diff --git a/docs/BUG_HISTORY.md b/docs/BUG_HISTORY.md index abaf5850..61ac2ab2 100644 --- a/docs/BUG_HISTORY.md +++ b/docs/BUG_HISTORY.md @@ -5318,8 +5318,8 @@ - 根因：第一件带日期事件之后下一问固定走感情 → 事业 → 家人 → 职业。精度阶段/分盘差异问句不挡出牌，`session_outcome=adopt_representative` 时把它们放进 `deferred_followup`。探针年份锁在 choice_frame 里，但问句轮不到。 - 修复：已有带日期事件且仍有大运冲突探针时，`source=event_probe` 先问该年前事并挡住出牌。A 写入并重算以筛窗。年龄带探针不插队。空账本仍只自然语言收集。不打开唯一分钟确认门。折进未提交的 Skill 10.0.11。 - 验证：`frontend/tests/rectification-eight-method.test.ts` 锁定教育事件 + 事业激活探针先问 2018 年入职且 `session_outcome` 保持收集；年龄带探针不插在感情收集前。 -- 防复发：大运冲突探针在采用前必须挡住出牌。不得在空账本挂 `choice_frame`。不得把 `confirmation_allowed` 改成 true。 -- 相关记录：BUG-348、BUG-350 +- 防复发：大运冲突探针在采用前必须挡住出牌。不得在空账本挂 `choice_frame`。不得把 `confirmation_allowed` 改成 true。不得在采用门事件/领域未齐时插队（见 BUG-386）。 +- 相关记录：BUG-348、BUG-350、BUG-386 - 复发自：BUG-348（收集后再区分；区分问句被方法轮询和出牌 defer 排到采用之后） - 修复版本：9873ba42 @@ -5600,13 +5600,13 @@ - 状态：resolved - 首次发现：2026-08-25 - 最近更新：2026-08-25 -- 影响面：`rectification-record-evidence-batch`、`projectTurnDecision`、`persistServerOwnedFocus`、生时纠正系统提示 +- 影响面：`rectification-record-evidence-batch`、`projectTurnDecision`、`persistServerOwnedFocus`、`bindSpokenToOpenQuestion`、生时纠正运行器 - 用户现象：刚记下大学毕业后，气泡问「2016 年前后入学考试有没有发挥失常」，点选卡却是「2023 年前后有没有入职或职责加重」。题干和选项是服务器模板，没有跟口语对齐。 - 触发条件：账本已有带日期学业事件；引擎发出 dasha 冲突探针并盖戳点选卡；本轮只调了 `record-evidence-batch`，没有 `compare-candidates`。 -- 根因：(1) 证据写入后的自动重算会 `persistServerOwnedFocus`，但 batch 工具结果没有把 `open_question.prompt` 回给模型。(2) `turn_decision.current_question` 只有 id，没有题干。模型按上一轮学业对话另起了考试质量问，界面 GET 却展示已盖戳的事业卡。(3) 选题必须用刚算完的 decision receipt；若只信 persist RPC 回包，探针可能被丢掉。点选卡仍由服务器出，保证点选直接改后验，这不是本次缺陷。 -- 修复：batch / confirm 重算后返回 `open_question.prompt`；选题用刚写入的 receipt（含探针）。`current_question` 带上同一题干。系统提示要求正文只问这一句，不得另起高考发挥。不改 Skill `10.0.11`。 -- 验证：`frontend/tests/rectification-eight-method.test.ts`、`frontend/tests/rectification-answer-choice.test.ts`、`frontend/tests/rectification-v9-agent.test.ts`。 -- 防复发：有点选卡时口语必须是 `open_question.prompt` / `current_question.prompt`，不得另问一层。证据写入后的工具结果必须带回已持久化题干。不得改已哈希 Skill `10.0.11`。 +- 根因：(1) 证据写入后的自动重算会 `persistServerOwnedFocus`，但 batch 工具结果没有把 `open_question.prompt` 回给模型。(2) `turn_decision.current_question` 只有 id，没有题干。模型按上一轮学业对话另起了考试质量问，界面 GET 却展示已盖戳的事业卡。(3) 选题必须用刚算完的 decision receipt；若只信 persist RPC 回包，探针可能被丢掉。(4) 用系统提示逐条禁止高考发挥/入学年份堵不住用户回答的多种写法。点选卡仍由服务器出，保证点选直接改后验。 +- 修复：batch / confirm 重算后返回 `open_question.prompt`；选题用刚写入的 receipt。有已盖戳题干时，运行器丢掉口语里的其它问句并接到这句题干，不在提示词里枚举禁止主题。不改 Skill `10.0.11`。 +- 验证：`frontend/tests/rectification-eight-method.test.ts`、`frontend/tests/rectification-answer-choice.test.ts`、`frontend/tests/rectification-v9-agent.test.ts`、`frontend/tests/rectification-v9-stream.test.ts`。 +- 防复发：有点选卡时口语由运行器接到 `open_question.prompt` / `current_question.prompt`，不得靠提示词枚举禁止追问。证据写入后的工具结果必须带回已持久化题干。不得改已哈希 Skill `10.0.11`。 - 相关记录：BUG-375、BUG-379、BUG-382 - 复发自：BUG-375（点选卡已由服务器盖戳，但模型仍按探针列表另写一问） - 修复版本：待发布 @@ -5627,6 +5627,22 @@ - 复发自：BUG-381（避让点选卡时把居中改成了靠右） - 修复版本：待发布 +## BUG-386 | 采用门未齐就按冲突时间出反推前事点选卡 + +- 状态：resolved +- 首次发现：2026-08-25 +- 最近更新：2026-08-25 +- 影响面：`method_followup_plan`、`meetsAcceptanceEventQuality`、生时纠正系统提示 +- 用户现象：刚记下学业等一两件带日期经历后，点选卡就按 dasha 冲突年反问另一领域「某年前后有没有入职」。方法层感情/事业还没问完，采用门也还没到 3 件可评分事件、2 个领域。 +- 触发条件：账本已有至少一件可评分事件，但不足 3 件或不足 2 个领域；decision receipt 含 `dasha_activation` / `dasha_boundary` 探针。 +- 根因：BUG-351 在第一件带日期事件之后就让冲突探针插队并盖戳点选卡。引擎采用门仍要 3 件/2 领域，访谈却提前进入反推前事。 +- 修复：冲突探针插队改到与采用门相同的事件质量门槛：至少 3 件主评分事件、至少 2 个领域。`occupation_note` 与外貌/疤痕/占问不计。未齐时继续感情→事业→家人→职业自然语言收集，不盖冲突卡。齐了之后仍按 BUG-351 插队并挡出牌。不改 Skill `10.0.11`，不打开 `confirmation_allowed`。 +- 验证：`frontend/tests/rectification-eight-method.test.ts`、`frontend/tests/rectification-v9-agent.test.ts`、`frontend/tests/rectification-agentic-entry.test.ts`。 +- 防复发：不得在采用门事件/领域未齐时把 dasha 冲突探针写成 `source=event_probe` 或挂 `choice_frame`。不得把 `occupation_note` 算进 3 件。不得改已哈希 Skill `10.0.11`。 +- 相关记录：BUG-351、BUG-361、BUG-379 +- 复发自：BUG-351（一件带日期事件后冲突探针插队，早于采用门） +- 修复版本：待发布 + ## BUG-379 | 生时纠正已记入学后仍编造高考年并再问入学 - 状态：resolved diff --git a/frontend/src/lib/rectification-agentic/v9/agent-run.ts b/frontend/src/lib/rectification-agentic/v9/agent-run.ts index f9936755..97ce07e8 100644 --- a/frontend/src/lib/rectification-agentic/v9/agent-run.ts +++ b/frontend/src/lib/rectification-agentic/v9/agent-run.ts @@ -44,7 +44,12 @@ import { createStepAnswerState, flushStepAnswerOnStreamFinish, } from "./step-answer"; -import { composeRectificationTurnNarration, publicNarrationDtoFromDossier } from "./turn-narration"; +import { + bindSpokenToOpenQuestion, + composeRectificationTurnNarration, + openQuestionPromptFromToolResult, + publicNarrationDtoFromDossier, +} from "./turn-narration"; import { defaultMessageOrigin, isRectificationMessageOrigin, @@ -630,11 +635,16 @@ export async function runV9AgentTurn(options: V9AgentRunOptions): Promise | null = null; const stepAnswer = createStepAnswerState(); + let persistedPrompt: string | null = null; + const heldSpoken: string[] = []; + const publishSpokenStep = async (pieces: readonly string[]) => { - // The model's terminal text-delta is the user-visible reply. Do not - // regex-split it, and do not replace it with Case narration. const spoken = pieces.join("").trim(); if (!spoken || !caseLoaded) return; + if (persistedPrompt) { + heldSpoken.push(spoken); + return; + } answerText += spoken; answerDeltas.push(spoken); await emit({ type: "answer.delta", text: spoken }); @@ -663,6 +673,9 @@ export async function runV9AgentTurn(options: V9AgentRunOptions): Promise = new Set([ "horary_query", ]); +/** Notes that may cover a method layer but do not count as primary scoring events. */ +export const AUXILIARY_EVIDENCE_KINDS: ReadonlySet = new Set([ + "appearance_note", + "birthmark_or_scar", + "occupation_note", +]); + +export const NON_PRIMARY_SCORING_DOMAINS: ReadonlySet = new Set([ + "appearance", + "marks", + "occupation", + "horary", + "other", +]); + +/** Same floors as `scripts/rectification/decision_policy.py`. */ +export const MIN_ACCEPTANCE_EVENTS = 3; +export const MIN_ACCEPTANCE_DOMAINS = 2; + export function isBackgroundEvidenceKind(kind: EvidenceKind): boolean { return BACKGROUND_ONLY_KINDS.has(kind); } +export function isPrimaryScoreableEvidence(item: { + status: string; + domain: string; + datePrecision: string; + occurredFrom: string | null; + occurredTo: string | null; + eventKind?: string | null; +}): boolean { + if (item.status !== "confirmed") return false; + if (item.datePrecision === "unknown") return false; + if (!item.occurredFrom && !item.occurredTo) return false; + if (NON_PRIMARY_SCORING_DOMAINS.has(item.domain)) return false; + const kind = item.eventKind; + if ( + kind === "other" + || kind === "horary_query" + || kind === "appearance_note" + || kind === "birthmark_or_scar" + || kind === "occupation_note" + ) { + return false; + } + return true; +} + +/** Reverse-inference / conflict probes wait until the engine could accept. */ +export function meetsAcceptanceEventQuality( + evidence: readonly Readonly<{ + status: string; + domain: string; + datePrecision: string; + occurredFrom: string | null; + occurredTo: string | null; + eventKind?: string | null; + }>[], +): boolean { + const scoreable = evidence.filter(isPrimaryScoreableEvidence); + const domains = new Set(scoreable.map((item) => item.domain)); + return scoreable.length >= MIN_ACCEPTANCE_EVENTS && domains.size >= MIN_ACCEPTANCE_DOMAINS; +} + /** Ledger/engine subject follows the domain. Family events must not stay on the tool default `self`. */ export function evidenceSubjectForDomain( domain: string, diff --git a/frontend/src/lib/rectification-agentic/v9/method-followup.ts b/frontend/src/lib/rectification-agentic/v9/method-followup.ts index 9e0f8e0d..e65e6ce5 100644 --- a/frontend/src/lib/rectification-agentic/v9/method-followup.ts +++ b/frontend/src/lib/rectification-agentic/v9/method-followup.ts @@ -24,9 +24,10 @@ * Appearance and marks are skipped_by_policy. Horary does not block offering * time cards. Occupation does block cards until a note exists. * Method coverage asks for dated events in natural language. - * After the first dated event, remaining dasha conflict probes - * (year/activation differences) are asked before more method rotation - * and they block offering time cards so the window can be filtered. + * Dasha conflict probes wait until acceptance event quality + * (3 primary scoreable events in 2 domains), then jump ahead of + * remaining method rotation and block offering time cards so the + * window can be filtered. * Once blocking methods are covered, move into candidate discrimination. * Coverage complete never means adopt. Horary does not block cards. * A/B/C/D choice frames attach only when candidates already diverge @@ -52,6 +53,7 @@ import { type CandidateDiscriminatorProbe, } from "../core/candidate-contrast-packet.ts"; import type { SessionOutcomeKind } from "./confirmation-gate.ts"; +import { meetsAcceptanceEventQuality } from "./evidence-model"; import type { DiscriminatingEventProbe, NakshatraBoundary, @@ -764,7 +766,7 @@ export function buildMethodFollowupPlan(input: { ...(input.askedProbeKeys ?? []), ...askedKeysFromLedgerEvidence(input.evidence), ]); - const conflictProbe = dashaCovered + const conflictProbe = dashaCovered && meetsAcceptanceEventQuality(input.evidence) ? remainingConflictProbes(input.eventProbes, input.evidence, declined, askedKeys)[0] ?? null : null; if (!dashaCovered) { diff --git a/frontend/src/lib/rectification-agentic/v9/turn-narration.ts b/frontend/src/lib/rectification-agentic/v9/turn-narration.ts index 82dd9923..71a1a113 100644 --- a/frontend/src/lib/rectification-agentic/v9/turn-narration.ts +++ b/frontend/src/lib/rectification-agentic/v9/turn-narration.ts @@ -30,3 +30,47 @@ export function composeRectificationTurnNarration(dto: RectificationNarrationDto } return parts.join(""); } + +function promptFromQuestionField(value: unknown): string | null { + if (!value || typeof value !== "object" || Array.isArray(value)) return null; + const prompt = (value as { prompt?: unknown }).prompt; + if (typeof prompt !== "string") return null; + const text = prompt.trim(); + return text.length > 0 ? text : null; +} + +export function readOpenQuestionPrompt(value: unknown): string | null { + if (!value || typeof value !== "object" || Array.isArray(value)) return null; + const row = value as Record; + return promptFromQuestionField(row.open_question) + ?? promptFromQuestionField(row.current_question); +} + +export function openQuestionPromptFromToolResult(chunk: { + type?: string; + payload?: { result?: unknown; output?: unknown }; + object?: unknown; +}): string | null { + if (chunk.type !== "tool-result") return null; + return readOpenQuestionPrompt(chunk.payload?.result) + ?? readOpenQuestionPrompt(chunk.payload?.output) + ?? readOpenQuestionPrompt(chunk.object) + ?? readOpenQuestionPrompt(chunk.payload); +} + +/** + * When a choice card is already stamped, the persisted prompt is the only + * follow-up. Keep non-interrogative acknowledgements; drop any other ask. + * This is not a topic denylist — user answers stay free text or taps. + */ +export function bindSpokenToOpenQuestion(spoken: string, nextQuestion: string | null): string { + const question = nextQuestion?.trim() ?? ""; + if (!question) return spoken.trim(); + const ack = spoken + .split(/\n{2,}/) + .flatMap((block) => block.split(/(?