From 4e8efa041de3b4c0d4493c17a8fc20e7de05ca52 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Tue, 18 Aug 2026 13:51:01 -0700
Subject: [PATCH] fix(advisor): exclude in-sequence system rows from the
 advisor sub-call context

---
 .../messages/interceptors/advisor.py          |  7 ++-
 .../messages/test_advisor_orchestration.py    | 53 +++++++++++++++++++
 2 files changed, 59 insertions(+), 1 deletion(-)

diff --git a/litellm/llms/anthropic/experimental_pass_through/messages/interceptors/advisor.py b/litellm/llms/anthropic/experimental_pass_through/messages/interceptors/advisor.py
index 6f732db1548..701211049db 100644
--- a/litellm/llms/anthropic/experimental_pass_through/messages/interceptors/advisor.py
+++ b/litellm/llms/anthropic/experimental_pass_through/messages/interceptors/advisor.py
@@ -302,6 +302,11 @@ def _build_advisor_context(
 
     tool_use blocks are excluded because Anthropic requires tool_use to be
     immediately followed by tool_result — not the advisor question.
+
+    In-sequence system rows (e.g. Claude Code SessionStart hook output) are
+    excluded: they are executor-directed, and a trailing one becomes invalid
+    once the question turn is appended after it (a system row must precede an
+    assistant message or end the array).
     """
     question: Final = (advisor_use_block.get("input") or {}).get("question") or (
         "Please provide guidance on the current task."
@@ -313,7 +318,7 @@ def _build_advisor_context(
         for block in raw_content
         if isinstance(block, dict) and block.get("type") == "text"
     ]
-    result: Final = list(messages)
+    result: Final = [m for m in messages if m.get("role") != "system"]
     if executor_text_blocks:
         result.append({"role": "assistant", "content": executor_text_blocks})
     result.append({"role": "user", "content": question})
diff --git a/tests/test_litellm/llms/anthropic/messages/test_advisor_orchestration.py b/tests/test_litellm/llms/anthropic/messages/test_advisor_orchestration.py
index 019eb4355c2..da5b5ac3867 100644
--- a/tests/test_litellm/llms/anthropic/messages/test_advisor_orchestration.py
+++ b/tests/test_litellm/llms/anthropic/messages/test_advisor_orchestration.py
@@ -1294,3 +1294,56 @@ async def test_advisor_sub_call_client_override_bypasses_router():
     assert len(advisor_sub_calls) == 1
     assert advisor_sub_calls[0]["api_key"] == "client-key"
     assert advisor_sub_calls[0]["api_base"] == "https://client.example.com"
+
+
+# ---------------------------------------------------------------------------
+# 16. In-sequence system rows (e.g. Claude Code SessionStart hook output) are
+#     excluded from the advisor sub-call context but kept for the executor: a
+#     trailing system row followed by the appended question turn is rejected
+#     upstream ("role 'system' must precede an 'assistant' message or end the
+#     array").
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.asyncio
+async def test_advisor_context_excludes_in_sequence_system_rows():
+    from litellm.llms.anthropic.experimental_pass_through.messages.interceptors.advisor import (
+        AdvisorOrchestrationHandler,
+    )
+
+    messages_with_system_row = [
+        *MESSAGES,
+        {"role": "system", "content": "SessionStart hook output: prefer functional style."},
+    ]
+
+    sub_calls = []
+
+    async def mock_call(model, messages, tools, stream, max_tokens, **kwargs):
+        sub_calls.append({"messages": messages, "tools": tools})
+        if len(sub_calls) == 1:
+            return _make_advisor_tool_use_response()
+        if tools is None:
+            return _make_text_response("Advice.", model="claude-opus-4-6")
+        return _make_text_response("Final answer.")
+
+    with patch(
+        "litellm.llms.anthropic.experimental_pass_through.messages.interceptors.advisor._call_messages_handler",
+        side_effect=mock_call,
+    ):
+        h = AdvisorOrchestrationHandler()
+        await h.handle(
+            model="openai/gpt-4o-mini",
+            messages=messages_with_system_row,
+            tools=[ADVISOR_TOOL],
+            stream=False,
+            max_tokens=512,
+            custom_llm_provider="openai",
+        )
+
+    assert len(sub_calls) == 3
+    advisor_messages = sub_calls[1]["messages"]
+    assert sub_calls[1]["tools"] is None
+    assert [m["role"] for m in advisor_messages if m["role"] == "system"] == []
+    assert advisor_messages[-1]["role"] == "user"
+    executor_roles = [m["role"] for m in sub_calls[0]["messages"]]
+    assert "system" in executor_roles