dsh-llm-verifier 0.1.13 → 0.1.14

package/lib/index.js

@@ -31,8 +31,10 @@ const Config = z.object({
 	autoRouteMaxInputChars: z.number().step(1).min(1e3).default(6e4),
 	autoMaxModelCallsPerTask: z.number().step(1).min(1).default(48),
 	autoMaxModelCallsPerSession: z.number().step(1).min(1).default(160),
+	autoVerifyTeamTasks: z.boolean().default(true),
+	autoVerifyPlanMode: z.boolean().default(true),
 	provider: z.string().default("deepseek-official"),
-	model: z.string().default("deepseek-v4-flash"),
+	model: z.string().default("deepseek-flash"),
 	reasoningEffort: z.string(),
 	maxTokens: z.number().step(1).min(1).default(32768),
 	timeoutMs: z.number().step(1).min(1).default(3e5),
@@ -46,7 +48,7 @@ const Config = z.object({
 });
 function resolveConfig(config = {}) {
 	const provider = (config.provider ?? "deepseek-official").trim();
-	const model = (config.model ?? "deepseek-v4-flash").trim();
+	const model = (config.model ?? "deepseek-flash").trim();
 	if (!provider) throw new Error("llm-verifier: provider must be non-empty");
 	if (!model) throw new Error("llm-verifier: model must be non-empty");
 	const autoVerifyMode = config.autoVerifyMode ?? "smart";
@@ -100,6 +102,8 @@ function resolveConfig(config = {}) {
 		autoRouteSemantic: config.autoRouteSemantic ?? true,
 		autoRouteMinConfidence,
 		autoTrackCompletionThreshold,
+		autoVerifyTeamTasks: config.autoVerifyTeamTasks ?? true,
+		autoVerifyPlanMode: config.autoVerifyPlanMode ?? true,
 		provider,
 		model,
 		...reasoningEffort ? { reasoningEffort } : {},
@@ -604,10 +608,17 @@ async function loadVerifierImages(inputs, signal) {
 //#region src/session.ts
 function textOf(blocks) {
 	const parts = [];
-	for (const block of blocks) if (block.type === "text") parts.push(block.text);
-	else if (block.type === "reasoning") parts.push("[Reasoning] " + block.text);
-	else if (block.type === "tool-call") parts.push("[Tool Call] " + block.name + " " + block.arguments);
-	else if (block.type === "tool-result") parts.push("[Tool Result] " + textOf(block.content));
+	for (const block of blocks) {
+		const blockType = block.type;
+		if (blockType === "text") parts.push(block.text);
+		else if (blockType === "reasoning") parts.push("[Reasoning] " + block.text);
+		else if (blockType === "tool-call") parts.push("[Tool Call] " + block.name + " " + block.arguments);
+		else if (blockType === "tool-result") parts.push("[Tool Result] " + textOf(block.content));
+		else if (blockType === "file") {
+			const fileData = block;
+			parts.push("[File] " + (fileData.path ?? fileData.filename ?? fileData.title ?? "attachment"));
+		}
+	}
 	return parts.join("\n");
 }
 const DEFAULT_REDACT_PATTERNS = ["Bearer\\s+[A-Za-z0-9._~+\\/=-]+", "(?:api[_-]?key|token|password|secret)\\s*[=:]\\s*[\"']?[^\\s,\"';}]+"];
@@ -643,24 +654,35 @@ async function extractSession(agent, loadImage, options = {}) {
 	let problem = "";
 	const trace = [];
 	const images = [];
-	for (const event of events) if (event.type === "user/message") {
-		if (event.data.source.kind !== "user") continue;
-		const text = textOf(event.data.content);
-		if (!problem && text.trim()) problem = text.trim();
-		for (const block of event.data.content) if (block.type === "image") images.push(await loadImage(block.attachment));
-		trace.push("--- User seq " + event.seq + " ---\n" + text);
-	} else if (event.type === "assistant/message" && options.includeAssistantText !== false) trace.push("--- Assistant turn " + event.data.turn + " step " + event.data.step + " ---\n" + textOf(event.data.message.content));
-	else if (event.type === "tool/call") trace.push("--- Tool Call turn " + event.data.turn + " step " + event.data.step + " ---\n[Command] " + event.data.name + " " + event.data.arguments);
-	else if (event.type === "tool/result") trace.push("--- Tool Result turn " + event.data.turn + " step " + event.data.step + " ---\n[Output] " + textOf(event.data.message.content));
-	else if (event.type === "tool/code-dispatch") {
-		const data = event.data;
