"""模型路由与预算测试。 守的是这一层的核心判断: **大模型按关键时刻分配,不按时间分配。** 如果哪天有人把 tick 路由到 high,test_heartbeat_never_uses_model 会红。 如果哪天预算紧张把主对话降成"无模型",test_dialogue_floor_is_low 会红。 """ from __future__ import annotations import pytest from cybergril.providers import EchoLLM from cybergril.think import ( BudgetConfig, BudgetGuard, ModelRouter, Scene, Tier, build_router, floor_for, is_model_scene, ) from cybergril.think.scenes import CONTEXT_BUDGET, NEVER_NONE @pytest.fixture def providers() -> dict[Tier, EchoLLM | None]: return { Tier.HIGH: EchoLLM(prefix="[high]"), Tier.STANDARD: EchoLLM(prefix="[std]"), Tier.LOW: EchoLLM(prefix="[low]"), Tier.NONE: None, } @pytest.fixture def router(providers) -> ModelRouter: return build_router(providers=providers) # ---------------------------------------------------------------- 路由基表 @pytest.mark.parametrize( "scene", [Scene.TICK, Scene.FILLER, Scene.REACH_OUT, Scene.AVATAR_DRIVE], ) def test_heartbeat_and_rules_never_use_model(router: ModelRouter, scene: Scene) -> None: """心跳、填充词、主动性判定、表情:不调用任何模型。""" d = router.route(scene) assert d.effective is Tier.NONE assert d.provider is None assert d.is_rule_based assert d.context_budget == 0 def test_dialogue_main_uses_high(router: ModelRouter) -> None: d = router.route(Scene.DIALOGUE_MAIN) assert d.effective is Tier.HIGH assert d.provider is not None assert d.context_budget == CONTEXT_BUDGET[Tier.HIGH] def test_background_tasks_use_low(router: ModelRouter) -> None: for scene in (Scene.MEMORY_GIST, Scene.FACT_EXTRACT, Scene.LIFE_EVENT): d = router.route(scene) assert d.effective is Tier.LOW, scene def test_is_model_scene_helper() -> None: assert not is_model_scene(Scene.TICK) assert is_model_scene(Scene.DIALOGUE_MAIN) # ---------------------------------------------------------------- 预算降级 def test_budget_tight_drops_to_floor(providers) -> None: """到达硬阈值时直接降到该场景下限,而不是温和过渡。""" guard = BudgetGuard(BudgetConfig(daily_tokens=1000, hard_stop_ratio=0.95), today="2026-01-01") guard.record(960, "test") # 96% router = build_router(providers=providers) router.guard = guard d = router.route(Scene.DIALOGUE_MAIN) assert d.requested is Tier.HIGH assert d.effective is Tier.LOW # dialogue 的下限 assert d.degraded assert "预算" in (d.degraded_reason or "") def test_dialogue_floor_is_low_not_none(providers) -> None: """预算耗尽时,主对话可以降成小模型,但绝不能没有台词。""" guard = BudgetGuard(BudgetConfig(daily_tokens=1000, hard_stop_ratio=0.5), today="2026-01-01") guard.record(999, "test") router = build_router(providers=providers) router.guard = guard d = router.route(Scene.DIALOGUE_MAIN) assert d.effective is Tier.LOW assert d.provider is not None # 还有模型可用 assert floor_for(Scene.DIALOGUE_MAIN) is Tier.LOW def test_crisis_never_downgrades(providers) -> None: """危机场景即使预算耗尽也保持高档位。 这是产品意义上的强制约束,不是性能优化。 """ guard = BudgetGuard(BudgetConfig(daily_tokens=100, hard_stop_ratio=0.1), today="2026-01-01") guard.record(1000, "test") router = build_router(providers=providers) router.guard = guard d = router.route(Scene.CRISIS) assert d.effective is Tier.HIGH assert floor_for(Scene.CRISIS) is Tier.HIGH def test_never_none_scenes_are_declared() -> None: assert Scene.DIALOGUE_MAIN in NEVER_NONE assert Scene.CRISIS in NEVER_NONE def test_estimated_tokens_downgrades_one_step(providers) -> None: """单次请求放不下时只降一档——降级要可预测。""" guard = BudgetGuard(BudgetConfig(daily_tokens=1000, hard_stop_ratio=0.95), today="2026-01-01") guard.record(900, "test") # 90%,还没到硬阈值 router = build_router(providers=providers) router.guard = guard d = router.route(Scene.DIALOGUE_MAIN, estimated_tokens=500) assert d.effective is Tier.STANDARD assert "降一档" in (d.degraded_reason or "") # ---------------------------------------------------------------- 缺 provider def test_missing_tier_falls_through(providers) -> None: """某一档没配 provider 时应该往下找,而不是崩。""" providers[Tier.HIGH] = None router = build_router(providers=providers) d = router.route(Scene.PROACTIVE_MESSAGE) assert d.requested is Tier.HIGH assert d.effective is Tier.STANDARD assert d.provider is not None assert d.degraded # ---------------------------------------------------------------- 预算记账 def test_budget_records_by_scene() -> None: guard = BudgetGuard(BudgetConfig(daily_tokens=10_000), today="2026-01-01") guard.record(100, Scene.DIALOGUE_MAIN.value) guard.record(300, Scene.MEMORY_GIST.value) guard.record(50, Scene.DIALOGUE_MAIN.value) assert guard.used == 450 top = guard.top_scenes() assert top[0][0] == Scene.MEMORY_GIST.value assert top[0][1] == 300 def test_budget_usage_dict() -> None: guard = BudgetGuard(BudgetConfig(daily_tokens=10_000), today="2026-01-01") guard.record_usage({"prompt_tokens": 100, "completion_tokens": 50}, "x") assert guard.used == 150 def test_budget_ratio_and_remaining() -> None: guard = BudgetGuard(BudgetConfig(daily_tokens=1000), today="2026-01-01") guard.record(250, "x") assert guard.ratio() == 0.25 assert guard.remaining() == 750 assert not guard.is_tight() assert not guard.is_warning() def test_budget_warning_threshold() -> None: guard = BudgetGuard(BudgetConfig(daily_tokens=1000, warn_ratio=0.7), today="2026-01-01") guard.record(750, "x") assert guard.is_warning() assert not guard.is_tight() def test_budget_summary_readable() -> None: guard = BudgetGuard(BudgetConfig(daily_tokens=1000), today="2026-01-01") guard.record(300, "dialogue_main") text = guard.summary() assert "300" in text assert "dialogue_main" in text # ---------------------------------------------------------------- 上下文预算 def test_context_budget_decreases_with_tier() -> None: assert CONTEXT_BUDGET[Tier.HIGH] > CONTEXT_BUDGET[Tier.STANDARD] assert CONTEXT_BUDGET[Tier.STANDARD] > CONTEXT_BUDGET[Tier.LOW] assert CONTEXT_BUDGET[Tier.NONE] == 0 def test_routing_decision_str(router: ModelRouter) -> None: assert "dialogue_main" in str(router.route(Scene.DIALOGUE_MAIN)) assert "tick" in str(router.route(Scene.TICK))