffect(() => {
     if (skillPackagesSupported) return;
     form.setFieldsValue({ skill_package_ids: [] });
@@ -220,6 +249,8 @@ export default function CreateFromExpertDrawer({
       agent_id: undefined,
       welcome_message: defaults.welcome_message,
       default_model: MODEL_AUTO_VALUE,
+      default_conversation_mode: "craft",
+      default_knowledge_base_ids: [],
       backend_choice: DEFAULT_BACKEND,
       composite_default: DEFAULT_BACKEND,
       skill_package_ids: [],
@@ -343,6 +374,8 @@ export default function CreateFromExpertDrawer({
         ...(welcomeText ? { welcome_message: welcomeText } : {}),
         ...buildAgentRuntimeRequest(values),
         enable_trajectory: values.enable_trajectory === true,
+        default_conversation_mode: values.default_conversation_mode,
+        default_knowledge_base_ids: values.default_knowledge_base_ids ?? [],
       };
 
       let body: { agent_id: string; name: string };
@@ -573,6 +606,36 @@ export default function CreateFromExpertDrawer({
           />
         </Form.Item>
 
+        <Form.Item
+          name="default_conversation_mode"
+          label={t("experts.defaultConversationModeLabel", "默认对话模式")}
+          extra={t(
+            "experts.defaultConversationModeHint",
+            "新对话的初始权限模式；可在聊天输入栏随时切换。",
+          )}
+        >
+          <Select options={conversationModeSelectOptions(t)} />
+        </Form.Item>
+        <Form.Item
+          name="default_knowledge_base_ids"
+          label={t("experts.defaultKnowledgeBasesLabel", "默认知识库")}
+          extra={t(
+            "experts.defaultKnowledgeBasesHint",
+            "新对话未手动选择知识库时自动附带；仍可在输入栏增减。",
+          )}
+        >
+          <Select
+            mode="multiple"
+            allowClear
+            optionFilterProp="label"
+            options={knowledgeBaseOptions}
+            placeholder={t(
+              "experts.defaultKnowledgeBasesPlaceholder",
+              "选择知识库",
+            )}
+          />
+        </Form.Item>
+
         <AgentBackendFields
           backends={backends}
           backendsLoading={backendsLoading}
diff --git a/dashboard/src/pages/Experts/components/EditAgentDrawer.tsx b/dashboard/src/pages/Experts/components/EditAgentDrawer.tsx
index 112c60c18..888cbe73a 100644
--- a/dashboard/src/pages/Experts/components/EditAgentDrawer.tsx
+++ b/dashboard/src/pages/Experts/components/EditAgentDrawer.tsx
@@ -78,12 +78,16 @@ import AgentBackendFields from "./AgentBackendFields";
 import ExpertComposerDefaultsFields from "./ExpertComposerDefaultsFields";
 import SubagentCatalogDrawer from "./SubagentCatalogDrawer";
 import styles from "../index.module.less";
+import { conversationModeSelectOptions } from "../../Chat/utils/conversationMode";
+import { knowledgeBasesApi } from "../../../api/modules/knowledgeBases";
 
 interface AgentDetail {
   id: string;
   name: string;
   description: string | null;
   default_model: string | null;
+  default_conversation_mode?: "ask" | "plan" | "craft" | null;
+  default_knowledge_base_ids?: string[] | null;
   color?: string | null;
   icon_url?: string | null;
   max_iters?: number | null;
@@ -129,6 +133,8 @@ interface EditFormValues {
   welcome_message?: string;
   is_shared?: boolean;
   default_model: string;
+  default_conversation_mode: "ask" | "plan" | "craft";
+  default_knowledge_base_ids?: string[];
   backend_choice: string;
   composite_default: string;
   root_dir?: string;
@@ -153,6 +159,8 @@ interface EditAgentDrawerProps {
       | "name"
       | "description"
       | "default_model"
+      | "default_conversation_mode"
+      | "default_knowledge_base_ids"
       | "is_shared"
       | "color"
       | "icon_url"
@@ -230,6 +238,9 @@ function EditAgentDrawerBody({
     useAgentFormResources(true);
   const [pathMappings, setPathMappings] = useState<PathMapping[]>([]);
   const [agentConfig, setAgentConfig] = useState<Record<string, unknown>>({});
+  const [knowledgeBaseOptions, setKnowledgeBaseOptions] = useState<
+    Array<{ value: string; label: string }>
+  >([]);
   const [colorPalette, setColorPalette] = useState<string>(
     () => parseStoredColor(agent.color) ?? DEFAULT_PALETTE,
   );
@@ -268,6 +279,27 @@ function EditAgentDrawerBody({
     workspaceDir: workspaceDirFromConfig,
   });
 
+  useEffect(() => {
+    let cancelled = false;
+    void knowledgeBasesApi
+      .list()
+      .then((bases) => {
+        if (cancelled) return;
+        setKnowledgeBaseOptions(
+          bases.map((base) => ({
+            value: base.id,
+            label: base.name?.trim() || base.id,
+          })),
+        );
+      })
+      .catch(() => {
+        if (!cancelled) setKnowledgeBaseOptions([]);
+      });
+    return () => {
+      cancelled = true;
+    };
+  }, []);
+
   useEffect(() => {
     let cancelled = false;
     setLoading(true);
@@ -303,6 +335,23 @@ function EditAgentDrawerBody({
             typeof ag.welcome_message === "string" ? ag.welcome_message : "",
           is_shared: agent.is_shared ?? false,
           default_model: defaultModelToForm(ag.default_model),
+          default_conversation_mode:
+            ag.default_conversation_mode === "ask" ||
+            ag.default_conversation_mode === "plan" ||
+            ag.default_conversation_mode === "craft"
+              ? ag.default_conversation_mode
+              : "craft",
+          default_knowledge_base_ids: Array.isArray(
+            ag.default_knowledge_base_ids,
+          )
+            ? ag.default_knowledge_base_ids.filter(
+                (id): id is string => typeof id === "string" && Boolean(id),
+              )
+            : Array.isArray(cfg.default_knowledge_base_ids)
+            ? (cfg.default_knowledge_base_ids as unknown[]).filter(
+                (id): id is string => typeof id === "string" && Boolean(id),
+              )
+            : [],
           backend_choice: parsedBackend.backendChoice,
           composite_default: parsedBackend.compositeDefault,
           root_dir: parsedBackend.rootDir,
@@ -408,6 +457,8 @@ function EditAgentDrawerBody({
         ...agentConfig,
         backend: backendSpec,
         enable_trajectory: values.enable_trajectory === true,
+        default_conversation_mode: values.default_conversation_mode,
+        default_knowledge_base_ids: values.default_knowledge_base_ids ?? [],
       });
       delete nextConfig.color;
       delete nextConfig.icon_name;
@@ -483,6 +534,8 @@ function EditAgentDrawerBody({
         name: values.name,
         description: values.description || null,
         default_model: defaultModel,
+        default_conversation_mode: values.default_conversation_mode,
+        default_knowledge_base_ids: values.default_knowledge_base_ids ?? [],
         is_shared: values.is_shared ?? false,
         color: nextColor,
         icon_url: iconUrl,
@@ -799,6 +852,38 @@ function EditAgentDrawerBody({
                   }
                 />
               </Form.Item>
+              <Form.Item
+                name="default_conversation_mode"
+                label={t(
+                  "experts.defaultConversationModeLabel",
+                  "默认对话模式",
+                )}
+                extra={t(
+                  "experts.defaultConversationModeHint",
+                  "新对话的初始权限模式；可在聊天输入栏随时切换。",
+                )}
+              >
+                <Select options={conversationModeSelectOptions(t)} />
+              </Form.Item>
+              <Form.Item
+                name="default_knowledge_base_ids"
+                label={t("experts.defaultKnowledgeBasesLabel", "默认知识库")}
+                extra={t(
+                  "experts.defaultKnowledgeBasesHint",
+                  "新对话未手动选择知识库时自动附带；仍可在输入栏增减。",
+                )}
+              >
+                <Select
+                  mode="multiple"
+                  allowClear
+                  optionFilterProp="label"
+                  options={knowledgeBaseOptions}
+                  placeholder={t(
+                    "experts.defaultKnowledgeBasesPlaceholder",
+                    "选择知识库",
+                  )}
+                />
+              </Form.Item>
 
               <AgentBackendFields
                 backends={backends}
diff --git a/dashboard/src/pages/Experts/index.tsx b/dashboard/src/pages/Experts/index.tsx
index 2f9f7ff7c..4e5bab76d 100644
--- a/dashboard/src/pages/Experts/index.tsx
+++ b/dashboard/src/pages/Experts/index.tsx
@@ -224,6 +224,7 @@ export default function ExpertsPage() {
         | "name"
         | "description"
         | "default_model"
+        | "default_conversation_mode"
         | "is_shared"
         | "color"
         | "icon_url"
@@ -238,6 +239,7 @@ export default function ExpertsPage() {
                 name: updated.name,
                 description: updated.description,
                 default_model: updated.default_model,
+                default_conversation_mode: updated.default_conversation_mode,
                 is_shared: updated.is_shared,
                 color: updated.color,
                 icon_url: updated.icon_url,
diff --git a/docs/api.md b/docs/api.md
index aa56953c4..b78c830fd 100644
--- a/docs/api.md
+++ b/docs/api.md
@@ -80,9 +80,9 @@ routes until the wizard finishes.
 
 | Method | Path | Auth | Notes |
 |--------|------|------|-------|
-| `GET`    | `/agents` | user | `[{id, agent_id, name, persona_mbti, state, unread_count, ...}]` |
-| `POST`   | `/agents` | user | body `{name, persona_mbti?, default_model?, system_prompt?, description?, icon?, template_name?, config?}` → `201` |
-| `GET`    | `/agents/{id}` | owner | full agent row |
+| `GET`    | `/agents` | user | `[{id, agent_id, name, …, default_conversation_mode, default_knowledge_base_ids}]` |
+| `POST`   | `/agents` | user | body `{name, persona_mbti?, default_model?, system_prompt?, description?, icon?, template_name?, config?}` → `201`; optional `config.default_conversation_mode`, `config.default_knowledge_base_ids` |
+| `GET`    | `/agents/{id}` | owner | full agent row (includes top-level `default_conversation_mode`, `default_knowledge_base_ids`) |
 | `PATCH`  | `/agents/{id}` | owner | body subset of create body |
 | `DELETE` | `/agents/{id}` | owner | `204` |
 | `POST`   | `/agents/{id}/start` | owner | `204` |
@@ -90,7 +90,7 @@ routes until the wizard finishes.
 | `POST`   | `/agents/{id}/reload` | owner | `204` (rebuild harness runtime) |
 | `POST`   | `/agents/{id}/read` | owner | `204` (mark unread badge cleared) |
 | `GET`    | `/agents/{id}/status` | owner | `{state, last_error?, ...}` |
-| `POST`   | `/agents/from-expert/{expert_id}` | user | body `{name, ...}` → `201` (creates from bundled expert template) |
+| `POST`   | `/agents/from-expert/{expert_id}` | user | body `{name, …, default_conversation_mode?, enable_trajectory?}` → `201` (creates from bundled expert template) |
 | `GET`    | `/agents/{id}/tool-settings` | owner | built-in + installed plugin tools with enable / disableable / available flags |
 | `PUT`    | `/agents/{id}/tool-settings` | owner | body `{disabled_builtin: string[], plugins?}` — persists denylist + plugin flags (hot-sync, no reload) |
 | `PATCH`  | `/agents/{id}/tool-settings/{tool_name}` | owner | body `{enabled, source, plugin_id?}` — toggle one tool (hot-sync) |
@@ -110,7 +110,7 @@ destination synchronized after the request completes.
 
 | Path | Auth | Notes |
 |------|------|-------|
-| `WS /agents/{id}/chat/ws?token=<jwt>` | owner | Primary dashboard turn endpoint. Send `{"type":"user_turn", ...}` frames; server replies with harness stream chunks ending in `{"type":"done"}` or `{"type":"error","message":"..."}`. `{"type":"ping"}` → `{"type":"pong"}`. `{"type":"subscribe","thread_id"}` → `{"type":"turn_status","thread_id","active"}` (attach to an in-flight turn without cancelling on disconnect). `{"type":"cancel","thread_id"}` stops the active turn (explicit stop; disconnect alone does **not** cancel). |
+| `WS /agents/{id}/chat/ws?token=<jwt>` | owner | Primary dashboard turn endpoint. Send `{"type":"user_turn", …, "conversation_mode"?: "ask"|"plan"|"craft", "plan_brief"?: string}` frames; server replies with harness stream chunks ending in `{"type":"done"}` or `{"type":"error","message":"..."}`. Omit `conversation_mode` → agent `default_conversation_mode` (else craft); IM `/mode` sticky override also applies when meta omits mode. `plan_brief` is injected into the turn system prompt (Craft handoff) and is not required in chat text. `{"type":"ping"}` → `{"type":"pong"}`. `{"type":"subscribe","thread_id"}` → `{"type":"turn_status","thread_id","active"}` (attach to an in-flight turn without cancelling on disconnect). `{"type":"cancel","thread_id"}` stops the active turn (explicit stop; disconnect alone does **not** cancel). |
 | `GET /agents/{id}/chat/welcome` | agent access | `{welcome_message, quick_prompts, task_examples}`; `task_examples` is `null` when the workspace field is absent |
 | `POST /agents/{id}/chat/polish` | owner | body `{text, default_model?}` → `{text}` (one-shot prompt refinement) |
 | `POST /agents/{id}/chat/hitl/resume` | owner | body `{thread_id, decisions: [...]}` → SSE chunk stream; finishes with `{"type":"done"}` |
diff --git a/src/octop/api/routers/agents.py b/src/octop/api/routers/agents.py
index c74625a8b..ff43f3fdc 100644
--- a/src/octop/api/routers/agents.py
+++ b/src/octop/api/routers/agents.py
@@ -21,6 +21,10 @@
     read_workspace_avatar,
     write_workspace_avatar,
 )
+from octop.infra.agents.conversation_mode import (
+    default_conversation_mode_from_config,
+    normalize_config_default_conversation_mode,
+)
 from octop.infra.agents.profile import (
     id_list_from_row,
     parse_config_json,
@@ -33,6 +37,10 @@
     agent_runtime_values,
 )
 from octop.infra.errors import ErrorCode, OctopError
+from octop.infra.knowledge.default_open import (
+    default_knowledge_base_ids_from_config,
+    normalize_config_default_knowledge_base_ids,
+)
 from octop.infra.users.permissions import user_has_permission
 
 logger = logging.getLogger(__name__)
@@ -40,6 +48,17 @@
 router = APIRouter()
 
