diff --git a/litellm/router_strategy/complexity_router/complexity_router.py b/litellm/router_strategy/complexity_router/complexity_router.py index 2f4305756e9..1aeedaa97b2 100644 --- a/litellm/router_strategy/complexity_router/complexity_router.py +++ b/litellm/router_strategy/complexity_router/complexity_router.py @@ -479,6 +479,33 @@ def _last_human_ask_index( ) +def _newest_turn_is_human_ask( + messages: Sequence[Mapping[str, object]] | None, + marker_pairs: tuple[tuple[str, str], ...] = _DEFAULT_REMINDER_MARKERS, +) -> bool: + """Whether the request's newest turn carries a real human ask, i.e. this is a new ask rather + than an agent loop's continuation traffic. + + Anchored on `_last_human_ask_index` so every surface's plumbing reads as a continuation: + chat-completions tool turns are role=tool, Messages-surface tool_result turns flatten to empty + human text, and a hybrid turn carrying an ask alongside a tool_result still counts as an ask. + Compared against the newest non-system message rather than the raw tail, because Claude Code + appends a system-role reminder after the human turn; that trailing plumbing is neither an ask + nor loop traffic and must not turn a fresh ask into a continuation. An unreadable request (no + messages) is treated as a continuation: there is no ask to classify, which is the same reading + `_extract_current_ask_and_system_prompt` gives it downstream. + """ + if not messages: + return False + newest_non_system: Final = next( + (index for index in range(len(messages) - 1, -1, -1) if messages[index].get("role") != "system"), + None, + ) + if newest_non_system is None: + return False + return _last_human_ask_index(messages, marker_pairs) == newest_non_system + + def _iter_system_scope_texts( body_system: object, messages: Sequence[Mapping[str, object]], @@ -2247,14 +2274,18 @@ class ComplexityRouter(CustomLogger): @property def _uses_tier_pin(self) -> bool: - return bool(self.config.session_affinity and not self.config.plugins) + """classification_mode 'user_turn' implies the tier pin machinery: the pin write after each + pinnable classification is what gives a continuation a held decision to replay.""" + return bool( + (self.config.session_affinity or self.config.classification_mode == "user_turn") and not self.config.plugins + ) @property def _uses_deployment_pin(self) -> bool: - """session_affinity implies the deployment pin: a session frozen onto one model + """The tier pin implies the deployment pin: a session frozen onto one model group but load-balanced across its deployments would still go cache-cold, which is the exact failure both flags exist to prevent.""" - return bool((self.config.deployment_affinity or self.config.session_affinity) and not self.config.plugins) + return bool(self.config.deployment_affinity and not self.config.plugins) or self._uses_tier_pin def _with_session_deployment_affinity( self, response: PreRoutingHookResponse | None @@ -2282,6 +2313,11 @@ class ComplexityRouter(CustomLogger): pins the model chosen on the session's first turn and reuses it for every later turn, skipping classification entirely. Otherwise delegates to `_classify_and_route`. + When `classification_mode` is 'user_turn', the same pin is replayed only on + continuation turns (an agent loop's tool traffic); a new human ask always falls + through to classification, so the session can still move tiers between asks. + With both knobs on, session_affinity's pin-first behavior wins. + Skipped entirely when `plugins` are configured: reusing a stale pin would bypass the plugin pipeline on every turn after the first, since a pinned model was never re-checked against a policy plugin whose decision can change between turns (e.g. a @@ -2305,7 +2341,13 @@ class ComplexityRouter(CustomLogger): session_id: Final = self._get_session_id_from_request_kwargs(request_kwargs) if use_session_affinity else None cache_key = self._get_session_affinity_cache_key(session_id, request_kwargs) if session_id is not None else None - if cache_key is not None: + # In 'user_turn' mode a held pin is replayed only on continuation turns; a new human + # ask falls through and re-classifies. session_affinity restores pin-first for asks too. + pin_replay_allowed: Final = bool(self.config.session_affinity) or not _newest_turn_is_human_ask( + resolved_messages, self._reminder_markers + ) + + if cache_key is not None and pin_replay_allowed: pinned_value: Final = await self.litellm_router_instance.cache.async_get_cache(key=cache_key) pinned_pin: Final = _parse_session_affinity_pin(pinned_value) if pinned_pin is not None: @@ -2354,10 +2396,11 @@ class ComplexityRouter(CustomLogger): kwargs_metadata: