Raise LLM read timeout to 300s; prefer LAN endpoint in env example

Midnight failures were read timeouts against the public URL — hairpin NAT
and/or the request queuing behind other traffic on the shared model.
LAN endpoint + longer timeout covers both.

.env.example

@@ -1,7 +1,8 @@
 # Copy to .env and adjust. docker compose reads this automatically.
 
-# Any OpenAI-compatible endpoint (llamaswap example below)
-OPENAI_BASE_URL=http://swap.obahan.xyz/v1
+# Any OpenAI-compatible endpoint. Use the LAN address from the minisforum —
+# the public URL hairpins through the router and can stall.
+OPENAI_BASE_URL=http://minisforum:65000/v1
 OPENAI_MODEL=halogen-qwen3.8-flash-next
 OPENAI_API_KEY=***
 

app/llm_client.py

@@ -124,7 +124,9 @@ def generate_jokes(article_title: str, article_extract: str) -> list[str]:
         "max_tokens": 800,
     }
 
-    resp = httpx.post(url, headers=headers, json=body, timeout=120)
+    # 300s: the model may be shared (e.g. serving this agent too), so a
+    # request can legitimately queue behind other traffic.
+    resp = httpx.post(url, headers=headers, json=body, timeout=300)
     resp.raise_for_status()
     payload = resp.json()
     content = payload["choices"][0]["message"]["content"]