From b87f9c96b55e7ff0e7ea0e19d350ad14e2e3118b Mon Sep 17 00:00:00 2001
From: Aryan Rajesh Ganjoo <51299383+AryanRG@users.noreply.github.com>
Date: Fri, 16 Oct 2026 15:28:50 +0530
Subject: [PATCH] fix: allow explicit thread render overrides

---
 py/autoevals/llm.py      |  6 ++-
 py/autoevals/test_llm.py | 87 ++++++++++++++++++++++++++++++++++++++++
 2 files changed, 91 insertions(+), 2 deletions(-)

diff --git a/py/autoevals/llm.py b/py/autoevals/llm.py
index 1a38712..7693eb8 100644
--- a/py/autoevals/llm.py
+++ b/py/autoevals/llm.py
@@ -460,7 +460,8 @@ def _request_args(self, output, expected, **kwargs):
             thread_vars = self._compute_thread_vars_sync(trace)
 
         # Thread vars come first so explicit render args can override.
-        return super()._request_args(output, expected, **thread_vars, **kwargs)
+        render_args = {**thread_vars, **kwargs}
+        return super()._request_args(output, expected, **render_args)
 
     async def _request_args_async(self, output, expected, **kwargs):
         trace = kwargs.get("trace")
@@ -469,7 +470,8 @@ async def _request_args_async(self, output, expected, **kwargs):
             thread_vars = await self._compute_thread_vars_async(trace)
 
         # Thread vars come first so explicit render args can override.
-        return super()._request_args(output, expected, **thread_vars, **kwargs)
+        render_args = {**thread_vars, **kwargs}
+        return super()._request_args(output, expected, **render_args)
 
     @classmethod
     def from_spec(cls, name: str, spec: ModelGradedSpec, client: Client | None = None, **kwargs):
diff --git a/py/autoevals/test_llm.py b/py/autoevals/test_llm.py
index 2bcecef..29e8db0 100644
--- a/py/autoevals/test_llm.py
+++ b/py/autoevals/test_llm.py
@@ -188,6 +188,93 @@ async def test_llm_classifier_request_args_async_keep_thread_filtered_and_thread
     assert system_marker in full_thread
 
 
+def test_llm_classifier_request_args_explicit_thread_overrides():
+    system_marker = "PY_AUTOEVALS_SYSTEM_MARKER"
+    trace = _FakeTrace(
+        [
+            {"role": "system", "content": system_marker},
+            {"role": "user", "content": "Hello"},
+            {"role": "assistant", "content": "Hi there"},
+        ]
+    )
+    classifier = LLMClassifier(
+        "test",
+        "Thread:\n{{thread}}\nCount: {{thread_count}}\n\nFull thread:\n{{thread_with_system}}",
+        {"Yes": 1, "No": 0},
+        use_cot=False,
+    )
+
+    request_args = classifier._request_args(
+        output="",
+        expected="",
+        trace=trace,
+        thread=[{"role": "user", "content": "EXPLICIT_THREAD_SENTINEL"}],
+        thread_count=99,
+    )
+    rendered_prompt = request_args["messages"][0]["content"]
+
+    overridden_thread, full_thread = rendered_prompt.split("\n\nFull thread:\n", 1)
+    assert "EXPLICIT_THREAD_SENTINEL" in overridden_thread
+    assert "Count: 99" in overridden_thread
+    assert "Hello" not in overridden_thread
+    assert system_marker in full_thread
+
+
+@pytest.mark.asyncio
+async def test_llm_classifier_request_args_async_explicit_thread_overrides():
+    system_marker = "PY_AUTOEVALS_SYSTEM_MARKER_ASYNC"
+    trace = _FakeTrace(
+        [
+            {"role": "system", "content": system_marker},
+            {"role": "user", "content": "Hello"},
+            {"role": "assistant", "content": "Hi there"},
+        ]
+    )
+    classifier = LLMClassifier(
+        "test",
+        "Thread:\n{{thread}}\nCount: {{thread_count}}\n\nFull thread:\n{{thread_with_system}}",
+        {"Yes": 1, "No": 0},
+        use_cot=False,
+    )
+
+    request_args = await classifier._request_args_async(
+        output="",
+        expected="",
+        trace=trace,
+        thread=[{"role": "user", "content": "EXPLICIT_THREAD_SENTINEL"}],
+        thread_count=99,
+    )
+    rendered_prompt = request_args["messages"][0]["content"]
+
+    overridden_thread, full_thread = rendered_prompt.split("\n\nFull thread:\n", 1)
+    assert "EXPLICIT_THREAD_SENTINEL" in overridden_thread
+    assert "Count: 99" in overridden_thread
+    assert "Hello" not in overridden_thread
+    assert system_marker in full_thread
+
+
+def test_llm_classifier_request_args_explicit_thread_count_overrides():
+    trace = _FakeTrace(
+        [
+            {"role": "user", "content": "Hello"},
+            {"role": "assistant", "content": "Hi there"},
+        ]
+    )
+    classifier = LLMClassifier(
+        "test",
+        "Thread:\n{{thread}}\nCount: {{thread_count}}",
+        {"Yes": 1, "No": 0},
+        use_cot=False,
+    )
+
+    request_args = classifier._request_args(output="", expected="", trace=trace, thread_count=99)
+    rendered_prompt = request_args["messages"][0]["content"]
+
+    assert "Count: 99" in rendered_prompt
+    assert "Hello" in rendered_prompt
+    assert "Hi there" in rendered_prompt
+
+
 def test_openai():
     e = OpenAILLMClassifier(
         "title",