<=[。！])\s*/u)) + .map((part) => part.trim()) + .filter((part) => part.length > 0 && !/[？?]/.test(part) && !part.includes(question)) + .slice(0, 2); + return [...ack, question].join("\n\n"); +} diff --git a/frontend/src/mastra/agentic-rectification.ts b/frontend/src/mastra/agentic-rectification.ts index 21c2e76d..3edd7b84 100644 --- a/frontend/src/mastra/agentic-rectification.ts +++ b/frontend/src/mastra/agentic-rectification.ts @@ -69,9 +69,9 @@ const agenticRectificationInstructions = `你是 Jyotisha，只服务当前绑 6. 工具执行过程保持静默。思考过程必须用简体中文，只写在思维链里：可以说你在核对哪类经历，禁止写工具名、错误码、参数、内部 ID、评分或密钥。对用户说的话必须自己写在正文里，不要只写规划等服务器代写。正文像正常人说话，不写“本轮做了什么”，不描述 Skill、Case、Dossier、工具、内部 Activity、参数、错误或推理过程；完成凭证完全由服务端公开 Activity/receipt 展示。 7. 只基于成功 attempt 输出正文。工具失败时说明面向用户的边界，不声称未执行的方法或结果。 8. 当前轮新事件一律走 rectification-record-evidence-batch（一件也可以）。优先传 source 原文的 quoteStart/quoteEnd，不要改写 quote。rectification-confirm-evidence 只用于用户对已有 pending 明确说“对/是”。不得要求用户把已说清的事件再发一遍。 -9. 不得在同一回复中一边要求继续补证据，一边提供候选采用。落实 next_user_action：id=verify_adopted_time 时本轮只核一件前事，A 走 batch 并 compare，C 关闭该问，不要 offer 也不要 start_consultation。id=start_consultation 时请用户用当前采用时间看盘，对不上同时请改选其他候选。id 不是 adopt_representative 时不得调用 rectification-offer-candidates，也不得请用户采用。selection_allowed 只表示可以采用代表性时间，不是本轮必须出示卡片；propose_allowed 才是提出门。挡住出牌的方法层未齐时，source=event_probe 的冲突前事继续问并挡住出牌。方法覆盖已齐只进入候选区分，不等于 adopt。无日期 occupation_note 算职业已覆盖，不要再问职业，也不要因它出牌。id=ask_candidate_discriminator 或 session_outcome=discriminate_candidates 时按 candidate_contrast_packet / next_followup 问一件能拆开候选的前事，不得 offer。id=ask_holdout_validation 时做盘外核对，不得 offer。id=offer_provisional_range 时说明并列可信区间，不要称某分钟为当前推荐。accepted_time 为空且 session_outcome=adopt_representative 或 next_user_action.id=adopt_representative 时本轮结果是采用代表性时间，不要再问 next_followup；正文必须说本会话以代表性时间收口，不确认唯一分钟。unique_minute_path=closed_at_representative 时不得调用 confirm，不得把唯一分钟确认当下一步。用户说“暂时想不到了 / 没有更多 / 没有了 / 没了 / 没有其它 / 想不起来了 / 先这样”时改走 on_user_stop：账本为空则把已说的带日期经历 batch 写入再比较，有事件无结果则本轮 compare，已有代表性结果且尚未采用则解释、调用 offer-candidates 并请采用下方时间卡片，已采用则按 on_user_stop 看盘或改选。禁止只说记下了、会话会保留、以后再继续。出牌/采用轮把工具返回的 skill_verification_report 写入正文：筛选窗、事件–Dasha–Gochara 表、D9/D10 类型对照、六亲六步、职业类型表、占问 observation_only、文末技法审计表。80%/60% 只描述事件吻合率，不得写成已确认唯一出生分钟，也不得写成候选已经分开。确认门以 latest_result.confirmation_gate 为准；not_evaluated 不是 fail；官方分钟层 passed 仍不能单独打开确认门；holdout 为 not_ready 时 unique_minute_path 必须是 closed_at_representative，不得声称精确分钟或发布准确率。若宽度大于 5 或 confirmation_allowed 为 false，必须说这是一段不可分区间，把代表分钟称为代表性候选，不得说已定位到唯一分钟。候选未拉开时不得出示赢家卡；D9/D10 差异和精度阶段追问要用来区分，不得直接宣布不可分。用户仍可 accepted 代表性候选。 +9. 不得在同一回复中一边要求继续补证据，一边提供候选采用。落实 next_user_action：id=verify_adopted_time 时本轮只核一件前事，A 走 batch 并 compare，C 关闭该问，不要 offer 也不要 start_consultation。id=start_consultation 时请用户用当前采用时间看盘，对不上同时请改选其他候选。id 不是 adopt_representative 时不得调用 rectification-offer-candidates，也不得请用户采用。selection_allowed 只表示可以采用代表性时间，不是本轮必须出示卡片；propose_allowed 才是提出门。采用门所需的可评分事件未齐（至少 3 件、2 个领域）时继续按方法层收集，不要根据冲突探针出点选卡或改问冲突年。齐了之后，source=event_probe 的冲突前事继续问并挡住出牌。方法覆盖已齐只进入候选区分，不等于 adopt。无日期 occupation_note 算职业已覆盖，不要再问职业，也不要因它出牌。id=ask_candidate_discriminator 或 session_outcome=discriminate_candidates 时按 candidate_contrast_packet / next_followup 问一件能拆开候选的前事，不得 offer。id=ask_holdout_validation 时做盘外核对，不得 offer。id=offer_provisional_range 时说明并列可信区间，不要称某分钟为当前推荐。accepted_time 为空且 session_outcome=adopt_representative 或 next_user_action.id=adopt_representative 时本轮结果是采用代表性时间，不要再问 next_followup；正文必须说本会话以代表性时间收口，不确认唯一分钟。unique_minute_path=closed_at_representative 时不得调用 confirm，不得把唯一分钟确认当下一步。用户说“暂时想不到了 / 没有更多 / 没有了 / 没了 / 没有其它 / 想不起来了 / 先这样”时改走 on_user_stop：账本为空则把已说的带日期经历 batch 写入再比较，有事件无结果则本轮 compare，已有代表性结果且尚未采用则解释、调用 offer-candidates 并请采用下方时间卡片，已采用则按 on_user_stop 看盘或改选。