dsh-advisor-group 0.1.0 → 0.1.2

package/lib/index.mjs

@@ -1,12 +1,12 @@
 import Schema from "@deepseek-ai/schemastery";
 import { randomUUID } from "node:crypto";
-import { promises, readFileSync } from "node:fs";
+import { appendFileSync, mkdirSync, promises, readFileSync, readdirSync, rmSync, statSync, unlinkSync, writeFileSync } from "node:fs";
 import { homedir } from "node:os";
-import { dirname, join } from "node:path";
+import { dirname, isAbsolute, join } from "node:path";
+import { KNOWN_SESSION_EVENT_TYPES, Session } from "@deepseek-ai/dsh-session";
 import { createUserMessage } from "@deepseek-ai/dsh-llm";
 import { defineTool } from "@deepseek-ai/dsh-tools";
 import { SettingsConflictError } from "@deepseek-ai/dsh-settings";
-import { KNOWN_SESSION_EVENT_TYPES } from "@deepseek-ai/dsh-session";
 //#region src/defaults.ts
 const DEFAULT_ADVISOR_PROMPT = `
 你是顾问群中的资深专家。你的任务不是替用户执行，而是提供高质量的专业意见供主模型参考。
@@ -47,7 +47,19 @@ const AdvisorConfig = Schema.object({
 	clearApiKey: Schema.boolean().description("Transient client intent: clear the current provider direct key. Stripped before persistence."),
 	apiKeyEnv: Schema.string().description("Optional env var name for direct-http fallback."),
 	protocol: ProviderProtocol,
-	authMode: ProviderAuthMode.description("Anthropic auth header style: x-api-key or bearer.")
+	authMode: ProviderAuthMode.description("Anthropic auth header style: x-api-key or bearer."),
+	thinking: Schema.union([Schema.const("enabled"), Schema.const("disabled")]).description("Per-advisor thinking mode; unset = provider default. disabled forces no thinking block (DSH built-in channel maps it to reasoningEffort off)."),
+	reasoningEffort: Schema.union([
+		Schema.const("off"),
+		Schema.const("low"),
+		Schema.const("high"),
+		Schema.const("max")
+	]).description("Per-advisor thinking strength; unset = provider/model default. Standard DSH effort ids; off disables thinking per request."),
+	tools: Schema.union([
+		Schema.const("readonly"),
+		Schema.const("all"),
+		Schema.const("off")
+	]).description("Per-advisor tool calling scope override; unset = follow the global discussion.advisorTools default.")
 });
 const DriverModel = Schema.object({
 	provider: Schema.string().required().description("Provider route of the fallback driver model."),
@@ -60,12 +72,34 @@ const DiscussionConfig = Schema.object({
 	autoDeepen: Schema.boolean().default(true).description("Run the auto-deepen pipeline: after each round the driver model asks a deeper follow-up, and a final conclusion closes the consultation."),
 	driverModel: Schema.union([DriverModel, Schema.const(void 0)]).description("Fallback driver model (provider/model) used when the session header cannot be read."),
 	stopOnConsensus: Schema.boolean().default(false).description("Deprecated: dormant under the auto-deepen pipeline; kept for stored-config compatibility."),
-	advisorTimeoutMs: Schema.number().min(1e3).max(6e5).default(12e4).description("Per-advisor call timeout in milliseconds (default 120000).")
+	advisorTimeoutMs: Schema.number().min(1e3).max(6e5).default(6e5).description("Per-advisor call timeout in milliseconds (default 600000 = 10 min). Long-reasoning models can exhaust shorter budgets on the thinking chain and lose the answer body; a timeout keeps the partial thinking chain and marks the message as truncated."),
+	driverTimeoutMs: Schema.number().min(1e3).max(12e5).default(6e5).description("Driver generation timeout in milliseconds (deep-question / conclusion; default 600000 = 10 min). Long discussions with many tool results need a real budget."),
+	advisorTools: Schema.union([
+		Schema.const("readonly"),
+		Schema.const("all"),
+		Schema.const("off")
+	]).default("readonly").description("Global default advisor tool calling scope for advisors that do not set their own `tools` field: readonly (default; read-only session tools), all (every session-visible tool incl. writable ones), off (disabled). Only the direct-http OpenAI/Anthropic channel supports tool calling.")
 });
 const TriggerConfig = Schema.object({
 	requireClassifier: Schema.boolean().default(true).description("Run the pre-classifier before starting a consultation."),
 	allowWebFallback: Schema.boolean().default(true).description("When classifier says web search is better, return that hint to the main model."),
-	confidenceThreshold: Schema.number().min(0).max(1).default(.6).description("Below this confidence score, ask_advisors should escalate even without domain keywords.")
+	confidenceThreshold: Schema.number().min(0).max(1).default(.6).description("Below this confidence score, ask_advisors should escalate even without domain keywords."),
+	jev: Schema.object({
+		enabled: Schema.boolean().default(false).description("Use Jev for semantic pre-classification when configured."),
+		provider: Schema.union([Schema.const("typesafe"), Schema.const("openrouter")]).default("typesafe"),
+		model: Schema.string().default("jev-latest"),
+		baseURL: Schema.string().description("Optional custom Jev API endpoint."),
+		apiKey: Schema.string().role("secret").description("Jev API key; prefer apiKeyEnv when possible."),
+		apiKeyEnv: Schema.string().description("Optional environment variable containing the Jev API key."),
+		timeoutMs: Schema.number().min(1e3).max(6e4).default(1e4),
+		confidenceThreshold: Schema.number().min(0).max(1).default(.6),
+		useEnglishState: Schema.boolean().default(false).description("Judge with the caller-provided English gist (questionEn) as the Jev state instead of the Chinese question; display and session records stay Chinese."),
+		highRiskThreshold: Schema.number().min(0).max(1).description("Threshold for the high_risk question; unset means follow confidenceThreshold."),
+		apiKeyMeta: Schema.object({
+			configured: Schema.boolean(),
+			last4: Schema.string()
+		}).description("Server-only API key presence metadata.")
+	}).default({})
 });
 const UiConfig = Schema.object({
 	theme: Schema.union([
@@ -76,13 +110,24 @@ const UiConfig = Schema.object({
 	showTimestamps: Schema.boolean().default(true),
 	autoExpand: Schema.boolean().default(true)
 });
+const QuotaConfig = Schema.object({
+	enabled: Schema.boolean().default(true).description("Enable the daily new-consultation cap (cost safety valve)."),
+	maxPerDay: Schema.natural().min(1).max(1e5).default(50).description("Max new consultations per UTC day (1–100000); ignored when enabled is false.")
+});
 const Config = Schema.object({
-	enabled: Schema.boolean().default(true),
-	discussion: DiscussionConfig,
-	trigger: TriggerConfig,
-	ui: UiConfig,
-	advisors: Schema.array(AdvisorConfig).default([])
+	enabled: Schema.boolean().default(true).volatile(),
+	discussion: DiscussionConfig.volatile(),
+	trigger: TriggerConfig.volatile(),
+	ui: UiConfig.volatile(),
+	quota: QuotaConfig.volatile(),
+	advisors: Schema.array(AdvisorConfig).default([]).volatile()
 });
+function unwrapVolatileConfig(value) {
+	if (value && typeof value === "object" && "get" in value && typeof value.get === "function") return unwrapVolatileConfig(value.get());
+	if (Array.isArray(value)) return value.map((item) => unwrapVolatileConfig(item));
+	if (value && typeof value === "object") return Object.fromEntries(Object.entries(value).map(([key, item]) => [key, unwrapVolatileConfig(item)]));
+	return value;
+}
 //#endregion
 //#region src/providers/advisor-prompt.ts
 /**
@@ -96,7 +141,23 @@ const ADVISOR_OUTPUT_POLICY = `
 回答要求：
 - 先给结论，再简要说明依据；
 - 在回答末尾用一行标注：置信度（0-10）或不确定性；
-- 如果没有把握，请明确写“不确定”，不要强行编造。
+- 如果没有把握，请明确写“不确定”，不要强行编造；
+- **全程使用中文输出**：包括调用工具前的行动说明、正文与最终结论（代码、路径、专业术语可保留原文）；不要把“我接下来要做什么”的说明写成英文；
+- 思考（思维链）请保持精简：只列关键推理要点，切勿让思考链占满输出预算——必须保证正文完整（历史教训：长思考+长上下文会耗尽输出上限，导致正文丢失/中途截断）；
+- **最终答复必须输出一段完整的中文正文回答**：不要只输出工具调用描述/标记，不要输出空白或空消息——即使本轮没有新结论，也要用正文说明你的判断（这是硬性输出约束）。
+`;
+/**
+* Tool-calling guidance appended to an advisor's system prompt ONLY when the
+* advisor actually has tools configured and the session-visible set resolved
+* (callAdvisor adds it on the direct-http tool loop path).
+*/
+const ADVISOR_TOOL_GUIDANCE = `
+
+工具使用指引：
+- 你可以调用当前会话提供的工具（检索/读取/搜索等）来辅助回答；
+- 对于你不了解、不熟悉的内容，尤其是项目背景、代码细节、仓库状态等不在你知识范围内的情况，请优先使用联网工具（如 web_search、web_fetch）搜索核实，再基于实际信息给出意见；
+- 仅在确有必要时调用工具：先了解现状，再下结论；不要编造你未核实的事实；
+- **调用工具前必须先用一句中文说明你准备做什么**（行动说明），例如“我先读取 package.json 确认版本”；这段行动说明会显示在卡片上，不要用英文写。
 `;
 /**
 * Sequential-relay role hint appended to one advisor's system prompt.
@@ -129,6 +190,18 @@ function withTimeout(ms, signal) {
 	if (!signal) return timeout;
 	return AbortSignal.any([signal, timeout]);
 }
+/**
+* Same combination, but keeps the timeout signal reachable so the caller can
+* classify an aborted stream: timeout fired -> graceful truncation (keep the
+* partial thinking, mark `truncated`); caller signal fired -> cancellation.
+*/
+function timeoutSignalPair(ms, signal) {
+	const timeout = AbortSignal.timeout(ms);
+	return {
+		signal: signal ? AbortSignal.any([signal, timeout]) : timeout,
+		isTimeout: () => timeout.aborted
+	};
+}
 //#endregion
 //#region src/providers/ctx-llm.ts
 function formatTranscript(entries) {
@@ -140,6 +213,13 @@ function formatTranscript(entries) {
 * This is the primary path: it reuses whatever provider routes the user has
 * already configured in DSH (deepseek-official, llm-pi-ai providers, etc.),
 * including credentials resolved by the harness.
+*
+* Timeout semantics (root cause of "thinking cut mid-sentence, no body"): a
+* long-reasoning model can spend the whole advisor budget on its thinking
+* chain. When `timeoutMs` fires, the stream simply stops (it does not throw),
+* which used to produce a silent empty body. We now detect the timeout and
+* report `truncated` so the caller can surface it to the user instead of
+* treating the reply as complete. A caller-initiated abort still propagates.
 */
 async function callViaCtxLlm(ctx, advisor, transcript, signal, onDelta, timeoutMs) {
 	const userText = formatTranscript(transcript);
@@ -148,26 +228,54 @@ async function callViaCtxLlm(ctx, advisor, transcript, signal, onDelta, timeoutM
 			type: "text",
 			text: userText
 		}],
-		source: {
-			kind: "plugin",
-			plugin: "dsh-advisor-group"
-		}
+		source: { kind: "user" }
 	})];
+	const pair = timeoutSignalPair(timeoutMs ?? 12e4, signal);
+	const reasoningEffort = advisor.thinking === "disabled" ? "off" : advisor.reasoningEffort;
 	const options = {
 		provider: advisor.provider,
 		model: advisor.model,
 		system: `${advisor.systemPrompt}${ADVISOR_OUTPUT_POLICY}`,
 		messages,
+		...reasoningEffort === void 0 ? {} : { reasoningEffort },
 		temperature: advisor.temperature,
-		maxTokens: advisor.maxTokens,
-		signal: withTimeout(timeoutMs ?? 12e4, signal)
+		maxTokens: advisor.maxTokens ?? 16384,
+		signal: pair.signal
 	};
 	let text = "";
-	for await (const chunk of ctx.llm.stream(options)) if (chunk.type === "text-delta") {
-		text += chunk.text;
-		onDelta?.({ text: chunk.text });
-	} else if (chunk.type === "reasoning-delta") onDelta?.({ thinking: chunk.text });
-	return text;
+	let thinking = "";
+	try {
+		for await (const chunk of ctx.llm.stream(options)) if (chunk.type === "text-delta") {
+			text += chunk.text;
+			onDelta?.({ text: chunk.text });
+		} else if (chunk.type === "reasoning-delta") {
+			thinking += chunk.text;
+			onDelta?.({ thinking: chunk.text });
+		}
+	} catch (error) {
+		if (pair.isTimeout()) return {
+			content: text,
+			thinking,
+			truncated: {
+				reason: "timeout",
+				atMs: Date.now()
+			}
+		};
+		throw error;
+	}
+	if (pair.isTimeout()) return {
+		content: text,
+		thinking,
+		truncated: {
+			reason: "timeout",
+			atMs: Date.now()
+		}
+	};
+	if (signal?.aborted) throw new DOMException("Aborted", "AbortError");
+	return {
+		content: text,
+		thinking
+	};
 }
 //#endregion
 //#region src/providers/presets.ts
@@ -178,7 +286,11 @@ const PROVIDER_PRESETS = {
 		protocol: "openai",
 		baseURL: "https://api.deepseek.com",
 		apiKeyEnv: "DEEPSEEK_API_KEY",
-		defaultModels: ["deepseek-v4-flash", "deepseek-v4-pro"],
+		defaultModels: [
+			"deepseek-v4-flash",
+			"deepseek-v4-pro",
+			"deepseek-v4-flash-vision-exp"
+		],
 		notes: "Also available as deepseek-official via dsh-llm-deepseek, and Anthropic-compatible at https://api.deepseek.com/anthropic."
 	},
 	"deepseek-anthropic": {
@@ -187,7 +299,11 @@ const PROVIDER_PRESETS = {
 		protocol: "anthropic",
 		baseURL: "https://api.deepseek.com/anthropic",
 		apiKeyEnv: "DEEPSEEK_API_KEY",
-		defaultModels: ["deepseek-v4-flash", "deepseek-v4-pro"],
+		defaultModels: [
+			"deepseek-v4-flash",
+			"deepseek-v4-pro",
+			"deepseek-v4-flash-vision-exp"
+		],
 		authMode: "x-api-key",
 		modelsBaseURL: "https://api.deepseek.com",
 		modelsAuthMode: "bearer",
@@ -199,7 +315,12 @@ const PROVIDER_PRESETS = {
 		protocol: "openai",
 		baseURL: "https://api.moonshot.cn/v1",
 		apiKeyEnv: "MOONSHOT_API_KEY",
-		defaultModels: ["kimi-k2.6", "kimi-k2.5"],
+		defaultModels: [
+			"kimi-k3",
+			"kimi-k2.7-code",
+			"kimi-k2.7-code-highspeed",
+			"kimi-k2.6"
+		],
 		notes: "OpenAI-compatible; Anthropic-compatible endpoint at https://api.moonshot.cn/anthropic"
 	},
 	"moonshot-anthropic": {
@@ -208,7 +329,12 @@ const PROVIDER_PRESETS = {
 		protocol: "anthropic",
 		baseURL: "https://api.moonshot.cn/anthropic",
 		apiKeyEnv: "MOONSHOT_API_KEY",
-		defaultModels: ["kimi-k2.6", "kimi-k2.5"],
+		defaultModels: [
+			"kimi-k3",
+			"kimi-k2.7-code",
+			"kimi-k2.7-code-highspeed",
+			"kimi-k2.6"
+		],
 		authMode: "bearer",
 		modelsBaseURL: "https://api.moonshot.cn/v1",
 		modelsAuthMode: "bearer",
@@ -221,14 +347,14 @@ const PROVIDER_PRESETS = {
 		baseURL: "https://api.kimi.com/coding/v1",
 		apiKeyEnv: "KIMI_CODE_API_KEY",
 		defaultModels: [
-			"k3",
-			"k3-256k",
-			"kimi-for-coding",
-			"kimi-for-coding-highspeed"
+			"kimi-k3",
+			"kimi-k2.7-code",
+			"kimi-k2.7-code-highspeed",
+			"kimi-for-coding"
 		],
 		modelsBaseURL: "https://api.kimi.com/coding/v1",
 		modelsAuthMode: "bearer",
-		notes: "Kimi Code 会员 API；模型 ID 为 k3 / k3-256k / kimi-for-coding / kimi-for-coding-highspeed。"
+		notes: "Kimi Code 会员 API；可用模型以会员页为准，常用 kimi-k3 / kimi-k2.7-code / kimi-k2.7-code-highspeed，旧 kimi-for-coding 仍兼容。"
 	},
 	"kimi-code-anthropic": {
 		id: "kimi-code-anthropic",
@@ -237,10 +363,10 @@ const PROVIDER_PRESETS = {
 		baseURL: "https://api.kimi.com/coding",
 		apiKeyEnv: "KIMI_CODE_API_KEY",
 		defaultModels: [
-			"k3",
-			"k3-256k",
-			"kimi-for-coding",
-			"kimi-for-coding-highspeed"
+			"kimi-k3",
+			"kimi-k2.7-code",
+			"kimi-k2.7-code-highspeed",
+			"kimi-for-coding"
 		],
 		authMode: "bearer",
 		modelsBaseURL: "https://api.kimi.com/coding/v1",
@@ -254,9 +380,9 @@ const PROVIDER_PRESETS = {
 		baseURL: "https://dashscope.aliyuncs.com/compatible-mode/v1",
 		apiKeyEnv: "DASHSCOPE_API_KEY",
 		defaultModels: [
-			"qwen3.7-plus",
-			"qwen3.7-max",
-			"qwen-max"
+			"qwen3.8-max",
+			"qwen3.8-flash",
+			"qwen3.7-plus"
 		],
 		notes: "DashScope OpenAI-compatible endpoint（北京公共域名）。"
 	},
@@ -267,9 +393,9 @@ const PROVIDER_PRESETS = {
 		baseURL: "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
 		apiKeyEnv: "DASHSCOPE_API_KEY",
 		defaultModels: [
-			"qwen3.7-plus",
-			"qwen3.7-max",
-			"qwen-max"
+			"qwen3.8-max",
+			"qwen3.8-flash",
+			"qwen3.7-plus"
 		],
 		notes: "DashScope 新加坡公共域名 OpenAI 兼容端点。"
 	},
@@ -280,9 +406,9 @@ const PROVIDER_PRESETS = {
 		baseURL: "https://dashscope-us.aliyuncs.com/compatible-mode/v1",
 		apiKeyEnv: "DASHSCOPE_API_KEY",
 		defaultModels: [
-			"qwen3.7-plus",
-			"qwen3.7-max",
-			"qwen-max"
+			"qwen3.8-max",
+			"qwen3.8-flash",
+			"qwen3.7-plus"
 		],
 		notes: "DashScope 美国（弗吉尼亚）公共域名 OpenAI 兼容端点。"
 	},
@@ -293,13 +419,13 @@ const PROVIDER_PRESETS = {
 		baseURL: "https://{WorkspaceId}.cn-beijing.maas.aliyuncs.com/apps/anthropic",
 		apiKeyEnv: "DASHSCOPE_API_KEY",
 		defaultModels: [
-			"qwen3.7-max",
+			"qwen3.8-max",
+			"qwen3.8-flash",
 			"qwen3.7-plus",
-			"qwen3.6-plus",
 			"deepseek-v4-pro",
 			"deepseek-v4-flash",
-			"kimi-k2.6",
-			"glm-5.1"
+			"kimi-k3",
+			"glm-5.3"
 		],
 		authMode: "bearer",
 		modelsBaseURL: "https://dashscope.aliyuncs.com/compatible-mode/v1",
@@ -313,13 +439,13 @@ const PROVIDER_PRESETS = {
 		baseURL: "https://dashscope.aliyuncs.com/apps/anthropic",
 		apiKeyEnv: "DASHSCOPE_API_KEY",
 		defaultModels: [
-			"qwen3.7-max",
+			"qwen3.8-max",
+			"qwen3.8-flash",
 			"qwen3.7-plus",
-			"qwen3.6-plus",
 			"deepseek-v4-pro",
 			"deepseek-v4-flash",
-			"kimi-k2.6",
-			"glm-5.1"
+			"kimi-k3",
+			"glm-5.3"
 		],
 		authMode: "bearer",
 		modelsBaseURL: "https://dashscope.aliyuncs.com/compatible-mode/v1",
@@ -333,13 +459,13 @@ const PROVIDER_PRESETS = {
 		baseURL: "https://dashscope-intl.aliyuncs.com/apps/anthropic",
 		apiKeyEnv: "DASHSCOPE_API_KEY",
 		defaultModels: [
-			"qwen3.7-max",
+			"qwen3.8-max",
+			"qwen3.8-flash",
 			"qwen3.7-plus",
-			"qwen3.6-plus",
 			"deepseek-v4-pro",
 			"deepseek-v4-flash",
-			"kimi-k2.6",
-			"glm-5.1"
+			"kimi-k3",
+			"glm-5.3"
 		],
 		authMode: "bearer",
 		modelsBaseURL: "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
@@ -353,13 +479,13 @@ const PROVIDER_PRESETS = {
 		baseURL: "https://dashscope-us.aliyuncs.com/apps/anthropic",
 		apiKeyEnv: "DASHSCOPE_API_KEY",
 		defaultModels: [
-			"qwen3.7-max",
+			"qwen3.8-max",
+			"qwen3.8-flash",
 			"qwen3.7-plus",
-			"qwen3.6-plus",
 			"deepseek-v4-pro",
 			"deepseek-v4-flash",
-			"kimi-k2.6",
-			"glm-5.1"
+			"kimi-k3",
+			"glm-5.3"
 		],
 		authMode: "bearer",
 		modelsBaseURL: "https://dashscope-us.aliyuncs.com/compatible-mode/v1",
@@ -373,13 +499,13 @@ const PROVIDER_PRESETS = {
 		baseURL: "https://token-plan.cn-beijing.maas.aliyuncs.com/apps/anthropic",
 		apiKeyEnv: "DASHSCOPE_TOKEN_PLAN_API_KEY",
 		defaultModels: [
-			"qwen3.7-max",
+			"qwen3.8-max",
+			"qwen3.8-flash",
 			"qwen3.7-plus",
-			"qwen3.6-plus",
 			"deepseek-v4-pro",
 			"deepseek-v4-flash",
-			"kimi-k2.6",
-			"glm-5.1"
+			"kimi-k3",
+			"glm-5.3"
 		],
 		authMode: "bearer",
 		modelsBaseURL: "https://dashscope.aliyuncs.com/compatible-mode/v1",
@@ -393,13 +519,13 @@ const PROVIDER_PRESETS = {
 		baseURL: "https://coding.dashscope.aliyuncs.com/apps/anthropic",
 		apiKeyEnv: "DASHSCOPE_CODING_PLAN_API_KEY",
 		defaultModels: [
-			"qwen3.7-max",
+			"qwen3.8-max",
+			"qwen3.8-flash",
 			"qwen3.7-plus",
-			"qwen3.6-plus",
 			"deepseek-v4-pro",
 			"deepseek-v4-flash",
-			"kimi-k2.6",
-			"glm-5.1"
+			"kimi-k3",
+			"glm-5.3"
 		],
 		authMode: "bearer",
 		modelsBaseURL: "https://dashscope.aliyuncs.com/compatible-mode/v1",
@@ -414,12 +540,9 @@ const PROVIDER_PRESETS = {
 		apiKeyEnv: "ZHIPU_API_KEY",
 		defaultModels: [
 			"glm-5.3",
+			"glm-5.3-flash",
 			"glm-5.2",
-			"glm-5",
-			"glm-4.6",
-			"glm-4.6v",
-			"glm-4.6v-flashx",
-			"glm-4.5v"
+			"glm-5.1"
 		],
 		modelsBaseURL: "https://open.bigmodel.cn/api/paas/v4",
 		modelsAuthMode: "bearer",
@@ -433,12 +556,9 @@ const PROVIDER_PRESETS = {
 		apiKeyEnv: "ZHIPU_API_KEY",
 		defaultModels: [
 			"glm-5.3",
+			"glm-5.3-flash",
 			"glm-5.2",
-			"glm-5",
-			"glm-4.6",
-			"glm-4.6v",
-			"glm-4.6v-flashx",
-			"glm-4.5v"
+			"glm-5.1"
 		],
 		authMode: "bearer",
 		modelsBaseURL: "https://open.bigmodel.cn/api/paas/v4",
@@ -451,7 +571,14 @@ const PROVIDER_PRESETS = {
 		protocol: "openai",
 		baseURL: "https://api.openai.com/v1",
 		apiKeyEnv: "OPENAI_API_KEY",
-		defaultModels: ["gpt-5", "gpt-5-mini"],
+		defaultModels: [
+			"gpt-6-astra",
+			"gpt-5.6-sol",
+			"gpt-5.6-terra",
+			"gpt-5.6-luna",
+			"gpt-5.5",
+			"gpt-5.4"
+		],
 		notes: "Official OpenAI API."
