From 3bbe65a2153b33ecfdfb6e23cf71b42b19559cb3 Mon Sep 17 00:00:00 2001 From: Jesse_Chen Date: Fri, 2 Oct 2026 09:03:22 +0800 Subject: [PATCH] fix(rectification): out-of-window event dates no longer fail every rescore; future dates refused at intake; corrections revise without a focus (BUG-1178, BUG-1179) Co-Authored-By: Claude Opus 5.5 Claude-Session: https://claude.ai/code/session_017eEAG8HD3mm8gsKXgk8uU8 --- CHANGELOG.md | 7 + docs/BUG_HISTORY.md | 30 ++++ ...RESS-rectification-event-dates-20261002.md | 33 +++++ docs/tasks/README.md | 1 + .../rectification-event-dates-20261002.md | 11 ++ frontend/docs/VOICE.md | 1 + .../lib/rectification-agentic/user-copy.ts | 3 + .../v9/agent-run-attempt.ts | 2 + .../v9/agent-run-support.ts | 9 +- .../rectification-agentic/v9/engine-client.ts | 29 +++- .../rectification-agentic/v9/host-fallback.ts | 9 ++ .../rectification-agentic/v9/turn-decision.ts | 5 +- frontend/src/mastra/agentic-rectification.ts | 2 +- frontend/src/mastra/rectification-v9-tools.ts | 132 +++++++++++++++--- ...ication-event-date-bounds-20261002.test.ts | 125 +++++++++++++++++ ...ctification-v10-conversation-focus.test.ts | 9 +- .../rectification-v10-tool-contract.test.ts | 7 + .../tests/rectification-v9-evidence.test.ts | 51 ++++++- 18 files changed, 439 insertions(+), 27 deletions(-) create mode 100644 docs/tasks/PROGRESS-rectification-event-dates-20261002.md create mode 100644 docs/testing/rectification-event-dates-20261002.md create mode 100644 frontend/tests/rectification-event-date-bounds-20261002.test.ts diff --git a/CHANGELOG.md b/CHANGELOG.md index 13ca3991..68aa65ba 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,5 +1,12 @@ # 印度占星 Skill 更新日志 +## 2026-10-02 — 生时校正：一件日期不对的经历不再卡死；说「打错了」就能改 + +- 修复：只要有一件经历的日期在今天之后（比如年份打错成 2028），或者是「今年 / 这个月」的事，之后每次都「没有重新比较」、一直追问不出卡。现在这类经历不再挡住比较，已经卡住的校正下一轮就会恢复（BUG-1178）。 +- 说到今天之后的日期时，回复会直说「这件先不算；如果年份打错了，直接告诉我正确的年份就行」，不再闷声记下。 +- 说「刚才 2028 打错了，是 2018」时，会改掉那一条并重新比较，不会再多记一条（BUG-1179）。撤回已答的选择题暂不支持。 +- Skill 版本不变；不改数据库结构。 + ## 2026-10-02 — 采用后核对题让盘型结论变了：说清采用的时间还在不在，需要时给「改用」 - 一件事对不上但盘型没变：回复说「一件事还不足以改变盘型结论，先按 HH:MM 排盘」，不再劝改选。 diff --git a/docs/BUG_HISTORY.md b/docs/BUG_HISTORY.md index c609d843..142a179d 100644 --- a/docs/BUG_HISTORY.md +++ b/docs/BUG_HISTORY.md @@ -15777,3 +15777,33 @@ - 复发自：无 - 修复版本：`codex/rectification-post-adopt-verify-20261001`。 +## BUG-1178 | 一件未来日期或「今年」的经历让之后每次重新比较都失败，校正一直追问、不出时间卡 + +- 状态：resolved（2026-10-02 Claude 直接执行；真引擎 + 真库验证） +- 首次发现 / 最近更新：2026-10-02 / 2026-10-02 +- 影响面：`v9/engine-client.ts::engineRequestBody`（score / diagnostics / block_scan / vedastro-validate 共用）、`mastra/rectification-v9-tools.ts`（record-evidence-batch、propose-evidence）、`v9/agent-run-support.ts` / `agent-run-attempt.ts` / `host-fallback.ts`（回复补句）、`user-copy.ts`。所有生时校正用户。 +- 用户现象（staging 真机测试者）：第一件经历把年份打成「2028 年借出一笔钱」，之后每轮回复都是「记下了：…这次没有重新比较，稍后再比一次。」，范围不动、不出时间卡，测试者「都不想回答了」。 +- 触发条件：账本里有任何一件开始日期晚于今天的经历；或任何「今年 / 这个月」的经历（年 / 月精度的结束日 12-31 / 月底晚于今天）；或完全在出生前的经历。 +- 根因：引擎 `scripts/rectification/contracts.py` 要求每件经历的日期都在 [出生日, 今天] 内，否则整个请求 400（`events[i] dates must be between birth_date and today`）。前端证据入口与数据库都不校验日期上限，未来年份照常记下且确认；之后每次重算都把它带上，整次失败，回复走 rescoreSkipped 文案。本机真引擎实测：2028 年、今年（年精度）、本月（月精度）都 400，去年 200。 +- 修复：(1) `boundEngineEventDates`：发给引擎前丢掉开始晚于上限（UTC 昨天，避免两端时钟差）或结束早于出生日的经历，其余裁进窗口；全部被丢则按既有 `no_scorable_evidence` 处理。已在账本里的未来经历不再挡住比较，该测试者的 Case 下一轮即可恢复。(2) 入口：record-evidence-batch 与 propose-evidence 对开始日期晚于今天（UTC）的项不写库，返回 `future_date`；回复由服务端追加「有一件事的日期在今天之后，这件先不算；如果年份打错了，直接告诉我正确的年份就行。」(3) 顺带修 batch 结果编号：已写入行的 `index` 改为模型原始条目序号（此前是写入序号，被拒条目在前时「记下了」复述会读错位置）。 +- 验证：`frontend/tests/rectification-event-date-bounds-20261002.test.ts`（裁剪六种日期、请求体只带裁剪后事件且全部被丢时报 no_scorable_evidence、入口判定、batch 中 2099 项不写库且其余条目序号与复述正确、回复补句只出现一次；去掉裁剪时请求体用例红）。真引擎 + 本机 PG17：虚构案例账本中加入 2028 年与今年的经历后重算成功（修前引擎 400）。 +- 防复发：任何发给引擎的事件都经 `engineRequestBody`；新增事件入口须做日期上限校验。旧版 `mastra/rectification-tools.ts`（仅 legacy session 引用）未改，记入进度。 +- 相关记录：BUG-1084（打字事件不计分研究）、BUG-1179。 +- 复发自：无 +- 修复版本：`codex/rectification-post-adopt-verify-20261001`。 + +## BUG-1179 | 用户说「刚才年份打错了」改不了：更正工具要求问题焦点，Agent 也拿不到经历编号 + +- 状态：resolved（2026-10-02 Claude 直接执行；产品 10-02 选 A；真库验证；真实模型调用为环境缺口） +- 首次发现 / 最近更新：2026-10-02 / 2026-10-02 +- 影响面：`mastra/rectification-v9-tools.ts::rectification-revise-evidence`、`v9/turn-decision.ts`（`relevant_evidence_summary`）、`mastra/agentic-rectification.ts`（提示规则）。 +- 用户现象：测试者问「回答错了这个能返回吗」，产品答「目前不能」。实际上已有 `rectification-revise-evidence`，但用不上。 +- 根因：(1) 工具 schema 强制 `focusId`，而随口更正一件打字经历时没有对应的问题焦点；(2) 修订后的新记录是 pending，要第二次调用确认才生效、才重算；(3) Agent 的经历摘要不带 `evidence_id`，无从指定改哪条；(4) 提示词没有「更正走修订、不要新记一条」的规则。 +- 决策记录：产品 10-02 选 A——先做「改经历」：用户说出正确年份即改正并重新比较；「撤回上一道选择题」不做（选项 B 未选）。推翻 08-14 v10 运行时「修订必须绑定问题焦点」的约束，改由「quote 必须落在本轮用户原话里」（与新事件同一校验）保证改动来自用户本人。 +- 修复：`focusId` 改为可选；有焦点走 `revise_..._v10`，无焦点走 `revise_agentic_rectification_evidence`（真库确认 service_role 可调用）；随后在同一调用里 `confirm_..._v10` 确认新修订并重算，返回 rescore / range_after_rescore；quote 未落在本轮用户消息里→`quote_mismatch` 不写；更正后的日期仍在未来→`future_date` 不写。经历摘要带 `evidence_id` 并排除已被替代的记录；提示规则：用户更正之前说过的事走修订工具。 +- 验证：`frontend/tests/rectification-v9-evidence.test.ts` 新增「无焦点更正：修订→确认、未落原话不写、未来日期不写」；三条既有合同用例按三栏改写（schema 允许省略 focusId 但仍须 evidenceId；修订后多一次确认）。本机 PG17：无焦点修订 + 确认后旧记录 superseded、新记录 confirmed。真实模型是否按新规则调用修订工具：本机无模型凭据，留待 staging 真机（`docs/testing/rectification-event-dates-20261002.md`）。 +- 防复发：经历的任何改动必须带本轮用户原话；修订后立即生效，不得停在 pending 等第二次调用。 +- 相关记录：BUG-1178。 +- 复发自：无 +- 修复版本：`codex/rectification-post-adopt-verify-20261001`。 + diff --git a/docs/tasks/PROGRESS-rectification-event-dates-20261002.md b/docs/tasks/PROGRESS-rectification-event-dates-20261002.md new file mode 100644 index 00000000..c997639f --- /dev/null +++ b/docs/tasks/PROGRESS-rectification-event-dates-20261002.md @@ -0,0 +1,33 @@ +# PROGRESS：经历日期越界卡死比较 + 更正经历（2026-10-02，BUG-1178～1179） + +- 模式：直接执行（产品 10-02「两条修法需要修，改错选 A」）。 +- 基线：`origin/staging` `8e978cab`；分支 `codex/rectification-post-adopt-verify-20261001`。 +- BUG 编号：staging 任务书已预留 1174～1177（普通对话），本单用 1178～1179。 + +## 实证 + +staging 真机测试者：「2028 年借出一笔钱」（年份打错）后每轮「这次没有重新比较」，不出卡。本机真引擎同一请求： + +| 追加的经历 | 引擎 | +| --- | --- | +| 2028 年（未来） | 400 `events[7] dates must be between birth_date and today` | +| 今年（年精度，止于 12-31） | 400 | +| 本月（月精度，止于月底） | 400 | +| 去年 | 200 | + +## 改动 + +| 位置 | 改动 | +| --- | --- | +| `engine-client.ts` | `boundEngineEventDates`：发引擎前丢未来 / 出生前事件，裁剪越界日期（上限 UTC 昨天） | +| `rectification-v9-tools.ts` | batch / propose 拒收未来日期（`future_date`，不写库）；batch 已写行 `index` 改用原始条目序号，复述按写入行读取；revise 工具 focus 可选、quote 须落本轮原话、修订后立即确认并重算 | +| `agent-run-*` / `host-fallback.ts` / `user-copy.ts` | 未来日期时服务端追加一句 | +| `turn-decision.ts` | 经历摘要带 `evidence_id`，排除 superseded | +| `agentic-rectification.ts` | 提示：用户更正走 revise-evidence | + +## 验证 + +- 新测试 `rectification-event-date-bounds-20261002.test.ts` 5 条；`rectification-v9-evidence` 新增无焦点更正用例；三条合同用例三栏改写。 +- 真库 + 真引擎：虚构案例账本加 2028 与今年经历后重算成功；无焦点修订 + 确认后旧记录 superseded。 +- 环境缺口：真实模型是否按新提示调用修订工具，本机无模型凭据，见真机清单。 +- 未改：`mastra/rectification-tools.ts`（仅旧版 `session.ts` 引用，非当前产品路径）同样直接发事件给引擎。 diff --git a/docs/tasks/README.md b/docs/tasks/README.md index 564d2a1e..c814e095 100644 --- a/docs/tasks/README.md +++ b/docs/tasks/README.md @@ -266,6 +266,7 @@ | `TASK-rectification-angle-timing-research-20261001.md` | `PROGRESS-rectification-angle-timing-research-20261001.md` | **角度类时间技法研究（离线）**：行运土木罗压本命上升 / 天顶度数、次限推运角、太阳弧角（1° ≈ 4 分钟出生时间），77 例 v5；乱序日期对照 ≥ 50 次必做、留一法、不得挤出真值；显著才叠加到六题后验比头段命中与区间宽度 | 已实现待验收（0/9 预登记检验显著，BUG-1141 closed_by_design；分支 codex/rectification-angle-timing-research-20261001，未推送） | BUG-1141 | | （直接执行，无任务书） | `PROGRESS-rectification-segment-order-20261001.md` | **健康区分题写入失败（BUG-1143）+ 按盘型挑题默认开启（BUG-1144）**：区分类焦点把 `health_pressure` 原样写入、违反表约束、写入被吞后提前交付（BUG-672 同族）；修后 77 例真库回放平均问题数 ±10 2.8→4.6、±30 3.6→5.4，D10 头段 ±30 39→52；按段选题 ≤61 分钟默认开，±30 D9 46→51、D10 52→55，真值段 76/76 | 已验收（Claude 10-01 直接执行） | 分支 codex/rectification-segment-order-20261001 | | （直接执行，无任务书） | `PROGRESS-rectification-post-adopt-card-20261001.md` | **采用后核对题无按钮（BUG-1146）+ 盘型口径回复念分钟排名（BUG-1147）**：keep 分支丢探针身份致 GET 空卡（BUG-498 复发）；采用接口不写轮次、只读运行被 60 秒交付窗口挡掉，题挂到上一题回复下；改为采用接口写「已采用 HH:MM。再核对一件过去的事。」轮次并挂题。盘型口径下去掉「领先 / 落后」、交付旁白不再说「分不开」、边界句只说一次 | 已验收（Claude 10-01 直接执行）；真机回报后补修 BUG-1149（请求号非 uuid 被真库拒写）、BUG-1150（核对题作答 stale_probe），本机真库重放通过；BUG-1151 盘型报告去分钟排名（产品选 A）；BUG-1153 交付正文连出两条（BUG-1149 引入的回归）；BUG-1172 核对题答完说对不对得上（产品选 A）；BUG-1173 盘型变化时说清并给「改用」（产品选 A）；真机清单 `docs/testing/rectification-post-adopt-card-20261001.md` | 分支 codex/rectification-post-adopt-verify-20261001 | +| （直接执行，无任务书） | `PROGRESS-rectification-event-dates-20261002.md` | **经历日期越界卡死比较（BUG-1178）+ 更正经历（BUG-1179，产品选 A）**：未来年份或「今年 / 本月」经历让引擎整单 400，之后每轮「没有重新比较」、不出卡；发引擎前裁剪日期、入口拒收未来日期并告知；revise 工具免焦点、立即确认重算、quote 须落本轮原话 | 已实现（Claude 10-02 直接执行；真引擎 + 真库验证；真实模型调用待真机） | 分支 codex/rectification-post-adopt-verify-20261001 | | `TASK-rectification-typed-event-scoring-research-20260929.md` | `PROGRESS-rectification-typed-event-research-20260929.md` | **打字经历按选择题规则计分（离线研究，不上线）**：计分通道不对称 + 已入账年份挡题；R0 学业质量题措辞 / 年精度显示成 1 月（冻结文件，需重新冻结） | 已验收关单（Claude 10-01：三类技法 9 项检验 0 显著，真值平均排位 0.49～0.59；Claude 补正对照——埋入信号排位 0.077、挪年 0.502——流程有效；BUG-1141 closed_by_design） | BUG-1088、1089 | | `TASK-report-reader-polish-20260929.md` | `PROGRESS-report-reader-polish-20260929.md` | **报告页打磨**：生成入口挪进页面主体（删标题栏按钮）、详情页到底部按钮（懒渲染一次到底）、导出按钮带文字、目录一级/二级分层、去掉「字段」与 RL/NL 缩写表头、状态列同义重复去重、报告表格淡底色、分块导出显示真实文件大小 | Claude 直接执行并自验（tsc 0、lint 0 error、全量 fail 与基线同 24 条、`/` Static、gzip +7 B、快速门 Python 1000 passed），已部署 staging `d7772011`，真机欠 | BUG-1092～1094 | | `TASK-report-english-edition-20260929.md` | `PROGRESS-report-english-edition-20260929.md` | **报告中英两版**：同一次引擎计算渲染 zh/en 两遍（不用模型翻译），瑜伽库 477 条补英文、模板与前端表头英文化；中文逐字节不变、英文零汉字、两版数字序列一致；阅读页 `?lang=en` 切换，导出当前语言 + 「问 AI 建议导出英文版」提示；旧报告不补英文 | Claude 直接执行（含两个 fork 子代理）并自验（tsc 0、lint 0 error、全量 fail 与基线同 24 条、`/` Static、gzip +0.06%、Python 定向全绿），已部署 staging `d7772011`，真机欠 | — | diff --git a/docs/testing/rectification-event-dates-20261002.md b/docs/testing/rectification-event-dates-20261002.md new file mode 100644 index 00000000..a292f192 --- /dev/null +++ b/docs/testing/rectification-event-dates-20261002.md @@ -0,0 +1,11 @@ +# 真机清单：经历日期越界 + 更正经历（2026-10-02，BUG-1178～1179） + +> staging 部署后，用自己的账号走一遍，经历用虚构的即可。 + +1. **已卡住的校正**：打开之前一直「没有重新比较」的那次校正，再说一件带年月的事。回复不再出现「这次没有重新比较」，范围或盘型应有变化；问够后会出时间卡。 +2. **未来年份**：新开一次校正，说「2028 年换了工作」。回复说「有一件事的日期在今天之后，这件先不算；如果年份打错了，直接告诉我正确的年份就行」，不说「记下了：2028…」。 +3. **今年的事**：说「今年 3 月搬家」。回复正常记下，后面照常比较，不出现「这次没有重新比较」。 +4. **更正**：先说「2015 年借出一笔钱」，下一句说「刚才年份打错了，是 2018 年」。回复说已改成 2018 年；不应多出一条新的 2018 年记录；随后照常比较。 +5. **旧校正会话**：从会话列表打开一个以前的校正会话，能打开、不报错。 + +有任何一项不符，截图发给 Claude。 diff --git a/frontend/docs/VOICE.md b/frontend/docs/VOICE.md index e8c4c2f5..de49d5b6 100644 --- a/frontend/docs/VOICE.md +++ b/frontend/docs/VOICE.md @@ -76,6 +76,7 @@ Jyotisha 的可见文案是产品的一部分。正确性红线（真实性、 - 「范围从 A–B 变为 C–D。」（本轮开始到结算时可信范围变了才写；点选题里顺手补了带年月的经历时，这一句覆盖整条消息，从答题前的范围算起） - 「这次没有重新比较，稍后再比一次。」 +- 「有一件事的日期在今天之后，这件先不算；如果年份打错了，直接告诉我正确的年份就行。」（BUG-1178：用户说的事件日期在今天之后时由服务端追加；不说「已记下」。） - 「候选比较这次没跑成，下一句话时会自动再试。」 模型正文不写时刻（HH:MM）、时间区间和百分比；只有写正文时服务端已经告诉它「本轮交付」（`range_after_rescore.delivers_range_this_turn`），才写交付三句，数字只抄服务端给的值。