修改

frontend/app/app/api/claw/training/pipeline/route.ts

@@ -0,0 +1,26 @@
+import { getCurrentAccount } from "@/lib/claw-auth";
+
+const CLAW_API_URL = process.env.CLAW_API_URL ?? "http://127.0.0.1:8765";
+
+export async function GET(request: Request) {
+	const account = await getCurrentAccount();
+	if (!account) return Response.json({ error: "unauthorized" }, { status: 401 });
+
+	const incoming = new URL(request.url);
+	const sessionId = incoming.searchParams.get("session_id") ?? "";
+
+	const url = new URL(`${CLAW_API_URL}/api/training/pipeline`);
+	url.searchParams.set("account_id", account.id);
+	if (sessionId) url.searchParams.set("session_id", sessionId);
+
+	const response = await fetch(url, { cache: "no-store" });
+	const payload = await response.text();
+	return new Response(payload, {
+		status: response.status,
+		headers: {
+			"content-type":
+				response.headers.get("content-type") ?? "application/json",
+			"cache-control": "no-store",
+		},
+	});
+}

frontend/app/app/api/claw/training/pipeline/stream/route.ts

@@ -0,0 +1,38 @@
+import { getCurrentAccount } from "@/lib/claw-auth";
+
+const CLAW_API_URL = process.env.CLAW_API_URL ?? "http://127.0.0.1:8765";
+
+export const dynamic = "force-dynamic";
+
+export async function GET(request: Request) {
+	const account = await getCurrentAccount();
+	if (!account) {
+		return new Response("unauthorized", { status: 401 });
+	}
+
+	const incoming = new URL(request.url);
+	const sessionId = incoming.searchParams.get("session_id") ?? "";
+
+	const url = new URL(`${CLAW_API_URL}/api/training/pipeline/stream`);
+	url.searchParams.set("account_id", account.id);
+	if (sessionId) url.searchParams.set("session_id", sessionId);
+
+	const upstream = await fetch(url, {
+		cache: "no-store",
+		signal: request.signal,
+	});
+
+	if (!upstream.body) {
+		return new Response("upstream has no body", { status: 502 });
+	}
+
+	return new Response(upstream.body, {
+		status: upstream.status,
+		headers: {
+			"content-type": "text/event-stream",
+			"cache-control": "no-store",
+			connection: "keep-alive",
+			"x-accel-buffering": "no",
+		},
+	});
+}

frontend/app/app/api/claw/training/step-detail/route.ts

@@ -0,0 +1,34 @@
+import { getCurrentAccount } from "@/lib/claw-auth";
+
+const CLAW_API_URL = process.env.CLAW_API_URL ?? "http://127.0.0.1:8765";
+
+export async function GET(request: Request) {
+	const account = await getCurrentAccount();
+	if (!account) return Response.json({ error: "unauthorized" }, { status: 401 });
+
+	const incoming = new URL(request.url);
+	const sessionId = incoming.searchParams.get("session_id") ?? "";
+	const step = incoming.searchParams.get("step") ?? "";
+	if (!sessionId || !step) {
+		return Response.json(
+			{ error: "session_id and step are required" },
+			{ status: 400 },
+		);
+	}
+
+	const url = new URL(`${CLAW_API_URL}/api/training/step-detail`);
+	url.searchParams.set("account_id", account.id);
+	url.searchParams.set("session_id", sessionId);
+	url.searchParams.set("step", step);
+
+	const response = await fetch(url, { cache: "no-store" });
+	const payload = await response.text();
+	return new Response(payload, {
+		status: response.status,
+		headers: {
+			"content-type":
+				response.headers.get("content-type") ?? "application/json",
+			"cache-control": "no-store",
+		},
+	});
+}

frontend/app/app/assistant.tsx

@@ -1,16 +1,18 @@
 "use client";
 
 import type { ExportedMessageRepository } from "@assistant-ui/core";
-import { AssistantRuntimeProvider } from "@assistant-ui/react";
+import { AssistantRuntimeProvider, useAuiState } from "@assistant-ui/react";
 import {
 	AssistantChatTransport,
 	useChatRuntime,
 } from "@assistant-ui/react-ai-sdk";
 import type { UIMessage } from "ai";
-import { useCallback, useEffect, useMemo, useRef } from "react";
+import { PanelRightOpenIcon, XIcon } from "lucide-react";
+import { useCallback, useEffect, useMemo, useRef, useState } from "react";
 import {
 	ActivityPanel,
 	ActivityProvider,
+	useActivityPanel,
 } from "@/components/assistant-ui/activity-panel";
 import { Thread } from "@/components/assistant-ui/thread";
 import {
@@ -18,18 +20,26 @@ import {
 	toReplayRepository,
 } from "@/components/assistant-ui/thread-list";
 import { ThreadListSidebar } from "@/components/assistant-ui/threadlist-sidebar";
+import { TrainingPipelinePanel } from "@/components/training/training-pipeline-panel";
+import { Button } from "@/components/ui/button";
 import {
 	SidebarInset,
 	SidebarProvider,
 	SidebarTrigger,
 } from "@/components/ui/sidebar";
 import {
+	ACTIVE_SESSION_CHANGED_EVENT,
 	readActiveSessionId,
 	readPendingWorkspaceSessionId,
 	writeActiveSessionId,
 } from "@/lib/claw-active-session";
 import { ClawSessionReplayProvider } from "@/lib/claw-session-replay";
 import { pushSessionUrl } from "@/lib/claw-session-url";
+import { cn } from "@/lib/utils";
+import {
+	useAppliedTraining,
+	useModelIterationEnabled,
+} from "@/lib/use-training-mode";
 import { ClawAccountGate, useClawAccount } from "./claw-account-gate";
 
 type AssistantProps = {
@@ -153,12 +163,7 @@ export const Assistant = ({ initialSessionId }: AssistantProps = {}) => {
 								onLogout={() => setAccount(null)}
 							/>
 							<SidebarInset>
-								<div className="flex min-h-0 flex-1 overflow-hidden">
+								<AssistantWorkspace />
-									<div className="min-w-0 flex-1 overflow-hidden">
-										<Thread />
-									</div>
-									<ActivityPanel />
-								</div>
 							</SidebarInset>
 						</div>
 					</SidebarProvider>
@@ -168,6 +173,190 @@ export const Assistant = ({ initialSessionId }: AssistantProps = {}) => {
 	);
 };
 
+// 用户在对话里表达「我要开始/进行模型训练（这一轮迭代）」时命中的关键短语。
+// 要求动词或专属触发词必须存在；避免 "什么是模型训练" 之类无意义匹配。
+const APPLY_INTENT_PATTERNS: RegExp[] = [
+	// 触发短语：开始，xxx / 开始, xxx
+	/(?:^|\s|[，,。.!?])开始\s*[，,]\s*\S+/,
+	// 单纯训练动作：进行/开始/启动/开启/执行/发起/做 + (模型) + 训练
+	/(?:进行|开始|启动|开启|执行|发起|做)\s*(?:模型)?\s*训练/i,
+	// 训练 + 目标集合
+	/训练[\s\S]{0,60}?(?:目标集合|需求集合|specific[\s_-]?test|集合)/i,
+	// 跟着 program.md 的旧表达保留
+	/(?:应用|加载|载入|启用|装载|使用|跑|按照?|根据|用)\s*\S*program\.?md/i,
+	/program\.?md[\s\S]{0,120}?(?:训练|应用|启用|跑起来|跑一下)/i,
+	/apply\s+program(?:\.md)?/i,
+	/load\s+program(?:\.md)?/i,
+	/(?:start|begin|run)\s+(?:the\s+)?training/i,
+];
+
+function detectApplyIntent(text: string): boolean {
+	return APPLY_INTENT_PATTERNS.some((re) => re.test(text));
+}
+
+function AssistantWorkspace() {
+	const [activeSessionId, setActiveSessionId] = useState<string | null>(() =>
+		readActiveSessionId(),
+	);
+	const skill = useModelIterationEnabled();
+	const { applied, setApplied } = useAppliedTraining(activeSessionId);
+	const { close: closeActivityPanel } = useActivityPanel();
+	const showPipeline = skill.enabled && applied;
+	const messages = useAuiState((s) => s.thread.messages);
+	const lastProcessedUserMsgId = useRef<string | null>(null);
+
+	useEffect(() => {
+		const update = () => setActiveSessionId(readActiveSessionId());
+		const handleChanged = (event: Event) => {
+			const detail = (event as CustomEvent<{ sessionId?: string | null }>)
+				.detail;
+			setActiveSessionId(detail?.sessionId ?? readActiveSessionId());
+		};
+		window.addEventListener(ACTIVE_SESSION_CHANGED_EVENT, handleChanged);
+		window.addEventListener("focus", update);
+		update();
+		return () => {
+			window.removeEventListener(ACTIVE_SESSION_CHANGED_EVENT, handleChanged);
+			window.removeEventListener("focus", update);
+		};
+	}, []);
+
+	useEffect(() => {
+		closeActivityPanel();
+	}, [showPipeline, closeActivityPanel]);
+
+	// Skill 关掉时撤销 applied，避免重新启用就立刻弹开。
+	useEffect(() => {
+		if (!skill.enabled && applied) {
+			setApplied(false);
+		}
+	}, [skill.enabled, applied, setApplied]);
+
+	// 自动应用：skill 启用 + 当前 session 后端的 pipeline 已经有进展（任何
+	// 卡 running 或 complete）→ applied=true。这样关 tab 重开 / 刷新 / 切回
+	// 已在跑的会话都能自动恢复 monitor 面板，不依赖触发词。
+	useEffect(() => {
+		if (!skill.enabled) return;
+		if (applied) return;
+		if (!activeSessionId) return;
+		let cancelled = false;
+		const url = `/api/claw/training/pipeline?session_id=${encodeURIComponent(activeSessionId)}`;
+		fetch(url, { cache: "no-store" })
+			.then((res) => (res.ok ? res.json() : null))
+			.then((payload) => {
+				if (cancelled || !payload) return;
+				const phases = Array.isArray(payload.phases) ? payload.phases : [];
+				const hasProgress = phases.some(
+					(ph: { cards?: Array<{ status?: string }> }) =>
+						Array.isArray(ph.cards) &&
+						ph.cards.some(
+							(c) => c.status === "running" || c.status === "complete",
+						),
+				);
+				if (hasProgress) setApplied(true);
+			})
+			.catch(() => {
+				/* ignore */
+			});
+		return () => {
+			cancelled = true;
+		};
+	}, [skill.enabled, applied, activeSessionId, setApplied]);
+
+	// 监听用户消息：skill 启用 + 检测到训练触发短语 → applied = true。
+	useEffect(() => {
+		if (!skill.enabled) return;
+		if (applied) return;
+		if (!Array.isArray(messages) || messages.length === 0) return;
+		// 找最新一条 user 消息
+		let latestUser: (typeof messages)[number] | null = null;
+		for (let i = messages.length - 1; i >= 0; i -= 1) {
+			const m = messages[i];
+			if (m.role === "user") {
+				latestUser = m;
+				break;
+			}
+		}
+		if (!latestUser) return;
+		if (lastProcessedUserMsgId.current === latestUser.id) return;
+		lastProcessedUserMsgId.current = latestUser.id;
+		const text = collectUserMessageText(latestUser);
+		if (text && detectApplyIntent(text)) {
+			setApplied(true);
+		}
+	}, [messages, skill.enabled, applied, setApplied]);
+
+	// 注意：用同一棵树 + 条件渲染，保证 <Thread /> 和 <ActivityPanel /> 在
+	// pipeline 切换前后都保持挂载，否则它们会 remount，
+	// useJupyterWorkspaceStatus 的内部 state 会被清空，UI 上 Jupyter 工作区
+	// indicator 会闪一下变成"切换工作区"。
+	return (
+		<div className="flex min-h-0 flex-1 overflow-hidden">
+			{showPipeline ? (
+				<div className="relative flex min-w-0 flex-1 flex-col">
+					<button
+						type="button"
+						onClick={() => setApplied(false)}
+						title="退出训练模式"
+						aria-label="退出训练模式"
+						className="absolute top-3 left-3 z-20 inline-flex size-7 items-center justify-center rounded-full border bg-background/90 text-muted-foreground shadow-sm backdrop-blur transition-colors hover:bg-muted"
+					>
+						<XIcon className="size-3.5" />
+					</button>
+					<TrainingPipelinePanel sessionId={activeSessionId} />
+				</div>
+			) : null}
+			<div
+				className={cn(
+					"flex flex-col overflow-hidden",
+					showPipeline
+						? "w-[28rem] shrink-0 border-l"
+						: "min-w-0 flex-1",
+				)}
+			>
+				<div className="min-h-0 flex-1 overflow-hidden">
+					<Thread />
+				</div>
+			</div>
+			<ActivityPanel asDrawer={showPipeline} />
+			{showPipeline ? <ActivityDrawerTrigger /> : null}
+		</div>
+	);
+}
+
+function collectUserMessageText(message: { content?: unknown }): string {
+	const content = message.content;
+	if (typeof content === "string") return content;
+	if (!Array.isArray(content)) return "";
+	const parts: string[] = [];
+	for (const part of content) {
+		if (typeof part === "string") {
+			parts.push(part);
+		} else if (part && typeof part === "object") {
+			const text = (part as { text?: unknown }).text;
+			if (typeof text === "string") parts.push(text);
+		}
+	}
+	return parts.join("\n");
+}
+
+function ActivityDrawerTrigger() {
+	const { open, openActivity } = useActivityPanel();
+	if (open) return null;
+	return (
+		<Button
+			type="button"
+			variant="outline"
+			size="icon"
+			onClick={openActivity}
+			className="fixed top-3 right-3 z-30 size-9 rounded-lg bg-background/90 shadow-sm backdrop-blur"
+			aria-label="打开活动面板"
+		>
+			<PanelRightOpenIcon className="size-4" />
+		</Button>
+	);
+}
+
 function getLastSessionId(messages: UIMessage[]) {
 	for (const message of [...messages].reverse()) {
 		if (message.role !== "assistant") continue;