-		const status = data.isError ? " [Error]" : "";
-		const args = data.arguments !== void 0 ? " " + (typeof data.arguments === "string" ? data.arguments : JSON.stringify(data.arguments)) : "";
-		const content = Array.isArray(data.content) ? textOf(data.content) : "";
-		if (Array.isArray(data.content)) {
-			for (const block of data.content) if (block.type === "image") images.push(await loadImage(block.attachment));
+	for (const rawEvent of events) {
+		const event = rawEvent;
+		if (event.type === "user/message") {
+			const sourceKind = event.data?.source?.kind;
+			if (sourceKind !== "user" && sourceKind !== "team-message") continue;
+			const text = textOf(event.data.content);
+			if (!problem && sourceKind === "user" && text.trim()) problem = text.trim();
+			for (const block of event.data.content) if (block.type === "image") images.push(await loadImage(block.attachment));
+			const tag = sourceKind === "team-message" ? "Team Message" : "User";
+			trace.push("--- " + tag + " seq " + event.seq + " ---\n" + text);
+		} else if (event.type === "assistant/message" && options.includeAssistantText !== false) trace.push("--- Assistant turn " + event.data.turn + " step " + event.data.step + " ---\n" + textOf(event.data.message.content));
+		else if (event.type === "tool/call") trace.push("--- Tool Call turn " + event.data.turn + " step " + event.data.step + " ---\n[Command] " + event.data.name + " " + event.data.arguments);
+		else if (event.type === "tool/result") trace.push("--- Tool Result turn " + event.data.turn + " step " + event.data.step + " ---\n[Output] " + textOf(event.data.message.content));
+		else if (event.type === "tool/ptc-dispatch" || event.type === "tool/code-dispatch") {
+			const data = event.data;
+			const status = data.isError ? " [Error]" : "";
+			const args = data.arguments !== void 0 ? " " + (typeof data.arguments === "string" ? data.arguments : JSON.stringify(data.arguments)) : "";
+			const content = Array.isArray(data.content) ? textOf(data.content) : "";
+			if (Array.isArray(data.content)) {
+				for (const block of data.content) if (block.type === "image") images.push(await loadImage(block.attachment));
+			}
+			const label = event.type === "tool/ptc-dispatch" ? "PTC Dispatch " : "Code Dispatch ";
+			trace.push("--- " + label + data.name + status + " seq " + event.seq + " ---\n[Command] " + data.name + args + "\n[Output] " + content);
+		} else if (event.type === "team/message/queued") {
+			const data = event.data;
+			const sender = data?.message?.senderName ?? "Teammate";
+			const content = Array.isArray(data?.message?.content) ? textOf(data.message.content) : "";
+			trace.push("--- Team Message Queued from " + sender + " seq " + event.seq + " ---\n" + content);
 		}
-		trace.push("--- Code Dispatch " + data.name + status + " seq " + event.seq + " ---\n[Command] " + data.name + args + "\n[Output] " + content);
 	}
 	const raw = redactText(trace.join("\n\n"), patterns);
 	const maxChars = options.maxChars ?? 2e5;
@@ -895,21 +917,32 @@ function buildEvidenceIndex(events) {
 	const calls = /* @__PURE__ */ new Map();
 	const results = /* @__PURE__ */ new Map();
 	const todos = /* @__PURE__ */ new Map();
+	const teamTasks = /* @__PURE__ */ new Map();
 	const paired = /* @__PURE__ */ new Map();
-	for (const event of relevant) if (event.type === "tool/call") calls.set(String(event.data.callId), event);
-	else if (event.type === "tool/result" && successful(event)) results.set(String(event.data.message.source.callId), event);
-	else if (event.type === "todo/write") todos.set(event.seq, event.data.todos);
-	else if (event.type === "tool/code-dispatch") {
-		const data = event.data;
-		if (data.isError !== true && (!Array.isArray(data.content) || data.content.every((b) => b.isError !== true))) {
-			const subCallId = String(data.subCallId ?? "code:" + event.seq);
-			const content = Array.isArray(data.content) ? data.content : [];
-			paired.set(subCallId, {
-				name: data.name,
-				callSeq: event.seq,