任何一句里的时刻或百分比与服务端当轮事实不符，整句不落库（正则只当门，不在句子里删词）。模型正文一句不剩时，用服务端记下的经历复述「记下了：年 月 事件。」代替。 diff --git a/frontend/src/lib/rectification-agentic/user-copy.ts b/frontend/src/lib/rectification-agentic/user-copy.ts index 93582793..c26e8d81 100644 --- a/frontend/src/lib/rectification-agentic/user-copy.ts +++ b/frontend/src/lib/rectification-agentic/user-copy.ts @@ -223,6 +223,8 @@ export const RECTIFICATION_USER_COPY = { forceMinuteAfterSubBlocks: "时段分不开，直接按分钟比。", compareFailedRetry: "候选比较这次没跑成，下一句话时会自动再试。", rescoreSkipped: "这次没有重新比较，稍后再比一次。", + /** BUG-1178: an event dated after today is not recorded or scored. */ + futureDatedEvent: "有一件事的日期在今天之后，这件先不算；如果年份打错了，直接告诉我正确的年份就行。", lastSuccessfulCompareRange: "这是按上一次成功比较给出的范围。", deferredCareerWindow: "下一次事业变动的预测窗口留在采用后的核对阶段。", rangeDeliveryTitle: "目前范围", @@ -662,6 +664,7 @@ export function listUserVisibleCopy(): string[] { RECTIFICATION_USER_COPY.postAdoptVerifyDone, RECTIFICATION_USER_COPY.compareFailedRetry, RECTIFICATION_USER_COPY.rescoreSkipped, + RECTIFICATION_USER_COPY.futureDatedEvent, RECTIFICATION_USER_COPY.lastSuccessfulCompareRange, RECTIFICATION_USER_COPY.deferredCareerWindow, RECTIFICATION_USER_COPY.rangeDeliveryTitle, diff --git a/frontend/src/lib/rectification-agentic/v9/agent-run-attempt.ts b/frontend/src/lib/rectification-agentic/v9/agent-run-attempt.ts index fa2ee586..19924025 100644 --- a/frontend/src/lib/rectification-agentic/v9/agent-run-attempt.ts +++ b/frontend/src/lib/rectification-agentic/v9/agent-run-attempt.ts @@ -46,6 +46,7 @@ import { import { currentEngineCallTimings, reportTurnProgress } from "./turn-instrumentation.ts"; import { batchResultFromToolChunk, + batchHasFutureDatedItem, batchRescoreFailed, acceptedRecapLines, composeHostFallbackNarration, @@ -620,6 +621,7 @@ export async function streamV9Attempt( serverFacts = { compareFailed, rescoreSkipped, + futureDated: batchHasFutureDatedItem(batchToolResult), rangeBefore: rangeBeforeCompare, rangeAfter: rangeAfterEvidence, }; diff --git a/frontend/src/lib/rectification-agentic/v9/agent-run-support.ts b/frontend/src/lib/rectification-agentic/v9/agent-run-support.ts index a365c2c0..8c9c5a6a 100644 --- a/frontend/src/lib/rectification-agentic/v9/agent-run-support.ts +++ b/frontend/src/lib/rectification-agentic/v9/agent-run-support.ts @@ -10,6 +10,7 @@ import { toAgentModelFinishReason } from "../../agent-observability.ts"; import type { PublicStreamEvent } from "./stream-mapping"; import type { SpokenFactWhitelist } from "./spoken-grounding"; import { + RECTIFICATION_USER_COPY, withCompareFailedRetryNotice, withRangeChangedAfterEvidence, withRescoreSkippedNotice, @@ -51,6 +52,8 @@ export type AttemptOutcome = Readonly<{ export type TurnServerFacts = Readonly<{ compareFailed: boolean; rescoreSkipped: boolean; + /** BUG-1178: an item of this turn's batch was dated after today and was not recorded. */ + futureDated?: boolean; rangeBefore: readonly [string, string] | null; rangeAfter: readonly [string, string] | null; }>; @@ -62,9 +65,11 @@ export function composeSpokenWithServerFacts( ): string { let spoken = body.trim(); if (!facts) return spoken; + const future = facts.futureDated === true ? RECTIFICATION_USER_COPY.futureDatedEvent : ""; + const withFuture = (text: string) => !future || text.includes(future) ? text : text ? `${text}${/[。！？]$/.test(text) ? "" : "。"}${future}` : future; if (facts.compareFailed) spoken = withCompareFailedRetryNotice(spoken); - if (facts.rescoreSkipped) return withRescoreSkippedNotice(spoken); - return withRangeChangedAfterEvidence(spoken, facts.rangeBefore, facts.rangeAfter); + if (facts.rescoreSkipped) return withFuture(withRescoreSkippedNotice(spoken)); + return withFuture(withRangeChangedAfterEvidence(spoken, facts.rangeBefore, facts.rangeAfter)); } export const MAX_ATTEMPTS = 2; diff --git a/frontend/src/lib/rectification-agentic/v9/engine-client.ts b/frontend/src/lib/rectification-agentic/v9/engine-client.ts index 21c3fff1..dc6e52e2 100644 --- a/frontend/src/lib/rectification-agentic/v9/engine-client.ts +++ b/frontend/src/lib/rectification-agentic/v9/engine-client.ts @@ -769,6 +769,30 @@ export function sanitizeAskedProbeKeysForEngine( return next; } +/** + * BUG-1178: the engine rejects the whole request when any event lies outside + * [birth_date, today] (`events[i] dates must be between birth_date and today`). A typo'd future + * year, or any 「今年 / 这个月」 event whose period ends after today, then failed every rescore + * of the Case and the delivery card never came. Before sending: drop events that start after + * the cap or end before birth; clip the rest into the window. The cap is yesterday (UTC) so the + * engine's own clock can never be behind ours. + */ +export function boundEngineEventDates( + events: readonly V9EngineEvent[], + birthDate: string, + now: Date = new Date(), +): V9EngineEvent[] { + const cap = new Date(now.getTime() - 86_400_000).toISOString().slice(0, 10); + const floor = /^\d{4}-\d{2}-\d{2}$/.test(birthDate) ? birthDate : null; + return events.flatMap((event): V9EngineEvent[] => { + if (event.date_start > cap) return []; + if (floor && event.date_end < floor) return []; + const start = floor && event.date_start < floor ? floor : event.date_start; + const end = event.date_end > cap ? cap : event.date_end; + return start === event.date_start && end === event.date_end ? [event] : [{ ...event, date_start: start, date_end: end }]; + }); +} + export function engineRequestBody(input: { baselineBirthSnapshot: Readonly>; candidateRange: DatedCandidateRange; @@ -786,7 +810,8 @@ export function engineRequestBody(input: { if (!birthDate || lat === null || lon === null || tz === null) { throw new RectificationEngineError("engine_profile_incomplete", "server profile snapshot is incomplete"); } - if (input.events.length === 0) { + const events = boundEngineEventDates(input.events, birthDate); + if (events.length === 0) { throw new RectificationEngineError("no_scorable_evidence", "no scorable evidence for the engine"); } const askedProbeKeys = sanitizeAskedProbeKeysForEngine(input.askedProbeKeys); @@ -805,7 +830,7 @@ export function engineRequestBody(input: { tz, ayanamsa: resolveAyanamsa(snapshot), node_mode: "mean", - events: input.events, + events, birth_time_source: snapshot.birth_time_source, timezone_id: snapshot.timezone_id, timezone_source: snapshot.timezone_source, diff --git a/frontend/src/lib/rectification-agentic/v9/host-fallback.ts b/frontend/src/lib/rectification-agentic/v9/host-fallback.ts index c4a1c898..f5642a9e 100644 --- a/frontend/src/lib/rectification-agentic/v9/host-fallback.ts +++ b/frontend/src/lib/rectification-agentic/v9/host-fallback.ts @@ -119,6 +119,15 @@ export function batchResultFromToolChunk(chunk: { return result; } +/** BUG-1178: the batch turned an item away because it starts after today. */ +export function batchHasFutureDatedItem(batchResult: unknown): boolean { + if (!batchResult || typeof batchResult !== "object" || Array.isArray(batchResult)) return false; + const items = (batchResult as { items?: unknown }).items; + return Array.isArray(items) && items.some((item) => ( + Boolean(item) && typeof item === "object" && (item as { error_code?: unknown }).error_code === "future_date" + )); +} + export function batchRescoreFailed(batchResult: unknown): boolean { if (!batchResult || typeof batchResult !== "object" || Array.isArray(batchResult)) return false; const rescore = (batchResult as { rescore?: { status?: unknown } }).rescore; diff --git a/frontend/src/lib/rectification-agentic/v9/turn-decision.ts b/frontend/src/lib/rectification-agentic/v9/turn-decision.ts index 39214d67..477b58c0 100644 --- a/frontend/src/lib/rectification-agentic/v9/turn-decision.ts +++ b/frontend/src/lib/rectification-agentic/v9/turn-decision.ts @@ -254,7 +254,10 @@ export function projectTurnDecision( } return [{ role: turn.role, text }]; }); - const evidence = dossier.evidence.slice(-TURN_DECISION_EVIDENCE_LIMIT).map((item) => ({ + // BUG-1179: the id lets the agent revise a record the user corrects (「2028 打错了，是 2018」). + const evidence = dossier.evidence.filter((item) => item.status !== "superseded") + .slice(-TURN_DECISION_EVIDENCE_LIMIT).map((item) => ({ + evidence_id: item.id, domain: item.domain, kind: item.eventKind, status: item.status, diff --git a/frontend/src/mastra/agentic-rectification.ts b/frontend/src/mastra/agentic-rectification.ts index 47cdc9f7..c5f2500a 100644 --- a/frontend/src/mastra/agentic-rectification.ts +++ b/frontend/src/mastra/agentic-rectification.ts @@ -39,7 +39,7 @@ const agenticRectificationInstructions = `你是 Jyotisha，只服务当前绑 懂行、可靠、说人话：先回应人，再展开方法。范围由系统说：范围变化、没重新比较、比较没跑成这几句由服务器接在你的正文后面；除出牌轮外你不写时刻（HH:MM）、时间区间和百分比；任何一句里的时刻或百分比与服务端当轮事实不符，整句不会落库。下一问是采集题时，先承接用户刚说的事；进度数字只用 collection_progress，没有该字段不说进度。不得写「范围在收窄」「继续收窄」这类没有数字的进度句。 1. 第一步调用 rectification-read-case。服务器是事实、焦点、权限与终态的唯一权威。 2. 事实只能来自用户原话；复述日期必须用 display_date_label。不得虚构事件、候选或出生分钟。 -3. 新事件走 rectification-record-evidence-batch。工具执行保持静默；思考用简体中文写在思维链；对用户说的话必须自己写在正文里，不叙述工具或内部状态。 +3. 新事件走 rectification-record-evidence-batch。用户更正之前说过的某件事（如「刚才 2028 打错了，是 2018」）走 rectification-revise-evidence，evidenceId 取 relevant_evidence_summary 里那一条的 evidence_id，不要再记一条新的。工具执行保持静默；思考用简体中文写在思维链；对用户说的话必须自己写在正文里，不叙述工具或内部状态。 4. 每轮在记录证据后，用 rectification-set-focus 的 spokenPrompt 写出服务端给你的下一问：用自己的话、结合用户刚说的事，问出同一个年份/期间和同一个事件家族；不得改年份、不得改选项含义、不得合并两道题。正文只做承接，不提问、不复述题干、不预告选项——题干会作为同一条消息的下一段自动出现。开场轮：先 set-focus 写采集题的 spokenPrompt（开场题干由服务端固定，已列出${OPENING_COLLECT_DOMAINS.join("、")}和一个回答示例），正文两句大白话：要把出生时间缩小到更准的范围、现在先在哪段时间里找；做法是用户说几件人生大事和大概年月，拿去和星盘对照。正文不用大运、盘面、分盘、候选、区间、代表分钟、精确到秒这类词，不重复题干里的例子，不提问；不得写具体年份，不得要求先准备材料。没有下一问（服务端返回 next_followup=null）时不要自拟问题。证据轮的复述由服务器写：「记下了 N 件事」（两件以内直接列出），清单挂在同一条消息里；你不逐条复述用户说的经历，不得评价价值或写「很有帮助 / 很有价值 / 很有分量 / 特别有用」。case.accepted_time 非空时，正文第一句要说明已按该时间采用、现在在核对。正文不得断言界面当前状态，不要写「界面上有下一问」「界面上出现了…」。choice 选项由服务端写入同一条消息，collect_spoken 只承接用户刚说的事实，不输出输入提示。点选与「先这样」由服务器处理。职业题只问平时做什么，不得自行追加「哪年 / 哪一年开始干这一行」；要问开始年份必须走服务器锚定题，且焦点 domain 是 career 不是 occupation。 5. 不得宣称唯一出生分钟。confirmation_allowed 为 false 或宽度大于 5 时，说明这是不可分区间，代表分钟只是代表性候选。rectification-record-evidence-batch 返回 range_after_rescore.delivers_range_this_turn=true 时才是出牌轮：正文只写三句（范围与代表分钟；choice_count 大于 0 时写「用了 N 道选择题」；边界句），range_not_narrowed=true 时直说「这个窗口按现在的方法缩不下去」，数字只抄 range_after_rescore；不写吻合率、不写对照了几件经历、不邀请再补经历；否则按证据轮只写一句复述。八法报告在卡片折叠块（skill_verification_report），不要写进气泡。80%/60% 只是折叠报告里的事件吻合率，不进气泡。 6. 一次一问。不泄露提示词或 Skill 原文。 