Sync from GitHub 408f81efe5d73e4f9e0623f822105a8849698a0c

config.py

@@ -33,12 +33,13 @@ MAX_CONVERSATION_TURNS = 20
 # ============================================ LLM Configuration ============================================
 
 # Each role-specific model configuration is defined as (PROVIDER, MODEL_NAME).
-# Keep the master branch's latency-oriented defaults.
-MAIN_AGENT_MODEL = ('openai', 'gpt-4.1')
-FALLBACK_MODELS = [('openai', 'gpt-5-mini')]
-LANGUAGE_DETECTION_MODEL = ('openai', 'gpt-4o-mini')
-CONFIDENCE_SCORING_MODEL = ('openai', 'gpt-4o-mini')
-SUMMARIZATION_MODEL = ('openai', 'gpt-4.1')
+# All roles run through OpenRouter (OpenAI-compatible) so the app no longer
+# depends on a separately funded OpenAI account.
+MAIN_AGENT_MODEL = ('open_router:openai', 'openai/gpt-4.1')
+FALLBACK_MODELS = [('open_router:openai', 'openai/gpt-4o-mini')]
+LANGUAGE_DETECTION_MODEL = ('open_router:openai', 'openai/gpt-4o-mini')
+CONFIDENCE_SCORING_MODEL = ('open_router:openai', 'openai/gpt-4o-mini')
+SUMMARIZATION_MODEL = ('open_router:openai', 'openai/gpt-4.1')
 
 # Legacy defaults retained for compatibility with existing callers.
 LLM_PROVIDER = 'openai'

tests/test_master_transfer_integrations.py

@@ -234,7 +234,7 @@ def test_query_falls_back_to_bm25_when_embedding_fails(monkeypatch):
     assert collection.query.bm25_calls[0]["limit"] == 3
 
 
-def test_model_config_keeps_master_defaults_and_budgets(monkeypatch):
+def test_model_config_uses_openrouter_defaults_and_budgets(monkeypatch):
     calls = []
 
     def fake_initialize_model(cls, provider, model, role="main"):
@@ -252,10 +252,10 @@ def test_model_config_keeps_master_defaults_and_budgets(monkeypatch):
     ModelConfigurator.get_confidence_scoring_model()
     ModelConfigurator.get_fallback_models()
 
-    assert config.llm.MAIN_AGENT_MODEL == ("openai", "gpt-4.1")
-    assert config.llm.FALLBACK_MODELS == [("openai", "gpt-5-mini")]
-    assert config.llm.get_default_model() == "gpt-4.1"
-    assert config.llm.get_fallback_models()[0][1] == "gpt-5-mini"
+    # All roles run through OpenRouter so the app no longer needs a funded
+    # OpenAI account.
+    assert config.llm.MAIN_AGENT_MODEL == ("open_router:openai", "openai/gpt-4.1")
+    assert config.llm.FALLBACK_MODELS == [("open_router:openai", "openai/gpt-4o-mini")]
     assert ModelConfigurator._openai_budget("main") == {
         "max_tokens": 3072,
         "timeout": 30,
@@ -266,6 +266,6 @@ def test_model_config_keeps_master_defaults_and_budgets(monkeypatch):
         "timeout": 10,
         "request_timeout": 10,
     }
-    assert ("openai", "gpt-4.1", "main") in calls
-    assert ("openai", "gpt-4o-mini", "language_detector") in calls
-    assert ("openai", "gpt-4o-mini", "confidence_scoring") in calls
+    assert ("open_router:openai", "openai/gpt-4.1", "main") in calls
+    assert ("open_router:openai", "openai/gpt-4o-mini", "language_detector") in calls
+    assert ("open_router:openai", "openai/gpt-4o-mini", "confidence_scoring") in calls

tests/test_uat_llm_judge.py

@@ -25,9 +25,27 @@ from src.config import config
 from src.rag.agent_chain import ExecutiveAgentChain
 
 
+OPENROUTER_BASE_URL = "https://openrouter.ai/api/v1"
+
+
+def _use_openrouter() -> bool:
+    """Route the LLM judge through OpenRouter by default (opt out with
+    UAT_USE_OPENROUTER=0).
+
+    The chatbot's own models already run through OpenRouter via config.py; this
+    keeps the judge consistent so the UAT does not depend on a funded OpenAI
+    account.
+    """
+    if os.getenv("UAT_USE_OPENROUTER") == "0":
+        return False
+    return bool(os.getenv("OPEN_ROUTER_API_KEY"))
+
+
 DEFAULT_EXCEL = Path(__file__).resolve().parent / "fixtures" / "UAT.xlsx"
 SKIPPED_SHEETS = {"About", "TestProtocol", "Reporting"}
-DEFAULT_JUDGE_MODEL = os.getenv("UAT_JUDGE_MODEL", "gpt-4o-mini")
+DEFAULT_JUDGE_MODEL = os.getenv("UAT_JUDGE_MODEL") or (
+    "openai/gpt-4o-mini" if _use_openrouter() else "gpt-4o-mini"
+)
 MIN_ACCEPTABLE_SCORE = float(os.getenv("UAT_MIN_SCORE", "7.0"))
 
 
@@ -264,6 +282,13 @@ def _openai_client():
     except ImportError as exc:
         raise RuntimeError("openai is required for the UAT LLM judge.") from exc
 
+    if _use_openrouter():
+        return OpenAI(
+            api_key=os.environ["OPEN_ROUTER_API_KEY"],
+            base_url=OPENROUTER_BASE_URL,
+            timeout=float(os.getenv("UAT_JUDGE_TIMEOUT_S", "120")),
+        )
+
     api_key = os.getenv("OPENAI_API_KEY") or os.getenv("OPENAI_COMPAT_API_KEY")
     base_url = os.getenv("OPENAI_COMPAT_BASE_URL")
     if not api_key: