From 6320afbf5dbaa48a80f915d186d53106527bb386 Mon Sep 17 00:00:00 2001 From: Jesse_Chen Date: Sun, 26 Jul 2026 19:55:38 +0800 Subject: [PATCH] fix: decouple rectification narrative from scoring state --- docs/BUG_HISTORY.md | 15 ++ .../narrative-agent.ts | 92 ++++++++++- .../conversational-narrative-agent.test.ts | 88 ++++++++--- ...ational-rectification-orchestrator.test.ts | 147 ++++++++++++------ 4 files changed, 260 insertions(+), 82 deletions(-) diff --git a/docs/BUG_HISTORY.md b/docs/BUG_HISTORY.md index 5831723f..22bbd43c 100644 --- a/docs/BUG_HISTORY.md +++ b/docs/BUG_HISTORY.md @@ -1474,3 +1474,18 @@ - 防复发：模型自然语言不得成为事件保存的硬依赖；私有路由元数据为空时应省略可选状态，而不是生成违反持久化合同的半合法对象。 - 相关记录：BUG-075、BUG-078 - 修复版本：本次修复提交 + +## BUG-080 | 开放叙事仍被技术上下文塑造成流程播报 + +- 状态：resolved +- 首次发现：2026-07-26 +- 最近更新：2026-07-26 +- 影响面：生时校正普通叙事轮、自然对话、追问状态持久化 +- 用户现象：用户叙述“2016 年离家去外地上大学”后，Agent 仍回答“先记为、对校时有价值、参与候选核对、不会确认某个分钟”，像记录员而不是自然交谈。 +- 触发条件：普通 `intermediate` 轮生成叙事时仍向模型提供 Technical Packet、候选分钟、分盘、事件台账和未决证据。 +- 根因：此前开放叙事只取消了“每轮必须提问”，没有隔离普通聊天与技术收敛上下文；模型继续模仿内部流程措辞。叙事没有可见问题时，模型返回的隐藏 `evidenceRequest` 还会进入下一轮状态。普通提示不再暴露事件 ID 后，旧 authored schema 又仍强制模型提交已有 evidenceId，形成合同矛盾。 +- 修复：普通叙事轮只向模型提供最近真实对话和本轮用户原话，禁止主动播报记录、校时价值、候选范围、分盘、评分与收敛；只有用户主动询问技术结果时才提供受约束技术包。没有可见问题的隐藏 `evidenceRequest` 统一清空；自然追问只由模型表达问题意图，目标事件 ID 由服务器依据活动事件和未决证据绑定。首轮、最终轮及主动技术询问仍保留技术边界。 +- 验证：Narrative Agent 与 Orchestrator 联合回归 99/99 通过，覆盖普通 prompt 不含 packet/eventLedger、隐藏追问清空、服务器绑定 follow-up、相对日期承接、fallback、暂停恢复和并发幂等。 +- 防复发：普通聊天的可见措辞不得由技术 packet 或事件台账驱动；内部收敛状态只能记录和计算，不能成为 Agent 每轮必须复述的脚本。 +- 相关记录：BUG-074、BUG-075、BUG-078、BUG-079 +- 修复版本：本次修复提交 diff --git a/frontend/src/lib/conversational-rectification/narrative-agent.ts b/frontend/src/lib/conversational-rectification/narrative-agent.ts index a7f373b6..e21e0c4e 100644 --- a/frontend/src/lib/conversational-rectification/narrative-agent.ts +++ b/frontend/src/lib/conversational-rectification/narrative-agent.ts @@ -65,6 +65,7 @@ const labeledYearChoicesPattern = /A\s*[.、:：)]?[\s\S]{0,80}(?:19|20)\d{2}\s* const affirmativeAnswerPattern = /^\s*(?:是(?:的)?|对(?:的)?|没错|正确|确认|就是|嗯+|没问题)\s*[。.!！,，]?\s*$/u; const negativeAnswerPattern = /^\s*(?:不是|不对|错了|并不是|否)\s*[。.!！,，]?\s*$/u; const proposedDateQuestionPattern = /(?:19|20)\d{2}\s*年(?:\s*(?:1[0-2]|0?[1-9])\s*月)?(?:\s*(?:3[01]|[12]\d|0?[1-9])\s*(?:日|号))?[\s\S]{0,30}(?:吗|是否|是不是|确认|对不对|正确)/u; +const technicalDiscussionRequestPattern = /(?:校时|生时(?:校正|纠正)|出生时间|候选(?:时间|范围|分钟)|收敛(?:结果|进度)?|技术(?:分析|结果)|分盘|\bD\d{1,3}\b|\b(?:UL|A7|A10|KP)\b)/iu; const domainLabels = { career: "事业", education: "学业", @@ -98,6 +99,18 @@ export const rectificationNarrativeOutputSchema = z.object({ }).strict().nullable(), }).strict(); +const authoredFollowUpSchema = z.object({ + kind: z.enum(["new_event", "event_date", "event_detail"]), + // Event ids are server-owned. Ordinary conversation prompts deliberately do + // not expose the event ledger, so authored outputs may omit the target id. + evidenceId: z.string().uuid().nullable().optional(), + answerMode: z.enum(["free_text", "yes_no"]).optional(), + proposedDate: z.object({ + value: z.string().regex(/^\d{4}(?:-(?:0[1-9]|1[0-2])(?:-(?:0[1-9]|[12]\d|3[01]))?)?