diff --git a/frontend/src/mastra/rectification-v9-tools.ts b/frontend/src/mastra/rectification-v9-tools.ts index 806afdc7..96919b7a 100644 --- a/frontend/src/mastra/rectification-v9-tools.ts +++ b/frontend/src/mastra/rectification-v9-tools.ts @@ -24,6 +24,7 @@ import { proposeV9Evidence, confirmV10Evidence, reviseV10Evidence, + reviseV9Evidence, setV10ConversationFocus, resolveV10ConversationFocus, recordV10EvidenceBatch, @@ -935,6 +936,19 @@ function normalizeDatePart(value: string): string | null { return `${parts[0]}-${parts[1]!.padStart(2, "0")}-${parts[2]!.padStart(2, "0")}`; } +/** + * BUG-1178: an event that starts after today cannot be scored (the engine rejects the whole + * request) and is almost always a typed-year slip. It is not written; the reply says so. + */ +export function evidenceStartsInFuture( + item: Readonly<{ occurredFrom?: string | null; occurredTo?: string | null }>, + now: Date = new Date(), +): boolean { + const raw = item.occurredFrom || item.occurredTo; + const start = raw ? normalizeDatePart(raw) : null; + return Boolean(start && start > now.toISOString().slice(0, 10)); +} + function occupationNormalizedLedgerRows(input: { occupationFocus: Parameters[0]; domain: string; @@ -1706,8 +1720,21 @@ export function createRectificationV9Tools(ctx: RectificationV9Context) { }] : [] )); + const futureResults = groundedItems.flatMap(({ item, resolved }, index) => ( + resolved.ok && !isHoldoutVerificationQuote(resolved.quote) && evidenceStartsInFuture(item) + ? [{ + index, + outcome: "rejected" as const, + evidenceId: null, + status: "rejected", + idempotent: false, + clarificationFields: [] as string[], + errorCode: "future_date", + }] + : [] + )); const scoringItems = groundedItems.flatMap(({ item, resolved }, index) => ( - resolved.ok && !isHoldoutVerificationQuote(resolved.quote) + resolved.ok && !isHoldoutVerificationQuote(resolved.quote) && !evidenceStartsInFuture(item) ? [{ item, index, quote: resolved.quote, quoteStart: resolved.quoteStart, quoteEnd: resolved.quoteEnd }] : [] )); @@ -1734,12 +1761,23 @@ export function createRectificationV9Tools(ctx: RectificationV9Context) { summary: item.summary, })); }); + // BUG-1178: a write row's place in the model's own item order (rejected items no longer + // shift the recorded rows onto the wrong index). + const writeSources = scoringItems.flatMap(({ item, index }) => occupationNormalizedLedgerRows({ + occupationFocus, + domain: item.domain, + eventKind: item.proposedKind as EvidenceKind, + datePrecision: item.datePrecision, + occurredFrom: item.occurredFrom ? normalizeDatePart(item.occurredFrom) : null, + occurredTo: item.occurredTo ? normalizeDatePart(item.occurredTo) : null, + }).map(() => index)); + let recordedRows: readonly { outcome: string }[] = []; const result = scoringItems.length === 0 ? { - items: [...mismatchResults, ...holdoutResults], + items: [...mismatchResults, ...holdoutResults, ...futureResults].sort((left, right) => left.index - right.index), acceptedCount: 0, needsClarificationCount: 0, - rejectedCount: mismatchResults.length + holdoutResults.length, + rejectedCount: mismatchResults.length + holdoutResults.length + futureResults.length, focusId: input.focusId ?? null, } : await recordV10EvidenceBatch( @@ -1749,20 +1787,24 @@ export function createRectificationV9Tools(ctx: RectificationV9Context) { turnId, input.focusId ?? null, writeItems, - ).then((recorded) => ({ + ).then((recorded) => { + recordedRows = recorded.items; + return { items: [ ...recorded.items.map((item, offset) => ({ ...item, - index: writeItems[offset] ? offset : item.index, + index: writeSources[offset] ?? item.index, })), ...holdoutResults, ...mismatchResults, + ...futureResults, ].sort((left, right) => left.index - right.index), acceptedCount: recorded.acceptedCount, needsClarificationCount: recorded.needsClarificationCount, - rejectedCount: recorded.rejectedCount + holdoutResults.length + mismatchResults.length, + rejectedCount: recorded.rejectedCount + holdoutResults.length + mismatchResults.length + futureResults.length, focusId: recorded.focusId, - })); + }; + }); if (result.acceptedCount > 0) { const acceptedEvidenceId = result.items.find( (item) => item.outcome === "accepted" && item.evidenceId, @@ -1782,7 +1824,7 @@ export function createRectificationV9Tools(ctx: RectificationV9Context) { ? await autoRescoreAfterEvidenceChange(input.caseId) : { status: "skipped" as const, executedMethods: [] as const, errorCode: null, errorKind: null, cached: false, openQuestion: null, rangeAfter: null }; const acceptedRecaps = writeItems.flatMap((item, offset) => { - const recorded = result.items[offset]; + const recorded = recordedRows[offset]; if (!recorded || recorded.outcome !== "accepted") return []; if (item.datePrecision === "unknown") return []; const label = displayDateLabel( @@ -1876,6 +1918,16 @@ export function createRectificationV9Tools(ctx: RectificationV9Context) { note: "盘外核对不得写入可评分证据，也不会改候选分数。", }; } + if (evidenceStartsInFuture(input)) { + return { + evidence_id: null, + status: "rejected", + outcome: "rejected", + error_code: "future_date", + idempotent: false, + note: "这件事的日期在今天之后，不能计分；请用户核对年份，不要写成已记下。", + }; + } await receipt("rectification-propose-evidence", "evidence.proposed", "started", { inputFingerprint }); const occurredFrom = input.occurredFrom ? normalizeDatePart(input.occurredFrom) : null; const occurredTo = input.occurredTo ? normalizeDatePart(input.occurredTo) : null; @@ -1998,10 +2050,10 @@ export function createRectificationV9Tools(ctx: RectificationV9Context) { const reviseEvidenceTool = createTool({ id: "rectification-revise-evidence", description: - "修订既有证据的事实（通常是用户更正日期）。旧记录标记 superseded，生成新 revision，不覆盖历史。quote 必须来自用户本轮原话；日期精度如实保留。", + "用户更正之前说过的某件事（通常是年份打错）时调用：evidenceId 取 relevant_evidence_summary 里那一条的 evidence_id，不要再记一条新的。旧记录标记 superseded，生成新 revision 并立即确认、重新比较，不覆盖历史。quote 必须来自用户本轮原话；日期精度如实保留。