Archived
审核通过与否读基座数;「重做」不再被指向 finish;审核卡摆成验收工单(W05 1.4.7)
用户两报:①「我上一步明确要求重做了,怎么会继续向下推进」②「验收工单弹出的是需要你定,明显不对」。 ①病根两处合谋:人两次答 custom:重做一次,模型把一模一样的问法又发了一遍被闸拦下;而拒回的话 不分答案一律说「照这个选择把结论落进台账那面旗,再 finish」—— 对通过是对的,对重做是把模型往错路上推, 它照办写了 review_approved_total = 1,而 S4 的哨兵读的正是这面模型自写的旗(被叫的那一方自己记账,A6 那一课)。 - 基座新读数 decision_last / decision_<值>_total(按 card-reply 数);S4 哨兵改读 decision_last = "approve", 人最后一次真点了通过才开门;静态审认这几个名字。 - 拒回 / 观察 / 事实行三处同一段 chosenGuidance 按答案分岔:custom: 不是通过 —— 按他写的改、换问法重发、 不许写通过旗、不许 finish;选项值指回说明书那一档。 - S4 剧本驳回分支明写不许写旗、ask 开头必须写明这一轮改了什么(字面一样的卡会被吞)。 ②平台的验收工单是终局,中间审核只能靠决策卡循环 —— 但不该穿「需要你定」的衣服:契约 decisionCard.variant:'review', PlanStep.cardVariant 扩到 'review',cards.decide 带 itemId,引擎按剧本决定;对话区摆成「工单 · 验收 / 审核」。 2106 测 / typecheck / lint / depcruise 绿;api-report 已刷;CONTRACT.md 同步。 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FJcX8A6TWp4GRs41TYryEh
This commit is contained in:
xcs
@@ -146,7 +146,8 @@ describe('说明书与剧本的骨架', () => {
|
||||
for (const item of ['S0d_p08_master', 'S0d_p10_document']) {
|
||||
expect(ASSESSMENT_STEPS.find((s) => s.item === item)?.cardVariant, item).toBe('data');
|
||||
}
|
||||
for (const step of ASSESSMENT_STEPS) if (!['S0d_p08_master', 'S0d_p10_document'].includes(step.item)) expect(step.cardVariant, step.item).toBeUndefined();
|
||||
expect(ASSESSMENT_STEPS.find((s) => s.item === 'S4_review')?.cardVariant).toBe('review');
|
||||
for (const step of ASSESSMENT_STEPS) if (!['S0d_p08_master', 'S0d_p10_document', 'S4_review'].includes(step.item)) expect(step.cardVariant, step.item).toBeUndefined();
|
||||
expect(NARRATIVE).toContain('"title":"数据工单 · 芯片主数据创建(P-08)');
|
||||
expect(NARRATIVE).toContain('"title":"数据工单 · 原厂技术资料创建(P-10)');
|
||||
expect(NARRATIVE).toContain('"title":"评估输入与资料补齐确认');
|
||||
@@ -330,6 +331,14 @@ describe('说明书与剧本的骨架', () => {
|
||||
}
|
||||
});
|
||||
|
||||
it('审核通过与否读基座数 decision_last,不由模型自写的旗说了算;驳回分支明写不许写旗、不许 finish(2026-08-27 实测:人答「重做」它照样推到放行)', () => {
|
||||
expect(CMMN_XML).toContain('decision_last = "approve"');
|
||||
expect(CMMN_XML).toContain('<caseFileItem id="decision_last"');
|
||||
const g = ASSESSMENT_STEPS.find((s) => s.item === 'S4_review')?.guide ?? '';
|
||||
expect(g).toMatch(/不许写 review_approved_total、不许 finish/);
|
||||
expect(g).toMatch(/ask 的开头必须写明这一轮改了什么/);
|
||||
});
|
||||
|
||||
it('五段进度,每段名 ≤10 字;除放行外每项恰属一段', () => {
|
||||
expect(ASSESSMENT_PHASES).toHaveLength(5);
|
||||
for (const p of ASSESSMENT_PHASES) expect(p.name.length, p.name).toBeLessThanOrEqual(10);
|
||||
|
||||
@@ -64,6 +64,7 @@ export const CMMN_XML = `<?xml version="1.0" encoding="UTF-8"?>
|
||||
<caseFileItem id="tool_part_substitution_assessment_compare_parameters_calls" name="比对工具成功调用次数(基座现数)"/>
|
||||
