dsh-llm-verifier 0.8.0 → 0.8.2

package/README.md

@@ -514,6 +514,8 @@ Workflow 候选在被记录后的**下一步主模型请求之前**就会评审
 
 候选选择产生 steering 后，Agent 必须实施胜出候选；进度未达阈值也会 steering 继续。任何自动 `compare/select/track` 成功都会设置 `finalVerificationRequired`，下一停止边界即使没有传统写入类工具也必须执行 `verifier_current_session`；只有针对最新快照且达到阈值的最终验收才能清除要求（计划预审通过**不会**设置该标记：批准计划不等于完成工作）。路由采用 reservation/commit/fail 状态，取消、过期或失败不会被误记为成功；strict 会维持阻断并 fail closed。自动反馈按**真实结果**措辞：唯一最高分才报 "Winner / Proceed with"，最高分并列会列出并列集合并说明没有唯一依据，逐字节相同的候选说明"未做质量比较"——**不会**把稳定排序的第一项说成赢家；最终验收不通过时会列出未达标的判据、判官实际读到的区间，以及被长度上限省略的字符数（没有逐项判据时如实说明缺少定位依据）。以上文案是确定性的，不额外调用模型。**交付阶段**（本任务 Todo 全部完成且存在一次真实验证运行）在停止边界跳过进度 track 直接最终验收：Todo 完成只决定"送去验收"，不决定通过，验证失败也照常送到裁判眼前；同一完成信号只消费一次，新工作或新验证结果才会重新激活，且未处理的候选选择优先于该快路径。
 
+**计划模式（plan mode）激活期间，自动门控整体挂起**：停止边界的所有自动路由、Team 任务验收与最终验收、`agent/pre-step` 的候选早评审与过程选优登记全部跳过（以会话日志中最后一条 `plan/mode` 事件折叠判定，空日志视为未激活）。此时 Agent 的职责是调研与提案，唯一仍运行的门是 `exit_plan_mode` 计划预审——否则验收必然对着尚未实施的方案判低分，并 steering「去实际执行」，等于在用户批准计划之前就命令 Agent 动手。计划获批退出计划模式后，下一停止边界照常恢复验收。
+
 显式调用 `verifier_current_session` **不会**无条件关闭自动门控：只有该次复核真的达到通过阈值（`winner === 'A'`、分数不低于阈值，**且每一项标准各自不低于阈值**——与自动门控同一条 `sessionAccepted` 规则，否则显式调用会变成绕过单项失败的漏洞），并且之后没有再产生实质工作（写文件、执行命令等）时，才会被当作"已完成验收"而放行。判决失败、低于阈值、结果无法解析、或通过之后又继续改动的复核，都不会解除门控——下一停止边界仍会照常验收。此外，Agent Teams 的队友会话以 `team-message` 作为任务边界，因此队友会话同样可以触发（并遵守）任务级门控；Team 任务验收与计划预审都要求裁判给出显式的 `Verdict: <A-T>` 行，解析不出来时：计划预审放行（预算耗尽时 deny 会死锁 plan mode，只能放行），**Team 任务验收在 strict 下照常阻断**，与其它自动阶段一致。
 
 `track` 采用**离线**打分：判官在给较早的检查点打分时已经看到了它之后的内容，因此曲线偏乐观。门控只用**最新**检查点（它之后没有内容，不存在"未来泄露"），曲线展示时请把它读作"判官回看整段会话时的判断"，而不是逐步在线估计；上游的 `ProgressTracker` 是逐步在线形态，代价是每一步一次调用，本插件不采用。