 
+def _validated_agent_config(config: dict[str, Any] | None) -> dict[str, Any] | None:
+    """Normalize conversation-mode / KB default fields; raise OctopError on invalid values."""
+    if config is None:
+        return None
+    try:
+        out = normalize_config_default_conversation_mode(config)
+        return normalize_config_default_knowledge_base_ids(out)
+    except ValueError as exc:
+        raise OctopError(ErrorCode.INTERNAL_ERROR, str(exc), status=400) from exc
+
+
 class AgentCreateBody(AgentRuntimeFields):
     name: str
     description: str | None = None
@@ -150,6 +169,8 @@ def _row_dict(
         "description": row.description,
         "persona_mbti": row.persona_mbti,
         "default_model": row.default_model,
+        "default_conversation_mode": default_conversation_mode_from_config(cfg),
+        "default_knowledge_base_ids": default_knowledge_base_ids_from_config(cfg),
         "system_prompt": row.system_prompt,
         "state": row.last_state or "unknown",
         "last_error": row.last_error,
@@ -258,12 +279,14 @@ async def create_agent(
     from octop.infra.agents.manager import AgentCreateSpec  # noqa: PLC0415
 
     assert server.app_runtime is not None
+    config = body.config if isinstance(body.config, dict) else {}
     if isinstance(body.config, dict):
         assert_user_backend_root_dirs(
             user,
             body.config.get("backend"),
             policy_repo=server.services.user_policy_repo,
         )
+        config = _validated_agent_config(body.config) or {}
     knowledge_ids = (
         server.app_runtime.agent_registry.validate_knowledge_base_ids(
             user.id, body.knowledge_base_ids
@@ -283,7 +306,7 @@ async def create_agent(
         persona_mbti=body.persona_mbti,
         default_model=body.default_model,
         system_prompt=body.system_prompt,
-        config=body.config,
+        config=config,
         runtime_config=runtime_field_updates(body, exclude_unset=True),
         icon=body.icon,
         template_name=body.template_name,
@@ -376,7 +399,8 @@ async def patch_agent(
         }
     }
     if body.config is not None:
-        updates["config_json"] = json.dumps(body.config)
+        validated = _validated_agent_config(body.config if isinstance(body.config, dict) else {})
+        updates["config_json"] = json.dumps(validated)
     if body.welcome_message is not None:
         updates["welcome_message"] = body.welcome_message
     if updates:
diff --git a/src/octop/api/routers/chat/models.py b/src/octop/api/routers/chat/models.py
index 54f6a5804..06a93ef39 100644
--- a/src/octop/api/routers/chat/models.py
+++ b/src/octop/api/routers/chat/models.py
@@ -46,6 +46,20 @@ class ChatTurnBody(BaseModel):
     )
     reasoning_mode: Literal["auto", "enabled", "disabled"] | None = None
     reasoning_effort: str | None = None
+    conversation_mode: Literal["ask", "plan", "craft"] | None = Field(
+        default=None,
+        description=(
+            "Turn permission mode: ask (read-only), plan (plan then confirm), "
+            "craft (full tools). Omit to default to craft."
+        ),
+    )
+    plan_brief: str | None = Field(
+        default=None,
+        description=(
+            "Optional approved plan text for a Craft handoff turn. "
+            "Injected into the system prompt by the gateway — not required in chat text."
+        ),
+    )
     target_agent_ids: list[str] | None = Field(
         default=None,
         description="Optional agent ids to involve via @mention (same user only).",
@@ -87,6 +101,14 @@ def from_ws_payload(cls, payload: dict[str, Any]) -> ChatTurnBody:
             if isinstance(payload.get("reasoning_effort"), str)
             and str(payload["reasoning_effort"]).strip()
             else None,
+            conversation_mode=payload.get("conversation_mode")
+            if payload.get("conversation_mode") in ("ask", "plan", "craft")
+            else None,
+            plan_brief=(
+                str(payload["plan_brief"]).strip()
+                if isinstance(payload.get("plan_brief"), str) and str(payload["plan_brief"]).strip()
+                else None
+            ),
             target_agent_ids=(
                 [str(x) for x in payload["target_agent_ids"]]
                 if isinstance(payload.get("target_agent_ids"), list)
@@ -107,6 +129,8 @@ class UserTurnWsFrame(BaseModel):
     default_model: str | None = None
     reasoning_mode: Literal["auto", "enabled", "disabled"] | None = None
     reasoning_effort: str | None = None
+    conversation_mode: Literal["ask", "plan", "craft"] | None = None
+    plan_brief: str | None = None
     mcp_servers: list[str] | None = None
     knowledge_base_ids: list[str] | None = None
     skills: list[str] | None = None
diff --git a/src/octop/api/routers/chat/serialize.py b/src/octop/api/routers/chat/serialize.py
index 21f1e9711..3d873a65e 100644
--- a/src/octop/api/routers/chat/serialize.py
+++ b/src/octop/api/routers/chat/serialize.py
@@ -23,6 +23,7 @@
     INBOUND_ATTACHMENTS_KEY,
     STREAM_ERROR_CODE_KEY,
     STREAM_ERROR_FLAG,
+    UI_HIDDEN_KEY,
     parse_checkpoint_ts_ms,
 )
 from octop.infra.utils.llm_text import strip_thinking as _strip_thinking
@@ -1000,6 +1001,8 @@ def _serialize_history_message(
             entry["composer_context"] = raw_ctx
         if has_user_attachments:
             entry["inbound_attachments"] = raw_att
+        if additional_kwargs.get(UI_HIDDEN_KEY):
+            entry["ui_hidden"] = True
     if additional_kwargs.get(STREAM_ERROR_FLAG):
         entry["status"] = "error"
         code = additional_kwargs.get(STREAM_ERROR_CODE_KEY)
diff --git a/src/octop/api/routers/chat/turn.py b/src/octop/api/routers/chat/turn.py
index f8b0d14c6..a46f908dc 100644
--- a/src/octop/api/routers/chat/turn.py
+++ b/src/octop/api/routers/chat/turn.py
@@ -124,9 +124,11 @@ def _message_content_nonempty(content: Any) -> bool:
 
 
 def turn_has_content(turn: ChatTurnBody) -> bool:
-    """True when the turn has user text or structured message blocks."""
+    """True when the turn has user text, attachments, or a plan_brief handoff."""
     if _turn_plain_text(turn):
         return True
+    if isinstance(turn.plan_brief, str) and turn.plan_brief.strip():
+        return True
     for msg in reversed(turn.messages):
         if str(msg.get("role") or "").lower() != "user":
             continue
@@ -172,54 +174,58 @@ def content_parts_from_dashboard_turn(turn: ChatTurnBody) -> list[ContentPart]:
         user_content = msg.get("content")
         break
 
-    if not isinstance(user_content, list):
-        return parts
-
-    seen_text = bool(text)
-    for block in user_content:
-        if not isinstance(block, dict):
-            continue
-        btype = str(block.get("type") or "")
-        if btype == "text":
-            t = str(block.get("text") or "").strip()
-            if t and not seen_text:
-                parts.append(TextContent(text=t))
-                seen_text = True
-            continue
-
-        if btype not in ("image", "file", "image_url"):
-            continue
-
-        path = _workspace_path_from_block(block)
-        if not path:
-            # Upload always returns workspace_path; ignore preview-only blocks.
-            continue
-
-        media_type = _mime_from_block(block)
-        filename = str(
-            block.get("filename") or block.get("name") or Path(path).name or "attachment"
-        )
-        use_vision = is_vision_attachment(
-            kind="image" if btype in ("image", "image_url") else btype,
-            media_type=media_type,
-            path=path,
-        )
-        if use_vision:
-            parts.append(
-                ImageContent(
-                    local_path=path,
-                    mime_type=media_type if media_type.startswith("image/") else "image/png",
-                    alt_text=filename,
-                )
+    if isinstance(user_content, list):
+        seen_text = bool(text)
+        for block in user_content:
+            if not isinstance(block, dict):
+                continue
+            btype = str(block.get("type") or "")
+            if btype == "text":
+                t = str(block.get("text") or "").strip()
+                if t and not seen_text:
+                    parts.append(TextContent(text=t))
+                    seen_text = True
+                continue
+
+            if btype not in ("image", "file", "image_url"):
+                continue
+
+            path = _workspace_path_from_block(block)
+            if not path:
+                # Upload always returns workspace_path; ignore preview-only blocks.
+                continue
+
+            media_type = _mime_from_block(block)
+            filename = str(
+                block.get("filename") or block.get("name") or Path(path).name or "attachment"
             )
-        else:
-            parts.append(
-                FileContent(
-                    local_path=path,
-                    filename=filename,
-                    mime_type=media_type or "application/octet-stream",
-                )
+            use_vision = is_vision_attachment(
+                kind="image" if btype in ("image", "image_url") else btype,
+                media_type=media_type,
+                path=path,
             )
+            if use_vision:
+                parts.append(
+                    ImageContent(
+                        local_path=path,
+                        mime_type=media_type if media_type.startswith("image/") else "image/png",
+                        alt_text=filename,
+                    )
+                )
+            else:
+                parts.append(
+                    FileContent(
+                        local_path=path,
+                        filename=filename,
+                        mime_type=media_type or "application/octet-stream",
+                    )
+                )
+
+    # Silent Plan→Craft: plan lives in metadata / system hint; still need a HumanMessage.
+    if not parts and isinstance(turn.plan_brief, str) and turn.plan_brief.strip():
+        from octop.infra.agents.plan_artifact import PLAN_EXECUTE_USER_TRIGGER  # noqa: PLC0415
+
+        parts.append(TextContent(text=PLAN_EXECUTE_USER_TRIGGER))
     return parts
 
 
@@ -332,6 +338,10 @@ def build_dashboard_inbound(
         metadata["reasoning_mode"] = turn.reasoning_mode
     if turn.reasoning_effort:
         metadata["reasoning_effort"] = turn.reasoning_effort
+    if turn.conversation_mode is not None:
+        metadata["conversation_mode"] = turn.conversation_mode
+    if turn.plan_brief:
+        metadata["plan_brief"] = turn.plan_brief
     if prepared.composer_context:
         metadata[COMPOSER_CTX_KEY] = prepared.composer_context
     if prepared.inbound_attachments:
diff --git a/src/octop/api/routers/experts.py b/src/octop/api/routers/experts.py
index 6da5a12ff..442667aba 100644
--- a/src/octop/api/routers/experts.py
+++ b/src/octop/api/routers/experts.py
@@ -12,7 +12,7 @@
 from __future__ import annotations
 
 import asyncio
-from typing import Any
+from typing import Any, Literal
 
 from fastapi import APIRouter, Depends, Response
 from pydantic import BaseModel, Field
@@ -25,6 +25,7 @@
     display_published_expert_icon_url,
     read_snapshot_avatar,
 )
+from octop.infra.agents.conversation_mode import apply_default_conversation_mode
 from octop.infra.agents.experts.catalog import (
     MANIFEST_FILENAME,
     build_create_spec_from_expert,
@@ -63,6 +64,7 @@
     fetch_skillset,
 )
 from octop.infra.errors import ErrorCode, OctopError
+from octop.infra.knowledge.default_open import apply_default_knowledge_base_ids
 from octop.infra.trajectory.settings import apply_enable_trajectory
 from octop.infra.utils.locale import resolve_user_locale
 
@@ -116,6 +118,17 @@ class FromExpertBody(AgentRuntimeFields):
     )
     welcome_message: str | None = None
     enable_trajectory: bool = True
+    default_conversation_mode: Literal["ask", "plan", "craft"] | None = Field(
+        default=None,
+        description=(
+            "Initial chat permission mode for new threads "
+            "(ask / plan / craft). Omit to default to craft."
+        ),
+    )
+    default_knowledge_base_ids: list[str] | None = Field(
+        default=None,
+        description="Knowledge bases auto-attached when a turn omits an explicit list.",
+    )
 
 
 class PublishExpertBody(BaseModel):
@@ -152,6 +165,17 @@ class InstallPublishedExpertBody(AgentRuntimeFields):
     )
     welcome_message: str | None = None
     enable_trajectory: bool = True
+    default_conversation_mode: Literal["ask", "plan", "craft"] | None = Field(
+        default=None,
+        description=(
+            "Initial chat permission mode for new threads "
+            "(ask / plan / craft). Omit to default to craft."
+        ),
+    )
+    default_knowledge_base_ids: list[str] | None = Field(
+        default=None,
+        description="Knowledge bases auto-attached when a turn omits an explicit list.",
+    )
 
 
 class LocalizedTextResponse(BaseModel):
@@ -537,6 +561,8 @@ async def install_published_expert(
             welcome_message=body.welcome_message,
             runtime_config=runtime_field_updates(body, exclude_unset=True),
             enable_trajectory=body.enable_trajectory,
+            default_conversation_mode=body.default_conversation_mode,
+            default_knowledge_base_ids=body.default_knowledge_base_ids,
         ),
     )
 