<caseFileItem id="tool_part_substitution_assessment_recommend_calls" name="定级工具成功调用次数(基座现数)"/>
|
||||
<caseFileItem id="review_approved_total" name="差异结论已审核通过(仅通过时置一)"/>
|
||||
<caseFileItem id="decision_last" name="最后一张决策卡人选的值(基座按留痕数)"/>
|
||||
<caseFileItem id="landing_prepared_total" name="拟落账已备好(S5 写完时置一)"/>
|
||||
</caseFileModel>
|
||||
<casePlanModel id="plan" name="差异评估" autoComplete="true">
|
||||
@@ -113,7 +114,12 @@ export const CMMN_XML = `<?xml version="1.0" encoding="UTF-8"?>
|
||||
<!-- ⑥ 审核已通过 —— 拟落账的入场前提 -->
|
||||
<sentry id="s_reviewed">
|
||||
<onPart id="op_r" caseFileItemRef="review_approved_total"><standardEvent>update</standardEvent></onPart>
|
||||
<ifPart id="if_r"><condition>review_approved_total > 0 and substitution_difference_item_missing = 0</condition></ifPart>
|
||||
<!--
|
||||
审核通过与否**不由模型自写的旗说了算**(2026-08-27 实测:人两次答「重做」,模型照拒回的指引写了旗,
|
||||
S4 被收掉一路推到放行)。\`decision_last\` 是基座按 card-reply 数出来的:人最后一次真点了 approve 才开门。
|
||||
旗留着记账,但它单独说了不算。
|
||||
-->
|
||||
<ifPart id="if_r"><condition>review_approved_total > 0 and decision_last = "approve" and substitution_difference_item_missing = 0</condition></ifPart>
|
||||
</sentry>
|
||||
<!-- ⑦ 拟落账已备好:每条关系都到 FAE 确认、P-01 回写五项齐 —— 放行门的入场前提 -->
|
||||
<sentry id="s_landing_ready">
|
||||
@@ -540,6 +546,8 @@ difference_status / compare_result / risk_level / impact_statement / unknown_dis
|
||||
},
|
||||
{
|
||||
item: 'S4_review',
|
||||
/* 中间审核摆成验收工单的样子(机制是可循环的决策卡);由剧本声明,模型抄不抄都一样 */
|
||||
cardVariant: 'review' as const,
|
||||
intent: '差异评估结论审核:决策卡循环,通过或驳回 + AI 按意见改台账重发',
|
||||
writes: ['substitution.relation', 'substitution.difference_item', 'validation.record', `${D}.verdict`],
|
||||
reads: ['substitution.relation', 'substitution.difference_item', 'validation.record', `${D}.verdict`],
|
||||
@@ -550,11 +558,13 @@ difference_status / compare_result / risk_level / impact_statement / unknown_dis
|
||||
轨迹里找这一项已答过、还没处理的 [card-reply]:
|
||||
· 没有 → 照说明书⑥发第一张决策卡(ask 里把候选数、比对条数、四态各几条、逐候选推荐等级说全),发完这一拍结束。
|
||||
· "approve" → 写 \`${D}.verdict.review_approved_total = 1\`,finish 这一项。别再发卡。
|
||||
· "reject" / "custom:…" → **驳回**。按意见改台账(同 PK 覆盖写,不开新行、不改行数):
|
||||
· "reject" / "custom:…" → **驳回**。**不许写 review_approved_total、不许 finish** —— 下一步的门读的是基座数
|
||||
\`decision_last\`(人最后一次真点了 approve 才开),写了旗也开不了门,只会把自己绕进死循环。按意见改台账(同 PK 覆盖写,不开新行、不改行数):
|
||||
· 意见点名某个参数 → 改那一行 \`substitution.difference_item\` 的 impact_statement / unknown_disposition / evidence_ref;要挂验证的补一行 \`validation.record\`;
|
||||
· 意见说推荐理由站不住 → 改 \`substitution.relation.recommendation_reason\`(等级不许改 —— 那是工具算的;要改先回 S3 补证据重比,本活动不回头,写清楚为什么维持);
|
||||
· 没写具体意见("reject")→ 自己挑最站不住的一处改实(证据最含糊的那条 evidence_ref、去向最空的那条 unknown_disposition),**至少改对一处**;
|
||||
然后**换一种问法**再发一张决策卡(把改了什么写进 ask),发完结束。
