@tiens.nguyen/gonext-local-worker 1.0.138 → 1.0.140

package/gonext-local-worker.mjs

@@ -1526,14 +1526,15 @@ async function runAgentChatJob(job) {
       emailFrom: payload?.emailFrom ?? "",
       emailAllowList: payload?.emailAllowList ?? "",
     });
-    // 10 min max for an agent run: 5 ReAct steps on a 14B with a cold prompt cache —
-    // or a remote Ollama box cold-loading a big model — can exceed the old caps. Must
-    // stay BELOW the web client's hard WS deadline (660s) so our clean timeout error
-    // reaches the user instead of a generic WS timeout. On expiry the python child is
-    // SIGKILLed, which also shows up as a BrokenPipeError in the model server log
-    // (expected). Liveness during long model calls comes from the python heartbeat
-    // (a keepalive step chunk every 45s), not from this cap.
-    const timeoutMs = 600_000;
+    // 30 min max for an agent run: multi-step ReAct on a 14B/31B with a cold prompt
+    // cache — or a remote Ollama box on slow GPU cold-loading a big model — can run
+    // for many minutes per step. Must stay BELOW the web client's hard WS deadline
+    // (1_860_000 ms) so our clean timeout error reaches the user instead of a generic
+    // WS timeout. On expiry the python child is SIGKILLed, which also shows up as a
+    // BrokenPipeError in the model server log (expected). Liveness during long model
+    // calls comes from the python heartbeat (a keepalive step chunk every 45s), not
+    // from this cap.
+    const timeoutMs = 1_800_000;
 
     let inThink = false;
     let finalText = "";

package/gonext_agent_chat.py

@@ -398,6 +398,24 @@ def _list_model_ids(base_url, api_key=""):
             if isinstance(d, dict) and isinstance(d.get("id"), str) and d["id"].strip()]
 
 
+def _is_ollama_server(base_url):
+    """True if the OpenAI-compatible base URL is an Ollama server.
+
+    Ollama's root path answers GET / with the plain-text banner "Ollama is
+    running". Used to gate Ollama-only request params (reasoning_effort) that an
+    MLX server might reject.
+    """
+    root = re.sub(r"/v1/?$", "", (base_url or "").rstrip("/"))
+    if not root:
+        return False
+    req = urllib.request.Request(root, headers={"User-Agent": "gonext-agent/1.0"})
+    try:
+        with urllib.request.urlopen(req, timeout=6, context=_ssl_context()) as resp:
+            return b"ollama" in resp.read(200).lower()
+    except Exception:  # noqa: BLE001
+        return False
+
+
 def _detect_model_id(base_url, api_key=""):
     """Ask an OpenAI-compatible server which model it serves (first reported id).
 
@@ -1948,16 +1966,27 @@ def run_agent_chat(cfg):
             return ChatMessage(role=MessageRole.ASSISTANT, content=text)
 
     try:
+        # Ollama's OpenAI-compat endpoint ignores Qwen3's /no_think soft-switch (its
+        # template forces thinking mode; "think": false is also ignored there) but DOES
+        # honor reasoning_effort="none". Thinking burned ~430 of 487 output tokens in
+        # the live test — minutes per step on a slow GPU — so disable it explicitly.
+        # Sniffed per-server because MLX might reject the unknown param.
+        extra_model_kwargs = {}
+        if _is_ollama_server(coding_base_url):
+            _log("coding server is Ollama → sending reasoning_effort='none' to disable thinking")
+            extra_model_kwargs["reasoning_effort"] = "none"
         model = _LoggingModel(
             model_id=coding_model_id,
             api_base=coding_base_url,
             api_key=agent_api_key,
-            # Cap each HTTP attempt at 240s so one hung request (remote server
-            # cold-loading a big model) fails into OUR retry loop instead of eating
-            # the whole worker budget; by the retry the model is usually loaded.
+            # Cap each HTTP attempt at 600s so a genuinely slow single call (a
+            # remote M6000/Vulkan box doing a cold ~4-5min prompt-eval) can finish
+            # on the first try, while a truly hung request still fails into OUR
+            # retry loop rather than eating the whole 30min worker budget.
             # max_retries=0: the openai client's own silent retries would otherwise
             # stack multiplicatively with ours (3 × 3 attempts).
-            client_kwargs={"timeout": 240.0, "max_retries": 0},
+            client_kwargs={"timeout": 600.0, "max_retries": 0},
+            **extra_model_kwargs,
         )
         agent_tools = [http_request, web_search, fetch_url, calculate,
                        get_current_datetime, create_pdf]

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.138",
+  "version": "1.0.140",
   "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",