$/), + precision: z.enum(["year", "month", "day"]), + }).strict().nullable().optional(), +}).strict(); + export const rectificationNarrativeAuthoredOutputSchema = z.object({ narrative: z.string().trim().min(1).max(12_000), evidenceRequest: z.object({ @@ -106,7 +119,7 @@ export const rectificationNarrativeAuthoredOutputSchema = z.object({ domains: z.array(domainSchema).min(1).max(4).optional(), datePrecision: z.enum(["month_preferred", "year_accepted"]), prompt: z.string().trim().min(1).max(1_000), - followUp: rectificationFollowUpSchema.default({ kind: "new_event", evidenceId: null }), + followUp: authoredFollowUpSchema.default({ kind: "new_event", evidenceId: null }), }).strict().nullable(), }).strict(); @@ -177,11 +190,39 @@ function groundedEvidenceRequest( return domains.length > 0 ? { ...request, domains } : null; } +function serverOwnedFollowUp( + followUp: z.infer, + context: RectificationNarrativeContext, +): RectificationFollowUp { + if (followUp.kind === "new_event") return { kind: "new_event", evidenceId: null }; + const active = context.eventLedger?.filter((item) => item.active) ?? []; + const allowedIds = new Set(active.map((item) => item.id)); + const evidenceId = followUp.evidenceId && allowedIds.has(followUp.evidenceId) + ? followUp.evidenceId + : followUp.kind === "event_date" + ? context.unresolvedEvidence?.at(-1)?.id ?? active.at(-1)?.id + : active.at(-1)?.id; + if (!evidenceId) return { kind: "new_event", evidenceId: null }; + if (followUp.kind === "event_date" && followUp.answerMode === "yes_no" && followUp.proposedDate) { + return rectificationFollowUpSchema.parse({ + kind: followUp.kind, + evidenceId, + answerMode: "yes_no", + proposedDate: followUp.proposedDate, + }); + } + return rectificationFollowUpSchema.parse({ kind: followUp.kind, evidenceId }); +} + function completeAuthoredOutput( output: z.infer, packet: RectificationTechnicalPacket, + context: RectificationNarrativeContext, ): RectificationNarrativeModelOutput { - const evidenceRequest = groundedEvidenceRequest(output.evidenceRequest, packet); + const groundedRequest = groundedEvidenceRequest(output.evidenceRequest, packet); + const evidenceRequest = groundedRequest + ? { ...groundedRequest, followUp: serverOwnedFollowUp(groundedRequest.followUp, context) } + : null; return { ...output, evidenceRequest, @@ -200,6 +241,7 @@ function completeAuthoredOutput( function parseModelOutput( text: string, packet: RectificationTechnicalPacket, + context: RectificationNarrativeContext, ): RectificationNarrativeModelOutput { const normalized = text.trim().replace(/^```(?:json)?\s*/i, "").replace(/\s*```$/, ""); const start = normalized.indexOf("{"); @@ -208,15 +250,18 @@ function parseModelOutput( const parsed: unknown = JSON.parse(normalized.slice(start, end + 1)); const legacy = rectificationNarrativeOutputSchema.safeParse(parsed); if (legacy.success) { + const groundedRequest = groundedEvidenceRequest(legacy.data.evidenceRequest, packet); return { ...legacy.data, // The next conversational topic is authored by the model, but the // scoring-domain allowlist remains server-owned. If no grounded routing // domain exists, keep the prose and omit only the optional follow-up state. - evidenceRequest: groundedEvidenceRequest(legacy.data.evidenceRequest, packet), + evidenceRequest: groundedRequest + ? { ...groundedRequest, followUp: serverOwnedFollowUp(groundedRequest.followUp, context) } + : null, }; } - return completeAuthoredOutput(rectificationNarrativeAuthoredOutputSchema.parse(parsed), packet); + return completeAuthoredOutput(rectificationNarrativeAuthoredOutputSchema.parse(parsed), packet, context); } function narrativeTimes(value: string): string[] { @@ -296,6 +341,13 @@ export function validateNarrativeAgainstPacket( } } if (output.evidenceRequest) { + if (phase === "intermediate" && !/[?？]/u.test(output.narrative)) { + issues.push("evidence request is not visibly asked in narrative"); + } + if (phase === "intermediate" + && !normalizedQuestion(output.narrative).includes(normalizedQuestion(output.evidenceRequest.prompt))) { + issues.push("evidence request prompt is not present in narrative"); + } for (const domain of output.evidenceRequest.domains) { if (!allowedDomains.has(domain)) issues.push(`evidence domain ${domain} is not packet-grounded`); } @@ -518,6 +570,24 @@ function promptFor( retryIssues: boundedReceiptIssues(retryIssues), }); } + if (phase === "intermediate" && !technicalDiscussionRequestPattern.test(context.latestUserText ?? "")) { + return JSON.stringify({ + task: "像正常的人一样接住用户刚才的叙述并自由交谈。