|
||||
然后**换一种问法**再发一张决策卡 —— **ask 的开头必须写明这一轮改了什么**(与上一张字面不同;
|
||||
字面一样的卡会被当成同一张吞掉,连吞两次就算被拒),发完结束。
|
||||
**基座对着字面比对意见与草稿**:观察里点名了哪一行,就改那一行,不挪到别的行头上。`,
|
||||
},
|
||||
{
|
||||
|
||||
@@ -89,7 +89,7 @@ export const manifest: ActivityManifest = {
|
||||
identity: {
|
||||
activityId: ACTIVITY_ID,
|
||||
name: 'ACT-WORK-05 · 差异评估',
|
||||
packageVersion: '1.4.6',
|
||||
packageVersion: '1.4.7',
|
||||
category: '需求分析与方案匹配',
|
||||
sdkVersion: '0.1.0-m5',
|
||||
checksum: 'dev-checksum-part-substitution-assessment',
|
||||
|
||||
+9
@@ -1825,6 +1825,9 @@ export declare const AI_TAG_LABEL: Readonly<Record<AiTag, string>>;
|
||||
/** 决策卡:分支条件选择表;含自定义收尾项 */
|
||||
export declare const decisionCardSchema: z.ZodObject<{
|
||||
k: z.ZodLiteral<"decision">;
|
||||
variant: z.ZodOptional<z.ZodEnum<{
|
||||
review: "review";
|
||||
}>>;
|
||||
ask: z.ZodString;
|
||||
basis: z.ZodArray<z.ZodObject<{
|
||||
k: z.ZodString;
|
||||
@@ -1937,6 +1940,9 @@ export declare const ratingCardSchema: z.ZodObject<{
|
||||
export type RatingCard = z.infer<typeof ratingCardSchema>;
|
||||
export declare const chatCardSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
||||
k: z.ZodLiteral<"decision">;
|
||||
variant: z.ZodOptional<z.ZodEnum<{
|
||||
review: "review";
|
||||
}>>;
|
||||
ask: z.ZodString;
|
||||
basis: z.ZodArray<z.ZodObject<{
|
||||
k: z.ZodString;
|
||||
@@ -2070,6 +2076,9 @@ export declare const chatMessageSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
|
||||
at: z.ZodISODateTime;
|
||||
}, z.core.$strip>, z.ZodObject<{
|
||||
k: z.ZodLiteral<"decision">;
|
||||
variant: z.ZodOptional<z.ZodEnum<{
|
||||
review: "review";
|
||||
}>>;
|
||||
ask: z.ZodString;
|
||||
basis: z.ZodArray<z.ZodObject<{
|
||||
k: z.ZodString;
|
||||
|
||||
@@ -204,6 +204,7 @@ export declare const planStepSchema: z.ZodObject<{
|
||||
doneWhen: z.ZodOptional<z.ZodString>;
|
||||
cardVariant: z.ZodOptional<z.ZodEnum<{
|
||||
data: "data";
|
||||
review: "review";
|
||||
}>>;
|
||||
}, z.core.$strip>;
|
||||
export type PlanStep = z.infer<typeof planStepSchema>;
|
||||
@@ -267,6 +268,7 @@ export declare const loopModelSchema: z.ZodObject<{
|
||||
doneWhen: z.ZodOptional<z.ZodString>;
|
||||
cardVariant: z.ZodOptional<z.ZodEnum<{
|
||||
data: "data";
|
||||
review: "review";
|
||||
}>>;
|
||||
}, z.core.$strip>>>;
|
||||
phases: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
||||
@@ -324,6 +326,7 @@ export declare const businessModelSchema: z.ZodObject<{
|
||||
doneWhen: z.ZodOptional<z.ZodString>;
|
||||
cardVariant: z.ZodOptional<z.ZodEnum<{
|
||||
data: "data";
|
||||
review: "review";
|
||||
}>>;
|
||||
}, z.core.$strip>>>;
|
||||
phases: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
