BPOS-MVP/apps/api/src/app/__tests__/agent-runner-retry.test.ts

/**
 * 输出不成形的就地重问(2026-08-03 用户实测送来的单逼出来的)。
 *
 * 病:关思考的模型偶发把一拍输出写坏(整串行动一个都不认识),从前一次没认出
 * 就直接发「需要你判断下一步」的兜底卡,还把 JSON 原话整段摆在卡面上 ——
 * 人打开工单看到的是一坨代码(用户原话:「卡片里又是乱七八糟的东西」)。
 *
 * 法:第一次没认出 → 原文经 `note` 存进工作日志(内部痕),附「只输出 JSON」
 * 就地再问一次;第二次仍不成形才交人,卡面只说人话、不摆原话。
 * 这里用假网关钉死三件事:重问只有一次、原文进痕、好输出不白重问。
 * (ModelGateway 是外界 —— LLM 厂商 —— 的缝,录制回放替身合规。)
 */
import { describe, expect, it } from 'vitest';
import type { EnabledSet } from '@modules/loop-engine';
import type { ModelCall, ModelGateway, ModelResult } from '@modules/llm';
import type { TaskId } from '@contracts';
import { createAgentRunner } from '../agent-runner.ts';

const ENABLED: EnabledSet = {
  entries: [{ id: 'T1', name: '干活', type: 'humanTask', requiresManualStart: false, discretionary: false }],
  inProgress: [],
  judgements: [],
  gaps: ['缺记录'],
  activeStages: [],
};

/** 按序回放给定出参的假网关 —— 超出名单就重复最后一个 */
function gatewayOf(outputs: Array<{ object: unknown; text: string }>): { gateway: ModelGateway; calls: ModelCall[] } {
  const calls: ModelCall[] = [];
  const gateway: ModelGateway = {
    mode: 'replay',
    resolve: () => Promise.reject(new Error('测试里用不到 resolve')),
    generate: (call) => {
      calls.push(call);
      const next = outputs[Math.min(calls.length - 1, outputs.length - 1)];
      if (next === undefined) return Promise.reject(new Error('名单空着'));
      const result: ModelResult = {
        text: next.text,
        object: next.object,
        usage: { inputTokens: 10, outputTokens: 5, latencyMs: 1 },
        binding: { name: 'fake', provider: 'deepseek', model: 'fake' },
        mode: 'replay',
      };
      return Promise.resolve(result);
    },
    withMode: () => gateway,
  };
  return { gateway, calls };
}

function runnerOf(outputs: Array<{ object: unknown; text: string }>) {
  const { gateway, calls } = gatewayOf(outputs);
  const notes: Array<{ title: string; raw: string }> = [];
  const runner = createAgentRunner({
    gateway,
    modelOf: () => Promise.resolve('fake'),
    onUsage: () => Promise.resolve(),
    note: (_taskId, title, raw) => {
      notes.push({ title, raw });
      return Promise.resolve();
    },
  });
  const decide = () =>
    runner.decide({
      taskId: 'task-1' as TaskId,
      narrative: '(空)',
      enabled: ENABLED,
      scope: {},
      writable: [],
      tools: [],
    });
  return { decide, calls, notes };
}

const GOOD = { object: { say: '收工', actions: [{ kind: 'finish', itemId: 'T1' }] }, text: '{"say":"收工"}' };
const BAD = { object: { say: '', actions: [{ kind: '胡编的' }] }, text: '再连贯的散文也不是行动串' };

describe('输出不成形先就地重问,别一抖就叫人', () => {
  it('第一拍坏、重问好 → 两次调用、原文进痕一条、返回的是好行动', async () => {
    const { decide, calls, notes } = runnerOf([BAD, GOOD]);
    const turn = await decide();
    expect(calls).toHaveLength(2);
    // 重问那一次要把「只输出 JSON」的硬提示带上 —— 不带等于原样再撞一次
    expect(calls[1]?.prompt).toContain('不是合法的行动 JSON');
    expect(notes).toHaveLength(1);
    expect(notes[0]?.raw).toContain('散文');
    expect(turn.actions[0]?.kind).toBe('finish');
  });

  it('两拍都坏 → 就收手交人,不无限重问;卡面不摆模型原话', async () => {
    const { decide, calls } = runnerOf([BAD, BAD]);
    const turn = await decide();
    expect(calls).toHaveLength(2);
    const action = turn.actions[0];
    expect(action?.kind).toBe('human');
    if (action?.kind !== 'human') return;
    expect(action.card['modelReply']).toBeUndefined();
    expect(String(action.card['reason'])).toContain('工作日志');
  });

  it('第一拍就是好的 → 只调一次,不白烧重问', async () => {
    const { decide, calls, notes } = runnerOf([GOOD]);
    const turn = await decide();
    expect(calls).toHaveLength(1);
    expect(notes).toHaveLength(0);
    expect(turn.actions[0]?.kind).toBe('finish');
  });
});