 	},
 	anthropic: {
@@ -460,7 +587,13 @@ const PROVIDER_PRESETS = {
 		protocol: "anthropic",
 		baseURL: "https://api.anthropic.com",
 		apiKeyEnv: "ANTHROPIC_API_KEY",
-		defaultModels: ["claude-sonnet-4-5", "claude-opus-4-1"],
+		defaultModels: [
+			"claude-opus-5",
+			"claude-sonnet-5",
+			"claude-opus-4-8",
+			"claude-opus-4-7",
+			"claude-sonnet-4-6"
+		],
 		authMode: "x-api-key",
 		notes: "Anthropic Messages API; docs in D:/X/X/14-claude."
 	},
@@ -470,7 +603,12 @@ const PROVIDER_PRESETS = {
 		protocol: "gemini",
 		baseURL: "https://generativelanguage.googleapis.com",
 		apiKeyEnv: "GEMINI_API_KEY",
-		defaultModels: ["gemini-3-flash-preview", "gemini-3-pro-preview"],
+		defaultModels: [
+			"gemini-3.8-flash",
+			"gemini-3.7-flash",
+			"gemini-3.1-pro-preview",
+			"gemini-3-flash-preview"
+		],
 		notes: "Google Gemini API; docs in D:/X/X/13-gemini."
 	},
 	siliconflow: {
@@ -479,7 +617,11 @@ const PROVIDER_PRESETS = {
 		protocol: "openai",
 		baseURL: "https://api.siliconflow.cn/v1",
 		apiKeyEnv: "SILICONFLOW_API_KEY",
-		defaultModels: ["deepseek-ai/DeepSeek-V3", "Qwen/Qwen3-235B-A22B"],
+		defaultModels: [
+			"deepseek-ai/DeepSeek-V3.2",
+			"Qwen/Qwen3-235B-A22B",
+			"zai-org/GLM-5.1"
+		],
 		notes: "OpenAI-compatible; also supports Anthropic-compatible at /v1/messages."
 	},
 	"siliconflow-anthropic": {
@@ -490,8 +632,8 @@ const PROVIDER_PRESETS = {
 		apiKeyEnv: "SILICONFLOW_API_KEY",
 		defaultModels: [
 			"deepseek-ai/DeepSeek-V3.2",
-			"moonshotai/Kimi-K2-Instruct",
-			"Qwen/Qwen3-32B"
+			"Qwen/Qwen3-32B",
+			"zai-org/GLM-5.1"
 		],
 		authMode: "bearer",
 		modelsBaseURL: "https://api.siliconflow.cn/v1",
@@ -505,8 +647,9 @@ const PROVIDER_PRESETS = {
 		baseURL: "https://aihubmix.com/v1",
 		apiKeyEnv: "AIHUBMIX_API_KEY",
 		defaultModels: [
-			"gpt-5",
+			"gpt-5.2",
 			"claude-sonnet-4-5",
+			"claude-opus-4-5",
 			"gemini-3-flash-preview"
 		],
 		notes: "Aggregator; also supports Anthropic-compatible at https://aihubmix.com."
@@ -519,8 +662,8 @@ const PROVIDER_PRESETS = {
 		apiKeyEnv: "AIHUBMIX_API_KEY",
 		defaultModels: [
 			"claude-sonnet-4-5",
-			"claude-opus-4-1",
-			"gpt-5"
+			"claude-opus-4-6",
+			"claude-opus-4-5"
 		],
 		authMode: "x-api-key",
 		modelsBaseURL: "https://aihubmix.com/v1",
@@ -534,8 +677,9 @@ const PROVIDER_PRESETS = {
 		baseURL: "https://openrouter.ai/api/v1",
 		apiKeyEnv: "OPENROUTER_API_KEY",
 		defaultModels: [
-			"openai/gpt-5",
-			"anthropic/claude-sonnet-4.5",
+			"openai/gpt-5.6",
+			"openai/gpt-5.5",
+			"anthropic/claude-sonnet-4.6",
 			"deepseek/deepseek-chat"
 		],
 		notes: "OpenAI-compatible; docs in D:/X/X/06-OpenRouter."
@@ -547,9 +691,9 @@ const PROVIDER_PRESETS = {
 		baseURL: "https://openrouter.ai/api",
 		apiKeyEnv: "OPENROUTER_API_KEY",
 		defaultModels: [
-			"anthropic/claude-sonnet-4.5",
-			"anthropic/claude-opus-4-1",
-			"openai/gpt-5",
+			"anthropic/claude-sonnet-4.6",
+			"anthropic/claude-opus-4-6",
+			"openai/gpt-5.6",
 			"deepseek/deepseek-chat"
 		],
 		authMode: "bearer",
@@ -562,52 +706,282 @@ function getProviderPreset(id) {
 	return PROVIDER_PRESETS[id];
 }
 //#endregion
+//#region src/dsml.ts
+const OUTER_OPEN = /<[｜|]{0,2}(?:DSML[｜|]{0,2})?(?:tool_calls|function_calls)\s*>/g;
+const OUTER_CLOSE_RE = /<\/[｜|]{0,2}(?:DSML[｜|]{0,2})?[｜|]{0,2}(?:tool_calls|function_calls)\s*>/;
+const INVOKE_RE = /<[｜|]{0,2}(?:DSML[｜|]{0,2})?[｜|]{0,2}\s*invoke\s+name="([^"]+)">([\s\S]*?)<\/[｜|]{0,2}(?:DSML[｜|]{0,2})?[｜|]{0,2}\s*invoke\s*>/g;
+const PARAM_RE = /<[｜|]{0,2}(?:DSML[｜|]{0,2})?[｜|]{0,2}\s*parameter\s+name="([^"]+)"(?:\s+string="true")?>([\s\S]*?)<\/[｜|]{0,2}(?:DSML[｜|]{0,2})?[｜|]{0,2}\s*parameter\s*>/g;
+const PARAM_OPEN_RE = /<[｜|]{0,2}(?:DSML[｜|]{0,2})?[｜|]{0,2}\s*parameter\s+name="[^"]+"/g;
+function insideFence(content, index) {
+	const fences = content.slice(0, index).match(/```/g);
+	return Boolean(fences && fences.length % 2 === 1);
+}
+/** Parse every invoke inside one outer block; null when any gate fails. */
+function parseOuterBlock(block, allowedNames) {
+	const calls = [];
+	INVOKE_RE.lastIndex = 0;
+	let invoke;
+	let invokeCount = 0;
+	while ((invoke = INVOKE_RE.exec(block)) !== null) {
+		invokeCount += 1;
+		const name = invoke[1];
+		if (!allowedNames.has(name)) return null;
+		const body = invoke[2];
+		const params = {};
+		PARAM_RE.lastIndex = 0;
+		let param;
+		let paramCount = 0;
+		while ((param = PARAM_RE.exec(body)) !== null) {
+			paramCount += 1;
+			if (PARAM_OPEN_RE.test(param[2])) return null;
+			params[param[1]] = param[2];
+			PARAM_OPEN_RE.lastIndex = 0;
+		}
+		if ((body.match(PARAM_OPEN_RE) ?? []).length !== paramCount) return null;
+		PARAM_OPEN_RE.lastIndex = 0;
+		calls.push({
+			id: `text-${calls.length}`,
+			name,
+			argumentsJson: paramCount > 0 ? JSON.stringify(params) : "{}"
+		});
+	}
+	if (invokeCount === 0) return null;
+	return calls;
+}
+/** Extract text tool calls; blocks outside code fences, whitelisted names. */
+function parseTextToolCalls(content, allowedNames) {
+	if (!content || allowedNames.size === 0) return {
+		toolCalls: [],
+		cleaned: content
+	};
+	const ranges = [];
+	const allCalls = [];
+	OUTER_OPEN.lastIndex = 0;
+	let match;
+	while ((match = OUTER_OPEN.exec(content)) !== null) {
+		const start = match.index;
+		if (insideFence(content, start)) continue;
+		const rest = content.slice(start);
+		const close = rest.match(OUTER_CLOSE_RE);
+		const calls = parseOuterBlock(close && close.index !== void 0 ? rest.slice(0, close.index) : rest, allowedNames);
+		if (calls === null) continue;
+		const end = close && close.index !== void 0 ? start + close.index + close[0].length : content.length;
+		allCalls.push(...calls);
+		ranges.push([start, end]);
+	}
+	if (allCalls.length === 0) return {
+		toolCalls: [],
+		cleaned: content
+	};
+	let cleaned = content;
+	for (let i = ranges.length - 1; i >= 0; i -= 1) {
+		const [from, to] = ranges[i];
+		cleaned = cleaned.slice(0, from) + cleaned.slice(to);
+	}
+	return {
+		toolCalls: allCalls,
+		cleaned: cleaned.trim()
+	};
+}
+//#endregion
 //#region src/providers/direct-http.ts
+/** True when the user asked to force thinking off, through either spelling. */
+function thinkingOff(advisor) {
+	return advisor.thinking === "disabled" || advisor.reasoningEffort === "off";
+}
+/**
+* OpenAI-compatible reasoning params, provider-aware (2026-09-05 知识 X 最新):
+* - DeepSeek: `thinking {type}` switch + `reasoning_effort` low/high/max.
+* - Kimi/Kimi Code: kimi-k3 = top-level `reasoning_effort` low/high/max (no
+*   `thinking`); kimi-k2.7-code = always thinking, no params; kimi-k2.6 =
+*   `thinking {type}` switch only.
+* - OpenAI official: `reasoning_effort` minimal|low|medium|high (no max, no off).
+* - 阿里云百炼: `enable_thinking` bool (Qwen chat completions).
+* - 智谱: `thinking {type}` switch + `reasoning_effort` (GLM-5.3 force-enabled,
+*   disabled→enabled+low; GLM-5.2/5.1/5 can disable).
+* - 硅基流动: `enable_thinking` + `thinking_budget`.
+* - Others (AIHubMix/OpenRouter/custom): best-effort `reasoning_effort`.
+*/
+function openAiThinkingParams(advisor) {
+	const provider = advisor.provider;
+	const model = advisor.model;
+	const effort = advisor.reasoningEffort;
+	if (provider === "deepseek") {
+		if (thinkingOff(advisor)) return { thinking: { type: "disabled" } };
+		return {
+			...advisor.thinking === "enabled" ? { thinking: { type: "enabled" } } : {},
+			...effort && effort !== "off" ? { reasoning_effort: effort } : {}
+		};
+	}
+	if (provider === "moonshot" || provider === "kimi-code") {
+		if (/^(k3|kimi-k3)(-|$)/i.test(model)) {
+			if (effort && effort !== "off") return { reasoning_effort: effort };
+			return {};
+		}
+		if (/kimi-k2\.7-code|kimi-for-coding/i.test(model)) return {};
+		if (advisor.thinking === "disabled") return { thinking: { type: "disabled" } };
+		if (advisor.thinking === "enabled") return { thinking: { type: "enabled" } };
+		return {};
+	}
+	if (provider === "openai") {
+		if (effort === "low") return { reasoning_effort: "low" };
+		if (effort === "high" || effort === "max") return { reasoning_effort: "high" };
+		return {};
+	}
+	if (provider === "bailian" || provider === "bailian-openai-singapore" || provider === "bailian-openai-us") {
+		if (thinkingOff(advisor)) return { enable_thinking: false };
+		if (advisor.thinking === "enabled") return { enable_thinking: true };
+		return {};
+	}
+	if (provider === "zhipu") {
+		if (/glm-5\.3/i.test(model)) return {
+			thinking: { type: "enabled" },
+			reasoning_effort: thinkingOff(advisor) ? "low" : effort && effort !== "off" ? effort : "max"
+		};
+		if (/glm-5/i.test(model)) {
+			if (thinkingOff(advisor)) return { thinking: { type: "disabled" } };
+			return {
+				...advisor.thinking === "enabled" ? { thinking: { type: "enabled" } } : {},
+				...effort && effort !== "off" ? { reasoning_effort: effort } : {}
+			};
+		}
+		if (advisor.thinking === "disabled") return { thinking: { type: "disabled" } };
+		if (advisor.thinking === "enabled") return { thinking: { type: "enabled" } };
+		return {};
+	}
+	if (provider === "siliconflow") {
+		if (thinkingOff(advisor)) return { enable_thinking: false };
+		const budget = effort === "low" ? 1024 : effort === "high" ? 8192 : effort === "max" ? 16384 : 0;
+		return {
+			...advisor.thinking === "enabled" || budget > 0 ? { enable_thinking: true } : {},
+			...budget > 0 ? { thinking_budget: budget } : {}
+		};
+	}
+	if (effort && effort !== "off") return { reasoning_effort: effort };
+	return {};
+}
+/** Anthropic Messages reasoning params, provider-aware.
+*  Official Anthropic: adaptive (`thinking:{type:'adaptive'}` + `output_config.effort`)
+*  for Opus 4.6/4.7/Sonnet 4.6/Mythos; manual `enabled`+`budget_tokens` for older
+*  models. DeepSeek's Anthropic-compatible endpoint uses `reasoning.effort` +
+*  `output_config.effort` (its `budget_tokens` is ignored). Other Anthropic skins
+*  follow the official Anthropic wire shape. */
+function anthropicThinkingParams(advisor) {
+	if (advisor.provider === "deepseek-anthropic") {
+		if (thinkingOff(advisor)) return { reasoning: { effort: "none" } };
+		if (advisor.reasoningEffort && advisor.reasoningEffort !== "off") return {
+			reasoning: { effort: advisor.reasoningEffort },
+			output_config: { effort: advisor.reasoningEffort }
+		};
+		return {};
+	}
+	if (thinkingOff(advisor)) return { thinking: { type: "disabled" } };
+	if (advisor.reasoningEffort && advisor.reasoningEffort !== "off") {
+		if (/(?:opus-4-[6-8]|opus-5|sonnet-4-6|sonnet-5|fable|mythos)/i.test(advisor.model)) return {
+			thinking: { type: "adaptive" },
+			output_config: { effort: advisor.reasoningEffort }
+		};
+		return { thinking: {
+			type: "enabled",
+			budget_tokens: advisor.reasoningEffort === "low" ? 2048 : advisor.reasoningEffort === "high" ? 8192 : 16384
+		} };
+	}
+	return {};
+}
+/** Gemini `generateContent` thinking config. Disabling thinking uses
+*  `thinkingBudget: 0` (NOT `includeThoughts:false`, which only hides returned
+*  thoughts). Effort uses `thinkingBudget`, the backward-compatible knob that
+*  works across Gemini 2.5 and Gemini 3 per Google's migration note. */
+function geminiThinkingParams(advisor) {
+	if (thinkingOff(advisor)) return { thinkingConfig: { thinkingBudget: 0 } };
+	const budget = advisor.reasoningEffort === "low" ? 1024 : advisor.reasoningEffort === "high" ? 8192 : advisor.reasoningEffort === "max" ? 16384 : 0;
+	if (budget > 0) return { thinkingConfig: { thinkingBudget: budget } };
+	return {};
+}
+/** Validate the per-advisor thinking fields at the direct-http boundary so a
+*  malformed stored config fails loudly instead of silently ignoring. */
+function assertThinkingParams(advisor) {
+	if (advisor.thinking !== void 0 && advisor.thinking !== "enabled" && advisor.thinking !== "disabled") throw new Error(`Unsupported thinking mode: ${String(advisor.thinking)}`);
+	if (advisor.reasoningEffort !== void 0 && ![
+		"off",
+		"low",
+		"high",
+		"max"
+	].includes(advisor.reasoningEffort)) throw new Error(`Unsupported reasoningEffort: ${String(advisor.reasoningEffort)}`);
+}
 /** Stream an advisor response through direct HTTP (OpenAI/Anthropic SSE). */
-async function streamDirectHttp(advisor, transcript, onDelta, signal, timeoutMs) {
+async function streamDirectHttp(advisor, transcript, onDelta, signal, timeoutMs, tools, parseToolNames) {
 	const preset = advisor.provider ? getProviderPreset(advisor.provider) : void 0;
 	const baseURL = (advisor.baseURL ?? preset?.baseURL ?? "").replace(/\/+$/, "");
 	const apiKey = advisor.apiKey ?? process.env[advisor.apiKeyEnv ?? preset?.apiKeyEnv ?? ""] ?? "";
 	const protocol = advisor.protocol ?? preset?.protocol ?? "openai";
 	const authMode = advisor.authMode ?? preset?.authMode ?? "x-api-key";
+	assertThinkingParams(advisor);
 	if (!baseURL || !apiKey) throw new Error(`Direct HTTP fallback is not configured for advisor "${advisor.name}" (${advisor.provider}). Add baseURL/apiKey or configure the provider in DSH.`);
 	const transcriptText = transcript.map((entry) => `[${entry.role === "main" ? "主模型" : entry.name}]\n${entry.content}`).join("\n\n");
+	let result;
 	switch (protocol) {
-		case "openai": return streamOpenAICompatible(baseURL, advisor, apiKey, transcriptText, onDelta, signal, timeoutMs);
-		case "anthropic": return streamAnthropic(baseURL, advisor, apiKey, transcriptText, authMode, onDelta, signal, timeoutMs);
+		case "openai":
+			result = await streamOpenAICompatible(baseURL, advisor, apiKey, transcriptText, onDelta, signal, timeoutMs, tools);
+			break;
+		case "anthropic":
+			result = await streamAnthropic(baseURL, advisor, apiKey, transcriptText, authMode, onDelta, signal, timeoutMs, tools);
+			break;
 		case "gemini": {
 			const content = await callGemini(baseURL, advisor, apiKey, transcriptText, signal, timeoutMs);
 			onDelta({ text: content });
-			return {
+			result = {
 				content,
 				thinking: ""
 			};
+			break;
 		}
 		default: throw new Error(`Unsupported protocol: ${String(protocol)}`);
 	}
+	const parsed = parseTextToolCalls(result.content, new Set(parseToolNames ?? (tools ?? []).map((tool) => tool.name)));
+	if (parsed.toolCalls.length > 0) {
+		console.warn(`[dsh-advisor-group] 检测到文本工具调用（DSML/非结构流），已还原并交给 runAdvisorToolLoop 执行：${parsed.toolCalls.map((tool) => tool.name).join("、")}`);
+		return {
+			...result,
+			content: parsed.cleaned,
+			toolCalls: [...result.toolCalls ?? [], ...parsed.toolCalls]
+		};
+	}
+	return result;
 }
-async function streamOpenAICompatible(baseURL, advisor, apiKey, transcriptText, onDelta, signal, timeoutMs) {
+async function streamOpenAICompatible(baseURL, advisor, apiKey, transcriptText, onDelta, signal, timeoutMs, tools) {
 	const endpoint = baseURL.endsWith("/chat/completions") ? baseURL : `${baseURL.replace(/\/+$/, "")}/chat/completions`;
+	const pair = timeoutSignalPair(timeoutMs ?? 12e4, signal);
+	const body = {
+		model: advisor.model,
+		messages: [{
+			role: "system",
+			content: advisor.systemPrompt + ADVISOR_OUTPUT_POLICY
+		}, {
+			role: "user",
+			content: transcriptText
+		}],
+		temperature: advisor.temperature ?? .3,
+		max_tokens: advisor.maxTokens ?? 16384,
+		...openAiThinkingParams(advisor),
+		stream: true
+	};
+	if (tools && tools.length > 0) body.tools = tools.map((tool) => ({
+		type: "function",
+		function: {
+			name: tool.name,
+			...tool.description === void 0 ? {} : { description: tool.description },
+			parameters: tool.parameters ?? { type: "object" }
+		}
+	}));
 	const response = await fetch(endpoint, {
 		method: "POST",
 		headers: {
 			"content-type": "application/json",
 			authorization: `Bearer ${apiKey}`
 		},
-		body: JSON.stringify({
-			model: advisor.model,
-			messages: [{
-				role: "system",
-				content: advisor.systemPrompt + ADVISOR_OUTPUT_POLICY
-			}, {
-				role: "user",
-				content: transcriptText
-			}],
-			temperature: advisor.temperature ?? .3,
-			max_tokens: advisor.maxTokens ?? 16384,
-			stream: true
-		}),
-		signal: withTimeout(timeoutMs ?? 12e4, signal)
+		body: JSON.stringify(body),
+		signal: pair.signal
 	});
 	if (!response.ok) throw new Error(`OpenAI-compatible request failed: ${response.status} ${await response.text()}`);
 	if (!response.body) throw new Error("OpenAI-compatible response has no body");
@@ -616,6 +990,7 @@ async function streamOpenAICompatible(baseURL, advisor, apiKey, transcriptText,
 	let buffer = "";
 	let content = "";
 	let thinking = "";
+	const toolCalls = /* @__PURE__ */ new Map();
 	while (true) {
 		let done;
 		let value;
@@ -624,8 +999,28 @@ async function streamOpenAICompatible(baseURL, advisor, apiKey, transcriptText,
 			done = read.done;
 			value = read.value;
 		} catch (error) {
+			const errName = error?.name ?? "UnknownError";
+			const errMessage = error instanceof Error ? error.message : String(error);
+			console.warn(`[dsh-advisor-group] direct-http stream read error: advisor=${advisor.name} base=${baseURL} name=${errName} msg=${errMessage} isTimeout=${pair.isTimeout()} signalAborted=${signal?.aborted ?? false} collectedChars=${content.length + thinking.length}`);
+			if (pair.isTimeout()) return {
+				content,
+				thinking,
+				toolCalls: collectToolCalls(toolCalls),
+				truncated: {
+					reason: "timeout",
+					atMs: Date.now()
+				}
+			};
 			if (signal?.aborted || error?.name === "AbortError") throw error;
-			break;
+			return {
+				content,
+				thinking,
+				toolCalls: collectToolCalls(toolCalls),
+				truncated: {
+					reason: "network",
+					atMs: Date.now()
+				}
+			};
 		}
 		if (done) break;
 		buffer += decoder.decode(value, { stream: true });
@@ -641,6 +1036,18 @@ async function streamOpenAICompatible(baseURL, advisor, apiKey, transcriptText,
 					const delta = JSON.parse(data).choices?.[0]?.delta ?? {};
 					const text = typeof delta.content === "string" ? delta.content : "";
 					const think = typeof delta.reasoning_content === "string" ? delta.reasoning_content : typeof delta.reasoning === "string" ? delta.reasoning : "";
+					if (delta.tool_calls) for (const tc of delta.tool_calls) {
+						const idx = tc.index ?? 0;
+						const current = toolCalls.get(idx) ?? {
+							id: "",
+							name: "",
+							args: ""
+						};
+						if (tc.id) current.id = tc.id;
+						if (tc.function?.name) current.name = tc.function.name;
+						if (tc.function?.arguments) current.args += tc.function.arguments;
+						toolCalls.set(idx, current);
+					}
 					if (text) {
 						content += text;
 						onDelta({ text });
@@ -655,10 +1062,24 @@ async function streamOpenAICompatible(baseURL, advisor, apiKey, transcriptText,
 	}
 	return {
 		content,
-		thinking
+		thinking,
+		toolCalls: collectToolCalls(toolCalls)
 	};
 }
-async function streamAnthropic(baseURL, advisor, apiKey, transcriptText, authMode, onDelta, signal, timeoutMs) {
+/** Map<index, {id,name,args}> → model-facing tool call list. */
+function collectToolCalls(byIndex) {
+	const calls = [];
+	for (const [, current] of byIndex) {
+		if (!current.name) continue;
+		calls.push({
+			id: current.id || `tool-${calls.length}`,
+			name: current.name,
+			argumentsJson: current.args
+		});
+	}
+	return calls.length > 0 ? calls : void 0;
+}
+async function streamAnthropic(baseURL, advisor, apiKey, transcriptText, authMode, onDelta, signal, timeoutMs, tools) {
 	const endpoint = baseURL.endsWith("/v1/messages") ? baseURL : baseURL.endsWith("/v1") ? `${baseURL}/messages` : `${baseURL.replace(/\/+$/, "")}/v1/messages`;
 	const headers = {
 		"content-type": "application/json",
@@ -666,21 +1087,29 @@ async function streamAnthropic(baseURL, advisor, apiKey, transcriptText, authMod
 	};
 	if (authMode === "bearer") headers["authorization"] = `Bearer ${apiKey}`;
 	else headers["x-api-key"] = apiKey;
+	const body = {
+		model: advisor.model,
+		system: advisor.systemPrompt + ADVISOR_OUTPUT_POLICY,
+		messages: [{
+			role: "user",
+			content: transcriptText
+		}],
+		temperature: advisor.temperature ?? .3,
+		max_tokens: advisor.maxTokens ?? 16384,
+		...anthropicThinkingParams(advisor),
+		stream: true
+	};
+	if (tools && tools.length > 0) body.tools = tools.map((tool) => ({
+		name: tool.name,
+		...tool.description === void 0 ? {} : { description: tool.description },
+		input_schema: tool.parameters ?? { type: "object" }
+	}));
+	const pair = timeoutSignalPair(timeoutMs ?? 12e4, signal);
 	const response = await fetch(endpoint, {
 		method: "POST",
 		headers,
-		body: JSON.stringify({
-			model: advisor.model,
-			system: advisor.systemPrompt + ADVISOR_OUTPUT_POLICY,
-			messages: [{
-				role: "user",
-				content: transcriptText
-			}],
-			temperature: advisor.temperature ?? .3,
-			max_tokens: advisor.maxTokens ?? 16384,
-			stream: true
-		}),
-		signal: withTimeout(timeoutMs ?? 12e4, signal)
+		body: JSON.stringify(body),
+		signal: pair.signal
 	});
 	if (!response.ok) throw new Error(`Anthropic request failed: ${response.status} ${await response.text()}`);
 	if (!response.body) throw new Error("Anthropic response has no body");
@@ -689,6 +1118,9 @@ async function streamAnthropic(baseURL, advisor, apiKey, transcriptText, authMod
 	let buffer = "";
 	let content = "";
 	let thinking = "";
+	let currentTool = null;
+	let activeTool = false;
+	const toolCalls = [];
 	while (true) {
 		let done;
 		let value;
@@ -697,8 +1129,28 @@ async function streamAnthropic(baseURL, advisor, apiKey, transcriptText, authMod
 			done = read.done;
 			value = read.value;
 		} catch (error) {
+			const errName = error?.name ?? "UnknownError";
+			const errMessage = error instanceof Error ? error.message : String(error);
+			console.warn(`[dsh-advisor-group] anthropic stream read error: advisor=${advisor.name} base=${baseURL} name=${errName} msg=${errMessage} isTimeout=${pair.isTimeout()} signalAborted=${signal?.aborted ?? false} collectedChars=${content.length + thinking.length}`);
+			if (pair.isTimeout()) return {
+				content,
+				thinking,
+				toolCalls: toolCalls.length > 0 ? toolCalls : void 0,
+				truncated: {
+					reason: "timeout",
+					atMs: Date.now()
+				}
+			};
 			if (signal?.aborted || error?.name === "AbortError") throw error;
-			break;
+			return {
+				content,
+				thinking,
+				toolCalls: toolCalls.length > 0 ? toolCalls : void 0,
+				truncated: {
+					reason: "network",
+					atMs: Date.now()
+				}
+			};
 		}
 		if (done) break;
 		buffer += decoder.decode(value, { stream: true });
@@ -712,13 +1164,32 @@ async function streamAnthropic(baseURL, advisor, apiKey, transcriptText, authMod
 				if (!data) continue;
 				try {
 					const json = JSON.parse(data);
-					if (json.type === "content_block_delta" && json.delta) {
+					if (json.type === "content_block_start" && json.content_block?.type === "tool_use") {
+						currentTool = currentTool === null ? {
+							id: json.content_block.id ?? "",
+							name: json.content_block.name ?? "",
+							args: ""
+						} : currentTool;
+						activeTool = true;
+					} else if (json.type === "content_block_delta" && json.delta) {
 						if (json.delta.type === "text_delta" && typeof json.delta.text === "string") {
 							content += json.delta.text;
 							onDelta({ text: json.delta.text });
 						} else if (json.delta.type === "thinking_delta" && typeof json.delta.thinking === "string") {
 							thinking += json.delta.thinking;
 							onDelta({ thinking: json.delta.thinking });
+						} else if (json.delta.type === "input_json_delta" && typeof json.delta.partial_json === "string") {
+							if (currentTool) currentTool.args += json.delta.partial_json;
+						}
+					} else if (json.type === "content_block_stop") {
+						if (activeTool && currentTool) {
+							toolCalls.push({
+								id: currentTool.id,
+								name: currentTool.name,
+								argumentsJson: currentTool.args
+							});
+							currentTool = null;
+							activeTool = false;
 						}
 					}
 				} catch {}
@@ -727,7 +1198,8 @@ async function streamAnthropic(baseURL, advisor, apiKey, transcriptText, authMod
 	}
 	return {
 		content,
-		thinking
+		thinking,
+		toolCalls: toolCalls.length > 0 ? toolCalls : void 0
 	};
 }
 async function callGemini(baseURL, advisor, apiKey, transcriptText, signal, timeoutMs) {
@@ -745,7 +1217,8 @@ async function callGemini(baseURL, advisor, apiKey, transcriptText, signal, time
 			systemInstruction: { parts: [{ text: advisor.systemPrompt + ADVISOR_OUTPUT_POLICY }] },
 			generationConfig: {
 				temperature: advisor.temperature ?? .3,
-				maxOutputTokens: advisor.maxTokens ?? 16384
+				maxOutputTokens: advisor.maxTokens ?? 16384,
+				...geminiThinkingParams(advisor)
 			}
 		}),
 		signal: withTimeout(timeoutMs ?? 12e4, signal)
@@ -764,7 +1237,7 @@ async function callGemini(baseURL, advisor, apiKey, transcriptText, signal, time
 * stalls. Any generation error degrades to the fallback text, never throws.
