From c82c9cbcedf43661521270740273d6437d76c098 Mon Sep 17 00:00:00 2001 From: tin-berri Date: Wed, 9 Sep 2026 11:04:11 -0700 Subject: [PATCH] fix(router): strip encrypted reasoning on an auto-router tier change instead of a 503 (#40280) A Responses API follow-up that replays reasoning.encrypted_content is pinned to the deployment that minted it. Behind an auto-router the pre-routing hook rebinds the model to the tier it picked before the candidate pool is built, so a turn that classifies into a different tier never finds the origin and the affinity check raised its fail-fast 503, whose text claims a cooldown that does not exist When the deployment that minted the reasoning is not a member of the model group this turn is routed to, strip the encrypted reasoning (keeping any readable summary, string or block form) and dispatch to the routed group. Membership is tested by deployment id against the candidate set the router itself resolved for the route (routing group, model_name, team, and pattern alike), not by model-group name, so an alias, a provider-qualified spelling, a team-public name, or a pattern route of the same group is not misread as a tier change. An unknown origin (a removed deployment, or a forged/unauthenticated marker) is handled the same as a cross-group one and its reasoning is stripped, so a real cross-group id and a nonexistent id return the same response and cannot be used to enumerate deployment ids. Unavailability within the origin's own group keeps the existing 429/503 fail-fast, so the cooldown contract is unchanged Resolves LIT-7195 Claude-Session: https://claude.ai/code/session_01KAumQbhzk6jdWWHFLA8Jar Co-authored-by: Claude Opus 4.8 --- litellm/responses/utils.py | 43 ++ litellm/router.py | 34 ++ .../encrypted_content_affinity_check.py | 73 ++- .../test_encrypted_content_affinity_check.py | 561 ++++++++++++------ tests/test_litellm/test_router.py | 42 ++ 5 files changed, 557 insertions(+), 196 deletions(-) diff --git a/litellm/responses/utils.py b/litellm/responses/utils.py index 540d492beec..599e978df6a 100644 --- a/litellm/responses/utils.py +++ b/litellm/responses/utils.py @@ -543,6 +543,49 @@ class ResponsesAPIRequestUtils: return request_input + @staticmethod + def strip_encrypted_reasoning_from_input(request_input: object) -> None: + """Drop reasoning items the routed deployment cannot decrypt, keeping their readable summary. + + Mutates ``request_input`` in place: the router's fallback snapshot shares this + list object, so a rebound list would replay the stripped items on the fallback hop. + """ + if not isinstance(request_input, list): + return + items: Final = cast(list[object], request_input) # cast-ok: untyped client json + stripped: Final = tuple(ResponsesAPIRequestUtils._without_encrypted_reasoning(item) for item in items) + items[:] = (item for item in stripped if item is not None) # rebind-ok: list shared with fallback snapshot + + @staticmethod + def _without_encrypted_reasoning(item: object) -> object | None: + if not isinstance(item, dict): + return item + reasoning: Final = cast(Mapping[str, object], item) # cast-ok: untyped client json + if reasoning.get("type") != "reasoning" or not reasoning.get("encrypted_content"): + return reasoning + readable: Final = any( + ResponsesAPIRequestUtils._has_readable_text(reasoning.get(key)) for key in ("summary", "content") + ) + if not readable: + return None + kept: Final[dict[str, object]] = { # mutable-ok: request item rebuilt without the undecryptable keys + key: value for key, value in reasoning.items() if key not in ("encrypted_content", "id") + } + return kept + + @staticmethod + def _has_readable_text(value: object) -> bool: + """A reasoning item's ``summary``/``content`` carries readable text: a non-empty string, or a + list holding at least one block with a non-empty ``text`` field (summary_text / output_text).""" + if isinstance(value, str): + return bool(value.strip()) + if isinstance(value, list): + return any( + isinstance(block, dict) and bool(cast(Mapping[str, object], block).get("text")) # cast-ok: untyped json + for block in value + ) + return False + @staticmethod def _build_responses_api_response_id( custom_llm_provider: str | None, diff --git a/litellm/router.py b/litellm/router.py index 2fbdf541487..68ace283949 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -11167,6 +11167,40 @@ class Router: return ids + def get_candidate_model_ids_for_route(self, model: str, team_id: str | None = None) -> frozenset[str]: + """ + Deployment ids that could serve ``model`` for ``team_id``, unioned across the paths + the router resolves a route through: ``model_group_alias``, a routing group, the + ``model_name`` and team indexes, and wildcard pattern routes. Read-only and + side-effect-free, unlike ``_common_checks_available_deployment`` which also applies + fallbacks and can raise. Lets a pre-call check tell a genuine cross-group route from + same-group unavailability without re-deriving that precedence at the call site, and + without leaking deployment ids into request kwargs bound for the provider. + """ + resolved: Final = self._get_model_from_alias(model=model) or model + routing_group_members: Final = self._get_routing_group_deployments(model=resolved, team_id=team_id) + if routing_group_members is not None: + return self._deployment_ids(routing_group_members) + if resolved in self.model_names: + return self._deployment_ids(self._get_all_deployments(model_name=resolved, team_id=team_id)) + team_router: Final = self.team_pattern_routers.get(team_id) if team_id is not None else None + return self._deployment_ids( + ( + *self._get_all_deployments(model_name=resolved, team_id=team_id), + *(self.pattern_router.route(resolved) or ()), + *((team_router.route(resolved) or ()) if team_router is not None else ()), + ) + ) + + @staticmethod + def _deployment_ids(deployments: Sequence[Mapping[str, object]]) -> frozenset[str]: + return frozenset( + str(model_info["id"]) + for deployment in deployments + for model_info in (deployment.get("model_info"),) + if isinstance(model_info, Mapping) and model_info.get("id") is not None + ) + def has_model_id(self, candidate_id: str) -> bool: """ O(1) membership check for a deployment ID without allocating large lists. diff --git a/litellm/router_utils/pre_call_checks/encrypted_content_affinity_check.py b/litellm/router_utils/pre_call_checks/encrypted_content_affinity_check.py index b623e31ce06..d10153881d9 100644 --- a/litellm/router_utils/pre_call_checks/encrypted_content_affinity_check.py +++ b/litellm/router_utils/pre_call_checks/encrypted_content_affinity_check.py @@ -37,13 +37,13 @@ Safe to enable globally: """ import time +from collections.abc import Mapping from typing import TYPE_CHECKING, Final, Optional, Protocol, cast import httpx from litellm._logging import verbose_router_logger from litellm.exceptions import ( - BadRequestError, RateLimitError, ServiceUnavailableError, ) @@ -158,6 +158,23 @@ class EncryptedContentAffinityCheck(CustomLogger): return deployment return None + @staticmethod + def _request_team_id(request_kwargs: Mapping[str, object]) -> str | None: + containers: Final = (request_kwargs.get("metadata"), request_kwargs.get("litellm_metadata")) + team_ids: Final = (c.get("user_api_key_team_id") for c in containers if isinstance(c, Mapping)) + return next((tid for tid in team_ids if isinstance(tid, str)), None) + + def _routed_group_candidate_model_ids(self, request_kwargs: Mapping[str, object], model: str) -> frozenset[str]: + """ + Deployment ids that could serve this turn's routed ``model``, as the router + resolves a route (model_group_alias / routing group / model_name / team / + pattern). Delegates to the router so the full precedence is not re-derived here + and no deployment ids are written into request kwargs bound for the provider. + """ + if self.router is None: + return frozenset() + return self.router.get_candidate_model_ids_for_route(model=model, team_id=self._request_team_id(request_kwargs)) + @staticmethod def _encryption_boundary_key( litellm_params: object, @@ -225,10 +242,14 @@ class EncryptedContentAffinityCheck(CustomLogger): """ If the request ``input`` contains litellm-encoded item IDs, decode the embedded ``model_id`` and pin the request to that deployment. Raises - ``RateLimitError`` / ``ServiceUnavailableError`` / ``BadRequestError`` - when the originating deployment is unavailable and no encryption-boundary - peer exists, rather than dispatching a doomed request to a non-peer - deployment. The 429/503 split mirrors the originating cooldown's status: + ``RateLimitError`` / ``ServiceUnavailableError`` when the originating + deployment is a member of the routed model group but currently unavailable + and no encryption-boundary peer exists, rather than dispatching a doomed + request to a non-peer deployment. When the origin is not a member of the + routed group (an auto-router tier change, a model switch with no peer, a + removed deployment, or an unknown/forged marker), the encrypted reasoning is + stripped and the request dispatches with its readable history instead. The + 429/503 split mirrors the originating cooldown's status: a 429-induced cooldown surfaces as 429 (with ``Retry-After`` set to the remaining cooldown window) so OpenAI-compatible clients back off and retry after the deployment is eligible again. @@ -285,12 +306,34 @@ class EncryptedContentAffinityCheck(CustomLogger): request_kwargs["_encrypted_content_affinity_pinned"] = True return boundary_matches - # Dispatching to a non-peer would guarantee an upstream - # `invalid_encrypted_content` 400, so fail fast with a clearer error. + # The origin cannot serve this turn's routed group and no peer shares the boundary, so its + # encrypted reasoning can never decrypt here. Strip it, keep the readable history, and dispatch + # to the routed group instead of failing. Membership is tested by deployment id against the set + # the router actually resolved for this route, not by model-group name, so an alias, a + # provider-qualified spelling, a team-public name, or a pattern route of the same group is not + # mistaken for a tier change. An unknown origin (a removed deployment, or a forged marker) is + # treated the same as a cross-group one, which also denies an authenticated caller a + # deployment-id existence oracle: a real cross-group id and a nonexistent id both strip and + # dispatch rather than returning distinguishable responses. Only a genuine same-group member + # that is currently unavailable falls through to the fail-fast, preserving the cooldown contract. + routed_group_model_ids: Final = ( + self._routed_group_candidate_model_ids(request_kwargs, model) if originating is not None else frozenset() + ) + if str(model_id) not in routed_group_model_ids: + verbose_router_logger.debug( + "EncryptedContentAffinityCheck: model_id=%s is not a candidate for the routed group %s; " + "forwarding without its encrypted reasoning", + model_id, + model, + ) + ResponsesAPIRequestUtils.strip_encrypted_reasoning_from_input(request_input) + return typed_healthy_deployments + + # The origin is a member of the routed group but currently unavailable (cooled down); fail fast + # rather than dispatching to a non-peer, which would guarantee an upstream 400. raise await self._unavailable_origin_error( model=model, model_id=model_id, - originating=originating, parent_otel_span=parent_otel_span, ) @@ -298,25 +341,11 @@ class EncryptedContentAffinityCheck(CustomLogger): self, model: str, model_id: str, - originating: Deployment | None, parent_otel_span: Span | None, ) -> Exception: # Public error messages intentionally omit the originating ``model_id`` so # an authenticated caller forging encrypted-content markers cannot use the # error surface to enumerate which deployment IDs exist on this router. - if originating is None: - return BadRequestError( - message=( - "The deployment that produced this encrypted_content is no " - "longer configured on this router, and no deployment on the " - "same encryption