-				resultSeq: event.seq,
-				text: blockText(content)
-			});
+	const currentTeamTasks = /* @__PURE__ */ new Map();
+	for (const rawEvent of relevant) {
+		const event = rawEvent;
+		if (event.type === "tool/call") calls.set(String(event.data.callId), rawEvent);
+		else if (event.type === "tool/result" && successful(rawEvent)) results.set(String(event.data.message.source.callId), rawEvent);
+		else if (event.type === "todo/write") todos.set(event.seq, event.data.todos);
+		else if (event.type === "team/task") {
+			const data = event.data;
+			if (data?.task) {
+				currentTeamTasks.set(data.task.id, { ...data.task });
+				teamTasks.set(event.seq, [...currentTeamTasks.values()]);
+			}
+		} else if (event.type === "tool/ptc-dispatch" || event.type === "tool/code-dispatch") {
+			const data = event.data;
+			if (data.isError !== true && (!Array.isArray(data.content) || data.content.every((b) => b.isError !== true))) {
+				const subCallId = String(data.subCallId ?? "code:" + event.seq);
+				const content = Array.isArray(data.content) ? data.content : [];
+				paired.set(subCallId, {
+					name: data.name,
+					callSeq: event.seq,
+					resultSeq: event.seq,
+					text: blockText(content)
+				});
+			}
 		}
 	}
 	for (const [callId, call] of calls) {
@@ -924,7 +957,8 @@ function buildEvidenceIndex(events) {
 	return {
 		problemSeq: taskStartSeq,
 		calls: paired,
-		todos
+		todos,
+		teamTasks
 	};
 }
 function parseTrustedWorkflow(value, callId, callSeq, resultSeq, maxCandidates, maxItemChars) {
@@ -978,6 +1012,23 @@ function canonicalTodoSnapshots(index) {
 	}
 	return values;
 }
+function canonicalTeamTaskSnapshots(index) {
+	const values = [];
+	let previous = "";
+	for (const [seq, tasks] of index.teamTasks) {
+		const canonical = JSON.stringify(tasks.map((t) => ({
+			id: t.id,
+			status: t.status,
+			revision: t.revision
+		})));
+		if (canonical !== previous) values.push({
+			seq,
+			tasks
+		});
+		previous = canonical;
+	}
+	return values;
+}
 function analyzeStructuredRoute(events, maxCandidates = 8, maxItemChars = 2e4) {
 	const index = buildEvidenceIndex(events);
 	if (!index) return void 0;
@@ -1031,13 +1082,32 @@ function analyzeStructuredRoute(events, maxCandidates = 8, maxItemChars = 2e4) {
 				evidenceSeqs: snapshots.map((snapshot) => snapshot.seq)
 			};
 		}
+		const teamSnapshots = canonicalTeamTaskSnapshots(index);
+		if (teamSnapshots.length >= 2) {
+			const steps = teamSnapshots.map((snapshot) => sanitizeVerifierText("Team task checkpoint seq " + snapshot.seq + ":\n" + snapshot.tasks.map((task) => "- [" + task.status + "] " + task.subject + (task.description ? " (" + task.description + ")" : "")).join("\n"), maxItemChars));
+			const checkpoints = steps.map((_, i) => i + 1);
+			return {
+				kind: "track",
+				source: "structured",
+				confidence: 1,
+				reason: "changed durable team tasks",
+				fingerprint: stableHash({
+					kind: "track",
+					teamSnapshots
+				}),
+				steps,
+				checkpoints,
+				evidenceSeqs: teamSnapshots.map((snapshot) => snapshot.seq)
+			};
+		}
 	}
 }
 function semanticRouteHint(events) {
 	const index = buildEvidenceIndex(events);
 	if (!index) return false;
-	if ([...index.calls.values()].some((pair) => pair.name === "subagent" || pair.name === "subagent_fork" || pair.name === "workflow")) return true;
+	if ([...index.calls.values()].some((pair) => pair.name === "subagent" || pair.name === "subagent_fork" || pair.name === "workflow" || pair.name === "exit_plan_mode")) return true;
 	if (canonicalTodoSnapshots(index).length >= 2) return true;
+	if (canonicalTeamTaskSnapshots(index).length >= 2) return true;
 	return false;
 }
 function buildSemanticRoutePrompt(problem, events, maxCandidates, maxItemChars = 2e4) {
@@ -1232,6 +1302,124 @@ var AutoVerifierRouter = class {
 	}
 };
 //#endregion
+//#region src/plan-gate.ts
+function detectPlanExit(events, fromSeq = 0) {
+	for (let i = events.length - 1; i >= 0; i--) {
+		const event = events[i];