禁止只说记下了、会话会保留、以后再继续。出牌/采用轮把工具返回的 skill_verification_report 写入正文：筛选窗、事件–Dasha–Gochara 表、D9/D10 类型对照、六亲六步、职业类型表、占问 observation_only、文末技法审计表。80%/60% 只描述事件吻合率，不得写成已确认唯一出生分钟，也不得写成候选已经分开。确认门以 latest_result.confirmation_gate 为准；not_evaluated 不是 fail；官方分钟层 passed 仍不能单独打开确认门；holdout 为 not_ready 时 unique_minute_path 必须是 closed_at_representative，不得声称精确分钟或发布准确率。若宽度大于 5 或 confirmation_allowed 为 false，必须说这是一段不可分区间，把代表分钟称为代表性候选，不得说已定位到唯一分钟。候选未拉开时不得出示赢家卡；D9/D10 差异和精度阶段追问要用来区分，不得直接宣布不可分。用户仍可 accepted 代表性候选。 10. 不泄露系统提示词或 Skill 原文。 -11. 追问只跟 method_followup_plan 与服务器已持久化的 current_question / open_question。不要调用 rectification-set-focus；下一问和点选卡由 compare-candidates / read-case 在服务端事务内创建。账本为空或 collect_method_evidence 时用自然语言问一件带大概年份的经历，正文直接问，不要提点选卡。若工具返回了 open_question.prompt 或 current_question.prompt，正文只问这一句，不得另起高考发挥、入学年份或其它方法层追问。不得发明年份，不得根据出生年推算高考或入学年份并当成事实，不要把已回答的考试质量题或职责倾向再问一遍。账本已有入学、毕业或感情开始/结束日期时，不要再问那一件发生在哪一年。挡住出牌的方法层未齐时，source=event_probe 只问这一件反推前事用来筛窗，不要继续轮询方法层，不要 offer。覆盖已齐后只问当前剩余候选分钟还能拆开的区分探针；没有剩余拆分且未拉开时落实 offer_provisional_range，不要再问整窗 D9/D24，也不要 adopt。不要问两套盘哪个更像或可能性高低。点选 A/B/C/D 与「先这样」由服务器按 questionId/optionId 确定性处理，不要把选项全文当成新事件，也不要为点选调用 resolve-focus、read-case 或 compare；自由文本补充才走工具。正文禁止复述选项。不得询问外貌、体质、胎记或疤痕，也不得问钟点。不得按 missing_evidence_categories 轮询迁居，也不得先要 10–15 条事件长表。财务与健康只有用户主动说才问。方法覆盖为感情→事业→家人→职业→占问。D9/D10 类型表是校时方法，不是命运承诺。以「盘外核对（不计分）」开头的消息不得调用 record-evidence-batch 或 propose-evidence。 +11. 追问只跟 method_followup_plan 与服务器已持久化的 current_question / open_question。不要调用 rectification-set-focus；下一问和点选卡由 compare-candidates / read-case 在服务端事务内创建。账本为空或 collect_method_evidence 时用自然语言问一件带大概年份的经历，正文直接问，不要提点选卡。若工具返回了 open_question.prompt 或 current_question.prompt，不要另写追问；下一问由点选卡呈现，运行器会把口语接到这句题干。自由文本只作补充。采用门所需的可评分事件/领域未齐时不要走 event_probe，忽略 receipt 里的 dasha 冲突探针。齐了之后 source=event_probe 只问这一件反推前事用来筛窗，不要继续轮询方法层，不要 offer。覆盖已齐后只问当前剩余候选分钟还能拆开的区分探针；没有剩余拆分且未拉开时落实 offer_provisional_range，不要再问整窗 D9/D24，也不要 adopt。不要问两套盘哪个更像或可能性高低。点选 A/B/C/D 与「先这样」由服务器按 questionId/optionId 确定性处理，不要把选项全文当成新事件，也不要为点选调用 resolve-focus、read-case 或 compare；自由文本补充才走工具。正文禁止复述选项。不得询问外貌、体质、胎记或疤痕，也不得问钟点。不得按 missing_evidence_categories 轮询迁居，也不得先要 10–15 条事件长表。财务与健康只有用户主动说才问。方法覆盖为感情→事业→家人→职业→占问。D9/D10 类型表是校时方法，不是命运承诺。以「盘外核对（不计分）」开头的消息不得调用 record-evidence-batch 或 propose-evidence。 12. 证据有效变化后由服务器重算候选。不要等用户说“没有更多了”才比较，也不要对同一证据指纹再 compare。分钟扫描只在服务端，结果只是候选或平台，不得宣布确认。 13. 落实 start_consultation：前事核对结束或用户先这样后，请用户用当前采用时间看盘；对不上同时请改选其他候选。解释事件–Dasha 账本、双轨是否一致、换升时刻、精度阶段、D9/D10 类型对照和相对支持时，仍必须说候选范围不是出生时间真值。`; diff --git a/frontend/tests/rectification-agentic-entry.test.ts b/frontend/tests/rectification-agentic-entry.test.ts index d9ea620c..9a3d8ee4 100644 --- a/frontend/tests/rectification-agentic-entry.test.ts +++ b/frontend/tests/rectification-agentic-entry.test.ts @@ -243,6 +243,8 @@ test("usage completes or releases without hiding settlement failures", () => { assert.match(run, /thinkingTokens: 8_192/); assert.match(run, /activity.changed/); assert.match(run, /if \(!answerText\.trim\(\)\) \{[\s\S]*composeRectificationTurnNarration/); + assert.match(run, /bindSpokenToOpenQuestion/); + assert.match(run, /openQuestionPromptFromToolResult/); assert.doesNotMatch(run, /splitRectificationSpokenAndThinking/); assert.match(run, /emit\(event: PublicStreamEvent\): Promise \| void;/); assert.match(run, /agentGenerationSettings\(options\.generationModel, \{[\s\S]*thinking: "enabled"/); @@ -637,6 +639,8 @@ test("the Agent prompt cannot offer candidates while asking for more evidence", assert.match(agent, /id 不是 adopt_representative 时不得调用 rectification-offer-candidates/); assert.match(agent, /verify_adopted_time/); assert.match(agent, /event_probe/); + assert.match(agent, /至少 3 