boundary is available. Re-issue the request " - "without the stale encrypted_content items, or restore the " - "originating deployment." - ), - model=model, - llm_provider="", - ) - cooldown: Final = await self._get_origin_cooldown(model_id=model_id, parent_otel_span=parent_otel_span) if cooldown is not None and str(cooldown.get("status_code")) == "429": diff --git a/tests/test_litellm/router_utils/pre_call_checks/test_encrypted_content_affinity_check.py b/tests/test_litellm/router_utils/pre_call_checks/test_encrypted_content_affinity_check.py index dac991a41c4..3961c8d74c2 100644 --- a/tests/test_litellm/router_utils/pre_call_checks/test_encrypted_content_affinity_check.py +++ b/tests/test_litellm/router_utils/pre_call_checks/test_encrypted_content_affinity_check.py @@ -16,12 +16,10 @@ The mechanism works without any cache and supports two encoding strategies: """ import time -from typing import List, Optional from unittest.mock import AsyncMock, patch import pytest - import litellm from litellm.responses.utils import ResponsesAPIRequestUtils from litellm.types.llms.openai import ResponsesAPIResponse @@ -68,9 +66,7 @@ class TestEncryptedItemIdCodec: def test_roundtrip(self): model_id = "deployment-1" original_item_id = "rs_abc123def456" - encoded = ResponsesAPIRequestUtils._build_encrypted_item_id( - model_id, original_item_id - ) + encoded = ResponsesAPIRequestUtils._build_encrypted_item_id(model_id, original_item_id) assert encoded.startswith("encitem_") decoded = ResponsesAPIRequestUtils._decode_encrypted_item_id(encoded) assert decoded is not None @@ -81,9 +77,7 @@ class TestEncryptedItemIdCodec: """Decoding must succeed even if base64 padding (=) was stripped in transit.""" model_id = "gpt-5.1-codex-openai-2" original_item_id = "rs_0efb96cb222403210069a01d5d52588196a9dc394ffdb89d00" - encoded = ResponsesAPIRequestUtils._build_encrypted_item_id( - model_id, original_item_id - ) + encoded = ResponsesAPIRequestUtils._build_encrypted_item_id(model_id, original_item_id) # Strip any trailing '=' to simulate what happens in transit stripped = encoded.rstrip("=") decoded = ResponsesAPIRequestUtils._decode_encrypted_item_id(stripped) @@ -100,9 +94,7 @@ class TestEncryptedItemIdCodec: """item_id values containing ';' must survive the roundtrip.""" model_id = "deployment-1" original_item_id = "rs_part1;part2;part3" - encoded = ResponsesAPIRequestUtils._build_encrypted_item_id( - model_id, original_item_id - ) + encoded = ResponsesAPIRequestUtils._build_encrypted_item_id(model_id, original_item_id) decoded = ResponsesAPIRequestUtils._decode_encrypted_item_id(encoded) assert decoded is not None assert decoded["item_id"] == original_item_id @@ -118,11 +110,7 @@ class TestUpdateEncryptedContentItemIds: {"id": "rs_xyz", "type": "reasoning", "encrypted_content": "secret"}, ], } - result = ( - ResponsesAPIRequestUtils._update_encrypted_content_item_ids_in_response( - response, model_id - ) - ) + result = ResponsesAPIRequestUtils._update_encrypted_content_item_ids_in_response(response, model_id) # Plain message item untouched assert result["output"][0]["id"] == "msg_abc" # Reasoning item with encrypted_content gets encoded @@ -133,16 +121,8 @@ class TestUpdateEncryptedContentItemIds: assert decoded["item_id"] == "rs_xyz" def test_no_op_when_model_id_is_none(self): - response = { - "output": [ - {"id": "rs_xyz", "type": "reasoning", "encrypted_content": "secret"} - ] - } - result = ( - ResponsesAPIRequestUtils._update_encrypted_content_item_ids_in_response( - response, None - ) - ) + response = {"output": [{"id": "rs_xyz", "type": "reasoning", "encrypted_content": "secret"}]} + result = ResponsesAPIRequestUtils._update_encrypted_content_item_ids_in_response(response, None) assert result["output"][0]["id"] == "rs_xyz" @@ -151,9 +131,7 @@ class TestEncryptedContentWrapping: """Test wrapping encrypted_content with model_id metadata.""" model_id = "deployment-1" original_content = "gAAAAABpnW_yEYmSNEyOG_original_encrypted_data" - wrapped = ResponsesAPIRequestUtils._wrap_encrypted_content_with_model_id( - original_content, model_id - ) + wrapped = ResponsesAPIRequestUtils._wrap_encrypted_content_with_model_id(original_content, model_id) assert wrapped.startswith("litellm_enc:") assert wrapped != original_content @@ -170,9 +148,7 @@ class TestEncryptedContentWrapping: ( model_id, content, - ) = ResponsesAPIRequestUtils._unwrap_encrypted_content_with_model_id( - plain_content - ) + ) = ResponsesAPIRequestUtils._unwrap_encrypted_content_with_model_id(plain_content) assert model_id is None assert content == plain_content @@ -189,11 +165,7 @@ class TestEncryptedContentWrapping: }, ], } - result = ( - ResponsesAPIRequestUtils._update_encrypted_content_item_ids_in_response( - response, model_id - ) - ) + result = ResponsesAPIRequestUtils._update_encrypted_content_item_ids_in_response(response, model_id) assert result["output"][0].get("encrypted_content") is None wrapped = result["output"][1]["encrypted_content"] assert wrapped.startswith("litellm_enc:") @@ -210,19 +182,13 @@ class TestRestoreEncryptedContentItemIds: def test_restores_encoded_ids(self): model_id = "deployment-1" original_id = "rs_encrypted_item_456" - encoded_id = ResponsesAPIRequestUtils._build_encrypted_item_id( - model_id, original_id - ) + encoded_id = ResponsesAPIRequestUtils._build_encrypted_item_id(model_id, original_id) request_input = [ {"type": "message", "id": "msg_abc123", "role": "assistant"}, {"type": "reasoning", "id": encoded_id, "encrypted_content": "secret"}, ] - restored = ( - ResponsesAPIRequestUtils._restore_encrypted_content_item_ids_in_input( - request_input - ) - ) + restored = ResponsesAPIRequestUtils._restore_encrypted_content_item_ids_in_input(request_input) assert restored[0]["id"] == "msg_abc123" assert restored[1]["id"] == original_id @@ -230,33 +196,21 @@ class TestRestoreEncryptedContentItemIds: """Test that wrapped encrypted_content is unwrapped before forwarding.""" model_id = "deployment-1" original_content = "gAAAAABpnW_yEYmSNEyOG_original" - wrapped_content = ( - ResponsesAPIRequestUtils._wrap_encrypted_content_with_model_id( - original_content, model_id - ) - ) + wrapped_content = ResponsesAPIRequestUtils._wrap_encrypted_content_with_model_id(original_content, model_id) request_input = [ {"type": "reasoning", "encrypted_content": wrapped_content}, ] - restored = ( - ResponsesAPIRequestUtils._restore_encrypted_content_item_ids_in_input( - request_input - ) - ) + restored = ResponsesAPIRequestUtils._restore_encrypted_content_item_ids_in_input(request_input) assert restored[0]["encrypted_content"] == original_content def test_no_op_for_plain_string_input(self): - result = ResponsesAPIRequestUtils._restore_encrypted_content_item_ids_in_input( - "Hello world" - ) + result = ResponsesAPIRequestUtils._restore_encrypted_content_item_ids_in_input("Hello world") assert result == "Hello world" def test_no_op_for_unencoded_ids(self): request_input = [{"type": "message", "id": "msg_plain"}] - result = ResponsesAPIRequestUtils._restore_encrypted_content_item_ids_in_input( - request_input - ) + result = ResponsesAPIRequestUtils._restore_encrypted_content_item_ids_in_input(request_input) assert result[0]["id"] == "msg_plain" @@ -283,9 +237,7 @@ async def test_encrypted_content_affinity_tracks_and_routes(): "id": "msg_abc123", "status": "completed", "role": "assistant", - "content": [ - {"type": "output_text", "text": "Hello!", "annotations": []} - ], + "content": [{"type": "output_text", "text": "Hello!", "annotations": []}], }, { "type": "reasoning", @@ -347,9 +299,9 @@ async def test_encrypted_content_affinity_tracks_and_routes(): # The response must have rewritten the encrypted item's ID to encoded form encoded_item_id = _extract_encoded_item_id(first_response) - assert encoded_item_id.startswith( - "encitem_" - ), f"Expected output item ID to be rewritten to encitem_... but got {encoded_item_id!r}" + assert encoded_item_id.startswith("encitem_"), ( + f"Expected output item ID to be rewritten to encitem_... but got {encoded_item_id!r}" + ) # Verify the encoded ID decodes back to the correct deployment + original ID decoded = ResponsesAPIRequestUtils._decode_encrypted_item_id(encoded_item_id) @@ -371,9 +323,9 @@ async def test_encrypted_content_affinity_tracks_and_routes(): ) second_model_id = second_response._hidden_params["model_id"] - assert ( - second_model_id == first_model_id - ), f"Expected affinity to route to {first_model_id}, but got {second_model_id}" + assert second_model_id == first_model_id, ( + f"Expected affinity to route to {first_model_id}, but got {second_model_id}" + ) @pytest.mark.asyncio @@ -478,9 +430,7 @@ async def test_encrypted_content_affinity_bypasses_rpm_limits(): # Extract encoded item ID from the first response output encoded_item_id = _extract_encoded_item_id(first_response) - assert encoded_item_id.startswith( - "encitem_" - ), f"Expected encitem_... but got {encoded_item_id!r}" + assert encoded_item_id.startswith("encitem_"), f"Expected encitem_... but got {encoded_item_id!r}" # Follow-up with the encoded item ID — should pin to same deployment second_response = await router.aresponses( @@ -628,17 +578,13 @@ async def test_encrypted_content_affinity_with_wrapped_content_no_id(): if hasattr(first_item, "encrypted_content") else first_item.get("encrypted_content") ) - assert wrapped_content.startswith( - "litellm_enc:" - ), f"Expected wrapped content but got {wrapped_content[:50]}..." + assert wrapped_content.startswith("litellm_enc:"), f"Expected wrapped content but got {wrapped_content[:50]}..." # Verify we can extract model_id from wrapped content ( extracted_model_id, _, - ) = ResponsesAPIRequestUtils._unwrap_encrypted_content_with_model_id( - wrapped_content - ) + ) = ResponsesAPIRequestUtils._unwrap_encrypted_content_with_model_id(wrapped_content) assert extracted_model_id == first_model_id # Second request: use wrapped encrypted_content WITHOUT an ID (Codex behavior) @@ -653,9 +599,9 @@ async def test_encrypted_content_affinity_with_wrapped_content_no_id(): ) second_model_id = second_response._hidden_params["model_id"] - assert ( - second_model_id == first_model_id - ), f"Expected affinity to route to {first_model_id}, but got {second_model_id}" + assert second_model_id == first_model_id, ( + f"Expected affinity to route to {first_model_id}, but got {second_model_id}" + ) def test_encrypted_content_wrapping_preserves_original_content(): @@ -664,13 +610,9 @@ def test_encrypted_content_wrapping_preserves_original_content(): This is critical for streaming responses where content must round-trip correctly. """ model_id = "test-deployment-1" - original_encrypted_content = ( - "gAAAAABpnW_yEYmSNEyOG_streaming_test_content_with_special_chars==+/" - ) + original_encrypted_content = "gAAAAABpnW_yEYmSNEyOG_streaming_test_content_with_special_chars==+/" - wrapped = ResponsesAPIRequestUtils._wrap_encrypted_content_with_model_id( - original_encrypted_content, model_id - ) + wrapped = ResponsesAPIRequestUtils._wrap_encrypted_content_with_model_id(original_encrypted_content, model_id) assert wrapped.startswith("litellm_enc:") assert wrapped != original_encrypted_content @@ -691,9 +633,7 @@ def test_encrypted_content_wrapping_with_multiple_semicolons(): model_id = "deployment-with-semicolons" original_content = "gAAAAAB;some;content;with;semicolons" - wrapped = ResponsesAPIRequestUtils._wrap_encrypted_content_with_model_id( - original_content, model_id - ) + wrapped = ResponsesAPIRequestUtils._wrap_encrypted_content_with_model_id(original_content, model_id) ( extracted_model_id, @@ -764,9 +704,7 @@ async def test_encrypted_content_affinity_preserves_litellm_metadata_for_respons request_kwargs=request_kwargs, ) - assert ( - request_kwargs["litellm_metadata"]["encrypted_content_affinity_enabled"] is True - ) + assert request_kwargs["litellm_metadata"]["encrypted_content_affinity_enabled"] is True assert request_kwargs["litellm_metadata"]["model_info"] == {"id": "dep-1"} @@ -777,9 +715,7 @@ def test_encrypted_content_wrapping_empty_string(): model_id = "test-deployment" original_content = "" - wrapped = ResponsesAPIRequestUtils._wrap_encrypted_content_with_model_id( - original_content, model_id - ) + wrapped = ResponsesAPIRequestUtils._wrap_encrypted_content_with_model_id(original_content, model_id) assert wrapped.startswith("litellm_enc:") @@ -1132,9 +1068,7 @@ def test_boundary_key_accepts_pydantic_litellm_params_instance(): "api_key": "fake-azure-resource-key-a", } - pydantic_key = EncryptedContentAffinityCheck._encryption_boundary_key( - pydantic_params - ) + pydantic_key = EncryptedContentAffinityCheck._encryption_boundary_key(pydantic_params) plain_key = EncryptedContentAffinityCheck._encryption_boundary_key(plain_params) assert pydantic_key is not None @@ -1161,18 +1095,8 @@ def test_boundary_key_rejects_non_dict_like_inputs(): for bad in (None, [], "not a dict", 42, object()): assert EncryptedContentAffinityCheck._encryption_boundary_key(bad) is None - assert ( - EncryptedContentAffinityCheck._encryption_boundary_key( - {"api_base": "", "api_key": "k"} - ) - is None - ) - assert ( - EncryptedContentAffinityCheck._encryption_boundary_key( - {"api_base": "https://x"} - ) - is None - ) + assert EncryptedContentAffinityCheck._encryption_boundary_key({"api_base": "", "api_key": "k"}) is None + assert EncryptedContentAffinityCheck._encryption_boundary_key({"api_base": "https://x"}) is None # --------------------------------------------------------------------------- @@ -1180,10 +1104,11 @@ def test_boundary_key_rejects_non_dict_like_inputs(): # --------------------------------------------------------------------------- -def _make_originating_mock(api_base: str, api_key: str): +def _make_originating_mock(api_base: str, api_key: str, model_name: str = "gpt-5.4"): from unittest.mock import MagicMock originating = MagicMock() + originating.model_name = model_name originating.litellm_params.model_dump.return_value = { "api_base": api_base, "api_key": api_key, @@ -1192,19 +1117,23 @@ def _make_originating_mock(api_base: str, api_key: str): def _make_router_mock_with_cooldown( - originating, cooldown_entries: Optional[List[tuple]] = None + originating, + cooldown_entries: list[tuple] | None = None, + routed_group_model_ids: list[str] | None = None, ): """ Build a MagicMock router whose ``cooldown_cache.async_get_active_cooldowns`` - returns ``cooldown_entries`` (defaulting to ``[]`` — no active cooldown). + returns ``cooldown_entries`` (defaulting to ``[]`` — no active cooldown), and + whose ``get_candidate_model_ids_for_route`` returns ``routed_group_model_ids`` + (the deployment ids the router resolves for the routed model; defaulting to ``[]`` + — origin absent from the routed group, i.e. a tier change). """ from unittest.mock import AsyncMock, MagicMock mock_router = MagicMock() mock_router.get_deployment.return_value = originating - mock_router.cooldown_cache.async_get_active_cooldowns = AsyncMock( - return_value=list(cooldown_entries or []) - ) + mock_router.cooldown_cache.async_get_active_cooldowns = AsyncMock(return_value=list(cooldown_entries or [])) + mock_router.get_candidate_model_ids_for_route.return_value = frozenset(routed_group_model_ids or []) return mock_router @@ -1235,15 +1164,15 @@ async def test_affinity_raises_service_unavailable_when_origin_cooled_for_non_42 }, ) ], + routed_group_model_ids=["deployment-a-cooled", "deployment-b"], ) check = EncryptedContentAffinityCheck(router=mock_router) - encoded_id = ResponsesAPIRequestUtils._build_encrypted_item_id( - "deployment-a-cooled", "rs_test" - ) + encoded_id = ResponsesAPIRequestUtils._build_encrypted_item_id("deployment-a-cooled", "rs_test") healthy_only_b = [ { "model_info": {"id": "deployment-b"}, + "model_name": "gpt-5.4", "litellm_params": { "api_base": "https://account-b.openai.azure.com/", "api_key": "key-b", @@ -1297,15 +1226,15 @@ async def test_affinity_raises_rate_limit_with_retry_after_when_origin_cooled_fo }, ) ], + routed_group_model_ids=["deployment-a-cooled-429", "deployment-b"], ) check = EncryptedContentAffinityCheck(router=mock_router) - encoded_id = ResponsesAPIRequestUtils._build_encrypted_item_id( - "deployment-a-cooled-429", "rs_test" - ) + encoded_id = ResponsesAPIRequestUtils._build_encrypted_item_id("deployment-a-cooled-429", "rs_test") healthy_only_b = [ { "model_info": {"id": "deployment-b"}, + "model_name": "gpt-5.4", "litellm_params": { "api_base": "https://account-b.openai.azure.com/", "api_key": "key-b", @@ -1345,15 +1274,16 @@ async def test_affinity_raises_service_unavailable_when_origin_filtered_without_ ) originating = _make_originating_mock("https://account-a.openai.azure.com/", "key-a") - mock_router = _make_router_mock_with_cooldown(originating, cooldown_entries=[]) + mock_router = _make_router_mock_with_cooldown( + originating, cooldown_entries=[], routed_group_model_ids=["deployment-a-filtered", "deployment-b"] + ) check = EncryptedContentAffinityCheck(router=mock_router) - encoded_id = ResponsesAPIRequestUtils._build_encrypted_item_id( - "deployment-a-filtered", "rs_test" - ) + encoded_id = ResponsesAPIRequestUtils._build_encrypted_item_id("deployment-a-filtered", "rs_test") healthy_only_b = [ { "model_info": {"id": "deployment-b"}, + "model_name": "gpt-5.4", "litellm_params": { "api_base": "https://account-b.openai.azure.com/", "api_key": "key-b", @@ -1377,15 +1307,18 @@ async def test_affinity_raises_service_unavailable_when_origin_filtered_without_ @pytest.mark.asyncio -async def test_affinity_raises_bad_request_when_origin_removed(): +async def test_affinity_strips_and_dispatches_when_origin_is_unknown_or_removed(): """ - Originating deployment was removed from the router config and no boundary - peer is available. This is permanent (the stale encrypted_content cannot - be honored), so surface a 400 with actionable text. + A removed deployment, or a forged/unknown affinity marker, resolves to no + originating deployment. It is handled like a cross-group origin: the encrypted + reasoning is stripped and the request dispatches with