BerriAI/litellm · error · BadRequestError

BedrockException - . Enable 'litellm.modify_params=True'…

Error message

BedrockException - {error_str}
. Enable 'litellm.modify_params=True' (for PROXY do: `litellm_settings::modify_params: True`) to insert a dummy assistant message and fix this error.

What it means

LiteLLM throws BadRequestError when Bedrock's Converse API rejects the message list because conversation blocks and tool result blocks appear in the same turn (a user turn carries tool results while assistant content is also present). LiteLLM can auto-repair this shape by inserting a dummy assistant message, but only when litellm.modify_params is enabled; the exception message spells out that exact setting. It is a client-side (400) error: same payload plus modify_params=True succeeds.

Solutions

Enable the fix: litellm.modify_params = True in code (env LITELLM_MODIFY_PARAMS=True), or on the proxy litellm_settings: modify_params: True, so LiteLLM inserts the dummy assistant message
Restructure messages yourself: put tool results in their own turn per Bedrock Converse rules (assistant tool_use turn, then tool-result turn, then next user text)
Upgrade litellm - Bedrock tool-turn transformation receives ongoing fixes
If you must keep modify_params off, insert a minimal assistant message ({'role':'assistant','content':'.'}) before the tool results

Example fix

# before
litellm.completion(model="bedrock/anthropic.claude-3-sonnet-...")  # raises: Conversation blocks and tool result blocks in same turn

# after
import litellm
litellm.modify_params = True
litellm.completion(model="bedrock/anthropic.claude-3-sonnet-...", messages=messages, tools=tools)

Defensive patterns

Strategy: validation

Validate before calling

import litellm
litellm.modify_params = True  # let LiteLLM insert the dummy assistant message

def bedrock_safe_messages(messages: list) -> list:
    out = []
    for m in messages:
        if m.get('role') == 'tool' and (not out or out[-1].get('role') != 'assistant'):
            out.append({'role': 'assistant', 'content': '.'})  # dummy assistant turn
        out.append(m)
    return out

Type guard

import litellm

def is_bedrock_bad_request(e: BaseException) -> bool:
    return isinstance(e, litellm.BadRequestError) and 'tool result blocks' in str(e)

Try / catch

try:
    resp = litellm.completion(model='bedrock/...', messages=messages, tools=tools)
except litellm.BadRequestError as e:
    if 'tool result blocks' in str(e):
        litellm.modify_params = True
        resp = litellm.completion(model='bedrock/...', messages=messages, tools=tools)
    else:
        raise

Prevention

Keep litellm.modify_params=True when doing tool calling on Bedrock Converse models
Always order tool turns as assistant tool_use -> tool result -> next user message
Test agent tool loops against Bedrock specifically, not just OpenAI
After trimming history, re-validate that each tool result is preceded by an assistant turn

When it happens

Trigger: Calling litellm.completion(model='bedrock/...', tools=[...]) for the turn after a tool call, where the follow-up messages array places tool results and conversation blocks in one turn - typical when hand-building tool-response messages with OpenAI-style role='tool' ordering instead of the Converse-required assistant tool_use -> tool result -> user sequence.

Common situations: Agent loops that append tool output as a user message right after the assistant tool_use turn; porting OpenAI tool-calling code to Bedrock without reordering messages; default-safe litellm.modify_params=False so the dummy assistant message is never inserted.

Related errors

AI-assisted analysis of BerriAI/litellm@77b7c6c40c (2026-08-18). Data as JSON: /api/errors/f35684f225f67b94. Report an issue: GitHub.

Appendix: source

Thrown at litellm/litellm_core_utils/exception_mapping_utils.py:838

    exception_type: str,
    exception_provider: str,
    extra_information: str,
) -> None:
    if (
        "too many tokens" in error_str
        or "expected maxLength:" in error_str
        or "Input is too long" in error_str
        or "prompt is too long" in error_str
        or "prompt: length: 1.." in error_str
        or "Too many input tokens" in error_str
    ):
        raise ContextWindowExceededError(
            message=f"BedrockException: Context Window Error - {error_str}",
            model=model,
            llm_provider="bedrock",
        )
    elif "Conversation blocks and tool result blocks cannot be provided in the same turn." in error_str:
        raise BadRequestError(
            message=f"BedrockException - {error_str}\n. Enable 'litellm.modify_params=True' (for PROXY do: `litellm_settings::modify_params: True`) to insert a dummy assistant message and fix this error.",
            model=model,
            llm_provider="bedrock",
            response=getattr(original_exception, "response", None),
        )
    elif "Malformed input request" in error_str:
        raise BadRequestError(
            message=f"BedrockException - {error_str}",
            model=model,
            llm_provider="bedrock",
            response=getattr(original_exception, "response", None),
        )
    elif "A conversation must start with a user message." in error_str:
        raise BadRequestError(
            message=f"BedrockException - {error_str}\n. Pass in default user message via `completion(..,user_continue_message=)` or enable `litellm.modify_params=True`.\nFor Proxy: do via `litellm_settings::modify_params: True` or user_continue_message under `litellm_params`",
            model=model,
            llm_provider="bedrock",
            response=getattr(original_exception, "response", None),

View on GitHub (pinned to 77b7c6c40c)