@@ -640,6 +666,8 @@ async def install_expert_hub_item(
                 knowledge_base_ids=kb_ids,
                 mcp_servers=servers,
                 enable_trajectory=body.enable_trajectory,
+                default_conversation_mode=body.default_conversation_mode,
+                default_knowledge_base_ids=body.default_knowledge_base_ids,
                 **runtime_field_updates(body, exclude_unset=False),
             ),
         )
@@ -720,6 +748,8 @@ async def create_agent_from_expert(
     if body.backend:
         config_extra["backend"] = body.backend
     apply_enable_trajectory(config_extra, body.enable_trajectory)
+    apply_default_conversation_mode(config_extra, body.default_conversation_mode)
+    apply_default_knowledge_base_ids(config_extra, body.default_knowledge_base_ids)
 
     locale = resolve_user_locale(
         user_repo=server.services.user_repo,
diff --git a/src/octop/i18n/domains/conversation.py b/src/octop/i18n/domains/conversation.py
new file mode 100644
index 000000000..96916eb3c
--- /dev/null
+++ b/src/octop/i18n/domains/conversation.py
@@ -0,0 +1,47 @@
+"""``conversation.*`` — Ask / Plan / Craft mode system hints (#616)."""
+
+from __future__ import annotations
+
+from octop.i18n.loader import tr
+from octop.infra.agents.conversation_mode import ConversationMode
+from octop.infra.utils.locale import Locale
+
+__all__ = [
+    "conversation_mode_plan_brief_block",
+    "conversation_mode_system_hint",
+    "conversation_mode_tool_blocked",
+]
+
+
+def conversation_mode_system_hint(
+    mode: ConversationMode,
+    locale: str | Locale = "en",
+) -> str:
+    """Localized system hint for the turn's conversation mode."""
+    return tr(f"conversation.mode.{mode}_system_hint", locale)
+
+
+def conversation_mode_plan_brief_block(
+    brief: str,
+    locale: str | Locale = "en",
+) -> str:
+    """Wrap an approved plan brief for system-prompt injection (Craft handoff)."""
+    cleaned = brief.strip()
+    if not cleaned:
+        return ""
+    prefix = tr("conversation.mode.plan_brief_prefix", locale)
+    return f"{prefix}\n\n{cleaned}"
+
+
+def conversation_mode_tool_blocked(
+    tool_name: str,
+    mode: ConversationMode,
+    locale: str | Locale = "en",
+) -> str:
+    """Error text when a denylisted tool is invoked under Ask/Plan."""
+    return tr(
+        "conversation.mode.tool_blocked",
+        locale,
+        tool_name=tool_name,
+        mode=mode,
+    )
diff --git a/src/octop/i18n/en.json b/src/octop/i18n/en.json
index d9f1b5b0a..e87e0ace8 100644
--- a/src/octop/i18n/en.json
+++ b/src/octop/i18n/en.json
@@ -111,6 +111,13 @@
       "cleared": "Model override cleared.",
       "set": "Model for this thread → {model}."
     },
+    "mode": {
+      "override": "Conversation mode override: {mode}",
+      "current_default": "Conversation mode: {mode} (agent default)",
+      "usage": "Usage: /mode ask|plan|craft|reset (aliases: /ask /plan /craft)",
+      "cleared": "Conversation mode override cleared.",
+      "set": "Conversation mode for this thread → {mode}."
+    },
     "compact": {
       "done": "Model context compacted (thread `{short}`): summarized {count} older message(s), kept the recent tail.\n\nChat history is unchanged; from the next turn the model only sees the summary plus recent messages.",
       "done_offload": "Model context compacted (thread `{short}`): summarized {count} older message(s), kept the recent tail.\n\n- Saved: `{path}`\n- Chat history is unchanged; from the next turn the model only sees the summary plus recent messages.",
@@ -214,6 +221,10 @@
         "label": "Model",
         "description": "Show, set, or reset the model override for this thread"
       },
+      "mode": {
+        "label": "Conversation mode",
+        "description": "Show, set, or reset Ask / Plan / Craft for this thread"
+      },
       "new": {
         "label": "New chat",
         "description": "Clear context and start a new conversation thread"
@@ -682,5 +693,14 @@
     "row_total": "Total",
     "axis_tokens": "Tokens",
     "filename_prefix": "token-usage"
+  },
+  "conversation": {
+    "mode": {
+      "ask_system_hint": "You are in Ask mode (问答): answer with read-only tools (knowledge search, file read, web fetch). Do not create, edit, delete, execute, or write todos/plans. Do not send files, run browser/desktop/mobile actions, or delegate via task / ACP / ask_agent. If the user needs changes, say they can use Plan or Default mode — do not tell them how to operate the UI.",
+      "plan_system_hint": "You are in Plan mode (计划): explore with read-only tools and produce a clear step-by-step plan/spec only. Prefer write_todos and structured plan text. Do not execute, edit files, or delegate via task / ACP / ask_agent. When the plan is ready, stop — the product UI will offer confirmation to execute. Do not ask the user to change mode controls or paste internal prompts.",
+      "craft_system_hint": "You are in Default / Craft mode (默认): you may use the full set of allowed tools to carry out the user's request, subject to existing safety and HITL rules.",
+      "plan_brief_prefix": "The user confirmed the following approved plan. Execute it now in Default mode. Do not repeat the plan verbatim unless needed for clarification:",
+      "tool_blocked": "Tool `{tool_name}` is not available in {mode} mode. Switch to Default (Craft) to run mutating or delegated tools."
+    }
   }
 }
diff --git a/src/octop/i18n/zh.json b/src/octop/i18n/zh.json
index ba28405f8..13f71765c 100644
--- a/src/octop/i18n/zh.json
+++ b/src/octop/i18n/zh.json
@@ -111,6 +111,13 @@
       "cleared": "已清除模型覆盖。",
       "set": "此话题模型 → {model}。"
     },
+    "mode": {
+      "override": "对话模式覆盖：{mode}",
+      "current_default": "对话模式：{mode}（Agent 默认）",
+      "usage": "用法：/mode ask|plan|craft|reset（别名：/ask /plan /craft）",
+      "cleared": "已清除对话模式覆盖。",
+      "set": "此话题对话模式 → {mode}。"
+    },
     "compact": {
       "done": "已压缩模型上下文（话题 `{short}`）：较早消息 {count} 条已总结，保留最近对话。\n\n聊天记录仍完整保留；下一轮起模型只使用摘要 + 近期对话。",
       "done_offload": "已压缩模型上下文（话题 `{short}`）：较早消息 {count} 条已总结，保留最近对话。\n\n- 原文：`{path}`\n- 聊天记录仍完整保留；下一轮起模型只使用摘要 + 近期对话。",
@@ -214,6 +221,10 @@
         "label": "切换模型",
         "description": "查看、设置或重置当前话题的模型覆盖"
       },
+      "mode": {
+        "label": "对话模式",
+        "description": "查看、设置或重置当前话题的问答 / 计划 / 默认模式"
+      },
       "new": {
         "label": "新对话",
         "description": "清空上下文并开始新的对话线程"
@@ -682,5 +693,14 @@
     "row_total": "合计",
     "axis_tokens": "Token 数",
     "filename_prefix": "Token用量"
+  },
+  "conversation": {
+    "mode": {
+      "ask_system_hint": "当前为问答（Ask）模式：只用只读能力（知识库检索、读文件、网页抓取等）回答与解释。不要创建、编辑、删除、执行，也不要写 todos/计划清单；不要发送文件或操作浏览器/桌面/手机，也不要通过 task / ACP / ask_agent 委派执行。若用户需要实际改动，可说明可使用计划或默认模式——不要指导用户如何操作界面控件。",
+      "plan_system_hint": "当前为计划（Plan）模式：用只读工具探索，只产出清晰的分步计划/规格说明。优先用 write_todos 与结构化计划文本。不要执行、改文件，也不要通过 task / ACP / ask_agent 委派。计划就绪后停止即可，产品界面会提供确认执行入口。不要让用户去切换模式按钮，也不要输出内部切换提示词。",
+      "craft_system_hint": "当前为默认（Craft）模式：在现有安全与 HITL 规则下，可使用已启用的完整工具集执行用户请求。",
+      "plan_brief_prefix": "用户已确认以下计划。请在默认模式下立即执行。除非澄清需要，否则不要原样复述整份计划：",
+      "tool_blocked": "当前为 {mode} 模式，工具 `{tool_name}` 不可用。如需执行改动或委派类操作，请切换到默认（Craft）模式。"
+    }
   }
 }
diff --git a/src/octop/infra/agents/conversation_mode.py b/src/octop/infra/agents/conversation_mode.py
new file mode 100644
index 000000000..7ba625f22
--- /dev/null
+++ b/src/octop/infra/agents/conversation_mode.py
@@ -0,0 +1,145 @@
+"""Conversation modes (Ask / Plan / Craft) — turn-scoped tool overlays (#616).
+
+Ask / Plan denylists cover **builtin** Octop / harness tool names only. MCP and
+custom connector tools are not name-denylisted here; mutating MCP tools may still
+appear unless blocked by agent ``tools_disabled`` or connector policy. Prefer an
+allowlist / mutating tag on MCP tools for stricter Ask if needed.
+"""
+
+from __future__ import annotations
+
+from typing import Literal
+
+from octop.infra.agents.tool_catalog import CRITICAL_TOOLS
+
+ConversationMode = Literal["ask", "plan", "craft"]
+
+DEFAULT_CONVERSATION_MODE: ConversationMode = "craft"
+
+_VALID_MODES: frozenset[str] = frozenset({"ask", "plan", "craft"})
+
+# Mutating / side-effect tools blocked in Ask and Plan. Must not intersect CRITICAL_TOOLS.
+_ASK_PLAN_TOOLS_DISABLED: frozenset[str] = frozenset(
+    {
+        "write_file",
+        "edit_file",
+        "execute",
+        "write_env_file",
+        "send_file_to_user",
+        "browser_use",
+        "desktop_screenshot",
+        "generate_image",
+        "generate_video",
+        "mobile_screenshot",
+        "mobile_tap",
+        "mobile_swipe",
+        "mobile_launch_app",
+        "mobile_ui_dump",
+        "mobile_handoff_to_user",
+        # cron mutations
+        "cronjob_create",
+        "cronjob_update",
+        "cronjob_delete",
+        "cronjob_run_now",
+    }
+)
+
+# Ask-only: no structured plan todos (Plan keeps write_todos).
+_ASK_ONLY_TOOLS_DISABLED: frozenset[str] = frozenset({"write_todos"})
+
+# Delegation / ACP / team tools that would escape Ask/Plan into a full-autonomy child.
+# ``task`` is CRITICAL for agent settings (cannot be permanently disabled) but MUST be
+# turn-blocked in Ask/Plan — same idea as Cursor Plan: explore + write the plan only.
+_ORCHESTRATION_ESCAPE_TOOLS: frozenset[str] = frozenset(
+    {
+        "task",
+        "acp_runner",
+        "ask_agent",
+    }
+)
+
+assert _ASK_PLAN_TOOLS_DISABLED.isdisjoint(CRITICAL_TOOLS)
+# write_todos is CRITICAL (cannot be permanently disabled in agent settings) but Ask
+# still turn-blocks it via middleware — same pattern as ``task``.
+
+
+def parse_conversation_mode(value: object | None) -> ConversationMode:
+    """Return a valid mode. ``None`` → craft; unknown non-null → ``ValueError``."""
+    if value is None:
+        return DEFAULT_CONVERSATION_MODE
+    if not isinstance(value, str) or value not in _VALID_MODES:
+        raise ValueError(f"invalid conversation_mode: {value!r}")
+    return value  # type: ignore[return-value]
+
+
+def conversation_mode_tools_disabled(mode: ConversationMode) -> frozenset[str]:
+    """Turn-scoped builtin denylist.
+
+    ``craft`` → empty; Plan → mutating + no orchestration escape; Ask → Plan set
+    plus ``write_todos``. Does **not** cover MCP/custom tool names (see module doc).
+    """
+    if mode == "craft":
+        return frozenset()
+    base = _ASK_PLAN_TOOLS_DISABLED | _ORCHESTRATION_ESCAPE_TOOLS
+    if mode == "ask":
+        return base | _ASK_ONLY_TOOLS_DISABLED
+    return base
+
+
+def resolve_conversation_mode(
+    *,
+    explicit: object | None = None,
+    thread_override: ConversationMode | None = None,
+    agent_default: ConversationMode | None = None,
+) -> ConversationMode:
+    """Resolve turn mode: explicit → thread sticky → agent default → craft.
+
+    Unknown non-null *explicit* values resolve to craft (compat) and do not fall
+    through to sticky/agent defaults.
+    """
+    if isinstance(explicit, str) and explicit in _VALID_MODES:
+        return explicit  # type: ignore[return-value]
+    if explicit is not None:
+        return DEFAULT_CONVERSATION_MODE
+    if thread_override is not None:
+        return thread_override
+    if agent_default is not None:
+        return agent_default
+    return DEFAULT_CONVERSATION_MODE
+
+
+DEFAULT_CONVERSATION_MODE_CONFIG_KEY = "default_conversation_mode"
+
+
+def default_conversation_mode_from_config(cfg: object | None) -> ConversationMode:
+    """Read agent default mode from config; missing/null → craft."""
+    if not isinstance(cfg, dict):
+        return DEFAULT_CONVERSATION_MODE
+    raw = cfg.get(DEFAULT_CONVERSATION_MODE_CONFIG_KEY)
+    if raw is None:
+        return DEFAULT_CONVERSATION_MODE
+    return parse_conversation_mode(raw)
+
+
+def normalize_config_default_conversation_mode(cfg: dict[str, object]) -> dict[str, object]:
+    """Validate/normalize ``default_conversation_mode`` in-place copy. Raises ``ValueError``."""
+    if DEFAULT_CONVERSATION_MODE_CONFIG_KEY not in cfg:
+        return dict(cfg)
+    out = dict(cfg)
+    raw = out.get(DEFAULT_CONVERSATION_MODE_CONFIG_KEY)
+    if raw is None:
+        out.pop(DEFAULT_CONVERSATION_MODE_CONFIG_KEY, None)
+        return out
+    out[DEFAULT_CONVERSATION_MODE_CONFIG_KEY] = parse_conversation_mode(raw)
+    return out
+
+
+def apply_default_conversation_mode(
+    config: dict[str, object],
+    mode: object | None,
+) -> dict[str, object]:
+    """Write ``default_conversation_mode`` when *mode* is set. Mutates and returns *config*."""
+    if mode is None:
+        return config
+    config[DEFAULT_CONVERSATION_MODE_CONFIG_KEY] = parse_conversation_mode(mode)
+    return config
diff --git a/src/octop/infra/agents/experts/market_creation.py b/src/octop/infra/agents/experts/market_creation.py
index a23c8907d..583148a4e 100644
--- a/src/octop/infra/agents/experts/market_creation.py
+++ b/src/octop/infra/agents/experts/market_creation.py
@@ -7,6 +7,7 @@
 from dataclasses import dataclass
 from typing import Any, Literal
 