 */
 const FALLBACK_DEEPEN_QUESTION = "请基于以上全部讨论，针对仍然存在的分歧、未验证的假设与主要风险，提出一个有深度的后续问题。";
-const FALLBACK_CONCLUSION = "请综合以上全部讨论，给出结论：各位顾问的共识、分歧，以及最可靠的行动建议。";
+const FALLBACK_CONCLUSION = "（⚠ 驱动模型未能生成综合结论：请主模型以上述顾问讨论为基础，自行总结共识、分歧与最可靠的行动建议。）";
 const DRIVER_SYSTEM_PROMPT = `
 你是顾问群的主持人（驱动模型）。你的职责：
 1. 阅读项目背景、主模型问题与全部顾问回答；
@@ -795,7 +1268,7 @@ function transcriptOf(session) {
 		return `[${message.role === "main" ? "主模型" : message.advisorName ?? message.advisorId ?? "顾问"}]\n${message.content}`;
 	}).join("\n\n").slice(-24e3);
 }
-async function generateWith(ctx, source, system, prompt, signal) {
+async function generateWith(ctx, source, sessionId, system, prompt, signal, timeoutMs = 6e5) {
 	try {
 		const options = {
 			provider: source.provider,
@@ -806,31 +1279,56 @@ async function generateWith(ctx, source, system, prompt, signal) {
 					type: "text",
 					text: prompt
 				}],
-				source: {
-					kind: "plugin",
-					plugin: "dsh-advisor-group"
-				}
+				source: { kind: "user" }
 			})],
 			temperature: .7,
-			maxTokens: 600,
-			signal: withTimeout(6e4, signal)
+			maxTokens: 2048,
+			signal: withTimeout(timeoutMs, signal)
 		};
 		let text = "";
-		for await (const chunk of ctx.llm.stream(options)) if (chunk.type === "text-delta") text += chunk.text;
-		return text.trim() || void 0;
-	} catch {
+		let chunks = 0;
+		let textDeltas = 0;
+		let reasoningDeltas = 0;
+		for await (const chunk of ctx.llm.stream(options)) {
+			chunks += 1;
+			if (chunk.type === "text-delta") {
+				textDeltas += 1;
+				text += chunk.text;
+			} else reasoningDeltas += 1;
+		}
+		const trimmed = text.trim();
+		if (!trimmed) console.warn(`[dsh-advisor-group] 驱动模型空响应（${source.provider}/${source.model}）：`, sessionId, `chunks=${chunks} textDeltas=${textDeltas} reasoningDeltas=${reasoningDeltas}；建议配置 discussion.driverModel 或提高 maxTokens`);
+		return trimmed || void 0;
+	} catch (error) {
+		if (signal?.aborted) throw error;
+		console.warn(`[dsh-advisor-group] 驱动模型生成失败（${source.provider}/${source.model}）：`, sessionId, error instanceof Error ? error.message : String(error));
 		return;
 	}
 }
 /** One deepening question for the next round (never throws; returns fallback text). */
-async function generateDeepenQuestion(ctx, session, source, signal) {
+async function generateDeepenQuestion(ctx, session, source, signal, timeoutMs) {
 	if (!source) return FALLBACK_DEEPEN_QUESTION;
-	return await generateWith(ctx, source, DRIVER_SYSTEM_PROMPT, `以下是一轮顾问群的讨论记录：\n\n${transcriptOf(session)}\n\n请给出下一步的深入追问。`, signal) || "请基于以上全部讨论，针对仍然存在的分歧、未验证的假设与主要风险，提出一个有深度的后续问题。";
+	return await generateWith(ctx, source, session.id, DRIVER_SYSTEM_PROMPT, `以下是一轮顾问群的讨论记录：\n\n${transcriptOf(session)}\n\n请给出下一步的深入追问。`, signal, timeoutMs) || "请基于以上全部讨论，针对仍然存在的分歧、未验证的假设与主要风险，提出一个有深度的后续问题。";
 }
 /** Final synthesis after all rounds (never throws; returns fallback text). */
-async function generateConclusion(ctx, session, source, signal) {
+async function generateConclusion(ctx, session, source, signal, timeoutMs) {
 	if (!source) return FALLBACK_CONCLUSION;
-	return await generateWith(ctx, source, CONCLUSION_SYSTEM_PROMPT, `以下是顾问群全部 ${session.messages.filter((m) => m.role === "advisor").length} 条顾问回答的讨论记录：\n\n${transcriptOf(session)}\n\n请给出综合结论。`, signal) || "请综合以上全部讨论，给出结论：各位顾问的共识、分歧，以及最可靠的行动建议。";
+	return await generateWith(ctx, source, session.id, CONCLUSION_SYSTEM_PROMPT, `以下是顾问群全部 ${session.messages.filter((m) => m.role === "advisor").length} 条顾问回答的讨论记录：\n\n${transcriptOf(session)}\n\n请给出综合结论。`, signal, timeoutMs) || "（⚠ 驱动模型未能生成综合结论：请主模型以上述顾问讨论为基础，自行总结共识、分歧与最可靠的行动建议。）";
+}
+//#endregion
+//#region src/risk-notes.ts
+const TRUNCATION_NOTE = "有顾问输出在流式过程中被截断（超时或网络中断），其正文可能不完整，且该顾问本轮可能只提供了部分意见。";
+const CANCEL_NOTE = "本次顾问群讨论已被用户或主模型取消，结论可能不完整。";
+/**
+* 汇总风险提示（两条规则：截断 / 取消）。
+* 判定只看消息上的截断标志与 cancelled 标志，**不读正文**——没有基于文本的猜测，
+* 因此不会误报。content 字段保留在入参类型里只为兼容既有调用方（本函数不再读它）。
+*/
+function buildRiskNotes(messages, cancelled) {
+	const notes = [];
+	if (messages.some((message) => message.truncated !== void 0)) notes.push(TRUNCATION_NOTE);
+	if (cancelled) notes.push(CANCEL_NOTE);
+	return notes;
 }
 //#endregion
 //#region src/session-log.ts
@@ -889,7 +1387,11 @@ function appendAdvisorMessage(log, sessionId, message) {
 		...message.advisorId === void 0 ? {} : { advisorId: message.advisorId },
 		...message.advisorName === void 0 ? {} : { advisorName: message.advisorName },
 		...message.round === void 0 ? {} : { round: message.round },
-		...message.thinking === void 0 ? {} : { thinking: message.thinking }
+		...message.thinking === void 0 ? {} : { thinking: message.thinking },
+		...message.thinkingSegments === void 0 ? {} : { thinkingSegments: message.thinkingSegments },
+		...message.actionDescriptions === void 0 ? {} : { actionDescriptions: message.actionDescriptions },
+		...message.toolSteps === void 0 ? {} : { toolSteps: message.toolSteps },
+		...message.truncated === void 0 ? {} : { truncated: message.truncated }
 	};
 	log.append("advisor-group/message", data);
 }
@@ -904,10 +1406,19 @@ function appendAdvisorDelta(log, sessionId, delta) {
 		...delta.round === void 0 ? {} : { round: delta.round },
 		...delta.contentDelta === void 0 ? {} : { contentDelta: delta.contentDelta },
 		...delta.thinkingDelta === void 0 ? {} : { thinkingDelta: delta.thinkingDelta },
+		...delta.toolStep === void 0 ? {} : { toolStep: delta.toolStep },
 		...delta.done === void 0 ? {} : { done: delta.done }
 	};
 	log.append("advisor-group/delta", data);
 }
+function appendAdvisorResume(log, sessionId) {
+	const { turn, step } = latestTurnStep(log);
+	log.append("advisor-group/resume", {
+		sessionId,
+		turn,
+		step
+	});
+}
 function appendAdvisorEnd(log, session, summary) {
 	const { turn, step } = latestTurnStep(log);
 	const data = {
@@ -921,11 +1432,68 @@ function appendAdvisorEnd(log, session, summary) {
 				advisorName,
 				corePoints
 			})),
-			...summary.riskNotes.length > 0 ? { riskNotes: summary.riskNotes } : {}
+			...summary.riskNotes.length > 0 ? { riskNotes: summary.riskNotes } : {},
+			...summary.stopped === void 0 ? {} : { stopped: summary.stopped },
+			...summary.conclusion === void 0 ? {} : { conclusion: summary.conclusion }
 		}
 	};
 	log.append("advisor-group/end", data);
 }
+/**
+* Open the cross-session ask card (waiting state). Reuses the registered
+* `advisor-group/start` event with `kind:'ask'` so no new event type needs
+* persistence registration; the client branches on `kind` for the ask layout.
+*/
+function appendAskStart(log, info) {
+	const { turn, step } = latestTurnStep(log);
+	log.append("advisor-group/start", {
+		sessionId: info.askId,
+		turn,
+		step,
+		kind: "ask",
+		question: info.question,
+		...info.context === void 0 ? {} : { context: info.context },
+		advisors: [{
+			id: "ask-sender",
+			name: info.senderName
+		}, {
+			id: "ask-target",
+			name: info.targetName
+		}]
+	});
+}
+/**
+* Close the cross-session ask card: the target's answer as an advisor-role
+* bubble, then the end event whose `summary.conclusion` is the note line.
+* Message is appended BEFORE end so the assembled state has the answer in
+* place when the status flips to completed.
+*/
+function appendAskAnswer(log, info) {
+	const { turn, step } = latestTurnStep(log);
+	log.append("advisor-group/message", {
+		sessionId: info.askId,
+		turn,
+		step,
+		role: "advisor",
+		advisorId: "ask-target",
+		advisorName: info.targetName,
+		content: info.answer.trim() ? info.answer : "（未收到回复，详见下方投递说明）"
+	});
+	log.append("advisor-group/end", {
+		sessionId: info.askId,
+		turn,
+		step,
+		summary: {
+			question: info.question,
+			advisors: [{
+				advisorId: "ask-target",
+				advisorName: info.targetName,
+				corePoints: []
+			}],
+			conclusion: info.note
+		}
+	});
+}
 //#endregion
 //#region src/stream-channel.ts
 const BOOT_ID = `${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 8)}`;
@@ -998,12 +1566,212 @@ function publish(sessionId, delta) {
 	for (const res of set) writeFrame(res, frame);
 }
 //#endregion
+//#region src/advisor-tools.ts
+/** Raw tool-call / DSML markup that some providers echo as plain text instead
+*  of a structured tool call. Observed shapes (deepseek-v4-pro, anthropic
+*  channel):
+*   - plain `<tool_calls>…</tool_calls>` (line-start block)
+*   - piped `<｜｜tool_calls>\n<｜｜DSML｜…` without a closing tag (U+FF5C)
+*   - MID-LINE occurrence: `证据。<tool_calls><invoke name="bash">…` — the
+*     marker is glued to prose, so line-start anchoring alone is NOT enough. */
+const RAW_TOOL_CALL_BLOCK = /<(?:[｜|]{1,2})?(?:tool_calls|DSML)[\s\S]*?<\/(?:[｜|]{1,2})?(?:tool_calls|DSML)\s*>/g;
+/** Unclosed tail: cut from the first marker that is actually a STRUCTURED
+*  block (followed by `<invoke`/`<parameter`/`<｜｜DSML` within the next ~60
+*  chars). An isolated prose/code mention like `` `<tool_calls>` `` is never
+*  followed by those children, so it stays untouched. */
+const RAW_TOOL_CALL_TAIL = /<(?:[｜|]{1,2})?(?:tool_calls|DSML)[\s\S]{0,60}?(?:<invoke|<parameter|<[｜|]{1,2}DSML)[\s\S]*$/g;
+/**
+* Strip raw `<tool_calls>` / piped `<｜｜tool_calls>` / `<｜｜DSML｜` markup from
+* an advisor's final content (line-start OR mid-line, closed or unclosed).
+* Returns the cleaned text; when the content was only leaked tool markup, a
+* short notice replaces it so the bubble is never empty/raw (the thinking
+* panel and tool rows still carry the real substance).
+*/
+function sanitizeAdvisorContent(content) {
+	if (!content.trim()) return content;
+	return content.replace(RAW_TOOL_CALL_BLOCK, "").replace(RAW_TOOL_CALL_TAIL, "").trim() || "（顾问未生成正文：输出仅为工具调用描述，请参考其思维链与工具调用记录。）";
+}
+/** Conservative read-only whitelist for the default advisor tool scope. */
+const READONLY_TOOL_NAMES = /* @__PURE__ */ new Set([
+	"read",
+	"grep",
+	"glob",
+	"web_search",
+	"web_fetch",
+	"scan_discover",
+	"list_imported_sessions"
+]);
+const MAX_TOOL_ROUNDS = 4;
+const TOOL_EXEC_TIMEOUT_MS = 3e4;
+const TOOL_RESULT_MAX_CHARS = 8e3;
+/**
+* Resolve the advisor-visible tool schemas. `agent` (when provided) makes the
+* registry resolve the scoped view — exactly the tools that session's Agent
+* loop serves, honouring presets and restrict() filters.
+*/
+function resolveAdvisorToolSchemas(ctx, agent, mode) {
+	if (mode === "off") return [];
+	const schemas = ctx.tools?.schemas?.(agent) ?? [];
+	let list = schemas;
+	if (mode === "readonly") list = schemas.filter((schema) => typeof schema === "object" && schema !== null && READONLY_TOOL_NAMES.has(schema.name));
+	return list.map((schema) => {
+		const { name, description, parameters } = schema;
+		return {
+			name,
+			...typeof description === "string" ? { description } : {},
+			...parameters === void 0 ? {} : { parameters }
+		};
+	});
+}
+/** Human-readable one-line summary of a tool result (for the thinking panel). */
+function summarizeToolResult(result) {
+	let text = "";
+	if (result.isError) text = `（错误）${result.error?.message ?? "unknown"}`;
+	else {
+		text = (result.content ?? []).map((block) => typeof block === "object" && block !== null && "text" in block && typeof block.text === "string" ? block.text : "").join("").trim();
+		if (!text && result.value !== void 0) try {
+			text = JSON.stringify(result.value);
+		} catch {
+			text = String(result.value);
+		}
+	}
+	return text.length > TOOL_RESULT_MAX_CHARS ? `${text.slice(0, TOOL_RESULT_MAX_CHARS)}…（已截断）` : text;
+}
+/**
+* Execute one tool via the official pipeline (scoped dispatch when `agent` is
+* provided). Returns the model-facing text; never throws.
+*/
+async function executeAdvisorTool(ctx, agent, call, signal) {
+	const tools = ctx.tools;
+	if (!tools?.execute) return "（顾问工具执行不可用：当前环境未提供工具执行服务）";
+	let args = {};
+	try {
+		args = call.argumentsJson ? JSON.parse(call.argumentsJson) : {};
+	} catch {
+		return "（工具参数不是合法 JSON）";
+	}
+	try {
+		return summarizeToolResult(await tools.execute({
+			callId: call.id,
+			name: call.name,
+			arguments: args,
+			...agent === void 0 ? {} : { agent },
+			signal: withTimeout(TOOL_EXEC_TIMEOUT_MS, signal)
+		}));
+	} catch (error) {
+		return `（工具执行失败：${error instanceof Error ? error.message : String(error)}）`;
+	}
+}
+/**
+* Drive the tool-calling loop: stream with `tools` → execute every
+* model-requested call (each step surfaced via `onToolStep` as its own row) →
+* feed a compact "already executed" block back → repeat until the model
+* answers (or MAX_TOOL_ROUNDS, then one final no-tools round to force a text
+* answer).
+*/
+async function runAdvisorToolLoop(tools, streamOnce, executeTool, onToolStep) {
+	let extraContext = "";
+	let truncated;
+	const executedMemo = /* @__PURE__ */ new Map();
+	const executeDeduped = async (call) => {
+		const key = `${call.name}:${call.argumentsJson}`;
+		const cached = executedMemo.get(key);
+		if (cached !== void 0) {
+			console.warn(`[dsh-advisor-group] 同参工具调用去重（复用已执行结果）：${call.name}`);
+			return {
+				outcome: cached,
+				reused: true
+			};
+		}
+		const outcome = await executeTool(call);
+		executedMemo.set(key, outcome);
+		return {
+			outcome,
+			reused: false
+		};
+	};
+	/** Result fed back to the model: capped so extraContext never balloons
+	*  (display rows keep their own 240-char preview). */
+	const capOutcome = (outcome) => outcome.length > 800 ? `${outcome.slice(0, 800)}…` : outcome;
+	for (let round = 0; round < MAX_TOOL_ROUNDS; round += 1) {
+		const result = await streamOnce(tools, extraContext);
+		if (result.toolCalls.length === 0) return {
+			content: result.content,
+			truncated: result.truncated
+		};
+		truncated = result.truncated ?? truncated;
+		const executed = [];
+		for (const call of result.toolCalls) {
+			const argsPreview = call.argumentsJson.length > 120 ? `${call.argumentsJson.slice(0, 120)}…` : call.argumentsJson;
+			onToolStep({
+				kind: "call",
+				name: call.name,
+				text: argsPreview || "无参数"
+			});
+			if (!READONLY_TOOL_NAMES.has(call.name)) console.warn(`[dsh-advisor-group] 顾问调用非只读工具（${call.name}）：`, argsPreview);
+			const { outcome, reused } = await executeDeduped(call);
+			if (outcome.trim()) {
+				const flat = outcome.replace(/\n+/g, " ").trim();
+				const preview = flat.length > 240 ? `${flat.slice(0, 240)}…` : flat;
+				onToolStep({
+					kind: "result",
+					name: call.name,
+					text: `${reused ? "（复用已执行结果）" : ""}${preview}`
+				});
+			}
+			executed.push(`- ${call.name}${call.argumentsJson ? ` 参数：${call.argumentsJson.slice(0, 200)}` : ""}\n结果：${capOutcome(outcome)}`);
+		}
+		extraContext = `${extraContext}\n\n【顾问已执行工具】\n${executed.join("\n\n")}`;
+	}
+	const final = await streamOnce([], extraContext);
+	if (final.toolCalls.length > 0) {
+		const executed = [];
+		for (const call of final.toolCalls) {
+			const argsPreview = call.argumentsJson.length > 120 ? `${call.argumentsJson.slice(0, 120)}…` : call.argumentsJson;
+			onToolStep({
+				kind: "call",
+				name: call.name,
+				text: argsPreview || "无参数"
+			});
+			if (!READONLY_TOOL_NAMES.has(call.name)) console.warn(`[dsh-advisor-group] 顾问调用非只读工具（${call.name}）：`, argsPreview);
+			const { outcome, reused } = await executeDeduped(call);
+			if (outcome.trim()) {
+				const flat = outcome.replace(/\n+/g, " ").trim();
+				const preview = flat.length > 240 ? `${flat.slice(0, 240)}…` : flat;
+				onToolStep({
+					kind: "result",
+					name: call.name,
+					text: `${reused ? "（复用已执行结果）" : ""}${preview}`
+				});
+			}
+			executed.push(`- ${call.name}${call.argumentsJson ? ` 参数：${call.argumentsJson.slice(0, 200)}` : ""}\n结果：${capOutcome(outcome)}`);
+		}
+		const final2 = await streamOnce([], `${extraContext}\n\n【顾问已执行工具】\n${executed.join("\n\n")}`);
+		return {
+			content: final2.content,
+			truncated: final2.truncated ?? final.truncated ?? truncated
+		};
+	}
+	return {
+		content: final.content,
+		truncated: final.truncated ?? truncated
+	};
+}
+//#endregion
 //#region src/service.ts
 function isAbortError(error, signal) {
 	if (signal?.aborted) return true;
 	return error instanceof Error && error.name === "AbortError";
 }
-const MAX_DAILY_CONSULTATIONS = 50;
+const UUID_RE = /^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/i;
+/** Real consult ids are server-minted uuids; anything else is rejected. */
+function isSafeConsultId(id) {
+	return UUID_RE.test(id);
+}
+/** Agent DSH session ids look like `session-<uuid>` or a plain uuid. */
+function isSafeDshSessionId(id) {
+	return UUID_RE.test(id) || UUID_RE.test(id.replace(/^session-/i, ""));
+}
 function extractUserMessageText(event) {
 	if (event.type !== "user/message") return null;
 	const data = event.data;
@@ -1021,6 +1789,11 @@ var AdvisorGroupService = class {
 	/** Durable daily-guard counter fixture under the DSH home storage area. */
 	dailyGuardPath;
 	dailyWriteChain = Promise.resolve();
+	/** Per-session stop controllers for the auto-deepen pipeline. */
+	activeAborts = /* @__PURE__ */ new Map();
+	/** Durable consultation snapshots (usable after a dsh restart). */
+	sessionsDir;
+	sessionWriteChain = Promise.resolve();
 	enabled;
 	config;
 	constructor(ctx, config) {
@@ -1029,7 +1802,9 @@ var AdvisorGroupService = class {
 		this.enabled = config.enabled;
 		const dshHome = process.env.DSH_HOME || join(homedir(), ".dsh");
 		this.dailyGuardPath = join(dshHome, "storages", "advisor-group", "daily-guard.json");
+		this.sessionsDir = join(dshHome, "storages", "advisor-group", "sessions");
 		this.loadDailyGuard();
+		this.loadStoredSessions();
 		ctx.on("session/event", (session, event) => {
 			const text = extractUserMessageText(event);
 			if (!text) return;
@@ -1065,19 +1840,241 @@ var AdvisorGroupService = class {
 			console.warn("[dsh-advisor-group] 每日护栏持久化失败：", error instanceof Error ? error.message : String(error));
 		});
 	}
-	getRepeatPressure(sessionId) {
-		if (!sessionId) return void 0;
-		const tracked = this.repeatTracker.get(sessionId);
-		if (!tracked) return void 0;
+	/**
+	* Restore interrupted consultations from durable snapshots so the card's
+	* 「▶ 继续聊天」 button still works after a dsh restart. Completed sessions
+	* are not restored (they cannot resume) and their snapshot is cleaned up.