Final = request_kwargs.setdefault("metadata", {}) if isinstance(kwargs_metadata, dict): kwargs_metadata[ADAPTIVE_ROUTER_CHOSEN_MODEL_KEY] = routed_model + replay_cause: Final[RoutingDecisionCause] = ( + "session_affinity_pin" if self.config.session_affinity else "user_turn_continuation" + ) cause: RoutingDecisionCause = ( - "plan_mode" - if plan_floored - else ("session_affinity_escalation" if escalated else "session_affinity_pin") + "plan_mode" if plan_floored else ("session_affinity_escalation" if escalated else replay_cause) ) verbose_router_logger.info( "ComplexityRouter: routing decision cause=%s, routed_model=%s", cause, routed_model diff --git a/litellm/router_strategy/complexity_router/config.py b/litellm/router_strategy/complexity_router/config.py index 335de11e669..0abe962edf1 100644 --- a/litellm/router_strategy/complexity_router/config.py +++ b/litellm/router_strategy/complexity_router/config.py @@ -839,6 +839,21 @@ class ComplexityRouterConfig(BaseModel): description="Minimum cosine similarity for a semantic keyword match", ) + classification_mode: Literal["every_request", "user_turn"] = Field( + default="every_request", + description=( + "When to run the complexity classifier. 'every_request' (the default) classifies every " + "inference request, including the tool-result continuation turns of an agentic loop. " + "'user_turn' classifies only requests whose newest turn is a new human ask and replays " + "the session's held routing decision on continuation turns, which cuts classifier " + "spend and eliminates mid-loop model switches. Continuations with no held decision to " + "replay (no resolvable session_id, expired pin, fresh restart) still classify. Unlike " + "session_affinity, a new human ask always re-classifies, so a session can still move " + "tiers between asks. Suppressed when plugins are configured, for the same reason " + "session_affinity is: a replayed decision would bypass the plugin pipeline." + ), + ) + # Session affinity: pin the first turn's routed model for the rest of the session session_affinity: bool = Field( default=False, diff --git a/litellm/types/utils.py b/litellm/types/utils.py index 4bf8289d725..14749ecde6a 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -2842,6 +2842,11 @@ RoutingDecisionCause = Literal[ "housekeeping", "session_affinity_pin", "session_affinity_escalation", + # classification_mode 'user_turn': the request is an agent loop's continuation turn (no new + # human ask), so the session's held routing decision was replayed and the classifier was never + # called. Distinct from "session_affinity_pin", which reports the session_affinity flag pinning + # every turn including new asks; this cause only appears when session_affinity is off. + "user_turn_continuation", "default_fallback", "keyword", "quality_tier", diff --git a/tests/test_litellm/router_strategy/test_complexity_router.py b/tests/test_litellm/router_strategy/test_complexity_router.py index eee4e9aa185..97f60ec8e57 100644 --- a/tests/test_litellm/router_strategy/test_complexity_router.py +++ b/tests/test_litellm/router_strategy/test_complexity_router.py @@ -4425,6 +4425,265 @@ class _DummyPlugin: return context +class TestClassificationMode: + """Test classification_mode='user_turn': classify only requests whose newest turn is a new + human ask; tool-loop continuation turns replay the session's held routing decision.""" + + REASONING_ASK = { + "role": "user", + "content": "Let's think step by step and reason through this problem carefully.", + } + SIMPLE_ASK = {"role": "user", "content": "Hello!"} + ASSISTANT_ANSWER = {"role": "assistant", "content": "the answer"} + TOOL_CALL_1 = { + "role": "assistant", + "content": None, + "tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "read_file", "arguments": "{}"}}], + } + TOOL_RESULT_1 = {"role": "tool", "tool_call_id": "call_1", "content": "file contents"} + TOOL_CALL_2 = { + "role": "assistant", + "content": None, + "tool_calls": [{"id": "call_2", "type": "function", "function": {"name": "run_tests", "arguments": "{}"}}], + } + TOOL_RESULT_2 = {"role": "tool", "tool_call_id": "call_2", "content": "3 passed"} + + @pytest.fixture + def user_turn_config(self, basic_config) -> dict: + return {**basic_config, "classification_mode": "user_turn"} + + @staticmethod + def _request_kwargs(session_id: str) -> dict: + return {"metadata": {"session_id": session_id}} + + def _router(self, mock_router_instance, config: dict) -> ComplexityRouter: + mock_router_instance.cache = DualCache() + return ComplexityRouter( + model_name="test-router", + litellm_router_instance=mock_router_instance, + complexity_router_config=config, + ) + + def _tool_loop_turns(self) -> list[list[dict]]: + return [ + [self.REASONING_ASK], + [self.REASONING_ASK, self.TOOL_CALL_1, self.TOOL_RESULT_1], + [self.REASONING_ASK, self.TOOL_CALL_1, self.TOOL_RESULT_1, self.TOOL_CALL_2, self.TOOL_RESULT_2], + ] + + def test_default_mode_is_every_request(self, complexity_router): + assert complexity_router.config.classification_mode == "every_request" + + def test_invalid_classification_mode_rejected(self, mock_router_instance, basic_config): + with pytest.raises(ValidationError): + ComplexityRouter( + model_name="test-router", + litellm_router_instance=mock_router_instance, + complexity_router_config={**basic_config, "classification_mode": "sometimes"}, + ) + + @pytest.mark.asyncio + async def test_user_turn_mode_classifies_tool_loop_once(self, mock_router_instance, user_turn_config): + """The mutation check: a 3-request tool loop drives exactly one classification, and both + continuation turns hold the classified model under the user_turn_continuation cause.""" + router = self._router(mock_router_instance, user_turn_config) + with patch.object(router, "_classify_and_route", wraps=router._classify_and_route) as spy: + responses = [ + await router.async_pre_routing_hook( + model="test-model", request_kwargs=self._request_kwargs("loop-1"), messages=turn + ) + for turn in self._tool_loop_turns() + ] + assert spy.call_count == 1 + assert [r.model for r in responses] == ["o1-preview", "o1-preview", "o1-preview"] + assert [r.routing_decision["cause"] for r in responses[1:]] == [ + "user_turn_continuation", + "user_turn_continuation", + ] + + @pytest.mark.asyncio + async def test_every_request_default_classifies_every_tool_loop_turn(self, mock_router_instance, basic_config): + """Pins today's default: every request classifies, including tool-loop continuations.""" + router = self._router(mock_router_instance, basic_config) + with patch.object(router, "_classify_and_route", wraps=router._classify_and_route) as spy: + responses = [ + await router.async_pre_routing_hook( + model="test-model", request_kwargs=self._request_kwargs("loop-2"), messages=turn + ) + for turn in self._tool_loop_turns() + ] + assert spy.call_count == 3 + assert [r.model for r in responses] == ["o1-preview", "o1-preview", "o1-preview"] + assert all(r.routing_decision["cause"] != "user_turn_continuation" for r in responses) + + @pytest.mark.asyncio + async def test_continuation_without_session_id_still_classifies(self, mock_router_instance, user_turn_config): + """No resolvable session id means no held decision to replay, so every request classifies.""" + router = self._router(mock_router_instance, user_turn_config) + with patch.object(router, "_classify_and_route", wraps=router._classify_and_route) as spy: + responses = [ + await router.async_pre_routing_hook(model="test-model", request_kwargs={}, messages=turn) + for turn in self._tool_loop_turns() + ] + assert spy.call_count == 3 + assert [r.model for r in responses] == ["o1-preview", "o1-preview", "o1-preview"] + assert all(r.routing_decision["cause"] != "user_turn_continuation" for r in responses) + + @pytest.mark.asyncio + async def test_plugins_suppress_user_turn_gate(self, mock_router_instance, basic_config): + """A replayed decision would bypass the plugin pipeline, so plugins force every request + through _classify_and_route, exactly as they do for session_affinity.""" + router = self._router( + mock_router_instance, + {**basic_config, "classification_mode": "user_turn", "plugins": [_DummyPlugin()]}, + ) + with patch.object(router, "_classify_and_route", wraps=router._classify_and_route) as spy: + responses = [ + await router.async_pre_routing_hook( + model="test-model", request_kwargs=self._request_kwargs("loop-3"), messages=turn + ) + for turn in self._tool_loop_turns() + ] + assert spy.call_count == 3 + assert [r.model for r in responses] == ["o1-preview", "o1-preview", "o1-preview"] + assert all(r.routing_decision["cause"] != "user_turn_continuation" for r in responses) + + @pytest.mark.asyncio + async def test_new_human_ask_reclassifies_and_repins(self, mock_router_instance, user_turn_config): + """Unlike session_affinity, a new human ask never short-circuits on the pin: the session + re-classifies, moves tier, and the moved decision becomes the next held decision.""" + router = self._router(mock_router_instance, user_turn_config) + first = await router.async_pre_routing_hook( + model="test-model", request_kwargs=self._request_kwargs("s-repin"), messages=[self.REASONING_ASK] + ) + second = await router.async_pre_routing_hook( + model="test-model", + request_kwargs=self._request_kwargs("s-repin"), + messages=[self.REASONING_ASK, self.ASSISTANT_ANSWER, self.SIMPLE_ASK], + ) + third = await router.async_pre_routing_hook( + model="test-model", + request_kwargs=self._request_kwargs("s-repin"), + messages=[self.REASONING_ASK, self.ASSISTANT_ANSWER, self.SIMPLE_ASK, self.TOOL_CALL_1, self.TOOL_RESULT_1], + ) + assert first.model == "o1-preview" + assert second.model == "gpt-4o-mini" + assert third.model == "gpt-4o-mini" + assert third.routing_decision["cause"] == "user_turn_continuation" + + @pytest.mark.asyncio + async def test_new_ask_with_trailing_system_reminder_reclassifies(self, mock_router_instance, user_turn_config): + """Claude Code appends a system-role reminder after the human turn; that trailing plumbing + must not turn a new ask into a continuation, and a continuation turn carrying the same + trailing reminder stays a continuation.""" + router = self._router(mock_router_instance, user_turn_config) + reminder = {"role": "system", "content": "100 tokens left"} + first = await router.async_pre_routing_hook( + model="test-model", request_kwargs=self._request_kwargs("s-reminder"), messages=[self.REASONING_ASK] + ) + second = await router.async_pre_routing_hook( + model="test-model", + request_kwargs=self._request_kwargs("s-reminder"), + messages=[self.REASONING_ASK, self.ASSISTANT_ANSWER, self.SIMPLE_ASK, reminder], + ) + third = await router.async_pre_routing_hook( + model="test-model", + request_kwargs=self._request_kwargs("s-reminder"), + messages=[ + self.REASONING_ASK, + self.ASSISTANT_ANSWER, + self.SIMPLE_ASK, + reminder, + self.TOOL_CALL_1, + self.TOOL_RESULT_1, + reminder, + ], + ) + assert first.model == "o1-preview" + assert second.model == "gpt-4o-mini" + assert second.routing_decision["cause"] != "user_turn_continuation" + assert third.model == "gpt-4o-mini" + assert third.routing_decision["cause"] == "user_turn_continuation" + + @pytest.mark.asyncio + async def test_escalation_keyword_turn_is_a_new_ask(self, mock_router_instance, user_turn_config): + """An escalation keyword arrives as human text, so the turn classifies and escalates + instead of replaying the held decision.""" + router = self._router(mock_router_instance, user_turn_config) + first = await router.async_pre_routing_hook( + model="test-model", request_kwargs=self._request_kwargs("s-esc"), messages=[self.SIMPLE_ASK] + ) + second = await router.async_pre_routing_hook( + model="test-model", + request_kwargs=self._request_kwargs("s-esc"), + messages=[self.SIMPLE_ASK, self.ASSISTANT_ANSWER, {"role": "user", "content": "LITELLM ESCALATE"}], + ) + assert first.model == "gpt-4o-mini" + assert second.model == "gpt-4o" + assert second.routing_decision["escalated"] is True + + @pytest.mark.asyncio + async def test_messages_surface_tool_result_shapes(self, mock_router_instance, user_turn_config): + """Messages-surface shapes: a tool_result-only user turn is a continuation, while an ask + riding alongside a tool_result in the same turn is a new ask.""" + router = self._router(mock_router_instance, user_turn_config) + tool_use = {"role": "assistant", "content": [{"type": "tool_use", "id": "x", "name": "t", "input": {}}]} + tool_result = {"type": "tool_result", "tool_use_id": "x", "content": "ok"} + first = await router.async_pre_routing_hook( + model="test-model", request_kwargs=self._request_kwargs("s-msgs"), messages=[self.REASONING_ASK] + ) + pure = await router.async_pre_routing_hook( + model="test-model", + request_kwargs=self._request_kwargs("s-msgs"), + messages=[self.REASONING_ASK, tool_use, {"role": "user", "content": [tool_result]}], + ) + hybrid = await router.async_pre_routing_hook( + model="test-model", + request_kwargs=self._request_kwargs("s-msgs"), + messages=[ + self.REASONING_ASK, + tool_use, + {"role": "user", "content": [tool_result, {"type": "text", "text": "Hello!"}]}, + ], + ) + assert first.model == "o1-preview" + assert pure.model == "o1-preview" + assert pure.routing_decision["cause"] == "user_turn_continuation" + assert hybrid.model == "gpt-4o-mini" + + @pytest.mark.asyncio + async def test_session_affinity_wins_when_both_knobs_are_on(self, mock_router_instance, user_turn_config): + """With session_affinity also on, the pin short-circuits new asks too and keeps its own + cause, so the session stays on turn 1's model.""" + router = self._router(mock_router_instance, {**user_turn_config, "session_affinity": True}) + first = await router.async_pre_routing_hook( + model="test-model", request_kwargs=self._request_kwargs("s-both"), messages=[self.REASONING_ASK] + ) + second = await router.async_pre_routing_hook( + model="test-model", + request_kwargs=self._request_kwargs("s-both"), + messages=[self.REASONING_ASK, self.ASSISTANT_ANSWER, self.SIMPLE_ASK], + ) + assert