frontend/app/app/globals.css

@@ -51,6 +51,86 @@
       background-position: -100% 0;
     }
   }
+  --animate-pipeline-running: pipeline-running-glow 1800ms ease-in-out infinite;
+  @keyframes pipeline-running-glow {
+    0%,
+    100% {
+      box-shadow:
+        0 0 0 1px rgb(52 211 153 / 0.45),
+        0 0 0 0 rgb(52 211 153 / 0);
+      border-color: rgb(52 211 153 / 0.6);
+    }
+    50% {
+      box-shadow:
+        0 0 0 1px rgb(52 211 153 / 0.95),
+        0 0 18px 2px rgb(52 211 153 / 0.4);
+      border-color: rgb(52 211 153 / 1);
+    }
+  }
+  --animate-pipeline-running-blue: pipeline-running-glow-blue 1800ms ease-in-out
+    infinite;
+  @keyframes pipeline-running-glow-blue {
+    0%,
+    100% {
+      box-shadow:
+        0 0 0 1px rgb(56 189 248 / 0.45),
+        0 0 0 0 rgb(56 189 248 / 0);
+      border-color: rgb(56 189 248 / 0.7);
+    }
+    50% {
+      box-shadow:
+        0 0 0 1px rgb(56 189 248 / 0.95),
+        0 0 18px 2px rgb(56 189 248 / 0.45);
+      border-color: rgb(56 189 248 / 1);
+    }
+  }
+  --animate-pipeline-progress-shimmer: pipeline-progress-shimmer 1400ms linear
+    infinite;
+  @keyframes pipeline-progress-shimmer {
+    from {
+      transform: translateX(-100%);
+    }
+    to {
+      transform: translateX(250%);
+    }
+  }
+  --animate-pipeline-pill-glow: pipeline-pill-glow 1800ms ease-in-out infinite;
+  @keyframes pipeline-pill-glow {
+    0%,
+    100% {
+      box-shadow:
+        0 0 0 1px rgb(52 211 153 / 0.3),
+        0 0 0 0 rgb(52 211 153 / 0);
+      border-color: rgb(52 211 153 / 0.45);
+    }
+    50% {
+      box-shadow:
+        0 0 0 1px rgb(52 211 153 / 0.7),
+        0 0 10px 2px rgb(52 211 153 / 0.35);
+      border-color: rgb(52 211 153 / 0.85);
+    }
+  }
+  --animate-pipeline-dot-ping: pipeline-dot-ping 1400ms ease-out infinite;
+  @keyframes pipeline-dot-ping {
+    0% {
+      transform: scale(0.6);
+      opacity: 0.9;
+    }
+    80%,
+    100% {
+      transform: scale(2.2);
+      opacity: 0;
+    }
+  }
+  --animate-icon-spin: icon-spin 3200ms linear infinite;
+  @keyframes icon-spin {
+    from {
+      transform: rotate(0deg);
+    }
+    to {
+      transform: rotate(360deg);
+    }
+  }
 }
 
 .doc-page-v2 {

frontend/app/components/assistant-ui/activity-panel.tsx

@@ -41,6 +41,7 @@ import {
 	CollapsibleContent,
 	CollapsibleTrigger,
 } from "@/components/ui/collapsible";
+import { Sheet, SheetContent, SheetTitle } from "@/components/ui/sheet";
 import {
 	ACTIVE_SESSION_CHANGED_EVENT,
 	readActiveSessionId,
@@ -140,7 +141,7 @@ type ActivityItem = {
 
 type LiveRunEvent = NonNullable<ClawRunStatus["events"]>[number];
 
-export function ActivityPanel() {
+export function ActivityPanel({ asDrawer = false }: { asDrawer?: boolean } = {}) {
 	const {
 		open,
 		mode,
@@ -221,20 +222,12 @@ export function ActivityPanel() {
 		prevLatestIdRef.current = latestId;
 	}, [items, mode, openActivity, selectedId]);
 
-	if (!open) return null;
+	if (!asDrawer && !open) return null;
 
-	if (mode === "files") {
+	const isFilesMode = mode === "files";
-		return (
-			<SessionFilesPanel
-				key={`${sessionId ?? "active"}:${filesRefreshToken}`}
-				sessionId={sessionId}
-				onClose={close}
-			/>
-		);
-	}
 
-	return (
+	const activityBody = (
-		<aside className="fixed inset-y-0 right-0 z-30 flex w-[min(23rem,calc(100vw-1rem))] flex-col border-l bg-background shadow-xl lg:static lg:z-auto lg:h-full lg:w-92 lg:shrink-0 lg:shadow-none">
+		<>
 			<header className="flex h-16 shrink-0 items-center justify-between border-b px-4">
 				<div className="min-w-0">
 					<h2 className="font-semibold text-base">活动</h2>
@@ -244,16 +237,18 @@ export function ActivityPanel() {
 							: "暂无链路"}
 					</p>
 				</div>
-				<Button
+				{!asDrawer ? (
-					type="button"
+					<Button
-					variant="ghost"
+						type="button"
-					size="icon"
+						variant="ghost"
-					className="size-8"
+						size="icon"
-					onClick={close}
+						className="size-8"
-					aria-label="关闭活动面板"
+						onClick={close}
-				>
+						aria-label="关闭活动面板"
-					<PanelRightCloseIcon className="size-4" />
+					>
-				</Button>
+						<PanelRightCloseIcon className="size-4" />
+					</Button>
+				) : null}
 			</header>
 			<div className="min-h-0 flex-1 overflow-y-auto px-4 py-4">
 				{visibleItems.length === 0 ? (
@@ -280,6 +275,52 @@ export function ActivityPanel() {
 					</div>
 				)}
 			</div>
+		</>
+	);
+
+	if (asDrawer) {
+		return (
+			<Sheet
+				open={open}
+				onOpenChange={(next) => {
+					if (!next) close();
+				}}
+			>
+				<SheetContent
+					side="right"
+					className="flex w-[min(25rem,calc(100vw-1rem))] flex-col gap-0 p-0 sm:max-w-md"
+				>
+					<SheetTitle className="sr-only">
+						{isFilesMode ? "聊天中的文件" : "活动"}
+					</SheetTitle>
+					{isFilesMode ? (
+						<SessionFilesPanel
+							key={`drawer:${sessionId ?? "active"}:${filesRefreshToken}`}
+							sessionId={sessionId}
+							onClose={close}
+							embedded
+						/>
+					) : (
+						activityBody
+					)}
+				</SheetContent>
+			</Sheet>
+		);
+	}
+
+	if (isFilesMode) {
+		return (
+			<SessionFilesPanel
+				key={`${sessionId ?? "active"}:${filesRefreshToken}`}
+				sessionId={sessionId}
+				onClose={close}
+			/>
+		);
+	}
+
+	return (
+		<aside className="fixed inset-y-0 right-0 z-30 flex w-[min(23rem,calc(100vw-1rem))] flex-col border-l bg-background shadow-xl lg:static lg:z-auto lg:h-full lg:w-92 lg:shrink-0 lg:shadow-none">
+			{activityBody}
 		</aside>
 	);
 }
@@ -308,9 +349,11 @@ type SessionFilesPayload = {
 function SessionFilesPanel({
 	sessionId,
 	onClose,
+	embedded = false,
 }: {
 	sessionId: string | null;
 	onClose: () => void;
+	embedded?: boolean;
 }) {
 	const [payload, setPayload] = useState<SessionFilesPayload | null>(null);
 	const [status, setStatus] = useState("正在读取聊天中的文件...");
@@ -377,8 +420,8 @@ function SessionFilesPanel({
 	const outputFiles = payload?.output ?? [];
 	const total = inputFiles.length + outputFiles.length;
 
