Summary
/v1/messages cross-provider translation is text-only on the request side: image, tool_use, tool_result, document, and thinking content blocks are silently dropped when the target is a non-Anthropic upstream (crates/aisix-provider-anthropic/src/wire.rs:976-984; the module docstring at wire.rs:900-901 already promises this follow-up). Consequence: multi-turn tool loops (e.g. Claude Code against a non-Anthropic model) lose their tool history, and vision/document inputs vanish. Matrix row "Anthropic Messages 跨厂商 = 仅文本块" in api7/AISIX-Cloud#873 section ⑤.
Scope (map exactly to LiteLLM LiteLLMAnthropicMessagesAdapter.translate_anthropic_messages_to_openai, litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py:353)
Request → OpenAI chat shape:
image (base64 → data: URL; url → passthrough) → image_url content partdocument/PDF → file content part (image_url-style per LiteLLM)- assistant
tool_use → tool_calls[] entry (name 64-char truncation, arguments JSON-encoded)
tool_result → role:"tool" message keyed by tool_use_id; string content, single-block, and multi-block (text+image combined) formsthinking/redacted_thinking history blocks → dropped for OpenAI-wire upstreams (same effective behavior as LiteLLM's OpenAI provider transform; document it)
Response → Anthropic shape: text + tool_use already work (non-stream + stream); verify nothing regresses and add coverage.
Tests
- Unit tests on the wire translation match arms per block type.
- DP e2e: anthropic-shape multi-turn tool loop against a mock OpenAI upstream (tool_use + tool_result history preserved); vision request round-trip.
Part of the section-⑤ gap closure for api7/AISIX-Cloud#873.
Summary
/v1/messagescross-provider translation is text-only on the request side:image,tool_use,tool_result,document, andthinkingcontent blocks are silently dropped when the target is a non-Anthropic upstream (crates/aisix-provider-anthropic/src/wire.rs:976-984; the module docstring at wire.rs:900-901 already promises this follow-up). Consequence: multi-turn tool loops (e.g. Claude Code against a non-Anthropic model) lose their tool history, and vision/document inputs vanish. Matrix row "Anthropic Messages 跨厂商 = 仅文本块" in api7/AISIX-Cloud#873 section ⑤.Scope (map exactly to LiteLLM
LiteLLMAnthropicMessagesAdapter.translate_anthropic_messages_to_openai,litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py:353)Request → OpenAI chat shape:
image(base64 →data:URL; url → passthrough) →image_urlcontent partdocument/PDF → file content part (image_url-style per LiteLLM)tool_use→tool_calls[]entry (name 64-char truncation, arguments JSON-encoded)tool_result→role:"tool"message keyed bytool_use_id; string content, single-block, and multi-block (text+image combined) formsthinking/redacted_thinkinghistory blocks → dropped for OpenAI-wire upstreams (same effective behavior as LiteLLM's OpenAI provider transform; document it)Response → Anthropic shape: text + tool_use already work (non-stream + stream); verify nothing regresses and add coverage.
Tests
Part of the section-⑤ gap closure for api7/AISIX-Cloud#873.