debug: raw decode

app.py

@@ -120,8 +120,9 @@ def generate(system, user, max_tokens, temperature):
     try:
         inputs = _build_inputs(system, user)
         out = _model.generate(**_gen_kwargs(inputs, max_tokens, temperature))
-        text = _tok.decode(out[0][_prompt_len(inputs):], skip_special_tokens=True)
-        return text.strip() or "[EMPTY OUTPUT]"
+        # DEBUG: raw decode (special tokens kept) to inspect the reasoning/response structure.
+        raw = _tok.decode(out[0][_prompt_len(inputs):], skip_special_tokens=False)
+        return "[RAW]\n" + raw
     except Exception:  # noqa: BLE001
         import traceback
         return "[ERROR]\n" + traceback.format_exc()