不要把自己写成记录员、问卷或校时流程播报器。用户可以继续叙述，也可以自然转到别的话题。", + phase, + conversation: { + recentConversation: context.recentConversation?.slice(-40), + latestUserText: context.latestUserText, + previousAssistantNarrative: context.previousAssistantNarrative, + }, + rules: [ + "优先回应用户话里的真实内容、感受、选择或转折，不要复述成档案摘要。", + "除非用户主动询问校时进度或技术结果，不要说已记录、先记为、对校时有价值、参与候选时间核对、不会因单一事件确认分钟，也不要主动谈候选范围、分盘、评分、收敛或内部处理。", + "不必每轮提问。需要提问时只问自然推进对话真正需要的问题，不要机械补年月、结果或转折。", + "只返回 narrative 和 evidenceRequest。没有在 narrative 中逐字提出一个用户可见的明确问题时，evidenceRequest 必须为 null；若提出问题，evidenceRequest.prompt 必须与 narrative 中的问题文字完全一致。followUp 只表达 kind、answerMode 和 proposedDate，不要生成 evidenceId，目标事件由服务器绑定。", + ], + retryIssues: boundedReceiptIssues(retryIssues), + }); + } return JSON.stringify({ task: "write_grounded_rectification_narrative", phase, @@ -568,11 +638,12 @@ function promptFor( function fallbackNarrative(packet: RectificationTechnicalPacket, phase: RectificationNarrativePhase): string { const candidate = packet.candidate; + if (phase === "intermediate") { + return "我听到了。你可以顺着这段经历继续说，也可以自然讲下一件想到的事。"; + } const phaseLine = phase === "final" && candidate.status === "ready_for_confirmation" ? "当前证据已形成候选总结，但仍有残余不确定性；只有明确确认后才会替换当前排盘时间。" - : phase === "final" - ? "当前证据只能支持候选范围，系统验证尚未闭环；本次不会替换当前排盘时间，也不再强制追问更多人生事件。" - : `我收到了这段叙述。你可以继续讲这段经历，也可以按自己的节奏说下一件想到的事。`; + : "当前证据只能支持候选范围，系统验证尚未闭环；本次不会替换当前排盘时间，也不再强制追问更多人生事件。"; return [ `当前仍在核对 ${candidate.range.startTime}–${candidate.range.endTime} 的候选范围，还不能把其中某一分钟当作确定出生时间。`, phaseLine, @@ -740,7 +811,12 @@ export async function generateRectificationNarrative(input: { issues, ), { signal, attempt }); const modelId = modelIdSchema.parse(generated.modelId ?? defaultModelId); - const output = parseModelOutput(generated.text, input.packet); + const parsedOutput = parseModelOutput(generated.text, input.packet, input.context ?? {}); + const output = input.phase === "intermediate" + && parsedOutput.evidenceRequest + && !/[?？]/u.test(parsedOutput.narrative) + ? { ...parsedOutput, evidenceRequest: null } + : parsedOutput; const validation = validateNarrativeAgainstPacket( output, input.packet, diff --git a/frontend/tests/conversational-narrative-agent.test.ts b/frontend/tests/conversational-narrative-agent.test.ts index 776c53eb..2db931e7 100644 --- a/frontend/tests/conversational-narrative-agent.test.ts +++ b/frontend/tests/conversational-narrative-agent.test.ts @@ -235,7 +235,7 @@ test("keeps a natural intermediate acknowledgement without appending a question" assert.doesNotMatch(result.narrative, /[？?]/); }); -test("hides an uninvoked technique inventory during evidence collection", async () => { +test("keeps technical packets and event bookkeeping out of ordinary narrative turns", async () => { const prompts: string[] = []; const packet = { ...syntheticTechnicalPacket(), @@ -261,13 +261,54 @@ test("hides an uninvoked technique inventory during evidence collection", async await generateRectificationNarrative({ phase: "intermediate", packet, - generator: generator([richOutput()], prompts), + context: { + latestUserText: "2016 年离家去外地上大学", + recentConversation: [{ role: "user", text: "2016 年离家去外地上大学" }], + eventLedger: [{ + id: "00000000-0000-4000-8000-000000000801", + rawText: "2016 年离家去外地上大学", + dateLabel: "2016", + summary: "离家去外地上大学", + domain: "education", + extractionStatus: "clear", + active: true, + correctsEvidenceIds: [], + }], + }, + generator: generator([{ + narrative: "第一次长期离开家，生活节奏应该一下子变了很多。