XiaChangSong/JX-BPOS
1
0
Fork 0
forked from ChaiHao/BPOS-MVP
Code Pull Requests Activity

度量八指标全部接真:DMN 覆盖率与单位成本不再恒为 0

Browse Source 
两处度量是空的,查下来都是「没人记」而不是「算错了」:

① **DMN 覆盖率恒 0**:判据只认标题里的「求值:」,把跑得最多的**纯工具步**整个漏掉。
   DMN 决策表与机检正是落在纯工具/工具步上。改为按步类型分:
   确定性求值(pure-tool / tool)÷ 判断总数(再加 LLM 步)——「临场搬进确定」的进度表。
   顺带把**步类型显式写进留痕**:判据该由引擎给出,不靠解析标题后缀。
   现读数 66%。

② **单位成本恒 0**:任务消耗只记步数,token / 活跃时长 / 调人次数**从来没人累加** ——
   于是八指标里的单位成本永远是 0,预算熔断与打扰上限也形同虚设。
   · tick 计这一拍真正花掉的时间(挂起等人的不算活跃,那是世界等待);
   · flow 的 LLM 步与 loop 的每一回合都把 token 计入;
   · 每发一张卡计一次调人。
   现读数:A2 单轮 12k token、A3 3.4k;单位成本 73 token 每放行单。