||||
@@ -390,6 +393,7 @@ export declare const manifestAssetsSchema: z.ZodObject<{
|
||||
doneWhen: z.ZodOptional<z.ZodString>;
|
||||
cardVariant: z.ZodOptional<z.ZodEnum<{
|
||||
data: "data";
|
||||
review: "review";
|
||||
}>>;
|
||||
}, z.core.$strip>>>;
|
||||
phases: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
||||
@@ -667,6 +671,7 @@ export declare const activityManifestSchema: z.ZodObject<{
|
||||
doneWhen: z.ZodOptional<z.ZodString>;
|
||||
cardVariant: z.ZodOptional<z.ZodEnum<{
|
||||
data: "data";
|
||||
review: "review";
|
||||
}>>;
|
||||
}, z.core.$strip>>>;
|
||||
phases: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
||||
|
||||
@@ -20,7 +20,7 @@
|
||||
*/
|
||||
import { describe, expect, it } from 'vitest';
|
||||
import { emptyGate, findKnownRow, gateAddressesOf, gateOf, parseRowKey, resolveRowKey, rowChanged, writesOf } from '../routes/desk.ts';
|
||||
import { createOnlyViolation, missingReadings } from '../wiring/engines.ts';
|
||||
import { createOnlyViolation, decisionReadings, missingReadings } from '../wiring/engines.ts';
|
||||
|
||||
/** A1 受理记录的声明(裁到这一段用得着的几列):PN 与年用量必填 */
|
||||
const RECORD = {
|
||||
@@ -574,3 +574,19 @@ describe('createOnly:键已在正式表 → 拒并指路', () => {
|
||||
expect(createOnlyViolation(sub, { internal_pn: 'XL2596S-5.0', param_name: '输出电压' }, rows)).toBeUndefined();
|
||||
});
|
||||
});
|
||||
|
||||
/** 决策卡的回答是留痕里的事实(2026-08-27:审核哨兵不再读模型自写的旗) */
|
||||
describe('decisionReadings:按 card-reply 数', () => {
|
||||
it('最后一次选的值 + 按值计数;custom: 一律归 custom', () => {
|
||||
const trail = [
|
||||
{ kind: 'card-sent', detail: {} },
|
||||
{ kind: 'card-reply', detail: { value: 'custom:重做一次' } },
|
||||
{ kind: 'card-reply', detail: { value: 'custom:再改一下' } },
|
||||
{ kind: 'card-reply', detail: { value: 'approve' } },
|
||||
];
|
||||
expect(decisionReadings(trail)).toEqual({ decision_last: 'approve', decision_approve_total: 1, decision_reject_total: 0, decision_custom_total: 2 });
|
||||
});
|
||||
it('一张都没答 → last 为空串、各计数铺 0(哨兵里 decision_last = "approve" 恒假,门关着)', () => {
|
||||
expect(decisionReadings([])).toEqual({ decision_last: '', decision_approve_total: 0, decision_reject_total: 0, decision_custom_total: 0 });
|
||||
});
|
||||
});
|
||||
|
||||
@@ -443,6 +443,31 @@ export function touchedRows(
|
||||
* 声明过的工具**一律先铺 0** —— 与 `_rows` 同一个理由:未绑定的名字参与比较求值为假,
|
||||
* 「没调过」与「调了 0 次」在业务上是同一件事,读数上也得是。
|
||||
*/
|
||||
/**
|
||||
* **决策卡的回答是留痕里的事实,不是模型的记账**(2026-08-27 W05 实测逼出):
|
||||
* 审核那一步的入场哨兵读的是模型自写的 `review_approved_total`,人答了「重做」它照写 1,一路推到放行。
|
||||
* 与 A6「该不该叫人不能由被叫的那一方说了算」同一课。基座按 card-reply 数:
|
||||
* `decision_last` —— 这一单最后一张决策卡人选的值(原样;`custom:…` 保留前缀)
|
||||
* `decision_<值>_total` —— 按值计数(`custom:` 开头的一律归 `decision_custom_total`)