its readable history, + rather than returning a distinguishable error. That uniform handling denies an + authenticated caller a deployment-id existence oracle, an existing cross-group id + and a nonexistent id both strip and proceed, so responses cannot be told apart. + The membership lookup is skipped entirely when the origin is unknown. """ from unittest.mock import MagicMock - from litellm.exceptions import BadRequestError from litellm.router_utils.pre_call_checks.encrypted_content_affinity_check import ( EncryptedContentAffinityCheck, ) @@ -1394,12 +1327,11 @@ async def test_affinity_raises_bad_request_when_origin_removed(): mock_router.get_deployment.return_value = None check = EncryptedContentAffinityCheck(router=mock_router) - encoded_id = ResponsesAPIRequestUtils._build_encrypted_item_id( - "deployment-removed", "rs_test" - ) - healthy_only_b = [ + wrapped = ResponsesAPIRequestUtils._wrap_encrypted_content_with_model_id("gAAAAA-blob", "deployment-removed") + routed_pool = [ { "model_info": {"id": "deployment-b"}, + "model_name": "gpt-5.4", "litellm_params": { "api_base": "https://account-b.openai.azure.com/", "api_key": "key-b", @@ -1408,18 +1340,28 @@ async def test_affinity_raises_bad_request_when_origin_removed(): } ] request_kwargs = { - "input": [{"id": encoded_id, "type": "reasoning"}], + "litellm_metadata": {}, + "input": [ + {"role": "user", "content": "why is the sky blue?"}, + { + "type": "reasoning", + "encrypted_content": wrapped, + "summary": [{"type": "summary_text", "text": "scattering"}], + }, + {"role": "user", "content": "and sunsets?"}, + ], } - with pytest.raises(BadRequestError) as excinfo: - await check.async_filter_deployments( - model="gpt-5.4", - healthy_deployments=healthy_only_b, - messages=None, - request_kwargs=request_kwargs, - ) + result = await check.async_filter_deployments( + model="gpt-5.4", + healthy_deployments=routed_pool, + messages=None, + request_kwargs=request_kwargs, + ) - assert "deployment-removed" not in str(excinfo.value) + assert result is routed_pool + assert not any(isinstance(item, dict) and item.get("encrypted_content") for item in request_kwargs["input"]) + mock_router.get_candidate_model_ids_for_route.assert_not_called() @pytest.mark.asyncio @@ -1444,9 +1386,7 @@ async def test_affinity_does_not_raise_when_boundary_peer_available(): mock_router.get_deployment.return_value = originating check = EncryptedContentAffinityCheck(router=mock_router) - encoded_id = ResponsesAPIRequestUtils._build_encrypted_item_id( - "deployment-a", "rs_test" - ) + encoded_id = ResponsesAPIRequestUtils._build_encrypted_item_id("deployment-a", "rs_test") peer = { "model_info": {"id": "deployment-a-peer"}, "litellm_params": { @@ -1490,9 +1430,7 @@ async def test_model_group_affinity_config_enables_encrypted_content_affinity(): }, target_deployment, ] - encoded_id = ResponsesAPIRequestUtils._build_encrypted_item_id( - "deployment-b", "rs_test" - ) + encoded_id = ResponsesAPIRequestUtils._build_encrypted_item_id("deployment-b", "rs_test") request_kwargs = { "input": [{"type": "reasoning", "id": encoded_id}], "litellm_metadata": {}, @@ -1536,9 +1474,7 @@ async def test_model_group_affinity_config_does_not_disable_global_encrypted_con }, target_deployment, ] - encoded_id = ResponsesAPIRequestUtils._build_encrypted_item_id( - "deployment-b", "rs_test" - ) + encoded_id = ResponsesAPIRequestUtils._build_encrypted_item_id("deployment-b", "rs_test") request_kwargs = { "input": [{"type": "reasoning", "id": encoded_id}], "litellm_metadata": {}, @@ -1600,15 +1536,9 @@ async def test_model_group_encrypted_content_affinity_overrides_global_deploymen try: callbacks = router.optional_callbacks or [] - deployment_callback = next( - cb for cb in callbacks if isinstance(cb, DeploymentAffinityCheck) - ) - encrypted_content_callback = next( - cb for cb in callbacks if isinstance(cb, EncryptedContentAffinityCheck) - ) - assert callbacks.index(encrypted_content_callback) < callbacks.index( - deployment_callback - ) + deployment_callback = next(cb for cb in callbacks if isinstance(cb, DeploymentAffinityCheck)) + encrypted_content_callback = next(cb for cb in callbacks if isinstance(cb, EncryptedContentAffinityCheck)) + assert callbacks.index(encrypted_content_callback) < callbacks.index(deployment_callback) assert encrypted_content_callback.enable_global_affinity is False cache_key = DeploymentAffinityCheck.get_affinity_cache_key( @@ -1620,9 +1550,7 @@ async def test_model_group_encrypted_content_affinity_overrides_global_deploymen value={"model_id": "deployment-a"}, ttl=60, ) - encoded_id = ResponsesAPIRequestUtils._build_encrypted_item_id( - "deployment-b", "rs_test" - ) + encoded_id = ResponsesAPIRequestUtils._build_encrypted_item_id("deployment-b", "rs_test") request_kwargs = { "input": [ { @@ -1643,16 +1571,301 @@ async def test_model_group_encrypted_content_affinity_overrides_global_deploymen ) assert after_deployment_affinity == [deployment_a, deployment_b] - after_encrypted_content_affinity = ( - await encrypted_content_callback.async_filter_deployments( - model=model_group, - healthy_deployments=after_deployment_affinity, - messages=None, - request_kwargs=request_kwargs, - ) + after_encrypted_content_affinity = await encrypted_content_callback.async_filter_deployments( + model=model_group, + healthy_deployments=after_deployment_affinity, + messages=None, + request_kwargs=request_kwargs, ) assert after_encrypted_content_affinity == [deployment_b] assert request_kwargs.get("_encrypted_content_affinity_pinned") is True finally: router.discard() + + +class TestStripEncryptedReasoningFromInput: + def test_keeps_summary_and_drops_encrypted_content_and_id(self): + wrapped = ResponsesAPIRequestUtils._wrap_encrypted_content_with_model_id("gAAAAA-blob", "deployment-a") + encoded_id = ResponsesAPIRequestUtils._build_encrypted_item_id("deployment-a", "rs_1") + request_input = [ + {"role": "user", "content": "first turn"}, + { + "type": "reasoning", + "id": encoded_id, + "encrypted_content": wrapped, + "summary": [{"type": "summary_text", "text": "thought