Merge 737f2dce41 into 2dccc0dc79

litellm/proxy/guardrails/guardrail_hooks/headroom/headroom.py

@@ -88,17 +88,22 @@ def _flatten_messages_for_compression(messages: list[dict[str, object]]) -> list
     at its part), so merging text across a non-text part would move a later
     breakpoint to the other side of it. Rows with non-text parts are sent
     unchanged and pass through the service untouched.
+
+    An explicit ``"tool_calls": null`` (what OpenAI SDK message objects
+    serialize to) is dropped: the service iterates the field and older
+    releases fail on ``None``.
     """
-    flattened: Final[list[dict[str, object]]] = []
-    for msg in messages:
-        content = msg.get("content")
-        if is_all_text_parts(content):
-            text = content_to_text(content)
-            if text:
-                flattened.append({**msg, "content": text})
-                continue
-        flattened.append(msg)
-    return flattened
+    return [_compression_row(msg) for msg in messages]
+
+
+def _compression_row(message: Mapping[str, object]) -> dict[str, object]:
+    content: Final = message.get("content")
+    text: Final = content_to_text(content) if is_all_text_parts(content) else ""
+    return {
+        key: text if key == "content" and text else value
+        for key, value in message.items()
+        if key != "tool_calls" or value is not None
+    }
 
 
 def _restore_content_shapes(

tests/test_litellm/proxy/guardrails/guardrail_hooks/test_headroom.py

@@ -2008,6 +2008,41 @@ async def test_apply_guardrail_sends_textless_parts_rows_unflattened(
     assert wire_messages[1]["content"] == "D" * 5000
 
 
+@pytest.mark.asyncio
+async def test_apply_guardrail_drops_null_tool_calls_but_keeps_real_ones(
+    guardrail: HeadroomGuardrail,
+):
+    real_tool_calls = [{"id": "call_1", "type": "function", "function": {"name": "lookup", "arguments": "{}"}}]
+    history = [
+        {"role": "user", "content": "E" * 5000},
+        {"role": "assistant", "content": None, "tool_calls": real_tool_calls},
+        {"role": "tool", "tool_call_id": "call_1", "content": "result"},
+        {"role": "assistant", "content": "summary", "tool_calls": None, "function_call": None},
+    ]
+    inputs = GenericGuardrailAPIInputs(
+        texts=["E" * 5000],
+        structured_messages=json.loads(json.dumps(history))
+        + [{"role": "assistant", "content": "last turn"}, {"role": "user", "content": "and now?"}],
+    )
+    mock_response = _make_compress_response(json.loads(json.dumps(history)))
+
+    with patch.object(
+        guardrail.async_handler,
+        "post",
+        new_callable=AsyncMock,
+        return_value=mock_response,
+    ) as mock_post:
+        await guardrail.apply_guardrail(
+            inputs=inputs,
+            request_data={"model": "claude-fable-5"},
+            input_type="request",
+        )
+
+    wire_messages = mock_post.call_args.kwargs["json"]["messages"]
+    assert wire_messages[1]["tool_calls"] == real_tool_calls
+    assert wire_messages[3] == {"role": "assistant", "content": "summary", "function_call": None}
+
+
 @pytest.mark.asyncio
 async def test_fail_open_returns_original_parts_shapes():
     guardrail = _make_guardrail(unreachable_fallback="fail_open")