|
||||
* 值的语义(approve 是通过)归活动的哨兵去写;这里只把事实摆出来。
|
||||
*/
|
||||
export function decisionReadings(
|
||||
trail: ReadonlyArray<{ readonly kind: string; readonly detail?: Record<string, unknown> | undefined }>,
|
||||
): Record<string, number | string> {
|
||||
const out: Record<string, number | string> = { decision_last: '', decision_approve_total: 0, decision_reject_total: 0, decision_custom_total: 0 };
|
||||
for (const entry of trail) {
|
||||
if (entry.kind !== 'card-reply') continue;
|
||||
const raw = entry.detail?.['value'];
|
||||
if (typeof raw !== 'string' || raw.trim() === '') continue;
|
||||
const value = raw.trim();
|
||||
out['decision_last'] = value;
|
||||
const token = value.startsWith('custom:') ? 'custom' : value.replace(/[^0-9a-zA-Z]+/g, '_');
|
||||
const key = `decision_${token}_total`;
|
||||
out[key] = Number(out[key] ?? 0) + 1;
|
||||
}
|
||||
return out;
|
||||
}
|
||||
|
||||
export function toolCallReadings(
|
||||
trail: ReadonlyArray<{ readonly kind: string; readonly title: string }>,
|
||||
declared: ReadonlyArray<{ readonly name: string }>,
|
||||
@@ -935,6 +960,7 @@ export function wireEngines(deps: EngineWiringDeps): EngineWiring {
|
||||
...tallyReadings(mine, declared),
|
||||
...distinctReadings(mine, declared),
|
||||
...toolCallReadings(trail, manifest.declarations.tools),
|
||||
...decisionReadings(trail),
|
||||
...mine,
|
||||
...touched,
|
||||
};
|
||||
@@ -1325,6 +1351,7 @@ export function wireEngines(deps: EngineWiringDeps): EngineWiring {
|
||||
}>;
|
||||
basis?: ReadonlyArray<{ k: string; v: string }>;
|
||||
allowCustom?: boolean;
|
||||
itemId?: string | undefined;
|
||||
}) {
|
||||
const cardSpan = deps.trace.beginAction({
|
||||
taskId: input.taskId,
|
||||
@@ -1333,10 +1360,17 @@ export function wireEngines(deps: EngineWiringDeps): EngineWiring {
|
||||
input: { options: input.options.map((o) => o.value) },
|
||||
});
|
||||
const recommended = input.options.find((o) => o.recommended === true);
|
||||
/* 审核卡(剧本 cardVariant:'review'):按验收工单的样子摆,机制仍是决策卡 —— 剧本说了算,不靠模型抄 */
|
||||
const reviewFace =
|
||||
input.itemId !== undefined &&
|
||||
(await deps.registry.manifestOf((await taskContext.of(input.taskId)).activityId)).assets.businessModel.steps?.find(
|
||||
(st) => st.item === input.itemId,
|
||||
)?.cardVariant === 'review';
|
||||
const issued = await deps.humanTool.issueChatCard(
|
||||
{ taskId: input.taskId, ownerId: input.ownerId, humanCallCap: input.humanCallCap },
|
||||
{
|
||||
k: 'decision',
|
||||
...(reviewFace ? { variant: 'review' as const } : {}),
|
||||
id: cardIdSchema.parse(ids.next()),
|
||||
at: clock.now().toISOString(),
|
||||
ask: input.ask,
|
||||
