fix: agent/run uses direct Groq→Cerebras→OpenRouter fallback chain (no Databricks)

app_routes_extension.py

@@ -94,23 +94,47 @@ class AgentRunRequest(BaseModel):
 
 @ext_router.post("/api/agent/run")
 async def agent_run(body: AgentRunRequest) -> dict[str, Any]:
-    """Route agent prompts through the LLM gateway."""
-    t0 = time.monotonic()
-    try:
-        from backend.llm.gateway import ModelGateway
-        gateway = ModelGateway()
-        result = await gateway.complete(
-            prompt=body.message,
-            intent="agent",
-        )
-        return {
-            "response": result,
-            "provider": body.provider,
-            "latency_ms": round((time.monotonic() - t0) * 1000),
-        }
-    except Exception as exc:
-        logger.error("agent_run failed: %s", exc)
-        return {"error": str(exc), "latency_ms": round((time.monotonic() - t0) * 1000)}
+    """Route agent prompts through the production LLM fallback chain (Groq→Cerebras→OpenRouter)."""
+    import os, httpx, time as _time
+    t0 = _time.monotonic()
+
+    messages = [{"role": "user", "content": body.message}]
+
+    chain = []
+    if os.environ.get("GROQ_API_KEY"):
+        chain.append(("groq", "https://api.groq.com/openai/v1/chat/completions",
+                       "llama-3.3-70b-versatile", f"Bearer {os.environ['GROQ_API_KEY']}"))
+    if os.environ.get("CEREBRAS_API_KEY"):
+        chain.append(("cerebras", "https://api.cerebras.ai/v1/chat/completions",
+                       "llama3.1-70b", f"Bearer {os.environ['CEREBRAS_API_KEY']}"))
+    if os.environ.get("OPENROUTER_API_KEY"):
+        chain.append(("openrouter", "https://openrouter.ai/api/v1/chat/completions",
+                       "meta-llama/llama-3.1-8b-instruct:free",
+                       f"Bearer {os.environ['OPENROUTER_API_KEY']}"))
+
+    last_error = "No LLM providers configured"
+    async with httpx.AsyncClient(timeout=30.0) as client:
+        for provider_name, url, model, auth in chain:
+            try:
+                resp = await client.post(url,
+                    headers={"Authorization": auth, "Content-Type": "application/json"},
+                    json={"model": model, "messages": messages, "max_tokens": 1024},
+                )
+                if resp.status_code == 200:
+                    data = resp.json()
+                    text = data["choices"][0]["message"]["content"]
+                    return {
+                        "response": text,
+                        "provider": provider_name,
+                        "model": model,
+                        "latency_ms": round((_time.monotonic() - t0) * 1000),
+                    }
+                last_error = f"{provider_name} returned {resp.status_code}: {resp.text[:200]}"
+            except Exception as exc:
+                last_error = f"{provider_name} error: {exc}"
+                continue
+
+    return {"error": last_error, "latency_ms": round((_time.monotonic() - t0) * 1000)}
 
 
 def register(app) -> None: