[Bug] Claude Sonnet-4.6 incompatibility with Agno agents (assistant prefill breaking change)

[parthiv-middleware]

Description

🐛 Issue: Claude Sonnet 4.6 Breaking Change – Assistant Prefill Not Supported

Overview

While migrating from claude-sonnet-4-5 to claude-sonnet-4-6, Agno agents and teams fail with a 400 error related to assistant message prefilling.

This issue appears to stem from a breaking change introduced in Claude Sonnet 4.6, where assistant message prefilling is no longer supported.

---

🔗 Reference (Official Migration Guide)

According to the official Anthropic migration guide:
https://platform.claude.com/docs/en/about-claude/models/migration-guide

"Prefilling assistant messages returns a 400 error on Sonnet 4.6. Use structured outputs, system prompt instructions, or output_config.format instead."

---

🔁 Steps to Reproduce

1. Create an Agno agent or team
2. Configure the model as claude-sonnet-4-6
3. Execute the agent or team workflow
4. Observe failure with a 400 invalid_request_error

---

🔍 Analysis

Agno appears to construct conversations where the final message role is assistant (e.g., agent outputs or internal chaining).

This pattern was valid in Claude Sonnet 4.5 but is now rejected in 4.6.

Claude Sonnet 4.6 Constraints:

• Conversation must end with a user message
• Assistant prefill is no longer supported

---

🔄 Regression

This behavior worked correctly with:

• claude-sonnet-4-5

Steps to Reproduce

1. Create an Agno agent or team
2. Configure the model as claude-sonnet-4-6
3. Execute the agent or team workflow
4. Observe failure with a 400 invalid_request_error

Agent Configuration (if applicable)

No response

Expected Behavior

Agno agents and teams should work seamlessly with Claude Sonnet 4.6 without requiring manual restructuring of message roles.

Actual Behavior

Execution fails with:

• invalid_request_error
• Message:
  "This model does not support assistant message prefill. The conversation must end with a user message."

Screenshots or Logs (if applicable)

ERROR Claude API error (status 400): Error code: 400 - {
'type': 'error',
'error': {
'type': 'invalid_request_error',
'message': 'This model does not support assistant message prefill. The conversation must end with a user message.'
}
}

ERROR Non-retryable model provider error: Error code: 400 - {
'type': 'error',
'error': {
'type': 'invalid_request_error',
'message': 'This model does not support assistant message prefill. The conversation must end with a user message.'
}
}

ERROR Error in Team run: Error code: 400 - {
'type': 'error',
'error': {
'type': 'invalid_request_error',
'message': 'This model does not support assistant message prefill. The conversation must end with a user message.'
}
}

Environment

- **Framework:** Agno  
- **Model:** claude-sonnet-4-6  
- **Provider:** Anthropic API  
- **OS:** Linux

Possible Solutions (optional)

• Adapt Agno’s message construction to ensure the final role is user
• OR provide built-in compatibility handling for Claude 4.6 models

Additional Context

No response

[ron-42]

Hey @harshsinha03 , thanks for taking this up! I was actually already working on this issue and ended up raising a PR for it (#7022 ).
Didn’t realize it got self-assigned 😅
Would be great if you could take a quick look, happy to make any changes if needed.

[kausmeows]

Hey @harshsinha03 , thanks for taking this up! I was actually already working on this issue and ended up raising a PR for it (#7022 ). Didn’t realize it got self-assigned 😅 Would be great if you could take a quick look, happy to make any changes if needed.

Thanks @ron-42, @harshsinha03 will help review and take this through, but pls go through the updated contribution guidelines. We mostly wont be taking PRs seriously if they are just raised without first communicating in the comments on the issue section. This is to avoid noise that we've been getting with duplicate and AI PRs which dont really solve anything.. i hope you understand. And thanks a lot for taking this though, really appreciate it!

[harshsinha03]

Hey, thanks for raising the PR! I’ll take a look at it shortly. Assigning this issue to you 👍

[ysolanky]

@ron-42 can you please share a script we can use to replicate the error? Thanks

[ron-42]

@ysolanky

`
from os import getenv
from pathlib import Path

from agno.models.anthropic import Claude
from agno.models.message import Message
from dotenv import load_dotenv

load_dotenv()

MODEL_ID = "claude-sonnet-4-6"

SYSTEM_PROMPT = (
"You are a customer support triage assistant. "
"Return a compact JSON object with the keys: "
"priority, category, customer_sentiment, and needs_human_agent."
)

USER_TICKET = (
"Ticket ID: CS-18427\n"
"Customer: Our production checkout is failing for multiple users after today's deploy. "
"Payments are being authorized, but the order confirmation page crashes and no receipt is sent. "
"This is affecting revenue and we need an urgent response.\n\n"
"Classify this ticket and return only JSON."
)

def run_invalid_prefill_example(model: Claude) -> None:
print("\n=== Invalid Claude 4.6 Pattern ===")
invalid_messages = [
Message(role="system", content=SYSTEM_PROMPT),
Message(role="user", content=USER_TICKET),
Message(role="assistant", content='{"priority":'),
]

try:
    model.response(messages=invalid_messages)
except ValueError as exc:
    print(exc)

def main() -> None:
if not getenv("ANTHROPIC_API_KEY"):
raise SystemExit("Set ANTHROPIC_API_KEY before running this example.")

model = Claude(id=MODEL_ID, max_tokens=128)

run_invalid_prefill_example(model)

if name == "main":
main()`

[wolyosung]

I'm experiencing the same issue.

[harshsinha03]

Hey @parthiv-middleware @wolyosung @ron-42 Thank you for reporting this! We have the fix out and we will release it soon.

Appreciate you for using Agno!