@@ -1744,7 +1778,8 @@ export function wireEngines(deps: EngineWiringDeps): EngineWiring {
|
||||
: (await deps.registry.manifestOf((await taskContext.of(input.taskId)).activityId)).assets.businessModel.steps?.find(
|
||||
(s) => s.item === input.itemId,
|
||||
)?.cardVariant;
|
||||
const variant = stepVariant ?? (input.card['variant'] === 'data' ? ('data' as const) : undefined);
|
||||
// 执行工单这一路只认 data(review 是决策卡那一路的面孔)
|
||||
const variant = stepVariant === 'data' || input.card['variant'] === 'data' ? ('data' as const) : undefined;
|
||||
const issued = await deps.humanTool.issueExecution(context, {
|
||||
title,
|
||||
basis: basis.length > 0 ? basis : [{ k: '来源', v: '案例模型当下局面' }],
|
||||
|
||||
@@ -311,6 +311,8 @@ function boundNames(manifest: AcceptanceAuditInput): ReadonlySet<string> {
|
||||
for (const tool of manifest.declarations.tools ?? []) {
|
||||
out.add(`tool_${tool.name.replace(/[^0-9a-zA-Z]+/g, '_')}_calls`);
|
||||
}
|
||||
// 决策卡回答读数(2026-08-27):人在决策卡上选了什么,基座按留痕数 —— 审核通过与否不再由模型自写的旗说了算
|
||||
for (const name of ['decision_last', 'decision_approve_total', 'decision_reject_total', 'decision_custom_total']) out.add(name);
|
||||
for (const id of manifest.assets.businessModel.cmmnXml.match(/caseFileItem\s+id="([^"]+)"/g) ?? []) {
|
||||
const m = /id="([^"]+)"/.exec(id);
|
||||
if (m?.[1] !== undefined) out.add(m[1]);
|
||||
|
||||
@@ -132,3 +132,7 @@
|
||||
(`PlanStep.cardVariant`:这一步派的是数据工单),**不靠模型抄卡骨架**(实测它会把 `variant` 抄丢)。
|
||||
引擎自己发的卡(验收 / 兜底)没有项,不传。
|
||||
|
||||
近变(2026-08-27):`cards.decide` 同样多一项可选 `itemId`(`cardVariant: 'review'` = 审核卡摆成验收工单的样子,机制仍是决策卡);
|
||||
决策卡的回答进基座读数 `decision_last` / `decision_<值>_total`(按 card-reply 数)—— 审核通过与否不再由模型自写的旗说了算;
|
||||
拒回 / 观察 / 事实行按答案分岔(`chosenGuidance`):`custom:` 不是通过,不许写通过旗、不许 finish。
|
||||
|
||||
|
||||
@@ -1729,7 +1729,16 @@ describe('「这张卡人已经答过了」摆成每拍都在的事实行', () =
|
||||
const line = answeredCardFact([sent('请审核结论'), { kind: 'card-reply', title: '请审核结论', detail: { value: 'approve' } }]);
|
||||
expect(line).toContain('人已经答过了');
|
||||
expect(line).toContain('他选的是「approve」');
|
||||
expect(line).toContain('finish');
|
||||
expect(line).toContain('通过才收步');
|
||||
});
|
||||
|
||||
it('**custom: 的答复不是通过**(2026-08-27 W05:人两次答「重做一次」,旧话术「把结论落进台账再 finish」把模型推去写了通过旗,一路推到放行)', () => {
|
||||
const line = answeredCardFact([sent('请审核结论'), { kind: 'card-reply', title: '请审核结论', detail: { value: 'custom:重做一次' } }]);
|
||||
expect(line).toContain('写了自己的话:「重做一次」');
|
||||
expect(line).toContain('这不是通过');
|
||||
expect(line).toContain('换一种问法');
|
||||
expect(line).toContain('不许 finish');
|
||||
expect(line).not.toContain('把结论落进台账');
|
||||
});
|
||||
|
||||
it('执行工单没有值 → 说清「空答复 = 照原样确认」,不编一个选择出来', () => {
|
||||
|
||||
@@ -330,7 +330,28 @@ export function answeredCardFact(
