import request from "supertest"; import { afterEach, describe, expect, it, vi } from "vitest"; const FLAG_KEYS = [ "FEATURE_ASSISTANT_INVESTIGATION_STATE_V1", "FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1", "FEATURE_ASSISTANT_CONTRACTS_V11", "FEATURE_ASSISTANT_PROBLEM_UNITS_V1", "FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1", "FEATURE_ASSISTANT_ANSWER_POLICY_V11", "FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1" ] as const; const ORIGINAL_FLAGS: Record = 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 createAppWithFlags(flags: { state: "0" | "1"; binding: "0" | "1"; contracts?: "0" | "1"; problemUnits?: "0" | "1"; continuity?: "0" | "1"; answerPolicy?: "0" | "1"; problemCentric?: "0" | "1"; }) { process.env.FEATURE_ASSISTANT_INVESTIGATION_STATE_V1 = flags.state; process.env.FEATURE_ASSISTANT_STATE_FOLLOWUP_BINDING_V1 = flags.binding; process.env.FEATURE_ASSISTANT_CONTRACTS_V11 = flags.contracts ?? "1"; process.env.FEATURE_ASSISTANT_PROBLEM_UNITS_V1 = flags.problemUnits ?? "0"; process.env.FEATURE_ASSISTANT_PROBLEM_UNIT_CONTINUITY_V1 = flags.continuity ?? "0"; process.env.FEATURE_ASSISTANT_ANSWER_POLICY_V11 = flags.answerPolicy ?? "0"; process.env.FEATURE_ASSISTANT_PROBLEM_CENTRIC_ANSWER_V1 = flags.problemCentric ?? "0"; vi.resetModules(); const { createApp } = await import("../src/server"); return createApp(); } describe.sequential("assistant follow-up state binding", () => { afterEach(() => { restoreFlags(); vi.resetModules(); }); it("applies investigation_state binding in follow-up flow when flags are ON", async () => { const app = await createAppWithFlags({ state: "1", binding: "1" }); const sessionId = `asst-wave2-on-${Date.now()}`; const first = await request(app).post("/api/assistant/message").send({ session_id: sessionId, useMock: true, promptVersion: "normalizer_v2_0_2", user_message: "Разложи цепочку документов по контрагентам." }); expect(first.status).toBe(200); expect(first.body.debug?.followup_state_usage).toBeUndefined(); expect(first.body.debug?.investigation_state_snapshot?.turn_index).toBe(1); const second = await request(app).post("/api/assistant/message").send({ session_id: sessionId, useMock: true, promptVersion: "normalizer_v2_0_2", user_message: "И по периоду 2020-06." }); expect(second.status).toBe(200); expect(second.body.debug?.followup_state_usage?.applied).toBe(true); expect(second.body.debug?.followup_state_usage?.context_patch?.business_context_from_state).toBe(true); expect(second.body.debug?.followup_state_usage?.state_turn_index).toBe(1); expect( (second.body.debug?.routes ?? []).some((item: { route?: string }) => item.route && item.route !== "no_route") ).toBe(true); expect(second.body.debug?.investigation_state_snapshot?.turn_index).toBe(2); }); it("does not apply follow-up binding when binding flag is OFF", async () => { const app = await createAppWithFlags({ state: "1", binding: "0" }); const sessionId = `asst-wave2-off-${Date.now()}`; const first = await request(app).post("/api/assistant/message").send({ session_id: sessionId, useMock: true, promptVersion: "normalizer_v2_0_2", user_message: "Разложи цепочку документов по контрагентам." }); expect(first.status).toBe(200); const second = await request(app).post("/api/assistant/message").send({ session_id: sessionId, useMock: true, promptVersion: "normalizer_v2_0_2", user_message: "И по периоду 2020-06." }); expect(second.status).toBe(200); expect(second.body.debug?.followup_state_usage).toBeUndefined(); expect((second.body.debug?.routes ?? []).every((item: { route?: string }) => item.route === "no_route")).toBe(true); }); it("keeps legacy-like behavior when investigation state flag is OFF", async () => { const app = await createAppWithFlags({ state: "0", binding: "1" }); const response = await