about it"}], + }, + {"type": "reasoning", "id": encoded_id, "encrypted_content": wrapped}, + {"type": "reasoning", "encrypted_content": wrapped, "summary": []}, + {"type": "message", "id": "msg_1", "role": "assistant", "content": "hi"}, + {"role": "user", "content": "second turn"}, + ] + ResponsesAPIRequestUtils.strip_encrypted_reasoning_from_input(request_input) + assert request_input == [ + {"role": "user", "content": "first turn"}, + { + "type": "reasoning", + "summary": [{"type": "summary_text", "text": "thought about it"}], + }, + {"type": "message", "id": "msg_1", "role": "assistant", "content": "hi"}, + {"role": "user", "content": "second turn"}, + ] + + def test_keeps_string_form_summary_when_stripping(self): + wrapped = ResponsesAPIRequestUtils._wrap_encrypted_content_with_model_id("gAAAAA-blob", "deployment-a") + request_input = [ + {"type": "reasoning", "encrypted_content": wrapped, "summary": "plain string thought"}, + { + "type": "reasoning", + "encrypted_content": wrapped, + "content": [{"type": "output_text", "text": "in content"}], + }, + {"type": "reasoning", "encrypted_content": wrapped, "summary": "", "content": []}, + ] + ResponsesAPIRequestUtils.strip_encrypted_reasoning_from_input(request_input) + assert request_input == [ + {"type": "reasoning", "summary": "plain string thought"}, + {"type": "reasoning", "content": [{"type": "output_text", "text": "in content"}]}, + ] + + def test_leaves_input_untouched_when_no_encrypted_reasoning(self): + request_input = [ + {"role": "user", "content": "first turn"}, + {"type": "reasoning", "summary": [{"type": "summary_text", "text": "no blob here"}]}, + {"role": "user", "content": "second turn"}, + ] + before = [dict(item) for item in request_input] + ResponsesAPIRequestUtils.strip_encrypted_reasoning_from_input(request_input) + assert request_input == before + + +def _cross_group_request_kwargs(): + wrapped = ResponsesAPIRequestUtils._wrap_encrypted_content_with_model_id("gAAAAA-blob", "deployment-a") + return { + "litellm_metadata": {}, + "input": [ + {"role": "user", "content": "ZEBRA: why is the sky blue?"}, + { + "type": "reasoning", + "encrypted_content": wrapped, + "summary": [{"type": "summary_text", "text": "scattering"}], + }, + {"type": "message", "role": "assistant", "content": "Rayleigh scattering."}, + {"role": "user", "content": "KIWI: and sunsets?"}, + ], + } + + +@pytest.mark.asyncio +async def test_affinity_strips_encrypted_reasoning_when_routed_to_another_model_group(): + """ + An auto-router tier change (or a model switch with no boundary peer): the + routed pool holds no deployment of the origin's model group. The origin is + healthy, so a 503 would be wrong; the follow-up dispatches to the routed + pool with the origin's encrypted reasoning stripped and its summary kept. + """ + from litellm.router_utils.pre_call_checks.encrypted_content_affinity_check import ( + EncryptedContentAffinityCheck, + ) + + originating = _make_originating_mock(None, "key-a", model_name="gpt-reasoning-tier") + mock_router = _make_router_mock_with_cooldown( + originating, cooldown_entries=[], routed_group_model_ids=["deployment-b"] + ) + check = EncryptedContentAffinityCheck(router=mock_router) + routed_pool = [ + { + "model_info": {"id": "deployment-b"}, + "model_name": "gpt-simple-tier", + "litellm_params": { + "api_base": "https://gateway.example/v1", + "api_key": "key-b", + "model": "openai/gpt-5-nano", + }, + } + ] + request_kwargs = _cross_group_request_kwargs() + original_input = request_kwargs["input"] + + result = await check.async_filter_deployments( + model="gpt-simple-tier", + healthy_deployments=routed_pool, + messages=None, + request_kwargs=request_kwargs, + ) + + assert result is routed_pool + assert "_encrypted_content_affinity_pinned" not in request_kwargs + assert request_kwargs["litellm_metadata"]["encrypted_content_affinity_enabled"] is True + assert request_kwargs["input"] is original_input + assert [item.get("type") or item["role"] for item in original_input] == [ + "user", + "reasoning", + "message", + "user", + ] + assert original_input[1] == { + "type": "reasoning", + "summary": [{"type": "summary_text", "text": "scattering"}], + } + assert not any(isinstance(item, dict) and item.get("encrypted_content") for item in original_input) + + +@pytest.mark.asyncio +async def test_affinity_fails_fast_within_the_origins_own_group(): + """ + Negative class for the tier-change discriminator: the routed group IS the + origin's group (a same-group cooldown, not a tier change), so even with a + healthy non-origin sibling that cannot decrypt the content, the request + still fails fast and the encrypted reasoning is left intact rather than + stripped. Preserves the LIT-3051 cooldown contract. + """ + from litellm.exceptions import ServiceUnavailableError + from litellm.router_utils.pre_call_checks.encrypted_content_affinity_check import ( + EncryptedContentAffinityCheck, + ) + + originating = _make_originating_mock( + "https://account-a.openai.azure.com/", "key-a", model_name="gpt-reasoning-tier" + ) + mock_router = _make_router_mock_with_cooldown( + originating, cooldown_entries=[], routed_group_model_ids=["deployment-a", "deployment-a2"] + ) + check = EncryptedContentAffinityCheck(router=mock_router) + sibling_pool = [ + { + "model_info": {"id": "deployment-a2"}, + "model_name": "gpt-reasoning-tier", + "litellm_params": { + "api_base": "https://account-a2.openai.azure.com/", + "api_key": "key-a2", + "model": "azure/gpt-5.4", + }, + } + ] + request_kwargs = _cross_group_request_kwargs() + + with pytest.raises(ServiceUnavailableError): + await check.async_filter_deployments( + model="gpt-reasoning-tier", + healthy_deployments=sibling_pool, + messages=None, + request_kwargs=request_kwargs, + ) + + assert request_kwargs["input"][1].get("encrypted_content") + + +@pytest.mark.asyncio +async def test_affinity_does_not_strip_when_group_is_spelled_differently_but_same_by_id(): + """ + The discriminator