|
||||
const head = `【这张卡人已经答过了】「${reply.title.length <= 60 ? reply.title : `${reply.title.slice(0, 60)}…`}」`;
|
||||
return chose === undefined
|
||||
? `${head} → 人已交回(**空答复 = 照原样确认**,不是没答)。别再发一张一样的卡 —— 照说明书里「人交回之后」那一段办,办完 finish 这一项。`
|
||||
: `${head} → **他选的是「${chose}」**。别再发一张一样的卡 —— 照说明书处理这个选择(通常是把结论落进台账的那一面记账旗,再 finish 这一项)。`;
|
||||
: `${head} → ${chosenGuidance(chose)}`;
|
||||
}
|
||||
|
||||
/**
|
||||
* **人选了什么,下一步就该怎么走** —— 拒回、观察、事实行三处同一句话,按答案分岔。
|
||||
*
|
||||
* 从前不分岔,一律说「照这个选择把结论落进台账那面旗,再 finish」—— 对「通过」是对的,
|
||||
* 对「重做」是把模型往错路上推:2026-08-27 W05 实测,人两次答 `custom:重做一次`,模型照这句
|
||||
* 写了 review_approved_total = 1,审核那一步就被收掉、一路推到放行。
|
||||
* `custom:` 是平台定义的自由答复,平台知道它**不是任何一个选项**;选项值的语义平台不知道,只能指回说明书。
|
||||
*/
|
||||
export function chosenGuidance(chose: string | undefined): string {
|
||||
if (chose === undefined) return '他选了什么就在你的轨迹里(card-reply 那一条,detail 带着选中的值),照说明书里那一档办。';
|
||||
if (chose.startsWith('custom:')) {
|
||||
const said = chose.slice(7).trim();
|
||||
return (
|
||||
`**他没选任何一个选项,写了自己的话:「${said}」** —— 这不是通过。按他写的去改台账(同 PK 覆盖写),` +
|
||||
'改完**换一种问法**再发一张新卡(标题里写明这一轮改了什么;字面一样的卡会被当成同一张吞掉);' +
|
||||
'**不许写「已通过」那类记账旗、不许 finish 这一项**。'
|
||||
);
|
||||
}
|
||||
return `**他选的是「${chose}」** —— 照说明书里「${chose}」那一档办(通过才收步;驳回就改台账、换问法重发)。`;
|
||||
}
|
||||
|
||||
/**
|
||||
@@ -1211,11 +1232,8 @@ export function createTurnRunner(deps: TurnRunnerDeps): LoopEngineService {
|
||||
if (verdict.kind === 'reject') {
|
||||
throw new Error(
|
||||
`这个岔路口人已经选过了(「${verdict.answered}」)—— 别再发一张一样的决策卡。` +
|
||||
(verdict.chose === undefined
|
||||
? '他选了什么就在你的轨迹里(card-reply 那一条,detail 带着选中的值):'
|
||||
: `**他选的是「${verdict.chose}」** —— `) +
|
||||
'照这个选择往下办:照说明书把结论落进台账(通常是那一面记账旗),' +
|
||||
'再 finish 这一项。真有**另一个**岔路口要人定,换一件事、换一个问法。',
|
||||
chosenGuidance(verdict.chose) +
|
||||
'真有**另一个**岔路口要人定,换一件事、换一个问法。',
|
||||
);
|
||||
}
|
||||
await deps.memory.append({
|
||||
@@ -1224,10 +1242,7 @@ export function createTurnRunner(deps: TurnRunnerDeps): LoopEngineService {
|
||||
title: echoTitle,
|
||||
detail: {
|
||||
internal: true,
|
||||
note:
|
||||
verdict.chose === undefined
|
||||
? '人已经在这个岔路口选过了(选中的值在 card-reply 那一条)。别再发 —— 照他的选择接着办,办完把这一步 finish 掉。'
|
||||
: `人已经在这个岔路口选过了,他选的是「${verdict.chose}」。别再发 —— 照这个选择把结论落进台账,再 finish 这一项。`,
|
||||
note: '人已经在这个岔路口选过了。别再发同一张 —— ' + chosenGuidance(verdict.chose),
|
||||
},
|
||||
});
|
||||
// 什么也没做,局面原样交回
|
||||
@@ -1236,6 +1251,7 @@ export function createTurnRunner(deps: TurnRunnerDeps): LoopEngineService {
|
||||
const advanced = readyToWork((await loaded(taskId)).parsed, runtime, action.itemId);
|
||||
const card = await deps.cards.decide({
|
||||
taskId,
|
||||
itemId: action.itemId,
|
||||
ownerId: context.ownerId,
