@tiens.nguyen/gonext-local-worker 1.0.123 → 1.0.125

package/gonext_agent_chat.py

@@ -904,20 +904,33 @@ def run_agent_chat(cfg):
     raw_coding_base = (cfg.get("codingBaseURL") or "").strip()
     raw_coding_model = (cfg.get("codingModelId") or "").strip()
     if raw_coding_base:
-        # A dedicated coding server is configured. If no model name was given,
-        # ask the server which model it serves (mlx_lm.server otherwise tries to
-        # download a mismatched name from HF and 404s).
-        detected = raw_coding_model or _detect_model_id(raw_coding_base, agent_api_key)
-        if detected:
-            coding_base_url = raw_coding_base
-            coding_model_id = detected
-        else:
+        same_server = raw_coding_base.rstrip("/") == (agent_base_url or "").rstrip("/")
+        if same_server and not raw_coding_model:
+            # The coding URL points at the SAME server as the chat model. Do NOT
+            # auto-detect: /v1/models can list several cached models and return a
+            # SMALLER one first (e.g. a 3B), silently downgrading the agent's reasoning.
+            # Reuse the chat model id directly — it's the model actually loaded here.
             _log(
-                f"coding model id unresolved for {raw_coding_base!r}; "
-                "falling back to chat model"
+                f"coding base == chat base ({raw_coding_base}); "
+                f"reusing chat model {agent_model_id!r} (skipping auto-detect)"
             )
             coding_base_url = agent_base_url
             coding_model_id = agent_model_id
+        else:
+            # A DIFFERENT dedicated coding server. If no model name was given, ask the
+            # server which model it serves (mlx_lm.server otherwise tries to download a
+            # mismatched name from HF and 404s).
+            detected = raw_coding_model or _detect_model_id(raw_coding_base, agent_api_key)
+            if detected:
+                coding_base_url = raw_coding_base
+                coding_model_id = detected
+            else:
+                _log(
+                    f"coding model id unresolved for {raw_coding_base!r}; "
+                    "falling back to chat model"
+                )
+                coding_base_url = agent_base_url
+                coding_model_id = agent_model_id
     else:
         coding_base_url = agent_base_url
         coding_model_id = agent_model_id
@@ -1187,15 +1200,13 @@ def run_agent_chat(cfg):
         "\n"
         "http_request RETURN FORMAT: 'HTTP 200\\n{body}' — first line is 'HTTP <code>', body follows.\n"
         "\n"
-        "BASIC AUTH — ALWAYS use username= and password=, NEVER construct headers manually:\n"
-        "  response = http_request('GET', 'https://api.example.com/data',\n"
-        "                          username='alice@example.com', password='secret123')\n"
-        "  final_answer(response)\n"
-        "The function handles base64 encoding automatically. NEVER write 'Basic ' + anything.\n"
-        "\n"
-        "BEARER TOKEN — use headers:\n"
-        "  response = http_request('GET', url, headers='{\"Authorization\": \"Bearer TOKEN\"}')\n"
-        "  final_answer(response)\n"
+        "AUTH — ONLY when the TASK itself provides credentials or a token. If it does "
+        "NOT, call http_request with NO username/password/headers. NEVER invent or copy "
+        "the placeholder values below.\n"
+        "  BASIC AUTH — pass username= and password= (auto base64; never build a 'Basic ' header):\n"
+        "      http_request('GET', url, username=<username from the task>, password=<password from the task>)\n"
+        "  BEARER TOKEN — pass it in headers:\n"
+        "      http_request('GET', url, headers='{\"Authorization\": \"Bearer <token from the task>\"}')\n"
         "\n"
         "CHOOSING A TOOL (match the TASK, not these examples):\n"
         "- ONLY a date/time question (e.g. 'what is the date today') -> get_current_datetime().\n"
@@ -1203,6 +1214,8 @@ def run_agent_chat(cfg):
         "knowledge -> web_search(query).\n"
         "- 'read this page' / 'open this link' / 'what does <URL> say' -> fetch_url(url).\n"
         "- any arithmetic, percentage, or 'how much is' math -> calculate(expression).\n"
+        "- a live/current NUMBER with no URL given (price, exchange rate, weather, score, "
+        "stock/crypto) -> web_search(query). Do NOT guess an API URL for these.\n"
         "- 'create/make/generate/export a PDF' of some text/data -> create_pdf(text, title).\n"
         + _pdf_read_choose_line + _email_choose_line +
         "- A specific known API/URL was given -> http_request().\n"
@@ -1212,8 +1225,11 @@ def run_agent_chat(cfg):
         "If nothing works, call final_answer explaining what you need — do NOT make up an answer.\n"
         "- Only report what a tool ACTUALLY returned. Never fabricate a response, body, or status code.\n"
         "- Pass an http_request response DIRECTLY to final_answer — do NOT split, parse, or index it.\n"
-        "- If a response starts with 'HTTP 2' it SUCCEEDED — call final_answer immediately.\n"
-        "- If a tool returns 'Error:' or HTTP 4xx/5xx, try a DIFFERENT approach, not the same URL.\n"
+        "- When a call's result is UNCERTAIN (any network call), do NOT call final_answer "
+        "in the SAME block. Call the tool alone, READ the Observation, THEN answer next step.\n"
+        "- If a response starts with 'HTTP 2' it SUCCEEDED — answer with it next step.\n"
+        "- NEVER pass an 'Error:' / HTTP 4xx / 5xx result to final_answer. On an error, try a "
+        "DIFFERENT approach next step (e.g. web_search); only give up after a few tries.\n"
         "- Do NOT put final_answer outside the code block.\n\n"
     )
     # Lead with the TASK so the weak model anchors on what's actually being asked —
@@ -1558,22 +1574,58 @@ def run_agent_chat(cfg):
     # literal messages array sent to /v1/chat/completions.
     class _LoggingModel(OpenAIServerModel):
         def generate(self, *args, **kwargs):
-            # Strip Qwen3 <think>…</think> reasoning traces from each step's reply
-            # BEFORE smolagents parses it for the code block / final_answer. Thinking
-            # stays on (better reasoning), but the trace never confuses the parser or
-            # leaks to the user.
+            # Two safeguards on each step's reply BEFORE smolagents parses it for the
+            # code block / final_answer:
+            #   1. content=None guard — Qwen3 (and reasoning models generally) can return
+            #      message.content=None on a turn (e.g. it spent the whole completion on a
+            #      <think> trace and never emitted a code block). smolagents has NO null
+            #      guard (models.py: `content = response.choices[0].message.content`) and
+            #      feeds it straight into parse_code_blobs → `re` gets None →
+            #      "expected string or bytes-like object, got 'NoneType'" → the whole
+            #      loop spirals on parse errors and trips the WebSocket timeout. Coercing
+            #      None→"" turns that fatal crash into a normal "no code block" retry.
+            #   2. strip any leftover Qwen3 <think>…</think> trace (belt-and-suspenders;
+            #      thinking is disabled via /no_think in _prepare_completion_kwargs, but a
+            #      stray trace must never reach the parser or leak to the user).
             msg = super().generate(*args, **kwargs)
             try:
-                if isinstance(getattr(msg, "content", None), str) and "</think>" in msg.content.lower():
-                    cleaned = _strip_think(msg.content)
-                    _log(f"stripped <think> trace ({len(msg.content) - len(cleaned)} chars)")
+                content = getattr(msg, "content", None)
+                if content is None:
+                    _log("model returned content=None (empty/thinking-only turn) → coercing to ''")
+                    msg.content = ""
+                elif isinstance(content, str) and "</think>" in content.lower():
+                    cleaned = _strip_think(content)
+                    _log(f"stripped <think> trace ({len(content) - len(cleaned)} chars)")
                     msg.content = cleaned
             except Exception as e:  # noqa: BLE001
-                _log(f"think-strip error: {e}")
+                _log(f"content-normalize error: {e}")
             return msg
 
