dsh-llm-verifier 0.6.8 → 0.7.0

package/lib/{auto-DmyrNZXK.js → auto-b3qPKifm.js}

@@ -361,7 +361,7 @@ function blockText$1(blocks) {
 * @param blocks - content blocks of an `assistant/message` event.
 * @returns The joined text, trimmed.
 */
-function narrativeText(blocks) {
+function narrativeText$1(blocks) {
 	const parts = [];
 	const visit = (items) => {
 		for (const block of items) if (block.type === "text") parts.push(block.text);
@@ -433,7 +433,7 @@ function buildEvidenceIndex(events) {
 			}
 		} else if (event.type === "assistant/message") {
 			const blocks = event.data?.message?.content ?? event.data?.content;
-			const text = narrativeText(Array.isArray(blocks) ? blocks : []);
+			const text = narrativeText$1(Array.isArray(blocks) ? blocks : []);
 			if (text) narration = {
 				seq: event.seq,
 				text
@@ -1770,13 +1770,22 @@ const PASSIVE_TOOLS = /* @__PURE__ */ new Set([
 	"glob",
 	"grep",
 	"web_search",
+	"web_fetch",
 	"ssh_list",
 	"job_list",
 	"job_output",
 	"list_agents",
+	"list_subagent_models",
+	"list_mcp_resources",
+	"list_mcp_resource_templates",
+	"read_mcp_resource",
 	"get_goal",
 	"skill",
-	"mcp__codegraph__codegraph_explore"
+	"mcp__codegraph__codegraph_explore",
+	"ask_user_question",
+	"todo_write",
+	"present",
+	"run_code"
 ]);
 const VERIFIER_TOOLS = /* @__PURE__ */ new Set([
 	"verifier_compare",
@@ -1790,7 +1799,6 @@ const CONSEQUENTIAL_TOOLS = /* @__PURE__ */ new Set([
 	"write",
 	"pwsh",
 	"bash",
-	"run_code",
 	"codex_image_generate",
 	"ssh_exec",
 	"ssh_upload",
@@ -1993,6 +2001,126 @@ function hasPendingSubagents(events, taskStartSeq) {
 	}
 	return pendingContinuable.size > 0 || pendingJobs.size > 0 || anonymousPendingCount > 0;
 }
+function parseArgumentsObject(value) {
+	if (typeof value === "string") try {
+		const parsed = JSON.parse(value);
+		return typeof parsed === "object" && parsed !== null ? parsed : {};
+	} catch {
+		return {};
+	}
+	return typeof value === "object" && value !== null ? value : {};
+}
+const USER_QUESTION_PATTERNS = [
+	/[?？]\s*["'）)』」]*\s*$/u,
+	/(?:请|麻烦您?)(?:告知|确认|指示|选择|提供|决定|回复)/u,
+	/(?:等待|静待|等)(?:您的|你的|您|你|用户)?(?:确认|指示|回复|决定|指令|反馈|选择|输入)/u,
+	/(?:如果您?|如|若)(?:希望|需要|想)(?:继续|执行).*(?:请|告诉我|告知)/u,
+	/(?:是否|要不要|可否)(?:需要我?|继续|同意|允许|采用).*[？?]?/u,
+	/(?:暂时|先|已)?暂停(?:工作|执行|后续).*(?:等|指示|确认|决定|用户)/u,
+	/\b(?:please\s+(?:confirm|let\s+me\s+know|advise|choose|select|provide|indicate|tell\s+me|reply))\b/iu,
+	/\b(?:waiting\s+for|awaiting)\s+(?:your\s+|user\s+)?(?:input|instructions?|reply|response|confirmation|decision|guidance|feedback)\b/iu,
+	/\b(?:would\s+you\s+like|should\s+i|do\s+you\s+want\s+me\s+to|how\s+would\s+you\s+like|which\s+(?:option|approach|strategy)\s+do\s+you\s+prefer)\b/iu,
+	/\b(?:let\s+me\s+know\s+(?:how|what|if|whether|when))\b/iu,
+	/\b(?:paused?\s+(?:work|execution|here|for\s+now)|stopping\s+here)\b/iu
+];
+function isAwaitingUserText(text) {
+	const trimmed = text.trim();
+	if (!trimmed) return false;
+	const lastLine = trimmed.split("\n").filter((l) => l.trim()).at(-1) ?? "";
+	return USER_QUESTION_PATTERNS.some((p) => p.test(lastLine) || p.test(trimmed));
+}
+function narrativeText(blocks) {
+	if (typeof blocks === "string") return blocks.trim();
+	if (!Array.isArray(blocks)) return "";
+	const parts = [];
+	for (const block of blocks) if (block?.type === "text" && typeof block.text === "string") parts.push(block.text);
+	else if (block?.type === "tool-result") parts.push(blockText(block.content));
+	return parts.join("\n").trim();
+}
+/**
+* Whether the agent has paused to ask the user a question, obtain confirmation, or await user instructions.
+*
+* During task execution, an agent may legitimately pause to ask the operator a question
+* (e.g. calling `ask_user_question`, pausing/blocking a goal, or concluding a turn with prose
+* awaiting user guidance). Gating or steering in this state forces the model to keep executing,
+* overriding the user interaction boundary and locking the user out of providing guidance.