+	*/
+	loadStoredSessions() {
+		let files;
+		try {
+			files = readdirSync(this.sessionsDir).filter((name) => name.endsWith(".json"));
+		} catch {
+			return;
+		}
+		for (const file of files) try {
+			const raw = JSON.parse(readFileSync(join(this.sessionsDir, file), "utf8"));
+			if (raw.version !== 1 || typeof raw.id !== "string" || typeof raw.question !== "string") continue;
+			if (!isSafeConsultId(raw.id)) continue;
+			if (raw.dshSessionId !== void 0 && !isSafeDshSessionId(raw.dshSessionId)) continue;
+			if (raw.status === "completed") {
+				unlinkSync(join(this.sessionsDir, file));
+				continue;
+			}
+			const advisors = this.resolveAdvisors(raw.advisorIds);
+			if (advisors.length === 0) continue;
+			const session = {
+				id: raw.id,
+				status: "cancelled",
+				question: raw.question,
+				context: raw.context,
+				advisors,
+				maxRounds: raw.maxRounds,
+				messages: raw.messages.map((message) => ({
+					role: message.role,
+					advisorId: message.advisorId,
+					advisorName: message.advisorName,
+					content: message.content,
+					thinking: message.thinking,
+					thinkingSegments: message.thinkingSegments,
+					actionDescriptions: message.actionDescriptions,
+					round: message.round,
+					truncated: message.truncated,
+					toolSteps: message.toolSteps,
+					ts: message.ts
+				})),
+				cwd: raw.cwd,
+				dshSessionId: raw.dshSessionId,
+				driverSource: raw.driverSource,
+				stopReason: raw.stopReason ?? "abort-error",
+				createdAt: raw.createdAt,
+				updatedAt: raw.updatedAt
+			};
+			this.sessions.set(session.id, session);
+		} catch (error) {
+			console.warn("[dsh-advisor-group] 会话快照读取失败，已跳过：", file, error instanceof Error ? error.message : String(error));
+		}
+	}
+	/** Serialization-safe snapshot: advisor identity only, NEVER credentials. */
+	snapshotPayload(session) {
 		return {
-			count: tracked.count,
-			text: tracked.text
+			version: 1,
+			id: session.id,
+			status: session.status,
+			question: session.question,
+			...session.context === void 0 ? {} : { context: session.context },
+			...session.cwd === void 0 ? {} : { cwd: session.cwd },
+			...session.dshSessionId === void 0 ? {} : { dshSessionId: session.dshSessionId },
+			...session.driverSource === void 0 ? {} : { driverSource: session.driverSource },
+			...session.stopReason === void 0 ? {} : { stopReason: session.stopReason },
+			advisorIds: session.advisors.map((advisor) => advisor.id),
+			maxRounds: session.maxRounds,
+			createdAt: session.createdAt,
+			updatedAt: session.updatedAt,
+			messages: session.messages.map((message) => ({
+				role: message.role,
+				...message.advisorId === void 0 ? {} : { advisorId: message.advisorId },
+				...message.advisorName === void 0 ? {} : { advisorName: message.advisorName },
+				content: message.content,
+				...message.thinking === void 0 ? {} : { thinking: message.thinking },
+				...message.thinkingSegments === void 0 ? {} : { thinkingSegments: message.thinkingSegments },
+				...message.actionDescriptions === void 0 ? {} : { actionDescriptions: message.actionDescriptions },
+				...message.round === void 0 ? {} : { round: message.round },
+				...message.truncated === void 0 ? {} : { truncated: message.truncated },
+				...message.toolSteps === void 0 ? {} : { toolSteps: message.toolSteps },
+				ts: message.ts
+			}))
 		};
 	}
-	getConfig() {
-		return this.config;
-	}
-	setConfig(config) {
+	/** Serialized, atomic (tmp + rename) consultation snapshot persistence. */
+	persistSession(session) {
+		if (!isSafeConsultId(session.id)) {
+			console.warn("[dsh-advisor-group] 拒绝持久化非法会话 id：", session.id);
+			return;
+		}
+		const payload = JSON.stringify(this.snapshotPayload(session));
+		const target = join(this.sessionsDir, `${session.id}.json`);
+		this.sessionWriteChain = this.sessionWriteChain.then(async () => {
+			await promises.mkdir(this.sessionsDir, { recursive: true });
+			const tmp = `${target}.tmp`;
+			await promises.writeFile(tmp, payload, "utf8");
+			await promises.rename(tmp, target);
+		}).catch((error) => {
+			console.warn("[dsh-advisor-group] 咨询快照持久化失败：", session.id, error instanceof Error ? error.message : String(error));
+		});
+	}
+	/**
+	* Card-facing advisor failure text: credentials redacted, collapsed to one
+	* line, length-capped. The full raw error is logged server-side instead.
+	*/
+	summarizeAdvisorError(raw) {
+		const singleLine = raw.replace(/\b(api[_-]?key["']?\s*[:=]\s*["']?)[^\s"',}&]+/gi, "$1[已隐藏]").replace(/([?&]key=)[^\s&"']+/gi, "$1[已隐藏]").replace(/\b(Bearer|Basic)\s+[A-Za-z0-9._~+/=-]+/gi, "$1 [已隐藏]").replace(/\bsk-[A-Za-z0-9_-]{8,}/g, "[已隐藏]").replace(/\s+/g, " ").trim();
+		if (singleLine.length <= 300) return singleLine;
+		return `${singleLine.slice(0, 300)}…（完整错误已记录到服务端日志）`;
+	}
+	/** Wait for queued durable writes before a host or test tears down storage. */
+	async flushPersistence() {
+		await Promise.all([this.dailyWriteChain, this.sessionWriteChain]);
+	}
+	/** Structural view of the DSH in-memory session store (may be absent). */
+	get sessionStore() {
+		let store;
+		try {
+			store = this.ctx.sessions;
+		} catch {
+			return;
+		}
+		if (!store || typeof store !== "object") return void 0;
+		const typed = store;
+		if (typeof typed.get !== "function" || typeof typed.prepare !== "function" || typeof typed.enter !== "function" || typeof typed.flush !== "function") return;
+		return store;
+	}
+	/** Best-effort durability barrier for the advisor-group log events. */
+	async flushLog(sessionLog) {
+		if (!sessionLog) return;
+		const store = this.sessionStore;
+		if (!store) return;
+		try {
+			await store.flush(sessionLog);
+		} catch {}
+	}
+	/** Build a contiguous seed log from the snapshot (cross-restart resume). */
+	buildSeedEvents(session) {
+		const events = [];
+		let seq = 0;
+		events.push({
+			type: "advisor-group/start",
+			seq: seq++,
+			time: session.createdAt,
+			data: {
+				sessionId: session.id,
+				turn: 0,
+				step: 0,
+				question: session.question,
+				...session.context === void 0 ? {} : { context: session.context },
+				advisors: session.advisors.map(({ id, name, avatar }) => avatar === void 0 ? {
+					id,
+					name
+				} : {
+					id,
+					name,
+					avatar
+				})
+			}
+		});
+		for (const message of session.messages) {
+			if (message.role === "system") continue;
+			events.push({
+				type: "advisor-group/message",
+				seq: seq++,
+				time: message.ts || Date.now(),
+				data: {
+					sessionId: session.id,
+					turn: 0,
+					step: 0,
+					role: message.role,
+					content: message.content,
+					...message.advisorId === void 0 ? {} : { advisorId: message.advisorId },
+					...message.advisorName === void 0 ? {} : { advisorName: message.advisorName },
+					...message.thinking === void 0 ? {} : { thinking: message.thinking },
+					...message.round === void 0 ? {} : { round: message.round },
+					...message.truncated === void 0 ? {} : { truncated: message.truncated }
+				}
+			});
+		}
+		return events;
+	}
+	/**
+	* Obtain the DSH Session handle for durable log events. Prefers the live
+	* store entry (same process). After a restart the store is empty, so the
+	* session is rebuilt with `prepare` + `enter` (append hooks publish events
+	* and the persistence layer picks them up; the session is NOT announced, so
+	* no agent lifecycle reactivation is triggered). Falls back to a detached
+	* `Session.create` (live SSE + snapshot only, no durable log) when neither
+	* works.
+	*/
+	resolveSessionHandle(session) {
+		const store = this.sessionStore;
+		const resumeId = session.dshSessionId !== void 0 && isSafeDshSessionId(session.dshSessionId) ? session.dshSessionId : isSafeConsultId(session.id) ? session.id : "";
+		if (!resumeId) return void 0;
+		const live = store?.get(resumeId);
+		if (live) return live;
+		if (!store) return this.buildDetachedSession(session, resumeId);
+		try {
+			const meta = { cwd: session.cwd !== void 0 && isAbsolute(session.cwd) ? session.cwd : process.cwd() };
+			const prepared = store.prepare(resumeId, {
+				seed: this.buildSeedEvents(session),
+				meta
+			});
+			store.enter(prepared);
+			return prepared;
+		} catch (error) {
+			console.warn("[dsh-advisor-group] 重建会话句柄失败（降级为 detached）：", session.id, error instanceof Error ? error.message : String(error));
+			const again = store.get(resumeId);
+			if (again) return again;
+			return this.buildDetachedSession(session, resumeId);
+		}
+	}
+	buildDetachedSession(session, sessionId = session.id) {
+		try {
+			return Session.create(sessionId, this.buildSeedEvents(session), void 0, 0);
+		} catch {
+			return;
+		}
+	}
+	getRepeatPressure(sessionId) {
+		if (!sessionId) return void 0;
+		const tracked = this.repeatTracker.get(sessionId);
+		if (!tracked) return void 0;
+		return {
+			count: tracked.count,
+			text: tracked.text
+		};
+	}
+	getConfig() {
+		return this.config;
+	}
+	setConfig(config) {
 		this.config = config;
 		this.enabled = config.enabled;
 	}
@@ -1098,26 +2095,35 @@ var AdvisorGroupService = class {
 		this.setEnabled(enabled);
 		if (this.persistEnabledCallback) try {
 			await this.persistEnabledCallback(enabled);
-		} catch {}
+		} catch (error) {
+			console.warn("[dsh-advisor-group] 启用开关保存失败（内存态未回滚）：", error instanceof Error ? error.message : String(error));
+		}
 		return this.enabled;
 	}
 	/**
 	* Current daily-guard facts for display (does NOT increment — read-only).
 	* Rolls the in-memory counter to a new UTC day view without persisting.
+	* `enabled` reflects the configurable cap; `remaining` is -1 when the cap
+	* is disabled (the counter still accumulates for visibility).
 	*/
 	getDailyGuard() {
 		const today = (/* @__PURE__ */ new Date()).toISOString().slice(0, 10);
 		const used = this.dailyConsultationDate === today ? this.dailyConsultationCount : 0;
+		const enabled = this.config.quota.enabled;
+		const limit = this.config.quota.maxPerDay;
 		return {
 			used,
-			limit: MAX_DAILY_CONSULTATIONS,
-			remaining: Math.max(0, MAX_DAILY_CONSULTATIONS - used)
+			enabled,
+			limit,
+			remaining: enabled ? Math.max(0, limit - used) : -1
 		};
 	}
 	/**
 	* Minimal cost guard: atomically check and increment the per-day new
 	* consultation counter. Node is single-threaded, so doing both inside one
 	* synchronous block closes the TOCTOU window a two-step check would leave.
+	* The cap is configurable (`quota.enabled` / `quota.maxPerDay`); when the
+	* cap is disabled the counter still increments (used for display only).
 	* The counter is persisted (UTC day key) so a harness restart does not
 	* reset the quota; persistence is fire-and-forget after the sync block.
 	*/
@@ -1127,9 +2133,9 @@ var AdvisorGroupService = class {
 			this.dailyConsultationDate = today;
 			this.dailyConsultationCount = 0;
 		}
-		if (this.dailyConsultationCount >= MAX_DAILY_CONSULTATIONS) return {
+		if (this.config.quota.enabled && this.dailyConsultationCount >= this.config.quota.maxPerDay) return {
 			ok: false,
-			reason: "今日顾问咨询次数已达上限，请明天再试或调整配置。"
+			reason: "今日顾问咨询次数已达上限，请明天再试，或在设置中调整每日咨询上限。"
 		};
 		this.dailyConsultationCount += 1;
 		this.persistDailyGuard();
@@ -1138,9 +2144,10 @@ var AdvisorGroupService = class {
 	getSession(id) {
 		return this.sessions.get(id);
 	}
-	createSession(question, context, advisorIds = []) {
+	createSession(question, context, advisorIds = [], cwd, dshSessionId) {
 		const maxAdvisors = this.config.discussion.maxAdvisorsPerCall;
 		const advisors = this.resolveAdvisors(advisorIds).slice(0, maxAdvisors);
+		const now = Date.now();
 		const session = {
 			id: randomUUID(),
 			status: "active",
@@ -1151,20 +2158,23 @@ var AdvisorGroupService = class {
 			messages: [{
 				role: "system",
 				content: `顾问群已建立，共 ${advisors.length} 位顾问。`,
-				ts: Date.now()
+				ts: now
 			}, {
 				role: "main",
 				content: context ? `${context}\n\n${question}` : question,
-				ts: Date.now()
+				ts: now
 			}],
-			createdAt: Date.now(),
-			updatedAt: Date.now()
+			cwd,
+			dshSessionId,
+			createdAt: now,
+			updatedAt: now
 		};
 		if (this.sessions.size >= 100) {
 			const oldest = this.sessions.keys().next().value;
 			if (oldest !== void 0) this.sessions.delete(oldest);
 		}
 		this.sessions.set(session.id, session);
+		this.persistSession(session);
 		this.ctx.emit("advisor-group/session-start", session);
 		return session;
 	}
@@ -1179,22 +2189,60 @@ var AdvisorGroupService = class {
 		session.messages.push(message);
 		session.updatedAt = Date.now();
 		if (sessionLog) appendAdvisorMessage(sessionLog, session.id, message);
+		this.persistSession(session);
 		this.ctx.emit("advisor-group/message", {
 			sessionId: session.id,
 			message
 		});
 		return session;
 	}
-	async runOneRound(session, signal, sessionLog) {
+	/**
+	* The next round that still needs work, derived from the durable messages:
+	* when the last round is incomplete (an advisor answered, a later one was
+	* cut by a stop), that round resumes with ONLY the missing advisors; when
+	* it is complete, the next round is fresh. `undefined` means finished.
+	*/
+	nextPendingRound(session) {
+		const advisorMessages = session.messages.filter((message) => message.role === "advisor");
+		const maxRound = advisorMessages.reduce((max, message) => Math.max(max, message.round ?? 1), 0);
 		const advisorCount = Math.max(1, session.advisors.length);
-		const round = Math.floor(session.messages.filter((message) => message.role === "advisor").length / advisorCount) + 1;
+		if (maxRound === 0) return {
+			round: 1,
+			answeredIds: []
+		};
+		if (maxRound > session.maxRounds) return void 0;
+		const answeredInMax = advisorMessages.filter((message) => (message.round ?? 1) === maxRound).map((message) => message.advisorId ?? "").filter((id) => id !== "");
+		const unique = [...new Set(answeredInMax)];
+		if (unique.length < advisorCount) return {
+			round: maxRound,
+			answeredIds: unique
+		};
+		if (maxRound + 1 > session.maxRounds) return void 0;
+		return {
+			round: maxRound + 1,
+			answeredIds: []
+		};
+	}
+	/** Number of main-model messages: the opening question + all deep-questions. */
+	mainMessageCount(session) {
+		return session.messages.filter((message) => message.role === "main").length;
+	}
+	/** Run exactly one round: every advisor that has NOT answered this round. */
+	async runRoundFrom(session, round, answeredIds, signal, sessionLog, agent) {
 		for (let index = 0; index < session.advisors.length; index++) {
 			const advisor = session.advisors[index];
+			if (answeredIds.includes(advisor.id)) continue;
 			const transcript = this.buildTranscript(session);
-			await this.callAdvisor(session, advisor, transcript, round, index + 1, signal, sessionLog);
+			await this.callAdvisor(session, advisor, transcript, round, index + 1, signal, sessionLog, agent);
 		}
 		session.status = "active";
 		session.updatedAt = Date.now();
+		this.persistSession(session);
+	}
+	async runOneRound(session, signal, sessionLog, agent) {
+		const pending = this.nextPendingRound(session);
+		if (!pending) return;
+		await this.runRoundFrom(session, pending.round, pending.answeredIds, signal, sessionLog, agent);
 	}
 	/**
 	* Auto-deepen pipeline (2026-09-05): one ask_advisors call runs the whole
@@ -1202,21 +2250,159 @@ var AdvisorGroupService = class {
 	*   [driver deep-question (after the first)] → advisor A → advisor B (sees A)
 	*   → advisor C (sees A+B) → …
 	* and closes with a driver-generated conclusion. The pipeline is synchronous
-	* (SSE keeps the card streaming); `exec.signal` aborts the whole run.
+	* (SSE keeps the card streaming); `exec.signal` aborts the whole run and the
+	* user may stop it explicitly via `stopConsultation(sessionId)` (POST
+	* /advisor-group/stop) — a stop degrades to a graceful partial summary.
+	*
+	* Resumption: after a stop (or a dsh restart that restored the snapshot)
+	* `resumeConsultation()` re-enters the SAME pipeline; `nextPendingRound()`
+	* detects the interruption point — a partially-answered round is completed
+	* with only the missing advisors (the deep-question is NOT regenerated for
+	* that round), then remaining rounds run as usual.
 	*/
-	async runAutoPipeline(session, signal, sessionLog) {
-		while (!this.hasReachedMaxRounds(session)) {
-			if (this.getRoundCount(session) + 1 > 1 && this.config.discussion.autoDeepen) {
-				const question = await generateDeepenQuestion(this.ctx, session, resolveDriverSource(sessionLog, this.config.discussion.driverModel), signal);
-				this.appendMainMessage(session, question, sessionLog);
+	async runAutoPipeline(session, signal, sessionLog, agent) {
+		const stop = new AbortController();
+		this.activeAborts.set(session.id, stop);
+		const combined = signal ? AbortSignal.any([signal, stop.signal]) : stop.signal;
+		if (!session.driverSource) {
+			const source = resolveDriverSource(sessionLog, this.config.discussion.driverModel);
+			if (source) {
+				session.driverSource = source;
+				this.persistSession(session);
+			}
+		}
+		try {
+			while (true) {
+				const pending = this.nextPendingRound(session);
+				if (!pending) break;
+				const freshRound = pending.answeredIds.length === 0;
+				const questionAlreadyAsked = this.mainMessageCount(session) >= pending.round;
+				if (pending.round > 1 && freshRound && this.config.discussion.autoDeepen && !questionAlreadyAsked) {
+					const question = await generateDeepenQuestion(this.ctx, session, session.driverSource ?? resolveDriverSource(sessionLog, this.config.discussion.driverModel), combined, this.config.discussion.driverTimeoutMs);
+					this.appendMainMessage(session, question, sessionLog);
+				}
+				await this.runRoundFrom(session, pending.round, pending.answeredIds, combined, sessionLog, agent);
+			}
+			session.status = "completed";
+			session.stopReason = void 0;
+			const summary = await this.generateFinalSummary(session, combined, sessionLog, false);
+			this.persistSession(session);
+			await this.flushLog(sessionLog);
+			return summary;
+		} catch (error) {
+			if (isAbortError(error, combined) || stop.signal.aborted) {
+				if (stop.signal.aborted) session.stopReason = "user-stop";
+				else if (signal?.aborted) session.stopReason = "exec-cancel";
+				else session.stopReason = "abort-error";
+				session.status = "cancelled";
+				this.persistSession(session);
+				const summary = await this.generateFinalSummary(session, void 0, sessionLog, true);
+				await this.flushLog(sessionLog);
+				return summary;
 			}
-			await this.runOneRound(session, signal, sessionLog);
+			throw error;
+		} finally {
+			this.activeAborts.delete(session.id);
+		}
+	}
+	/**
+	* User-initiated stop: aborts the in-flight consultation pipeline. Returns
+	* whether a running consultation was found and aborted.
+	*/
+	stopConsultation(sessionId) {
+		const stop = this.activeAborts.get(sessionId);
+		if (!stop) return false;
+		stop.abort();
+		return true;
+	}
+	/** True while a pipeline run (auto-deepen or a follow-up round) is in flight. */
+	isSessionRunning(sessionId) {
+		return this.activeAborts.has(sessionId);
+	}
+	/**
+	* Resume a stopped/interrupted consultation from its interruption point.
+	* Works for a user stop (live session) and after a dsh restart (snapshot
+	* restored; the DSH session handle is rebuilt via `prepare` + `enter`).
+	* Returns immediately; the pipeline runs in the background and streams over
+	* the existing SSE channel, then closes with a fresh `advisor-group/end`.
+	*/
+	resumeConsultation(sessionId) {
+		const session = this.sessions.get(sessionId);
+		if (!session) return {
+			ok: false,
+			reason: "咨询会话不存在或已过期，请重新发起咨询。"
+		};
+		if (this.activeAborts.has(sessionId)) return {
+			ok: false,
+			reason: "该咨询正在运行中，无需继续。"
+		};
+		if (session.status === "completed" || this.nextPendingRound(session) === void 0) return {
+			ok: false,
+			reason: "该咨询已完成，无需继续。"
+		};
+		if (session.advisors.length === 0) return {
+			ok: false,
+			reason: "恢复失败：当前配置无法匹配该咨询的原顾问（顾问配置可能已被修改）。"
+		};
+		if (!this.acquireResumeLock(session.id)) return {
+			ok: false,
+			reason: "该咨询正在另一实例中恢复，请稍后再试（若确认无其他实例运行，10 分钟后会自动接管该锁）。"
+		};
+		const dshSession = this.resolveSessionHandle(session);
+		const agentRegistry = this.ctx.agents;
+		const resumeAgent = session.dshSessionId !== void 0 && agentRegistry?.get ? agentRegistry.get(session.dshSessionId) : void 0;
+		session.status = "active";
+		session.stopReason = void 0;
+		session.updatedAt = Date.now();
+		this.persistSession(session);
+		if (dshSession) appendAdvisorResume(dshSession, session.id);
+		this.ctx.emit("advisor-group/resume", { sessionId: session.id });
+		this.runAutoPipeline(session, void 0, dshSession, resumeAgent).catch((error) => {
+			console.error("[dsh-advisor-group] 续跑流水线失败：", session.id, error instanceof Error ? error.message : String(error));
+		}).finally(() => this.releaseResumeLock(session.id));
+		return { ok: true };
+	}
+	/** Atomic cross-instance lock (mkdir-based, with stale takeover). */
+	acquireResumeLock(sessionId) {
+		const lockDir = join(this.sessionsDir, `${sessionId}.lock`);
+		try {
+			mkdirSync(this.sessionsDir, { recursive: true });
+			mkdirSync(lockDir);
+			try {
+				writeFileSync(join(lockDir, "owner"), String(process.pid), "utf8");
+			} catch {}
+			return true;
+		} catch {
+			try {
+				const stat = statSync(lockDir);
+				if (Date.now() - stat.mtimeMs > 6e5) {
+					rmSync(lockDir, {
+						recursive: true,
+						force: true
+					});
+					mkdirSync(lockDir);
+					return true;
+				}
+			} catch {}
+			return false;
 		}
-		session.status = "completed";
+	}
+	releaseResumeLock(sessionId) {
+		try {
+			rmSync(join(this.sessionsDir, `${sessionId}.lock`), {
+				recursive: true,
+				force: true
+			});
+		} catch {}
+	}
+	async generateFinalSummary(session, signal, sessionLog, stopped) {
 		session.updatedAt = Date.now();
-		const summary = this.buildSummary(session, false);
-		const conclusion = await generateConclusion(this.ctx, session, resolveDriverSource(sessionLog, this.config.discussion.driverModel), signal);
-		const finalSummary = {
+		const summary = this.buildSummary(session, stopped);
+		const conclusion = stopped ? "" : await generateConclusion(this.ctx, session, session.driverSource ?? resolveDriverSource(sessionLog, this.config.discussion.driverModel), signal, this.config.discussion.driverTimeoutMs);
+		const finalSummary = stopped ? {
+			...summary,
+			stopped: true
+		} : {
 			...summary,
 			conclusion
 		};
@@ -1235,6 +2421,7 @@ var AdvisorGroupService = class {
 		};
 		session.messages.push(message);
 		if (sessionLog) appendAdvisorMessage(sessionLog, session.id, message);
+		this.persistSession(session);
 		this.ctx.emit("advisor-group/message", {
 			sessionId: session.id,
 			message
@@ -1246,17 +2433,38 @@ var AdvisorGroupService = class {
 	* calling ask_advisors again with sessionId + followUp, never by the
 	* advisor models talking to themselves back-to-back.
