Add modal_custom provider for OpenAI-compatible endpoints

agent/llm_client.py

@@ -113,6 +113,8 @@ class LLMClient:
             return "anthropic"
         elif self.openai_key:
             return "openai"
+        elif os.getenv("MODAL_CUSTOM_BASE_URL"):
+            return "modal_custom"
         elif self.blaxel_key:
             return "blaxel"
         elif self.sambanova_key:
@@ -151,6 +153,8 @@ class LLMClient:
             return self._call_nebius(messages, temperature, max_tokens)
         elif self.provider == "modal":
             return self._call_modal(messages, temperature, max_tokens)
+        elif self.provider == "modal_custom":
+            return self._call_modal_custom(messages, temperature, max_tokens)
         elif self.provider == "huggingface":
             return self._call_huggingface(messages, temperature, max_tokens)
     
@@ -258,6 +262,54 @@ class LLMClient:
             if monitor:
                 monitor.complete_call(call_id, success=False, error_message=str(e))
             raise
+
+    def _call_modal_custom(self, messages, temperature, max_tokens):
+        """
+        Call an OpenAI-compatible endpoint hosted on Modal (custom vLLM server).
+        Uses MODAL_CUSTOM_BASE_URL, MODAL_CUSTOM_MODEL, MODAL_CUSTOM_API_KEY.
+        """
+        call_id = str(uuid.uuid4())
+        model = (
+            os.getenv("MODAL_CUSTOM_MODEL")
+            or os.getenv("OG_OPENAI_MODEL")
+            or "network-llm"
+        )
+        base_url = os.getenv("MODAL_CUSTOM_BASE_URL") or os.getenv("OPENAI_BASE_URL")
+        api_key = (
+            os.getenv("MODAL_CUSTOM_API_KEY")
+            or os.getenv("OPENAI_API_KEY")
+            or os.getenv("OPENAI_MCP_1ST_BDAY")
+            or ""
+        )
+        if not base_url:
+            raise RuntimeError("MODAL_CUSTOM_BASE_URL is not set for modal_custom provider.")
+
+        if monitor:
+            monitor.start_call(call_id, "llm", "modal_custom", model, temperature=temperature)
+
+        try:
+            from openai import OpenAI
+            client = OpenAI(api_key=api_key, base_url=base_url)
+            response = client.chat.completions.create(
+                model=model,
+                messages=[{"role": m.role, "content": m.content} for m in messages],
+                temperature=temperature,
+                max_tokens=max_tokens,
+            )
+            content = response.choices[0].message.content if response.choices else ""
+            if monitor and getattr(response, "usage", None):
+                monitor.complete_call(
+                    call_id,
+                    success=True,
+                    input_tokens=response.usage.prompt_tokens,
+                    output_tokens=response.usage.completion_tokens,
+                )
+            return content
+        except Exception as e:
+            logger.error(f"Modal custom endpoint error: {e}")
+            if monitor:
+                monitor.complete_call(call_id, success=False, error_message=str(e))
+            raise
     
     def _call_anthropic(self, messages, temperature, max_tokens, stream):
         """Call Anthropic API"""