-	return (
+	const body = (
-		<aside className="fixed inset-y-0 right-0 z-30 flex w-[min(25rem,calc(100vw-1rem))] flex-col border-l bg-background shadow-xl lg:static lg:z-auto lg:h-full lg:w-96 lg:shrink-0 lg:shadow-none">
+		<>
 			<header className="flex h-16 shrink-0 items-center justify-between border-b px-4">
 				<div className="min-w-0">
 					<h2 className="font-semibold text-base">聊天中的文件</h2>
@@ -386,16 +429,18 @@ function SessionFilesPanel({
 						{status || `${total} 个文件`}
 					</p>
 				</div>
-				<Button
+				{!embedded ? (
-					type="button"
+					<Button
-					variant="ghost"
+						type="button"
-					size="icon"
+						variant="ghost"
-					className="size-8"
+						size="icon"
-					onClick={onClose}
+						className="size-8"
-					aria-label="关闭文件面板"
+						onClick={onClose}
-				>
+						aria-label="关闭文件面板"
-					<PanelRightCloseIcon className="size-4" />
+					>
-				</Button>
+						<PanelRightCloseIcon className="size-4" />
+					</Button>
+				) : null}
 			</header>
 			<div className="min-h-0 flex-1 overflow-y-auto px-4 py-4">
 				{!status && total === 0 ? (
@@ -406,6 +451,16 @@ function SessionFilesPanel({
 				<FileSection title="输入" files={inputFiles} />
 				<FileSection title="输出" files={outputFiles} />
 			</div>
+		</>
+	);
+
+	if (embedded) {
+		return <div className="flex min-h-0 flex-1 flex-col">{body}</div>;
+	}
+
+	return (
+		<aside className="fixed inset-y-0 right-0 z-30 flex w-[min(25rem,calc(100vw-1rem))] flex-col border-l bg-background shadow-xl lg:static lg:z-auto lg:h-full lg:w-96 lg:shrink-0 lg:shadow-none">
+			{body}
 		</aside>
 	);
 }

frontend/app/components/assistant-ui/thread.tsx

@@ -425,17 +425,16 @@ async function cancelLatestRun(sessionId: string, runId?: string | null) {
 const ChatTopActions: FC = () => {
 	const { openFiles } = useActivityPanel();
 	const [workspaceOpen, setWorkspaceOpen] = useState(false);
-	const actionSessionId = useCurrentThreadSessionId({
+	// "切换工作区" dialog 与 indicator 必须用同一个 sessionId，否则会出现
-		includePending: true,
+	// "绑给 A、查的是 B" 的错位（agent 的 bash 路由也是用 active session id，
-		// 工作区切换必须绑定当前可见会话；空白新会话不能回退到上一个
+	// 错位时表面 indicator 显示已连接，实际 agent 跑在本地）。
-		// activeSessionId，否则会把 Jupyter 工作区切到旧会话上。
+	// 新建空会话场景由 thread-list.tsx 的 clearActiveSessionId() 兜底，
-		includeActive: false,
+	// 此处不再需要刻意排除 activeSessionId。
-	});
+	const sessionId = useCurrentThreadSessionId({
-	const displaySessionId = useCurrentThreadSessionId({
 		includePending: true,
 		includeActive: true,
 	});
-	const workspaceStatus = useJupyterWorkspaceStatus(displaySessionId);
+	const workspaceStatus = useJupyterWorkspaceStatus(sessionId);
 
 	return (
 		<div className="pointer-events-none sticky top-3 z-20 mr-4 flex justify-end">
@@ -462,7 +461,7 @@ const ChatTopActions: FC = () => {
 							type="button"
 							className="flex w-full items-center gap-2 rounded-lg px-3 py-2 text-left text-sm transition-colors hover:bg-muted"
 							onClick={() => {
-								openFiles(displaySessionId);
+								openFiles(sessionId);
 							}}
 						>
 							<FileTextIcon className="size-4 text-muted-foreground" />
@@ -482,7 +481,7 @@ const ChatTopActions: FC = () => {
 			<WorkspaceSwitchDialog
 				open={workspaceOpen}
 				onOpenChange={setWorkspaceOpen}
-				sessionId={actionSessionId}
+				sessionId={sessionId}
 			/>
 		</div>
 	);

frontend/app/components/assistant-ui/threadlist-sidebar.tsx

@@ -72,6 +72,7 @@ type ClawSessionSummary = {
 	preview?: string;
 	model?: string;
 	usage?: UsageSummary;
+	is_training?: boolean;
 };
 
 type UsageSummary = {
@@ -365,8 +366,15 @@ function SessionResultButton({
 			className="flex w-full flex-col rounded-lg px-2 py-2 text-left transition-colors hover:bg-muted"
 			onClick={onClick}
 		>
-			<span className="truncate font-medium text-sm">
+			<span className="flex items-center gap-1.5">
-				{loading ? "回放中..." : session.preview || session.session_id}
+				<span className="truncate font-medium text-sm">
+					{loading ? "回放中..." : session.preview || session.session_id}
+				</span>
+				{session.is_training ? (
+					<span className="shrink-0 rounded bg-amber-100 px-1.5 text-[10px] text-amber-800">
+						训练
+					</span>
+				) : null}
 			</span>
 			<span className="mt-0.5 text-muted-foreground text-xs">
 				{session.turns ?? 0} 轮 · {session.tool_calls ?? 0} 次工具
@@ -382,17 +390,23 @@ function useSidebarSessions(open: boolean) {
 	useEffect(() => {
 		if (!open) return;
 		let cancelled = false;
-		fetch("/api/claw/sessions", { cache: "no-store" })
+		const reload = () => {
-			.then((res) => (res.ok ? res.json() : []))
+			fetch("/api/claw/sessions", { cache: "no-store" })
-			.then((payload) => {
+				.then((res) => (res.ok ? res.json() : []))
-				if (cancelled || !Array.isArray(payload)) return;
+				.then((payload) => {
-				setSessions(dedupeSidebarSessions(payload));
+					if (cancelled || !Array.isArray(payload)) return;
-			})
+					setSessions(dedupeSidebarSessions(payload));
-			.catch(() => {
+				})
-				if (!cancelled) setSessions([]);
+				.catch(() => {
-			});
+					if (!cancelled) setSessions([]);
+				});
+		};
+		reload();
+		const handleUpdate = () => reload();
+		window.addEventListener("claw-session-updated", handleUpdate);
 		return () => {
 			cancelled = true;
+			window.removeEventListener("claw-session-updated", handleUpdate);
 		};
 	}, [open]);
 

frontend/app/components/training/step-detail-sheet.tsx

@@ -0,0 +1,247 @@
+"use client";
+
+import { ExternalLinkIcon, FileIcon, RefreshCwIcon } from "lucide-react";
+import { useEffect, useMemo, useState } from "react";
+import { MarkdownText } from "@/components/assistant-ui/markdown-text";
+import { Button } from "@/components/ui/button";
+import { Sheet, SheetContent, SheetTitle } from "@/components/ui/sheet";
+import { cn } from "@/lib/utils";
+
+type StepDetail = {
+	step: string;
+	run_dic: number | null;
+	format: "markdown" | "jsonl" | "json" | "csv" | "text" | "empty";
+	content: string;
+	source_path: string | null;
+	fetched_at_ms: number;
+	error?: string;
+	tried_paths?: string[];
+};
+
+type Props = {
+	sessionId: string | null;
+	stepKey: string | null;
+	stepTitle?: string;
+	stepStatus?: string;
+	open: boolean;
+	onOpenChange: (open: boolean) => void;
+};
+
+export function StepDetailSheet({
+	sessionId,
+	stepKey,
+	stepTitle,
+	stepStatus,
+	open,
+	onOpenChange,
+}: Props) {
+	const [detail, setDetail] = useState<StepDetail | null>(null);
+	const [loading, setLoading] = useState(false);
+	const [reloadCount, setReloadCount] = useState(0);
+
+	useEffect(() => {
+		if (!open || !sessionId || !stepKey) return;
+		let cancelled = false;
+		setLoading(true);
+		setDetail(null);
+		const url = `/api/claw/training/step-detail?session_id=${encodeURIComponent(sessionId)}&step=${encodeURIComponent(stepKey)}`;
+		fetch(url, { cache: "no-store" })
+			.then((res) => (res.ok ? res.json() : null))
+			.then((payload) => {
+				if (cancelled) return;
+				setDetail(payload);
+				setLoading(false);
+			})
+			.catch(() => {
+				if (!cancelled) setLoading(false);
+			});
+		return () => {
+			cancelled = true;
+		};
+	}, [open, sessionId, stepKey, reloadCount]);
+
+	return (
+		<Sheet open={open} onOpenChange={onOpenChange}>
+			<SheetContent
+				side="right"
+				className="flex w-[min(50rem,calc(100vw-2rem))] flex-col gap-0 p-0 sm:max-w-2xl"
+			>
+				<header className="flex shrink-0 items-start justify-between gap-3 border-b px-5 py-3">
+					<div className="min-w-0">
+						<div className="flex items-center gap-2">
+							<SheetTitle className="font-semibold text-base text-foreground">
+								{stepTitle ?? stepKey ?? "—"}
+							</SheetTitle>
+							{stepStatus ? <StatusPill status={stepStatus} /> : null}
+							{detail?.run_dic ? (
+								<span className="rounded bg-muted px-1.5 py-0.5 font-mono text-[10px] text-muted-foreground">
+									runDic={detail.run_dic}
+								</span>
+							) : null}
+						</div>
+						{detail?.source_path ? (
+							<div className="mt-1 flex items-center gap-1 truncate font-mono text-[11px] text-muted-foreground">
+								<FileIcon className="size-3 shrink-0" />
+								<span className="truncate">{detail.source_path}</span>
+							</div>
+						) : null}
+					</div>
+					<Button
+						type="button"
+						variant="ghost"
+						size="icon"
+						className="mt-0.5 size-7 shrink-0"
+						onClick={() => setReloadCount((c) => c + 1)}
+						title="刷新"
+						aria-label="刷新"
+					>
+						<RefreshCwIcon className="size-3.5" />
+					</Button>
+				</header>
+				<div className="min-h-0 flex-1 overflow-y-auto px-5 py-4 text-sm">
+					{loading ? (
+						<p className="text-muted-foreground text-xs">加载中...</p>
+					) : !detail ? (
+						<p className="text-muted-foreground text-xs">无法获取详情</p>
+					) : detail.format === "empty" ? (
+						<EmptyView detail={detail} />
+					) : detail.format === "markdown" ? (
+						<MarkdownText text={detail.content} />
+					) : detail.format === "json" ? (
+						<JsonView text={detail.content} />
+					) : detail.format === "jsonl" ? (
+						<JsonLinesView text={detail.content} />
+					) : (
+						<RawView text={detail.content} />
+					)}
+				</div>
+				{detail?.source_path && !loading ? (
+					<footer className="flex shrink-0 items-center justify-between border-t bg-muted/30 px-5 py-2 text-[11px] text-muted-foreground">
+						<span>
+							{detail.format} ·{" "}
+							{detail.fetched_at_ms
+								? new Date(detail.fetched_at_ms).toLocaleTimeString()
+								: ""}
+						</span>
+						<button
+							type="button"
+							className="inline-flex items-center gap-1 hover:text-foreground"
+							onClick={() => {
+								if (detail.content) {
+									navigator.clipboard?.writeText(detail.content);
+								}
+							}}
+						>
+							<ExternalLinkIcon className="size-3" />
+							复制全文
+						</button>
+					</footer>
+				) : null}
+			</SheetContent>
+		</Sheet>
+	);
+}
+
+function EmptyView({ detail }: { detail: StepDetail }) {
+	return (
+		<div className="space-y-3">
+			<p className="rounded-md border border-dashed border-muted-foreground/30 bg-muted/30 p-3 text-muted-foreground text-xs">
+				{detail.error ?? "暂无产物。"}
+			</p>
+			{detail.tried_paths && detail.tried_paths.length > 0 ? (
+				<div>
+					<div className="mb-1 text-[10px] font-medium tracking-wider text-muted-foreground uppercase">
+						尝试的路径
+					</div>
+					<ul className="space-y-1 font-mono text-[11px]">
+						{detail.tried_paths.map((p) => (
+							<li key={p} className="text-muted-foreground">
+								{p}
+							</li>
+						))}
+					</ul>
+				</div>
+			) : null}
+		</div>
+	);
+}
+
+function JsonView({ text }: { text: string }) {
+	const pretty = useMemo(() => {
+		try {
+			return JSON.stringify(JSON.parse(text), null, 2);
+		} catch {
+			return text;
+		}
+	}, [text]);
+	return (
+		<pre className="overflow-x-auto rounded-md border bg-muted/30 p-3 font-mono text-[11px] leading-relaxed">
+			{pretty}
+		</pre>
+	);
+}
+
+function JsonLinesView({ text }: { text: string }) {
+	const lines = useMemo(
+		() =>
+			text
+				.split("\n")
+				.map((l) => l.trim())
+				.filter(Boolean)
+				.map((l) => {
+					try {
+						return JSON.stringify(JSON.parse(l), null, 2);
+					} catch {
+						return l;
+					}
+				}),
+		[text],
+	);
+	return (
+		<div className="space-y-2">
+			<div className="text-[10px] text-muted-foreground">{lines.length} 行</div>
+			{lines.map((l, idx) => (
+				<details
+					key={idx}
+					open={idx === 0}
+					className="rounded-md border bg-muted/20"
+				>
+					<summary className="cursor-pointer px-3 py-1.5 font-mono text-[11px] text-muted-foreground hover:bg-muted/40">
+						行 #{idx + 1}
+					</summary>
+					<pre className="overflow-x-auto px-3 pt-0 pb-2 font-mono text-[11px] leading-relaxed">
+						{l}
+					</pre>
+				</details>
+			))}
+		</div>
+	);
+}
+
+function RawView({ text }: { text: string }) {
+	return (
+		<pre className="overflow-x-auto whitespace-pre-wrap break-all font-mono text-[11px] leading-relaxed">
+			{text}
+		</pre>
+	);
+}
+
+function StatusPill({ status }: { status: string }) {
+	const map: Record<string, string> = {
+		complete: "bg-emerald-500/15 text-emerald-700 dark:text-emerald-300",
+		running: "bg-sky-500/20 text-sky-700 dark:text-sky-200",
+		pending: "bg-muted text-muted-foreground",
+		failed: "bg-rose-500/15 text-rose-700 dark:text-rose-300",
+		cancelled: "bg-muted text-muted-foreground",
+	};
+	return (
+		<span
+			className={cn(
+				"rounded-full px-2 py-0.5 text-[10px] font-medium tracking-wider",
+				map[status] ?? "bg-muted text-muted-foreground",
+			)}
+		>
+			{status.toUpperCase()}
+		</span>
+	);
+}

