fix(llm): pin LLMTR endpoint for a dedicated llmtr/ dedupe model

When the dedupe model uses llmtr/ but the main model is a different
provider, the global litellm.api_base (derived from the main model) is
not the LLMTR gateway, so the dedupe request would resolve to LiteLLM's
default openai/ endpoint and fail. Mirror the existing per-call dedupe
credential pattern: inject the LLMTR gateway base URL into the dedupe
request's extra_args when no explicit DEDUPE_LLM_API_BASE is set. An
explicit DEDUPE_LLM_API_BASE still wins.

Addresses Greptile review feedback on the mixed main/dedupe routing case.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ag2XdAsvrR7XKMmQsAjqSN

strix/report/dedupe.py

@@ -13,8 +13,10 @@ from openai.types.responses import ResponseOutputMessage
 
 from strix.config import load_settings
 from strix.config.models import (
+    LLMTR_API_BASE,
     StrixProvider,
     configure_sdk_model_defaults,
+    is_llmtr_model,
 )
 from strix.core.inputs import make_model_settings
 from strix.report.state import get_global_report_state
@@ -45,6 +47,12 @@ def _dedupe_extra_args(dedupe: DedupeSettings) -> dict[str, str]:
         extra["api_key"] = dedupe.api_key.strip()
     if dedupe.api_base and dedupe.api_base.strip():
         extra["api_base"] = dedupe.api_base.strip()
+    elif is_llmtr_model(dedupe.model):
+        # A dedicated llmtr/ dedupe model reaches the LLMTR gateway only through
+        # its endpoint. The global default is derived from the main model, which
+        # may be a different provider, so pin the gateway per call here rather
+        # than relying on the shared litellm.api_base.
+        extra["api_base"] = LLMTR_API_BASE
     return extra
 
 

tests/test_dedupe_model.py

@@ -7,7 +7,7 @@ from typing import TYPE_CHECKING
 
 from strix.config import loader
 from strix.config.settings import DedupeSettings
-from strix.report.dedupe import _dedupe_model_settings
+from strix.report.dedupe import _dedupe_extra_args, _dedupe_model_settings
 
 
 if TYPE_CHECKING:
@@ -44,6 +44,35 @@ def test_dedupe_endpoint_sent_per_call() -> None:
     assert (settings.extra_args or {})["api_key"] == "dedupe-key"
 
 
+def test_llmtr_dedupe_model_pins_gateway_endpoint_per_call() -> None:
+    # A dedicated llmtr/ dedupe model reaches the gateway only via its endpoint.
+    # The global default is keyed to the main model, which may be a different
+    # provider, so the endpoint must ride on the dedupe request itself.
+    dedupe = DedupeSettings(
+        STRIX_DEDUPE_MODEL="llmtr/openai/gpt-5.5",
+        DEDUPE_LLM_API_KEY="llmtr-key",
+    )
+    extra = _dedupe_extra_args(dedupe)
+    assert extra["api_base"] == "https://llmtr.com/v1"
+    assert extra["api_key"] == "llmtr-key"
+
+
+def test_llmtr_dedupe_model_respects_explicit_endpoint() -> None:
+    # An explicit DEDUPE_LLM_API_BASE (e.g. a proxy in front of LLMTR) wins over
+    # the gateway default.
+    dedupe = DedupeSettings(
+        STRIX_DEDUPE_MODEL="llmtr/openai/gpt-5.5",
+        DEDUPE_LLM_API_KEY="llmtr-key",
+        DEDUPE_LLM_API_BASE="https://proxy.example/v1",
+    )
+    assert _dedupe_extra_args(dedupe)["api_base"] == "https://proxy.example/v1"
+
+
+def test_non_llmtr_dedupe_model_gets_no_injected_endpoint() -> None:
+    dedupe = DedupeSettings(STRIX_DEDUPE_MODEL="openai/gpt-5.5", DEDUPE_LLM_API_KEY="k")
+    assert "api_base" not in _dedupe_extra_args(dedupe)
+
+
 def test_dedicated_dedupe_model_uses_own_headers_not_main() -> None:
     dedupe = DedupeSettings(
         STRIX_DEDUPE_MODEL="deepseek/cheap",