Skip to content

Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.

Fix tool_result ordering: no stray user message between tool_calls and tool

chopratejas7d25c872868c
diff --git a/headroom/backends/litellm.py b/headroom/backends/litellm.py
index c64daa56f..c2c81514c 100644
--- a/headroom/backends/litellm.py
+++ b/headroom/backends/litellm.py
@@ -511,9 +511,11 @@ class LiteLLMBackend(Backend):
# tool_result blocks → OpenAI "tool" role messages
if tool_result_blocks:
- # Include any text from the user message first
- if text_parts:
- converted.append({"role": "user", "content": "\n".join(text_parts)})
+ # Do NOT insert a separate user text message here — Bedrock
+ # requires tool role messages to appear immediately after the
+ # assistant tool_calls message with no intervening messages.
+ # Any text alongside tool_result is discarded (Claude Code
+ # doesn't send text with tool_result blocks in practice).
for tr in tool_result_blocks:
tr_content = tr.get("content", "")
if isinstance(tr_content, list):
diff --git a/tests/test_backend_bugs.py b/tests/test_backend_bugs.py
index a999c0d84..b04f27e1f 100644
--- a/tests/test_backend_bugs.py
+++ b/tests/test_backend_bugs.py
@@ -324,6 +324,78 @@ class TestConvertMessagesToolBlocks:
assert converted[1]["role"] == "tool"
assert converted[1]["tool_call_id"] == "toolu_b"
+ def test_tool_result_immediately_follows_tool_calls(self):
+ """Bedrock requires role=tool immediately after assistant tool_calls — no intervening messages.
+
+ Regression test for GitHub issue #70: a stray user text message was inserted
+ between the assistant tool_calls and the tool results, causing Bedrock to reject
+ the request with 'tool_use ids were found without tool_result blocks immediately after'.
+ """
+ backend = self._make_backend()
+ messages = [
+ {"role": "user", "content": "What's the weather in Paris and Tokyo?"},
+ {
+ "role": "assistant",
+ "content": [
+ {
+ "type": "tool_use",
+ "id": "toolu_01",
+ "name": "get_weather",
+ "input": {"city": "Paris"},
+ },
+ {
+ "type": "tool_use",
+ "id": "toolu_02",
+ "name": "get_weather",
+ "input": {"city": "Tokyo"},
+ },
+ ],
+ },
+ {
+ "role": "user",
+ "content": [
+ {"type": "tool_result", "tool_use_id": "toolu_01", "content": "Sunny, 22C"},
+ {"type": "tool_result", "tool_use_id": "toolu_02", "content": "Rainy, 18C"},
+ ],
+ },
+ ]
+ converted = backend._convert_messages_for_litellm(messages)
+
+ # Find the assistant message with tool_calls
+ assistant_idx = next(i for i, m in enumerate(converted) if m.get("tool_calls"))
+
+ # Every message after the assistant tool_calls must be role=tool
+ # with no intervening user/assistant messages
+ for i in range(assistant_idx + 1, len(converted)):
+ assert converted[i]["role"] == "tool", (
+ f"Message at index {i} has role={converted[i]['role']!r}, "
+ f"expected 'tool' — Bedrock requires tool results immediately "
+ f"after assistant tool_calls with no intervening messages"
+ )
+
+ def test_tool_result_with_text_does_not_insert_user_message(self):
+ """Text alongside tool_result should NOT produce a separate user message.
+
+ Bedrock rejects any message between assistant tool_calls and tool results.
+ """
+ backend = self._make_backend()
+ messages = [
+ {
+ "role": "user",
+ "content": [
+ {"type": "text", "text": "Here are the results:"},
+ {"type": "tool_result", "tool_use_id": "toolu_01", "content": "42"},
+ ],
+ },
+ ]
+ converted = backend._convert_messages_for_litellm(messages)
+
+ # Should only have the tool message, no user text message
+ assert len(converted) == 1
+ assert converted[0]["role"] == "tool"
+ assert converted[0]["tool_call_id"] == "toolu_01"
+ assert converted[0]["content"] == "42"
+
# =============================================================================
# Streaming tool_calls (GitHub Issue — Bug 1)