 	*/
-	async runOneRoundAndSummarize(session, signal, sessionLog) {
-		await this.runOneRound(session, signal, sessionLog);
-		session.status = "completed";
-		session.updatedAt = Date.now();
-		const summary = this.buildSummary(session, false);
-		if (sessionLog) appendAdvisorEnd(sessionLog, session, summary);
-		this.ctx.emit("advisor-group/session-end", {
-			sessionId: session.id,
-			summary
-		});
-		return summary;
+	async runOneRoundAndSummarize(session, signal, sessionLog, agent) {
+		const stop = new AbortController();
+		this.activeAborts.set(session.id, stop);
+		const combined = signal ? AbortSignal.any([signal, stop.signal]) : stop.signal;
+		try {
+			await this.runOneRound(session, combined, sessionLog, agent);
+			session.status = "completed";
+			session.updatedAt = Date.now();
+			const summary = this.buildSummary(session, false);
+			if (sessionLog) appendAdvisorEnd(sessionLog, session, summary);
+			this.persistSession(session);
+			await this.flushLog(sessionLog);
+			this.ctx.emit("advisor-group/session-end", {
+				sessionId: session.id,
+				summary
+			});
+			return summary;
+		} catch (error) {
+			if (isAbortError(error, combined) || stop.signal.aborted) {
+				if (stop.signal.aborted) session.stopReason = "user-stop";
+				else if (signal?.aborted) session.stopReason = "exec-cancel";
+				else session.stopReason = "abort-error";
+				session.status = "cancelled";
+				this.persistSession(session);
+				const summary = await this.generateFinalSummary(session, void 0, sessionLog, true);
+				await this.flushLog(sessionLog);
+				return summary;
+			}
+			throw error;
+		} finally {
+			this.activeAborts.delete(session.id);
+		}
 	}
 	getRoundCount(session) {
 		const advisorCount = Math.max(1, session.advisors.length);
@@ -1271,21 +2479,21 @@ var AdvisorGroupService = class {
 		return this.config.advisors.filter((advisor) => advisorIds.includes(advisor.id));
 	}
 	buildTranscript(session) {
-		return session.messages.filter((message) => message.role !== "system").slice(-20).map((message) => {
+		return session.messages.filter((message) => message.role !== "system").slice(-12).map((message) => {
 			if (message.role === "main") return {
 				role: "main",
 				name: "主模型",
-				content: message.content.slice(0, 4e3)
+				content: message.content.slice(0, 3e3)
 			};
 			if (message.role === "advisor" && message.advisorName) return {
 				role: "advisor",
 				name: message.advisorName,
-				content: message.content.slice(0, 4e3)
+				content: message.content.slice(0, 3e3)
 			};
 			return null;
 		}).filter((entry) => entry !== null);
 	}
-	async callAdvisor(session, advisor, transcript, round, joinIndex, signal, sessionLog) {
+	async callAdvisor(session, advisor, transcript, round, joinIndex, signal, sessionLog, agent) {
 		const relational = { ...advisor };
 		relational.systemPrompt = `${advisor.systemPrompt}\n${advisorJoinPrompt(joinIndex, session.advisors.length)}`;
 		const message = {
@@ -1323,18 +2531,21 @@ var AdvisorGroupService = class {
 				pendingText = "";
 				pendingThinking = "";
 			};
-			const emitDelta = (delta) => {
+			const emitDelta = (delta, phase = "answer") => {
 				if (delta.text) {
 					message.content += delta.text;
 					pendingText += delta.text;
 				}
 				if (delta.thinking) {
 					message.thinking = (message.thinking ?? "") + delta.thinking;
+					thinkingBuffer += delta.thinking;
 					pendingThinking += delta.thinking;
 				}
 				publish(session.id, {
 					advisorId: advisor.id,
 					advisorName: advisor.name,
+					round,
+					phase,
 					contentDelta: delta.text,
 					thinkingDelta: delta.thinking,
 					done: false
@@ -1342,28 +2553,105 @@ var AdvisorGroupService = class {
 				if (!timer) timer = setTimeout(flush, 200);
 			};
 			const advisorTimeoutMs = this.config.discussion.advisorTimeoutMs;
+			let truncated;
+			const thinkingSegments = [];
+			let thinkingBuffer = "";
+			const actionDescriptions = [];
 			if (!isDshLlmProvider || advisor.baseURL || advisor.apiKey || advisor.apiKeyEnv) {
-				const streamResult = await streamDirectHttp(relational, transcript, emitDelta, signal, advisorTimeoutMs);
-				flush();
-				content = streamResult.content;
-				message.thinking = streamResult.thinking;
+				const toolMode = advisor.tools ?? this.config.discussion.advisorTools ?? "readonly";
+				const toolSchemas = resolveAdvisorToolSchemas(this.ctx, agent, toolMode);
+				if (toolSchemas.length > 0) {
+					relational.systemPrompt = `${relational.systemPrompt}${ADVISOR_TOOL_GUIDANCE}`;
+					const messageToolSteps = [];
+					let narrationBuffer = "";
+					const emitToolStep = (step) => {
+						const full = {
+							...step,
+							atMs: Date.now()
+						};
+						const parts = [];
+						if (narrationBuffer.trim()) {
+							parts.push(narrationBuffer.trim());
+							narrationBuffer = "";
+						}
+						if (thinkingBuffer.trim()) parts.push(thinkingBuffer);
+						if (parts.length > 0) {
+							thinkingSegments.push(parts.join("\n\n"));
+							thinkingBuffer = "";
+						}
+						messageToolSteps.push(full);
+						if (sessionLog) appendAdvisorDelta(sessionLog, session.id, {
+							sessionId: session.id,
+							advisorId: advisor.id,
+							advisorName: advisor.name,
+							round,
+							toolStep: full
+						});
+						publish(session.id, {
+							advisorId: advisor.id,
+							advisorName: advisor.name,
+							round,
+							toolStep: full
+						});
+					};
+					const streamOnce = (tools, extra) => {
+						const withExtra = extra.trim() ? [...transcript, {
+							role: "main",
+							name: "已执行工具",
+							content: extra.trim()
+						}] : transcript;
+						const parseNames = toolSchemas.map((tool) => tool.name);
+						const emitRoundDelta = (delta) => emitDelta(delta, tools.length > 0 ? "tool" : "answer");
+						return streamDirectHttp(relational, withExtra, emitRoundDelta, signal, advisorTimeoutMs, tools, parseNames).then((result) => {
+							narrationBuffer = result.content.trim();
+							if (tools.length > 0 && narrationBuffer) actionDescriptions.push(narrationBuffer);
+							return {
+								content: result.content,
+								thinking: result.thinking,
+								toolCalls: result.toolCalls ?? [],
+								truncated: result.truncated
+							};
+						});
+					};
+					const loop = await runAdvisorToolLoop(toolSchemas, streamOnce, (call) => executeAdvisorTool(this.ctx, agent, call, signal), (step) => emitToolStep(step));
+					flush();
+					content = loop.content;
+					truncated = loop.truncated;
+					message.toolSteps = messageToolSteps;
+				} else {
+					const streamResult = await streamDirectHttp(relational, transcript, emitDelta, signal, advisorTimeoutMs);
+					flush();
+					content = streamResult.content;
+					message.thinking = streamResult.thinking;
+					truncated = streamResult.truncated;
+				}
 			} else {
-				content = await callViaCtxLlm(this.ctx, relational, transcript, signal, emitDelta, advisorTimeoutMs);
+				const result = await callViaCtxLlm(this.ctx, relational, transcript, signal, emitDelta, advisorTimeoutMs);
 				flush();
+				content = result.content;
+				truncated = result.truncated;
 			}
-			message.content = content;
+			message.content = sanitizeAdvisorContent(content);
+			if (truncated) message.truncated = truncated;
+			if (thinkingBuffer.trim()) thinkingSegments.push(thinkingBuffer);
+			if (thinkingSegments.length > 0) message.thinkingSegments = thinkingSegments;
+			if (actionDescriptions.length > 0) message.actionDescriptions = actionDescriptions;
 		} catch (error) {
 			if (isAbortError(error, signal)) throw error;
-			message.content = `（顾问调用失败：${error instanceof Error ? error.message : String(error)}）`;
+			const rawMessage = error instanceof Error ? error.message : String(error);
+			console.warn("[dsh-advisor-group] 顾问调用失败：", session.id, advisor.name, rawMessage);
+			message.content = `（顾问调用失败：${this.summarizeAdvisorError(rawMessage)}）`;
 		}
 		publish(session.id, {
 			advisorId: advisor.id,
 			advisorName: advisor.name,
+			round,
 			done: true
 		});
 		session.messages.push(message);
 		if (sessionLog) appendAdvisorMessage(sessionLog, session.id, message);
 		session.updatedAt = Date.now();
+		this.persistSession(session);
 		this.ctx.emit("advisor-group/message", {
 			sessionId: session.id,
 			message
@@ -1379,7 +2667,7 @@ var AdvisorGroupService = class {
 		}
 		const advisors = [...byAdvisor.entries()].map(([id, messages]) => {
 			const advisor = session.advisors.find((item) => item.id === id);
-			const meaningful = messages.filter((message) => message.content && message.content.trim() && !message.content.startsWith("（顾问调用失败"));
+			const meaningful = messages.filter((message) => message.content && message.content.trim() && !message.content.startsWith("（顾问调用失败") && !message.content.startsWith("（顾问未生成正文"));
 			const last = meaningful[meaningful.length - 1] ?? messages[messages.length - 1];
 			return {
 				advisorId: id,
@@ -1387,8 +2675,7 @@ var AdvisorGroupService = class {
 				corePoints: last ? this.extractCorePoints(last.content) : []
 			};
 		});
-		const riskNotes = advisorMessages.some((message) => /风险|注意|不确定|risk|uncertain|confidence/i.test(message.content)) ? ["部分顾问提到了风险、不确定性或置信度较低，请主模型谨慎采用。"] : [];
-		if (cancelled) riskNotes.push("本次顾问群讨论已被用户或主模型取消，结论可能不完整。");
+		const riskNotes = buildRiskNotes(advisorMessages, cancelled);
 		return {
 			sessionId: session.id,
 			question: session.question,
@@ -1743,267 +3030,1514 @@ function matchAdvisors(text, advisors) {
 	return [...matches];
 }
 //#endregion
-//#region src/tools.ts
-function registerAdvisorTools(service) {
-	return [defineTool({
-		name: "ask_advisors",
-		description: "Consult configured expert advisor models when the question is professional, long-tail world knowledge, high-risk, or the main model is uncertain. ALWAYS provide context with (a) a brief overview of the current project and (b) concrete details of the problem, including what has been tried and what went wrong. Also call immediately when the user @顾问群 or repeats the same question three times without resolution. Do NOT use for simple facts that web search can answer. Starting a new consultation runs the auto-deepen pipeline: up to maxRounds rounds of [driver deep-question → advisor A → advisor B (sees A) → advisor C (sees A+B) → …], closed by a driver-generated conclusion.",
-		parameters: {
-			question: {
-				type: "string",
-				description: "The question to ask the advisors. Required when starting a new consultation."
-			},
-			context: {
-				type: "string",
-				description: "Required background: (a) current project overview (domain/goal/stack/constraints) and (b) problem details (steps tried, expected vs actual, errors). Write \"该项背景未知\" for anything unknown; never leave it empty."
-			},
-			sessionId: {
-				type: "string",
-				description: "Existing consultation session id to continue."
-			},
-			followUp: {
-				type: "string",
-				description: "Follow-up question to send to the advisors in an existing session."
-			},
-			advisorIds: {
-				type: "array",
-				items: { type: "string" },
-				description: "Optional advisor ids to include; defaults to all configured advisors (capped by maxAdvisorsPerCall)."
+//#region src/providers/jev.ts
+const DEFAULT_ENDPOINTS = {
+	typesafe: "https://api.typesafe.ai/v1/systemone",
+	openrouter: "https://openrouter.ai/api/alpha/decisions"
+};
+function numberValue(value) {
+	return typeof value === "number" && Number.isFinite(value) ? value : void 0;
+}
+function answerIsTrue(answer, threshold) {
+	const noul = numberValue(answer?.noul);
+	if (noul !== void 0) return noul >= threshold;
+	const choice = typeof answer?.choice === "string" ? answer.choice.toLowerCase() : "";
+	if (choice === "yes" || choice === "true") return true;
+	if (choice === "no" || choice === "false") return false;
+	return false;
+}
+/**
+* An answer is usable only when it carries a finite noul or an explicit
+* yes/no choice. Anything else (missing key, string noul, "maybe") would be
+* read as `false` by answerIsTrue — which for needs_advisor/high_risk means a
+* silent rejection with no fallback to the local classifier.
+*/
+function answerUsable(answer) {
+	if (numberValue(answer?.noul) !== void 0) return true;
+	const choice = typeof answer?.choice === "string" ? answer.choice.toLowerCase() : "";
+	return choice === "yes" || choice === "true" || choice === "no" || choice === "false";
+}
+function parseResponse(body, config, advisors) {
+	if (!body || typeof body !== "object") throw new Error("Jev 响应不是对象");
+	const answers = body.answers;
+	if (!answers || typeof answers !== "object") throw new Error("Jev 响应缺少 answers");
+	const needsAdvisor = answers.needs_advisor;
+	const webSearch = answers.web_search;
+	const risk = answers.high_risk;
+	const domainAnswer = answers.domain;
+	if (!answerUsable(needsAdvisor)) throw new Error("Jev 响应缺少 needs_advisor 的有效判定");
+	if (!answerUsable(risk)) throw new Error("Jev 响应缺少 high_risk 的有效判定");
+	const domain = typeof domainAnswer?.choice === "string" ? domainAnswer.choice : "general";
+	const needsAdvisorNoul = numberValue(needsAdvisor?.noul);
+	const confidence = numberValue(needsAdvisor?.confidence) ?? needsAdvisorNoul;
+	const shouldEscalate = answerIsTrue(needsAdvisor, config.confidenceThreshold);
+	const suggestWebSearch = answerIsTrue(webSearch, config.confidenceThreshold);
+	const highRisk = answerIsTrue(risk, config.highRiskThreshold ?? config.confidenceThreshold);
+	const reportedModel = body.model;
+	const webSearchNoul = numberValue(webSearch?.noul);
+	const riskNoul = numberValue(risk?.noul);
+	const rawAnswers = {
+		...needsAdvisorNoul === void 0 ? {} : { needsAdvisor: needsAdvisorNoul },
+		...webSearchNoul === void 0 ? {} : { webSearch: webSearchNoul },
+		...riskNoul === void 0 ? {} : { highRisk: riskNoul },
+		domain
+	};
+	return {
+		shouldEscalate: shouldEscalate || highRisk,
+		suggestWebSearch: !shouldEscalate && !highRisk && suggestWebSearch,
+		suggestedAdvisors: matchAdvisors(`${domain}\n${domainAnswer?.choice ?? ""}`, advisors),
+		reason: highRisk ? "Jev 判断该问题包含高风险因素，建议启动顾问群。" : shouldEscalate ? `Jev 判断该问题需要顾问群（${confidence === void 0 ? "无置信度" : `置信度 ${confidence}`}）。` : suggestWebSearch ? "Jev 判断该问题更适合联网搜索。" : "Jev 判断该问题不需要启动顾问群。",
+		...confidence === void 0 ? {} : { confidence },
+		rawAnswers,
+		...typeof reportedModel === "string" ? { model: reportedModel } : {},
+		provider: config.provider
+	};
+}
+async function classifyWithJev(question, context, advisors, config, signal, questionEn) {
+	const endpoint = config.baseURL?.trim() || DEFAULT_ENDPOINTS[config.provider];
+	const storedKey = typeof config.apiKey === "string" ? config.apiKey : config.apiKey && typeof config.apiKey === "object" && "value" in config.apiKey ? String(config.apiKey.value ?? "") : void 0;
+	const apiKey = (config.apiKeyEnv ? process.env[config.apiKeyEnv] : void 0) || storedKey;
+	if (!apiKey) throw new Error("Jev 未配置 API key");
+	const controller = new AbortController();
+	const timeout = setTimeout(() => controller.abort(/* @__PURE__ */ new Error(`Jev 超时（${config.timeoutMs}ms）`)), config.timeoutMs);
+	const abort = () => controller.abort();
+	signal?.addEventListener("abort", abort, { once: true });
+	try {
+		const response = await fetch(endpoint, {
+			method: "POST",
+			headers: {
+				authorization: `Bearer ${apiKey}`,
+				"content-type": "application/json"
 			},
-			confidence: {
-				type: "number",
-				description: "Optional self-assessed confidence 0-1. Values below the configured threshold escalate even without domain keywords."
-			}
-		},
-		output: {
-			schema: {
-				type: "object",
-				additionalProperties: false,
-				properties: {
-					sessionId: {
-						type: "string",
-						required: true
+			body: JSON.stringify({
+				state: config.useEnglishState && questionEn?.trim() ? questionEn.trim() : `${question}\n\n背景：${context || "无"}`,
+				model: config.model || "jev-latest",
+				questions: {
+					needs_advisor: {
+						type: "noul",
+						instructions: "Does answering this question require expert-level knowledge?",
+						criteria: {
+							true: {
+								what: "The subject matter is professional, specialist, or long-tail knowledge a competent generalist would not reliably get right, or the field itself has no single settled answer without specialist judgment.",
+								examples: [
+									"professional-domain judgment such as accounting treatment, engineering trade-offs, or contract drafting",
+									"rare or long-tail situations with little public consensus",
+									"topics where reasonable experts disagree and picking correctly matters"
+								]
+							},
+							false: {
+								what: "A competent generalist can answer from common knowledge, widely documented facts, or ordinary reasoning without specialist judgment.",
+								examples: ["everyday facts, definitions, or casual opinions", "well-documented mainstream topics with an agreed answer"]
+							}
+						}
 					},
-					advice: {
-						type: "string",
-						required: true
+					web_search: {
+						type: "noul",
+						instructions: "Does answering this question primarily require current, real-time, or externally verifiable information?",
+						criteria: {
+							true: {
+								what: "The correct answer depends on facts that change over time or live outside the provided state, so it must be looked up rather than reasoned out.",
+								examples: [
+									"today’s weather, prices, exchange rates, or quotes",
+									"recent releases, news, or version-specific changelogs",
+									"local or niche facts with no authoritative source in memory"
+								]
+							},
+							false: {
+								what: "The answer follows from stable knowledge or reasoning already available in the state.",
+								examples: ["concepts, definitions, and established best practices", "questions fully answerable from the provided background"]
+							}
+						}
 					},
-					skipped: { type: "boolean" },
-					reason: { type: "string" }
-				}
-			},
-			render: (_args, value) => [{
-				type: "text",
-				text: value.advice
-			}]
-		},
-		async execute(args, exec) {
-			const cfg = service.getConfig();
-			if (args.confidence !== void 0 && (args.confidence < 0 || args.confidence > 1)) return {
-				sessionId: "",
-				advice: "confidence 参数必须在 0-1 之间。",
-				skipped: true,
-				reason: "invalid-confidence"
-			};
-			if (!service.isEnabled()) return {
-				sessionId: "",
-				advice: "顾问群插件已禁用。请先调用 toggle_advisor_group 启用。",
-				skipped: true,
-				reason: "disabled"
-			};
-			if (cfg.advisors.length === 0) return {
-				sessionId: "",
-				advice: "顾问群尚未配置任何顾问。请先到 设置 → 插件 → 顾问群 添加顾问。",
-				skipped: true,
-				reason: "no-advisors"
-			};
-			const withoutModel = advisorsMissingModel(cfg);
-			if (withoutModel.length > 0) return {
-				sessionId: "",
-				advice: `顾问群中有 ${withoutModel.length} 位顾问未配置模型（${withoutModel.map((a) => a.name).join("、")}）。请先到 设置 → 插件 → 顾问群 补全模型后再发起，避免整轮调用失败。`,
-				skipped: true,
-				reason: "advisor-model-missing"
-			};
-			const sessionLog = exec.agent?.session;
-			const questionText = typeof args.question === "string" ? args.question : "";
-			const contextText = typeof args.context === "string" ? args.context : "";
-			const mentionedAdvisorGroup = /@顾问群|@顧問群/.test(`${questionText}\n${contextText}`);
-			const forcedByRepeat = (service.getRepeatPressure(sessionLog?.id)?.count ?? 0) >= 3;
-			if (args.sessionId) {
-				if (!args.followUp) return {
-					sessionId: args.sessionId,
-					advice: "继续追问需要提供 followUp 参数。",
-					skipped: true,
-					reason: "missing-follow-up"
-				};
-				const session = service.getSession(args.sessionId);
-				if (!session) return {
-					sessionId: args.sessionId,
-					advice: "顾问群会话不存在或已过期，请重新发起 ask_advisors。",
-					skipped: true,
-					reason: "session-not-found"
-				};
-				if (service.hasReachedMaxRounds(session)) return {
-					sessionId: session.id,
-					advice: "本轮已是顾问群讨论的最后一轮，请综合现有顾问意见给出最终答复，不要再追问。",
-					skipped: true,
-					reason: "max-rounds-reached"
-				};
-				service.appendFollowUp(session.id, args.followUp, sessionLog);
-				await service.runOneRoundAndSummarize(session, exec.signal, sessionLog);
-				const canContinue = !service.hasReachedMaxRounds(session);
-				return {
-					sessionId: session.id,
-					advice: formatConversation(session) + (canContinue ? "\n\n（如需继续，可再次传入 sessionId 和 followUp 追问。）" : "\n\n（讨论轮数已达上限，请综合顾问意见给出最终答复。）"),
-					skipped: false
-				};
-			}
-			if (!args.question) return {
-				sessionId: "",
-				advice: "发起新讨论需要提供 question 参数。",
-				skipped: true,
-				reason: "missing-question"
-			};
-			if (cfg.trigger.requireClassifier && !mentionedAdvisorGroup && !forcedByRepeat) {
-				const classification = classifyRequest(args.question, args.context, cfg.advisors, args.confidence, cfg.trigger.confidenceThreshold);
-				appendShadowSample({
-					ts: Date.now(),
-					question: args.question,
-					...typeof args.confidence === "number" ? { confidence: args.confidence } : {},
-					shouldEscalate: classification.shouldEscalate,
-					reason: classification.reason,
-					suggestWebSearch: classification.suggestWebSearch,
-					launched: classification.shouldEscalate
-				});
-				if (!classification.shouldEscalate) {
-					if (classification.suggestWebSearch && cfg.trigger.allowWebFallback) return {
-						sessionId: "",
-						advice: `前置分类器建议不启动顾问群：${classification.reason}`,
-						skipped: true,
-						reason: "web-search-recommended"
-					};
-					return {
-						sessionId: "",
-						advice: `前置分类器建议不启动顾问群：${classification.reason}`,
-						skipped: true,
-						reason: "classifier-rejected"
-					};
+					high_risk: {
+						type: "noul",
+						instructions: "Does this question involve medical, legal, financial, safety, compliance, or other high-risk advice?",
+						criteria: {
+							true: {
+								what: "A wrong or careless answer could materially harm the asker or a third party in health, money, safety, or legal standing.",
+								examples: [
+									"medical symptoms, medication, diagnosis, or treatment decisions",
+									"legal rights, contracts, litigation, or regulatory compliance",
+									"investment, insurance, tax, or other money decisions with real downside",
+									"physical safety, hazardous actions, or regulated activities"
+								]
+							},
+							false: {
+								what: "Being wrong carries no material harm to health, money, safety, or legal standing.",
+								examples: ["general knowledge, opinions, or entertainment questions", "technical details whose worst case is wasted time"]
+							}
+						}
+					},
+					domain: {
+						type: "choice",
+						instructions: "Which domain best describes this question?",
+						criteria: {
+							legal: "law, contracts, litigation, compliance, or taxation",
+							medical: "medicine, symptoms, medication, diagnosis, or treatment",
+							finance: "investment, insurance, accounting, or financial risk",
+							code: "software, programming, architecture, debugging, or APIs",
+							general: "none of the listed domains"
+						}
+					}
 				}
-			}
-			if (args.advisorIds && args.advisorIds.length > 0) {
-				const configuredIds = new Set(cfg.advisors.map((advisor) => advisor.id));
-				if (args.advisorIds.filter((id) => configuredIds.has(id)).length === 0) return {
-					sessionId: "",
-					advice: "未匹配到任何指定顾问，请检查 advisorIds 是否与设置中的顾问 id 一致。",
-					skipped: true,
-					reason: "no-matching-advisors"
-				};
-			}
-			const costGuard = service.tryStartConsultation();
-			if (!costGuard.ok) return {
-				sessionId: "",
-				advice: costGuard.reason,
-				skipped: true,
-				reason: "daily-limit-reached"
-			};
-			const session = service.createSession(args.question, args.context, args.advisorIds);
-			if (sessionLog) appendAdvisorStart(sessionLog, session);
-			const summary = await service.runAutoPipeline(session, exec.signal, sessionLog);
-			return {
-				sessionId: session.id,
-				advice: formatConversation(session) + (summary.conclusion ? `\n\n【综合结论】\n${summary.conclusion}` : "") + (service.hasReachedMaxRounds(session) ? "\n\n（已达设置的最大讨论轮数，请参考上述顾问意见与综合结论给出最终答复。）" : ""),
-				skipped: false
-			};
-		}
-	}), defineTool({
-		name: "toggle_advisor_group",
-		description: "Enable or disable the advisor group plugin at runtime. Omit enabled to toggle the current state.",
-		parameters: { enabled: {
-			type: "boolean",
-			description: "Optional desired state."
