NODEDC_1C/llm_normalizer/backend/tests/assistantProblemUnitRuntimeRollout.test.ts

import request from "supertest";
import { afterEach, describe, expect, it, vi } from "vitest";

const FLAG_KEYS = [
  "FEATURE_ASSISTANT_PROBLEM_UNITS_V1",
  "FEATURE_ASSISTANT_ANSWER_POLICY_V11",
  "FEATURE_ASSISTANT_BROAD_GUARD_V1",
  "FEATURE_ASSISTANT_MIN_EVIDENCE_GATE_V1",
  "FEATURE_ASSISTANT_ANTI_GENERIC_RANKING_GUARD_V1"
] as const;

const ORIGINAL_FLAGS: Record<string, string | undefined> = Object.fromEntries(
  FLAG_KEYS.map((key) => [key, process.env[key]])
);

function restoreFlags(): void {
  for (const key of FLAG_KEYS) {
    const original = ORIGINAL_FLAGS[key];
    if (original === undefined) {
      delete process.env[key];
    } else {
      process.env[key] = original;
    }
  }
}

async function createAppWithProblemUnitsFlag(flagValue: "0" | "1") {
  process.env.FEATURE_ASSISTANT_PROBLEM_UNITS_V1 = flagValue;
  process.env.FEATURE_ASSISTANT_ANSWER_POLICY_V11 = "1";
  process.env.FEATURE_ASSISTANT_BROAD_GUARD_V1 = "1";
  process.env.FEATURE_ASSISTANT_MIN_EVIDENCE_GATE_V1 = "1";
  process.env.FEATURE_ASSISTANT_ANTI_GENERIC_RANKING_GUARD_V1 = "1";
  vi.resetModules();
  const { createApp } = await import("../src/server");
  return createApp();
}

function routedRetrievalResults(body: Record<string, unknown>): Record<string, unknown>[] {
  const results = Array.isArray((body.debug as { retrieval_results?: unknown[] } | undefined)?.retrieval_results)
    ? ((body.debug as { retrieval_results?: unknown[] }).retrieval_results as Record<string, unknown>[])
    : [];
  return results.filter((item) => String(item.route ?? "") !== "no_route");
}

describe.sequential("assistant problem-unit runtime rollout", () => {
  afterEach(() => {
    restoreFlags();
    vi.resetModules();
  });

  it("emits problem-unit layer on problem-heavy scenarios when flag is ON", async () => {
    const app = await createAppWithProblemUnitsFlag("1");
    const cases = [
      {
        tag: "chain",
        user_message: "Разложи цепочку документов и оплат по контрагентам за 2020-06, где разрыв механизма закрытия."
      },
      {
        tag: "anomaly",
        user_message: "Разложи lifecycle по счету 97 за 2020-06 и покажи аномалии списания по последовательности."
      },
      {
        tag: "contradiction",
        user_message: "Проверь НДС за 2020-06: где противоречия между документами, проводками и регистрами."
      },
      {
        tag: "period_risk",
        user_message: "Разложи по счетам 51 и 60 за 2020-06, что создаёт риск закрытия периода и где разрывы цепочки."
      }
    ];

    const observedTypes = new Set<string>();
    let scenariosWithProblemUnits = 0;
    for (const scenario of cases) {
      const response = await request(app).post("/api/assistant/message").send({
        useMock: true,
        promptVersion: "normalizer_v2_0_2",
        user_message: scenario.user_message
      });
      expect(response.status).toBe(200);

      const routed = routedRetrievalResults(response.body as Record<string, unknown>);
      expect(routed.length).toBeGreaterThan(0);

      const withProblemUnits = routed.filter((item) => Array.isArray(item.problem_units) && item.problem_units.length > 0);
      if (withProblemUnits.length === 0) {
        continue;
      }
      scenariosWithProblemUnits += 1;

      for (const result of withProblemUnits) {
        const summary = (result.summary as Record<string, unknown>) ?? {};
        const candidateEvidence = result.candidate_evidence as Array<Record<string, unknown>>;
        const problemUnits = result.problem_units as Array<Record<string, unknown>>;
        const problemSummary = (result.problem_unit_summary as Record<string, unknown>) ?? {};

        expect(summary.problem_units_enabled).toBe(true);
        expect(summary.candidate_evidence_count).toBe(candidateEvidence.length);
        expect(summary.problem_units_count).toBe(problemUnits.length);
        expect(Array.isArray(summary.problem_unit_types)).toBe(true);
        expect(typeof summary.problem_unit_duplicate_collapses).toBe("number");
        expect(problemSummary.units_total).toBe(problemUnits.length);

        for (const unit of problemUnits) {
          expect(typeof unit.problem_unit_id).toBe("string");
          expect(typeof unit.problem_unit_type).toBe("string");
          expect(typeof unit.mechanism_summary).toBe("string");
          expect(typeof unit.severity?.score).toBe("number");
          expect(typeof unit.confidence?.score).toBe("number");
          observedTypes.add(String(unit.problem_unit_type));
        }
      }
    }

    expect(scenariosWithProblemUnits).toBeGreaterThan(0);
    expect(observedTypes.size).toBeGreaterThan(0);
    expect(Array.from(observedTypes).every((item) =>
      [
        "document_conflict",
        "broken_chain_segment",
        "lifecycle_anomaly_node",
        "unresolved_settlement_cluster",
        "period_risk_cluster",
        "cross_branch_inconsistency_cluster"
      ].includes(item)
    )).toBe(true);
  });

  it("does not emit problem-unit layer when flag is OFF", async () => {
    const app = await createAppWithProblemUnitsFlag("0");

    const response = await request(app).post("/api/assistant/message").send({
      useMock: true,
      promptVersion: "normalizer_v2_0_2",
      user_message: "Разложи цепочку документов и оплат по контрагентам за 2020-06."
    });

    expect(response.status).toBe(200);
    const routed = routedRetrievalResults(response.body as Record<string, unknown>);
    expect(routed.length).toBeGreaterThan(0);
    for (const result of routed) {
      expect(result.raw_entities).toBeUndefined();
      expect(result.candidate_evidence).toBeUndefined();
      expect(result.problem_units).toBeUndefined();
      expect(result.problem_unit_summary).toBeUndefined();
      expect((result.summary as Record<string, unknown>).problem_units_enabled).toBeUndefined();
    }
  });
});