Separating agent

agents/code_writing_agent.py

@@ -0,0 +1,19 @@
+import os
+import smolagents
+import smolagents.models
+
+
+def create_code_writing_agent():
+    return smolagents.CodeAgent(
+        model=smolagents.models.OpenAIServerModel(
+            model_id=os.getenv("AGENT_MODEL", ""),
+            api_base=os.getenv("UPSTREAM_OPENAI_BASE", "").rstrip("/"),
+            api_key=os.getenv("OPENAI_API_KEY"),
+        ),
+        tools=[],  # no extra tools
+        add_base_tools=False,
+        max_steps=4,
+        verbosity_level=int(
+            os.getenv("AGENT_VERBOSITY", "1")
+        ),  # quieter by default; override via env
+    )

proxy.py

@@ -16,13 +16,11 @@ import fastapi.responses
 import io
 import contextlib
 
-# smolagents + OpenAI-compatible model wrapper
-import smolagents
-import smolagents.models
-
 # Upstream pass-through
 import httpx
 
+from agents.code_writing_agent import create_code_writing_agent
+
 # Logging setup
 logging.basicConfig(level=os.getenv("LOG_LEVEL", "INFO").upper())
 log = logging.getLogger(__name__)
@@ -44,22 +42,6 @@ if not UPSTREAM_BASE:
 if not HF_TOKEN:
     log.warning("HF_TOKEN is empty; upstream may 401/403 if it requires auth.")
 
-# ================== Agent ====================
-llm = smolagents.models.OpenAIServerModel(
-    model_id=AGENT_MODEL,
-    api_base=UPSTREAM_BASE,
-    api_key=HF_TOKEN,
-)
-agent = smolagents.CodeAgent(
-    model=llm,
-    tools=[],  # no extra tools
-    add_base_tools=False,
-    max_steps=4,
-    verbosity_level=int(
-        os.getenv("AGENT_VERBOSITY", "1")
-    ),  # quieter by default; override via env
-)
-
 # ================== FastAPI ==================
 app = fastapi.FastAPI()
 
@@ -405,7 +387,7 @@ async def run_agent_stream(task: str, agent_obj: typing.Optional[typing.Any] = N
     """
     loop = asyncio.get_running_loop()
     q: asyncio.Queue = asyncio.Queue()
-    agent_to_use = agent_obj or agent
+    agent_to_use = agent_obj or create_code_writing_agent
 
     stop_evt = threading.Event()
 
@@ -712,16 +694,7 @@ async def chat_completions(req: fastapi.Request):
         AGENT_MODEL + "-nothink",
     ) and isinstance(model_name, str):
         try:
-            req_llm = smolagents.models.OpenAIServerModel(
-                model_id=model_name, api_base=UPSTREAM_BASE, api_key=HF_TOKEN
-            )
-            agent_for_request = smolagents.CodeAgent(
-                model=req_llm,
-                tools=[],
-                add_base_tools=False,
-                max_steps=4,
-                verbosity_level=int(os.getenv("AGENT_VERBOSITY", "1")),
-            )
+            agent_for_request = create_code_writing_agent()
         except Exception:
             log.exception(
                 "Failed to construct agent for model '%s'; using default", model_name
@@ -934,4 +907,4 @@ if __name__ == "__main__":
 
     uvicorn.run(
         "app:app", host="0.0.0.0", port=int(os.getenv("PORT", "8000")), reload=False
-    )
+    )