+*
+* Returns undefined when the work has already reached its delivery phase (all todos completed
+* with verification evidence), because in that state the agent is delivering the task rather than
+* pausing for input.
+*/
+function inspectUserInteractionPause(events, taskStartSeq, currentTurn) {
+	const delivery = inspectDeliveryPhase(events);
+	if (delivery?.todosComplete && delivery.verification !== void 0) return;
+	let turnEvents = events;
+	if (currentTurn !== void 0) {
+		const turnStartIndex = events.findLastIndex((e) => e.type === "turn/start" && e.data.turn === currentTurn);
+		if (turnStartIndex >= 0) turnEvents = events.slice(turnStartIndex);
+		else {
+			const byTurn = events.filter((e) => e.data?.turn === currentTurn);
+			if (byTurn.length > 0) turnEvents = byTurn;
+		}
+	} else {
+		const lastTurnStart = events.findLastIndex((e) => e.type === "turn/start");
+		if (lastTurnStart >= 0) turnEvents = events.slice(lastTurnStart);
+	}
+	let lastAskUserSeq = -1;
+	for (const event of turnEvents) if (event.type === "tool/call") {
+		if (event.data.name === "ask_user_question") lastAskUserSeq = Math.max(lastAskUserSeq, event.seq);
+	} else if (CODE_DISPATCH_TYPES.has(event.type)) {
+		if (event.data?.name === "ask_user_question") lastAskUserSeq = Math.max(lastAskUserSeq, event.seq);
+	}
+	if (lastAskUserSeq >= 0) {
+		if (!turnEvents.some((event) => {
+			if (event.seq <= lastAskUserSeq) return false;
+			if (event.type === "tool/call") return isConsequential(event.data.name);
+			if (CODE_DISPATCH_TYPES.has(event.type)) {
+				const data = event.data;
+				return data ? isConsequential(data.name) : false;
+			}
+			return false;
+		})) return {
+			paused: true,
+			reason: "ask_user_question in current turn"
+		};
+	}
+	const taskEvents = events.filter((e) => e.seq >= taskStartSeq);
+	let goalPhase;
+	for (const event of taskEvents) if (event.type === "tool/call") {
+		const call = event;
+		if (call.data.name === "update_goal") {
+			const args = parseArgumentsObject(call.data.arguments);
+			if (args.action === "pause") goalPhase = "paused";
+			else if (args.action === "blocked") goalPhase = "blocked";
+			else if (args.action === "resume") goalPhase = "active";
+			else if (args.action === "complete") goalPhase = "complete";
+		}
+	} else if (CODE_DISPATCH_TYPES.has(event.type)) {
+		const data = event.data;
+		if (data?.name === "update_goal") {
+			const args = parseArgumentsObject(data.arguments);
+			if (args.action === "pause") goalPhase = "paused";
+			else if (args.action === "blocked") goalPhase = "blocked";
+			else if (args.action === "resume") goalPhase = "active";
+			else if (args.action === "complete") goalPhase = "complete";
+		}
+	}
+	if (goalPhase === "paused" || goalPhase === "blocked") return {
+		paused: true,
+		reason: "goal is " + goalPhase
+	};
+	const lastAssistant = turnEvents.filter((e) => e.type === "assistant/message").at(-1);
+	if (lastAssistant && lastAssistant.type === "assistant/message") {
+		const content = lastAssistant.data.message?.content ?? [];
+		if (!content.some((b) => b.type === "tool-call")) {
+			const text = narrativeText(content);
+			if (text && isAwaitingUserText(text)) return {
+				paused: true,
+				reason: "assistant awaiting user instructions"
+			};
+		}
+	}
+}
 function analyzeAutoTask(events, policy, sessionId) {
 	const taskStartSeq = latestDirectUserSeq(events);
 	if (taskStartSeq === void 0) return {
@@ -2003,6 +2131,7 @@ function analyzeAutoTask(events, policy, sessionId) {
 		hasManualSessionVerification: false,
 		manualVerificationAccepted: false,
 		pendingSubagents: false,
+		pendingUserInteraction: false,
 		eligible: false,
 		reason: "no-direct-user-task"
 	};
@@ -2061,6 +2190,7 @@ function analyzeAutoTask(events, policy, sessionId) {
 		}, policy.threshold) && !stale(verdict.toSeq);
 	});
 	const pendingSubagents = hasPendingSubagents(events, taskStartSeq);
+	const pendingUserInteraction = inspectUserInteractionPause(events, taskStartSeq) !== void 0;
 	if (policy.mode === "manual") return {
 		taskStartSeq,
 		toolCalls,
@@ -2069,6 +2199,7 @@ function analyzeAutoTask(events, policy, sessionId) {
 		hasManualSessionVerification,
 		manualVerificationAccepted,
 		pendingSubagents,
