bug: Anthropic /v1/messages -> Responses API adapter drops required detail field on images and mislabels thinking as output_text #37175

Related to #23841 (same file, different specific bugs).

Environment: litellm main-stable, target backend: self-hosted vLLM 0.27.1 (OpenAI-compatible /v1/responses), via /v1/messages -> litellm/llms/anthropic/experimental_pass_through/responses_adapters/transformation.py.

Bug 1: input_image missing required detail field -> 400

translate_messages_to_responses_input, btype == "image" branch:

user_parts.append({"type": "input_image", "image_url": url})

The Responses API's ResponseInputImageParam requires detail: "low"|"high"|"auto"|"original". Without it the backend rejects the request with a large Pydantic union-validation error, which litellm's own MaskedHTTPStatusError then hides -- confirmed by reproducing the identical error directly against vLLM with the same field missing, and confirmed fixed by adding it.

Repro: send any image content block (e.g. a PDF page rendered by Claude Code's Read tool) through /v1/messages to an OpenAI-Responses-API-compatible backend.

Fix:

- user_parts.append({"type": "input_image", "image_url": url})
+ user_parts.append({"type": "input_image", "image_url": url, "detail": "auto"})

Bug 2: assistant thinking blocks mislabeled as output_text

Same function, btype == "thinking" branch:

asst_parts.append({"type": "output_text", "text": thinking_text})

Prior reasoning content gets folded into the message's regular visible text instead of a proper reasoning item, so on multi-turn replay the model can no longer distinguish its own prior thinking from prior final answers. The response-side translation in this same file already builds real reasoning/ResponseReasoningItem items -- this is only the request-side/history-replay path.

Fix:

- asst_parts.append({"type": "output_text", "text": thinking_text})
+ input_items.append({"type": "reasoning", "id": block.get("signature") or f"reasoning_{len(input_items)}", "summary": [{"type": "summary_text", "text": thinking_text}]})

Happy to open a PR with both fixes if useful.