+		if (event.seq < fromSeq) break;
+		if (event.type === "tool/call" && event.data.name === "exit_plan_mode") {
+			let planText = "";
+			try {
+				const parsed = JSON.parse(event.data.arguments);
+				if (typeof parsed.plan === "string" && parsed.plan.trim()) planText = parsed.plan.trim();
+			} catch {}
+			if (!planText) for (let j = i - 1; j >= 0; j--) {
+				const prev = events[j];
+				if (prev.seq < fromSeq) break;
+				if (prev.type === "assistant/message") {
+					const msg = prev.data.message;
+					if (Array.isArray(msg?.content)) {
+						const texts = msg.content.filter((b) => b.type === "text").map((b) => b.text);
+						if (texts.length > 0) {
+							planText = texts.join("\n").trim();
+							break;
+						}
+					}
+				}
+			}
+			return {
+				hasExitPlanMode: true,
+				callSeq: event.seq,
+				planText
+			};
+		}
+	}
+	return { hasExitPlanMode: false };
+}
+function buildPlanPreReviewPrompt(problem, planText, maxChars = 2e4) {
+	return [
+		"You are an expert independent technical plan verifier. A candidate implementation plan was submitted for human review.",
+		"Evaluate whether the plan is sound, executable, and comprehensive.",
+		"Scoring criteria:",
+		"A: Flawless, thorough, with clear steps, rigorous verification, edge cases handled.",
+		"B-J: Mostly sound, minor gaps or nuances missing.",
+		"K-S: Noticeable architectural gaps, missing regression checks, or ambiguities.",
+		"T: Fundamentally flawed, dangerous, or missing core requirements.",
+		"",
+		"Task requirement:",
+		sanitizeVerifierText(problem, 4e3),
+		"",
+		"Proposed Plan:",
+		sanitizeVerifierText(planText, maxChars),
+		"",
+		"Output format:",
+		"Line 1: Verdict: <Single uppercase letter A-T>",
+		"Line 2: Summary: <One sentence assessment>",
+		"Line 3+: Key strengths, blind spots, and verification guidance."
+	].join("\n\n");
+}
+function parsePlanReviewVerdict(text) {
+	const match = /Verdict:\s*([A-T])/i.exec(text);
+	const verdict = match ? match[1].toUpperCase() : "B";
+	const index = verdict.charCodeAt(0) - 65;
+	return {
+		verdict,
+		score: Math.max(0, Math.min(1, 1 - index / 19)),
+		feedback: text.trim()
+	};
+}
+//#endregion
+//#region src/team-gate.ts
+function inspectTeamTasks(events, fromSeq = 0) {
+	let latestCompletedTask;
+	let completedSeq;
+	const taskMap = /* @__PURE__ */ new Map();
+	for (const rawEvent of events) {
+		const event = rawEvent;
+		if (event.type === "team/task") {
+			const data = event.data;
+			if (data?.task) {
+				taskMap.set(data.task.id, { ...data.task });
+				if (event.seq >= fromSeq && data.task.status === "completed") {
+					latestCompletedTask = { ...data.task };
+					completedSeq = event.seq;
+				}
+			}
+		}
+	}
+	return {
+		latestCompletedTask,
+		completedSeq,
+		activeTasks: [...taskMap.values()].filter((t) => t.status !== "deleted"),
+		hasRecentCompletedTask: latestCompletedTask !== void 0
+	};
+}
+function buildTeamTaskVerificationPrompt(task, executionTrace, maxChars = 2e4) {
+	return [
+		"You are an expert independent technical verifier reviewing a completed Agent Teams task.",
+		"Verify if the task goal has been satisfied by concrete evidence (code edits, test runs, successful commands).",
+		"Task Details:",
+		"ID: " + task.id,
+		"Subject: " + task.subject,
+		task.description ? "Description: " + task.description : "",
+		"",
+		"Execution Evidence & Trace:",
+		sanitizeVerifierText(executionTrace, maxChars),
+		"",
+		"Evaluate whether the task is genuinely completed with verified proof.",
+		"Scoring criteria:",
+		"A: Completely fulfilled with solid execution/test evidence.",
+		"B-J: Mostly complete, minor gaps or edge cases not fully asserted.",
+		"K-S: Incomplete, missing key requirements or untested.",
+		"T: Unfulfilled, failed, or contradictory evidence.",
+		"",
+		"Output format:",
+		"Line 1: Verdict: <Single uppercase letter A-T>",
+		"Line 2: Summary: <One sentence evaluation of task fulfillment>",
+		"Line 3+: Remaining gaps or next steps for the team."