-		} },
-		output: {
-			schema: {
-				type: "object",
-				additionalProperties: false,
-				properties: { enabled: {
-					type: "boolean",
-					required: true
-				} }
-			},
-			render: (_args, value) => [{
-				type: "text",
-				text: `顾问群插件已${value.enabled ? "启用" : "禁用"}。`
-			}]
-		},
-		async execute(args) {
-			const next = typeof args.enabled === "boolean" ? args.enabled : !service.isEnabled();
-			return { enabled: await service.toggleEnabled(next) };
-		}
-	})];
-}
-function formatConversation(session) {
-	const lines = [`顾问群对话（session: ${session.id}）`, ""];
-	for (const message of session.messages) {
-		if (message.role === "system") continue;
-		if (message.role === "main") lines.push(`主模型：${message.content}`);
-		else if (message.role === "advisor") {
-			const body = message.content && message.content.trim() ? message.content : message.thinking && message.thinking.trim() ? `（思维链）\n${message.thinking}` : "（顾问未返回正文）";
-			lines.push(`【${message.advisorName ?? message.advisorId ?? "顾问"}】：${body}`);
-		}
-		lines.push("");
+			}),
+			signal: controller.signal
+		});
+		if (!response.ok) throw new Error(`Jev HTTP ${response.status}`);
+		return parseResponse(await response.json(), config, advisors);
+	} finally {
+		clearTimeout(timeout);
+		signal?.removeEventListener("abort", abort);
 	}
-	lines.push("（以上是顾问模型与主模型的对话记录，主模型可直接基于其中内容继续思考或追问。）");
-	return lines.join("\n");
 }
 //#endregion
-//#region src/settings-api.ts
-function sendJson(res, status, body) {
-	res.writeHead(status, { "content-type": "application/json" });
-	res.end(JSON.stringify(body));
+//#region src/cross-session.ts
+/**
+* Cross-session ask (v0): let the current session deliver one question to
+* another existing ordinary session, wake it, and report what that session
+* said between the delivery receipt and its next whole-session idle.
+*
+* Facts this module is built on (DSH 0.2.0-rc.1/rc.2 types, verified):
+* - `ctx.sessionController.resolveAgent(sessionId)` yields the live Agent
+*   (`{ agent }` or `{ error }`), resuming a cold session.
+* - `agent.followup(createUserMessage({ content, source }))` queues one
+*   ordinary follow-up turn and wakes the driver.
+* - `ctx.sessionController.prompt({ sessionId, content, mode:'queue', requestId })`
+*   is the cheap fallback, but the Host hardcodes `source` to a user prompt, so
+*   that path is identity-dishonest and must be reported as such.
+* - There is NO API that returns "the answer to this message". The Host only
+*   guarantees delivery, so this module reports an INTERVAL, never a reply.
+* - The Host enforces no caller identity, no cycle guard, and hard-excludes
+*   subagent sessions, so every guard in `ask_session` lives here.
+*
+* The pure helpers below never touch `ctx`; only `askSession()` does I/O.
+*/
+/** Maximum combined UTF-8 bytes of question + context (design §六.5). */
+const ASK_CONTENT_MAX_BYTES = 32e3;
+/** Default and maximum wait for the target's next idle (design §十). */
+const ASK_WAIT_DEFAULT_MS = 18e4;
+const ASK_WAIT_MAX_MS = 6e5;
+/** Poll interval while waiting for the target to reach whole-session idle. */
+const ASK_POLL_INTERVAL_MS = 1e3;
+/**
+* Envelope the receiving model reads. The Host offers no registration point for
+* this framing (the official Agent Teams build their own too), so the identity
+* is carried in the content itself: a header block naming the asking session,
+* then the question, then optional background.
+* @param senderName - title (or cwd+time fallback) of the asking session.
+* @param question - the question to hand over.
+* @param context - optional background; permanently recorded in the target log.
+* @returns content blocks, envelope first.
+*/
+function buildAskEnvelope(senderName, question, context) {
+	const blocks = [{
+		type: "text",
+		text: `【来自会话「${senderName}」的提问】`
+	}, {
+		type: "text",
+		text: question
+	}];
+	if (context !== void 0 && context.trim().length > 0) blocks.push({
+		type: "text",
+		text: `【对方补充的背景】\n${context}`
+	});
+	return blocks;
 }
-/** Mask a stored API key so it never leaves the host in plain text. */
-function maskApiKey(apiKey) {
-	if (!apiKey) return apiKey;
-	if (apiKey.length <= 4) return "*".repeat(apiKey.length);
-	return `${"*".repeat(apiKey.length - 4)}${apiKey.slice(-4)}`;
+/**
+* Build the identified user message handed to `agent.followup()`.
+* @param envelope - content blocks produced by {@link buildAskEnvelope}.
+* @param source - producer-declared session-ask source.
+* @returns an immutable identified user message carrying our source kind.
+*/
+function buildAskMessage(envelope, source) {
+	return createUserMessage({
+		content: envelope,
+		source
+	});
 }
 /**
-* Whether an inbound apiKey is a mask of the previously stored key.
-*
-* Accepts both the current length-preserving mask and the legacy 8-character
-* `****<last4>` mask so an unrefreshed settings page does not overwrite the
-* real key with a mask string after upgrading.
+* UTF-8 byte length of the text a caller wants to hand over.
+* @param question - the question text.
+* @param context - optional background text.
+* @returns byte length of question + context.
 */
-function isMaskedApiKey(incoming, previous) {
-	if (!incoming || !previous) return false;
-	if (incoming === maskApiKey(previous)) return true;
-	if (previous.length <= 4) return incoming === "*".repeat(previous.length);
-	return incoming === `****${previous.slice(-4)}`;
+function askContentBytes(question, context) {
+	return Buffer.byteLength(question, "utf8") + (context === void 0 ? 0 : Buffer.byteLength(context, "utf8"));
+}
+/** Deterministic fallback label: working directory plus most recent activity. */
+function fallbackSessionName(cwd, updatedAt) {
+	const where = cwd === void 0 || cwd.length === 0 ? "（未知工作目录）" : cwd;
+	if (updatedAt === void 0 || !Number.isFinite(updatedAt)) return where;
+	const at = new Date(updatedAt);
+	return `${where} · ${Number.isNaN(at.getTime()) ? String(updatedAt) : at.toISOString()}`;
 }
 /**
-* Return a copy of the config safe to send to the browser.
-*
-* 0.1.2-rc.1 SecretField convergence: no key material (not even a mask) leaves
-* the host. `apiKey` is returned empty and per-provider presence facts
-* (`configured` + `last4`) ride in `apiKeyMetaByProvider`, so the card can
-* render the official "已配置/未配置" badge and publish new/cleared keys
-* without ever round-tripping a mask.
+* Model-visible name of one listed session. The title lives in the `title`
+* projection (a snapshot object), is optional, and may be a stale cache value,
+* so every failure degrades to the cwd+time fallback instead of throwing.
 */
-function sanitizeConfig(config) {
+function sessionDisplayName(summary) {
+	const raw = summary.projections?.values?.title;
+	const title = raw !== null && typeof raw === "object" && typeof raw.title === "string" ? raw.title.trim() : "";
+	if (title.length > 0) return title;
+	return fallbackSessionName(summary.cwd, summary.updatedAt);
+}
+function candidateOf(summary) {
 	return {
-		...config,
-		advisors: config.advisors.map((advisor) => {
-			const { apiKeysByProvider: _apiKeysByProvider, clearApiKey: _clearApiKey, apiKeyMetaByProvider: _meta, ...rest } = advisor;
-			const meta = {};
-			for (const [provider, key] of Object.entries(advisor.apiKeysByProvider ?? {})) if (key) meta[provider] = {
-				configured: true,
-				last4: key.slice(-4)
-			};
-			if (advisor.apiKey && advisor.provider) meta[advisor.provider] = {
-				configured: true,
-				last4: advisor.apiKey.slice(-4)
-			};
+		sessionId: summary.sessionId,
+		name: sessionDisplayName(summary),
+		hint: fallbackSessionName(summary.cwd, summary.updatedAt)
+	};
+}
+/** Most-recently-active first, running sessions ahead of idle ones. */
+function orderCandidates(rows) {
+	return [...rows].sort((a, b) => {
+		if (a.running !== b.running) return a.running ? -1 : 1;
+		return b.updatedAt - a.updatedAt;
+	});
+}
+/** Partial candidates are capped: a broad needle must stay a readable list. */
+const PARTIAL_CANDIDATE_LIMIT = 10;
+/**
+* Resolve a user/model-supplied target reference against the session list.
+* Ladder, in order: exact sessionId → exact display name → partial natural
+* name (substring of the display name OR of the workspace cwd, case-insensitive)
+* → missing. One hit resolves outright; several hits become candidates —
+* this function NEVER guesses. Partial candidates are ordered most-recently-
+* active first (running sessions first) and capped, so "ask the 极简遥控器
+* session" works without anyone digging session ids out of log directories.
+* @param sessions - rows from `sessionController.list()`.
+* @param reference - a session id, display name, or natural partial name.
+* @returns the unique target, the candidate list, or "missing".
+*/
+function matchTargetSession(sessions, reference) {
+	const needle = reference.trim();
+	if (needle.length === 0) return { kind: "missing" };
+	const byId = sessions.find((session) => session.sessionId === needle);
+	if (byId) return {
+		kind: "unique",
+		target: {
+			sessionId: byId.sessionId,
+			name: sessionDisplayName(byId)
+		}
+	};
+	const folded = needle.toLowerCase();
+	const exact = sessions.filter((session) => sessionDisplayName(session).toLowerCase() === folded);
+	if (exact.length === 1) return {
+		kind: "unique",
+		target: {
+			sessionId: exact[0].sessionId,
+			name: sessionDisplayName(exact[0])
+		}
+	};
+	if (exact.length > 1) return {
+		kind: "ambiguous",
+		candidates: orderCandidates(exact).map(candidateOf)
+	};
+	const partial = sessions.filter((session) => sessionDisplayName(session).toLowerCase().includes(folded) || session.cwd !== void 0 && session.cwd.toLowerCase().includes(folded));
+	if (partial.length === 1) return {
+		kind: "unique",
+		target: {
+			sessionId: partial[0].sessionId,
+			name: sessionDisplayName(partial[0])
+		}
+	};
+	if (partial.length === 0) return { kind: "missing" };
+	return {
+		kind: "ambiguous",
+		candidates: orderCandidates(partial).slice(0, PARTIAL_CANDIDATE_LIMIT).map(candidateOf)
+	};
+}
+/**
+* Hop guard counter: the hop of the question this session is about to send.
+* The value is carried by the asking message that reached this session, so it
+* is read from the most recent inbound `user/message` whose source is
+* `session-ask`; a later plain human prompt does not reset it, because the
+* question being asked here still descends from that relay. No session-ask
+* source anywhere means this is a fresh, top-level question (hop 0).
+* @param events - this session's log events in order.
+* @returns the hop to stamp on the outgoing message.
+*/
+function nextAskHop(events) {
+	for (let i = events.length - 1; i >= 0; i -= 1) {
+		const event = events[i];
+		if (event.type !== "user/message") continue;
+		const source = event.data.message?.source;
+		if (source?.kind !== "session-ask") continue;
+		return typeof source.hop === "number" && Number.isFinite(source.hop) ? source.hop + 1 : 1;
+	}
+	return 0;
+}
+/**
+* Collect what the target session said inside the interval. Only
+* `assistant/message` events strictly after the delivery watermark count, so a
+* reply that predates the delivery cannot be reported as an answer. The
+* watermark (highest seq already present at delivery) is the authoritative cut;
+* `fromMs` is kept only as the recorded interval start.
+* @param events - the target session's log events in order.
+* @param fromSeq - highest event seq already present at delivery time.
+* @returns the newest event time seen and the per-message text pieces.
+*/
+function collectAnswer(events, fromSeq) {
+	const pieces = [];
+	let latestEventTime = null;
+	for (const event of events) {
+		if (Number.isFinite(event.time) && (latestEventTime === null || event.time > latestEventTime)) latestEventTime = event.time;
+		if (event.type !== "assistant/message") continue;
+		if (!(event.seq > fromSeq)) continue;
+		const text = (event.data?.message?.content ?? []).filter((block) => block.type === "text" && typeof block.text === "string").map((block) => block.text).join("").trim();
+		if (text.length > 0) pieces.push(text);
+	}
+	return {
+		latestEventTime,
+		pieces
+	};
+}
+/** Join the interval's text pieces into the returned answer body. */
+function joinAnswerPieces(pieces) {
+	return pieces.join("\n\n");
+}
+/** Highest event seq present, or -1 for an empty log (the session-cursor convention). */
+function logWatermark(events) {
+	let highest = -1;
+	for (const event of events) if (typeof event.seq === "number" && event.seq > highest) highest = event.seq;
+	return highest;
+}
+/** Human-readable text the main model reads for one settled ask_session call. */
+function renderAskResult(value) {
+	return [{
+		type: "text",
+		text: `${value.answer.trim().length > 0 ? value.answer : "（该区间内对方没有产出正文。）"}\n\n${value.note}`
+	}];
+}
+/**
+* The Host hard-excludes subagent sessions with `session/agent-busy` and the
+* reason "use subagent delivery for this child session"; recognise it so the
+* caller gets the actionable `subagent-target` code instead of `delivery-failed`.
+*/
+function isSubagentDeliveryRejection(error) {
+	if (error === null || typeof error !== "object") return false;
+	const shape = error;
+	const reason = shape.details !== null && typeof shape.details === "object" ? shape.details.reason : void 0;
+	const haystack = `${typeof reason === "string" ? reason : ""} ${typeof shape.message === "string" ? shape.message : ""}`;
+	return haystack.includes("subagent delivery") || haystack.includes("owned by subagent routing");
+}
+/** One unsettled wait per session: at most one ask_session may be in flight. */
+const pendingWaits = /* @__PURE__ */ new Set();
+/** Whether this session already has an unsettled ask_session wait. */
+function isWaiting(sessionId) {
+	return pendingWaits.has(sessionId);
+}
+let askCounter = 0;
+/** Process-unique ask id (the receiver side may dedupe on it). */
+function nextAskId() {
+	askCounter += 1;
+	return `ask-${Date.now().toString(36)}-${askCounter}-${Math.random().toString(36).slice(2, 8)}`;
+}
+/**
+* Lazy host-service access: never a hard `inject` (older hosts lack this service).
+*
+* A bare `ctx.sessionController` read is gated by the plugin's `inject` list and
+* throws `cannot get property "sessionController" without inject` when the key is
+* not declared — so the property read itself must not be used. `ctx.get(name)` is
+* Cordis's documented read "without the inject requirement": it returns the service
+* or `undefined` when (not yet) provided.
+*/
+function getSessionController(ctx) {
+	if (ctx === null || ctx === void 0) return void 0;
+	try {
+		const context = ctx;
+		const controller = typeof context.get === "function" ? context.get.call(ctx, "sessionController") : context.sessionController;
+		if (controller === null || typeof controller !== "object") return void 0;
+		const candidate = controller;
+		if (typeof candidate.list !== "function" || typeof candidate.resolveAgent !== "function") return void 0;
+		return candidate;
+	} catch {
+		return;
+	}
+}
+/** Read a live Agent's own log through a defensive view (the log API is read-only here). */
+function agentEvents(agent) {
+	try {
+		return agent.session?.snapshotEvents() ?? [];
+	} catch {
+		return [];
+	}
+}
+async function sleep(ms, signal) {
+	if (signal.aborted) return;
+	await new Promise((resolve) => {
+		const timer = setTimeout(resolve, ms);
+		signal.addEventListener("abort", () => {
+			clearTimeout(timer);
+			resolve();
+		}, { once: true });
+	});
+}
+/**
+* Hand the question to the target session.
+*
+* Preferred path: `resolveAgent` + `agent.followup(...)`, which preserves our
+* own source kind. Fallback: `sessionController.prompt(...)`, which the Host
+* stamps as a user prompt — the envelope still names us, and the caller is told
+* in `note` that this identity-dishonest path was used.
+*/
+async function deliverAsk(controller, signal, input) {
+	const envelope = buildAskEnvelope(input.source.senderName, input.question, input.context);
+	try {
+		const resolved = await controller.resolveAgent(input.target.sessionId);
+		if ("agent" in resolved && resolved.agent) {
+			const agent = resolved.agent;
+			agent.followup(buildAskMessage(envelope, input.source));
+			return {
+				ok: true,
+				receipt: {
+					delivered: "native",
+					note: "本次投递走了首选通道（自定义来源），对方模型能看到这条消息来自另一个会话。",
+					fromSeq: logWatermark(agentEvents(agent)),
+					fromMs: Date.now(),
+					targetWasRunning: input.targetWasRunning
+				}
+			};
+		}
+		const reason = "error" in resolved ? resolved.error : void 0;
+		if (isSubagentDeliveryRejection(reason)) return {
+			ok: false,
+			code: "subagent-target",
+			message: `目标会话 ${input.target.sessionId} 属于子代理（subagent）会话，跨会话投递被宿主硬性排除。请改用子代理通道（subagent / send_message）。`
+		};
+		return deliverViaPrompt(controller, signal, envelope, input, `首选通道不可用（${describeHostError(reason)}）`);
+	} catch (error) {
+		if (signal.aborted) throw error;
+		if (isSubagentDeliveryRejection(error)) return {
+			ok: false,
+			code: "subagent-target",
+			message: `目标会话 ${input.target.sessionId} 属于子代理（subagent）会话，跨会话投递被宿主硬性排除。请改用子代理通道（subagent / send_message）。`
+		};
+		if (isTargetNotFound(error)) return {
+			ok: false,
+			code: "target-not-found",
+			message: `找不到会话 ${input.target.sessionId}（宿主报告 session/not-found）。`
+		};
+		return deliverViaPrompt(controller, signal, envelope, input, `首选通道报错（${describeHostError(error)}）`);
+	}
+}
+async function deliverViaPrompt(controller, _signal, envelope, input, why) {
+	try {
+		await controller.prompt({
+			requestId: input.source.askId,
+			sessionId: input.target.sessionId,
+			mode: "queue",
+			content: envelope
+		});
+		const floor = await watermarkViaResolve(controller, input.target.sessionId);
+		return {
+			ok: true,
+			receipt: {
+				delivered: "fallback",
+				note: `本次投递走了降级通道（宿主的 prompt 接口），宿主把来源写死成「用户输入」，对方模型会以为这是用户打的字；身份只能靠正文里的信封说明。原因：${why}`,
+				fromSeq: floor,
+				fromMs: Date.now(),
+				targetWasRunning: input.targetWasRunning
+			}
+		};
+	} catch (error) {
+		if (isSubagentDeliveryRejection(error)) return {
+			ok: false,
+			code: "subagent-target",
+			message: `目标会话 ${input.target.sessionId} 属于子代理（subagent）会话，跨会话投递被宿主硬性排除。请改用子代理通道（subagent / send_message）。`
+		};
+		if (isTargetNotFound(error)) return {
+			ok: false,
+			code: "target-not-found",
+			message: `找不到会话 ${input.target.sessionId}（宿主报告 session/not-found）。`
+		};
+		return {
+			ok: false,
+			code: "delivery-failed",
+			message: `投递失败：首选通道与降级通道都未能把问题送进目标会话。${why} 降级通道报错：${describeHostError(error)}`
+		};
+	}
+}
+async function watermarkViaResolve(controller, sessionId) {
+	try {
+		const resolved = await controller.resolveAgent(sessionId);
+		if ("agent" in resolved && resolved.agent) return logWatermark(agentEvents(resolved.agent));
+	} catch {}
+	return -1;
+}
+function isTargetNotFound(error) {
+	if (error === null || typeof error !== "object") return false;
+	return error.code === "session/not-found";
+}
+function describeHostError(error) {
+	if (error === null || error === void 0) return "未提供原因";
+	if (typeof error === "string") return error;
+	if (typeof error === "object") {
+		const shape = error;
+		const code = typeof shape.code === "string" ? shape.code : "";
+		const message = typeof shape.message === "string" ? shape.message : "";
+		if (code.length > 0 || message.length > 0) return `${code}${code && message ? " " : ""}${message}`;
+		try {
+			return JSON.stringify(error);
+		} catch {
+			return "宿主返回了无法序列化的错误";
+		}
+	}
+	return String(error);
+}
+/**
+* Wait for the target session's next whole-session idle after delivery.
+* Never cancels the target turn — a timeout stops only this session's waiting.
+*/
+async function waitForTargetIdle(controller, signal, input) {
+	const startedAt = Date.now();
+	let targetRunning = input.targetWasRunning;
+	let observedBusy = input.targetWasRunning;
+	const watermark = async () => {
+		try {
+			const resolved = await controller.resolveAgent(input.sessionId);
+			if (!("agent" in resolved) || !resolved.agent) return void 0;
+			return logWatermark(agentEvents(resolved.agent));
+		} catch {
+			return;
+		}
+	};
+	while (true) {
+		if (signal.aborted) return {
+			endedBecause: "error",
+			lastRunning: targetRunning,
+			note: "本次调用被取消，等待中止；没有取消对方的回合，对方可能仍在作答。"
+		};
+		const elapsed = Date.now() - startedAt;
+		if (elapsed >= input.waitMs) return {
+			endedBecause: "timeout",
+			lastRunning: targetRunning,
+			note: `等待达到上限（${input.waitMs}ms），已返回当前区间内收集到的内容；没有取消对方的回合，对方可能仍在作答。`
+		};
+		await sleep(Math.min(ASK_POLL_INTERVAL_MS, Math.max(0, input.waitMs - elapsed)), signal);
+		let row;
+		try {
+			row = (await controller.list({}, signal)).items.find((item) => item.sessionId === input.sessionId);
+		} catch (error) {
+			if (signal.aborted) return {
+				endedBecause: "error",
+				lastRunning: targetRunning,
+				note: "本次调用被取消，等待中止；没有取消对方的回合。"
+			};
+			return {
+				endedBecause: "error",
+				lastRunning: targetRunning,
+				note: `等待期间读取会话列表失败（${describeHostError(error)}），已返回当前区间内收集到的内容。`
+			};
+		}
+		if (!row) continue;
+		if (row.running) {
+			observedBusy = true;
+			targetRunning = true;
+			continue;
+		}
+		if (observedBusy) return {
+			endedBecause: "idle",
+			lastRunning: false,
+			note: "等待结束于该会话整体空闲（idle）。"
+		};
+		if (input.fromSeq !== void 0) {
+			const now = await watermark();
+			if (now !== void 0 && now > input.fromSeq) return {
+				endedBecause: "idle",
+				lastRunning: false,
+				note: "等待结束：该会话已追加新内容且当前不忙（对方的回合发生在两次轮询之间，忙的状态未被采到）。"
+			};
+		}
+		input.log(`waiting for ${input.sessionId} to start the delivered turn`);
+		targetRunning = false;
+	}
+}
+/** Read the target's log back and collect what it said inside the interval. */
+async function collectTargetAnswer(controller, sessionId, fromSeq) {
+	try {
+		const resolved = await controller.resolveAgent(sessionId);
+		if (!("agent" in resolved) || !resolved.agent) return {
+			pieces: [],
+			readError: describeHostError("error" in resolved ? resolved.error : void 0)
+		};
+		return { pieces: collectAnswer(agentEvents(resolved.agent), fromSeq).pieces };
+	} catch (error) {
+		return {
+			pieces: [],
+			readError: describeHostError(error)
+		};
+	}
+}
+/**
+* Every pre-delivery guard, in one place. Pure: takes plain facts, returns a
+* rejection or nothing. Order matters — identity before ownership, ownership
+* before budget, budget before hop.
+* @returns the rejection, or undefined when the ask may proceed.