+from octop.infra.agents.conversation_mode import apply_default_conversation_mode
 from octop.infra.agents.experts.catalog import (
     WORKSPACE_MANIFEST_PATH,
     build_create_spec_from_expert,
@@ -20,6 +21,7 @@
     install_skillset_template,
 )
 from octop.infra.errors import ErrorCode, OctopError
+from octop.infra.knowledge.default_open import apply_default_knowledge_base_ids
 from octop.infra.trajectory.settings import apply_enable_trajectory
 from octop.infra.utils.locale import resolve_user_locale
 
@@ -58,6 +60,8 @@ class SkillHubMarketAgentCreateOptions:
     top_p: float | None = None
     max_tokens: int | None = None
     enable_trajectory: bool = True
+    default_conversation_mode: str | None = None
+    default_knowledge_base_ids: list[str] | None = None
 
 
 @dataclass(frozen=True)
@@ -287,6 +291,8 @@ async def create_agent_from_skillhub_skillset(
     if options.backend:
         config_extra["backend"] = options.backend
     apply_enable_trajectory(config_extra, options.enable_trajectory)
+    apply_default_conversation_mode(config_extra, options.default_conversation_mode)
+    apply_default_knowledge_base_ids(config_extra, options.default_knowledge_base_ids)
 
     locale = resolve_user_locale(
         user_repo=server.services.user_repo,
diff --git a/src/octop/infra/agents/experts/published_creation.py b/src/octop/infra/agents/experts/published_creation.py
index ddfa9bcc9..45428ce42 100644
--- a/src/octop/infra/agents/experts/published_creation.py
+++ b/src/octop/infra/agents/experts/published_creation.py
@@ -13,6 +13,7 @@
 from psycopg import IntegrityError as PsycopgIntegrityError
 
 from octop.infra.agents.avatar import bind_workspace_avatar_icon_url
+from octop.infra.agents.conversation_mode import apply_default_conversation_mode
 from octop.infra.agents.experts.catalog import (
     MANIFEST_FILENAME,
     parse_task_examples,
@@ -29,6 +30,7 @@
 from octop.infra.agents.manager import AgentCreateSpec
 from octop.infra.db.repos.published_experts import PublishedExpertRow
 from octop.infra.errors import ErrorCode, OctopError
+from octop.infra.knowledge.default_open import apply_default_knowledge_base_ids
 from octop.infra.trajectory.settings import apply_enable_trajectory
 from octop.infra.users.identity import User
 from octop.infra.utils.ulid import new_ulid
@@ -50,6 +52,8 @@ class PublishedExpertInstallOptions:
     welcome_message: str | None = None
     runtime_config: dict[str, Any] | None = None
     enable_trajectory: bool = True
+    default_conversation_mode: str | None = None
+    default_knowledge_base_ids: list[str] | None = None
 
 
 def _snapshot_dir(services: Any, expert_id: str) -> Path:
@@ -342,6 +346,8 @@ async def install_published_expert(
     if options.backend:
         config_extra["backend"] = options.backend
     apply_enable_trajectory(config_extra, options.enable_trajectory)
+    apply_default_conversation_mode(config_extra, options.default_conversation_mode)
+    apply_default_knowledge_base_ids(config_extra, options.default_knowledge_base_ids)
 
     async def seed_snapshot(created_row: Any, workspace: Any) -> None:
         await seed_expert_directory(expert_dir=snapshot_dir, workspace=workspace)
diff --git a/src/octop/infra/agents/manager.py b/src/octop/infra/agents/manager.py
index 7c73bcdfb..4604e82f0 100644
--- a/src/octop/infra/agents/manager.py
+++ b/src/octop/infra/agents/manager.py
@@ -381,6 +381,8 @@ def __init__(
         self._mcp_tool_cache_guard = asyncio.Lock()
         # Sanitized plugin tool name → original label (per agent, rebuilt on reload).
         self._plugin_tool_labels: dict[str, dict[str, str]] = {}
+        # Sticky /mode overrides for IM (and slash) — not persisted across process restart.
+        self._thread_conversation_modes: dict[tuple[str, str], str] = {}
 
     def replace_persistence(self, repos: RepoBundle, config: OctopConfig) -> None:
         """Retarget repos/config and rebuild settings stores after control-plane rebind."""
@@ -1053,6 +1055,19 @@ def clear_thread_model(self, agent_id: str, thread_id: str) -> None:
         if self._harness_manager is not None:
             self._harness_manager.clear_thread_model(agent_id, thread_id)
 
+    def get_thread_conversation_mode(self, agent_id: str, thread_id: str) -> str | None:
+        """Sticky IM/slash conversation mode for *(agent_id, thread_id)*, if any."""
+        return self._thread_conversation_modes.get((agent_id, thread_id))
+
+    def set_thread_conversation_mode(self, agent_id: str, thread_id: str, mode: str) -> None:
+        from octop.infra.agents.conversation_mode import parse_conversation_mode
+
+        parsed = parse_conversation_mode(mode)
+        self._thread_conversation_modes[(agent_id, thread_id)] = parsed
+
+    def clear_thread_conversation_mode(self, agent_id: str, thread_id: str) -> None:
+        self._thread_conversation_modes.pop((agent_id, thread_id), None)
+
     def resolve_fallback_model_ref(self) -> str | None:
         """Settings active model when usable, else the first enabled catalog model."""
         name, model_id = self._repos.settings_repo.get_active_model()
@@ -2701,6 +2716,7 @@ def _build_harness_config(self, row: AgentRow) -> HarnessAgentConfig:
 
         from octop.infra.agents.middleware.binary_read_guard import BinaryReadGuardMiddleware
         from octop.infra.agents.middleware.browser_profile import BrowserProfileMiddleware
+        from octop.infra.agents.middleware.conversation_mode import ConversationModeMiddleware
         from octop.infra.agents.middleware.reasoning import ReasoningRequestMiddleware
         from octop.infra.agents.middleware.thread_artifacts import ThreadArtifactsMiddleware
         from octop.infra.agents.middleware.token_quota import TokenQuotaMiddleware
@@ -2713,6 +2729,7 @@ def _build_harness_config(self, row: AgentRow) -> HarnessAgentConfig:
         # BinaryReadGuard stays Octop-specific (inbound/attachment product policy).
         # ThreadArtifacts writes workspace paths onto threads after successful tools.
         # WorkspaceImageMaterialize expands path-only vision refs at model-call time.
+        # ConversationMode applies turn Ask/Plan denylist + system hint from configurable.
         agent_middleware: list[Any] = [
             *plugin_middleware,
             TokenQuotaMiddleware(
@@ -2721,6 +2738,7 @@ def _build_harness_config(self, row: AgentRow) -> HarnessAgentConfig:
             ),
             ReasoningRequestMiddleware(),
             KnowledgeSearchHintMiddleware(),
+            ConversationModeMiddleware(),
             BrowserProfileMiddleware(),
             BinaryReadGuardMiddleware(),
             WorkspaceImageMaterializeMiddleware(workspace=ws),
diff --git a/src/octop/infra/agents/middleware/conversation_mode.py b/src/octop/infra/agents/middleware/conversation_mode.py
new file mode 100644
index 000000000..562128cfa
--- /dev/null
+++ b/src/octop/infra/agents/middleware/conversation_mode.py
@@ -0,0 +1,198 @@
+"""Turn-scoped Ask / Plan / Craft: filter tools, inject system hint, block denylist calls.
+
+Denylist covers builtin tool names only — see ``conversation_mode`` module docs for MCP limits.
+"""
+
+from __future__ import annotations
+
+import logging
+from collections.abc import Awaitable, Callable
+from typing import Any
+
+from langchain.agents.middleware import AgentMiddleware, ModelRequest, ModelResponse
+from langchain_core.messages import SystemMessage, ToolMessage
+from langchain_core.tools.base import BaseTool
+from langgraph.config import get_config
+from langgraph.prebuilt.tool_node import ToolCallRequest
+from langgraph.types import Command
+
+from octop.infra.agents.conversation_mode import (
+    DEFAULT_CONVERSATION_MODE,
+    ConversationMode,
+    conversation_mode_tools_disabled,
+    parse_conversation_mode,
+)
+
+logger = logging.getLogger(__name__)
+
+CONFIG_MODE_KEY = "conversation_mode"
+CONFIG_HINT_KEY = "conversation_mode_hint"
+
+
+def _tool_name(tool: BaseTool | dict[str, Any]) -> str:
+    if isinstance(tool, dict):
+        fn = tool.get("function")
+        if isinstance(fn, dict) and fn.get("name"):
+            return str(fn["name"])
+        if tool.get("name"):
+            return str(tool["name"])
+        return ""
+    return str(getattr(tool, "name", "") or "")
+
+
+def mode_from_configurable(configurable: dict[str, Any] | None) -> ConversationMode:
+    """Resolve turn mode from LangGraph configurable; unknown → craft."""
+    raw = (configurable or {}).get(CONFIG_MODE_KEY)
+    if raw is None:
+        return DEFAULT_CONVERSATION_MODE
+    try:
+        return parse_conversation_mode(raw)
+    except ValueError:
+        return DEFAULT_CONVERSATION_MODE
+
+
+def _current_mode() -> ConversationMode:
+    cfg = get_config().get("configurable") or {}
+    if not isinstance(cfg, dict):
+        return DEFAULT_CONVERSATION_MODE
+    return mode_from_configurable(cfg)
+
+
+def _current_hint() -> str:
+    cfg = get_config().get("configurable") or {}
+    if not isinstance(cfg, dict):
+        return ""
+    raw = cfg.get(CONFIG_HINT_KEY)
+    return raw.strip() if isinstance(raw, str) else ""
+
+
+def _hint_already_present(content: str | list[Any] | Any, hint: str) -> bool:
+    if isinstance(content, str):
+        return hint in content
+    if isinstance(content, list):
+        for block in content:
+            if isinstance(block, str) and hint in block:
+                return True
+            if isinstance(block, dict):
+                text = block.get("text")
+                if isinstance(text, str) and hint in text:
+                    return True
+        return False
+    return hint in str(content)
+
+
+def _append_system_hint(request: ModelRequest[Any], hint: str) -> ModelRequest[Any]:
+    cleaned = hint.strip()
+    if not cleaned:
+        return request
+    existing = request.system_message
+    if existing is None:
+        return request.override(system_message=SystemMessage(content=cleaned))
+    content = existing.content
+    if _hint_already_present(content, cleaned):
+        return request
+    if isinstance(content, str):
+        merged: str | list[Any] = f"{content.rstrip()}\n\n{cleaned}"
+    elif isinstance(content, list):
+        merged = [*content, {"type": "text", "text": cleaned}]
+    else:
+        merged = f"{content}\n\n{cleaned}"
+    return request.override(system_message=SystemMessage(content=merged))
+
+
+def _filter_tools(request: ModelRequest[Any], disabled: frozenset[str]) -> ModelRequest[Any]:
+    if not disabled:
+        return request
+    tools_in = list(request.tools or [])
+    if not tools_in:
+        return request
+    filtered = [t for t in tools_in if _tool_name(t) not in disabled]
+    if len(filtered) == len(tools_in):
+        return request
+    return request.override(tools=filtered)
+
+
+def apply_conversation_mode_to_request(request: ModelRequest[Any]) -> ModelRequest[Any]:
+    """Filter Ask/Plan denylist tools and append the stamped system hint."""
+    mode = _current_mode()
+    out = _filter_tools(request, conversation_mode_tools_disabled(mode))
+    return _append_system_hint(out, _current_hint())
+
+
+def _blocked_tool_message(
+    *, tool_name: str, mode: ConversationMode, tool_call_id: str
+) -> ToolMessage:
+    from octop.i18n.domains.conversation import conversation_mode_tool_blocked
+
+    cfg = get_config().get("configurable") or {}
+    locale = "en"
+    if isinstance(cfg, dict):
+        raw_locale = cfg.get("locale")
+        if isinstance(raw_locale, str) and raw_locale.strip():
+            locale = raw_locale.strip()
+    return ToolMessage(
+        content=conversation_mode_tool_blocked(tool_name, mode, locale),
+        tool_call_id=tool_call_id,
+        status="error",
+    )
+
+
+class ConversationModeMiddleware(AgentMiddleware[Any, Any]):
+    """Apply turn ``conversation_mode`` from configurable — no agent-global mutation."""
+
+    def wrap_model_call(
+        self,
+        request: ModelRequest[Any],
+        handler: Callable[[ModelRequest[Any]], ModelResponse[Any]],
+    ) -> ModelResponse[Any]:
+        return handler(apply_conversation_mode_to_request(request))
+
+    async def awrap_model_call(
+        self,
+        request: ModelRequest[Any],
+        handler: Callable[[ModelRequest[Any]], Awaitable[ModelResponse[Any]]],
+    ) -> ModelResponse[Any]:
+        return await handler(apply_conversation_mode_to_request(request))
+
+    def wrap_tool_call(
+        self,
+        request: ToolCallRequest,
+        handler: Callable[[ToolCallRequest], ToolMessage | Command[Any]],
+    ) -> ToolMessage | Command[Any]:
+        mode = _current_mode()
+        disabled = conversation_mode_tools_disabled(mode)
+        tool_name = str(request.tool_call.get("name") or "")
+        if tool_name and tool_name in disabled:
+            logger.info("ConversationMode blocked tool %s (mode=%s)", tool_name, mode)
+            return _blocked_tool_message(
+                tool_name=tool_name,
+                mode=mode,
+                tool_call_id=str(request.tool_call.get("id") or ""),
+            )
+        return handler(request)
+
+    async def awrap_tool_call(
+        self,
+        request: ToolCallRequest,
+        handler: Callable[[ToolCallRequest], Awaitable[ToolMessage | Command[Any]]],
+    ) -> ToolMessage | Command[Any]:
+        mode = _current_mode()
+        disabled = conversation_mode_tools_disabled(mode)
+        tool_name = str(request.tool_call.get("name") or "")
+        if tool_name and tool_name in disabled:
+            logger.info("ConversationMode blocked tool %s (mode=%s)", tool_name, mode)
+            return _blocked_tool_message(
+                tool_name=tool_name,
+                mode=mode,
+                tool_call_id=str(request.tool_call.get("id") or ""),
+            )
+        return await handler(request)
+
+
+__all__ = [
+    "CONFIG_HINT_KEY",
+    "CONFIG_MODE_KEY",
+    "ConversationModeMiddleware",
+    "apply_conversation_mode_to_request",
+    "mode_from_configurable",
+]
diff --git a/src/octop/infra/agents/plan_artifact.py b/src/octop/infra/agents/plan_artifact.py
new file mode 100644
index 000000000..d9e0852c6
--- /dev/null
+++ b/src/octop/infra/agents/plan_artifact.py
@@ -0,0 +1,53 @@
+"""Plan brief artifact for Plan → Craft handoff (#616 P5)."""
+
+from __future__ import annotations
+
+from typing import Any
+
+# Minimal HumanMessage text when the dashboard sends plan_brief with no chat bubble.
+# Full plan lives in the system hint; this only gives the model a user-turn trigger.
+PLAN_EXECUTE_USER_TRIGGER = "Execute the approved plan now."
+
+
+def format_plan_brief(*, summary: str = "", todos: list[dict[str, Any]] | None = None) -> str:
+    """Serialize an approved plan into a Craft-turn brief (markdown)."""
+    lines: list[str] = ["## Approved plan", ""]
+    text = (summary or "").strip()
+    if text:
+        lines.extend([text, ""])
+    items = todos or []
+    if items:
+        lines.append("### Steps")
+        step_num = 0
+        for raw in items:
+            if isinstance(raw, str):
+                content = raw.strip()
+            elif isinstance(raw, dict):
+                content = str(
+                    raw.get("content") or raw.get("text") or raw.get("title") or ""
+                ).strip()
+            else:
+                content = ""
+            if not content:
+                continue
+            status = ""
+            if isinstance(raw, dict) and raw.get("status"):
+                status = f" ({raw['status']})"
+            step_num += 1
+            lines.append(f"{step_num}. {content}{status}")
+        lines.append("")
+    lines.append("Execute this plan now.")
+    return "\n".join(lines).strip() + "\n"
+
+
+def parse_plan_brief_steps(brief: str) -> list[str]:
+    """Extract numbered step lines from a brief produced by :func:`format_plan_brief`."""
+    steps: list[str] = []
+    for line in (brief or "").splitlines():
+        stripped = line.strip()
+        if len(stripped) >= 3 and stripped[0].isdigit() and ". " in stripped[:6]:
+            _, _, rest = stripped.partition(". ")
+            content = rest.strip()
+            if content:
+                steps.append(content)
+    return steps
diff --git a/src/octop/infra/gateway/process/message_keys.py b/src/octop/infra/gateway/process/message_keys.py
index de4f55b56..1bcb17e2d 100644
--- a/src/octop/infra/gateway/process/message_keys.py
+++ b/src/octop/infra/gateway/process/message_keys.py
@@ -11,6 +11,8 @@
 # Persisted on HumanMessage.additional_kwargs for dashboard history UI.
 COMPOSER_CTX_KEY = "octop_composer_context"
 INBOUND_ATTACHMENTS_KEY = "octop_inbound_attachments"
+# Plan→Craft silent handoff: history should not render this user turn as a bubble.
+UI_HIDDEN_KEY = "octop_ui_hidden"
 # Must match harness_agent.messages.CHECKPOINT_TS_KEY (epoch-ms).
 CHECKPOINT_TS_KEY = "checkpoint_ts"
 # Persisted on AIMessage.additional_kwargs when a stream fails mid-turn.
diff --git a/src/octop/infra/gateway/process/processor.py b/src/octop/infra/gateway/process/processor.py
index 358a5fe86..670b93c02 100644
--- a/src/octop/infra/gateway/process/processor.py
+++ b/src/octop/infra/gateway/process/processor.py
@@ -21,6 +21,9 @@
 from langchain_core.messages import AIMessage, HumanMessage
 
