OpenAI Chat Completions to Anthropic Messages

This page documents the conversion pair used when a client speaks OpenAI Chat Completions and the selected provider speaks Anthropic Messages.

Inbound

`openai_chat_completions` request payload

Normalize

`proxai-core/ingress` validates the structured Chat request

Translate request

Chat messages/tools -> Anthropic Messages request

Translate response

Anthropic Message/content blocks -> Chat choices

Outbound

`openai_chat_completions` response or SSE chunks

Request mapping

Chat concept	Anthropic Messages target	Notes
`messages[]`	`messages[]` plus normalized system placement	Chat system/developer instructions must not be duplicated as user-visible turns.
Assistant text	Assistant content blocks	Preserve ordered visible text.
Assistant `reasoning_content`	Anthropic thinking blocks when a ProxAI continuation is present	A versioned continuation suffix restores signed or redacted Anthropic thinking and is removed before forwarding. Plain unsigned reasoning has no Anthropic signature, so translation traces and skips it rather than forging one.
`tools[]` function definitions	`tools[]`	Preserve name, description, and JSON schema where possible.
Assistant `tool_calls[]`	Anthropic `tool_use` blocks	Chat `tool_calls[].index` is stream-local; Anthropic content blocks use content-block indices.
Tool result messages	Anthropic `tool_result` blocks	Preserve call id linkage where target protocol can represent it.

Anthropic concept	Chat target	Notes
Text content block	`choices[0].message.content`	Chat has a single assistant message per choice.
Tool use block	`choices[0].message.tool_calls[]`	Assign stable call ids and function names.
Thinking/reasoning block	`choices[0].message.reasoning_content`	Zed consumes this OpenAI-compatible extension as thinking output; ProxAI appends a versioned continuation suffix to preserve Anthropic signatures and redacted payloads for replay without mixing them into visible `content`.
`stop_reason`	`finish_reason`	Map natural turns, token limits, tool use, and refusals carefully.

Anthropic event	Chat SSE expectation
`message_start`	Initialize choice and assistant role delta.
Text `content_block_delta`	Emit `delta.content` chunks.
Thinking `content_block_delta`	Emit `delta.reasoning_content` chunks consumed by Zed as thinking; append the continuation suffix before the terminal chunk.
Tool `input_json_delta`	Emit `delta.tool_calls[].function.arguments` chunks keyed by Chat tool index.
`message_delta.stop_reason`	Set final `finish_reason` when terminal state is known.
`message_stop`	Emit final chunk and `[DONE]`.

Chat has no native output item graph.
Plain thinking text is preserved through the reasoning_content compatibility extension. ProxAI appends its versioned continuation suffix only for compatible clients that replay it; Chat itself still has no native per-item identity or signature field.
The continuation suffix can preserve redacted/encrypted thinking across ProxAI’s Anthropic conversion pair, but is not a portable Chat Completions feature.
Chat streaming uses tool_calls[].index; Anthropic streaming uses content block index.
Chat [DONE] is a stream carrier terminator; Anthropic message_stop is the semantic terminal event.