+	].filter(Boolean).join("\n\n");
+}
+//#endregion
 //#region src/statistics.ts
 const VERIFIER_TOOL_NAMES = [
 	"verifier_route_classify",
@@ -2010,8 +2198,7 @@ function apply(ctx, config = {}) {
 			summary: "Automatic verifier routed " + decision.kind
 		}
 	});
-	services.connection.rpc.handle("/llm-verifier", async (endpoint, payload) => {
-		if (endpoint !== "statistics") return rpcFailure("unknown llm-verifier endpoint");
+	const handleStatisticsQuery = async (payload) => {
 		if (typeof payload !== "object" || payload === null) return rpcFailure("statistics payload must be an object");
 		const row = payload;
 		const sessionId = typeof row.sessionId === "string" && row.sessionId.length > 0 ? row.sessionId : void 0;
@@ -2028,7 +2215,87 @@ function apply(ctx, config = {}) {
 		} catch (error) {
 			return rpcFailure(error instanceof Error ? error.message : String(error));
 		}
-	}, { authority: "loopback" });
+	};
+	const anyConn = services.connection;
+	if (anyConn && anyConn.fetch && typeof anyConn.fetch.register === "function") try {
+		anyConn.fetch.register({
+			path: "/api/llm-verifier/statistics",
+			methods: ["POST"],
+			requestBody: "buffered",
+			fetch: async (request) => {
+				let body;
+				try {
+					body = await request.json();
+				} catch {
+					return new Response("body is not JSON", { status: 400 });
+				}
+				const isRpcEnvelope = typeof body === "object" && body !== null && body.type === "client-request" && typeof body.rpcId === "string";
+				const rpcId = isRpcEnvelope ? body.rpcId : "direct";
+				const payload = isRpcEnvelope ? body.payload : body;
+				const outcome = await handleStatisticsQuery(payload);
+				if (isRpcEnvelope) return Response.json({
+					type: "server-response",
+					rpcId,
+					result: outcome
+				});
+				return Response.json(outcome);
+			}
+		});
+	} catch (e) {
+		ctx.logger.warn("failed to register /api/llm-verifier/statistics fetch route: " + String(e));
+	}
+	const mountWebServerRoute = (ws) => {
+		if (!ws || typeof ws.register !== "function") return;
+		try {
+			ws.register({
+				kind: "prefix",
+				path: "/llm-verifier",
+				handler: (req, res) => {
+					if (req.method !== "POST") {
+						res.writeHead(405);
+						res.end();
+						return;
+					}
+					let bodyStr = "";
+					req.on("data", (chunk) => {
+						bodyStr += chunk.toString();
+					});
+					req.on("end", async () => {
+						try {
+							const body = JSON.parse(bodyStr);
+							const isRpcEnvelope = typeof body === "object" && body !== null && body.type === "client-request" && typeof body.rpcId === "string";
+							const rpcId = isRpcEnvelope ? body.rpcId : "direct";
+							const payload = isRpcEnvelope ? body.payload : body;
+							const outcome = await handleStatisticsQuery(payload);
+							res.writeHead(200, { "content-type": "application/json" });
+							if (isRpcEnvelope) res.end(JSON.stringify({
+								type: "server-response",
+								rpcId,
+								result: outcome
+							}));
+							else res.end(JSON.stringify(outcome));
+						} catch (err) {
+							res.writeHead(400, { "content-type": "application/json" });
+							res.end(JSON.stringify({ error: String(err) }));
+						}
+					});
+				}
+			});
+		} catch {}
+	};
+	if (typeof ctx.inject === "function") try {
+		ctx.inject(["webServer"], (webCtx) => {
+			if (webCtx?.webServer) mountWebServerRoute(webCtx.webServer);
+		});
+	} catch {}
+	try {
+		services.connection?.rpc?.handle("/llm-verifier", async (endpoint, payload) => {
+			if (endpoint !== "statistics") return rpcFailure("unknown llm-verifier endpoint");
+			return handleStatisticsQuery(payload);
+		}, { authority: "loopback" });
+	} catch (e) {
+		ctx.logger.warn("failed to register /llm-verifier rpc handler: " + String(e));
+	}
 	ctx.on("agent/disposed", ({ agent }) => {
 		autoBudget.release(agent);
 		autoRouter.release(agent);
@@ -2047,6 +2314,114 @@ function apply(ctx, config = {}) {
 		const admittedLastSeq = agent.session.events.at(-1)?.seq ?? -1;
 		const snapshot = agent.session.events.filter((event) => event.seq <= admittedLastSeq);
