fix(llm): only attach prompt-cache points on routes LiteLLM serves (#1186)

Co-authored-by: Ahmed Allam <ahmed39652003@gmail.com>

strix/config/models.py

@@ -864,6 +864,22 @@ def is_claude_model(model_name: str) -> bool:
     return "claude" in (model_name or "").strip().lower()
 
 
+def routes_through_litellm(model_name: str | None) -> bool:
+    """Whether :class:`StrixProvider` sends this model through LiteLLM.
+
+    Bare names and the ``openai/``/``any-llm/`` prefixes are served by the SDK's
+    own clients, which raise ``TypeError`` on request fields they do not know,
+    so LiteLLM-only fields must not be attached there. A bare ``claude-...``
+    name is exactly that case: an ``LLM_API_BASE`` pointing at an
+    OpenAI-compatible gateway in front of Claude.
+    """
+    name = (model_name or "").strip()
+    if not name or codex.subscription_model(name):
+        return False
+    prefix, _, rest = name.partition("/")
+    return bool(rest) and prefix.lower() not in {"openai", "any-llm"}
+
+
 def is_bedrock_route(model_name: str) -> bool:
     name = (model_name or "").strip().lower()
     return name.startswith("bedrock/") or "anthropic." in name

strix/core/inputs.py

@@ -18,6 +18,7 @@ from strix.config.models import (
     is_openrouter_model,
     model_supports_reasoning,
     request_timeout_extra_args,
+    routes_through_litellm,
 )
 from strix.core.sessions import scrub_images_from_items
 
@@ -317,8 +318,13 @@ def _prompt_cache_extra_args(model_name: str) -> dict[str, Any] | None:
     it — elsewhere it leaks onto the wire and native Anthropic 400s). Unmapped
     Bedrock models get no points at all: Bedrock rejects the passed-through
     field outright.
+
+    The field is LiteLLM's own, consumed by its transform, so it only goes to
+    routes LiteLLM serves. A bare ``claude-...`` name is served by the SDK's
+    OpenAI client instead (a gateway in front of Claude), and that client raises
+    ``TypeError`` on request kwargs it does not know.
     """
-    if not is_claude_model(model_name):
+    if not is_claude_model(model_name) or not routes_through_litellm(model_name):
         return None
     if is_bedrock_route(model_name) and not bedrock_route_supports_prompt_caching(model_name):
         return None

tests/test_inputs.py

@@ -90,6 +90,16 @@ def test_make_model_settings_enables_prompt_cache_for_non_bedrock_claude(model_n
     ]
 
 
+@pytest.mark.parametrize(
+    "model_name",
+    ["claude-sonnet-4-5", "openai/claude-sonnet-4-5", "any-llm/anthropic/claude-sonnet-4-5"],
+)
+def test_no_prompt_cache_for_claude_off_the_litellm_route(model_name: str) -> None:
+    # These names are served by SDK clients that raise TypeError on LiteLLM-only
+    # request kwargs — e.g. a gateway in front of Claude reached with a bare name.
+    assert _cache_points(model_name) is None
+
+
 def test_tool_config_point_not_leaked_to_non_bedrock_claude() -> None:
     # LiteLLM only consumes tool_config on Bedrock; elsewhere it leaks onto the
     # wire and native Anthropic 400s.

tests/test_models.py

@@ -3,12 +3,17 @@
 from __future__ import annotations
 
 import pytest
+from agents.extensions.models.litellm_model import LitellmModel
 from agents.model_settings import ModelSettings
 
 from strix.config.models import (
     RECOMMENDED_MODEL_NAMES,
+    StrixProvider,
+    _NonStreamingModel,
+    _TurnGuardModel,
     is_recommended_or_frontier_model,
     request_timeout_extra_args,
+    routes_through_litellm,
     supports_strict_tool_schemas,
 )
 
@@ -112,3 +117,38 @@ def test_claude_routes_reject_strict_tool_schemas(model_name: str) -> None:
 )
 def test_other_routes_keep_strict_tool_schemas(model_name: str) -> None:
     assert supports_strict_tool_schemas(model_name)
+
+
+@pytest.mark.parametrize(
+    ("model_name", "litellm"),
+    [
+        ("claude-sonnet-4-5", False),
+        ("openai/claude-sonnet-4-5", False),
+        ("any-llm/anthropic/claude-sonnet-4-5", False),
+        ("anthropic/claude-sonnet-4-5", True),
+        ("litellm/anthropic/claude-sonnet-4-5", True),
+        ("bedrock/anthropic.claude-sonnet-4-5-20250929-v1:0", True),
+        ("ollama/llama3", True),
+    ],
+)
+def test_routes_through_litellm_matches_the_provider(
+    monkeypatch: pytest.MonkeyPatch, model_name: str, litellm: bool
+) -> None:
+    """The helper must agree with what StrixProvider actually builds.
+
+    Callers use it to decide whether a LiteLLM-only request field is safe to
+    attach; on the SDK's own clients such a field raises TypeError mid-turn, so
+    drift here breaks every request on that route.
+    """
+    monkeypatch.setenv("OPENAI_API_KEY", "test-key")
+    assert routes_through_litellm(model_name) is litellm
+    try:
+        model = StrixProvider().get_model(model_name)
+    except ImportError:
+        # any-llm's client is an optional dependency; reaching it at all already
+        # proves the route is not LiteLLM's.
+        assert not litellm
+        return
+    while isinstance(model, _NonStreamingModel | _TurnGuardModel):
+        model = model._inner
+    assert isinstance(model, LitellmModel) is litellm