bug: multiple issues in anthropic /v1/messages experimental pass-through to openai

[garnetlyx]

Bug: Multiple bugs in Anthropic /v1/messages experimental pass-through to OpenAI

Description

There are five related bugs in the Anthropic /v1/messages experimental pass-through implementation when routing requests to OpenAI/Azure models. These issues were discovered during integration testing with Claude Code CLI which uses input_text content block format and requires forcing routing through Chat Completions API when the downstream proxy doesn't support Responses API.

Bug 1: input_text type not supported in chat completions transformation

Claude Code CLI / Claude Agent SDK sends content blocks with {"type": "input_text"} (instead of plain {"type": "text"}) for user messages and system content. The current implementation only checks for "text", so input_text blocks are silently dropped → causes empty input → 422 validation error from downstream provider.

Location: litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py

Bug 2: input_text type not supported in Responses API transformation

Same issue exists on the Responses API conversion path. The code only checks for "text" in three locations:

• Top-level user content blocks
• Nested content inside tool_result
• System content array

Location: litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py

Bug 3: LITELLM_USE_CHAT_COMPLETIONS_URL_FOR_ANTHROPIC_MESSAGES not respected in responses_api_bridge_check

When the environment variable LITELLM_USE_CHAT_COMPLETIONS_URL_FOR_ANTHROPIC_MESSAGES=true is set (to force routing through /v1/chat/completions instead of /v1/responses), the responses_api_bridge_check function in main.py doesn't check this setting. It still activates Responses API mode for models with the responses/ prefix → request routed to wrong endpoint.

Location: litellm/main.py function responses_api_bridge_check()

Bug 4: LITELLM_USE_CHAT_COMPLETIONS_URL_FOR_ANTHROPIC_MESSAGES not respected in thinking auto-routing

When thinking parameter is enabled, the code automatically adds the responses/ prefix to route through Responses API (to get reasoning text output). But it didn't check the use_chat_completions_url_for_anthropic_messages setting before adding the prefix, ignoring user opt-out preference.

Location: litellm/llms/anthropic/experimental_pass_through/adapters/handler.py function _route_openai_thinking_to_responses_api_if_needed()

Bug 5: LITELLM_USE_CHAT_COMPLETIONS_URL_FOR_ANTHROPIC_MESSAGES not respected in token counting

When the environment variable is enabled, should_use_token_counting_api() still returns True for OpenAI provider, causing it to call /v1/responses/input_tokens endpoint which doesn't exist when routing through Chat Completions → 405 Method Not Allowed error.

Location: litellm/llms/openai/responses/count_tokens/token_counter.py function should_use_token_counting_api()

Expected behavior

1. input_text content blocks should be accepted and properly converted just like text blocks
2. The LITELLM_USE_CHAT_COMPLETIONS_URL_FOR_ANTHROPIC_MESSAGES setting should be respected consistently at all entry points where Responses API routing decision is made

Fix

All five bugs are fixed in the accompanying PR.

[krrish-berri-2]

@Sameerlite can you try to reproduce these issues. if you can, then please fix accordingly.