 		const stillCurrent = () => !signal.aborted && (agent.session.events.at(-1)?.seq ?? -1) === admittedLastSeq;
+		if (selected.autoVerifyPlanMode) {
+			const planDetection = detectPlanExit(snapshot, evidence.taskStartSeq);
+			if (planDetection.hasExitPlanMode && planDetection.planText) {
+				const planFingerprint = stableHash({
+					phase: "plan_review",
+					callSeq: planDetection.callSeq,
+					plan: planDetection.planText
+				});
+				const planReservation = autoRouter.reserve(agent, "plan_review", planFingerprint, 1, policy);
+				if (planReservation) try {
+					const prompt = buildPlanPreReviewPrompt((await extractTask(agent, evidence.taskStartSeq, admittedLastSeq, selected.autoVerifyMaxChars, signal)).problem, planDetection.planText, selected.autoRouteMaxInputChars);
+					const classified = await classifyRoute(agent, prompt, signal);
+					if (!stillCurrent()) {
+						autoRouter.fail(agent, planReservation, false);
+						return;
+					}
+					const verdict = parsePlanReviewVerdict(classified.text);
+					if (verdict.score >= selected.autoVerifyThreshold) autoRouter.commit(agent, planReservation, admittedLastSeq);
+					else {
+						autoRouter.fail(agent, planReservation, selected.autoVerifyMode === "strict");
+						agent.steer(createUserMessage({
+							content: [{
+								type: "text",
+								text: "[Automatic Verifier Plan Pre-review]\nPlan pre-review scored " + (verdict.score * 100).toFixed(1) + "% (Threshold: " + (selected.autoVerifyThreshold * 100).toFixed(0) + "%).\n" + verdict.feedback + "\nAddress the identified issues or improve the plan before proceeding."
+							}],
+							source: {
+								kind: "plugin",
+								plugin: "dsh-llm-verifier",
+								form: "notice",
+								summary: "Plan Pre-review Feedback"
+							}
+						}));
+						return;
+					}
+				} catch (error) {
+					autoRouter.fail(agent, planReservation, selected.autoVerifyMode === "strict");
+					ctx.logger.warn("llm-verifier plan pre-review failed: " + (error instanceof Error ? error.message : String(error)));
+					if (selected.autoVerifyMode === "strict" && !signal.aborted) {
+						agent.steer(createUserMessage({
+							content: [{
+								type: "text",
+								text: "[Automatic Verifier Plan Pre-review]\nPlan pre-review failed: " + (error instanceof Error ? error.message : String(error))
+							}],
+							source: {
+								kind: "plugin",
+								plugin: "dsh-llm-verifier"
+							}
+						}));
+						return;
+					}
+				}
+			}
+		}
+		if (selected.autoVerifyTeamTasks) {
+			const teamInspection = inspectTeamTasks(snapshot, evidence.taskStartSeq);
+			if (teamInspection.hasRecentCompletedTask && teamInspection.latestCompletedTask) {
+				const task = teamInspection.latestCompletedTask;
+				const taskFingerprint = stableHash({
+					phase: "team_task",
+					taskId: task.id,
+					status: task.status,
+					seq: teamInspection.completedSeq
+				});
+				const taskReservation = autoRouter.reserve(agent, "team_task", taskFingerprint, 1, policy);
+				if (taskReservation) try {
+					const prompt = buildTeamTaskVerificationPrompt(task, (await extractTask(agent, evidence.taskStartSeq, admittedLastSeq, selected.autoVerifyMaxChars, signal)).trace, selected.autoRouteMaxInputChars);
+					const classified = await classifyRoute(agent, prompt, signal);
+					if (!stillCurrent()) {
+						autoRouter.fail(agent, taskReservation, false);
+						return;
+					}
+					const verdict = parsePlanReviewVerdict(classified.text);
+					if (verdict.score >= selected.autoVerifyThreshold) autoRouter.commit(agent, taskReservation, admittedLastSeq);
+					else {
+						autoRouter.fail(agent, taskReservation, selected.autoVerifyMode === "strict");
+						agent.steer(createUserMessage({
+							content: [{
+								type: "text",
+								text: "[Automatic Verifier Team Task Gate]\nTask \"" + task.subject + "\" (#" + task.id + ") verification scored " + (verdict.score * 100).toFixed(1) + "% (Threshold: " + (selected.autoVerifyThreshold * 100).toFixed(0) + "%).\n" + verdict.feedback + "\nProvide verified execution evidence or resolve remaining issues before completing the task."