 from octop.i18n.domains.stream import format_stream_error
+from octop.infra.agents.conversation_mode import (
+    ConversationMode,
+)
 from octop.infra.agents.profile import parse_config_json
 from octop.infra.agents.providers.reasoning import reasoning_request_parameters
 from octop.infra.errors import OctopError
@@ -59,7 +62,6 @@
 from octop.infra.gateway.slash.ctx import SlashCtx, build_slash_ctx
 from octop.infra.gateway.slash.parser import parse_slash
 from octop.infra.gateway.slash.runner import try_handle_slash
-from octop.infra.knowledge.default_open import stamp_turn_knowledge_config
 from octop.infra.trajectory.settings import agent_trajectory_enabled
 from octop.infra.users.preferences import (
     get_model_reasoning_from_json,
@@ -851,6 +853,13 @@ async def __call__(self, msg: InboundMessage) -> AsyncIterator[MessageEvent]:
             locale=locale,
             agent_id=agent_id,
         )
+        self._attach_conversation_mode_config(
+            request,
+            msg.metadata,
+            locale=locale,
+            agent_id=agent_id,
+            thread_id=thread_id,
+        )
         if mcp_servers:
             request["mcp_servers"] = mcp_servers
 
@@ -1199,6 +1208,7 @@ async def _build_dashboard_request(
         from octop.infra.gateway.process.message_keys import (  # noqa: PLC0415
             COMPOSER_CTX_KEY,
             INBOUND_ATTACHMENTS_KEY,
+            UI_HIDDEN_KEY,
         )
 
         media_backend = media_backend_for_agent(self._agent_manager, agent_id)
@@ -1235,6 +1245,10 @@ async def _build_dashboard_request(
         attachments = meta.get(INBOUND_ATTACHMENTS_KEY)
         if isinstance(attachments, list) and attachments:
             message_kwargs[INBOUND_ATTACHMENTS_KEY] = attachments
+        raw_brief = meta.get("plan_brief")
+        if isinstance(raw_brief, str) and raw_brief.strip():
+            # Dashboard PlanReady CTA — do not render the trigger as a user bubble.
+            message_kwargs[UI_HIDDEN_KEY] = True
 
         explicit_mcp = meta.get("mcp_servers")
         # Dashboard always sends mcp_servers (possibly []); trust that list so
@@ -1294,6 +1308,13 @@ async def _build_dashboard_request(
             locale=locale,
             agent_id=agent_id,
         )
+        self._attach_conversation_mode_config(
+            request,
+            meta,
+            locale=locale,
+            agent_id=agent_id,
+            thread_id=thread_id,
+        )
 
         if mcp_servers:
             request["mcp_servers"] = mcp_servers
@@ -1319,16 +1340,88 @@ def _attach_turn_knowledge_config(
             if is_admin
             else self._knowledge_services.knowledge_repo.list_visible(user_id)
         )
-        extra_ids = self._agent_manager.default_knowledge_base_ids(agent_id) if agent_id else None
-        stamp_turn_knowledge_config(
-            request,
-            visible_bases=bases,
-            explicit_ids=explicit_ids,
+        agent_defaults: list[str] | None = None
+        if agent_id and self._agent_manager is not None:
+            from octop.infra.knowledge.default_open import (  # noqa: PLC0415
+                default_knowledge_base_ids_from_config,
+                merge_knowledge_base_ids,
+            )
+            from octop.infra.knowledge.hint import catalog_for_selected_bases  # noqa: PLC0415
+
+            agent_defaults = (
+                default_knowledge_base_ids_from_config(self._agent_manager.get_config(agent_id))
+                or None
+            )
+        selected_ids = merge_knowledge_base_ids(
+            bases,
+            explicit_ids,
             owner_user_id=user_id,
-            extra_ids=extra_ids,
-            is_admin=is_admin,
-            locale=locale,
+            agent_default_ids=agent_defaults,
         )
+        configurable = dict(request.get("configurable") or {})
+        configurable["knowledge_base_ids"] = selected_ids
+        configurable["knowledge_base_catalog"] = catalog_for_selected_bases(bases, selected_ids)
+        configurable["user_is_admin"] = is_admin
+        configurable["locale"] = locale
+        request["configurable"] = configurable
+
+    def _resolve_conversation_mode(
+        self,
+        meta: dict[str, Any] | None,
+        *,
+        agent_id: str | None = None,
+        thread_id: str | None = None,
+    ) -> ConversationMode:
+        """explicit meta → thread /mode sticky → agent default → craft."""
+        from octop.infra.agents.conversation_mode import (  # noqa: PLC0415
+            default_conversation_mode_from_config,
+            resolve_conversation_mode,
+        )
+
+        explicit = (meta or {}).get("conversation_mode")
+        thread_override: ConversationMode | None = None
+        if agent_id and thread_id and self._agent_manager is not None:
+            raw_sticky = self._agent_manager.get_thread_conversation_mode(agent_id, thread_id)
+            if raw_sticky in ("ask", "plan", "craft"):
+                thread_override = raw_sticky  # type: ignore[assignment]
+        agent_default: ConversationMode | None = None
+        if agent_id and self._agent_manager is not None:
+            agent_default = default_conversation_mode_from_config(
+                self._agent_manager.get_config(agent_id)
+            )
+        return resolve_conversation_mode(
+            explicit=explicit,
+            thread_override=thread_override,
+            agent_default=agent_default,
+        )
+
+    def _attach_conversation_mode_config(
+        self,
+        request: dict[str, Any],
+        meta: dict[str, Any] | None,
+        *,
+        locale: str | None = None,
+        agent_id: str | None = None,
+        thread_id: str | None = None,
+    ) -> None:
+        """Stamp resolved conversation_mode (+ localized system hint) onto configurable."""
+        from octop.i18n.domains.conversation import (
+            conversation_mode_plan_brief_block,
+            conversation_mode_system_hint,
+        )
+
+        mode = self._resolve_conversation_mode(meta, agent_id=agent_id, thread_id=thread_id)
+        configurable = dict(request.get("configurable") or {})
+        configurable["conversation_mode"] = mode
+        hint_locale = locale or str(configurable.get("locale") or "en")
+        hint = conversation_mode_system_hint(mode, hint_locale)
+        raw_brief = (meta or {}).get("plan_brief")
+        brief = raw_brief.strip() if isinstance(raw_brief, str) else ""
+        if brief:
+            configurable["plan_brief"] = brief
+            hint = f"{hint}\n\n{conversation_mode_plan_brief_block(brief, hint_locale)}"
+        configurable["conversation_mode_hint"] = hint
+        request["configurable"] = configurable
 
     async def _resolve_turn_mcp_servers(
         self,
diff --git a/src/octop/infra/gateway/slash/catalog.py b/src/octop/infra/gateway/slash/catalog.py
index 67a1f2ec6..d37d4865c 100644
--- a/src/octop/infra/gateway/slash/catalog.py
+++ b/src/octop/infra/gateway/slash/catalog.py
@@ -87,6 +87,14 @@ def label_for(self, locale: str) -> str:
         tone="violet",
         category="core",
     ),
+    SlashCommandSpec(
+        name="mode",
+        aliases=("ask", "plan", "craft"),
+        usage="/mode [ask|plan|craft|reset]",
+        icon="Shield",
+        tone="violet",
+        category="core",
+    ),
     SlashCommandSpec(
         name="new",
         aliases=("clear",),
diff --git a/src/octop/infra/gateway/slash/dispatcher.py b/src/octop/infra/gateway/slash/dispatcher.py
index 3820a5791..d0a7c8141 100644
--- a/src/octop/infra/gateway/slash/dispatcher.py
+++ b/src/octop/infra/gateway/slash/dispatcher.py
@@ -74,6 +74,21 @@ def clear_thread_model_override(self, ctx: SlashCtx, thread_id: str) -> None:
         if ctx.agent_manager is not None:
             ctx.agent_manager.clear_thread_model(ctx.agent_id, thread_id)
 
+    def get_thread_conversation_mode_override(self, ctx: SlashCtx, thread_id: str) -> str | None:
+        if ctx.agent_manager is None:
+            return None
+        return ctx.agent_manager.get_thread_conversation_mode(ctx.agent_id, thread_id)
+
+    def set_thread_conversation_mode_override(
+        self, ctx: SlashCtx, thread_id: str, mode: str
+    ) -> None:
+        if ctx.agent_manager is not None:
+            ctx.agent_manager.set_thread_conversation_mode(ctx.agent_id, thread_id, mode)
+
+    def clear_thread_conversation_mode_override(self, ctx: SlashCtx, thread_id: str) -> None:
+        if ctx.agent_manager is not None:
+            ctx.agent_manager.clear_thread_conversation_mode(ctx.agent_id, thread_id)
+
 
 def build_default_dispatcher() -> SlashDispatcher:
     from octop.infra.gateway.slash.handlers import register_all  # noqa: PLC0415
diff --git a/src/octop/infra/gateway/slash/handlers/composite.py b/src/octop/infra/gateway/slash/handlers/composite.py
index 926fa8c49..d4f3a3a67 100644
--- a/src/octop/infra/gateway/slash/handlers/composite.py
+++ b/src/octop/infra/gateway/slash/handlers/composite.py
@@ -322,10 +322,61 @@ async def cmd_model(d: SlashDispatcher, cmd: SlashCommand, ctx: SlashCtx, sink:
     await sink.text(tr("model.set", lang, model=name))
 