另修:执行卡的 ask 优先取 `card.ask`(此前只认 `reason`,新卡面格式下会退化成标题)。

第四轮端到端复跑:整条链闭环,消耗数据真实。

gates: tsc / eslint / depcruise(336 模块)/ api-report / 429 测试 全绿

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KeDBx2PHLavbURXB8LJJ7m
This commit is contained in:
ChaiHao
2026-07-27 11:44:10 +08:00
co-authored by Claude Opus 5
parent 94f31f267b
commit 61fe181678
8 changed files with 97 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files
apps/api/src/app/agent-runner.ts
+3
View File
@@ -51,6 +51,8 @@ export type AgentRunnerDeps = {
readonly gateway: ModelGateway;
/** 活动声明的自定义模型名 */
readonly modelOf: (taskId: TaskId) => Promise<string>;
/** token 回报:计入任务预算 —— 不计,预算熔断就永远不会响 */
readonly onUsage: (taskId: TaskId, tokens: number) => Promise<void>;
};
export function createAgentRunner(deps: AgentRunnerDeps): {
@@ -95,6 +97,7 @@ export function createAgentRunner(deps: AgentRunnerDeps): {
taskId,
label: 'loop-turn',
});
await deps.onUsage(taskId, result.usage.inputTokens + result.usage.outputTokens);
return normalize(result.object, enabled);
} catch (error) {
/*
apps/api/src/app/wiring/desk-elements.ts
+23 -12
View File
@@ -45,18 +45,29 @@ export function wireDeskElements(deps: DeskElementDeps): { minute: MinuteService
clock: deps.clock,
});
/** 归侧日志行 → 度量事实。判断类与断点由日志标题的确定形态识别,不另设埋点 */
const toLogFact = (taskId: string, line: { at: string; side: LogFact['side']; title: string }): LogFact => ({
taskId,
at: line.at,
side: line.side,
title: line.title,
isHumanToolCall: line.title.startsWith('发出:'),
// DMN / 机检的确定性求值在日志里是「求值:」;LLM 的判断落在「调用:」
isDeterministicJudgement: line.title.startsWith('求值:'),
isJudgement: line.title.startsWith('求值:') || line.title.startsWith('调用:'),
isBreakpoint: line.title.includes('断点') || line.title.includes('交集为空'),
});
/**
* 归侧日志行 → 度量事实。零埋点:判据全部取自结构化日志本身。
*
* **判断次数**包含两类:确定性求值(DMN 决策表 / 机检 —— 落在纯工具与工具步)
* 与临场判断(LLM 步、loop 的语义性前提评估)。
* DMN 覆盖率 = 前者 ÷ 两者之和 ——「临场搬进确定」的进度表。
* 早先只认标题里的「求值:」,把跑得最多的纯工具步整个漏掉了,覆盖率恒为 0。
*/
const toLogFact = (taskId: string, line: { at: string; side: LogFact['side']; title: string }): LogFact => {
const deterministic =
line.title.startsWith('求值:') || line.title.includes('(pure-tool)') || line.title.includes('(tool)');
const situational = line.title.includes('(llm)');
return {
taskId,
at: line.at,
side: line.side,
title: line.title,
isHumanToolCall: line.title.startsWith('发出:'),
isDeterministicJudgement: deterministic,
isJudgement: deterministic || situational,
isBreakpoint: line.title.includes('断点') || line.title.includes('交集为空'),
};
};
const report = createReportService({
store: createReportStore(deps.db),
apps/api/src/app/wiring/engines.ts
+10 -1
View File
@@ -270,6 +270,7 @@ export function wireEngines(deps: EngineWiringDeps): EngineWiring {
criteria: [{ label: '成果齐备', level: 'L5', pass: false }],
fields: [],
});
await taskStore.addConsumption(input.taskId, { humanCalls: 1 });
return { id: issued.order.id };
}
const fields = Array.isArray(input.card['fields'])
@@ -278,9 +279,11 @@ export function wireEngines(deps: EngineWiringDeps): EngineWiring {
const issued = await deps.humanTool.issueExecution(context, {
title,
basis: basis.length > 0 ? basis : [{ k: '来源', v: '步序声明' }],
ask: String(input.card['reason'] ?? title),
ask: String(input.card['ask'] ?? input.card['reason'] ?? title),
fields,
});
// 打扰率的计量对象:每发一张卡就是调了一次人 —— 不计就等于没有上限
await taskStore.addConsumption(input.taskId, { humanCalls: 1 });
return { id: issued.order.id };
},
};
@@ -348,6 +351,10 @@ export function wireEngines(deps: EngineWiringDeps): EngineWiring {
stepNo: input.stepNo,
label: 'flow-step',
});
// token 计入任务预算 —— 单位成本与预算熔断都靠它,不计就等于没有预算
await taskStore.addConsumption(input.taskId, {
tokens: result.usage.inputTokens + result.usage.outputTokens,
});
// 解析与校验归引擎(它才知道这一步声明了什么);这里只如实上交结果与原文
return {
ok: true,
@@ -373,6 +380,8 @@ export function wireEngines(deps: EngineWiringDeps): EngineWiring {
data: taskData,
agent: createAgentRunner({
gateway: deps.modelGateway,
// loop 的每一回合都要问模型 —— 不计 token,预算熔断就永远不会响
onUsage: (taskId, tokens) => taskStore.addConsumption(taskId, { tokens }).then(() => undefined),
modelOf: async (taskId) => {
const { activityId } = await taskContext.of(taskId);
const manifest = await deps.registry.manifestOf(activityId);
apps/api/src/modules/flow-engine/internal/usecase/flow-runner.ts
+2 -1
View File
@@ -128,7 +128,8 @@ export function createFlowRunner(deps: FlowRunnerDeps): FlowEngineService {
taskId,
kind: 'action',
title: `步 ${index + 1} · ${step.label}(${step.kind})`,
detail: { step: step.name, attempt, input: payload },
// 步类型显式入留痕:度量的判据由引擎给出,不靠解析标题后缀
detail: { step: step.name, stepKind: step.kind, attempt, input: payload },
});
switch (step.kind) {
apps/api/src/modules/task/internal/usecase/task-service.ts
+7 -2
View File
@@ -133,9 +133,14 @@ export function createTaskService(deps: TaskServiceDeps): TaskService {
return { task: finished, verdict, progressed: false };
}
const startedAt = deps.clock.now().getTime();
const outcome = await engineOf(task).advance(taskId);
// 步计量:loop 的一次决策循环 / flow 的一个声明步
const used = await deps.store.addConsumption(taskId, { steps: 1 });
// 步计量:loop 的一次决策循环 / flow 的一个声明步。
// 活跃时长按**这一拍真正花掉的时间**计 —— 挂起等人的时间不算活跃,那是世界等待。
const used = await deps.store.addConsumption(taskId, {
steps: 1,
activeMs: Math.max(0, deps.clock.now().getTime() - startedAt),
});
await deps.store.markProgress(taskId, deps.clock.now());
const runtime =