+							}],
+							source: {
+								kind: "plugin",
+								plugin: "dsh-llm-verifier",
+								form: "notice",
+								summary: "Team Task Gate Feedback"
+							}
+						}));
+						return;
+					}
+				} catch (error) {
+					autoRouter.fail(agent, taskReservation, selected.autoVerifyMode === "strict");
+					ctx.logger.warn("llm-verifier team task verification failed: " + (error instanceof Error ? error.message : String(error)));
+					if (selected.autoVerifyMode === "strict" && !signal.aborted) {
+						agent.steer(createUserMessage({
+							content: [{
+								type: "text",
+								text: "[Automatic Verifier Team Task Gate]\nTask verification failed: " + (error instanceof Error ? error.message : String(error))
+							}],
+							source: {
+								kind: "plugin",
+								plugin: "dsh-llm-verifier"
+							}
+						}));
+						return;
+					}
+				}
+			}
+		}
 		let decision = boundDecision(analyzeStructuredRoute(snapshot, selected.autoRouteMaxCandidates, selected.autoRouteMaxItemChars), policy);
 		if (decision === void 0 && selected.autoRouteSemantic && (selected.autoVerifyMode === "strict" || semanticRouteHint(snapshot))) {
 			const fingerprint = stableHash({
@@ -2551,6 +2926,6 @@ function apply(ctx, config = {}) {
 	}));
 }
 //#endregion
-export { AutoVerificationBudget, AutoVerifierRouter, Config, DEFAULT_CRITERIA, DEFAULT_GROUND_TRUTH_NOTE, GRANULARITY, LETTERS, RequestLimiter, SCALE_DESCRIPTION, ScoreCache, SingleFlight, StatisticsStore, VERIFIER_TOOL_NAMES, VerifierEngine, accumulatePairs, analyzeAutoTask, analyzeStructuredRoute, apply, automaticFeedback, boundDecision, bradleyTerry, buildEvidenceIndex, buildPairwisePrompt, buildProgressPrompt, buildSemanticRoutePrompt, callVerifier, emptyRunStats, errorDetails, extractProgressScore, extractScore, inject, latestDirectUserSeq, mergeStatisticsOverviews, name, normalizeCriteria, normalizeScoreLetter, parseSemanticRoute, pivotRoundPairs, rankScores, resolveCacheFile, resolveStatisticsFile, resolveTopicDataDir, ringCycle, seededRandom, semanticDecision, semanticRouteHint, stableHash, topPivots };
+export { AutoVerificationBudget, AutoVerifierRouter, Config, DEFAULT_CRITERIA, DEFAULT_GROUND_TRUTH_NOTE, GRANULARITY, LETTERS, RequestLimiter, SCALE_DESCRIPTION, ScoreCache, SingleFlight, StatisticsStore, VERIFIER_TOOL_NAMES, VerifierEngine, accumulatePairs, analyzeAutoTask, analyzeStructuredRoute, apply, automaticFeedback, boundDecision, bradleyTerry, buildEvidenceIndex, buildPairwisePrompt, buildPlanPreReviewPrompt, buildProgressPrompt, buildSemanticRoutePrompt, buildTeamTaskVerificationPrompt, callVerifier, detectPlanExit, emptyRunStats, errorDetails, extractProgressScore, extractScore, inject, inspectTeamTasks, latestDirectUserSeq, mergeStatisticsOverviews, name, normalizeCriteria, normalizeScoreLetter, parsePlanReviewVerdict, parseSemanticRoute, pivotRoundPairs, rankScores, resolveCacheFile, resolveStatisticsFile, resolveTopicDataDir, ringCycle, seededRandom, semanticDecision, semanticRouteHint, stableHash, topPivots };
 
 //# sourceMappingURL=index.js.map

package/lib/types/client-i18n.d.ts

@@ -122,6 +122,15 @@ export declare const zh: {
     'models.tokens': string;
     'models.empty': string;
     'slot.statistics': string;
+    'slot.globalDashboard': string;
+    'global.panelTitle': string;
