@tiens.nguyen/gonext-local-worker 1.0.124 → 1.0.125

package/gonext_agent_chat.py

@@ -1574,22 +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": "1.0.124",
+  "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",