 
+_MODE_ALIASES = {
+    "ask": "ask",
+    "plan": "plan",
+    "craft": "craft",
+    "default": "craft",
+    "问答": "ask",
+    "计划": "plan",
+    "默认": "craft",
+}
+
+
+async def cmd_mode(d: SlashDispatcher, cmd: SlashCommand, ctx: SlashCtx, sink: SlashSink) -> None:
+    """Set sticky conversation mode for this thread (IM + slash)."""
+    from octop.infra.agents.conversation_mode import (  # noqa: PLC0415
+        default_conversation_mode_from_config,
+    )
+
+    lang = lang_of(ctx)
+    tid = await ensure_thread_id(ctx)
+    # /ask /plan /craft as aliases — treat command name as the mode when args empty.
+    raw = cmd.args.strip() or (cmd.name if cmd.name in ("ask", "plan", "craft") else "")
+    if not raw:
+        override = d.get_thread_conversation_mode_override(ctx, tid)
+        agent_default = "craft"
+        if ctx.agent_manager is not None:
+            agent_default = default_conversation_mode_from_config(
+                ctx.agent_manager.get_config(ctx.agent_id)
+            )
+        if override:
+            await sink.text(tr("mode.override", lang, mode=override))
+        else:
+            await sink.text(tr("mode.current_default", lang, mode=agent_default))
+        return
+    key = raw.lower()
+    if key == "reset":
+        d.clear_thread_conversation_mode_override(ctx, tid)
+        if hasattr(sink, "action"):
+            await sink.action("clear_conversation_mode")
+        await sink.text(tr("mode.cleared", lang))
+        return
+    mode = _MODE_ALIASES.get(key) or _MODE_ALIASES.get(raw)
+    if mode is None:
+        await sink.text(tr("mode.usage", lang))
+        return
+    d.set_thread_conversation_mode_override(ctx, tid, mode)
+    if hasattr(sink, "action"):
+        await sink.action("set_conversation_mode", mode=mode)
+    await sink.text(tr("mode.set", lang, mode=mode))
+
+
 COMPOSITE_HANDLERS: dict[str, GatewayHandler] = {
     "compact": cmd_compact,
     "history": cmd_history,
     "status": cmd_status,
     "model": cmd_model,
     "models": cmd_model,
+    "mode": cmd_mode,
 }
diff --git a/src/octop/infra/knowledge/default_open.py b/src/octop/infra/knowledge/default_open.py
index 50d0d7b3a..0071a5ed3 100644
--- a/src/octop/infra/knowledge/default_open.py
+++ b/src/octop/infra/knowledge/default_open.py
@@ -7,6 +7,64 @@
 
 from octop.infra.db.repos.knowledge import KnowledgeBaseRow
 
+DEFAULT_KNOWLEDGE_BASE_IDS_CONFIG_KEY = "default_knowledge_base_ids"
+
+
+def default_knowledge_base_ids_from_config(cfg: object | None) -> list[str]:
+    """Read agent ``default_knowledge_base_ids``; missing/invalid → empty."""
+    if not isinstance(cfg, dict):
+        return []
+    raw = cfg.get(DEFAULT_KNOWLEDGE_BASE_IDS_CONFIG_KEY)
+    if not isinstance(raw, list):
+        return []
+    out: list[str] = []
+    seen: set[str] = set()
+    for item in raw:
+        if not isinstance(item, str):
+            continue
+        kid = item.strip()
+        if not kid or kid in seen:
+            continue
+        seen.add(kid)
+        out.append(kid)
+    return out
+
+
+def normalize_config_default_knowledge_base_ids(cfg: dict[str, object]) -> dict[str, object]:
+    """Validate/normalize ``default_knowledge_base_ids``; raises ``ValueError`` on bad types."""
+    if DEFAULT_KNOWLEDGE_BASE_IDS_CONFIG_KEY not in cfg:
+        return dict(cfg)
+    out = dict(cfg)
+    raw = out.get(DEFAULT_KNOWLEDGE_BASE_IDS_CONFIG_KEY)
+    if raw is None:
+        out.pop(DEFAULT_KNOWLEDGE_BASE_IDS_CONFIG_KEY, None)
+        return out
+    if not isinstance(raw, list):
+        raise ValueError("default_knowledge_base_ids must be a list of strings")
+    cleaned = default_knowledge_base_ids_from_config({DEFAULT_KNOWLEDGE_BASE_IDS_CONFIG_KEY: raw})
+    if cleaned:
+        out[DEFAULT_KNOWLEDGE_BASE_IDS_CONFIG_KEY] = cleaned
+    else:
+        out.pop(DEFAULT_KNOWLEDGE_BASE_IDS_CONFIG_KEY, None)
+    return out
+
+
+def apply_default_knowledge_base_ids(
+    config: dict[str, object],
+    ids: object | None,
+) -> dict[str, object]:
+    """Write ``default_knowledge_base_ids`` when *ids* is set. Mutates and returns *config*."""
+    if ids is None:
+        return config
+    if not isinstance(ids, list):
+        raise ValueError("default_knowledge_base_ids must be a list of strings")
+    cleaned = default_knowledge_base_ids_from_config({DEFAULT_KNOWLEDGE_BASE_IDS_CONFIG_KEY: ids})
+    if cleaned:
+        config[DEFAULT_KNOWLEDGE_BASE_IDS_CONFIG_KEY] = cleaned
+    else:
+        config.pop(DEFAULT_KNOWLEDGE_BASE_IDS_CONFIG_KEY, None)
+    return config
+
 
 def merge_knowledge_base_ids(
     visible_bases: Sequence[KnowledgeBaseRow],
@@ -14,12 +72,14 @@ def merge_knowledge_base_ids(
     *,
     owner_user_id: int,
     extra_ids: Sequence[str] | None = None,
+    agent_default_ids: Sequence[str] | None = None,
 ) -> list[str]:
     """Use the actor's own default-open bases when a turn omits a list.
 
     ``default_open`` is per-owner preference: shared bases marked default-open
     are auto-injected only for the creating user, not for other viewers.
-    ``extra_ids`` (expert composer picks) are unioned in when still visible.
+    ``extra_ids`` (expert composer picks) and ``agent_default_ids`` (agent
+    config defaults) are unioned in when still visible.
     """
     if explicit_ids is not None:
         return list(explicit_ids)
@@ -28,8 +88,11 @@ def merge_knowledge_base_ids(
     for base in visible_bases:
         if base.default_open and int(base.owner_user_id) == int(owner_user_id):
             selected.append(base.id)
-    for kb_id in extra_ids or []:
-        text = str(kb_id).strip()
+    extras: list[str] = []
+    extras.extend(str(kb_id) for kb_id in extra_ids or [])
+    extras.extend(str(kb_id) for kb_id in agent_default_ids or [])
+    for kb_id in extras:
+        text = kb_id.strip()
         if text and text in visible_ids and text not in selected:
             selected.append(text)
     return selected
diff --git a/tests/unit/agents/test_agent_manager.py b/tests/unit/agents/test_agent_manager.py
index 2e4d708ee..a46abc89d 100644
--- a/tests/unit/agents/test_agent_manager.py
+++ b/tests/unit/agents/test_agent_manager.py
@@ -242,12 +242,14 @@ def test_build_harness_config_includes_cronjob_tools_when_cron_manager_set(
 def test_build_harness_config_includes_search_knowledge_without_cron(
     manager: AgentManager,
 ) -> None:
+    from octop.infra.agents.middleware.conversation_mode import ConversationModeMiddleware
     from octop.infra.knowledge.hint import KnowledgeSearchHintMiddleware
 
     cfg = manager._build_harness_config(_row(agent_id="AGT001"))
     assert cfg.tools is not None
     assert {t.name for t in cfg.tools} == {"search_knowledge"}
     assert any(isinstance(item, KnowledgeSearchHintMiddleware) for item in (cfg.middleware or []))
+    assert any(isinstance(item, ConversationModeMiddleware) for item in (cfg.middleware or []))
 