+    'global.panelIntro': string;
+    'guide.verifier.title': string;
+    'guide.verifier.desc': string;
+    'field.autoVerifyTeamTasks.title': string;
+    'field.autoVerifyTeamTasks.help': string;
+    'field.autoVerifyPlanMode.title': string;
+    'field.autoVerifyPlanMode.help': string;
 };
 export type I18nDict = typeof zh;
 export declare const en: I18nDict;
@@ -250,6 +259,15 @@ export declare const dictionaries: {
         'models.tokens': string;
         'models.empty': string;
         'slot.statistics': string;
+        'slot.globalDashboard': string;
+        'global.panelTitle': string;
+        'global.panelIntro': string;
+        'guide.verifier.title': string;
+        'guide.verifier.desc': string;
+        'field.autoVerifyTeamTasks.title': string;
+        'field.autoVerifyTeamTasks.help': string;
+        'field.autoVerifyPlanMode.title': string;
+        'field.autoVerifyPlanMode.help': string;
     };
     en: {
         'settings.title': string;
@@ -375,6 +393,15 @@ export declare const dictionaries: {
         'models.tokens': string;
         'models.empty': string;
         'slot.statistics': string;
+        'slot.globalDashboard': string;
+        'global.panelTitle': string;
+        'global.panelIntro': string;
+        'guide.verifier.title': string;
+        'guide.verifier.desc': string;
+        'field.autoVerifyTeamTasks.title': string;
+        'field.autoVerifyTeamTasks.help': string;
+        'field.autoVerifyPlanMode.title': string;
+        'field.autoVerifyPlanMode.help': string;
     };
 };
 export declare const toolLabels: Record<'zh' | 'en', Record<string, string>>;

package/lib/types/client-i18n.js

@@ -137,6 +137,15 @@ export const zh = {
     'models.empty': '暂无模型调用',
     // Slot label
     'slot.statistics': '工具统计',
+    'slot.globalDashboard': 'Verifier 看板',
+    'global.panelTitle': 'LLM Verifier 全局监控中心',
+    'global.panelIntro': '跨会话全局统计、裁判模型效能大盘与任务验收总览。',
+    'guide.verifier.title': 'Verifier 统计',
+    'guide.verifier.desc': '大模型复核工具调用、耗时与成本统计',
+    'field.autoVerifyTeamTasks.title': '自动验收 Team 任务',
+    'field.autoVerifyTeamTasks.help': '当 Agent Teams 中的任务状态变更或进入完成状态时，自动路由进度追踪与任务验收。',
+    'field.autoVerifyPlanMode.title': '计划模式自动预审',
+    'field.autoVerifyPlanMode.help': '在 Agent 调用 exit_plan_mode 提交计划交付人类评审前，自动对计划完整度与风险进行独立预审。',
 };
 export const en = {
     // Settings Page
@@ -276,6 +285,15 @@ export const en = {
     'models.empty': 'No model calls',
     // Slot label
     'slot.statistics': 'Statistics',
+    'slot.globalDashboard': 'Verifier Dashboard',
+    'global.panelTitle': 'LLM Verifier Global Dashboard',
+    'global.panelIntro': 'Cross-session statistics, judge model metrics, and task acceptance overview.',
+    'guide.verifier.title': 'Verifier Stats',
+    'guide.verifier.desc': 'LLM verifier tool calls, duration and cost statistics',
+    'field.autoVerifyTeamTasks.title': 'Verify Team Tasks',
+    'field.autoVerifyTeamTasks.help': 'Automatically route progress and verify tasks when Agent Teams task status changes or completes.',
+    'field.autoVerifyPlanMode.title': 'Plan Mode Pre-verification',
+    'field.autoVerifyPlanMode.help': 'Automatically pre-verify plan feasibility and risks before exit_plan_mode presents it for user review.',
 };
 export const dictionaries = { zh, en };
 export const toolLabels = {