你可以接着讲后来发生的事。", + evidenceRequest: null, + }], prompts), }); - assert.match(prompts[0] ?? "", /expertWorkflow/); + assert.match(prompts[0] ?? "", /2016 年离家去外地上大学/); + assert.doesNotMatch(prompts[0] ?? "", /packet/); + assert.doesNotMatch(prompts[0] ?? "", /eventLedger/); + assert.doesNotMatch(prompts[0] ?? "", /expertWorkflow/); assert.doesNotMatch(prompts[0] ?? "", /KP cusp \/ sub-lord/); assert.doesNotMatch(prompts[0] ?? "", /minute_holdout_not_ready/); - assert.match(prompts[0] ?? "", /blockedOrNotEvaluatedTechniquesMustNeverBeClaimedAsUsed/); + assert.doesNotMatch(prompts[0] ?? "", /05:16|05:24|D9|D10/); +}); + +test("drops a hidden follow-up that the user cannot see", async () => { + const hidden = { + narrative: "第一次长期离开家，生活节奏应该一下子变了很多。你可以接着讲后来发生的事。", + evidenceRequest: { + datePrecision: "month_preferred" as const, + prompt: "2016 年离家去外地上大学大概是几月？", + followUp: { kind: "new_event" as const, evidenceId: null }, + }, + }; + const result = await generateRectificationNarrative({ + phase: "intermediate", + packet: syntheticTechnicalPacket(), + context: { latestUserText: "2016 年离家去外地上大学" }, + generator: generator([hidden]), + }); + + assert.equal(result.attempts, 1); + assert.equal(result.fallbackUsed, false); + assert.equal(result.output.evidenceRequest, null); }); test("appends the three auditable tables to every final Agent answer", async () => { @@ -400,25 +441,29 @@ test("rejects a model-authored event table that exposes private numeric scoring" assert.match(result.narrative, /\| 时间 \| 事件 \| 领域 \| 验证状态 \| 结论 \|/); }); -test("passes the user's latest concrete event to an intermediate skill-guided reply", async () => { +test("passes the user's latest words to an ordinary intermediate reply", async () => { const prompts: string[] = []; await generateRectificationNarrative({ phase: "intermediate", packet: syntheticTechnicalPacket(), context: { + latestUserText: "2023年9月离开家乡去上海开始第一份长期工作", latestEvidence: [{ dateLabel: "2023-09", summary: "离开家乡去上海开始第一份长期工作", domain: "career", }], }, - generator: generator([richOutput()], prompts), + generator: generator([{ + narrative: "第一次长期离开家去工作，适应过程应该不轻松。你可以接着讲。", + evidenceRequest: null, + }], prompts), }); assert.match(prompts[0] ?? "", /离开家乡去上海开始第一份长期工作/); - assert.match(prompts[0] ?? "", /acknowledgeAndReflectBeforeAnyClarification/); - assert.match(prompts[0] ?? "", /questionsAreOptional/); - assert.match(prompts[0] ?? "", /doNotRepeatCandidateBoundaryUnlessItChangedOrTheUserAsked/); + assert.match(prompts[0] ?? "", /不要把自己写成记录员/); + assert.doesNotMatch(prompts[0] ?? "", /latestEvidence/); + assert.doesNotMatch(prompts[0] ?? "", /候选范围.*05:16/); }); test("keeps internal event domains and suggested-domain routing out of the narrator context", async () => { @@ -448,18 +493,15 @@ test("keeps internal event domains and suggested-domain routing out of the narra }); const prompt = JSON.parse(prompts[0] ?? "{}") as { - conversationContext?: { - latestEvidence?: Array>; - eventLedger?: Array>; - }; + conversation?: Record; packet?: Record; }; - assert.equal(prompt.conversationContext?.latestEvidence?.[0]?.domain, undefined); - assert.equal(prompt.conversationContext?.eventLedger?.