fix(llm): when using action (such as look around), remove the thinking

block and increased token limit (from 512 to 1024) to make sure the full
asnwer won't be truncated anymore.

visualnovel/llm.py

@@ -123,8 +123,23 @@ class TransformersLLM:
         print(f"[llm] {config.LLM_REPO} ready")
         self.torch = torch
 
+    def _apply_template(self, messages: list[dict[str, str]], enable_thinking: bool = True) -> str:
+        """Apply chat template, disabling Qwen3 thinking mode when asked."""
+        try:
+            return self.tok.apply_chat_template(
+                messages,
+                tokenize=False,
+                add_generation_prompt=True,
+                enable_thinking=enable_thinking,
+            )
+        except TypeError:
+            # Tokenizer doesn't support enable_thinking (non-Qwen3 models)
+            return self.tok.apply_chat_template(
+                messages, tokenize=False, add_generation_prompt=True
+            )
+
     def complete(self, messages: list[dict[str, str]], **kw: Any) -> str:
-        text = self.tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+        text = self._apply_template(messages)
         ids = self.tok(text, return_tensors="pt").to(self.model.device)
         out = self.model.generate(**ids, max_new_tokens=kw.get("max_tokens", 512))
         return self.tok.decode(out[0][ids.input_ids.shape[1] :], skip_special_tokens=True)
@@ -163,14 +178,19 @@ class TransformersLLM:
             {**m, "content": m["content"] + hint} if m["role"] == "system" else m for m in messages
         ]
         for attempt in range(3):
-            raw = self.complete(augmented, **kw)
-            # Strip Qwen3 <think>...</think> reasoning blocks before JSON extraction
-            text = re.sub(r"<think>.*?</think>", "", raw, flags=re.DOTALL)
-            text = re.sub(r"```(?:json)?\s*", "", text).strip()
-            start, end = text.find("{"), text.rfind("}") + 1
+            # Disable Qwen3 thinking mode: <think> blocks waste tokens and can truncate the JSON
+            text = self._apply_template(augmented, enable_thinking=False)
+            ids = self.tok(text, return_tensors="pt").to(self.model.device)
+            max_new = kw.get("max_tokens", 512)
+            out = self.model.generate(**ids, max_new_tokens=max_new)
+            raw = self.tok.decode(out[0][ids.input_ids.shape[1] :], skip_special_tokens=True)
+            # Still strip any residual think tags (safety net)
+            raw = re.sub(r"<think>.*?</think>", "", raw, flags=re.DOTALL)
+            raw = re.sub(r"```(?:json)?\s*", "", raw).strip()
+            start, end = raw.find("{"), raw.rfind("}") + 1
             if start != -1 and end > start:
                 try:
-                    return json.loads(text[start:end])
+                    return json.loads(raw[start:end])
                 except json.JSONDecodeError as exc:
                     print(f"[llm] JSON parse error (attempt {attempt + 1}/3): {exc}")
         print("[llm] WARNING: all 3 retries failed, using empty schema fallback")

visualnovel/orchestrator.py

@@ -126,7 +126,7 @@ def direct_turn(
         schema=prompts.directive_schema(),
         temperature=0.7,
         top_p=0.9,
-        max_tokens=512,
+        max_tokens=1024,
     )
     out = DirectorOutput.model_validate(raw)
     return _repair(state, out)