diff --git a/tests/v1/test_e2e.py b/tests/v1/test_e2e.py
index ee2dadd742..3737831cef 100644
--- a/tests/v1/test_e2e.py
+++ b/tests/v1/test_e2e.py
@@ -18,6 +18,80 @@ def pair(a: str, b: str, id: str, *extra_marks):
     return pytest.param(a, b, marks=[*marks, *extra_marks], id=id)
 
 
+@pytest.mark.asyncio
+async def test_chat_harness_preserves_streamed_reasoning():
+    import json
+
+    import httpx
+    from openai import AsyncOpenAI
+
+    from verifiers.v1.harnesses.utils.core import chat
+
+    def chunk(text: str, finish_reason: str | None = None) -> dict:
+        return {
+            "id": "chatcmpl-test",
+            "object": "chat.completion.chunk",
+            "created": 0,
+            "model": "test-model",
+            "choices": [
+                {
+                    "index": 0,
+                    "delta": {
+                        "role": "assistant",
+                        "reasoning": text,
+                        "reasoning_content": text,
+                        "reasoning_details": [
+                            {
+                                "type": "reasoning.text",
+                                "index": 0,
+                                "id": "r1",
+                                "format": "unknown",
+                                "signature": "sig",
+                                "text": text,
+                            }
+                        ],
+                    },
+                    "finish_reason": finish_reason,
+                }
+            ],
+        }
+
+    events = [chunk("Plan: "), chunk("call ls", "stop")]
+    content = "".join(f"data: {json.dumps(event)}\n\n" for event in events)
+    content += "data: [DONE]\n\n"
+
+    async def respond(request: httpx.Request) -> httpx.Response:
+        return httpx.Response(
+            200,
+            content=content,
+            headers={"content-type": "text/event-stream"},
+            request=request,
+        )
+
+    async with httpx.AsyncClient(transport=httpx.MockTransport(respond)) as http_client:
+        client = AsyncOpenAI(
+            api_key="test",
+            base_url="https://example.test/v1",
+            http_client=http_client,
+        )
+        completion = await chat(client, "test-model", [], [])
+
+    message = completion.choices[0].message.model_dump(exclude_none=True)
+    assert message["role"] == "assistant"
+    assert message["reasoning"] == "Plan: call ls"
+    assert message["reasoning_content"] == "Plan: call ls"
+    assert message["reasoning_details"] == [
+        {
+            "type": "reasoning.text",
+            "index": 0,
+            "id": "r1",
+            "format": "unknown",
+            "signature": "sig",
+            "text": "Plan: call ls",
+        }
+    ]
+
+
 # harness x harness runtime: every harness once, both local runtimes hit (subprocess
 # only carries the in-house loops — the rest NEEDS_CONTAINER), one remote row per
 # provider. codex/claude-code are excluded here (unreliable on a no-op echo chat
diff --git a/verifiers/v1/harnesses/utils/core.py b/verifiers/v1/harnesses/utils/core.py
index 743131c55c..bdd2793ff3 100644
--- a/verifiers/v1/harnesses/utils/core.py
+++ b/verifiers/v1/harnesses/utils/core.py
@@ -181,6 +181,57 @@ def run_edit(path: str, old_str: str, new_str: str) -> str:
     return f"Edited {path}"
 
 
+_STREAMED_MESSAGE_FIELDS = (
+    "role",
+    "reasoning",
+    "reasoning_content",
+    "reasoning_details",
+)
+
+
+def _accumulate_streamed_message(accumulated: dict, delta: dict) -> None:
+    """Accumulate message fields whose stream semantics differ from the SDK defaults."""
+    if role := delta.get("role"):
+        accumulated["role"] = role
+
+    for field_name in ("reasoning", "reasoning_content"):
+        if value := delta.get(field_name):
+            accumulated[field_name] = accumulated.get(field_name, "") + value
+
+    delta_details = delta.get("reasoning_details") or []
+    if not delta_details:
+        return
+    reasoning_details = accumulated.setdefault("reasoning_details", [])
+    for detail in delta_details:
+        previous = reasoning_details[-1] if reasoning_details else {}
+        detail_type = detail.get("type")
+        content_field = {
+            "reasoning.summary": "summary",
+            "reasoning.text": "text",
+        }.get(detail_type)
+        if (
+            content_field
+            and detail_type == previous.get("type")
+            and all(
+                previous.get(field_name) is None
+                or detail.get(field_name) is None
+                or previous[field_name] == detail[field_name]
+                for field_name in ("id", "index", "format")
+            )
+        ):
+            previous[content_field] = (previous.get(content_field) or "") + (
+                detail.get(content_field) or ""
+            )
+            for field_name in ("id", "index", "signature", "format"):
+                if (
+                    previous.get(field_name) is None
+                    and detail.get(field_name) is not None
+                ):
+                    previous[field_name] = detail[field_name]
+        else:
+            reasoning_details.append(dict(detail))
+
+
 async def chat(
     client: AsyncOpenAI,
     model: str,
@@ -201,9 +252,18 @@ async def chat(
         raw_stream=raw_stream, response_format=omit, input_tools=[]
     ) as response:
         completion = None
+        message_overrides: dict[int, dict] = {}
         async for event in response:
             if event.type == "chunk":
                 completion = event.snapshot
+                for choice in event.chunk.choices:
+                    delta = choice.delta.model_dump(exclude_none=True)
+                    if any(
+                        delta.get(field_name) for field_name in _STREAMED_MESSAGE_FIELDS
+                    ):
+                        _accumulate_streamed_message(
+                            message_overrides.setdefault(choice.index, {}), delta
+                        )
         if (
             completion is None
             or not completion.choices
@@ -211,9 +271,10 @@ async def chat(
         ):
             raise RuntimeError("model stream ended before a completion finished")
         for choice in completion.choices:
-            # Some providers repeat the role in each delta. The SDK concatenates
-            # these strings, but the role is metadata, not incremental content.
-            choice.message.role = "assistant"
+            overrides = message_overrides.setdefault(choice.index, {})
+            overrides.setdefault("role", "assistant")
+            for field_name, value in overrides.items():
+                setattr(choice.message, field_name, value)
         return completion