Migrate upstream client to HuggingFace inference router

Switch config.py from the HY_API_KEY/HY_BASE_URL/HY_MODEL env vars to a
fixed HF router endpoint and a hardcoded model id (tencent/Hy3:deepinfra),
authenticating with HF_TOKEN. Stream-based chat path is unchanged.

Co-Authored-By: Claude <noreply@anthropic.com>

config.py

@@ -1,9 +1,11 @@
-"""Runtime configuration loaded from environment variables.
+"""Runtime configuration for the upstream OpenAI-compatible endpoint.
+
+The model is served via the HuggingFace inference router. The only secret
+we need is ``HF_TOKEN`` (a HuggingFace access token); the base URL and model
+name are fixed.
 
 Required:
-    HY_API_KEY        API key for the upstream OpenAI-compatible endpoint.
-    HY_BASE_URL       OpenAI-compatible base URL of the upstream endpoint.
-    HY_MODEL          Model name passed to chat.completions.create.
+    HF_TOKEN          HuggingFace access token (read access to the model).
 
 Optional:
     HY_LOG_LEVEL      One of DEBUG / INFO / WARNING / ERROR (default: WARNING).
@@ -17,9 +19,7 @@ import os
 from openai import OpenAI
 
 
-API_KEY = os.environ.get("HY_API_KEY", "").strip()
-BASE_URL = os.environ.get("HY_BASE_URL", "").strip()
-MODEL = os.environ.get("HY_MODEL", "").strip()
+MODEL = "tencent/Hy3:deepinfra"
 LOG_LEVEL = os.environ.get("HY_LOG_LEVEL", "WARNING").upper()
 
 
@@ -37,34 +37,19 @@ if not logging.getLogger().handlers:
 logger = logging.getLogger("Hy3")
 
 
-# Surface missing required env vars loudly at import time. We don't
-# raise — that would prevent the Gradio server from even starting up
-# (and HuggingFace Spaces would just show a blank build-failed page),
-# making it harder to debug. Instead: warn now, and the per-request
-# code paths will still fail cleanly with the upstream API's own
-# error message.
-_missing = [
-    name
-    for name, val in (
-        ("HY_API_KEY", API_KEY),
-        ("HY_BASE_URL", BASE_URL),
-        ("HY_MODEL", MODEL),
-    )
-    if not val
-]
-if _missing:
+# Surface a missing HF_TOKEN loudly at import time. We don't raise — that
+# would prevent the Gradio server from even starting up (and HuggingFace
+# Spaces would just show a blank build-failed page), making it harder to
+# debug. Instead: warn now, and the per-request code paths will still fail
+# cleanly with the upstream API's own 401 error message.
+if not os.environ.get("HF_TOKEN", "").strip():
     logger.warning(
-        "Required env var(s) not set: %s. "
-        "The app will boot but every chat request will fail until they are "
-        "configured (e.g. as HuggingFace Space secrets / variables).",
-        ", ".join(_missing),
+        "HF_TOKEN is not set. The app will boot but every chat request will "
+        "fail until it is configured (e.g. as a HuggingFace Space secret)."
     )
 
 
-# ``base_url=""`` would be passed through to httpx and produce undefined
-# request URLs. ``None`` makes the OpenAI SDK fall back to its default
-# (api.openai.com), which is at least a well-defined behaviour.
 client = OpenAI(
-    api_key=API_KEY or "missing-api-key",
-    base_url=BASE_URL or None,
+    base_url="https://router.huggingface.co/v1",
+    api_key=os.environ["HF_TOKEN"],
 )