Anthropic Messages to OpenAI Responses

This page documents the conversion path used when the provider speaks Anthropic Messages and the client expects OpenAI Responses output.

Provider response

`anthropic_messages` Message or SSE stream

Read upstream

`src/upstream/` reads body or ByteStream

Translate response

Anthropic content blocks -> Responses output items

Rebuild client response

`src/http_support/` emits Responses JSON or typed SSE

Client response

`openai_responses` envelope or events

Non-streaming response mapping

Anthropic concept	Responses target	Notes
`Message` envelope	Responses response envelope	Anthropic returns one final message; Responses returns an envelope with `output[]`.
Text block	Message output item/content part	Preserve visible text order.
`tool_use` block	`function_call` output item	Preserve call identity using Responses item/call identifiers.
`thinking` block	Reasoning output item when representable	Transform provider-specific metadata instead of raw round-tripping.
`stop_reason`	Responses terminal status and output state	Map terminal behavior, not just field names.

Anthropic event	Responses-side behavior
`message_start`	Initialize response and usage state.
`content_block_start`	Emit or stage a Responses output item depending on block type.
`content_block_delta`	Emit text, reasoning, or function-argument deltas.
`content_block_stop`	Finalize the current Responses item/content part.
`message_delta`	Update stop and usage metadata; may not have a standalone Responses event.
`message_stop`	Emit Responses terminal event such as `response.completed`.

Anthropic	Responses	Consequence
Content block `index`	Stable `item_id` plus output index	The translator must mint/track item ids for Responses events.
Tool `id` / tool result linkage	Function call item id and `call_id`	Do not expose raw provider ids as the only client identity.
Message-level terminal metadata	Response status plus item status	Terminal state may need to update both envelope and output items.

Anthropic content block index is stream-local; Responses item identity is client-visible and can be stable across turns.
Provider-specific thinking metadata may not have a Responses equivalent.
Anthropic pause_turn and refusal metadata need behavior-preserving target semantics.
Usage and stop metadata may arrive late in streams.