[0]?.domain, undefined); + assert.equal(prompt.conversation?.latestEvidence, undefined); + assert.equal(prompt.conversation?.eventLedger, undefined); assert.equal(prompt.packet?.suggestedDomains, undefined); }); -test("passes the active event ledger and unresolved facts to the intermediate agent", async () => { +test("keeps the event ledger and unresolved facts out of the ordinary intermediate agent", async () => { const prompts: string[] = []; await generateRectificationNarrative({ phase: "intermediate", @@ -494,13 +536,8 @@ test("passes the active event ledger and unresolved facts to the intermediate ag const prompt = prompts[0] ?? ""; assert.match(prompt, /23年关系结束后发生过一次交通事故/); - assert.match(prompt, /2024-08-08/); - assert.match(prompt, /一段重要关系结束/); - assert.match(prompt, /freeConversation/); - assert.match(prompt, /questionsAreOptional/); - assert.match(prompt, /acceptMultipleEventsInOneMessage/); - assert.match(prompt, /resolveDateContradictionsBeforeScoring/); - assert.match(prompt, /mergeSameEventDetailsWithoutDoubleCounting/); + assert.doesNotMatch(prompt, /2024-08-08/); + assert.doesNotMatch(prompt, /eventLedger|unresolvedEvidence|resolveDateContradictionsBeforeScoring/); }); test("rejects invented representative times, layers, and references", () => { @@ -564,6 +601,7 @@ test("allows packet-grounded technical tables during an intermediate turn", () = "| D9 | 在候选范围内呈分钟敏感差异 |", "下一步我想继续了解这段关系结束后的直接变化。", ].join("\n"), + evidenceRequest: null, } satisfies RectificationNarrativeModelOutput; assert.deepEqual(validateNarrativeAgainstPacket(output, packet, "intermediate"), { valid: true, issues: [] }); @@ -658,7 +696,7 @@ test("accepts a natural reply that mentions several known dates before asking a evidenceRequest: { domains: ["career" as const], datePrecision: "month_preferred" as const, - prompt: "毕业后的第一份工作是什么时候开始的？", + prompt: "毕业后的第一份工作是直接入职，还是先休息了一段时间？", }, } satisfies RectificationNarrativeModelOutput; @@ -827,6 +865,7 @@ test("allows natural discussion of an event after it has contributed to scoring" assert.ok(output.evidenceRequest); const conversational = { ...output, + narrative: "我理解，这次离开不只是换工作。你当时为什么辞职，是主动还是被动，对生活有什么影响？", evidenceRequest: { ...output.evidenceRequest, prompt: "你当时为什么辞职，是主动还是被动，对生活有什么影响？", @@ -924,6 +963,7 @@ test("replaces legacy model-selected evidence domains instead of rejecting the a packet, generator: generator([{ ...output, + narrative: `${output.narrative}\n${output.evidenceRequest?.prompt ?? ""}`, evidenceRequest: { ...output.evidenceRequest, domains: ["finance"], diff --git a/frontend/tests/conversational-rectification-orchestrator.test.ts b/frontend/tests/conversational-rectification-orchestrator.test.ts index 372ff745..10c98a1c 100644 --- a/frontend/tests/conversational-rectification-orchestrator.test.ts +++ b/frontend/tests/conversational-rectification-orchestrator.test.ts @@ -126,6 +126,10 @@ function validGenerator( } const request = JSON.parse(prompt) as { phase: "first" | "intermediate" | "final"; + conversation?: { + recentConversation?: Array<{ role: "assistant" | "user"; text: string }>; + latestUserText?: string; + }; conversationContext?: { recentConversation?: Array<{ role: "assistant" | "user"; text: string }>; latestEvidence?: Array<{ dateLabel: string; summary: string }>; @@ -140,45 +144,85 @@ function validGenerator( dateLabel: string; }>; }; - packet: Omit, "candidate"> & { + packet?: Omit, "candidate"> & { candidate: ReturnType["candidate"] & { rangeStart: string; rangeEnd: string; }; }; }; - const value = request.packet; - const hasRelationshipEvidence = request.conversationContext?.eventLedger - ?.some((item) => item.active && /关系|恋爱|分手|结婚|离婚|伴侣/.test(item.summary)) === true; - const domains = hasRelationshipEvidence - ? ["career" as const] - : ["relationship" as const]; - const nextDomain = domains[0] === "relationship" ? "重要关系" : "事业"; - const latest = request.conversationContext?.latestEvidence?.at(-1); + const fallbackPacket = packet(false); + const value = request.packet ?? { + ...fallbackPacket, + candidate: { + ...fallbackPacket.candidate, + rangeStart: fallbackPacket.candidate.range.startTime, + rangeEnd: fallbackPacket.candidate.range.endTime, + }, + }; + const ordinaryConversation = request.phase === "intermediate" && request.conversation !