+		pendingUserInteraction,
 		eligible: false,
 		reason: "manual-mode"
 	};
@@ -2080,6 +2211,7 @@ function analyzeAutoTask(events, policy, sessionId) {
 		hasManualSessionVerification,
 		manualVerificationAccepted,
 		pendingSubagents,
+		pendingUserInteraction,
 		eligible: false,
 		reason: "already-verified"
 	};
@@ -2091,9 +2223,22 @@ function analyzeAutoTask(events, policy, sessionId) {
 		hasManualSessionVerification,
 		manualVerificationAccepted,
 		pendingSubagents: true,
+		pendingUserInteraction,
 		eligible: false,
 		reason: "pending-subagents"
 	};
+	if (pendingUserInteraction) return {
+		taskStartSeq,
+		toolCalls,
+		completedToolResults,
+		consequentialToolCalls,
+		hasManualSessionVerification,
+		manualVerificationAccepted,
+		pendingSubagents,
+		pendingUserInteraction: true,
+		eligible: false,
+		reason: "user-interaction-paused"
+	};
 	if (consequentialToolCalls === 0) return {
 		taskStartSeq,
 		toolCalls,
@@ -2102,6 +2247,7 @@ function analyzeAutoTask(events, policy, sessionId) {
 		hasManualSessionVerification,
 		manualVerificationAccepted,
 		pendingSubagents,
+		pendingUserInteraction,
 		eligible: false,
 		reason: "no-consequential-work"
 	};
@@ -2113,6 +2259,7 @@ function analyzeAutoTask(events, policy, sessionId) {
 		hasManualSessionVerification,
 		manualVerificationAccepted,
 		pendingSubagents,
+		pendingUserInteraction,
 		eligible: false,
 		reason: "no-completed-evidence"
 	};
@@ -2124,6 +2271,7 @@ function analyzeAutoTask(events, policy, sessionId) {
 		hasManualSessionVerification,
 		manualVerificationAccepted,
 		pendingSubagents,
+		pendingUserInteraction,
 		eligible: false,
 		reason: "insufficient-tool-evidence"
 	};
@@ -2135,6 +2283,7 @@ function analyzeAutoTask(events, policy, sessionId) {
 		hasManualSessionVerification,
 		manualVerificationAccepted,
 		pendingSubagents,
+		pendingUserInteraction,
 		eligible: true,
 		reason: policy.mode + "-eligible"
 	};
@@ -2302,6 +2451,6 @@ function selectRouteFeedbackDetail(candidates, result, maxChars = MAX_ROUTE_FEED
 	].join("\n"));
 }
 //#endregion
-export { semanticRouteHint as A, itemBudget as C, routedRepeats as D, parseSemanticRoute as E, sessionEvents as F, ScoreCache as I, SingleFlight as L, verificationVerdict as M, extractSession as N, semanticDecision as O, sanitizeVerifierText as P, resolveCacheFile as R, inspectRecoverySignal as S, nextDiagnosticCycleId as T, buildEvidenceIndex as _, compareRouteFeedbackDetail as a, estimateRoutedCalls as b, isSubagentSession as c, topScoreIndices as d, AutoVerifierRouter as f, boundDecision as g, analyzeStructuredRoute as h, automaticFeedback as i, verificationFailed as j, semanticReferencesVisible as k, selectRouteFeedbackDetail as l, RECOVERY_FAILURE_CONTEXT_CHARS as m, PROPOSAL_FEEDBACK_NOTE as n, failedAcceptanceCriteria as o, MAX_ROUTED_CHECKPOINTS as p, analyzeAutoTask as r, hasPendingSubagents as s, MAX_ROUTE_FEEDBACK_CHARS as t, sessionAccepted as u, buildSemanticRoutePrompt as v, latestDirectUserSeq as w, inspectDeliveryPhase as x, buildSemanticRouteView as y, stableHash as z };
+export { semanticDecision as A, resolveCacheFile as B, inspectDeliveryPhase as C, nextDiagnosticCycleId as D, latestDirectUserSeq as E, extractSession as F, sanitizeVerifierText as I, sessionEvents as L, semanticRouteHint as M, verificationFailed as N, parseSemanticRoute as O, verificationVerdict as P, ScoreCache as R, estimateRoutedCalls as S, itemBudget as T, stableHash as V, analyzeStructuredRoute as _, compareRouteFeedbackDetail as a, buildSemanticRoutePrompt as b, inspectUserInteractionPause as c, selectRouteFeedbackDetail as d, sessionAccepted as f, RECOVERY_FAILURE_CONTEXT_CHARS as g, MAX_ROUTED_CHECKPOINTS as h, automaticFeedback as i, semanticReferencesVisible as j, routedRepeats as k, isAwaitingUserText as l, AutoVerifierRouter as m, PROPOSAL_FEEDBACK_NOTE as n, failedAcceptanceCriteria as o, topScoreIndices as p, analyzeAutoTask as r, hasPendingSubagents as s, MAX_ROUTE_FEEDBACK_CHARS as t, isSubagentSession as u, boundDecision as v, inspectRecoverySignal as w, buildSemanticRouteView as x, buildEvidenceIndex as y, SingleFlight as z };
 
-//# sourceMappingURL=auto-DmyrNZXK.js.map
+//# sourceMappingURL=auto-b3qPKifm.js.map