request(app).post("/api/assistant/message").send({ useMock: true, promptVersion: "normalizer_v2_0_2", user_message: "И по периоду 2020-06." }); expect(response.status).toBe(200); expect(response.body.debug?.investigation_state_snapshot).toBeNull(); expect(response.body.debug?.followup_state_usage).toBeUndefined(); }); it("applies problem continuity hints only when continuity flag is ON and follow-up has no strong new anchors", async () => { const app = await createAppWithFlags({ state: "1", binding: "1", problemUnits: "1", continuity: "1" }); const sessionId = `asst-wave4-problem-continuity-${Date.now()}`; const first = await request(app).post("/api/assistant/message").send({ session_id: sessionId, useMock: true, promptVersion: "normalizer_v2_0_2", user_message: "Разложи цепочку документов и оплат по контрагентам за 2020-06, где разрыв механизма закрытия." }); expect(first.status).toBe(200); expect(first.body.debug?.investigation_state_snapshot?.problem_unit_state).toBeTruthy(); const second = await request(app).post("/api/assistant/message").send({ session_id: sessionId, useMock: true, promptVersion: "normalizer_v2_0_2", user_message: "И по тому же разрыву добавь уточнение." }); expect(second.status).toBe(200); expect(second.body.debug?.followup_state_usage?.applied).toBe(true); expect(second.body.debug?.followup_state_usage?.context_patch?.problem_continuity_available).toBe(true); expect(second.body.debug?.followup_state_usage?.context_patch?.problem_continuity_applied).toBe(true); expect(second.body.debug?.followup_state_usage?.context_patch?.strong_new_anchor_detected).toBe(false); }); it("does not apply follow-up continuity when user gives strong new anchors", async () => { const app = await createAppWithFlags({ state: "1", binding: "1", problemUnits: "1", continuity: "1" }); const sessionId = `asst-wave4-strong-anchor-${Date.now()}`; const first = await request(app).post("/api/assistant/message").send({ session_id: sessionId, useMock: true, promptVersion: "normalizer_v2_0_2", user_message: "Разбери хвосты по счету 60 за 2020-06 и покажи проблемные цепочки." }); expect(first.status).toBe(200); expect(first.body.debug?.investigation_state_snapshot?.turn_index).toBe(1); const second = await request(app).post("/api/assistant/message").send({ session_id: sessionId, useMock: true, promptVersion: "normalizer_v2_0_2", user_message: "И отдельно по счету 97 за 2020-07." }); expect(second.status).toBe(200); expect(second.body.debug?.followup_state_usage).toBeUndefined(); expect(second.body.debug?.investigation_state_snapshot?.turn_index).toBe(2); }); it("isolates scope for independent cross-domain turn and does not carry stale period/domain", async () => { const app = await createAppWithFlags({ state: "1", binding: "1", problemUnits: "1", continuity: "1", answerPolicy: "1", problemCentric: "1" }); const sessionId = `asst-wave17-scope-isolation-${Date.now()}`; const first = await request(app).post("/api/assistant/message").send({ session_id: sessionId, useMock: true, promptVersion: "normalizer_v2_0_2", user_message: "Проверь хвосты по счету 60.01 за 2020-06 и почему не закрылся долг." }); expect(first.status).toBe(200); expect(first.body.debug?.investigation_state_snapshot?.question_scope_id).toContain("d:settlements_60_62"); expect(first.body.debug?.investigation_state_snapshot?.scope_origin).toBe("explicit_from_message"); const second = await request(app).post("/api/assistant/message").send({ session_id: sessionId, useMock: true, promptVersion: "normalizer_v2_0_2", user_message: "Проверь НДС по счету-фактуре: где разрыв в цепочке документ -> регистр -> книга?" }); expect(second.status).toBe(200); expect(second.body.debug?.followup_state_usage).toBeUndefined(); expect(second.body.debug?.investigation_state_snapshot?.scope_origin).toBe("explicit_from_message"); expect(String(second.body.debug?.investigation_state_snapshot?.question_scope_id ?? "")).toContain( "d:vat_document_register_book" ); expect(second.body.debug?.investigation_state_snapshot?.focus?.period).not.toBe("2020-06"); expect(String(second.body.debug?.investigation_state_snapshot?.focus?.domain ?? "")).not.toContain("settlements_60_62"); expect(second.body.debug?.investigation_state_snapshot?.followup_context?.question_scope_id).toBeTruthy(); expect(second.body.debug?.investigation_state_snapshot?.followup_context?.scope_origin).toBe("explicit_from_message"); }); it("rebinds follow-up domain away from settlements on fixed-asset amortization query", async () => { const app = await createAppWithFlags({ state: "1", binding: "1", problemUnits: "1", continuity: "1", answerPolicy: "1", problemCentric: "1" }); const sessionId = `asst-wave16-fa-domain-${Date.now()}`; const first = await request(app).post("/api/assistant/message").send({ session_id: sessionId, useMock: true, promptVersion: "normalizer_v2_0_2", user_message: "Почему деньги ушли, а долг по 60.01/62.02 остался?" }); expect(first.status).toBe(200); expect(first.body.debug?.investigation_state_snapshot?.followup_context?.active_domain).toBe("settlements_60_62"); const second = await request(app).post("/api/assistant/message").send({ session_id: sessionId, useMock: true, promptVersion: "normalizer_v2_0_2", user_message: "Полно ли начислена амортизация по объектам ОС за июль? Проверь по 01/02, нет ли пропущенных объектов." }); expect(second.status).toBe(200); const activeDomain = String(second.body.debug?.investigation_state_snapshot?.followup_context?.active_domain ?? ""); expect(activeDomain).not.toBe("settlements_60_62"); expect(activeDomain).toMatch(/fixed_asset_amortization|month_close_costs_20_44|no_route|hybrid_store_plus_live|fixed_asset/i); const settlementActions = second.body.debug?.investigation_state_snapshot?.followup_context?.settlement_next_actions; expect(Array.isArray(settlementActions) ? settlementActions.length : 0).toBe(0); }); it("keeps UTF-8 follow-up period refinement in-scope with soft continuity hints", async () => { const app = await createAppWithFlags({ state: "1", binding: "1", problemUnits: "1", continuity: "1", answerPolicy: "1", problemCentric: "1" }); const first = await request(app).post("/api/assistant/message").send({ useMock: true, promptVersion: "normalizer_v2_0_2", user_message: "Посмотри, пожалуйста, где по поставщикам сейчас хвосты уже похожи именно на проблему, а не просто на шум." }); expect(first.status).toBe(200); expect(first.body.reply_type).not.toBe("out_of_scope"); const second = await request(app).post("/api/assistant/message").send({ session_id: first.body.session_id, useMock: true, promptVersion: "normalizer_v2_0_2", user_message: "А если только за июнь 2020 смотреть, по кому это сильнее всего видно и что из этого реально может мешать закрытию?" }); expect(second.status).toBe(200); expect(second.body.reply_type).not.toBe("out_of_scope"); expect(second.body.debug?.followup_state_usage?.applied).toBe(true); expect(typeof second.body.debug?.followup_state_usage?.context_patch?.problem_continuity_applied).toBe("boolean"); expect(second.body.debug?.followup_state_usage?.context_patch?.strong_new_anchor_detected).toBe(false); expect( (second.body.debug?.routes ?? []).some((item: { route?: string }) => item.route && item.route !== "no_route") ).toBe(true); const third = await request(app).post("/api/assistant/message").send({ useMock: true, promptVersion: "normalizer_v2_0_2", user_message: "Проверь, пожалуйста, по 60-му счёту за июнь 2020, где есть самый явный проблемный участок по расчётам с поставщиками." }); expect(third.status).toBe(200); expect(third.body.reply_type).not.toBe("out_of_scope"); }); });