diff --git a/src/trinity/llm/openai_compatible_pool.py b/src/trinity/llm/openai_compatible_pool.py
index ea33d32..f10b5b7 100644
--- a/src/trinity/llm/openai_compatible_pool.py
+++ b/src/trinity/llm/openai_compatible_pool.py
@@ -96,6 +96,15 @@ def _parse_completion(data: dict, model: str) -> ChatResult:
     completion (``text=""``, ``finish_reason="error"``), mirroring the existing
     null-``content`` handling so a single odd reply degrades one turn instead of
     crashing the run. See JOURNAL 2026-07-08.
+
+    Reasoning models (e.g. the DeepSeek / GLM / Kimi "thinking" variants served
+    over OpenAI-compatible APIs) can return the answer in a separate
+    ``reasoning_content`` channel with ``content`` null or empty — most commonly
+    when the visible channel is cut off by the token cap
+    (``finish_reason="length"``). When ``content`` carries no text we fall back to
+    ``reasoning_content`` so that answer stays scoreable instead of collapsing the
+    turn to ``text=""`` (which zeroes the reward). A non-empty ``content`` always
+    wins, so ordinary completions are unaffected.
     """
     usage = data.get("usage") or {}
     prompt_tokens = int(usage.get("prompt_tokens", 0) or 0)
@@ -113,6 +122,8 @@ def _parse_completion(data: dict, model: str) -> ChatResult:
     choice = choices[0] or {}
     message = choice.get("message") or {}
     content = message.get("content")
+    if content is None or content == "":
+        content = message.get("reasoning_content")
     return ChatResult(
         model=model,
         text="" if content is None else str(content),
diff --git a/tests/test_pool_parse.py b/tests/test_pool_parse.py
index 5e8e682..40dcd40 100644
--- a/tests/test_pool_parse.py
+++ b/tests/test_pool_parse.py
@@ -64,3 +64,45 @@ def test_non_string_content_is_stringified():
     data = {"choices": [{"message": {"content": 42}, "finish_reason": "stop"}]}
     res = _parse_completion(data, "m")
     assert res.text == "42"
+
+
+def test_null_content_falls_back_to_reasoning_content():
+    # Reasoning models can put the answer in reasoning_content with content=null,
+    # commonly when the visible channel is truncated by the token cap.
+    data = {
+        "choices": [
+            {
+                "message": {"content": None, "reasoning_content": r"...so \boxed{42}"},
+                "finish_reason": "length",
+            }
+        ]
+    }
+    res = _parse_completion(data, "m")
+    assert res.text == r"...so \boxed{42}"
+    assert res.finish_reason == "length"
+
+
+def test_empty_content_falls_back_to_reasoning_content():
+    data = {
+        "choices": [
+            {"message": {"content": "", "reasoning_content": "the answer is B"}, "finish_reason": "stop"}
+        ]
+    }
+    res = _parse_completion(data, "m")
+    assert res.text == "the answer is B"
+
+
+def test_non_empty_content_wins_over_reasoning_content():
+    data = {
+        "choices": [
+            {"message": {"content": "final answer", "reasoning_content": "scratch work"}}
+        ]
+    }
+    res = _parse_completion(data, "m")
+    assert res.text == "final answer"
+
+
+def test_null_content_without_reasoning_content_stays_empty():
+    data = {"choices": [{"message": {"content": None}, "finish_reason": "stop"}]}
+    res = _parse_completion(data, "m")
+    assert res.text == ""