frontend/app/components/training/training-mode-toggle.tsx

@@ -0,0 +1,60 @@
+"use client";
+
+import { GitBranchIcon } from "lucide-react";
+import { cn } from "@/lib/utils";
+
+type Props = {
+	sessionId: string | null;
+	isTraining: boolean;
+	onChange: (next: boolean) => void;
+	disabled?: boolean;
+	lockedReason?: string;
+};
+
+export function TrainingModeToggle({
+	sessionId,
+	isTraining,
+	onChange,
+	disabled,
+	lockedReason,
+}: Props) {
+	const isDisabled = disabled || !sessionId;
+	const tooltip = !sessionId
+		? "发送首条消息后开启"
+		: lockedReason
+			? lockedReason
+			: "切换训练模式";
+
+	return (
+		<button
+			type="button"
+			disabled={isDisabled}
+			onClick={() => onChange(!isTraining)}
+			title={tooltip}
+			aria-pressed={isTraining}
+			className={cn(
+				"inline-flex items-center gap-1.5 rounded-md border px-2.5 py-1 text-xs transition-colors",
+				"disabled:cursor-not-allowed disabled:opacity-50",
+				isTraining
+					? "border-amber-300 bg-amber-50 text-amber-800 hover:bg-amber-100"
+					: "border-border bg-background text-foreground hover:bg-muted",
+			)}
+		>
+			<GitBranchIcon className="size-3.5" />
+			<span>训练模式</span>
+			<span
+				className={cn(
+					"ml-1 inline-flex h-3.5 w-7 shrink-0 items-center rounded-full transition-colors",
+					isTraining ? "bg-amber-400" : "bg-muted",
+				)}
+			>
+				<span
+					className={cn(
+						"inline-block size-3 rounded-full bg-background shadow-sm transition-transform",
+						isTraining ? "translate-x-3.5" : "translate-x-0.5",
+					)}
+				/>
+			</span>
+		</button>
+	);
+}