 
 def test_build_harness_config_defaults_local_shell_backend(manager: AgentManager) -> None:
diff --git a/tests/unit/agents/test_conversation_mode.py b/tests/unit/agents/test_conversation_mode.py
new file mode 100644
index 000000000..dd2064706
--- /dev/null
+++ b/tests/unit/agents/test_conversation_mode.py
@@ -0,0 +1,135 @@
+"""Unit tests for conversation mode parse + turn-scoped tool overlays (Sprint 0 / #616)."""
+
+from __future__ import annotations
+
+import pytest
+
+from octop.infra.agents.conversation_mode import (
+    DEFAULT_CONVERSATION_MODE,
+    conversation_mode_tools_disabled,
+    parse_conversation_mode,
+    resolve_conversation_mode,
+)
+
+
+def test_parse_conversation_mode_accepts_ask_plan_craft() -> None:
+    assert parse_conversation_mode("ask") == "ask"
+    assert parse_conversation_mode("plan") == "plan"
+    assert parse_conversation_mode("craft") == "craft"
+
+
+def test_parse_conversation_mode_rejects_unknown() -> None:
+    with pytest.raises(ValueError):
+        parse_conversation_mode("agent")
+    with pytest.raises(ValueError):
+        parse_conversation_mode("")
+    with pytest.raises(ValueError):
+        parse_conversation_mode(123)
+
+
+def test_parse_conversation_mode_none_defaults_to_craft() -> None:
+    assert parse_conversation_mode(None) == "craft"
+    assert parse_conversation_mode(None) == DEFAULT_CONVERSATION_MODE
+
+
+def test_ask_tools_disabled_includes_mutating_builtins() -> None:
+    disabled = conversation_mode_tools_disabled("ask")
+    for name in (
+        "write_file",
+        "edit_file",
+        "execute",
+        "write_env_file",
+        "send_file_to_user",
+        "browser_use",
+        "desktop_screenshot",
+        "generate_image",
+        "generate_video",
+        "mobile_tap",
+        "mobile_swipe",
+        "mobile_launch_app",
+        "mobile_screenshot",
+        "mobile_ui_dump",
+        "mobile_handoff_to_user",
+    ):
+        assert name in disabled, name
+
+
+def test_ask_tools_disabled_keeps_explore_tools_blocks_orchestration_escape() -> None:
+    disabled = conversation_mode_tools_disabled("ask")
+    for name in ("ls", "read_file", "glob", "grep"):
+        assert name not in disabled, name
+    assert "write_todos" in disabled  # Ask is Q&A only — no plan todos
+    assert "search_knowledge" not in disabled
+    assert "web_fetch" not in disabled
+    # CRITICAL ``task`` must still be turn-blocked so Plan/Ask cannot escape via subagents.
+    for name in ("task", "acp_runner", "ask_agent"):
+        assert name in disabled, name
+
+
+def test_craft_tools_disabled_is_empty() -> None:
+    assert conversation_mode_tools_disabled("craft") == frozenset()
+
+
+def test_plan_allows_write_todos_and_forbids_mutating_and_delegation() -> None:
+    """Plan keeps write_todos; blocks execute/write/edit and task/ACP escape."""
+    disabled = conversation_mode_tools_disabled("plan")
+    assert "write_todos" not in disabled
+    assert "execute" in disabled
+    assert "write_file" in disabled
+    assert "edit_file" in disabled
+    assert "task" in disabled
+    assert "acp_runner" in disabled
+    assert "ask_agent" in disabled
+    # Ask is stricter than Plan on write_todos.
+    assert "write_todos" in conversation_mode_tools_disabled("ask")
+
+
+def test_resolve_conversation_mode_priority() -> None:
+    """explicit → thread sticky → agent default → craft."""
+    assert resolve_conversation_mode(explicit="ask") == "ask"
+    assert resolve_conversation_mode(explicit=None, thread_override="plan") == "plan"
+    assert (
+        resolve_conversation_mode(explicit=None, thread_override=None, agent_default="ask") == "ask"
+    )
+    assert resolve_conversation_mode(explicit=None) == "craft"
+    # Unknown explicit stays craft (compat); does not fall through.
+    assert resolve_conversation_mode(explicit="nope", agent_default="ask") == "craft"
+    # Explicit craft wins over agent ask.
+    assert resolve_conversation_mode(explicit="craft", agent_default="ask") == "craft"
+
+
+def test_default_conversation_mode_from_config() -> None:
+    from octop.infra.agents.conversation_mode import default_conversation_mode_from_config
+
+    assert default_conversation_mode_from_config(None) == "craft"
+    assert default_conversation_mode_from_config({}) == "craft"
+    assert default_conversation_mode_from_config({"default_conversation_mode": "ask"}) == "ask"
+    with pytest.raises(ValueError):
+        default_conversation_mode_from_config({"default_conversation_mode": "agent"})
+
+
+def test_normalize_config_default_conversation_mode() -> None:
+    from octop.infra.agents.conversation_mode import (
+        normalize_config_default_conversation_mode,
+    )
+
+    assert normalize_config_default_conversation_mode({"x": 1}) == {"x": 1}
+    assert normalize_config_default_conversation_mode({"default_conversation_mode": "plan"}) == {
+        "default_conversation_mode": "plan"
+    }
+    cleaned = normalize_config_default_conversation_mode(
+        {"default_conversation_mode": None, "x": 1}
+    )
+    assert "default_conversation_mode" not in cleaned
+    assert cleaned["x"] == 1
+    with pytest.raises(ValueError):
+        normalize_config_default_conversation_mode({"default_conversation_mode": "nope"})
+
+
+def test_apply_default_conversation_mode() -> None:
+    from octop.infra.agents.conversation_mode import apply_default_conversation_mode
+
+    cfg: dict[str, object] = {"x": 1}
+    assert apply_default_conversation_mode(cfg, None) == {"x": 1}
+    apply_default_conversation_mode(cfg, "ask")
+    assert cfg["default_conversation_mode"] == "ask"
diff --git a/tests/unit/agents/test_conversation_mode_middleware.py b/tests/unit/agents/test_conversation_mode_middleware.py
new file mode 100644
index 000000000..5508b0ad0
--- /dev/null
+++ b/tests/unit/agents/test_conversation_mode_middleware.py
@@ -0,0 +1,128 @@
+"""Unit tests for ConversationModeMiddleware (#616)."""
+
+from __future__ import annotations
+
+from contextlib import contextmanager
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+from langchain.agents.middleware import ModelRequest
+from langchain_core.messages import SystemMessage, ToolMessage
+from langchain_core.tools import StructuredTool
+from langgraph.config import var_child_runnable_config
+from langgraph.prebuilt.tool_node import ToolCallRequest
+
+from octop.infra.agents.middleware.conversation_mode import (
+    ConversationModeMiddleware,
+    apply_conversation_mode_to_request,
+)
+
+
+@contextmanager
+def _configurable(**kwargs: object):
+    token = var_child_runnable_config.set({"configurable": kwargs})
+    try:
+        yield
+    finally:
+        var_child_runnable_config.reset(token)
+
+
+def _tool(name: str) -> StructuredTool:
+    return StructuredTool.from_function(
+        func=lambda: "x",
+        name=name,
+        description=name,
+    )
+
+
+def test_ask_filters_mutating_tools_and_keeps_read_tools() -> None:
+    request = ModelRequest(
+        model=MagicMock(),
+        messages=[],
+        system_message=SystemMessage(content="base"),
+        tools=[_tool("read_file"), _tool("write_file"), _tool("execute"), _tool("write_todos")],
+    )
+    with _configurable(
+        conversation_mode="ask",
+        conversation_mode_hint="ASK_HINT",
+    ):
+        out = apply_conversation_mode_to_request(request)
+    names = [getattr(t, "name", None) for t in (out.tools or [])]
+    assert names == ["read_file"]
+    assert out.system_message is not None
+    assert "base" in str(out.system_message.content)
+    assert "ASK_HINT" in str(out.system_message.content)
+
+
+def test_plan_allows_write_todos_and_blocks_execute() -> None:
+    request = ModelRequest(
+        model=MagicMock(),
+        messages=[],
+        tools=[_tool("write_todos"), _tool("execute"), _tool("edit_file")],
+    )
+    with _configurable(conversation_mode="plan", conversation_mode_hint="PLAN_HINT"):
+        out = apply_conversation_mode_to_request(request)
+    names = [getattr(t, "name", None) for t in (out.tools or [])]
+    assert names == ["write_todos"]
+    assert "PLAN_HINT" in str(out.system_message.content)
+
+
+def test_craft_is_passthrough_for_tools() -> None:
+    tools = [_tool("write_file"), _tool("execute")]
+    request = ModelRequest(
+        model=MagicMock(),
+        messages=[],
+        system_message=SystemMessage(content="base"),
+        tools=tools,
+    )
+    with _configurable(conversation_mode="craft", conversation_mode_hint="CRAFT_HINT"):
+        out = apply_conversation_mode_to_request(request)
+    assert [getattr(t, "name", None) for t in (out.tools or [])] == ["write_file", "execute"]
+    assert "CRAFT_HINT" in str(out.system_message.content)
+
+
+def test_hint_not_duplicated_on_repeat() -> None:
+    request = ModelRequest(
+        model=MagicMock(),
+        messages=[],
+        system_message=SystemMessage(content="base\n\nHINT"),
+        tools=[],
+    )
+    with _configurable(conversation_mode="ask", conversation_mode_hint="HINT"):
+        out = apply_conversation_mode_to_request(request)
+    assert str(out.system_message.content).count("HINT") == 1
+
+
+@pytest.mark.asyncio
+async def test_awrap_tool_call_blocks_ask_denylist() -> None:
+    mw = ConversationModeMiddleware()
+    request = ToolCallRequest(
+        tool_call={"name": "execute", "args": {}, "id": "c1"},
+        tool=None,  # type: ignore[arg-type]
+        state={},  # type: ignore[arg-type]
+        runtime=None,  # type: ignore[arg-type]
+    )
+    handler = AsyncMock(return_value=ToolMessage(content="ran", tool_call_id="c1"))
+    with _configurable(conversation_mode="ask", locale="en"):
+        result = await mw.awrap_tool_call(request, handler)
+    assert isinstance(result, ToolMessage)
+    assert result.status == "error"
+    assert "execute" in result.content
+    handler.assert_not_called()
+
+
+@pytest.mark.asyncio
+async def test_awrap_tool_call_allows_craft() -> None:
+    mw = ConversationModeMiddleware()
+    request = ToolCallRequest(
+        tool_call={"name": "execute", "args": {}, "id": "c1"},
+        tool=None,  # type: ignore[arg-type]
+        state={},  # type: ignore[arg-type]
+        runtime=None,  # type: ignore[arg-type]
+    )
+    ok = ToolMessage(content="ran", tool_call_id="c1")
+    handler = AsyncMock(return_value=ok)
+    with _configurable(conversation_mode="craft"):
+        result = await mw.awrap_tool_call(request, handler)
+    assert result is ok
+    handler.assert_awaited_once()
diff --git a/tests/unit/agents/test_plan_artifact.py b/tests/unit/agents/test_plan_artifact.py
new file mode 100644
index 000000000..c650f2ee7
--- /dev/null
+++ b/tests/unit/agents/test_plan_artifact.py
@@ -0,0 +1,40 @@
+"""Unit tests for Plan → Craft brief artifact (#616 P5)."""
+
+from __future__ import annotations
+
+from octop.infra.agents.plan_artifact import format_plan_brief, parse_plan_brief_steps
+
+
+def test_format_plan_brief_round_trip_steps() -> None:
+    brief = format_plan_brief(
+        summary="Ship conversation modes.",
+        todos=[
+            {"id": "1", "content": "Ask denylist", "status": "completed"},
+            {"id": "2", "content": "Plan CTA", "status": "pending"},
+        ],
+    )
+    assert "Approved plan" in brief
+    assert "Ship conversation modes." in brief
+    assert "Execute this plan now." in brief
+    steps = parse_plan_brief_steps(brief)
+    assert steps[0].startswith("Ask denylist")
+    assert "Plan CTA" in steps[1]
+
+
+def test_format_plan_brief_summary_only() -> None:
+    brief = format_plan_brief(summary="Just do the thing.")
+    assert "Just do the thing." in brief
+    assert parse_plan_brief_steps(brief) == []
+
+
+def test_format_plan_brief_skips_empty_todos_without_gaps() -> None:
+    brief = format_plan_brief(
+        todos=[
+            {"content": "First"},
+            {"content": ""},
+            {"content": "Third"},
+        ]
+    )
+    assert "1. First" in brief
+    assert "2. Third" in brief
+    assert "3." not in brief
diff --git a/tests/unit/api/test_agent_conversation_mode.py b/tests/unit/api/test_agent_conversation_mode.py
new file mode 100644
index 000000000..266af857a
--- /dev/null
+++ b/tests/unit/api/test_agent_conversation_mode.py
@@ -0,0 +1,64 @@
+"""Agent default_conversation_mode in API row payload (#616 M8)."""
+
+from __future__ import annotations
+
+from types import SimpleNamespace
+
+import pytest
+
+from octop.api.routers.agents import _row_dict, _validated_agent_config
+from octop.infra.errors import ErrorCode, OctopError
+
+
+def _agent_row(**kwargs: object) -> SimpleNamespace:
+    base = {
+        "id": 1,
+        "agent_id": "ag1",
+        "user_id": 1,
+        "name": "A",
+        "description": None,
+        "persona_mbti": None,
+        "default_model": None,
+        "system_prompt": None,
+        "last_state": "stopped",
+        "last_error": None,
+        "config_json": "{}",
+        "icon": None,
+        "template_name": None,
+        "icon_name": None,
+        "icon_url": None,
+        "color": None,
+        "skill_package_ids": None,
+        "published_expert_id": None,
+        "welcome_message": None,
+        "is_shared": 0,
+        "updated_at": None,
+    }
+    base.update(kwargs)
+    return SimpleNamespace(**base)
+
+
+def test_row_dict_exposes_default_conversation_mode_from_config() -> None:
+    import json
+
+    row = _agent_row(config_json=json.dumps({"default_conversation_mode": "ask"}))
+    payload = _row_dict(row, viewer_user_id=1)
+    assert payload["default_conversation_mode"] == "ask"
+    assert payload["config"]["default_conversation_mode"] == "ask"
+
+
+def test_row_dict_defaults_conversation_mode_to_craft() -> None:
+    payload = _row_dict(_agent_row(), viewer_user_id=1)
+    assert payload["default_conversation_mode"] == "craft"
+
+
+def test_validated_agent_config_rejects_invalid_mode() -> None:
+    with pytest.raises(OctopError) as ei:
+        _validated_agent_config({"default_conversation_mode": "agent"})
+    assert ei.value.code == ErrorCode.INTERNAL_ERROR
+    assert ei.value.status == 400
+
+
+def test_validated_agent_config_accepts_ask() -> None:
+    out = _validated_agent_config({"default_conversation_mode": "ask", "foo": 1})
+    assert out == {"default_conversation_mode": "ask", "foo": 1}
diff --git a/tests/unit/api/test_chat_conversation_mode.py b/tests/unit/api/test_chat_conversation_mode.py
new file mode 100644
index 000000000..b907d48a8
--- /dev/null
+++ b/tests/unit/api/test_chat_conversation_mode.py
@@ -0,0 +1,110 @@
+"""ChatTurnBody / inbound metadata for conversation_mode (#616 M1 API)."""
+
+from __future__ import annotations
+
+import pytest
+from pydantic import ValidationError
+
+from octop.api.routers.chat.models import ChatTurnBody
+
+
+def test_chat_turn_body_accepts_ask_plan_craft() -> None:
+    for mode in ("ask", "plan", "craft"):
+        body = ChatTurnBody(text="hi", conversation_mode=mode)  # type: ignore[arg-type]
+        assert body.conversation_mode == mode
+
+
+def test_chat_turn_body_rejects_unknown_mode() -> None:
+    with pytest.raises(ValidationError):
+        ChatTurnBody(text="hi", conversation_mode="agent")  # type: ignore[arg-type]
+
+
+def test_chat_turn_body_omitted_mode_is_none() -> None:
+    body = ChatTurnBody(text="hi")
+    assert body.conversation_mode is None
+
+
+def test_from_ws_payload_parses_conversation_mode() -> None:
+    body = ChatTurnBody.from_ws_payload({"text": "hi", "conversation_mode": "ask"})
+    assert body.conversation_mode == "ask"
+
+
+def test_from_ws_payload_drops_unknown_conversation_mode() -> None:
+    body = ChatTurnBody.from_ws_payload({"text": "hi", "conversation_mode": "agent"})
+    assert body.conversation_mode is None
+
+
+def test_build_dashboard_inbound_stamps_conversation_mode_metadata() -> None:
+    from harness_gateway.models import TextContent
+
+    from octop.api.routers.chat.turn import PreparedDashboardTurn, build_dashboard_inbound
+
+    prepared = PreparedDashboardTurn(