== undefined; + const latestUserText = request.conversation?.latestUserText ?? request.conversationContext?.recentConversation?.at(-1)?.text ?? ""; + const latest = request.conversationContext?.latestEvidence?.at(-1) + ?? (latestUserText ? { dateLabel: "", summary: latestUserText } : undefined); const latestActiveEvent = request.conversationContext?.eventLedger ?.filter((item) => item.active) .at(-1); const unresolved = request.conversationContext?.unresolvedEvidence?.at(-1); + const hasRelationshipEvidence = request.conversationContext?.eventLedger + ?.some((item) => item.active && /关系|恋爱|分手|结婚|离婚|伴侣/.test(item.summary)) === true; + const domains = hasRelationshipEvidence ? ["career" as const] : ["relationship" as const]; + const nextDomain = domains[0] === "relationship" ? "重要关系" : "事业"; const asksForLatestDetail = continueLatestEvent && request.phase === "intermediate" - && latestActiveEvent !== undefined; + && (ordinaryConversation ? latest !== undefined : latestActiveEvent !== undefined); + const asksForMissingDate = ordinaryConversation + && latest !== undefined + && !/(?:19|20)\d{2}|\d{1,2}\s*月|(?:同年|当年|那年|来年)/u.test(latestUserText) + && !/^(?:是|对|没错|确认|不是|不对)/u.test(latestUserText); const freeNarrative = request.phase !== "final" && freeNarrativeFromGeneration !== undefined && generation >= freeNarrativeFromGeneration; + + if (ordinaryConversation) { + const question = asksForMissingDate + ? "这大约发生在哪一年、哪一月？" + : asksForLatestDetail ? detailQuestion : null; + const narrative = freeNarrative + ? "我明白。你可以顺着这段经历继续说，也可以自然讲下一件想到的事。" + : [ + latest ? `你说的“${latest.summary}”，我理解了。` : "我在听。", + varyNarrative ? `这是第 ${generation} 次合成措辞。` : "", + question ?? "", + ].join(""); + return { text: JSON.stringify({ + narrative, + evidenceRequest: question ? { + domains, + datePrecision: "month_preferred", + prompt: question, + followUp: asksForMissingDate + ? { kind: "event_date" } + : mislabelLatestDetailAsNewEvent + ? { kind: "new_event", evidenceId: null } + : { kind: "event_detail" }, + } : null, + }) }; + } + const narrative = freeNarrative - ? "我明白，这段经历已经记下。你可以继续按自己的节奏讲。" + ? "我明白。你可以顺着这段经历继续说，也可以自然讲下一件想到的事。" : [ - request.phase === "intermediate" && latest - ? `记下了：${latest.dateLabel} · ${latest.summary}。` - : `当前仍在核对 ${value.candidate.rangeStart}–${value.candidate.rangeEnd} 的候选范围，不能视为已经确认的出生分钟。`, - varyNarrative ? `这是第 ${generation} 次合成措辞。` : "", - request.phase === "final" - ? "当前证据已形成候选总结。" - : unresolved?.dateLabel === "日期待补充" - ? `我先把“${unresolved.summary}”这件事补完整：它大约发生在哪一年、哪一月？` - : asksForLatestDetail - ? detailQuestion - : `先说一件已经发生的${nextDomain}经历好吗？请写明哪一年、哪一月以及发生了什么。`, + request.phase === "intermediate" && latest + ? `你刚才说到${latest.dateLabel ? `${latest.dateLabel} · ` : ""}${latest.summary}。` + : `当前仍在核对 ${value.candidate.rangeStart}–${value.candidate.rangeEnd} 的候选范围，不能视为已经确认的出生分钟。`, + varyNarrative ? `这是第 ${generation} 次合成措辞。` : "", + request.phase === "final" + ? "当前证据已形成候选总结。" + : unresolved?.dateLabel === "日期待补充" + ? `我先把“${unresolved.summary}”这件事补完整：它大约发生在哪一年、哪一月？` + : asksForLatestDetail + ? detailQuestion + : `先说一件已经发生的${nextDomain}经历好吗？请写明哪一年、哪一月以及发生了什么。`, ].join(""); return { text: JSON.stringify({ narrative, @@ -197,15 +241,15 @@ function validGenerator( prompt: unresolved?.dateLabel === "日期待补充" ? `“${unresolved.summary}”大约发生在哪一年、哪一月？` : asksForLatestDetail - ? detailQuestion - : `请说一件已经发生的${nextDomain}经历，并写明哪一年、哪一月以及发生了什么。`, + ? detailQuestion + : `请说一件已经发生的${nextDomain}经历，并写明哪一年、哪一月以及发生了什么。