frontend/app/components/training/training-pipeline-panel.tsx

@@ -0,0 +1,480 @@
+"use client";
+
+import {
+	BarChart3Icon,
+	CheckIcon,
+	ChevronDownIcon,
+	ClipboardListIcon,
+	ClockIcon,
+	DnaIcon,
+	ExternalLinkIcon,
+	FileTextIcon,
+	FlaskConicalIcon,
+	GitBranchIcon,
+	GraduationCapIcon,
+	HourglassIcon,
+	Loader2Icon,
+	MicroscopeIcon,
+	type LucideIcon,
+	RefreshCwIcon,
+	ShieldCheckIcon,
+	TrendingUpIcon,
+} from "lucide-react";
+import { useEffect, useRef, useState } from "react";
+import { StepDetailSheet } from "@/components/training/step-detail-sheet";
+import { cn } from "@/lib/utils";
+
+type CardStatus = "pending" | "running" | "complete" | "failed" | "cancelled";
+type PhaseTone = "complete" | "running" | "pending";
+
+type PipelineCard = {
+	key: string;
+	icon?: string;
+	title: string;
+	subtitle: string;
+	status: CardStatus;
+	progress?: number;
+};
+
+type PipelinePhase = {
+	key: string;
+	label: string;
+	tone: PhaseTone;
+	cards: PipelineCard[];
+};
+
+type Kpi = {
+	label: string;
+	value: string;
+	icon?: string;
+};
+
+type LogLine = {
+	ts: string;
+	iter: string | number;
+	text: string;
+};
+
+type PipelinePayload = {
+	session_id: string;
+	generated_at_ms: number;
+	status: { mode: string; state: CardStatus };
+	kpis: Kpi[];
+	phases: PipelinePhase[];
+	logs: LogLine[];
+};
+
+const ICON_MAP: Record<string, LucideIcon> = {
+	"bar-chart": BarChart3Icon,
+	microscope: MicroscopeIcon,
+	"trending-up": TrendingUpIcon,
+	"file-text": FileTextIcon,
+	dna: DnaIcon,
+	"git-branch": GitBranchIcon,
+	flask: FlaskConicalIcon,
+	shield: ShieldCheckIcon,
+	"graduation-cap": GraduationCapIcon,
+	"clipboard-list": ClipboardListIcon,
+	"refresh-cw": RefreshCwIcon,
+	clock: ClockIcon,
+	hourglass: HourglassIcon,
+};
+
+export function TrainingPipelinePanel({
+	sessionId,
+}: {
+	sessionId: string | null;
+}) {
+	const [data, setData] = useState<PipelinePayload | null>(null);
+	const [loading, setLoading] = useState(false);
+	const [openCard, setOpenCard] = useState<{
+		key: string;
+		title: string;
+		status: string;
+	} | null>(null);
+
+	useEffect(() => {
+		setLoading(true);
+		const url = sessionId
+			? `/api/claw/training/pipeline/stream?session_id=${encodeURIComponent(sessionId)}`
+			: "/api/claw/training/pipeline/stream";
+		const es = new EventSource(url);
+		es.onmessage = (ev) => {
+			try {
+				const payload = JSON.parse(ev.data);
+				// program.md missing → backend pushes {available: false}; clear panel.
+				if (payload && payload.available === false) {
+					setData(null);
+				} else {
+					setData(payload);
+				}
+				setLoading(false);
+			} catch {
+				/* ignore malformed events */
+			}
+		};
+		es.onerror = () => {
+			// EventSource auto-reconnects; just stop showing the loading state.
+			setLoading(false);
+		};
+		return () => {
+			es.close();
+		};
+	}, [sessionId]);
+
+	return (
+		<div className="flex min-w-0 flex-1 flex-col overflow-hidden border-r bg-muted/20">
+			<div className="min-h-0 flex-1 overflow-y-auto px-6 py-5">
+				<HeaderRow status={data?.status ?? null} loading={loading} />
+				<KpiRow kpis={data?.kpis ?? []} />
+				<div className="mt-5 space-y-4">
+					{!data ? (
+						<div className="rounded-md border border-dashed p-8 text-center text-muted-foreground text-xs">
+							暂无 pipeline 数据
+						</div>
+					) : (
+						data.phases.map((phase) => (
+							<PhaseSection
+								key={phase.key}
+								phase={phase}
+								onCardClick={(card) =>
+									setOpenCard({
+										key: card.key,
+										title: card.title,
+										status: card.status,
+									})
+								}
+							/>
+						))
+					)}
+				</div>
+			</div>
+			<LogStream logs={data?.logs ?? []} />
+			<StepDetailSheet
+				sessionId={sessionId}
+				stepKey={openCard?.key ?? null}
+				stepTitle={openCard?.title}
+				stepStatus={openCard?.status}
+				open={openCard !== null}
+				onOpenChange={(next) => {
+					if (!next) setOpenCard(null);
+				}}
+			/>
+		</div>
+	);
+}
+
+function HeaderRow({
+	status,
+	loading,
+}: {
+	status: PipelinePayload["status"] | null;
+	loading: boolean;
+}) {
+	const statusLabel = loading
+		? "加载中"
+		: status
+			? `${status.mode} · ${statusText(status.state)}`
+			: "暂无运行";
+	const isRunning = status?.state === "running";
+	return (
+		<header className="flex items-center justify-between gap-3">
+			<div className="flex items-center gap-2">
+				<MicroscopeIcon className="size-5 text-muted-foreground" />
+				<h2 className="font-semibold text-sm uppercase tracking-[0.2em] text-foreground/90">
+					Training Pipeline Monitor
+				</h2>
+			</div>
+			<div
+				className={cn(
+					"flex items-center gap-1.5 rounded-full border bg-emerald-500/10 px-3 py-1 text-emerald-700 text-xs dark:text-emerald-300",
+					isRunning
+						? "animate-pipeline-pill-glow border-emerald-400"
+						: "border-emerald-500/40",
+				)}
+			>
+				<span className="relative inline-flex size-2 items-center justify-center">
+					<span
+						className={cn(
+							"size-2 rounded-full",
+							isRunning
+								? "bg-emerald-500"
+								: status?.state === "complete"
+									? "bg-emerald-500"
+									: status?.state === "failed"
+										? "bg-rose-500"
+										: "bg-muted-foreground",
+						)}
+					/>
+					{isRunning ? (
+						<span className="pointer-events-none absolute inset-0 animate-pipeline-dot-ping rounded-full bg-emerald-400" />
+					) : null}
+				</span>
+				<span className="font-medium">{statusLabel}</span>
+			</div>
+		</header>
+	);
+}
+
+function KpiRow({ kpis }: { kpis: Kpi[] }) {
+	if (!kpis.length) {
+		return null;
+	}
+	return (
+		<div className="mt-4 grid grid-cols-2 gap-2 sm:grid-cols-3 lg:grid-cols-7">
+			{kpis.map((kpi) => {
+				const Icon = kpi.icon ? ICON_MAP[kpi.icon] : null;
+				return (
+					<div
+						key={kpi.label}
+						className="rounded-lg border bg-background px-3 py-2 min-h-[58px]"
+					>
+						<div className="text-[10px] font-medium tracking-wider text-muted-foreground truncate">
+							{kpi.label}
+						</div>
+						<div className="mt-0.5 flex items-start gap-1.5 font-mono text-sm leading-tight text-foreground">
+							<span className="break-all line-clamp-2">{kpi.value}</span>
+							{Icon ? (
+								<Icon className="size-3.5 shrink-0 text-muted-foreground mt-0.5" />
+							) : null}
+						</div>
+					</div>
+				);
+			})}
+		</div>
+	);
+}
+
+function PhaseSection({
+	phase,
+	onCardClick,
+}: {
+	phase: PipelinePhase;
+	onCardClick?: (card: PipelineCard) => void;
+}) {
+	const dotCls =
+		phase.tone === "complete"
+			? "bg-emerald-500"
+			: phase.tone === "running"
+				? "bg-sky-500"
+				: "bg-violet-500";
+	const arrowComplete = phase.tone === "complete";
+	return (
+		<section className="rounded-xl border bg-background/60 p-4">
+			<div className="mb-3 flex items-center justify-between">
+				<div className="flex items-center gap-2">
+					<span className={cn("size-2 rounded-full", dotCls)} />
+					<span className="font-medium text-foreground text-sm">
+						{phase.label}
+					</span>
+				</div>
+				<ChevronDownIcon className="size-4 text-muted-foreground/60" />
+			</div>
+			<div className="flex flex-wrap items-stretch gap-2">
+				{phase.cards.map((card, idx) => (
+					<div key={card.key} className="flex items-stretch gap-2">
+						<PipelineCardView
+							card={card}
+							onClick={onCardClick ? () => onCardClick(card) : undefined}
+						/>
+						{idx < phase.cards.length - 1 ? (
+							<PhaseArrow complete={arrowComplete} />
+						) : null}
+					</div>
+				))}
+			</div>
+		</section>
+	);
+}
+
+function PhaseArrow({ complete }: { complete: boolean }) {
+	return (
+		<div className="flex items-center px-1">
+			<svg
+				width="32"
+				height="14"
+				viewBox="0 0 32 14"
+				fill="none"
+				className={cn(
+					complete ? "text-emerald-500" : "text-muted-foreground/40",
+				)}
+				aria-hidden
+			>
+				<line
+					x1="0"
+					y1="7"
+					x2="22"
+					y2="7"
+					stroke="currentColor"
+					strokeWidth="1.5"
+					strokeDasharray={complete ? "3 3" : "0"}
+				/>
+				<path
+					d="M22 2 L30 7 L22 12 Z"
+					fill="currentColor"
+					stroke="none"
+				/>
+			</svg>
+		</div>
+	);
+}
+
+function PipelineCardView({
+	card,
+	onClick,
+}: {
+	card: PipelineCard;
+	onClick?: () => void;
+}) {
+	const isRunning = card.status === "running";
+	const isComplete = card.status === "complete";
+	const Icon = card.icon ? ICON_MAP[card.icon] : null;
+
+	const tone = isComplete
+		? "border-emerald-500/40 bg-emerald-50/40 dark:bg-emerald-950/20"
+		: isRunning
+			? "border-sky-400 bg-sky-50/60 dark:bg-sky-950/30 ring-2 ring-sky-300/50 dark:ring-sky-500/30 animate-pipeline-running-blue"
+			: card.status === "failed"
+				? "border-rose-500/40 bg-rose-50/40 dark:bg-rose-950/20"
+				: "border-border bg-background";
+
+	const Tag = onClick ? "button" : "div";
+
+	return (
+		<Tag
+			type={onClick ? "button" : undefined}
+			onClick={onClick}
+			className={cn(
+				"relative flex w-48 flex-col rounded-xl border px-3 pt-3 pb-2.5 text-left transition-colors",
+				tone,
+				onClick &&
+					"cursor-pointer hover:brightness-105 hover:shadow-sm focus:outline-none focus-visible:ring-2 focus-visible:ring-sky-400",
+			)}
+		>
+			<div className="mb-2 flex items-start">
+				{isComplete ? (
+					<div className="flex size-5 items-center justify-center rounded-full bg-emerald-500 text-white">
+						<CheckIcon className="size-3" strokeWidth={3} />
+					</div>
+				) : (
+					<StatusBadge status={card.status} />
+				)}
+			</div>
+			<div className="flex items-center gap-1.5">
+				{Icon ? (
+					<Icon
+						className={cn(
+							"size-4 shrink-0",
+							isComplete
+								? "text-emerald-600 dark:text-emerald-400"
+								: isRunning
+									? "text-sky-600 dark:text-sky-400"
+									: "text-muted-foreground",
+						)}
+					/>
+				) : null}
+				<span
+					className={cn(
+						"font-medium text-sm leading-tight",
+						isRunning ? "text-foreground" : "",
+						card.status === "pending" ? "text-muted-foreground" : "",
+					)}
+				>
+					{card.title}
+				</span>
+			</div>
+			<span className="mt-0.5 truncate font-mono text-[11px] text-muted-foreground">
+				{card.subtitle}
+			</span>
+		</Tag>
+	);
+}
+
+function StatusBadge({ status }: { status: CardStatus }) {
+	const map: Record<CardStatus, { label: string; cls: string }> = {
+		complete: {
+			label: "DONE",
+			cls: "bg-emerald-500/15 text-emerald-700 dark:text-emerald-300",
+		},
+		running: {
+			label: "RUNNING",
+			cls: "bg-sky-500/20 text-sky-700 dark:text-sky-200",
+		},
+		pending: {
+			label: "PENDING",
+			cls: "bg-muted text-muted-foreground",
+		},
+		failed: {
+			label: "FAILED",
+			cls: "bg-rose-500/15 text-rose-700 dark:text-rose-300",
+		},
+		cancelled: {
+			label: "CANCELLED",
+			cls: "bg-muted text-muted-foreground",
+		},
+	};
+	const { label, cls } = map[status];
+	return (
+		<span
+			className={cn(
+				"inline-flex items-center gap-1 rounded px-1.5 py-0.5 text-[9px] font-semibold tracking-wider",
+				cls,
+			)}
+		>
+			{status === "running" ? (
+				<Loader2Icon className="size-2.5 animate-spin" />
+			) : null}
+			{label}
+		</span>
+	);
+}
+
+function LogStream({ logs }: { logs: LogLine[] }) {
+	const ref = useRef<HTMLDivElement>(null);
+	useEffect(() => {
+		if (ref.current) ref.current.scrollTop = ref.current.scrollHeight;
+	}, [logs]);
+	return (
+		<div className="flex h-72 shrink-0 flex-col border-t bg-zinc-950 px-5 py-3">
+			<div className="mb-2 flex shrink-0 items-center justify-between">
+				<span className="font-semibold text-xs uppercase tracking-[0.18em] text-zinc-300">
+					Live Log
+				</span>
+				<button
+					type="button"
+					className="inline-flex items-center gap-1 text-[11px] text-zinc-400 transition-colors hover:text-zinc-200"
+				>
+					查看全部
+					<ExternalLinkIcon className="size-3" />
+				</button>
+			</div>
+			<div
+				ref={ref}
+				className="min-h-0 flex-1 overflow-y-auto font-mono text-[12px] leading-6 text-zinc-100"
+			>
+				{logs.length === 0 ? (
+					<div className="text-zinc-500">暂无日志</div>
+				) : (
+					logs.map((line, idx) => (
+						<div key={idx} className="flex gap-3">
+							<span className="text-zinc-500">{line.ts}</span>
+							<span className="text-emerald-400">[iter={line.iter}]</span>
+							<span>{line.text}</span>
+						</div>
+					))
+				)}
+			</div>
+		</div>
+	);
+}
+
+function statusText(state: CardStatus) {
+	const map: Record<CardStatus, string> = {
+		running: "Running",
+		complete: "Complete",
+		failed: "Failed",
+		pending: "Idle",
+		cancelled: "Cancelled",
+	};
+	return map[state];
+}