package/lib/{auto-CGbckoaX.js → auto-DpNooy0d.js}

@@ -2157,6 +2157,27 @@ function inspectUserInteractionPause(events, taskStartSeq, currentTurn) {
 		}
 	}
 }
+/**
+* Whether the session is currently in plan mode.
+*
+* The host logs one `plan/mode` event per committed transition and folds the log as
+* "empty log → inactive, last event wins" (its own projection does exactly this). Hosts
+* without plan mode never log the event, so nothing changes there. While planning, the
+* agent is expected to research and PROPOSE: any automatic route or acceptance verdict at
+* the turn-stopping boundary can only fail against the missing implementation and steer
+* "actually implement it" — commanding execution the human has not approved yet. The
+* `exit_plan_mode` pre-review is the one gate that still runs in this state.
+* @param events - Session event log (the whole log is folded; the mode may predate the task).
+* @returns True when the newest `plan/mode` event activated plan mode.
+*/
+function planModeActive(events) {
+	let active = false;
+	for (const event of events) {
+		if (event.type !== "plan/mode") continue;
+		active = event.data?.active === true;
+	}
+	return active;
+}
 function analyzeAutoTask(events, policy, sessionId) {
 	const taskStartSeq = latestDirectUserSeq(events);
 	if (taskStartSeq === void 0) return {
@@ -2168,6 +2189,7 @@ function analyzeAutoTask(events, policy, sessionId) {
 		manualVerificationAccepted: false,
 		pendingSubagents: false,
 		pendingUserInteraction: false,
+		planMode: planModeActive(events),
 		eligible: false,
 		reason: "no-direct-user-task"
 	};
@@ -2227,6 +2249,7 @@ function analyzeAutoTask(events, policy, sessionId) {
 	});
 	const pendingSubagents = hasPendingSubagents(events, taskStartSeq);
 	const pendingUserInteraction = inspectUserInteractionPause(events, taskStartSeq) !== void 0;
+	const planMode = planModeActive(events);
 	if (policy.mode === "manual") return {
 		taskStartSeq,
 		toolCalls,
@@ -2236,9 +2259,23 @@ function analyzeAutoTask(events, policy, sessionId) {
 		manualVerificationAccepted,
 		pendingSubagents,
 		pendingUserInteraction,
+		planMode,
 		eligible: false,
 		reason: "manual-mode"
 	};
+	if (planMode) return {
+		taskStartSeq,
+		toolCalls,
+		completedToolResults,
+		consequentialToolCalls,
+		hasManualSessionVerification,
+		manualVerificationAccepted,
+		pendingSubagents,
+		pendingUserInteraction,
+		planMode: true,
+		eligible: false,
+		reason: "plan-mode-active"
+	};
 	if (manualVerificationAccepted) return {
 		taskStartSeq,
 		toolCalls,
@@ -2248,6 +2285,7 @@ function analyzeAutoTask(events, policy, sessionId) {
 		manualVerificationAccepted,
 		pendingSubagents,
 		pendingUserInteraction,
+		planMode,
 		eligible: false,
 		reason: "already-verified"
 	};
@@ -2260,6 +2298,7 @@ function analyzeAutoTask(events, policy, sessionId) {
 		manualVerificationAccepted,
 		pendingSubagents: true,
 		pendingUserInteraction,
+		planMode,
 		eligible: false,
 		reason: "pending-subagents"
 	};
@@ -2272,6 +2311,7 @@ function analyzeAutoTask(events, policy, sessionId) {
 		manualVerificationAccepted,
 		pendingSubagents,
 		pendingUserInteraction: true,
+		planMode,
 		eligible: false,
 		reason: "user-interaction-paused"
 	};
@@ -2284,6 +2324,7 @@ function analyzeAutoTask(events, policy, sessionId) {
 		manualVerificationAccepted,
 		pendingSubagents,
 		pendingUserInteraction,
+		planMode,
 		eligible: false,
 		reason: "no-consequential-work"
 	};
@@ -2296,6 +2337,7 @@ function analyzeAutoTask(events, policy, sessionId) {
 		manualVerificationAccepted,
 		pendingSubagents,
 		pendingUserInteraction,
+		planMode,
 		eligible: false,
 		reason: "no-completed-evidence"
 	};
@@ -2308,6 +2350,7 @@ function analyzeAutoTask(events, policy, sessionId) {
 		manualVerificationAccepted,
 		pendingSubagents,
 		pendingUserInteraction,
+		planMode,
 		eligible: false,
 		reason: "insufficient-tool-evidence"
 	};
@@ -2320,6 +2363,7 @@ function analyzeAutoTask(events, policy, sessionId) {
 		manualVerificationAccepted,
 		pendingSubagents,
 		pendingUserInteraction,
+		planMode,
 		eligible: true,
 		reason: policy.mode + "-eligible"
 	};
@@ -2487,6 +2531,6 @@ function selectRouteFeedbackDetail(candidates, result, maxChars = MAX_ROUTE_FEED
 	].join("\n"));
 }
 //#endregion
-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 };
+export { routedRepeats as A, SingleFlight as B, estimateRoutedCalls as C, latestDirectUserSeq as D, itemBudget as E, verificationVerdict as F, stableHash as H, extractSession as I, sanitizeVerifierText as L, semanticReferencesVisible as M, semanticRouteHint as N, nextDiagnosticCycleId as O, verificationFailed as P, sessionEvents as R, buildSemanticRouteView as S, inspectRecoverySignal as T, resolveCacheFile as V, RECOVERY_FAILURE_CONTEXT_CHARS as _, compareRouteFeedbackDetail as a, buildEvidenceIndex as b, inspectUserInteractionPause as c, planModeActive as d, selectRouteFeedbackDetail as f, MAX_ROUTED_CHECKPOINTS as g, AutoVerifierRouter as h, automaticFeedback as i, semanticDecision as j, parseSemanticRoute as k, isAwaitingUserText as l, topScoreIndices as m, PROPOSAL_FEEDBACK_NOTE as n, failedAcceptanceCriteria as o, sessionAccepted as p, analyzeAutoTask as r, hasPendingSubagents as s, MAX_ROUTE_FEEDBACK_CHARS as t, isSubagentSession as u, analyzeStructuredRoute as v, inspectDeliveryPhase as w, buildSemanticRoutePrompt as x, boundDecision as y, ScoreCache as z };
 
-//# sourceMappingURL=auto-CGbckoaX.js.map
+//# sourceMappingURL=auto-DpNooy0d.js.map