         def _prepare_completion_kwargs(self, *args, **kwargs):
             ck = super()._prepare_completion_kwargs(*args, **kwargs)
+            # Disable Qwen3 "thinking" for the agent loop. Live runs showed thinking-on
+            # spends the whole completion budget on a <think> trace and returns
+            # content=None (no code block) → parse-error spiral → WS timeout. The
+            # `/no_think` soft switch (recognized by the Qwen3 chat template) forces a
+            # direct Thought+code reply every step. Injected into the SYSTEM message so it
+            # applies to EVERY step's request, not just the first user turn.
+            try:
+                msgs = ck.get("messages", []) or []
+                for m in msgs:
+                    role = m.get("role") if isinstance(m, dict) else getattr(m, "role", None)
+                    if str(role).endswith("system") or role == "system":
+                        c = m.get("content") if isinstance(m, dict) else getattr(m, "content", None)
+                        if isinstance(c, str) and "/no_think" not in c:
+                            m["content"] = c + "\n\n/no_think"
+                        elif isinstance(c, list):
+                            for part in c:
+                                if isinstance(part, dict) and isinstance(part.get("text"), str) \
+                                        and "/no_think" not in part["text"]:
+                                    part["text"] += "\n\n/no_think"
+                                    break
+                        break
+            except Exception as e:  # noqa: BLE001
+                _log(f"/no_think inject error: {e}")
             try:
                 msgs = ck.get("messages", []) or []
                 _log(f"=== MODEL REQUEST: {len(msgs)} message(s) sent to the model ===")

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tiens.nguyen/gonext-local-worker",
-  "version": "1.0.123",
+  "version": "1.0.125",
   "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",