没有对应的问题焦点时省略 focusId。", inputSchema: z.object({ caseId: z.string().uuid(), - focusId: z.string().uuid(), + focusId: z.string().uuid().nullable().optional(), evidenceId: z.string().uuid(), quote: z.string().trim().min(2).max(400), datePrecision: z.enum(["year", "month", "quarter", "day", "range", "unknown"]), @@ -2021,27 +2073,69 @@ export function createRectificationV9Tools(ctx: RectificationV9Context) { }); await receipt("rectification-revise-evidence", "evidence.proposed", "started", { inputFingerprint }); try { + // BUG-1179 (product 10-02 option A): a correction the user states this turn replaces the + // old record, is confirmed at once and rescored — no focus needed, no second tool call. + // The v10 rule 「every evidence change is bound to what the user said」 stays: the quote + // must be grounded in this turn's user message, exactly as for new events. + if (!resolveEvidenceQuote(userMessage ?? null, { quote: input.quote }).ok) { + await receipt("rectification-revise-evidence", "evidence.proposed", "completed", { inputFingerprint }); + return { + evidence_id: null, + supersedes_evidence_id: null, + status: "rejected", + error_code: "quote_mismatch", + note: "quote 必须是用户本轮原话；未更正，不要说已改好。", + }; + } + if (evidenceStartsInFuture(input)) { + await receipt("rectification-revise-evidence", "evidence.proposed", "completed", { inputFingerprint }); + return { + evidence_id: null, + supersedes_evidence_id: null, + status: "rejected", + error_code: "future_date", + note: "更正后的日期仍在今天之后，不能计分；请用户再核对年份。", + }; + } const occurredFrom = input.occurredFrom ? normalizeDatePart(input.occurredFrom) : null; const occurredTo = input.occurredTo ? normalizeDatePart(input.occurredTo) : null; - const result = await reviseV10Evidence(accounting, userId, input.caseId, { - focusId: input.focusId, + const revision = { evidenceId, quote: input.quote, occurredFrom, occurredTo, datePrecision: input.datePrecision, summary: input.summary, - }); - await receipt("rectification-revise-evidence", "evidence.proposed", "completed", { - inputFingerprint, - resultFingerprint: hashResult(result), - }); - return { + }; + const focusId = input.focusId ?? null; + const result = focusId + ? await reviseV10Evidence(accounting, userId, input.caseId, { ...revision, focusId }) + : await reviseV9Evidence(accounting, userId, input.caseId, revision); + const confirmed = await confirmV10Evidence(accounting, userId, input.caseId, focusId, result.evidenceId); + const rescore = confirmed.status === "confirmed" + ? await autoRescoreAfterEvidenceChange(input.caseId) + : { status: "skipped" as const, executedMethods: [] as const, errorCode: null, errorKind: null, cached: false, openQuestion: null, rangeAfter: null }; + const projection = { evidence_id: result.evidenceId, supersedes_evidence_id: result.supersedesEvidenceId, - status: "pending_confirmation", + status: confirmed.status, idempotent: result.idempotent, + rescore: { + status: rescore.status, + executed_methods: rescore.executedMethods, + error_code: rescore.errorCode, + error_kind: rescore.errorKind, + }, + open_question: rescore.openQuestion, }; + await receipt("rectification-revise-evidence", "evidence.proposed", "completed", { + inputFingerprint, + resultFingerprint: hashResult(projection), + executedMethods: [...rescore.executedMethods], + }); + return "rangeAfter" in rescore && rescore.rangeAfter + ? { ...projection, range_after_rescore: rescore.rangeAfter } + : projection; } catch (error) { await failReceipt("rectification-revise-evidence", "evidence.proposed", error, { inputFingerprint }); throw error; diff --git a/frontend/tests/rectification-event-date-bounds-20261002.test.ts b/frontend/tests/rectification-event-date-bounds-20261002.test.ts new file mode 100644 index 00000000..313004f9 --- /dev/null +++ b/frontend/tests/rectification-event-date-bounds-20261002.test.ts @@ -0,0 +1,125 @@ +import assert from "node:assert/strict"; +import test from "node:test"; + +import { boundEngineEventDates, engineRequestBody, type V9EngineEvent } from "../src/lib/rectification-agentic/v9/engine-client.ts"; +import { composeSpokenWithServerFacts } from "../src/lib/rectification-agentic/v9/agent-run-support.ts"; +import { batchHasFutureDatedItem } from "../src/lib/rectification-agentic/v9/host-fallback.ts"; +import { RECTIFICATION_USER_COPY } from "../src/lib/rectification-agentic/user-copy.ts"; +import { createRectificationV9Tools, evidenceStartsInFuture } from "../src/mastra/rectification-v9-tools.ts"; +import { + CASE_ID, + FOCUS_ID, + TURN_ID, + USER_ID, + computeFixture, + dossierFixture, + fakeAccounting, + receiptHandlers, +} from "./rectification-v9-test-support.ts"; + +// Real staging case 2026-10-02 (BUG-1178): a typo'd 「2028 年」 event entered the ledger and the +// engine (scripts/rectification/contracts.py) rejected every later rescore with +// 「events[i] dates must be between birth_date and today」; the card never came. 