件/); + assert.match(agent, /2 个领域/); assert.match(agent, /selection_allowed 只表示可以采用代表性时间/); const tools = readFileSync( new URL("../src/mastra/rectification-v9-tools.ts", import.meta.url), diff --git a/frontend/tests/rectification-eight-method.test.ts b/frontend/tests/rectification-eight-method.test.ts index c78fd892..0d0858ed 100644 --- a/frontend/tests/rectification-eight-method.test.ts +++ b/frontend/tests/rectification-eight-method.test.ts @@ -37,6 +37,8 @@ import { const THIRD_CANDIDATE_ID = "88888888-8888-4888-8888-888888888883"; const EDUCATION_ID = "44444444-4444-4444-8444-444444444441"; +const RELATIONSHIP_ID = "44444444-4444-4444-8444-444444444442"; +const FAMILY_ID = "44444444-4444-4444-8444-444444444443"; const UNIQUE_MINUTE_COPY = /±5 分钟确定性/; const CLASSIC_COVERAGE = [ @@ -50,6 +52,39 @@ const CLASSIC_COVERAGE = [ { status: "confirmed", domain: "horary", datePrecision: "day" as const, occurredFrom: "2024-01-01", occurredTo: null }, ]; +function datedEvidence( + domain: string, + year: string, + extra: { + eventKind?: string | null; + summary?: string | null; + datePrecision?: "year" | "month" | "day"; + } = {}, +) { + return { + status: "confirmed" as const, + domain, + datePrecision: extra.datePrecision ?? ("year" as const), + occurredFrom: `${year}-01-01`, + occurredTo: null, + ...(extra.eventKind !== undefined ? { eventKind: extra.eventKind } : {}), + ...(extra.summary !== undefined ? { summary: extra.summary } : {}), + }; +} + +const CAREER_CONFLICT_PROBE = { + year: 2018, + year_label: "2018 年前后", + domain: "career" as const, + event_family: "入职、升职或职责明显加重", + source: "dasha_activation" as const, + tracks: ["vimshottari", "narayana"] as const, + tracks_agree: true, + unique_minute_claim: false as const, + user_meaning: "年份锁定 2018 年前后。请写成一句自然语言，问是否入职或职责加重。", + role: "reverse_verify" as const, +}; + const ENGINE_SCORE = { success: true, endpoint: "rectification_v5_score", @@ -113,6 +148,36 @@ const educationEvidence = { created_at: "2026-08-12T10:00:06.000Z", }; +const relationshipEvidence = { + id: RELATIONSHIP_ID, + source_turn_id: TURN_ID, + subject: "self", + event_kind: "relationship_start", + domain: "relationship", + occurred_from: "2018-01-01", + occurred_to: null, + date_precision: "year", + summary: "2018年一段感情开始", + status: "confirmed", + supersedes_evidence_id: null, + created_at: "2026-08-12T10:00:07.000Z", +}; + +const familyEvidence = { + id: FAMILY_ID, + source_turn_id: TURN_ID, + subject: "family", + event_kind: "family_event", + domain: "family", + occurred_from: "2020-01-01", + occurred_to: null, + date_precision: "year", + summary: "2020年家人相关变化", + status: "confirmed", + supersedes_evidence_id: null, + created_at: "2026-08-12T10:00:08.000Z", +}; + function stubEngine(response: unknown) { const previous = globalThis.fetch; globalThis.fetch = (async () => ({ @@ -142,27 +207,25 @@ test("eight-method routing asks relationship after dated education, not relocati assert.equal(plan.methods.find((item) => item.method_id === "appearance")?.status, "skipped_by_policy"); }); -test("dated evidence plus dasha conflict probe asks that event before method rotation and blocks offer", () => { +test("dasha conflict probe does not jump ahead of method rotation before acceptance event quality", () => { const plan = buildMethodFollowupPlan({ - evidence: [{ - status: "confirmed", - domain: "education", - datePrecision: "year", - occurredFrom: "2016-01-01", - occurredTo: null, - }], - eventProbes: [{ - year: 2018, - year_label: "2018 年前后", - domain: "career", - event_family: "入职、升职或职责明显加重", - source: "dasha_activation", - tracks: ["vimshottari", "narayana"], - tracks_agree: true, - unique_minute_claim: false, - user_meaning: "年份锁定 2018 年前后。