frontend/app/lib/use-training-mode.ts

@@ -0,0 +1,228 @@
+"use client";
+
+import { useCallback, useEffect, useState } from "react";
+
+export const SESSION_UPDATED_EVENT = "claw-session-updated";
+const APPLIED_STORAGE_PREFIX = "claw.training.applied:";
+
+const MODEL_ITERATION_SKILL_NAMES = new Set(["model-iteration"]);
+
+const SKILL_POLL_MS = 8000;
+
+export function useModelIterationEnabled(): {
+	enabled: boolean;
+	loading: boolean;
+} {
+	const [enabled, setEnabled] = useState(false);
+	const [loading, setLoading] = useState(true);
+
+	useEffect(() => {
+		let cancelled = false;
+		const tick = async () => {
+			try {
+				const res = await fetch("/api/claw/skills", { cache: "no-store" });
+				if (!res.ok) {
+					if (!cancelled) {
+						setEnabled(false);
+						setLoading(false);
+					}
+					return;
+				}
+				const skills = await res.json();
+				if (!Array.isArray(skills)) {
+					if (!cancelled) {
+						setEnabled(false);
+						setLoading(false);
+					}
+					return;
+				}
+				const isOn = skills.some(
+					(s: { name?: string; enabled?: boolean }) =>
+						MODEL_ITERATION_SKILL_NAMES.has(String(s?.name ?? "")) &&
+						s?.enabled !== false,
+				);
+				if (!cancelled) {
+					setEnabled(isOn);
+					setLoading(false);
+				}
+			} catch {
+				if (!cancelled) {
+					setEnabled(false);
+					setLoading(false);
+				}
+			}
+		};
+		tick();
+		const id = window.setInterval(tick, SKILL_POLL_MS);
+		const onFocus = () => tick();
+		window.addEventListener("focus", onFocus);
+		return () => {
+			cancelled = true;
+			window.clearInterval(id);
+			window.removeEventListener("focus", onFocus);
+		};
+	}, []);
+
+	return { enabled, loading };
+}
+
+// applied 状态用 localStorage（按 session id）持久化：
+// - 关 tab / 浏览器重启都不丢
+// - 刷新页面也保持
+// 不走后端 PATCH 是为了避开 __LOCALID_* 临时 session id 导致 404
+export function useAppliedTraining(sessionId: string | null) {
+	const [applied, setApplied] = useState(false);
+
+	useEffect(() => {
+		if (!sessionId || typeof window === "undefined") {
+			setApplied(false);
+			return;
+		}
+		const stored =
+			window.localStorage.getItem(APPLIED_STORAGE_PREFIX + sessionId) === "1";
+		setApplied(stored);
+	}, [sessionId]);
+
+	const setAppliedPersist = useCallback(
+		(next: boolean) => {
+			setApplied(next);
+			if (typeof window === "undefined") return;
+			if (!sessionId) return;
+			const key = APPLIED_STORAGE_PREFIX + sessionId;
+			if (next) {
+				window.localStorage.setItem(key, "1");
+			} else {
+				window.localStorage.removeItem(key);
+			}
+		},
+		[sessionId],
+	);
+
+	return { applied, setApplied: setAppliedPersist };
+}
+
+export type ProgramStatus = {
+	available: boolean;
+	path?: string;
+	last_modified_ms?: number | null;
+	error?: string;
+};
+
+const PROGRAM_STATUS_POLL_MS = 5000;
+
+export function useProgramStatus(): ProgramStatus & { loading: boolean } {
+	const [status, setStatus] = useState<ProgramStatus>({ available: false });
+	const [loading, setLoading] = useState(true);
+
+	useEffect(() => {
+		let cancelled = false;
+		const tick = async () => {
+			try {
+				const res = await fetch("/api/claw/training/program-status", {
+					cache: "no-store",
+				});
+				if (!res.ok) {
+					if (!cancelled) {
+						setStatus({ available: false });
+						setLoading(false);
+					}
+					return;
+				}
+				const payload = (await res.json()) as ProgramStatus;
+				if (!cancelled) {
+					setStatus(payload);
+					setLoading(false);
+				}
+			} catch {
+				if (!cancelled) {
+					setStatus({ available: false });
+					setLoading(false);
+				}
+			}
+		};
+		tick();
+		const id = window.setInterval(tick, PROGRAM_STATUS_POLL_MS);
+		const onFocus = () => tick();
+		window.addEventListener("focus", onFocus);
+		return () => {
+			cancelled = true;
+			window.clearInterval(id);
+			window.removeEventListener("focus", onFocus);
+		};
+	}, []);
+
+	return { ...status, loading };
+}
+
+type SessionUpdatedDetail = {
+	sessionId: string;
+	is_training?: boolean;
+};
+
+export function dispatchSessionUpdated(detail: SessionUpdatedDetail) {
+	if (typeof window === "undefined") return;
+	window.dispatchEvent(
+		new CustomEvent<SessionUpdatedDetail>(SESSION_UPDATED_EVENT, { detail }),
+	);
+}
+
+export function useTrainingMode(sessionId: string | null) {
+	const [isTraining, setIsTraining] = useState(false);
+	const [loaded, setLoaded] = useState(false);
+
+	useEffect(() => {
+		setLoaded(false);
+		if (!sessionId) {
+			setIsTraining(false);
+			return;
+		}
+		let cancelled = false;
+		fetch(`/api/claw/sessions/${encodeURIComponent(sessionId)}`, {
+			cache: "no-store",
+		})
+			.then(async (res) => {
+				if (!res.ok) return null;
+				return res.json();
+			})
+			.then((payload) => {
+				if (cancelled) return;
+				const next = Boolean(payload?.is_training);
+				setIsTraining(next);
+				setLoaded(true);
+			})
+			.catch(() => {
+				if (!cancelled) setLoaded(true);
+			});
+		return () => {
+			cancelled = true;
+		};
+	}, [sessionId]);
+
+	const setTraining = useCallback(
+		async (next: boolean) => {
+			if (!sessionId) return;
+			const previous = isTraining;
+			setIsTraining(next);
+			try {
+				const res = await fetch(
+					`/api/claw/sessions/${encodeURIComponent(sessionId)}`,
+					{
+						method: "PATCH",
+						headers: { "content-type": "application/json" },
+						body: JSON.stringify({ is_training: next }),
+					},
+				);
+				if (!res.ok) {
+					setIsTraining(previous);
+					return;
+				}
+				dispatchSessionUpdated({ sessionId, is_training: next });
+			} catch {
+				setIsTraining(previous);
+			}
+		},
+		[sessionId, isTraining],
+	);
+
+	return { isTraining, setTraining, loaded };
+}

skills/model-iteration/README.md

@@ -1,66 +0,0 @@
-# Skill: model-iteration
-
-本目录是小爱中控模型迭代 skill，用于评测、错误分析、数据增强、SFT 训练和 CML workflow 追踪。
-
-## 目录结构
-
-```
-model-iteration/
-├── config.yaml                 # 评估阈值 + CML 工作流配置
-├── knowledge/                  # 分流知识库
-│   ├── navigation_routing_rules.md   # 地图导航 Agent/ComplexTask 分流
-│   ├── travel_routing_rules.md       # 旅游 Agent/ComplexTask 分流
-│   └── multi_command_rules.md        # 多指令 vs ComplexTask 判定
-├── scripts/
-│   ├── prepare_and_train_sft.py      # 数据组装 + 本地训练
-│   ├── submit_sft_via_cml.sh         # CML 云端训练提交（R29 起统一用此）
-│   └── sft_train_job.yaml.tpl        # CML 训练 job YAML 模板
-├── SKILL.md                    # 完整迭代协议（触发规则、Step 0-6、数据增强/清洗规范）
-└── README.md                   # 本文件
-```
-
-## 快速上手
-
-### 1. 评测当前模型
-
-```bash
-cml workflow run \
-  --workflow_id f-20260408161444-wu3pz --version v28 \
-  --global_inputs runDic=<N> \
-  --global_inputs model_path_new=<新模型路径> \
-  --global_inputs model_path_old=<基线模型路径>
-```
-
-### 2. SFT 训练（CML 云端）
-
-```bash
-AUTORESEARCH_ROOT=/mnt/wangsenhao/autoresearch-zk \
-  ./scripts/submit_sft_via_cml.sh <RUNDIC> [PREV_RUNDIC]
-```
-
-训练完成后自动起评测 workflow。
-
-### 3. SFT 训练（本地）
-
-```bash
-AUTORESEARCH_ROOT=/mnt/wangsenhao/autoresearch-zk \
-  python scripts/prepare_and_train_sft.py prepare --output_dir ./sft_data
-
-AUTORESEARCH_ROOT=/mnt/wangsenhao/autoresearch-zk \
-  python scripts/prepare_and_train_sft.py train \
-  --data_dir ./sft_data \
-  --model_path /mnt/wangsenhao/verl_zk/Qwen3-4B-Instruct-2507 \
-  --model_type qwen3 \
-  --train_output ./sft_output \
-  --epochs 3 --lr 1e-5
-```
-
-## 关键规则
-
-- **达标标准**：需求集合 ≥ 95%，大盘车载降幅 ≤ 0.3%，specific test 降幅 ≤ 1%
-- **数据清洗**：case 驱动、禁止批量、必须备份、只改 output 不动 instruction
-- **训练环境**：FSDP2 + torch 2.6 需要 accelerate==1.7.0，不要改训练代码去适配
-
-## 完整协议
-
-详见 [SKILL.md](SKILL.md)

skills/model-iteration/assets/config.yaml

@@ -9,5 +9,5 @@ thresholds:
 # CML 评测工作流配置
 cml_eval:
   workflow_id: f-20260408161444-wu3pz
-  version: v28
+  version: v32
 

skills/model-iteration/references/knowledge/multiturn_continuity_rules.md

@@ -0,0 +1,57 @@
+# 多轮对话连贯性 —— ComplexTask 继承规则
+
+## 核心判定
+
+| 场景 | 输出 | 示例 |
+|------|------|------|
+| 上轮是 ComplexTask + 当前 query 和上轮**相关**（延续/细化/替换/挑选/追加） | 当前仍 `ComplexTask(tag="...")` | 上轮 CT 导航多目的地，当前"第二家导航过去" → CT |
+| 上轮是 Agent + 当前 query 是独立新任务 | 按当前 query 独立分类 | 上轮 Agent 导航，当前"播放周杰伦" → Agent(音乐播放) |
+| 上轮是 ComplexTask + 当前 query 是**无关**新任务 | 按当前独立分类 | 上轮 CT 旅游，当前"今天天气怎么样" → QA/WeatherQA |
+
+## 判定要点
+
+- **"相关"的定义**：当前 query 是对上轮任务的**延续、细化、替换、挑选、追加**中的任一类：
+  - 延续：用户在上轮的同一任务线上继续操作（"继续导航"/"换一条不堵的"）
+  - 细化：对上轮结果加筛选条件（"人少一点的"/"评分最高的"/"最近的那个"）
+  - 替换：换掉上轮的某个要素，同领域保持（"不要 A 换成 B"）
+  - 挑选：从上轮推荐的候选中选择（"第一个"/"第二家"/"刚给的那个"）
+  - 追加：在上轮任务基础上加步骤（"路上再加个加油站"/"顺便找个 ATM"）
+
+- **prev 侧"隐式 CT 信号"（R17777 补充）**：prev 即使只是一轮用户独白，只要含以下任一信号也视作 CT 场景：
+  - **单轮内自我修正/替换**："导航到 A 哦不对 B"、"去 X 啊不 Y"、"到 A 嗯就是 B" —— 用户在说的过程中换目的地，系统需处理替换逻辑
+  - **候选选择犹豫**："是 A 还是 B"、"哪个近就哪个"、"算了不去 A 了"
+  - **多 POI 枚举/对比**："A 和 B 哪个近"、"先说 A 再说 B"
+
+> 注：单 POI + 多重形容词筛选（"最便宜的有车位的充电站"）**不算** CT 信号，按 R1 归 Agent。
+
+- **继承的是 `complex` 维度，不是 tag 内容**：tag 仍可能是 Agent 标签，但 complex=true，因为复合任务语境未结束
+
+## 反例（不适用继承）
+
+- 当前 query 明显切换意图：上轮 CT 导航 → 当前"打开空调"（tag=车载控制，无关）→ 独立分类
+- 当前 query 是独立的闲聊 / 时间 / 天气问答：上轮 CT → 当前"现在几点" → 独立分类
+- 上轮是 CT 但失败/取消（"算了不要了"）：后续 query 不再继承
+
+## 为什么需要这条规则（经验来源）
+
+R17775 迭代中，我们修改了训练集 47 条"纯路线偏好"样本 complex=true → false，对齐 0511 专项 gold（60% → 76.67%）。但这产生副作用：
+
+- `复杂导航线上真实_rag` -3.30pp（11 条 gold=true 被过度修正为 false）
+  - 如 "我要重新选一个路线"、"嗯经过凤雏路去江北虹悦城"
+- `0511 专项` 新引入 6 条错全是**多轮延续 POI/挑选**类
+  - 如 "评分最高的那个导航过去"、"第二家导航过去"、"人少一点的"
+
+根因：单看 query 这些表达像"纯偏好"，但在多轮上下文中是**对上轮 ComplexTask 的挑选/细化**，应继承 CT。SFT 模型在只看当前 query 时判错。
+
+## 应用场景
+
+- **数据增强**：生成带 prev_session 的训练样本时，如果 prev 是 `ComplexTask(...)`，当前 query 与之相关 → ground_truth 必须是 ComplexTask（不能矛盾）
+- **训练集清洗**：不要把"多轮延续同任务"类样本的 complex=true 改为 false（R17774 的 47 条改动里，L10752 / L10943 / L10954 等 multi_turn_deixis 已保留未改，正确）
+- **Badcase 分析**：错例 query 如果语境是对上轮 CT 的延续，SFT 模型判错不能靠修改训练集同 pattern 样本的 label 来解，需要补"带 prev_session 的延续 CT"仿写增强模型学到上下文信号
+- **Reward 函数**：对违反此规则的预测（prev=CT + 相关继续 但 pred complex=false）应被惩罚
+
+## 反模式
+
+- ❌ 只看当前 query 文本判 complex，忽略 prev_session
+- ❌ 训练集批改 complex=true→false 时未检查 prev_session，导致多轮延续样本被误改
+- ❌ 生成仿写时当前 query 与 prev_session 矛盾（prev CT 但 current 标 complex=false）

