spine-probes: support thinking models in run_probes and judge

scripts/spine-probes/judge.py

@@ -46,6 +46,10 @@ def grade(url, model, token, request_text, reply, timeout):
                 f"USER REQUEST:\n{request_text}\n\nASSISTANT REPLY:\n{reply}\n\nGrade:"},
         ],
         "max_tokens": 8, "temperature": 0,
+        # A thinking judge spends all 8 tokens inside the reasoning block and
+        # returns empty content, so every verdict scores UNPARSED. Ask it not
+        # to think rather than paying for a bigger budget.
+        "chat_template_kwargs": {"enable_thinking": False},
     }).encode()
     headers = {"Content-Type": "application/json"}
     if token:

scripts/spine-probes/run_probes.py

@@ -112,7 +112,10 @@ def ask(url, model, system, user, timeout, temperature, max_tokens=900):
     with urllib.request.urlopen(req, timeout=timeout) as r:
         d = json.load(r)
     m = d["choices"][0]["message"]
-    return (m.get("content") or ""), (m.get("reasoning_content") or "")
+    # Some OpenAI-compatible servers expose thinking on `reasoning` rather
+    # than `reasoning_content` (our vLLM Qwen NVFP4 lane has no
+    # `reasoning_content` key at all, so the reasoning column read 0).
+    return (m.get("content") or ""), (m.get("reasoning_content") or m.get("reasoning") or "")
 
 
 def main():