请写成一句自然语言，问是否入职或职责加重。", - role: "reverse_verify", - }], + evidence: [datedEvidence("education", "2016")], + eventProbes: [CAREER_CONFLICT_PROBE], + }); + assert.equal(plan.next_followup?.source, "method_coverage"); + assert.equal(plan.next_followup?.method_id, "d9_relationship"); + assert.equal(plan.next_followup?.choice_frame, null); + assert.notEqual(plan.next_followup?.source, "event_probe"); +}); + +test("dasha conflict probe jumps after three scoreable events in two domains and blocks offer", () => { + const plan = buildMethodFollowupPlan({ + evidence: [ + datedEvidence("education", "2016"), + datedEvidence("education", "2020"), + datedEvidence("relationship", "2018"), + ], + eventProbes: [CAREER_CONFLICT_PROBE], }); assert.equal(plan.next_followup?.source, "event_probe"); assert.equal(plan.next_followup?.intent, "distinguish_candidates"); @@ -181,6 +244,41 @@ test("dated evidence plus dasha conflict probe asks that event before method rot }), "collect_evidence"); }); +test("three scoreable events in one domain still rotate methods instead of reverse-inferring", () => { + const plan = buildMethodFollowupPlan({ + evidence: [ + datedEvidence("education", "2012"), + datedEvidence("education", "2016"), + datedEvidence("education", "2020"), + ], + eventProbes: [CAREER_CONFLICT_PROBE], + }); + assert.equal(plan.next_followup?.source, "method_coverage"); + assert.equal(plan.next_followup?.method_id, "d9_relationship"); + assert.equal(plan.next_followup?.choice_frame, null); +}); + +test("occupation_note does not count toward reverse-inference event quality", () => { + const plan = buildMethodFollowupPlan({ + evidence: [ + datedEvidence("education", "2016"), + datedEvidence("education", "2020"), + { + status: "confirmed", + domain: "occupation", + datePrecision: "unknown", + occurredFrom: null, + occurredTo: null, + eventKind: "occupation_note", + }, + ], + eventProbes: [CAREER_CONFLICT_PROBE], + }); + assert.equal(plan.next_followup?.source, "method_coverage"); + assert.equal(plan.next_followup?.method_id, "d9_relationship"); + assert.equal(plan.next_followup?.choice_frame, null); +}); + test("age-band probe does not jump ahead of uncovered relationship", () => { const plan = buildMethodFollowupPlan({ evidence: [{ @@ -689,7 +787,8 @@ test("evidence batch returns the persisted choice prompt as open_question", asyn const accounting = fakeAccounting({ ...receiptHandlers, get_agentic_rectification_case_dossier: () => dossierFixture({ - evidence: [educationEvidence], + evidence: [educationEvidence, relationshipEvidence, familyEvidence], + evidenceCount: 3, latestResult: null, }), get_agentic_rectification_case_compute: () => computeFixture(), @@ -1500,24 +1599,13 @@ test("没有了 is a user stop", () => { test("same domain different year still asks a conflict probe", () => { const plan = buildMethodFollowupPlan({ - evidence: [{ - status: "confirmed", - domain: "career", - datePrecision: "year", - occurredFrom: "2015-01-01", - occurredTo: null, - }], + evidence: [ + datedEvidence("education", "2016"), + datedEvidence("relationship", "2018"), + datedEvidence("career", "2015"), + ], eventProbes: [{ - year: 2018, - year_label: "2018 年前后", - domain: "career", - event_family: "入职、升职或职责明显加重", - source: "dasha_activation", - tracks: ["vimshottari", "narayana"], - tracks_agree: true, - unique_minute_claim: false, - user_meaning: "年份锁定 2018 年前后。