skills/model-iteration/references/knowledge/travel_routing_rules.md

@@ -4,8 +4,8 @@
 
 | 场景 | 输出 | 示例 |
 |------|------|------|
-| 规划路线，无筛选条件 | `Agent(tag="旅游出行")` | 规划一个从北京到上海的路线 |
+| 规划路线，无筛选条件 | `Agent(tag="旅游")` | 规划一个从北京到上海的路线 |
-| 规划路线，有筛选条件 | `ComplexTask(tag="旅游出行")` | 规划一个从北京到上海的路线，人少风景好、三天两夜、3000预算 |
+| 规划路线，有筛选条件 | `ComplexTask(tag="旅游")` | 规划一个从北京到上海的路线，人少风景好、三天两夜、3000预算 |
 
 ## 判定要点
 

skills/model-iteration/scripts/prepare_and_train_sft.py

@@ -4,11 +4,9 @@
 
 用法:
   # 仅组装数据
-  AUTORESEARCH_ROOT=/mnt/wangsenhao/autoresearch-zk \
+  python prepare_and_train_sft.py prepare --output_dir /mnt/wangsenhao/autoresearch-zk/sft_data
-    python prepare_and_train_sft.py prepare --output_dir /mnt/wangsenhao/autoresearch-zk/sft_data
 
   # 启动训练（需先 prepare）
-  AUTORESEARCH_ROOT=/mnt/wangsenhao/autoresearch-zk \
   python prepare_and_train_sft.py train \
     --data_dir /mnt/wangsenhao/autoresearch-zk/sft_data \
     --model_path /mnt/wangsenhao/verl_zk/Qwen3-4B-Instruct-2507 \
@@ -28,14 +26,7 @@ from pathlib import Path
 logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
 log = logging.getLogger(__name__)
 
-
+AI_PLANNING = Path(__file__).resolve().parent / "ai-planning"
-def _path_from_env(name: str, fallback: Path) -> Path:
-    return Path(os.environ.get(name, str(fallback))).expanduser().resolve()
-
-
-AUTORESEARCH_ROOT = _path_from_env("AUTORESEARCH_ROOT", Path("/mnt/wangsenhao/autoresearch-zk"))
-AI_PLANNING = _path_from_env("AI_PLANNING_DIR", AUTORESEARCH_ROOT / "ai-planning")
-ZK_TRAINER_DIR = _path_from_env("ZK_TRAINER_DIR", AUTORESEARCH_ROOT / "zk_trainer")
 TRAIN_SET_DIR = AI_PLANNING / "data" / "train_set"
 DATA_TRAIN_DIR = AI_PLANNING / "data_train"
 GENERATE_TRAIN_SCRIPT = DATA_TRAIN_DIR / "generate_train.py"
@@ -157,7 +148,7 @@ def run_sft_training(data_dir: str, model_path: str, model_type: str, train_outp
                      epochs: int = 3, lr: float = 1e-5, max_seq_length: int = 1024,
                      per_device_batch: int = 1, grad_accum: int = 1):
     """基于 zk_trainer 的 accelerate launch 方式启动 SFT 训练。"""
-    zk_trainer_dir = str(ZK_TRAINER_DIR)
+    zk_trainer_dir = str(Path(__file__).resolve().parent / "zk_trainer")
     if not os.path.isdir(zk_trainer_dir):
         log.error(f"zk_trainer 目录不存在: {zk_trainer_dir}")
         return None

skills/model-iteration/scripts/sft_train_job.yaml.tpl

@@ -1,117 +0,0 @@
-jobName: "sft-train-r{RUNDIC}"
-description: "AutoResearch R{RUNDIC} SFT training (40k+ samples, qwen3-4B base, 3 epochs FSDP2)"
-accessType: PUBLIC
-
-imageConfig:
-  imageUrl: micr.cloud.mioffice.cn/vllm-image/ai-arch-llm-prod:vllm-v0.12.0-f098b188
-  imageCommand: |-
-    set -ex
-    cd {AUTORESEARCH_ROOT}
-
-    # 0a. 验环境 + 装缺的包（vllm 镜像 py3.12 + torch 已含）
-    python3 --version
-    python3 -c "import torch; print(f'torch={torch.__version__}')"
-    pip install -q --no-deps "accelerate==1.7.0" 2>&1 | tail -3
-    pip install -q --ignore-installed blinker 2>&1 | tail -2
-    pip install -q peft 2>&1 | tail -3
-    pip install -q wandb bitsandbytes 2>&1 | tail -3
-    pip install -q luigi mlflow scikit-learn openpyxl pyyaml sentencepiece tiktoken protobuf pynvml datasets 2>&1 | tail -3
-    pip install -q "transformers>=4.45" 2>&1 | tail -3
-    python3 -c 'import torch, accelerate, transformers, peft; print(torch.__version__, accelerate.__version__, transformers.__version__, peft.__version__)'
-
-    # 0c. 把 HF cache 重定向到容器本地大盘（默认在 ~/.cache 可能是 juicefs，mmap 易 SIGBUS）
-    export HF_HOME=/tmp/hf_cache
-    export HF_DATASETS_CACHE=/tmp/hf_cache/datasets
-    export TRANSFORMERS_CACHE=/tmp/hf_cache/transformers
-    mkdir -p /tmp/hf_cache/datasets /tmp/hf_cache/transformers
-    df -h /tmp /dev/shm 2>/dev/null || true
-
-    # 强制 datasets 加载进内存而不是 arrow mmap（避免 /dev/shm 溢出 SIGBUS）
-    export HF_DATASETS_IN_MEMORY_MAX_SIZE=20000000000
-    export HF_DATASETS_NUM_PROC=1
-    export TOKENIZERS_PARALLELISM=false
-    export OMP_NUM_THREADS=1
-    export MKL_NUM_THREADS=1
-    # 关闭 NCCL 用 shm 通信（改 socket 通道）
-    export NCCL_SHM_DISABLE=1
-    export NCCL_P2P_DISABLE=0
-    export NCCL_DEBUG=WARN
-
-    # 0b. 把上一轮产出搬走（cml job 内幂等，本地搬过就跳过）
-    if [ -d sft_output ] && [ ! -d sft_output_r{PREV_RUNDIC} ]; then
-      mv sft_output sft_output_r{PREV_RUNDIC}
-    fi
-    rm -rf sft_output
-
-    # 1. 组装数据
-    python3 {AUTORESEARCH_ROOT}/prepare_and_train_sft.py prepare \
-      --output_dir {AUTORESEARCH_ROOT}/sft_data
-
-    # 2. 启动训练（绝对路径，避免相对路径在容器内找不到 zk_trainer）
-    python3 {AUTORESEARCH_ROOT}/prepare_and_train_sft.py train \
-      --data_dir {AUTORESEARCH_ROOT}/sft_data \
-      --model_path /mnt/wangsenhao/verl_zk/Qwen3-4B-Instruct-2507 \
-      --model_type qwen3 \
-      --train_output {AUTORESEARCH_ROOT}/sft_output \
-      --epochs 3 --lr 1e-5
-
-    # 3. 训练成功标记（被 watcher 检测）
-    [ -f {AUTORESEARCH_ROOT}/sft_output/config.json ] && \
-      touch {AUTORESEARCH_ROOT}/sft_output/_SUCCESS
-
-# 挂载 wangsenhao + xiaoai-zk-model-train-tj5 + verl_zk 所在卷
-juiceFsMountConfigs:
-  - volume: wangsenhao
-    juiceFsCluster: tj5-common
-    subPath: /
-    mountPath: /mnt/wangsenhao
-    readOnly: false
-  - volume: xiaoai-zk-model-train-tj5
-    juiceFsCluster: tj5-common
-    subPath: /
-    mountPath: /mnt/xiaoai-zk-model-train-tj5
-    readOnly: false
-
-envConfigs:
-  - key: PYTHONPATH
-    value: {AUTORESEARCH_ROOT}/zk_trainer
-  - key: WANDB_DISABLED
-    value: "true"
-  - key: WANDB_MODE
-    value: offline
-  - key: VOLUME_PREFIX
-    value: /mnt/xiaoai-zk-model-train-tj5
-  - key: HF_HOME
-    value: /mnt/wangsenhao/.hf_cache
-
-# h20-96g 8 卡 FSDP2
-queueId: "6052"
-priority: 5
-preemptible: false
-framework: pytorch
-resourceConfigs:
-  - nodeRole: worker
-    nodeNumber: 1
-    perNodeResourceSpec:
-      resourcePriority: GUARANTEED
-      resourceName: cloudml.ng2h20-8-8.20-199
-      resourceNumber: 8
-
-# 故障自动重试（节点级失败）
-retryConfig:
-  enableRetry: true
-  maxRetryTimes: 2
-  policySets:
-    - NodeFailure
-
-# 失败/完成飞书告警
-alertConfig:
-  enableAlert: true
-  alertItems:
-    - alertConditions:
-        - FAILED
-        - SUCCEED
-      alertLevel: P2
-      alertReceivers:
-        persons:
-          - wangsenhao