first.model == "o1-preview" + assert second.model == "o1-preview" + assert second.routing_decision["cause"] == "session_affinity_pin" + + def test_user_turn_mode_enables_tier_and_deployment_pins(self, mock_router_instance, basic_config): + """user_turn implies the tier pin machinery (the pin write is what gives a continuation + a held decision) and the tier pin implies the deployment pin; plugins suppress both.""" + default = self._router(mock_router_instance, basic_config) + enabled = self._router(mock_router_instance, {**basic_config, "classification_mode": "user_turn"}) + suppressed = self._router( + mock_router_instance, + {**basic_config, "classification_mode": "user_turn", "plugins": [_DummyPlugin()]}, + ) + assert default._uses_tier_pin is False + assert enabled._uses_tier_pin is True + assert enabled._uses_deployment_pin is True + assert suppressed._uses_tier_pin is False + assert suppressed._uses_deployment_pin is False + + class TestRoutingPlugins: """Test the `complexity_router_config.plugins` field: narrows the classified tier's candidate pool before a model is picked. Discussion: diff --git a/ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/RoutingDecisionCard.tsx b/ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/RoutingDecisionCard.tsx index 8c77b2db630..d2aa20901f5 100644 --- a/ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/RoutingDecisionCard.tsx +++ b/ui/litellm-dashboard/src/components/view_logs/LogDetailsDrawer/RoutingDecisionCard.tsx @@ -89,6 +89,7 @@ const CONSTANT_CAUSE_LABELS: Record = { semantic_keyword_match: "Semantic keyword match", session_affinity_pin: "Pinned to session", session_affinity_escalation: "Escalated from session pin", + user_turn_continuation: "Continuation turn, classifier skipped", quality_tier: "Quality tier mapping", bandit: "Adaptive bandit", default_fallback: "Default model, no route matched", diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index cc56483ac85..140130adc4b 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -34318,6 +34318,13 @@ export interface components { adaptive_eligible: "all" | "classified_tier"; /** @description Quality vs cost weights for adaptive selection (used when adaptive=True) */ adaptive_weights?: components["schemas"]["AdaptiveRouterWeights"]; + /** + * Classification Mode + * @description When to run the complexity classifier. 'every_request' (the default) classifies every inference request, including the tool-result continuation turns of an agentic loop. 'user_turn' classifies only requests whose newest turn is a new human ask and replays the session's held routing decision on continuation turns, which cuts classifier spend and eliminates mid-loop model switches. Continuations with no held decision to replay (no resolvable session_id, expired pin, fresh restart) still classify. Unlike session_affinity, a new human ask always re-classifies, so a session can still move tiers between asks. Suppressed when plugins are configured, for the same reason session_affinity is: a replayed decision would bypass the plugin pipeline. + * @default every_request + * @enum {string} + */ + classification_mode: "every_request" | "user_turn"; /** * Classification Prompt * @description Replaces the opening instructions of the LLM classifier rubric (the judging-criteria prose) for a custom tier set. The per-tier bullets and the trust-boundary paragraph telling the classifier to ignore tier requests embedded in quoted caller text are always appended after it and cannot be overridden. Requires tier_definitions; a built-in-tier router customizes its prompt via classifier_llm_config.system_prompt or classification_rubric instead. @@ -35590,7 +35597,7 @@ export interface components { * Cause * @enum {string} */ - cause?: "heuristic_scorer" | "reasoning_override" | "llm_classifier" | "heuristic_first_short_circuit" | "classifier_plugin" | "classifier_fallback" | "default_model_fallback" | "literal_keyword_match" | "semantic_keyword_match" | "plan_mode" | "housekeeping" | "session_affinity_pin" | "session_affinity_escalation" | "default_fallback" | "keyword" | "quality_tier" | "bandit"; + cause?: "heuristic_scorer" | "reasoning_override" | "llm_classifier" | "heuristic_first_short_circuit" | "classifier_plugin" | "classifier_fallback" | "default_model_fallback" | "literal_keyword_match" | "semantic_keyword_match" | "plan_mode" | "housekeeping" | "session_affinity_pin" | "session_affinity_escalation" | "user_turn_continuation" | "default_fallback" | "keyword" | "quality_tier" | "bandit"; /** Classifier Cost */ classifier_cost?: number; /** Classifier Model */