@tiens.nguyen/gonext-local-worker 1.0.182 → 1.0.184

package/gonext_agent_chat.py

@@ -649,6 +649,37 @@ _COMPACT_CODE_SYSTEM_PROMPT = (
 )
 
 
+def _is_model_not_found_err(e) -> bool:
+    """True when a chat completion failed because the server didn't recognize the model
+    name (mlx_lm.server treats an unknown name as an HF repo to fetch → 'Repository Not
+    Found' / 401 / 404), as opposed to a real network/auth failure."""
+    s = str(e).lower()
+    return ("repository not found" in s or "not found for url" in s
+            or "repo_id" in s or "does not exist" in s
+            or ("404" in s and "model" in s))
+
+
+def _chat_create(client, **kwargs):
+    """chat.completions.create with an mlx_lm safety net: if the server rejects the model
+    NAME as an unknown repo, retry once with 'default_model' — mlx_lm.server maps that
+    sentinel to whatever model it was launched with (`--model <path>`), so a server
+    started as e.g. `mlx_lm.server --model ~/mlx-models/Qwen3-14B-4bit` answers even when
+    we only know the bare name. Non-mlx servers surface their ORIGINAL error unchanged."""
+    try:
+        return client.chat.completions.create(**kwargs)
+    except Exception as e:  # noqa: BLE001
+        if kwargs.get("model") != "default_model" and _is_model_not_found_err(e):
+            _log(f"model {kwargs.get('model')!r} not found on server; retrying as "
+                 "'default_model' (mlx_lm preloaded model)")
+            try:
+                kwargs2 = dict(kwargs)
+                kwargs2["model"] = "default_model"
+                return client.chat.completions.create(**kwargs2)
+            except Exception:  # noqa: BLE001
+                raise e  # surface the clearer original error, not the retry's
+        raise
+
+
 def _route(task_text: str, base_url: str, api_key: str, model_id: str) -> bool:
     """Decide if the task needs the HTTP agent (True) or a plain chat reply (False).
 
@@ -668,7 +699,8 @@ def _route(task_text: str, base_url: str, api_key: str, model_id: str) -> bool:
         from openai import OpenAI
         client = OpenAI(base_url=base_url, api_key=api_key or "local",
                         max_retries=0, timeout=20)
-        resp = client.chat.completions.create(
+        resp = _chat_create(
+            client,
             model=model_id,
             messages=[
                 {"role": "system", "content": (
@@ -716,7 +748,8 @@ def _summarize_result(task_text: str, agent_output: str,
         from openai import OpenAI
         client = OpenAI(base_url=base_url, api_key=api_key or "local",
                         max_retries=0, timeout=30)
-        resp = client.chat.completions.create(
+        resp = _chat_create(
+            client,
             model=model_id,
             messages=[
                 {"role": "system", "content": (
@@ -773,7 +806,8 @@ def _plain_reply(messages: list, base_url: str, api_key: str, model_id: str,
             from openai import OpenAI
             client = OpenAI(base_url=b, api_key=api_key or "local",
                             max_retries=0, timeout=60)
-            resp = client.chat.completions.create(
+            resp = _chat_create(
+                client,
                 model=mid,
                 messages=chat_messages,
                 temperature=0.7,
@@ -783,8 +817,9 @@ def _plain_reply(messages: list, base_url: str, api_key: str, model_id: str,
         except Exception as e:  # noqa: BLE001
             last_err = e
             if i + 1 < len(targets):
-                _log(f"plain reply: primary model failed ({e}); retrying on fallback "
-                     f"{mid} @ {b}")
+                nb, nmid = targets[i + 1]  # the model we're about to try next
+                _log(f"plain reply: {mid} @ {b} failed ({e}); "
+                     f"falling back to {nmid} @ {nb}")
     return f"[Error: {last_err}]"
 
 
@@ -810,7 +845,8 @@ def _synthesize_document(gathered: list, task: str, base_url: str, api_key: str,
         from openai import OpenAI
         client = OpenAI(base_url=base_url, api_key=api_key or "local",
                         max_retries=0, timeout=120)
-        resp = client.chat.completions.create(
+        resp = _chat_create(
+            client,
             model=model_id,
             messages=[{"role": "system", "content": sys_prompt},
                       {"role": "user", "content": user_prompt}],
@@ -1083,7 +1119,8 @@ def _format_text_for_pdf(text: str, title: str, base_url: str, api_key: str,
             + (f"Requested change: {instruction}\n\n" if instruction else "")
             + f"Content to format:\n{text}"
         )
-        resp = client.chat.completions.create(
+        resp = _chat_create(
+            client,
             model=model_id,
             messages=[
                 {"role": "system", "content": system_content},

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.182",
+  "version": "1.0.184",
   "description": "Polls GoNext cloud API for async local LLM jobs and runs them against Ollama/OpenAI-compatible servers on this Mac",
   "type": "module",
   "license": "MIT",