Add read timeouts to llama-server clients so hangs fail fast

The OpenAI client defaults to a 600s read timeout, so a wedged
llama-server blocked requests for 10 minutes and never raised, which
meant with_fallbacks() to OpenAI never triggered. Set a per-chunk read
timeout (120s, configurable via LLAMA_TIMEOUT_SECONDS) plus one retry
on the ChatOpenAI and LLMClient clients. Context is preserved because
conversation history is rebuilt from Postgres on every request.

blueprints/conversation/agents.py

@@ -1,6 +1,7 @@
 import os
 from typing import cast
 
+import httpx
 from dotenv import load_dotenv
 from langchain.agents import create_agent
 from langchain.chat_models import BaseChatModel
@@ -20,12 +21,21 @@ from utils.ynab_service import YNABService
 load_dotenv()
 
 # Configure LLM with llama-server or OpenAI fallback
+# Read timeout between streamed chunks. The OpenAI client defaults to 600s,
+# which means a wedged llama-server blocks a request for 10 minutes before
+# erroring. With a tight timeout, a hang raises quickly so
+# with_fallbacks() below kicks in and the conversation (rebuilt from the
+# DB each request) continues seamlessly on OpenAI.
+LLAMA_READ_TIMEOUT = float(os.getenv("LLAMA_TIMEOUT_SECONDS", "120"))
+
 llama_url = os.getenv("LLAMA_SERVER_URL")
 if llama_url:
     llama_chat = ChatOpenAI(
         base_url=llama_url,
         api_key="not-needed",
         model=os.getenv("LLAMA_MODEL_NAME", "llama-3.1-8b-instruct"),
+        timeout=httpx.Timeout(LLAMA_READ_TIMEOUT, connect=10.0),
+        max_retries=1,
     )
 else:
     llama_chat = None

llm.py

@@ -1,6 +1,7 @@
 import os
 import logging
 
+import httpx
 from openai import OpenAI
 from dotenv import load_dotenv
 
@@ -13,7 +14,15 @@ class LLMClient:
     def __init__(self):
         llama_url = os.getenv("LLAMA_SERVER_URL")
         if llama_url:
-            self.client = OpenAI(base_url=llama_url, api_key="not-needed")
+            # Tight read timeout so a wedged llama-server fails fast instead
+            # of hanging for the OpenAI client's 600s default.
+            read_timeout = float(os.getenv("LLAMA_TIMEOUT_SECONDS", "120"))
+            self.client = OpenAI(
+                base_url=llama_url,
+                api_key="not-needed",
+                timeout=httpx.Timeout(read_timeout, connect=10.0),
+                max_retries=1,
+            )
             self.model = os.getenv("LLAMA_MODEL_NAME", "llama-3.1-8b-instruct")
             self.PROVIDER = "llama_server"
             logging.info("Using llama_server as LLM backend")