LiteLLM Append Trailing User Message

Append a trailing user turn so Claude 4.6+ models that reject assistant prefill work through LiteLLM.

Claude 4.6+ does not support assistant message prefill. The LiteLLM formatter appends a user turn when the final input message has the assistant role.

The source below supplies ordinary user messages and never exercises the append operation. Its separate reasoning_model stages are skipped because Agno does not recognize the LiteLLM wrapper as a supported native reasoning adapter. Run the current example below to exercise the formatter directly through an agent.

append_trailing_user_message.py
"""
LiteLLM Append Trailing User Message
=====================================

Claude 4.6+ does not support assistant message prefill. Enable
`append_trailing_user_message` to append a trailing user turn when the
conversation ends with an assistant message (e.g. during reasoning).

Use `trailing_user_message_content` to customise the appended text (defaults to "continue").

Note: Claude 4.6+ models auto-detect and enable this flag automatically.
"""

from agno.agent import Agent
from agno.models.litellm import LiteLLM

# ---------------------------------------------------------------------------
# Create Agent
# ---------------------------------------------------------------------------

agent = Agent(
    model=LiteLLM(
        id="anthropic/claude-sonnet-4-6",
        # Claude 4.6 rejects temperature + top_p together; drop top_p.
        top_p=None,
        append_trailing_user_message=True,
    ),
    reasoning_model=LiteLLM(id="anthropic/claude-opus-4-7", top_p=None),
    markdown=True,
)

# With custom trailing content
agent_custom = Agent(
    model=LiteLLM(
        id="anthropic/claude-sonnet-4-6",
        top_p=None,
        append_trailing_user_message=True,
        trailing_user_message_content="continue",
    ),
    reasoning_model=LiteLLM(id="anthropic/claude-opus-4-7", top_p=None),
    markdown=True,
)

# ---------------------------------------------------------------------------
# Run Agent
# ---------------------------------------------------------------------------
if __name__ == "__main__":
    agent.print_response("What is 15 + 27?")
    agent_custom.print_response("What is 15 + 27?")

Current Example

Each request ends with an assistant input message. The formatter appends continue on the first run and . on the second; this does not invoke a separate reasoning model.

trailing_current.py
from agno.agent import Agent
from agno.models.litellm import LiteLLM
from agno.models.message import Message

for trailing_text in ("continue", "."):
    agent = Agent(
        model=LiteLLM(
            id="anthropic/claude-sonnet-4-6",
            top_p=None,
            append_trailing_user_message=True,
            trailing_user_message_content=trailing_text,
        ),
        markdown=True,
    )
    agent.print_response([
        Message(role="user", content="What is 15 + 27?"),
        Message(role="assistant", content="I will add the two numbers."),
    ])

Run the Example

Set up your virtual environment

uv venv --python 3.12
source .venv/bin/activate

Install dependencies

uv pip install -U agno litellm

Export your Anthropic API key

unset LITELLM_API_KEY
export ANTHROPIC_API_KEY="your_anthropic_api_key_here"

Run the example

Save the Current Example as trailing_current.py, then run:

python trailing_current.py

Full source: cookbook/90_models/litellm/append_trailing_user_message.py