+        thread_id="t1",
+        session_key="sk",
+        mcp_servers=None,
+        skills=None,
+        model_ref=None,
+        inbound_content=[TextContent(text="hi")],
+        composer_context=None,
+        inbound_attachments=[],
+    )
+    inbound = build_dashboard_inbound(
+        agent_id="agent-1",
+        user_id=1,
+        prepared=prepared,
+        turn=ChatTurnBody(text="hi", conversation_mode="ask"),
+        ws_connection_id="conn",
+    )
+    assert inbound.metadata["conversation_mode"] == "ask"
+
+
+def test_build_dashboard_inbound_stamps_plan_brief_metadata() -> None:
+    from harness_gateway.models import TextContent
+
+    from octop.api.routers.chat.turn import PreparedDashboardTurn, build_dashboard_inbound
+
+    prepared = PreparedDashboardTurn(
+        thread_id="t1",
+        session_key="sk",
+        mcp_servers=None,
+        skills=None,
+        model_ref=None,
+        inbound_content=[TextContent(text="按计划执行")],
+        composer_context=None,
+        inbound_attachments=[],
+    )
+    inbound = build_dashboard_inbound(
+        agent_id="agent-1",
+        user_id=1,
+        prepared=prepared,
+        turn=ChatTurnBody(
+            text="按计划执行",
+            conversation_mode="craft",
+            plan_brief="## Approved plan\n\nDo it.",
+        ),
+        ws_connection_id="conn",
+    )
+    assert inbound.metadata["conversation_mode"] == "craft"
+    assert inbound.metadata["plan_brief"] == "## Approved plan\n\nDo it."
+
+
+def test_turn_has_content_accepts_plan_brief_only() -> None:
+    from octop.api.routers.chat.turn import turn_has_content
+
+    assert turn_has_content(ChatTurnBody(text="", plan_brief="## Approved plan\n\nDo it."))
+    assert not turn_has_content(ChatTurnBody(text=""))
+
+
+def test_content_parts_injects_trigger_for_plan_brief_only() -> None:
+    from harness_gateway.models import TextContent
+
+    from octop.api.routers.chat.turn import content_parts_from_dashboard_turn
+    from octop.infra.agents.plan_artifact import PLAN_EXECUTE_USER_TRIGGER
+
+    parts = content_parts_from_dashboard_turn(
+        ChatTurnBody(text="", plan_brief="## Approved plan\n\nDo it.")
+    )
+    assert len(parts) == 1
+    assert isinstance(parts[0], TextContent)
+    assert parts[0].text == PLAN_EXECUTE_USER_TRIGGER
diff --git a/tests/unit/gateway/test_processor_conversation_mode.py b/tests/unit/gateway/test_processor_conversation_mode.py
new file mode 100644
index 000000000..a87c5d394
--- /dev/null
+++ b/tests/unit/gateway/test_processor_conversation_mode.py
@@ -0,0 +1,247 @@
+"""GlobalProcessor turn-scoped conversation_mode configurable stamp (#616)."""
+
+from __future__ import annotations
+
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+from harness_gateway.models import ChannelSubject, InboundMessage, TextContent
+
+from octop.infra.gateway.process.processor import GlobalProcessor
+from octop.infra.gateway.slash.dispatcher import SlashDispatcher
+
+
+def _processor(agent_manager: MagicMock) -> GlobalProcessor:
+    thread_registry = MagicMock()
+    thread_registry.get_or_create_by_key = AsyncMock(return_value="thr-1")
+    thread_registry.touch_last_active = MagicMock()
+    thread_registry.set_title_if_null = MagicMock()
+    return GlobalProcessor(
+        agent_manager=agent_manager,
+        thread_registry=thread_registry,
+        audit_repo=MagicMock(),
+        agent_repo=MagicMock(get=MagicMock(return_value=MagicMock(user_id=1, default_model=None))),
+        user_repo=MagicMock(get=MagicMock(return_value=None)),
+        connector_repo=MagicMock(),
+        dispatcher=SlashDispatcher(),
+        usage_repo=None,
+        gateway=None,
+    )
+
+
+def _agent_manager(*, config: dict | None = None) -> MagicMock:
+    mgr = MagicMock()
+    mgr.merge_turn_mcp_servers = MagicMock(return_value=None)
+    mgr.prepare_chat_mcp = AsyncMock(return_value=[])
+    mgr.get_row = MagicMock(return_value=None)
+    mgr.get_config = MagicMock(return_value=config or {"tools_disabled": ["web_fetch"]})
+    mgr.get_thread_conversation_mode = MagicMock(return_value=None)
+    mgr.sync_effective_tools_disabled = MagicMock()
+    mgr.sync_tools_disabled = MagicMock()
+    mgr.providers = MagicMock()
+    mgr.providers.is_model_ref_usable = MagicMock(return_value=False)
+    mgr.providers.resolve_explicit_default_model = MagicMock(return_value=None)
+    mgr.providers.resolve_model_for_multimodal_turn = MagicMock(side_effect=lambda ref, **_k: ref)
+    mgr.providers.get_model_reasoning_capability = MagicMock(return_value=None)
+    mgr.get_thread_model = MagicMock(return_value=None)
+
+    async def fake_stream(_aid: str, _req: dict):
+        yield {"type": "token", "content": "ok"}
+
+    mgr.stream = fake_stream
+    return mgr
+
+
+@pytest.mark.asyncio
+async def test_dashboard_request_stamps_conversation_mode_ask() -> None:
+    processor = _processor(_agent_manager())
+    msg = InboundMessage(
+        channel_id="ws",
+        channel_type="dashboard",
+        tenant_id="agent-1",
+        channel_subject=ChannelSubject(subject_id="1"),
+        content=[TextContent(text="hi")],
+        metadata={"conversation_mode": "ask"},
+    )
+    request = await processor._build_dashboard_request(
+        msg,
+        agent_id="agent-1",
+        user_id=1,
+        session_key="sk",
+        thread_id="thr",
+        meta=msg.metadata or {},
+    )
+    assert (request.get("configurable") or {})["conversation_mode"] == "ask"
+    hint = (request.get("configurable") or {}).get("conversation_mode_hint")
+    assert isinstance(hint, str) and hint
+
+
+@pytest.mark.asyncio
+async def test_dashboard_request_stamps_plan_system_hint() -> None:
+    """Plan turn attaches plan-mode system hint for ConversationModeMiddleware."""
+    processor = _processor(_agent_manager())
+    msg = InboundMessage(
+        channel_id="ws",
+        channel_type="dashboard",
+        tenant_id="agent-1",
+        channel_subject=ChannelSubject(subject_id="1"),
+        content=[TextContent(text="plan it")],
+        metadata={"conversation_mode": "plan"},
+    )
+    request = await processor._build_dashboard_request(
+        msg,
+        agent_id="agent-1",
+        user_id=1,
+        session_key="sk",
+        thread_id="thr",
+        meta=msg.metadata or {},
+    )
+    cfg = request.get("configurable") or {}
+    assert cfg["conversation_mode"] == "plan"
+    hint = str(cfg.get("conversation_mode_hint") or "")
+    assert "plan" in hint.lower() or "想一想" in hint
+    assert "write_todos" in hint.lower() or "plan" in hint.lower()
+
+
+@pytest.mark.asyncio
+async def test_dashboard_request_defaults_conversation_mode_to_craft() -> None:
+    processor = _processor(_agent_manager())
+    msg = InboundMessage(
+        channel_id="ws",
+        channel_type="dashboard",
+        tenant_id="agent-1",
+        channel_subject=ChannelSubject(subject_id="1"),
+        content=[TextContent(text="hi")],
+        metadata={},
+    )
+    request = await processor._build_dashboard_request(
+        msg,
+        agent_id="agent-1",
+        user_id=1,
+        session_key="sk",
+        thread_id="thr",
+        meta={},
+    )
+    assert (request.get("configurable") or {})["conversation_mode"] == "craft"
+
+
+@pytest.mark.asyncio
+async def test_dashboard_request_falls_back_to_agent_default_conversation_mode() -> None:
+    processor = _processor(
+        _agent_manager(config={"default_conversation_mode": "ask", "tools_disabled": []})
+    )
+    request = await processor._build_dashboard_request(
+        InboundMessage(
+            channel_id="ws",
+            channel_type="dashboard",
+            tenant_id="agent-1",
+            channel_subject=ChannelSubject(subject_id="1"),
+            content=[TextContent(text="hi")],
+            metadata={},
+        ),
+        agent_id="agent-1",
+        user_id=1,
+        session_key="sk",
+        thread_id="thr",
+        meta={},
+    )
+    assert (request.get("configurable") or {})["conversation_mode"] == "ask"
+
+
+@pytest.mark.asyncio
+async def test_dashboard_request_thread_sticky_mode_beats_agent_default() -> None:
+    mgr = _agent_manager(config={"default_conversation_mode": "ask"})
+    mgr.get_thread_conversation_mode = MagicMock(return_value="plan")
+    processor = _processor(mgr)
+    request = await processor._build_dashboard_request(
+        InboundMessage(
+            channel_id="ws",
+            channel_type="dashboard",
+            tenant_id="agent-1",
+            channel_subject=ChannelSubject(subject_id="1"),
+            content=[TextContent(text="hi")],
+            metadata={},
+        ),
+        agent_id="agent-1",
+        user_id=1,
+        session_key="sk",
+        thread_id="thr",
+        meta={},
+    )
+    assert (request.get("configurable") or {})["conversation_mode"] == "plan"
+
+
+@pytest.mark.asyncio
+async def test_dashboard_request_explicit_mode_beats_sticky_and_agent_default() -> None:
+    mgr = _agent_manager(config={"default_conversation_mode": "ask"})
+    mgr.get_thread_conversation_mode = MagicMock(return_value="plan")
+    processor = _processor(mgr)
+    request = await processor._build_dashboard_request(
+        InboundMessage(
+            channel_id="ws",
+            channel_type="dashboard",
+            tenant_id="agent-1",
+            channel_subject=ChannelSubject(subject_id="1"),
+            content=[TextContent(text="hi")],
+            metadata={"conversation_mode": "craft"},
+        ),
+        agent_id="agent-1",
+        user_id=1,
+        session_key="sk",
+        thread_id="thr",
+        meta={"conversation_mode": "craft"},
+    )
+    assert (request.get("configurable") or {})["conversation_mode"] == "craft"
+
+
+@pytest.mark.asyncio
+async def test_dashboard_request_stamps_plan_brief_into_system_hint() -> None:
+    processor = _processor(_agent_manager())
+    brief = "## Approved plan\n\nWrite load.sh\n\nExecute this plan now."
+    msg = InboundMessage(
+        channel_id="ws",
+        channel_type="dashboard",
+        tenant_id="agent-1",
+        channel_subject=ChannelSubject(subject_id="1"),
+        content=[TextContent(text="按计划执行")],
+        metadata={"conversation_mode": "craft", "plan_brief": brief},
+    )
+    request = await processor._build_dashboard_request(
+        msg,
+        agent_id="agent-1",
+        user_id=1,
+        session_key="sk",
+        thread_id="thr",
+        meta=msg.metadata or {},
+    )
+    cfg = request.get("configurable") or {}
+    assert cfg["conversation_mode"] == "craft"
+    assert cfg.get("plan_brief") == brief
+    hint = str(cfg.get("conversation_mode_hint") or "")
+    assert "Write load.sh" in hint
+    assert "Approved plan" in hint
+    from octop.infra.gateway.process.message_keys import UI_HIDDEN_KEY
+
+    human = (request.get("messages") or [None])[0]
+    assert human is not None
+    assert human.additional_kwargs.get(UI_HIDDEN_KEY) is True
+
+
+@pytest.mark.asyncio
+async def test_iter_turn_chunks_does_not_mutate_agent_tools_disabled() -> None:
+    """Ask/Plan enforcement is middleware-scoped; processor must not hot-patch tools."""
+    mgr = _agent_manager()
+    processor = _processor(mgr)
+
+    msg = InboundMessage(
+        channel_id="ws",
+        channel_type="dashboard",
+        tenant_id="agent-1",
+        channel_subject=ChannelSubject(subject_id="1"),
+        content=[TextContent(text="hi")],
+        metadata={"conversation_mode": "ask", "thread_id": "thr-1"},
+    )
+    chunks = [c async for c in processor.iter_turn_chunks(msg)]
+    assert any(c.get("type") == "token" for c in chunks)
+    mgr.sync_tools_disabled.assert_not_called()
+    mgr.sync_effective_tools_disabled.assert_not_called()
diff --git a/tests/unit/i18n/test_conversation.py b/tests/unit/i18n/test_conversation.py
new file mode 100644
index 000000000..dee98f492
--- /dev/null
+++ b/tests/unit/i18n/test_conversation.py
@@ -0,0 +1,58 @@
+"""tests/unit/i18n/test_conversation.py — Ask/Plan/Craft hint key parity (#616 M7)."""
+
+from __future__ import annotations
+
+from octop.i18n import all_keys_for_locale, tr
+from octop.i18n.domains.conversation import conversation_mode_system_hint
+
+
+def test_conversation_mode_keys_parity() -> None:
+    en_keys = {k for k in all_keys_for_locale("en") if k.startswith("conversation.")}
+    zh_keys = {k for k in all_keys_for_locale("zh") if k.startswith("conversation.")}
+    assert en_keys == zh_keys
+    for mode in ("ask", "plan", "craft"):
+        key = f"conversation.mode.{mode}_system_hint"
+        assert key in en_keys
+        assert tr(key, "en")
+        assert tr(key, "zh")
+
+
+def test_conversation_mode_system_hint_helper() -> None:
+    ask_en = conversation_mode_system_hint("ask", "en")
+    ask_zh = conversation_mode_system_hint("ask", "zh")
+    assert "Ask" in ask_en or "read-only" in ask_en or "Ask mode" in ask_en
+    assert "问答" in ask_zh
+    assert ask_en != ask_zh
+    plan_en = conversation_mode_system_hint("plan", "en")
+    assert "task" in plan_en.lower() or "ACP" in plan_en or "delegate" in plan_en.lower()
+
+
+def test_conversation_mode_tool_blocked_helper() -> None:
+    from octop.i18n.domains.conversation import conversation_mode_tool_blocked
+
+    en = conversation_mode_tool_blocked("execute", "ask", "en")
+    zh = conversation_mode_tool_blocked("execute", "ask", "zh")
+    assert "execute" in en
+    assert "execute" in zh
+    assert en != zh
+
+
+def test_conversation_mode_plan_brief_block() -> None:
+    from octop.i18n.domains.conversation import conversation_mode_plan_brief_block
+
+    en = conversation_mode_plan_brief_block("## Approved plan\n\nDo it.", "en")
+    zh = conversation_mode_plan_brief_block("## Approved plan\n\nDo it.", "zh")
+    assert "Approved plan" in en
+    assert "Do it." in en
+    assert en != zh or "计划" in zh or "确认" in zh
+
+
+def test_plan_ready_cta_keys_parity() -> None:
+    """Backend keeps mode system hints; dashboard owns Ready-to-build CTA copy."""
+    for mode in ("ask", "plan", "craft"):
+        assert tr(f"conversation.mode.{mode}_system_hint", "en")
+        assert tr(f"conversation.mode.{mode}_system_hint", "zh")
+    assert tr("conversation.mode.tool_blocked", "en", tool_name="x", mode="ask")
+    assert tr("conversation.mode.tool_blocked", "zh", tool_name="x", mode="ask")
+    assert tr("conversation.mode.plan_brief_prefix", "en")
+    assert tr("conversation.mode.plan_brief_prefix", "zh")
diff --git a/tests/unit/knowledge/test_knowledge_default_open.py b/tests/unit/knowledge/test_knowledge_default_open.py
index a859327b1..cc7dbc269 100644
--- a/tests/unit/knowledge/test_knowledge_default_open.py
+++ b/tests/unit/knowledge/test_knowledge_default_open.py
@@ -35,6 +35,32 @@ def test_merge_knowledge_base_ids_default_open_only_for_owner() -> None:
     assert merge_knowledge_base_ids(visible, None, owner_user_id=2) == ["shared-default"]
 
 
+def test_merge_knowledge_base_ids_unions_agent_default_ids() -> None:
+    visible = [
+        SimpleNamespace(id="default", owner_user_id=1, default_open=True, shared=False),
+        SimpleNamespace(id="agent-kb", owner_user_id=1, default_open=False, shared=False),
+        SimpleNamespace(id="other", owner_user_id=1, default_open=False, shared=False),
+    ]
+    assert merge_knowledge_base_ids(
+        visible,
+        None,
+        owner_user_id=1,
+        agent_default_ids=["agent-kb", "missing"],
+    ) == ["default", "agent-kb"]
+
+
+def test_merge_knowledge_base_ids_falls_back_when_agent_defaults_invisible() -> None:
+    visible = [
+        SimpleNamespace(id="default", owner_user_id=1, default_open=True, shared=False),
+    ]
+    assert merge_knowledge_base_ids(
+        visible,
+        None,
+        owner_user_id=1,
+        agent_default_ids=["gone"],
+    ) == ["default"]
+
+
 def test_merge_knowledge_base_ids_unions_visible_extra_ids() -> None:
     visible = [
         SimpleNamespace(id="default", owner_user_id=1, default_open=True, shared=False),