`, followUp: unresolved?.dateLabel === "日期待补充" ? { kind: "event_date", evidenceId: unresolved.id } : asksForLatestDetail - ? mislabelLatestDetailAsNewEvent - ? { kind: "new_event", evidenceId: null } - : { kind: "event_detail", evidenceId: latestActiveEvent.id } - : { kind: "new_event", evidenceId: null }, + ? mislabelLatestDetailAsNewEvent + ? { kind: "new_event", evidenceId: null } + : { kind: "event_detail", evidenceId: latestActiveEvent?.id ?? null } + : { kind: "new_event", evidenceId: null }, }, }) }; }, @@ -1068,7 +1112,8 @@ test("a concrete event without a date is acknowledged and a date-only follow-up summary: item.summary, dateLabel: item.dateLabel, })), [{ summary: "我离开家去北京开始工作", dateLabel: "2023-03" }]); - assert.match(completed.narrative, /记下了：2023-03 · 我离开家去北京开始工作/); + assert.match(completed.narrative, /2023年3月/); + assert.doesNotMatch(completed.narrative, /记下了|已记录|校时价值|候选时间/); }); test("a descriptive date follow-up completes the targeted event instead of becoming a new event", async () => { @@ -1340,19 +1385,16 @@ test("an affirmative reply confirms the date proposed by the previous Agent turn assert.notEqual(completed.evidenceRequest?.followUp?.evidenceId, target.id); const prompt = JSON.parse(value.narrativePrompts.at(-1) ?? "{}") as { - conversationContext?: { + conversation?: { recentConversation?: Array<{ role: string; text: string }>; - previousEvidencePrompt?: string; - previousFollowUp?: { answerMode?: string; proposedDate?: { value: string } }; }; + conversationContext?: unknown; }; - assert.deepEqual(prompt.conversationContext?.recentConversation?.slice(-2), [ + assert.deepEqual(prompt.conversation?.recentConversation?.slice(-2), [ { role: "assistant", text: "你说实习到10月份然后辞职，这个10月是2020年10月吗？" }, { role: "user", text: "是的" }, ]); - assert.equal(prompt.conversationContext?.previousEvidencePrompt, "这个10月是2020年10月吗？"); - assert.equal(prompt.conversationContext?.previousFollowUp?.answerMode, "yes_no"); - assert.equal(prompt.conversationContext?.previousFollowUp?.proposedDate?.value, "2020-10"); + assert.equal(prompt.conversationContext, undefined); await value.service.answer(userId, { type: "answer", @@ -1624,7 +1666,7 @@ test("a dated independent event is not swallowed by a broad detail question misl assert.equal(next.evidenceRecap.some((item) => item.id === firstId), true); }); -test("the next evidence request moves past a domain the user already answered", async () => { +test("a completed event does not force the Agent into the next scoring domain", async () => { const value = harness({ readyAfterEvidenceCount: 99 }); await start(value, null); @@ -1636,9 +1678,8 @@ test("the next evidence request moves past a domain the user already answered", answer: "2020年5月结婚", }); - assert.deepEqual(turn.evidenceRequest?.domains, ["career"]); - assert.match(turn.narrative, /事业/); - assert.doesNotMatch(turn.narrative, /下一步[^\n]*重要关系/); + assert.equal(turn.evidenceRequest, null); + assert.doesNotMatch(turn.narrative, /下一步|事业|重要关系|校时价值|候选时间/); }); test("a rejected intermediate narrative falls back without discarding the event", async () => { @@ -1652,7 +1693,7 @@ test("a rejected intermediate narrative falls back without discarding the event" const stored = value.cases.get(startActionId)?.row; assert.equal(turn.turnVersion, 1); - assert.match(turn.narrative, /按自己的节奏/); + assert.match(turn.narrative, /顺着这段经历继续说|自然讲下一件/); assert.equal(stored?.privateCandidate.resultId, null); assert.equal(stored?.eventEvidence.length, 1); assert.equal(stored?.validationReceipts.length, 2); @@ -1689,7 +1730,7 @@ test("one through three supported events save and narrate before the fourth accu assert.equal(value.events.filter((event) => event === "narrative").length, 4); }); -test("intermediate narrative receives the complete active event ledger", async () => { +test("intermediate narrative receives