Parse <think> reasoning: show final answer, collapse reasoning, judge on answer

app.py

@@ -59,11 +59,10 @@ MODEL_ERROR: str | None = None
 MODEL_SETTINGS: dict[str, Any] = {}
 
 DEFAULT_SYSTEM = (
-    "You are First-Principle AI, a small reasoning model. Read every question "
-    "literally and reason from first principles. Watch for hidden assumptions, "
-    "misleading patterns, false premises, and familiar-looking riddles that have "
-    "been altered. State the correct answer first, then a one or two sentence "
-    "reason that names the trap."
+    "You are First-Principle AI, a small reasoning model. Reason briefly — a few "
+    "sentences at most — and watch for hidden assumptions, false premises, "
+    "misleading patterns, and familiar riddles that have been altered. Then end "
+    "with a clear final answer that names the trap in one line."
 )
 BLINDSPOT_SYSTEM = (
     "You are First-Principle AI acting as a blind-spot checker. The user describes "
@@ -583,9 +582,45 @@ VERDICT_TEXT = {
 }
 
 
-def _answer_md(trap: dict[str, Any], text: str, verdict: str) -> str:
+def _split_think(text: str) -> tuple[str, str]:
+    """Split a reasoning model's output into (reasoning, final answer)."""
+    text = (text or "").strip()
+    if "<think>" in text:
+        if "</think>" in text:
+            pre, rest = text.split("</think>", 1)
+            reasoning = pre.split("<think>", 1)[-1].strip()
+            answer = rest.strip()
+        else:
+            # generation was cut off before the model finished thinking
+            reasoning = text.split("<think>", 1)[-1].strip()
+            answer = ""
+    else:
+        reasoning, answer = "", text
+    # some chat templates prefix the final line with "Answer:"
+    return reasoning, answer
+
+
+def _render_reasoning(reasoning: str, answer: str) -> str:
+    parts: list[str] = []
+    if answer:
+        parts.append(answer)
+    else:
+        parts.append(
+            "_The model was still reasoning when it hit the token limit. "
+            "Raise **Max tokens** in the Engine Room for the full answer._"
+        )
+    if reasoning:
+        parts.append(
+            "\n\n<details><summary>🧠 Show first-principle reasoning</summary>\n\n"
+            f"{reasoning}\n\n</details>"
+        )
+    return "".join(parts)
+
+
+def _answer_md(trap: dict[str, Any], reasoning: str, answer: str, verdict: str) -> str:
     head = VERDICT_TEXT.get(verdict, VERDICT_TEXT["unclear"])
-    return f"{head}\n\n---\n\n**First-Principle AI says:**\n\n{text}"
+    body = _render_reasoning(reasoning, answer)
+    return f"{head}\n\n---\n\n**First-Principle AI says:**\n\n{body}"
 
 
 def _thinking_md() -> str:
@@ -673,10 +708,12 @@ def run_trap(
 
     try:
         text, meta = _generate(trap["prompt"], system_prompt, max_tokens, temperature, top_p, repeat_penalty)
-        verdict = judge(trap, text)
+        reasoning, answer = _split_think(text)
+        verdict = judge(trap, answer or reasoning)
         score = _bump(score, verdict)
     except Exception as exc:  # noqa: BLE001 - surface runtime issues in the UI
-        text = (
+        reasoning = ""
+        answer = (
             "Model load or inference failed.\n\n"
             f"```\n{exc}\n```\n\n"
             "The UI is live and the model artifact is published, but the runtime could not "
@@ -687,7 +724,7 @@ def run_trap(
 
     yield (
         trap_card_html(trap, revealed=True),
-        _answer_md(trap, text, verdict),
+        _answer_md(trap, reasoning, answer, verdict),
         scoreboard_html(score),
         _metrics_markdown(meta),
         score,