must key on deployment-id membership, not on the model-group + name string. Here the origin's configured group is spelled ``openai/gpt-5.4-mini`` + while the routed group is the canonical ``gpt-5.4-mini``: same group, different + spelling. A name compare (``originating.model_name != model``) would read this as + a tier change and strip the reasoning it did not have to. Because the origin's id + is a member of the routed group, this is a same-group cooldown instead: the request + fails fast and the encrypted reasoning is left intact. + """ + from litellm.exceptions import ServiceUnavailableError + from litellm.router_utils.pre_call_checks.encrypted_content_affinity_check import ( + EncryptedContentAffinityCheck, + ) + + originating = _make_originating_mock(None, "key-a", model_name="openai/gpt-5.4-mini") + mock_router = _make_router_mock_with_cooldown( + originating, cooldown_entries=[], routed_group_model_ids=["deployment-mini-a", "deployment-mini-b"] + ) + check = EncryptedContentAffinityCheck(router=mock_router) + wrapped = ResponsesAPIRequestUtils._wrap_encrypted_content_with_model_id("gAAAAA-blob", "deployment-mini-a") + sibling_pool = [ + { + "model_info": {"id": "deployment-mini-b"}, + "model_name": "gpt-5.4-mini", + "litellm_params": { + "api_base": "https://gateway.example/v1", + "api_key": "key-b", + "model": "openai/gpt-5.4-mini", + }, + } + ] + request_kwargs = { + "litellm_metadata": {}, + "input": [ + {"role": "user", "content": "why is the sky blue?"}, + { + "type": "reasoning", + "encrypted_content": wrapped, + "summary": [{"type": "summary_text", "text": "scattering"}], + }, + {"role": "user", "content": "and sunsets?"}, + ], + } + + with pytest.raises(ServiceUnavailableError): + await check.async_filter_deployments( + model="gpt-5.4-mini", + healthy_deployments=sibling_pool, + messages=None, + request_kwargs=request_kwargs, + ) + + assert request_kwargs["input"][1].get("encrypted_content") + + +@pytest.mark.asyncio +async def test_affinity_honors_router_candidate_ids_for_team_and_pattern_routes(): + """ + The exact `model_name` index does not include team-public or pattern routes, so a + same-group cooldown reached only through one of those would be misread as a tier change + and stripped. The check asks the router for the candidate ids it resolves for the route + (`get_candidate_model_ids_for_route`), which covers those paths, rather than the bare + index. Here that set marks the origin as a candidate, so the request fails fast with its + reasoning intact, and the routed group and team are passed through to the router. + """ + from litellm.exceptions import ServiceUnavailableError + from litellm.router_utils.pre_call_checks.encrypted_content_affinity_check import ( + EncryptedContentAffinityCheck, + ) + + originating = _make_originating_mock(None, "key-a", model_name="model_name_teamA_uuid") + mock_router = _make_router_mock_with_cooldown( + originating, cooldown_entries=[], routed_group_model_ids=["deployment-team-a", "deployment-team-b"] + ) + check = EncryptedContentAffinityCheck(router=mock_router) + wrapped = ResponsesAPIRequestUtils._wrap_encrypted_content_with_model_id("gAAAAA-blob", "deployment-team-a") + sibling_pool = [ + { + "model_info": {"id": "deployment-team-b"}, + "model_name": "team-public-model", + "litellm_params": { + "api_base": "https://gateway.example/v1", + "api_key": "key-b", + "model": "openai/gpt-5.4-mini", + }, + } + ] + request_kwargs = { + "litellm_metadata": {"user_api_key_team_id": "teamA"}, + "input": [ + {"role": "user", "content": "why is the sky blue?"}, + { + "type": "reasoning", + "encrypted_content": wrapped, + "summary": [{"type": "summary_text", "text": "scattering"}], + }, + {"role": "user", "content": "and sunsets?"}, + ], + } + + with pytest.raises(ServiceUnavailableError): + await check.async_filter_deployments( + model="team-public-model", + healthy_deployments=sibling_pool, + messages=None, + request_kwargs=request_kwargs, + ) + + assert request_kwargs["input"][1].get("encrypted_content") + mock_router.get_candidate_model_ids_for_route.assert_called_once_with(model="team-public-model", team_id="teamA") diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py index bc79c5f6589..80da922724d 100644 --- a/tests/test_litellm/test_router.py +++ b/tests/test_litellm/test_router.py @@ -14728,3 +14728,45 @@ def test_router_stays_quiet_when_a_deployment_drop_params_is_a_flag(value, caplo ) assert "is not a flag value" not in caplog.text + + +def test_get_candidate_model_ids_for_route_covers_model_name_and_pattern(): + """ + get_candidate_model_ids_for_route resolves a route the way the router does, so a + pre-call check can tell a genuine cross-group route from same-group unavailability. + A concrete model group returns its member ids; a wildcard/pattern deployment is + included for a concrete model it matches, which the bare model_name index misses. + Regression guard for the LIT-7195 tier-change discriminator's team/pattern gaps. + """ + router = Router( + model_list=[ + { + "model_name": "grp", + "litellm_params": {"model": "openai/gpt-4o", "api_key": "sk-a", "api_base": "https://x.invalid"}, + "model_info": {"id": "dep-a"}, + }, + { + "model_name": "grp", + "litellm_params": {"model": "openai/gpt-4o-mini", "api_key": "sk-b", "api_base": "https://x.invalid"}, + "model_info": {"id": "dep-b"}, + }, + { + "model_name": "openai/*", + "litellm_params": {"model": "openai/*", "api_key": "sk-c", "api_base": "https://x.invalid"}, + "model_info": {"id": "dep-wild"}, + }, + ] + ) + + assert router.get_candidate_model_ids_for_route(model="grp") == frozenset({"dep-a", "dep-b"}) + assert "dep-wild" in router.get_candidate_model_ids_for_route(model="openai/gpt-4o-some-new-model") + + +def test_deployment_ids_stringifies_ids_and_skips_entries_without_a_model_info_id(): + deployments = ( + {"model_info": {"id": "a"}}, + {"model_info": {"id": 2}}, + {"model_info": {}}, + {"no_model_info": True}, + ) + assert Router._deployment_ids(deployments) == frozenset({"a", "2"})