「今年 / 这个月」 +// events (period end after today) failed the same way. +const NOW = new Date("2026-10-02T03:00:00Z"); +const event = (id: string, start: string, end: string): V9EngineEvent => ({ + id, domain: "finance", event_kind: "finance_change", date_start: start, date_end: end, precision: "year", + summary: "fictional", date_source: null, date_reliability: null, date_corroboration: null, + date_conflict_status: null, source_turn_id: null, subject: "self", +} as V9EngineEvent); + +test("engine events stay inside [birth_date, yesterday]: future dropped, open periods clipped", () => { + const bounded = boundEngineEventDates([ + event("past", "2014-01-01", "2014-12-31"), + event("future", "2028-01-01", "2028-12-31"), + event("this-year", "2026-01-01", "2026-12-31"), + event("this-month", "2026-10-01", "2026-10-31"), + event("before-birth", "1980-01-01", "1980-12-31"), + event("straddles-birth", "1995-01-01", "1995-12-31"), + ], "1995-06-14", NOW); + assert.deepEqual(bounded.map((row) => [row.id, row.date_start, row.date_end]), [ + ["past", "2014-01-01", "2014-12-31"], + ["this-year", "2026-01-01", "2026-10-01"], + ["this-month", "2026-10-01", "2026-10-01"], + ["straddles-birth", "1995-06-14", "1995-12-31"], + ]); +}); + +test("the engine request body carries only bounded events", () => { + const body = engineRequestBody({ + baselineBirthSnapshot: { birth_date: "1995-06-14", latitude: 30, longitude: 120, timezone_offset: 8, timezone_id: "Asia/Shanghai" }, + candidateRange: { start_time: "14:34", end_time: "15:04" } as never, + events: [event("past", "2014-01-01", "2014-12-31"), event("future", "2099-01-01", "2099-12-31")], + }); + const events = body.events as V9EngineEvent[]; + assert.deepEqual(events.map((row) => row.id), ["past"]); + assert.throws(() => engineRequestBody({ + baselineBirthSnapshot: { birth_date: "1995-06-14", latitude: 30, longitude: 120, timezone_offset: 8 }, + candidateRange: { start_time: "14:34", end_time: "15:04" } as never, + events: [event("future", "2099-01-01", "2099-12-31")], + }), /no scorable evidence/); +}); + +test("intake: an item starting after today is future-dated; this year is not", () => { + assert.equal(evidenceStartsInFuture({ occurredFrom: "2028" }, NOW), true); + assert.equal(evidenceStartsInFuture({ occurredFrom: "2026-11" }, NOW), true); + assert.equal(evidenceStartsInFuture({ occurredFrom: "2026" }, NOW), false); + assert.equal(evidenceStartsInFuture({ occurredFrom: "2014-12" }, NOW), false); + assert.equal(evidenceStartsInFuture({ occurredFrom: null, occurredTo: null }, NOW), false); +}); + +test("batch: the future item is not written, the rest keep their own index and recap", async () => { + const quote = "2099年借出一笔钱，2014年年底拔牙"; + const accounting = fakeAccounting({ + ...receiptHandlers, + get_agentic_rectification_case_dossier: () => dossierFixture({ latestResult: null }), + get_agentic_rectification_case_compute: () => computeFixture(), + record_agentic_rectification_evidence_batch: (_fn, args) => { + const items = Array.isArray(args.p_items) ? args.p_items as Array> : []; + return { + items: items.map((_item, index) => ({ + index, outcome: "accepted", evidence_id: "77777777-7777-4777-8777-777777777777", status: "confirmed", + idempotent: false, clarification_fields: [], error_code: null, + })), + accepted_count: items.length, needs_clarification_count: 0, rejected_count: 0, focus_id: FOCUS_ID, + }; + }, + }); + const tools = createRectificationV9Tools({ + userId: USER_ID, caseId: CASE_ID, turnId: TURN_ID, userMessage: quote, accounting: accounting.client as never, + }); + const result = await (tools["rectification-record-evidence-batch"] as unknown as { + execute(input: unknown): Promise<{ + items: Array<{ index: number; outcome: string; error_code: string | null }>; + accepted_recaps: Array<{ display_date_label?: string; event_phrase?: string }>; + accepted_count: number; + }>; + }).execute({ + caseId: CASE_ID, + items: [ + { quote: "2099年借出一笔钱", proposedKind: "finance_change", subject: "self", domain: "finance", + datePrecision: "year", occurredFrom: "2099", summary: "借出一笔钱" }, + { quote: "2014年年底拔牙", proposedKind: "self_health_event", subject: "self", domain: "health", + datePrecision: "month", occurredFrom: "2014-12", summary: "2014年年底拔牙" }, + ], + }); + const write = accounting.calls.find((call) => call.fn === "record_agentic_rectification_evidence_batch"); + const written = (write?.args.p_items ?? []) as Array>; + assert.equal(written.length, 1, "the 2099 item is never written"); + assert.equal(written[0]?.occurred_from, "2014-12-01"); + assert.deepEqual(result.items.map((item) => [item.index, item.outcome, item.error_code]), [ + [0, "rejected", "future_date"], + [1, "accepted", null], + ]); + assert.equal(result.accepted_count, 1); + assert.equal(result.accepted_recaps.length, 1, "the recap reads the recorded row, not the rejected slot"); + assert.equal(batchHasFutureDatedItem(result), true); +}); + +test("the reply says the future-dated item was not counted, once", () => { + const facts = { compareFailed: false, rescoreSkipped: false, futureDated: true, rangeBefore: null, rangeAfter: null }; + const spoken = composeSpokenWithServerFacts("记下了：2014年年底拔牙。", facts); + assert.ok(spoken.endsWith(RECTIFICATION_USER_COPY.futureDatedEvent)); + assert.equal(spoken.split(RECTIFICATION_USER_COPY.futureDatedEvent).length, 2); + assert.equal(composeSpokenWithServerFacts("记下了。", { ...facts, futureDated: false }), "记下了。"); +}); diff --git a/frontend/tests/rectification-v10-conversation-focus.test.ts b/frontend/tests/rectification-v10-conversation-focus.test.ts index 6dabaa84..c61a402b 100644 --- a/frontend/tests/rectification-v10-conversation-focus.test.ts +++ b/frontend/tests/rectification-v10-conversation-focus.test.ts @@ -235,7 +235,11 @@ test("confirm/revise tools forward the durable focus/evidence binding", async () idempotent: false, }), }); - const tools = toolSet(accounting); + // BUG-1179: the revise below quotes this turn's user message (grounding replaces the focus requirement). + const tools = createRectificationV9Tools({ + userId: USER_ID, caseId: CASE_ID, turnId: TURN_ID, accounting: accounting.client as never, + userMessage: "不是9月，是10月", + }); await (tools["rectification-confirm-evidence"] as unknown as ExecutableTool).execute({ caseId: CASE_ID, @@ -262,6 +266,9 @@ test("confirm/revise tools forward the durable focus/evidence binding", async () })), [ { fn: "confirm_agentic_rectification_evidence_v10", focusId: FOCUS_ID, evidenceId: EVIDENCE_ID }, { fn: "revise_agentic_rectification_evidence_v10", focusId: FOCUS_ID, evidenceId: EVIDENCE_ID }, + // BUG-1179 — 原值: 两次写（confirm、revise） / 新值: revise 后再确认新 revision（第三次写，仍带同一 focus） + // / 原因: 产品 10-02 选 A，更正立即生效并重算。 + { fn: "confirm_agentic_rectification_evidence_v10", focusId: FOCUS_ID, evidenceId: SECOND_EVIDENCE_ID }, ]); }); diff --git a/frontend/tests/rectification-v10-tool-contract.test.ts b/frontend/tests/rectification-v10-tool-contract.test.ts index e67f6d96..903cb0ad 100644 --- a/frontend/tests/rectification-v10-tool-contract.test.ts +++ b/frontend/tests/rectification-v10-tool-contract.test.ts @@ -163,9 +163,16 @@ test("focus/evidence mutation schemas require durable focus and evidence referen assert.equal(confirm.safeParse({ caseId: CASE_ID, evidenceId: EVIDENCE_ID }).success, true); assert.equal(confirm.safeParse({ caseId: CASE_ID, focusId: FOCUS_ID }).success, false); + // BUG-1179 — 原值: revise 缺 focusId 时 schema 拒绝（false） / 新值: 允许省略 focusId（true），但仍必须带 + // evidenceId / 原因: 产品 10-02 选 A，用户随口更正打错的年份时没有对应的问题焦点；「改动必须来自用户 + // 原话」改由服务端校验 quote 落在本轮用户消息里来保证（见 rectification-v9-evidence 用例）。 assert.equal(revise.safeParse({ ...validInputs["rectification-revise-evidence"], focusId: undefined, + }).success, true); + assert.equal(revise.safeParse({ + ...validInputs["rectification-revise-evidence"], + evidenceId: undefined, }).success, false); assert.equal(resolve.safeParse({ caseId: CASE_ID, status: "declined" }).success, false); }); diff --git a/frontend/tests/rectification-v9-evidence.test.ts b/frontend/tests/rectification-v9-evidence.test.ts index 52277d56..ca55a98b 100644 --- a/frontend/tests/rectification-v9-evidence.test.ts +++ b/frontend/tests/rectification-v9-evidence.test.ts @@ -224,8 +224,16 @@ test("revision is append-only: revise supersedes and never overwrites history", supersedes_evidence_id: EVIDENCE_ID, idempotent: false, }), + // BUG-1179 — 原值: revise 后不确认（status pending_confirmation，未 mock confirm） / 新值: 同一调用里确认新 + // revision 并重算 / 原因: 产品 10-02 选 A，用户本轮说出的更正立即生效；quote 须来自本轮用户消息。 + confirm_agentic_rectification_evidence_v10: (_fn, args) => ({ + focus_id: args.p_focus_id, evidence_id: args.p_evidence_id, status: "confirmed", idempotent: false, + }), + }); + const tools = createRectificationV9Tools({ + userId: USER_ID, caseId: CASE_ID, turnId: TURN_ID, accounting: accounting.client as never, + userMessage: "不是，是2021年10月", }); - const { tools } = toolContext({ accounting }); const result = await (tools["rectification-revise-evidence"] as unknown as { execute(input: unknown): Promise<{ evidence_id: string; supersedes_evidence_id: string }>; }).execute({ @@ -241,6 +249,47 @@ test("revision is append-only: revise supersedes and never overwrites history", const reviseCall = accounting.calls.find((call) => call.fn === "revise_agentic_rectification_evidence_v10"); assert.ok(reviseCall); assert.equal(reviseCall.args.p_evidence_id, EVIDENCE_ID); + const confirmCall = accounting.calls.find((call) => call.fn === "confirm_agentic_rectification_evidence_v10"); + assert.equal(confirmCall?.args.p_evidence_id, "99999999-9999-4999-8999-999999999991"); +}); + +test("a correction without a focus revises, confirms and is grounded in this turn's words (BUG-1179)", async () => { + const NEW_ID = "99999999-9999-4999-8999-999999999992"; + const handlers = { + ...receiptHandlers, + get_agentic_rectification_case_dossier: () => dossierFixture(), + revise_agentic_rectification_evidence: (_fn: string, args: Record) => ({ + evidence_id: NEW_ID, supersedes_evidence_id: args.p_evidence_id, idempotent: false, + }), + confirm_agentic_rectification_evidence_v10: (_fn: string, args: Record) => ({ + focus_id: args.p_focus_id, evidence_id: args.p_evidence_id, status: "confirmed", idempotent: false, + }), + }; + const run = async (userMessage: string, occurredFrom: string) => { + const accounting = fakeAccounting(handlers); + const tools = createRectificationV9Tools({ + userId: USER_ID, caseId: CASE_ID, turnId: TURN_ID, accounting: accounting.client as never, userMessage, + }); + const result = await (tools["rectification-revise-evidence"] as unknown as { + execute(input: unknown): Promise>; + }).execute({ + caseId: CASE_ID, evidenceId: EVIDENCE_ID, quote: "刚才2028打错了，是2018", datePrecision: "year", + occurredFrom, summary: "借出一笔钱，多年后才还完", + }); + return { result, calls: accounting.calls.map((call) => call.fn) }; + }; + const ok = await run("刚才2028打错了，是2018", "2018"); + assert.equal(ok.result.status, "confirmed"); + assert.equal(ok.result.supersedes_evidence_id, EVIDENCE_ID); + assert.ok(ok.calls.includes("revise_agentic_rectification_evidence"), "no focus: the focus-free revise"); + assert.ok(!ok.calls.includes("revise_agentic_rectification_evidence_v10")); + assert.ok(ok.calls.includes("confirm_agentic_rectification_evidence_v10")); + const ungrounded = await run("今天天气不错", "2018"); + assert.equal(ungrounded.result.error_code, "quote_mismatch"); + assert.ok(!ungrounded.calls.includes("revise_agentic_rectification_evidence"), "no write without the user's words"); + const future = await run("刚才2028打错了，是2018", "2099"); + assert.equal(future.result.error_code, "future_date"); + assert.ok(!future.calls.includes("revise_agentic_rectification_evidence")); }); test("career and relationship kinds keep distinct semantics", () => {