+*/
+function checkAskGuards(input) {
+	const { guards, match } = input;
+	if (match.kind === "missing") return {
+		code: "target-not-found",
+		message: "没有找到匹配的会话。请用会话列表里的标题或 sessionId 指定目标；不确认时先列出候选再让用户选择。"
+	};
+	if (match.kind === "ambiguous") return {
+		code: "target-ambiguous",
+		message: `有 ${match.candidates.length} 个会话都叫这个名字，不能猜。请把候选交给用户挑选（用 sessionId 指定即可精确命中）。`,
+		candidates: match.candidates
+	};
+	const target = input.target ?? match.target;
+	if (target.sessionId === guards.currentSessionId) return {
+		code: "self-target",
+		message: "目标是当前会话自己，拒绝自问自答。"
+	};
+	if (input.sessions.find((session) => session.sessionId === target.sessionId)?.origin === "subagent") return {
+		code: "subagent-target",
+		message: "目标会话是子代理（subagent）会话，跨会话投递被宿主硬性排除。请改用子代理通道（subagent / send_message）。"
+	};
+	const bytes = askContentBytes(input.question, input.context);
+	if (bytes > 32e3) return {
+		code: "content-too-large",
+		message: `问题加背景共 ${bytes} 字节，超过上限 ${ASK_CONTENT_MAX_BYTES} 字节。请缩短问题，或把背景留在自己的会话里分次提问。`
+	};
+	if (guards.hop > 2) return {
+		code: "hop-limit",
+		message: `本次提问会形成第 ${guards.hop} 跳转发（上限 3 跳：0→2）。会话链式互问会绕成环，已拒绝；请由用户直接指定目标会话，或让当前会话自己作答。`
+	};
+}
+/** Normalise and validate the model-supplied wait budget. */
+function normalizeWaitMs(value) {
+	if (typeof value !== "number" || !Number.isFinite(value) || value <= 0) return ASK_WAIT_DEFAULT_MS;
+	return Math.min(Math.floor(value), ASK_WAIT_MAX_MS);
+}
+async function askSession(ctx, signal, input) {
+	const controller = getSessionController(ctx);
+	if (!controller) return {
+		ok: false,
+		error: "service-missing",
+		message: "此功能需要较新的 DSH 宿主：当前宿主没有提供 sessionController 服务，无法向其他会话投递问题。"
+	};
+	const waitMs = normalizeWaitMs(input.waitMs);
+	if (typeof input.waitMs === "number" && input.waitMs > 6e5) (input.log ?? (() => {}))(`ask_session waitMs 超出上限，已按 ${waitMs}ms 处理`);
+	if (isWaiting(input.currentSessionId)) return toToolError({
+		code: "already-waiting",
+		message: "当前会话已经在等另一个会话的答复（同一时刻只允许等一个），请等这次返回后再发起新的提问。"
+	});
+	pendingWaits.add(input.currentSessionId);
+	try {
+		const sessions = [];
+		let match = { kind: "missing" };
+		try {
+			const listed = await controller.list({}, signal);
+			sessions.push(...listed.items);
+			match = matchTargetSession(sessions, input.target ?? "");
+		} catch (error) {
+			return {
+				ok: false,
+				error: "service-missing",
+				message: `读取会话列表失败：${describeHostError(error)}`
+			};
+		}
+		return await runAsk(ctx, controller, signal, input, {
+			sessions,
+			match,
+			log: input.log ?? (() => {})
+		});
+	} finally {
+		pendingWaits.delete(input.currentSessionId);
+	}
+}
+/**
+* The ask itself, after the target list has been read and the wait latch is
+* held by the caller.
+*/
+async function runAsk(ctx, controller, signal, input, listed) {
+	const waitMs = normalizeWaitMs(input.waitMs);
+	const hop = nextAskHop(input.currentEvents);
+	const rejectionBase = checkAskGuards({
+		guards: {
+			currentSessionId: input.currentSessionId,
+			hop
+		},
+		match: listed.match,
+		sessions: listed.sessions,
+		question: input.question,
+		context: input.context
+	});
+	if (rejectionBase) return toToolError(rejectionBase);
+	const target = listed.match.kind === "unique" ? listed.match.target : {
+		sessionId: "",
+		name: ""
+	};
+	const targetRow = listed.sessions.find((session) => session.sessionId === target.sessionId);
+	const askId = nextAskId();
+	const source = {
+		kind: "session-ask",
+		form: "relay",
+		askId,
+		senderSessionId: input.currentSessionId,
+		senderName: input.senderName,
+		hop
+	};
+	const from = (/* @__PURE__ */ new Date()).toISOString();
+	const delivery = await deliverAsk(controller, signal, {
+		target,
+		question: input.question,
+		context: input.context,
+		source,
+		targetWasRunning: targetRow?.running === true
+	});
+	if (!delivery.ok) return {
+		ok: false,
+		error: delivery.code,
+		message: delivery.message
+	};
+	input.onDelivered?.({
+		askId,
+		question: input.question,
+		...input.context === void 0 ? {} : { context: input.context },
+		senderName: input.senderName,
+		targetName: target.name
+	});
+	const outcome = await waitForTargetIdle(controller, signal, {
+		sessionId: target.sessionId,
+		targetWasRunning: delivery.receipt.targetWasRunning,
+		waitMs,
+		log: listed.log,
+		fromSeq: delivery.receipt.fromSeq
+	});
+	const collected = await collectTargetAnswer(controller, target.sessionId, delivery.receipt.fromSeq);
+	const answer = joinAnswerPieces(collected.pieces);
+	const to = outcome.endedBecause === "idle" ? (/* @__PURE__ */ new Date()).toISOString() : null;
+	const note = [
+		"区间语义：从投递回执到对方会话下一次整体空闲，这两个时刻之间对方产出的文字都算在内——这是「这个区间里对方说的话」，不是逐条对应的回答。",
+		delivery.receipt.note,
+		outcome.note,
+		delivery.receipt.targetWasRunning ? "投递时对方正在忙，问题已排队，会在它当前回合结束后进入新回合。" : "投递时对方不忙，宿主已唤醒它开一个新回合。",
+		collected.readError === void 0 ? "" : `读取对方会话日志时出错，answer 可能不完整：${collected.readError}`
+	].filter((line) => line.length > 0).join("\n");
+	input.onSettled?.({
+		askId,
+		question: input.question,
+		targetName: target.name,
+		answer,
+		note
+	});
+	return {
+		ok: true,
+		target: {
+			sessionId: target.sessionId,
+			name: target.name
+		},
+		delivered: delivery.receipt.delivered,
+		askId,
+		interval: {
+			from,
+			to,
+			endedBecause: outcome.endedBecause
+		},
+		answer,
+		note
+	};
+}
+function toToolError(rejection) {
+	return {
+		ok: false,
+		error: rejection.code,
+		message: rejection.message,
+		...rejection.candidates ? { candidates: rejection.candidates } : {}
+	};
+}
+//#endregion
+//#region src/tools.ts
+function registerAdvisorTools(service, ctx) {
+	return [
+		defineTool({
+			name: "ask_advisors",
+			description: "Consult configured expert advisor models when the question is professional, long-tail world knowledge, high-risk, or the main model is uncertain. ALWAYS provide context with (a) a brief overview of the current project and (b) concrete details of the problem, including what has been tried and what went wrong. Also call immediately when the user @顾问群 or repeats the same question three times without resolution. Do NOT use for simple facts that web search can answer. Starting a new consultation runs the auto-deepen pipeline: up to maxRounds rounds of [driver deep-question → advisor A → advisor B (sees A) → advisor C (sees A+B) → …], closed by a driver-generated conclusion.",
+			parameters: {
+				question: {
+					type: "string",
+					description: "The question to ask the advisors. Required when starting a new consultation."
+				},
+				context: {
+					type: "string",
+					description: "Required background: (a) current project overview (domain/goal/stack/constraints) and (b) problem details (steps tried, expected vs actual, errors). Write \"该项背景未知\" for anything unknown; never leave it empty."
+				},
+				sessionId: {
+					type: "string",
+					description: "Existing consultation session id to continue."
+				},
+				followUp: {
+					type: "string",
+					description: "Follow-up question to send to the advisors in an existing session."
+				},
+				advisorIds: {
+					type: "array",
+					items: { type: "string" },
+					description: "Optional advisor ids to include; defaults to all configured advisors (capped by maxAdvisorsPerCall)."
+				},
+				confidence: {
+					type: "number",
+					description: "Optional self-assessed confidence 0-1. Values below the configured threshold escalate even without domain keywords."
+				},
+				questionEn: {
+					type: "string",
+					description: "Optional self-contained English gist of the question plus key background. Used as the Jev pre-classification input only when the「Jev 英文判定」setting is enabled; the displayed question, session records, and duplicate detection always keep using question/context in Chinese. Omit it otherwise."
+				}
+			},
+			output: {
+				schema: {
+					type: "object",
+					additionalProperties: false,
+					properties: {
+						sessionId: {
+							type: "string",
+							required: true
+						},
+						advice: {
+							type: "string",
+							required: true
+						},
+						skipped: { type: "boolean" },
+						reason: { type: "string" }
+					}
+				},
+				render: (_args, value) => [{
+					type: "text",
+					text: value.advice
+				}]
+			},
+			async execute(args, exec) {
+				const cfg = service.getConfig();
+				if (args.confidence !== void 0 && (args.confidence < 0 || args.confidence > 1)) return {
+					sessionId: "",
+					advice: "confidence 参数必须在 0-1 之间。",
+					skipped: true,
+					reason: "invalid-confidence"
+				};
+				if (!service.isEnabled()) return {
+					sessionId: "",
+					advice: "顾问群插件已禁用。请先调用 toggle_advisor_group 启用。",
+					skipped: true,
+					reason: "disabled"
+				};
+				if (cfg.advisors.length === 0) return {
+					sessionId: "",
+					advice: "顾问群尚未配置任何顾问。请先到 设置 → 插件 → 顾问群 添加顾问。",
+					skipped: true,
+					reason: "no-advisors"
+				};
+				const withoutModel = advisorsMissingModel(cfg);
+				if (withoutModel.length > 0) return {
+					sessionId: "",
+					advice: `顾问群中有 ${withoutModel.length} 位顾问未配置模型（${withoutModel.map((a) => a.name).join("、")}）。请先到 设置 → 插件 → 顾问群 补全模型后再发起，避免整轮调用失败。`,
+					skipped: true,
+					reason: "advisor-model-missing"
+				};
+				const sessionLog = exec.agent?.session;
+				const questionText = typeof args.question === "string" ? args.question : "";
+				const contextText = typeof args.context === "string" ? args.context : "";
+				const mentionedAdvisorGroup = /@顾问群|@顧問群/.test(`${questionText}\n${contextText}`);
+				const repeatCount = service.getRepeatPressure(sessionLog?.id)?.count ?? 0;
+				const forcedByRepeat = repeatCount >= 3;
+				let shadow;
+				let shadowWritten = false;
+				const recordShadow = (launched) => {
+					if (!shadow || shadowWritten) return;
+					shadowWritten = true;
+					appendShadowSample({
+						...shadow,
+						launched
+					});
+				};
+				if (args.sessionId) {
+					if (!args.followUp) return {
+						sessionId: args.sessionId,
+						advice: "继续追问需要提供 followUp 参数。",
+						skipped: true,
+						reason: "missing-follow-up"
+					};
+					const session = service.getSession(args.sessionId);
+					if (!session) return {
+						sessionId: args.sessionId,
+						advice: "顾问群会话不存在或已过期，请重新发起 ask_advisors。",
+						skipped: true,
+						reason: "session-not-found"
+					};
+					if (service.isSessionRunning(session.id)) return {
+						sessionId: session.id,
+						advice: "该顾问群会话正在运行中，请等本轮结束后再追问。",
+						skipped: true,
+						reason: "session-running"
+					};
+					if (service.hasReachedMaxRounds(session)) return {
+						sessionId: session.id,
+						advice: "本轮已是顾问群讨论的最后一轮，请综合现有顾问意见给出最终答复，不要再追问。",
+						skipped: true,
+						reason: "max-rounds-reached"
+					};
+					service.appendFollowUp(session.id, args.followUp, sessionLog);
+					const summary = await service.runOneRoundAndSummarize(session, exec.signal, sessionLog, exec.agent);
+					const canContinue = !service.hasReachedMaxRounds(session);
+					return {
+						sessionId: session.id,
+						advice: formatConversation(session) + formatRiskNotes(summary) + (canContinue ? "\n\n（如需继续，可再次传入 sessionId 和 followUp 追问。）" : "\n\n（讨论轮数已达上限，请综合顾问意见给出最终答复。）"),
+						skipped: false
+					};
+				}
+				if (!args.question) return {
+					sessionId: "",
+					advice: "发起新讨论需要提供 question 参数。",
+					skipped: true,
+					reason: "missing-question"
+				};
+				if (cfg.trigger.requireClassifier && !mentionedAdvisorGroup && !forcedByRepeat) {
+					let classification;
+					let jevResult;
+					let jevError;
+					let jevLatencyMs;
+					const jev = cfg.trigger.jev;
+					if (jev?.enabled) {
+						const jevStartedAt = Date.now();
+						try {
+							jevResult = await classifyWithJev(args.question, args.context, cfg.advisors, jev, exec.signal, typeof args.questionEn === "string" ? args.questionEn : void 0);
+							classification = jevResult;
+						} catch (error) {
+							if (exec.signal.aborted) throw error;
+							jevError = error instanceof Error ? error.message : String(error);
+							console.warn("[dsh-advisor-group] Jev 分类失败，回退到本地分类器：", sessionLog?.id ?? "no-session", jevError);
+						} finally {
+							jevLatencyMs = Date.now() - jevStartedAt;
+						}
+					}
+					classification ??= classifyRequest(args.question, args.context, cfg.advisors, args.confidence, cfg.trigger.confidenceThreshold);
+					shadow = {
+						ts: Date.now(),
+						question: args.question,
+						...typeof args.confidence === "number" ? { confidence: args.confidence } : {},
+						shouldEscalate: classification.shouldEscalate,
+						reason: classification.reason,
+						suggestWebSearch: classification.suggestWebSearch,
+						provider: jevResult ? "jev" : "local",
+						...jevError === void 0 ? {} : { jevError },
+						...jevLatencyMs === void 0 ? {} : { jevLatencyMs },
+						...repeatCount > 0 ? { repeatCount } : {},
+						...jevResult?.rawAnswers ? { scores: jevResult.rawAnswers } : {},
+						...jevResult?.model ? { model: jevResult.model } : {}
+					};
+					if (!classification.shouldEscalate) {
+						recordShadow(false);
+						if (classification.suggestWebSearch && cfg.trigger.allowWebFallback) return {
+							sessionId: "",
+							advice: `前置分类器建议不启动顾问群：${classification.reason}`,
+							skipped: true,
+							reason: "web-search-recommended"
+						};
+						return {
+							sessionId: "",
+							advice: `前置分类器建议不启动顾问群：${classification.reason}`,
+							skipped: true,
+							reason: "classifier-rejected"
+						};
+					}
+				} else if (cfg.trigger.requireClassifier) shadow = {
+					ts: Date.now(),
+					question: args.question,
+					shouldEscalate: true,
+					reason: mentionedAdvisorGroup ? "用户点名 @顾问群，跳过前置分类器。" : `同一问题已重复 ${repeatCount} 次，强制升级跳过前置分类器。`,
+					suggestWebSearch: false,
+					provider: "bypass",
+					bypass: mentionedAdvisorGroup ? "mention" : "repeat",
+					repeatCount
+				};
+				if (args.advisorIds && args.advisorIds.length > 0) {
+					const configuredIds = new Set(cfg.advisors.map((advisor) => advisor.id));
+					if (args.advisorIds.filter((id) => configuredIds.has(id)).length === 0) {
+						recordShadow(false);
+						return {
+							sessionId: "",
+							advice: "未匹配到任何指定顾问，请检查 advisorIds 是否与设置中的顾问 id 一致。",
+							skipped: true,
+							reason: "no-matching-advisors"
+						};
+					}
+				}
+				const costGuard = service.tryStartConsultation();
+				if (!costGuard.ok) {
+					recordShadow(false);
+					return {
+						sessionId: "",
+						advice: costGuard.reason,
+						skipped: true,
+						reason: "daily-limit-reached"
+					};
+				}
+				const session = service.createSession(args.question, args.context, args.advisorIds, sessionLog ? sessionLog.header.cwd : void 0, sessionLog ? sessionLog.id : void 0);
+				if (sessionLog) appendAdvisorStart(sessionLog, session);
+				recordShadow(true);
+				const summary = await service.runAutoPipeline(session, exec.signal, sessionLog, exec.agent);
+				return {
+					sessionId: session.id,
+					advice: formatConversation(session) + (summary.conclusion ? `\n\n【综合结论】\n${summary.conclusion}` : "") + formatRiskNotes(summary) + (service.hasReachedMaxRounds(session) ? "\n\n（已达设置的最大讨论轮数，请参考上述顾问意见与综合结论给出最终答复。）" : ""),
+					skipped: false
+				};
+			}
+		}),
+		defineTool({
+			name: "toggle_advisor_group",
+			description: "Enable or disable the advisor group plugin at runtime. Omit enabled to toggle the current state.",
+			parameters: { enabled: {
+				type: "boolean",
+				description: "Optional desired state."
+			} },
+			output: {
+				schema: {
+					type: "object",
+					additionalProperties: false,
+					properties: { enabled: {
+						type: "boolean",
+						required: true
+					} }
+				},
+				render: (_args, value) => [{
+					type: "text",
+					text: `顾问群插件已${value.enabled ? "启用" : "禁用"}。`
+				}]
+			},
+			async execute(args) {
+				const next = typeof args.enabled === "boolean" ? args.enabled : !service.isEnabled();
+				return { enabled: await service.toggleEnabled(next) };
+			}
+		}),
+		defineTool({
+			name: "ask_session",
+			description: "Ask another EXISTING ordinary session (another window/tab the user opened) a question, wake it, and return what that session said. Use ONLY when the user explicitly asks to ask another session/window (「去问另一个会话」「问一下那个窗口」). It wakes the target session and consumes its quota. The returned text is NOT a one-to-one reply: it is everything the target said between the delivery receipt and its next whole-session idle, so never present it as \"the answer to my message\". When the target title is ambiguous you MUST return the candidates and let the user choose — never guess. Never use it for subagent sessions or for yourself.",
+			parameters: {
+				target: {
+					type: "string",
+					description: "The target session: its full sessionId, its exact display title, or a natural partial name — the workspace/directory name the user mentioned is enough (「极简遥控器」 matches sessions under I:\\极简遥控器\\…). One unique match is used directly; several matches return the candidates (running / most recently active first) for the user to choose — never guess. Pass the name the user said — do NOT search log files or directories for session ids.",
+					required: true
+				},
+				question: {
+					type: "string",
+					description: "The question to hand over. It is permanently recorded in the target session log, so write it self-contained.",
+					required: true
+				},
+				context: {
+					type: "string",
+					description: "Optional background the target needs. Default: none. Anything sent here stays in the target session log forever, so send only what is necessary."
+				},
+				waitMs: {
+					type: "number",
+					description: `How long to wait for the target to finish, in milliseconds (default ${ASK_WAIT_DEFAULT_MS}, max ${ASK_WAIT_MAX_MS}). On timeout the already-collected text is returned and the target is NOT interrupted.`
+				}
+			},
+			output: {
+				schema: {
+					type: "object",
+					additionalProperties: false,
+					properties: {
+						ok: {
+							type: "boolean",
+							required: true
+						},
+						target: {
+							type: "object",
+							additionalProperties: false,
+							properties: {
+								sessionId: {
+									type: "string",
+									required: true
+								},
+								name: {
+									type: "string",
+									required: true
+								}
+							}
+						},
+						delivered: {
+							type: "string",
+							enum: ["native", "fallback"]
+						},
+						askId: { type: "string" },
+						interval: {
+							type: "object",
+							additionalProperties: false,
+							properties: {
+								from: {
+									type: "string",
+									required: true
+								},
+								to: { oneOf: [{ type: "string" }, { type: "null" }] },
+								endedBecause: {
+									type: "string",
+									enum: [
+										"idle",
+										"timeout",
+										"error"
+									],
+									required: true
+								}
+							},
+							required: true
+						},
+						answer: {
+							type: "string",
+							required: true
+						},
+						note: {
+							type: "string",
+							required: true
+						},
+						error: { type: "string" },
+						message: { type: "string" },
+						candidates: {
+							type: "array",
+							items: {
+								type: "object",
+								additionalProperties: false,
+								properties: {
+									sessionId: {
+										type: "string",
+										required: true
+									},
+									name: {
+										type: "string",
+										required: true
+									},
+									hint: {
+										type: "string",
+										required: true
+									}
+								}
+							}
+						}
+					}
+				},
+				render: (_args, value) => value.ok === true ? renderAskResult({
+					answer: String(value.answer ?? ""),
+					note: String(value.note ?? "")
+				}) : [{
+					type: "text",
+					text: `【ask_session 未完成：${String(value.error ?? "unknown")}】\n${String(value.message ?? "")}${Array.isArray(value.candidates) && value.candidates.length > 0 ? `\n候选会话：\n${value.candidates.map((item) => `- ${String(item.name ?? "")} · ${String(item.sessionId ?? "")}`).join("\n")}` : ""}`
+				}]
+			},
+			async execute(args, exec) {
+				const controller = getSessionController(ctx);
+				if (!controller) return {
+					ok: false,
+					error: "service-missing",
+					message: "此功能需要较新的 DSH 宿主：当前宿主没有提供 sessionController 服务，无法向其他会话投递问题。",
+					answer: "",
+					note: "",
+					interval: errorInterval()
+				};
+				const session = exec.agent?.session ?? getCurrentSessionFrom(exec);
+				const sessionId = session?.id;
+				if (typeof sessionId !== "string" || sessionId.length === 0) return {
+					ok: false,
+					error: "service-missing",
+					message: "无法确定当前会话的身份（工具执行上下文没有暴露会话），已拒绝投递。",
+					answer: "",
+					note: "",
+					interval: errorInterval()
+				};
+				const question = typeof args.question === "string" ? args.question : "";
+				const context = typeof args.context === "string" ? args.context : void 0;
+				let senderName = "未命名会话";
+				try {
+					const own = (await controller.list({}, exec.signal)).items.find((item) => item.sessionId === sessionId);
+					if (own) senderName = sessionDisplayName(own);
+				} catch {}
+				const cardLog = session;
+				const onDelivered = (info) => {
+					if (!cardLog) return;
+					try {
+						appendAskStart(cardLog, info);
+					} catch (error) {
+						console.warn("[dsh-advisor-group] 跨会话卡片（开始）写入失败：", error);
+					}
+				};
+				const onSettled = (info) => {
+					if (!cardLog) return;
+					try {
+						appendAskAnswer(cardLog, info);
+					} catch (error) {
+						console.warn("[dsh-advisor-group] 跨会话卡片（结束）写入失败：", error);
+					}
+				};
+				const value = await askSession(ctx, exec.signal, {
+					onDelivered,
+					onSettled,
+					currentSessionId: sessionId,
+					senderName,
+					currentEvents: readCurrentEvents(session),
+					target: typeof args.target === "string" ? args.target : "",
+					question,
+					context,
+					waitMs: args.waitMs,
+					log: (message) => console.log(`[dsh-advisor-group] ask_session: ${message}`)
+				});
+				if (value.ok !== true) return {
+					ok: false,
+					error: String(value.error ?? "delivery-failed"),
+					message: String(value.message ?? ""),
+					...Array.isArray(value.candidates) ? { candidates: value.candidates } : {},
+					answer: "",
+					note: "",
+					interval: errorInterval()
+				};
+				return {
+					ok: true,
+					target: value.target,
+					delivered: value.delivered,
+					askId: String(value.askId ?? ""),
+					interval: value.interval,
+					answer: String(value.answer ?? ""),
+					note: String(value.note ?? "")
+				};
+			}
+		})
+	];
+}
+/**
+* Current session id from the tool execution context. `exec.agent.session` is
+* the Agent on whose behalf the call runs (same access `ask_advisors` uses for
+* `exec.agent?.session`); the fallback covers hosts that only expose a raw
+* session object on the execution input.
+*/
+function getCurrentSessionFrom(exec) {
+	return exec?.session;
+}
+/** Read this session's own log for the inbound-source / hop guard; empty on failure. */
+function readCurrentEvents(session) {
+	try {
+		return session?.snapshotEvents() ?? [];
+	} catch {
+		return [];
+	}
+}
+/** Empty interval carried on every not-ok result (no reply window ever opened). */
+function errorInterval() {
+	return {
+		from: (/* @__PURE__ */ new Date()).toISOString(),
+		endedBecause: "error"
+	};
+}
+function formatRiskNotes(summary) {
+	return summary.riskNotes.length > 0 ? `\n\n【风险提示】\n${summary.riskNotes.join("\n")}` : "";
+}
+function formatConversation(session) {
+	const lines = [`顾问群对话（session: ${session.id}）`, ""];
+	for (const message of session.messages) {
+		if (message.role === "system") continue;
+		if (message.role === "main") lines.push(`主模型：${message.content}`);
+		else if (message.role === "advisor") {
+			const truncatedNote = message.truncated === void 0 ? "" : `（顾问输出在流式过程中被截断：${message.truncated.reason === "timeout" ? "超时" : "网络中断"}，正文可能不完整）\n`;
+			const toolNote = message.toolSteps && message.toolSteps.length > 0 ? `${message.toolSteps.map((step) => step.kind === "call" ? `  ⛭ 调用 ${step.name} · ${step.text}` : `  ↳ ${step.text}`).join("\n")}\n` : "";
+			const body = message.content && message.content.trim() ? truncatedNote + toolNote + message.content : truncatedNote + toolNote + (message.thinking && message.thinking.trim() ? `（思维链）\n${message.thinking}` : "（顾问未返回正文）");
+			lines.push(`【${message.advisorName ?? message.advisorId ?? "顾问"}】：${body}`);
+		}
+		lines.push("");
+	}
+	const stopNote = session.status === "cancelled" && session.stopReason ? `（注意：本次咨询被中断（${STOP_REASON_TEXT[session.stopReason] ?? session.stopReason}）；可在对话卡片点击「▶ 继续聊天」从断点续跑。）` : "";
+	lines.push(stopNote);
+	lines.push("（以上是顾问模型与主模型的对话记录，主模型可直接基于其中内容继续思考或追问。）");
+	return lines.join("\n");
+}
+const STOP_REASON_TEXT = {
+	"user-stop": "用户主动停止",
+	"exec-cancel": "宿主取消",
+	"abort-error": "流异常中断",
+	timeout: "顾问响应超时",
+	network: "网络中断"
+};
+//#endregion
+//#region src/prompt-guard.ts
+const REPLACEMENT = "�";
+const BOM = "";
+/**
+* Unambiguity check: same line count, and every character either equals the
+* template or is a replacement char (the byte-loss corruption signature).