conversation text without the private event ledger", async () => { const value = harness({ readyAfterEvidenceCount: 99 }); await start(value, null); await value.service.answer(userId, { @@ -1710,7 +1751,7 @@ test("intermediate narrative receives the complete active event ledger", async ( const prompt = value.narrativePrompts.at(-1) ?? ""; assert.match(prompt, /2020年9月底主动离开研究单位/); assert.match(prompt, /2023年4月进入下一家公司/); - assert.match(prompt, /eventLedger/); + assert.doesNotMatch(prompt, /eventLedger|unresolvedEvidence|expertWorkflow|candidateWeights/); }); test("a plateaued non-confirmable conversational case returns a bounded candidate after current domains are covered", async () => { @@ -1734,7 +1775,7 @@ test("a plateaued non-confirmable conversational case returns a bounded candidat }); if (index === 2) { assert.equal(latest.status, "active"); - assert.deepEqual(latest.evidenceRequest?.domains, ["relationship"]); + assert.equal(latest.evidenceRequest, null); } } @@ -1781,7 +1822,7 @@ test("an unanswered suggested domain keeps a plateaued candidate conversational" } assert.equal(latest?.status, "active"); - assert.deepEqual(latest?.evidenceRequest?.domains, ["finance"]); + assert.equal(latest?.evidenceRequest, null); assert.deepEqual(latest?.actions, ["answer", "pause", "abandon"]); }); @@ -1815,7 +1856,8 @@ test("answer persists after pause when the packet has no grounded follow-up doma assert.equal(turn.status, "active"); assert.equal(turn.evidenceRequest, null); - assert.match(turn.narrative, /记下了/); + assert.match(turn.narrative, /2020年4月进入研究院实习/); + assert.doesNotMatch(turn.narrative, /记下了|已记录|校时价值|候选时间/); }); test("system-only blockers return a bounded result without waiting for another plateau", async () => { @@ -1939,7 +1981,8 @@ test("vague, future, and unmatched answers stay conversational and never score", }); assert.equal(value.counts().packetBuilds, 1); assert.equal(turn.status, "active"); - assert.match(turn.narrative, /哪一年|哪一月|年月|已发生|已经发生|换个方向|未来/); + assert.ok(turn.narrative.length > 0); + assert.match(turn.narrative, new RegExp(answer.replace(/[.*+?^${}()|[\]\\]/g, "\\$&"))); assert.doesNotMatch(turn.narrative, /好的，我们不沿用不符合你的方向|已保存这段描述|我已保存你的原话|这条更正已保存/); assert.doesNotMatch(turn.narrative, /A[.、:]|B[.、:]|2006.?2011/); assert.equal(value.cases.get(startActionId)?.row.eventEvidence.at(-1)?.scoreable, false); @@ -1947,7 +1990,11 @@ test("vague, future, and unmatched answers stay conversational and never score", }); test("a free Agent reply clears the previous structured question instead of inheriting it", async () => { - const value = harness({ readyAfterEvidenceCount: 99, freeNarrativeFromGeneration: 2 }); + const value = harness({ + readyAfterEvidenceCount: 99, + continueLatestEvent: true, + freeNarrativeFromGeneration: 2, + }); await start(value, null); const first = await value.service.answer(userId, { @@ -1962,7 +2009,7 @@ test("a free Agent reply clears the previous structured question instead of inhe }); assert.equal(second.evidenceRequest, null); - assert.match(second.narrative, /按自己的节奏/); + assert.match(second.narrative, /顺着这段经历继续说|自然讲下一件/); assert.doesNotMatch(second.narrative, /[？?]/); }); @@ -2093,7 +2140,7 @@ test("a failed regenerate saves a fallback turn without changing evidence, candi const storedAfter = value.cases.get(startActionId)?.row; assert.equal(regenerated.turnVersion, answered.turnVersion + 1); - assert.match(regenerated.narrative, /按自己的节奏/); + assert.match(regenerated.narrative, /顺着这段经历继续说|自然讲下一件/); assert.deepEqual(storedAfter?.eventEvidence, evidenceBefore); assert.deepEqual(storedAfter?.privateCandidate, candidateBefore); assert.equal(value.counts().reserveCount, countsBefore.reserveCount);