请写成一句自然语言，问是否入职或职责加重。", - role: "reverse_verify", + ...CAREER_CONFLICT_PROBE, information_gain: 0.21, semantic_key: "career.2018.dasha_activation", }], diff --git a/frontend/tests/rectification-v9-agent.test.ts b/frontend/tests/rectification-v9-agent.test.ts index 94cce34e..c879bad6 100644 --- a/frontend/tests/rectification-v9-agent.test.ts +++ b/frontend/tests/rectification-v9-agent.test.ts @@ -79,14 +79,17 @@ test("system prompt carries only high-priority boundaries, never the method copy assert.match(prompt, /不要调用 rectification-set-focus/); assert.match(prompt, /open_question\.prompt/); assert.match(prompt, /current_question\.prompt/); - assert.match(prompt, /不得另起高考发挥/); + assert.match(prompt, /不要另写追问/); + assert.match(prompt, /运行器会把口语接到这句题干/); + assert.doesNotMatch(prompt, /不得另起高考发挥/); + assert.doesNotMatch(prompt, /不得根据出生年推算高考或入学年份/); + assert.doesNotMatch(prompt, /不要再问那一件发生在哪一年/); assert.match(prompt, /「先这样」由服务器/); assert.match(prompt, /盘外核对（不计分）/); assert.match(prompt, /verify_adopted_time/); assert.match(prompt, /event_probe/); - assert.match(prompt, /不得发明年份/); - assert.match(prompt, /不得根据出生年推算高考或入学年份/); - assert.match(prompt, /不要再问那一件发生在哪一年/); + assert.match(prompt, /至少 3 件/); + assert.match(prompt, /2 个领域/); assert.doesNotMatch(prompt, /两套盘各自的前事/); assert.doesNotMatch(prompt, /外貌、体质、胎记或疤痕可以问/); assert.doesNotMatch(prompt, /分盘句和宫位表由界面展示/); diff --git a/frontend/tests/rectification-v9-stream.test.ts b/frontend/tests/rectification-v9-stream.test.ts index 6e060fa8..56b354e4 100644 --- a/frontend/tests/rectification-v9-stream.test.ts +++ b/frontend/tests/rectification-v9-stream.test.ts @@ -23,6 +23,10 @@ import { import { RECTIFICATION_SKILL_NAME } from "../src/lib/rectification-agentic/v9/case-status.ts"; import { messageContentHash } from "../src/lib/rectification-agentic/v9/message-origin.ts"; import { safeToolErrorCode } from "../src/lib/rectification-agentic/v9/tool-service.ts"; +import { + bindSpokenToOpenQuestion, + openQuestionPromptFromToolResult, +} from "../src/lib/rectification-agentic/v9/turn-narration.ts"; import { createRectificationActivityReceiptState, receiptFromRectificationActivityState, @@ -1358,6 +1362,44 @@ test("Chinese interview planning stays on reasoning-delta; the spoken answer is assert.equal(emitted.some((event) => event.type === "thinking.delta"), false); }); +test("persisted choice prompt replaces a competing model follow-up without a topic denylist", async () => { + const spoken = "好的，2020 年 6 月毕业这条也记下了。\n\n再问你一件：2016 年前后那场重要的入学考试，你当时发挥明显失常、或者压力特别大，有没有发生过？"; + const prompt = "2023 年前后，有没有明显入职、升职或职责明显加重？"; + assert.equal( + bindSpokenToOpenQuestion(spoken, prompt), + `好的，2020 年 6 月毕业这条也记下了。\n\n${prompt}`, + ); + assert.equal(openQuestionPromptFromToolResult({ + type: "tool-result", + payload: { result: { open_question: { prompt } } }, + }), prompt); + + const { options, emitted } = runOptions({ + buildAgent: async () => fakeAgentStream([ + chunk("start"), + chunk("tool-call", { toolName: "skill", args: { name: RECTIFICATION_SKILL_NAME } }), + chunk("tool-result", { toolName: "skill" }), + chunk("tool-call", { toolName: "rectification-read-case", args: { caseId: CASE_ID } }), + chunk("tool-result", { toolName: "rectification-read-case" }), + chunk("tool-call", { toolName: "rectification-record-evidence-batch", args: { caseId: CASE_ID } }), + chunk("tool-result", { + toolName: "rectification-record-evidence-batch", + result: { accepted_count: 1, open_question: { prompt } }, + }), + chunk("text-delta", { text: spoken }), + chunk("finish"), + ]) as never, + }); + const result = await runV9AgentTurn(options); + assert.equal(result.ok, true); + assert.equal(result.answerText, `好的，2020 年 6 月毕业这条也记下了。\n\n${prompt}`); + assert.doesNotMatch(result.answerText, /入学考试/); + assert.deepEqual( + emitted.filter((event) => event.type === "answer.delta"), + [{ type: "answer.delta", text: result.answerText }], + ); +}); + test("model terminal text-delta is the reply even when Case narration could be composed", async () => { const spoken = "职业已经记下。你入职大概是哪一年？说个年份就行。"; const { options, emitted } = runOptions({