+*/
+function matchesTemplateWithReplacement(storedLines, templateLines) {
+	for (let i = 0; i < storedLines.length; i++) {
+		const stored = storedLines[i];
+		const template = templateLines[i];
+		if (stored === template) continue;
+		if (stored.length !== template.length) return false;
+		for (let j = 0; j < stored.length; j++) if (stored[j] !== template[j] && stored[j] !== REPLACEMENT) return false;
+	}
+	return true;
+}
+/**
+* Advisor system-prompt corruption guard (P1, 2026-09-27).
+* Fast path is free: prompts without replacement chars pass through
+* untouched (a customized prompt is legitimate and must never be overwritten).
+* A leading BOM alone is stripped (action 'stripped', no template involved).
+* Only on a replacement char do we compare against the built-in template:
+* unambiguous (line count + per-character match) -> heal to the template;
+* anything else keeps the stored value and reports it for manual handling.
+*/
+function healAdvisorSystemPrompt(stored, template = DEFAULT_ADVISOR_PROMPT) {
+	let value = stored;
+	let hadBom = false;
+	if (value.startsWith(BOM)) {
+		value = value.slice(1);
+		hadBom = true;
+	}
+	const replacements = (value.match(/\uFFFD/g) ?? []).length;
+	if (replacements === 0) {
+		if (hadBom) return {
+			action: "stripped",
+			value,
+			reason: "开头含 BOM，已剥除"
+		};
+		return {
+			action: "ok",
+			value: stored
+		};
+	}
+	const storedLines = value.split("\n");
+	const templateLines = template.split("\n");
+	if (storedLines.length !== templateLines.length) return {
+		action: "needs-manual",
+		value: stored,
+		reason: `含 ${replacements} 个替换符，行数 ${storedLines.length} 与模板 ${templateLines.length} 不一致，无法无歧义回填`
+	};
+	if (!matchesTemplateWithReplacement(storedLines, templateLines)) return {
+		action: "needs-manual",
+		value: stored,
+		reason: `含 ${replacements} 个替换符，但与模板逐位比对存在不一致（可能含自定义内容），未回填`
+	};
+	return {
+		action: "healed",
+		value: template,
+		reason: `含 ${replacements} 个替换符${hadBom ? "，另开头含 BOM" : ""}，行数与逐位比对均与模板一致，已回填内置模板`
+	};
+}
+/**
+* Load-time guard over every advisor prompt. Returns the original config
+* object untouched when nothing was healed, so callers can rely on identity
+* checks to detect "no change".
+*/
+function guardAdvisorPrompts(config) {
+	const findings = [];
+	let healedAny = false;
+	const advisors = config.advisors.map((advisor) => {
+		const result = healAdvisorSystemPrompt(advisor.systemPrompt);
+		if (result.action === "ok") return advisor;
+		findings.push({
+			advisorId: advisor.id,
+			action: result.action,
+			reason: result.reason ?? ""
+		});
+		if (result.action === "healed" || result.action === "stripped") {
+			healedAny = true;
+			return {
+				...advisor,
+				systemPrompt: result.value
+			};
+		}
+		return advisor;
+	});
+	return {
+		config: healedAny ? {
+			...config,
+			advisors
+		} : config,
+		findings
+	};
+}
+/**
+* Shared logger for both call sites: one line per finding, stating the action
+* actually taken (healed vs left untouched) plus the concrete reason.
+* Every line carries an ISO-8601 timestamp so the host console redirect
+* (which does not timestamp lines itself) stays forensically usable.
+*/
+function reportPromptGuardFindings(findings) {
+	const stamp = () => `[${(/* @__PURE__ */ new Date()).toISOString()}]`;
+	for (const finding of findings) if (finding.action === "healed") console.warn(`${stamp()} [dsh-advisor-group] 顾问提示词损坏，已按内置模板回填：${finding.advisorId}（${finding.reason}）`);
+	else if (finding.action === "stripped") console.warn(`${stamp()} [dsh-advisor-group] 顾问提示词开头含 BOM，已剥除、正文未改动：${finding.advisorId}（${finding.reason}）`);
+	else console.warn(`${stamp()} [dsh-advisor-group] 顾问提示词损坏且无法无歧义回填，未改动、需人工处理：${finding.advisorId}（${finding.reason}）`);
+}
+//#endregion
+//#region src/settings-api.ts
+function sendJson(res, status, body) {
+	res.writeHead(status, { "content-type": "application/json" });
+	res.end(JSON.stringify(body));
+}
+function normalizeJevConfig(input) {
+	const apiKey = typeof input?.apiKey === "string" ? input.apiKey : input?.apiKey && typeof input.apiKey === "object" && "value" in input.apiKey ? String(input.apiKey.value ?? "") : void 0;
+	return {
+		enabled: input?.enabled ?? false,
+		provider: input?.provider ?? "typesafe",
+		model: input?.model ?? "jev-latest",
+		...input?.baseURL ? { baseURL: input.baseURL } : {},
+		...apiKey ? { apiKey } : {},
+		...input?.apiKeyEnv ? { apiKeyEnv: input.apiKeyEnv } : {},
+		timeoutMs: input?.timeoutMs ?? 1e4,
+		confidenceThreshold: input?.confidenceThreshold ?? .6,
+		useEnglishState: input?.useEnglishState ?? false,
+		...typeof input?.highRiskThreshold === "number" ? { highRiskThreshold: input.highRiskThreshold } : {}
+	};
+}
+/** Mask a stored API key so it never leaves the host in plain text. */
+function maskApiKey(apiKey) {
+	if (!apiKey) return apiKey;
+	if (apiKey.length <= 4) return "*".repeat(apiKey.length);
+	return `${"*".repeat(apiKey.length - 4)}${apiKey.slice(-4)}`;
+}
+/**
+* Whether an inbound apiKey is a mask of the previously stored key.
+*
+* Accepts both the current length-preserving mask and the legacy 8-character
+* `****<last4>` mask so an unrefreshed settings page does not overwrite the
+* real key with a mask string after upgrading.
+*/
+function isMaskedApiKey(incoming, previous) {
+	if (!incoming || !previous) return false;
+	if (incoming === maskApiKey(previous)) return true;
+	if (previous.length <= 4) return incoming === "*".repeat(previous.length);
+	return incoming === `****${previous.slice(-4)}`;
+}
+/**
+* Return a copy of the config safe to send to the browser.
+*
+* SecretField convergence: no key material (not even a mask) leaves
+* the host. `apiKey` is returned empty and per-provider presence facts
+* (`configured` + `last4`) ride in `apiKeyMetaByProvider`, so the card can
+* render the official "已配置/未配置" badge and publish new/cleared keys
+* without ever round-tripping a mask.
+*/
+function sanitizeConfig(config) {
+	const jev = normalizeJevConfig(config.trigger.jev);
+	return {
+		...config,
+		trigger: {
+			...config.trigger,
+			jev: {
+				...jev,
+				apiKey: "",
+				...jev?.apiKey ? { apiKeyMeta: {
+					configured: true,
+					last4: jev.apiKey.slice(-4)
+				} } : {}
+			}
+		},
+		advisors: config.advisors.map((advisor) => {
+			const { apiKeysByProvider: _apiKeysByProvider, clearApiKey: _clearApiKey, apiKeyMetaByProvider: _meta, ...rest } = advisor;
+			const meta = {};
+			for (const [provider, key] of Object.entries(advisor.apiKeysByProvider ?? {})) if (key) meta[provider] = {
+				configured: true,
+				last4: key.slice(-4)
+			};
+			if (advisor.apiKey && advisor.provider) meta[advisor.provider] = {
+				configured: true,
+				last4: advisor.apiKey.slice(-4)
+			};
 			return {
 				...rest,
 				apiKey: "",
@@ -2032,8 +4566,19 @@ function hasSameCredentialScope(a, b) {
 */
 function reconcileApiKeys(incoming, previous) {
 	const previousById = new Map(previous.advisors.map((advisor) => [advisor.id, advisor]));
+	const incomingJev = normalizeJevConfig(incoming.trigger.jev);
+	const previousJev = normalizeJevConfig(previous.trigger.jev);
+	const jevApiKey = incomingJev?.apiKey || previousJev?.apiKey;
 	return {
 		...incoming,
+		trigger: {
+			...incoming.trigger,
+			jev: {
+				...incomingJev,
+				...jevApiKey ? { apiKey: jevApiKey } : {},
+				apiKeyMeta: void 0
+			}
+		},
 		advisors: incoming.advisors.map((advisor) => {
 			const { clearApiKey, ...advisorRest } = advisor;
 			const prev = previousById.get(advisor.id);
@@ -2101,7 +4646,14 @@ async function parseLookupParams(req) {
 	return params;
 }
 const ALLOWED_API_KEY_ENVS = new Set(Object.values(PROVIDER_PRESETS).map((preset) => preset.apiKeyEnv));
+ALLOWED_API_KEY_ENVS.add("TYPESAFE_API_KEY");
+ALLOWED_API_KEY_ENVS.add("OPENROUTER_API_KEY");
 function validateSecurity(input) {
+	if (input.trigger.jev?.baseURL) {
+		const url = input.trigger.jev.baseURL.trim();
+		if (!(url.startsWith("https://") || url.startsWith("http://127.0.0.1") || url.startsWith("http://localhost"))) return "Jev API 地址仅允许 https:// 或本机 http://127.0.0.1 / http://localhost";
+	}
+	if (input.trigger.jev?.apiKeyEnv && !/^[A-Za-z_][A-Za-z0-9_]*$/.test(input.trigger.jev.apiKeyEnv.trim())) return `Jev API Key 环境变量名无效：${input.trigger.jev.apiKeyEnv}`;
 	const ids = /* @__PURE__ */ new Set();
 	for (const advisor of input.advisors) {
 		if (ids.has(advisor.id)) return `顾问 id 重复：${advisor.id}`;
@@ -2293,10 +4845,21 @@ async function handleModelsRequest(ctx, service, req, res) {
 	} catch (error) {
 		sendJson(res, 400, {
 			ok: false,
-			error: error instanceof Error ? error.message : String(error)
+			error: describeConnectionError(error)
 		});
 	}
 }
+/** 网络/HTTP 类失败的设置页可读文案；原始信息保留在括号里。 */
+function describeConnectionError(error) {
+	const message = error instanceof Error ? error.message : String(error);
+	if (/HTTP (401|403)\b/i.test(message)) return `密钥无效或没有访问权限（${message}）`;
+	if (/HTTP 404\b/i.test(message)) return `地址不存在，检查 baseURL 是否写对（${message}）`;
+	if (/ENOTFOUND|EAI_AGAIN/i.test(message)) return `找不到这个地址，检查域名拼写与网络（${message}）`;
+	if (/ECONNREFUSED/i.test(message)) return `对方拒绝了连接，检查地址与端口（${message}）`;
+	if (/ETIMEDOUT|\btimeout\b|aborted/i.test(message)) return `连接超时（${message}）`;
+	if (/fetch failed/i.test(message)) return `网络连接失败，检查网络后重试（${message}）`;
+	return `连接失败（${message}）`;
+}
 async function handleTestConnectionRequest(ctx, service, req, res) {
 	const params = await parseLookupParams(req);
 	const provider = params.provider ?? "";
@@ -2373,7 +4936,7 @@ async function handleTestConnectionRequest(ctx, service, req, res) {
 	} catch (error) {
 		sendJson(res, 400, {
 			ok: false,
-			error: error instanceof Error ? error.message : String(error)
+			error: describeConnectionError(error)
 		});
 	}
 }
@@ -2407,12 +4970,95 @@ function handleStreamRequest(req, res) {
 		subscribe(sessionId, res, lastEventId, bootId);
 	});
 }
+/** First value of a possibly repeated header field. */
+function headerValue(value) {
+	return Array.isArray(value) ? value[0] ?? "" : value ?? "";
+}
+/** Origin of an http(s) URL; `''` for absent, opaque (`null`) or non-web values. */
+function webOrigin(value) {
+	if (value === "" || value === "null") return "";
+	try {
+		const parsed = new URL(value);
+		return parsed.protocol === "http:" || parsed.protocol === "https:" ? parsed.origin : "";
+	} catch {
+		return "";
+	}
+}
+/**
+* True when the handshake request may be handed the token. Only a *provably
+* foreign web page* is refused: an `Origin`/`Referer` that parses as http(s) on
+* another host. Same-origin pages pass, and so do callers carrying no web
+* provenance at all — the desktop shell, which forwards the page's request
+* without those headers, and any local program, which could forge them anyway.
+*
+* What actually keeps the token from a foreign page is not this check but the
+* absent CORS header: a cross-origin fetch cannot read a response body without
+* `Access-Control-Allow-Origin`, and this route never sends one. The check only
+* decides whether a foreign page is answered at all.
+*/
+function mayReadBootToken(req) {
+	const host = headerValue(req.headers.host);
+	const fromForeignPage = (value) => {
+		const origin = webOrigin(value);
+		if (origin === "" || host === "") return false;
+		return origin !== `http://${host}` && origin !== `https://${host}`;
+	};
+	if (fromForeignPage(headerValue(req.headers.origin))) return false;
+	if (fromForeignPage(headerValue(req.headers.referer))) return false;
+	return true;
+}
+/** Set once the process has logged an accepted handshake (refusals always log). */
+let bootTokenAcceptedLogged = false;
+/**
+* Records one handshake attempt in
+* `$DSH_HOME/storages/advisor-group/boot-token.log`. Whether the desktop shell's
+* request even reaches this route is otherwise invisible from outside the app,
+* so each decision leaves a line; the file keeps only its most recent lines.
+* Query strings are dropped and the referer is reduced to its origin, so no
+* token can land in the log. A write failure never breaks the handshake.
+*/
+function recordBootTokenAttempt(req, allowed) {
+	if (allowed && bootTokenAcceptedLogged) return;
+	bootTokenAcceptedLogged = true;
+	try {
+		const dshHome = process.env.DSH_HOME || join(homedir(), ".dsh");
+		const target = join(dshHome, "storages", "advisor-group", "boot-token.log");
+		const line = `${[
+			(/* @__PURE__ */ new Date()).toISOString(),
+			allowed ? "allow" : "deny",
+			`site=${headerValue(req.headers["sec-fetch-site"]) || "-"}`,
+			`origin=${headerValue(req.headers.origin) || "-"}`,
+			`referer=${webOrigin(headerValue(req.headers.referer)) || "-"}`,
+			`host=${headerValue(req.headers.host) || "-"}`,
+			`ua=${headerValue(req.headers["user-agent"]).slice(0, 48) || "-"}`
+		].join(" ")}\n`;
+		mkdirSync(dirname(target), { recursive: true });
+		appendFileSync(target, line);
+		if (statSync(target).size > 8192) writeFileSync(target, readFileSync(target, "utf8").split("\n").slice(-20).join("\n"));
+	} catch {}
+}
 function handleConfigRequest(ctx, scopedCtx, service, authToken) {
-	const namespace = "advisor-group";
+	const namespace = "dsh-advisor-group";
 	return async (req, res) => {
 		const rawUrl = req.url ?? "";
 		const url = new URL(rawUrl, "http://localhost");
 		const pathname = url.pathname.replace(/\/+$/, "") || "/";
+		if (req.method === "GET" && (pathname === "/advisor-group/boot-token" || pathname === "/advisor-group/boot-token/")) {
+			const allowed = mayReadBootToken(req);
+			recordBootTokenAttempt(req, allowed);
+			if (!allowed) {
+				sendJson(res, 403, {
+					ok: false,
+					error: "forbidden"
+				});
+				return;
+			}
+			sendJson(res, 200, {
+				ok: true,
+				token: authToken
+			});
+			return;
+		}
 		const token = String(req.headers["x-advisor-group-token"] ?? url.searchParams.get("token") ?? "");
 		if (authToken && token !== authToken) {
 			sendJson(res, 401, {
@@ -2437,6 +5083,55 @@ function handleConfigRequest(ctx, scopedCtx, service, authToken) {
 			await sendProviders(ctx, res);
 			return;
 		}
+		if (req.method === "POST" && (pathname === "/advisor-group/stop" || pathname === "/advisor-group/stop/")) {
+			try {
+				const raw = await readBody(req);
+				const parsed = JSON.parse(raw);
+				if (typeof parsed.sessionId !== "string" || !parsed.sessionId) {
+					sendJson(res, 400, {
+						ok: false,
+						error: "缺少 sessionId"
+					});
+					return;
+				}
+				sendJson(res, 200, {
+					ok: true,
+					stopped: service.stopConsultation(parsed.sessionId)
+				});
+			} catch (error) {
+				console.warn("[dsh-advisor-group] stop 请求处理失败：", error instanceof Error ? error.message : String(error));
+				sendJson(res, 400, {
+					ok: false,
+					error: error instanceof Error ? error.message : String(error)
+				});
+			}
+			return;
+		}
+		if (req.method === "POST" && (pathname === "/advisor-group/resume" || pathname === "/advisor-group/resume/")) {
+			try {
+				const raw = await readBody(req);
+				const parsed = JSON.parse(raw);
+				if (typeof parsed.sessionId !== "string" || !parsed.sessionId) {
+					sendJson(res, 400, {
+						ok: false,
+						error: "缺少 sessionId"
+					});
+					return;
+				}
+				const result = service.resumeConsultation(parsed.sessionId);
+				sendJson(res, result.ok ? 200 : 400, {
+					ok: result.ok,
+					...result.reason ? { error: result.reason } : {}
+				});
+			} catch (error) {
+				console.warn("[dsh-advisor-group] resume 请求处理失败：", error instanceof Error ? error.message : String(error));
+				sendJson(res, 400, {
+					ok: false,
+					error: error instanceof Error ? error.message : String(error)
+				});
+			}
+			return;
+		}
 		if (req.method === "GET" && (pathname === "/advisor-group/shadow" || pathname === "/advisor-group/shadow/")) {
 			sendJson(res, 200, {
 				ok: true,
@@ -2445,12 +5140,20 @@ function handleConfigRequest(ctx, scopedCtx, service, authToken) {
 			return;
 		}
 		if (req.method === "GET" && (pathname === "/advisor-group/config" || pathname === "/advisor-group/config/")) {
-			sendJson(res, 200, {
-				ok: true,
-				config: sanitizeConfig(service.getConfig()),
-				revision: currentRevision(scopedCtx, namespace),
-				dailyGuard: service.getDailyGuard()
-			});
+			try {
+				sendJson(res, 200, {
+					ok: true,
+					config: sanitizeConfig(service.getConfig()),
+					revision: currentRevision(scopedCtx, namespace),
+					dailyGuard: service.getDailyGuard()
+				});
+			} catch (error) {
+				console.error("[dsh-advisor-group] config read failed:", error instanceof Error ? error.message : String(error));
+				sendJson(res, 500, {
+					ok: false,
+					error: error instanceof Error ? error.message : String(error)
+				});
+			}
 			return;
 		}
 		if (req.method === "POST" && (pathname === "/advisor-group/config" || pathname === "/advisor-group/config/")) {
@@ -2459,7 +5162,7 @@ function handleConfigRequest(ctx, scopedCtx, service, authToken) {
 				const parsed = JSON.parse(raw);
 				const configInput = parsed.config ?? parsed;
 				const expectedRevision = typeof parsed.expectedRevision === "number" ? parsed.expectedRevision : currentRevision(scopedCtx, namespace);
-				const reconciled = reconcileApiKeys(Config(configInput), service.getConfig());
+				const reconciled = reconcileApiKeys(unwrapVolatileConfig(Config(configInput)), service.getConfig());
 				const securityError = validateSecurity(reconciled);
 				if (securityError) {
 					sendJson(res, 400, {
@@ -2468,8 +5171,10 @@ function handleConfigRequest(ctx, scopedCtx, service, authToken) {
 					});
 					return;
 				}
-				await scopedCtx.settings.replace(namespace, reconciled, expectedRevision);
-				service.setConfig(reconciled);
+				const guarded = guardAdvisorPrompts(reconciled);
+				reportPromptGuardFindings(guarded.findings);
+				await scopedCtx.settings.replace(namespace, guarded.config, expectedRevision);
+				service.setConfig(guarded.config);
 				sendJson(res, 200, {
 					ok: true,
 					config: sanitizeConfig(service.getConfig()),
@@ -2485,7 +5190,7 @@ function handleConfigRequest(ctx, scopedCtx, service, authToken) {
 				}
 				sendJson(res, 400, {
 					ok: false,
-					error: error instanceof Error ? error.message : String(error)
+					error: `保存失败：请检查刚才修改项的格式（原始信息：${error instanceof Error ? error.message : String(error)}）`
 				});
 			}
 			return;
@@ -2533,14 +5238,16 @@ function registerAdvisorSettingsAndRoutes(ctx, config, service) {
 	if (!inject) return;
 	inject(["settings", "webServer"], (scoped) => {
 		const scopedCtx = scoped;
-		const namespace = "advisor-group";
-		const scope = scopedCtx.settings.register(namespace, Config, { base: config });
-		service.setConfig(scope.get());
+		const namespace = "dsh-advisor-group";
+		service.setConfig(config);
+		const presentationDisposer = scopedCtx.settings.configure({ auto: false }, ctx.fiber);
 		service.setPersistEnabled(async (enabled) => {
 			await scopedCtx.settings.update(namespace, { enabled });
 		});
-		const watchDisposer = scope.watch((next) => {
-			service.setConfig(next);
+		const watchDisposer = ctx.on("loader/volatile-update", () => {
+			const guarded = guardAdvisorPrompts(unwrapVolatileConfig(ctx.fiber.config));
+			reportPromptGuardFindings(guarded.findings);
+			service.setConfig(guarded.config);
 		});
 		const authToken = randomUUID();
 		const indexTapDisposer = scopedCtx.webServer.tapIndex ? scopedCtx.webServer.tapIndex((html) => html.replace("<head>", `<head><script>globalThis.__ADVISOR_GROUP_TOKEN__ = ${JSON.stringify(authToken)}<\/script>`)) : null;
@@ -2553,6 +5260,7 @@ function registerAdvisorSettingsAndRoutes(ctx, config, service) {
 			watchDisposer();
 			routeDisposer();
 			indexTapDisposer?.();
+			presentationDisposer();
 		}, "dsh-advisor-group: settings watcher + config routes");
 	});
 }
@@ -2598,7 +5306,7 @@ function buildAdvisorPrompt(config, repeatPressure) {
 /**
 * Host-only registration for the advisor-group session event vocabulary.
 *
-* The official 0.1.2-rc.1 compatibility mechanism is the per-event
+* The official compatibility mechanism is the per-event
 * `SessionEvent.ignorable: true` envelope marker: `dsh-session-persistence`
 * refuses unknown event types unless `KNOWN_SESSION_EVENT_TYPES.has(type)`
 * or the stored event is marked ignorable. However `Session.append()` exposes
@@ -2615,7 +5323,8 @@ const ADVISOR_GROUP_EVENT_TYPES = [
 	"advisor-group/start",
 	"advisor-group/message",
 	"advisor-group/delta",
-	"advisor-group/end"
+	"advisor-group/end",
+	"advisor-group/resume"
 ];
 function ensureAdvisorGroupSessionEventTypes() {
 	const known = KNOWN_SESSION_EVENT_TYPES;
@@ -2630,12 +5339,15 @@ const name = "dsh-advisor-group";
 const inject = [
 	"tools",
 	"llm",
-	"systemPrompt"
+	"systemPrompt",
+	"sessions"
 ];
 function apply(ctx, config) {
-	const service = new AdvisorGroupService(ctx, config);
-	registerAdvisorSettingsAndRoutes(ctx, config, service);
-	for (const tool of registerAdvisorTools(service)) ctx.tools.register(tool);
+	const guarded = guardAdvisorPrompts(unwrapVolatileConfig(config));
+	reportPromptGuardFindings(guarded.findings);
+	const service = new AdvisorGroupService(ctx, guarded.config);
+	registerAdvisorSettingsAndRoutes(ctx, guarded.config, service);
+	for (const tool of registerAdvisorTools(service, ctx)) ctx.tools.register(tool);
 	const systemPrompt = ctx.systemPrompt;
 	if (systemPrompt) ctx.effect(() => systemPrompt.section({
 		name: "advisor-group-boundary",