|
||||
humanCallCap: context.humanCallCap,
|
||||
ask: action.ask,
|
||||
|
||||
@@ -208,6 +208,8 @@ export type CardGateway = {
|
||||
}>;
|
||||
basis?: ReadonlyArray<{ k: string; v: string }>;
|
||||
allowCustom?: boolean;
|
||||
/** 这张卡挂在哪个计划项上 —— 组装层按它查剧本声明(`cardVariant: 'review'` = 摆成验收工单) */
|
||||
itemId?: string | undefined;
|
||||
}): Promise<{ id: string }>;
|
||||
};
|
||||
|
||||
|
||||
@@ -504,13 +504,16 @@ function Msg({
|
||||
<Ag meta={`AI · ${hm(m.at)}`}>
|
||||
<div className="jc-ag-b card">
|
||||
<p className="jc-content">{m.ask}</p>
|
||||
<div className="jc-fc t-dec">
|
||||
{/*
|
||||
* 审核卡(契约 `variant:'review'`)摆成验收工单的样子(2026-08-27 用户:「验收工单弹出的是
|
||||
* 需要你定,明显不对」):平台的验收工单是终局,中间审核只能靠决策卡循环 —— 面孔换成
|
||||
* 「工单 · 验收」,选项就是通过 / 驳回,机制不变。
|
||||
*/}
|
||||
<div className={'jc-fc ' + (m.variant === 'review' ? 't-acc' : 't-dec')}>
|
||||
<div className="jc-fc-h">
|
||||
<Ico name="git-branch" s={13} />
|
||||
<b>需要你定</b>
|
||||
<em>
|
||||
{m.options.length + (m.allowCustom ? 1 : 0)} 选 1
|
||||
</em>
|
||||
<Ico name={m.variant === 'review' ? 'shield-check' : 'git-branch'} s={13} />
|
||||
<b>{m.variant === 'review' ? '工单 · 验收' : '需要你定'}</b>
|
||||
<em>{m.variant === 'review' ? '审核' : `${m.options.length + (m.allowCustom ? 1 : 0)} 选 1`}</em>
|
||||
</div>
|
||||
<div className="jc-fc-b">
|
||||
<Kv items={m.basis} />
|
||||
|
||||
@@ -520,6 +520,13 @@ const chatBase = { id: cardIdSchema, at: z.iso.datetime() };
|
||||
export const decisionCardSchema = z.object({
|
||||
...chatBase,
|
||||
k: z.literal('decision'),
|
||||
/**
|
||||
* **审核卡**(2026-08-27 用户:「验收工单弹出的是需要你定,明显不对,需要是验收工单」):
|
||||
* 平台的验收工单是终局(`submit` 一单只发一次,交出去就是放行),中间那道「审核 → 驳回 → 修订 → 再审」
|
||||
* 只能用决策卡循环 —— 但它不该穿着「需要你定」的衣服。`review` = 按验收工单的样子摆
|
||||
* (工单 · 验收 / 通过 · 驳回),机制仍是决策卡。由剧本 `cardVariant` 声明、引擎决定,不靠模型抄。
|
||||
*/
|
||||
variant: z.enum(['review']).optional(),
|
||||
ask: z.string(),
|
||||
basis: basisSchema,
|
||||
recommend: z.string().optional(),
|
||||
|
||||
@@ -160,13 +160,14 @@ export const planStepSchema = z.object({
|
||||
*/
|
||||
doneWhen: z.string().optional(),
|
||||
/**
|
||||
* **这一步派的执行工单是数据工单**(2026-08-27)—— 由剧本声明、引擎决定,不靠模型抄卡骨架。
|
||||
* **这一步派的卡穿哪副面孔**(2026-08-27)—— 由剧本声明、引擎决定,不靠模型抄卡骨架。
|
||||
* `data` = 执行工单摆成数据工单;`review` = 决策卡摆成验收工单(中间审核,可循环;终局验收仍是 submit)。
|
||||
*
|
||||
* 起因:说明书的卡骨架里写着 `"variant":"data"`,模型抄卡时把它丢了(看着像元数据),
|
||||
* 对话区于是照旧摆「执行工单」。与 `view` 同一条道理:这一步派的是什么单,是写剧本时就定死的事,
|
||||
* 归活动声明、不归模型临场发挥;模型写不写都一样。
|
||||
*/
|
||||
cardVariant: z.enum(['data']).optional(),
|
||||
cardVariant: z.enum(['data', 'review']).optional(),
|
||||
});
|
||||
export type PlanStep = z.infer<typeof planStepSchema>;
|
||||
|
||||
|
||||
Reference in New Issue
Block a user