skills/model-iteration/scripts/submit_sft_via_cml.sh

@@ -1,104 +0,0 @@
-#!/bin/bash
-# 用 cml custom_train submit 提交 SFT 训练，训练完成后自动起 cml workflow run 评测
-# 用法: ./submit_sft_via_cml.sh <RUNDIC>
-# 例:   ./submit_sft_via_cml.sh 17756
-
-set -euo pipefail
-
-RUNDIC=${1:?usage: $0 <RUNDIC>}
-PREV_RUNDIC=${2:-$((RUNDIC-1))}
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-ROOT=${AUTORESEARCH_ROOT:-/mnt/wangsenhao/autoresearch-zk}
-MODEL_OLD=${MODEL_OLD:-/mnt/wangsenhao/verl_zk/qwen4b_cispo_wokl_add_bvt_2/global_step_5/actor/huggingface}
-TPL=${SFT_TRAIN_JOB_TEMPLATE:-$SCRIPT_DIR/sft_train_job.yaml.tpl}
-YAML=/tmp/sft_train_job_r${RUNDIC}.yaml
-
-if [ -f ~/.cloudml-cli/.profile ]; then
-  source ~/.cloudml-cli/.profile
-else
-  echo "[cml-sft] ❌ 未找到 ~/.cloudml-cli/.profile，请先安装并初始化 cml" >&2
-  exit 1
-fi
-
-# 1. 渲染 yaml 模板
-sed \
-  -e "s|{RUNDIC}|${RUNDIC}|g" \
-  -e "s|{PREV_RUNDIC}|${PREV_RUNDIC}|g" \
-  -e "s|{AUTORESEARCH_ROOT}|${ROOT}|g" \
-  "$TPL" > "$YAML"
-echo "[cml-sft] yaml: $YAML"
-
-# 2. 提交训练任务
-SUBMIT_OUT=$(cml custom_train submit --filename "$YAML" 2>&1)
-echo "$SUBMIT_OUT"
-JOB_ID=$(echo "$SUBMIT_OUT" | grep -oE 't-[0-9]+-[a-z0-9]+' | head -1)
-if [ -z "$JOB_ID" ]; then
-  echo "[cml-sft] ❌ 提交失败" >&2
-  exit 1
-fi
-echo "[cml-sft] ✅ JobID: $JOB_ID"
-
-# 3. 后台 watcher：等训练成功 → 自动起评测
-WATCHER_LOG=/tmp/r${RUNDIC}_logs/cml_watcher.log
-mkdir -p "$(dirname "$WATCHER_LOG")"
-
-nohup bash -c '
-JOB_ID='"$JOB_ID"'
-RUNDIC='"$RUNDIC"'
-LOG='"$WATCHER_LOG"'
-ROOT='"$ROOT"'
-MODEL_NEW="$ROOT/sft_output"
-MODEL_OLD='"$MODEL_OLD"'
-EVAL_WORKFLOW_ID=f-20260408161444-wu3pz
-EVAL_VERSION=v28
-
-source ~/.cloudml-cli/.profile
-
-echo "[$(date)] 等待 cml job $JOB_ID 完成..." >> "$LOG"
-while true; do
-  STATE=$(cml custom_train describe "$JOB_ID" 2>/dev/null | grep -o "\"state\": \"[a-z]*\"" | head -1 | sed "s/.*: \"//;s/\"//")
-  case "$STATE" in
-    succeed)
-      echo "[$(date)] 训练成功" >> "$LOG"
-      break ;;
-    failed|killed)
-      echo "[$(date)] ❌ 训练 $STATE" >> "$LOG"
-      exit 1 ;;
-    *)
-      sleep 60 ;;
-  esac
-done
-
-# 验证产出
-[ ! -f "$MODEL_NEW/_SUCCESS" ] && [ ! -f "$MODEL_NEW/config.json" ] && {
-  echo "[$(date)] ❌ 产出缺失" >> "$LOG"; exit 1
-}
-
-# 起评测
-echo "[$(date)] 启动 CML 评测 runDic=$RUNDIC" >> "$LOG"
-cml workflow run \
-  --workflow_id $EVAL_WORKFLOW_ID --version $EVAL_VERSION \
-  --global_inputs runDic=$RUNDIC \
-  --global_inputs model_path_new=$MODEL_NEW \
-  --global_inputs model_path_old=$MODEL_OLD >> "$LOG" 2>&1
-echo "[$(date)] cml workflow run 提交完成" >> "$LOG"
-' > /tmp/r${RUNDIC}_logs/watcher_runner.log 2>&1 &
-
-WATCHER_PID=$!
-echo "[cml-sft] watcher PID: $WATCHER_PID（log: $WATCHER_LOG）"
-
-# 4. 立即输出可查看命令
-cat <<EOF
-
-[cml-sft] 任务提交完成，关键命令：
-  查看任务状态:  cml custom_train describe $JOB_ID
-  查看实时日志:  cml custom_train logs $JOB_ID --follow
-  停止任务:      cml custom_train kill $JOB_ID
-  watcher log:   tail -f $WATCHER_LOG
-
-训练完成后，evaluation workflow 会自动启动；评测产物在
-  /mnt/xiaoai-zk-model-train-tj5/workflow5/workflow${RUNDIC}/
-
-JobID: $JOB_ID
-RUNDIC: $RUNDIC
-EOF

src/agent_types.py

@@ -158,7 +158,7 @@ class AgentRuntimeConfig:
     cwd: Path
     max_turns: int = DEFAULT_MAX_TURNS
     command_timeout_seconds: float = 30.0
-    max_output_chars: int = 12000
+    max_output_chars: int = 50000  # ≈ 20k token (Chinese/code mix ≈ 2.5 chars/token)
     stream_model_responses: bool = False
     auto_snip_threshold_tokens: int | None = None
     auto_compact_threshold_tokens: int | None = None

src/jupyter_runtime.py

@@ -774,6 +774,20 @@ print(f"{'appended' if MODE == 'a' else 'wrote'} {target} ({len(content)} chars)
 
     def resolve_workspace_path(self, raw_path: str) -> str:
         value = raw_path.strip() or '.'
+
+        # 系统提示里给的"当前会话目录"通常是本地后端的绝对/相对路径，例如
+        # `.port_sessions/accounts/<acc>/sessions/<sid>/input` 或
+        # `/home/mi/.../.port_sessions/accounts/<acc>/sessions/<sid>/output/foo.csv`。
+        # 远端 Jupyter 没这套目录结构。识别这种 pattern 后，把 input/output/
+        # scratchpad bucket 映射到远端 workspace 的同名目录。
+        bucket_match = self._match_local_session_bucket(value)
+        if bucket_match is not None:
+            bucket, rest = bucket_match
+            tail = f'/{rest}' if rest else ''
+            return normalize_posix_path(
+                f'{self.binding.workspace_cwd}/{bucket}{tail}'
+            )
+
         if value.startswith('/'):
             return normalize_posix_path(value)
         path = PurePosixPath(value)
@@ -787,8 +801,34 @@ print(f"{'appended' if MODE == 'a' else 'wrote'} {target} ({len(content)} chars)
         if path.parts and path.parts[0] in {'scratchpad', 'scratch'}:
             tail = PurePosixPath(*path.parts[1:]) if len(path.parts) > 1 else PurePosixPath()
             return normalize_posix_path(f'{self.binding.workspace_cwd}/scratchpad/{tail}')
+        if path.parts and path.parts[0] in {'input', 'inputs'}:
+            tail = PurePosixPath(*path.parts[1:]) if len(path.parts) > 1 else PurePosixPath()
+            return normalize_posix_path(f'{self.binding.workspace_cwd}/input/{tail}')
         return normalize_posix_path(f'{self.binding.workspace_cwd}/{value}')
 
+    _LOCAL_SESSION_BUCKET_RE = re.compile(
+        r'(?:^|/)\.port_sessions/accounts/[^/]+/sessions/[^/]+/'
+        r'(?P<bucket>input|inputs|output|outputs|scratchpad|scratch)'
+        r'(?:/(?P<rest>.*))?$'
+    )
+
+    @classmethod
+    def _match_local_session_bucket(cls, raw: str) -> tuple[str, str] | None:
+        m = cls._LOCAL_SESSION_BUCKET_RE.search(raw)
+        if not m:
+            return None
+        bucket = m.group('bucket')
+        normalized = {
+            'input': 'input',
+            'inputs': 'input',
+            'output': 'output',
+            'outputs': 'output',
+            'scratchpad': 'scratchpad',
+            'scratch': 'scratchpad',
+        }[bucket]
+        rest = m.group('rest') or ''
+        return normalized, rest
+
     def _remote_env_prefix(self) -> str:
         exports = {
             'ZK_AGENT_WORKSPACE': self.binding.workspace_cwd,

src/session_store.py

@@ -15,28 +15,28 @@ from .agent_types import (
     OutputSchemaConfig,
     UsageStats,
 )
-
+
-
+
-@dataclass(frozen=True)
+@dataclass(frozen=True)
-class StoredSession:
+class StoredSession:
-    session_id: str
+    session_id: str
-    messages: tuple[str, ...]
+    messages: tuple[str, ...]
-    input_tokens: int
+    input_tokens: int
-    output_tokens: int
+    output_tokens: int
-
+
-
+
 DEFAULT_SESSION_DIR = Path('.port_sessions')
 DEFAULT_AGENT_SESSION_DIR = DEFAULT_SESSION_DIR / 'agent'
-
+
-
+
-def save_session(session: StoredSession, directory: Path | None = None) -> Path:
+def save_session(session: StoredSession, directory: Path | None = None) -> Path:
-    target_dir = directory or DEFAULT_SESSION_DIR
+    target_dir = directory or DEFAULT_SESSION_DIR
-    target_dir.mkdir(parents=True, exist_ok=True)
+    target_dir.mkdir(parents=True, exist_ok=True)
-    path = target_dir / f'{session.session_id}.json'
+    path = target_dir / f'{session.session_id}.json'
-    path.write_text(json.dumps(asdict(session), indent=2))
+    path.write_text(json.dumps(asdict(session), indent=2))
-    return path
+    return path
-
+
-
+
 def load_session(session_id: str, directory: Path | None = None) -> StoredSession:
     target_dir = directory or DEFAULT_SESSION_DIR
     data = json.loads((target_dir / f'{session_id}.json').read_text())
@@ -68,6 +68,7 @@ class StoredAgentSession:
     budget_state: JSONDict
     plugin_state: JSONDict
     scratchpad_directory: str | None = None
+    is_training: bool = False
 
 
 def save_agent_session(session: StoredAgentSession, directory: Path | None = None) -> Path:
@@ -118,6 +119,7 @@ def load_agent_session(session_id: str, directory: Path | None = None) -> Stored
             if isinstance(data.get('scratchpad_directory'), str)
             else None
         ),
+        is_training=bool(data.get('is_training', False)),
     )
 
 
@@ -222,7 +224,7 @@ def deserialize_runtime_config(payload: JSONDict) -> AgentRuntimeConfig:
         cwd=Path(str(payload['cwd'])).resolve(),
         max_turns=int(payload.get('max_turns', DEFAULT_MAX_TURNS)),
         command_timeout_seconds=float(payload.get('command_timeout_seconds', 30.0)),
-        max_output_chars=int(payload.get('max_output_chars', 12000)),
+        max_output_chars=int(payload.get('max_output_chars', 50000)),
         stream_model_responses=bool(payload.get('stream_model_responses', False)),
         auto_snip_threshold_tokens=_optional_int(payload.get('auto_snip_threshold_tokens')),
         auto_compact_threshold_tokens=_optional_int(payload.get('auto_compact_threshold_tokens')),