OpenAI Responses to Anthropic Messages

This page documents the high-value conversion pair used when a client speaks OpenAI Responses and the selected provider speaks Anthropic Messages.

Inbound

`openai_responses` request payload

Normalize

ingress parses Responses input and request controls

Translate request

Responses input/messages/tools -> Anthropic Messages request

Translate response

Anthropic Message/content blocks -> Responses output items

Outbound

`openai_responses` response or SSE events

Status

Area	Expected behavior
Non-streaming request	Supported when request fields can be represented in Anthropic Messages.
Non-streaming response	Anthropic returns one final `Message`; ProxAI converts it into one Responses response envelope with `output[]`.
Streaming response	Anthropic content-block events are accumulated and emitted as Responses-style typed SSE events.
Unsupported provider-specific fields	Fail or ignore deliberately; do not leak raw provider structures through Responses output.

Responses concept	Anthropic Messages target	Notes
`input` easy messages	`messages[]`	User/developer content becomes Anthropic message content blocks where possible.
System/developer instructions	`system` or equivalent normalized instruction placement	Preserve instruction intent without duplicating user-visible content.
`tools[]` function tools	`tools[]`	Function schema should preserve name, description, and JSON schema shape.
Hosted tools	Provider-compatible representation or explicit unsupported behavior	Do not pretend an upstream supports hosted tools if it does not.
`reasoning.effort`	`output_config.effort` when available	Supported effort values map directly; disabled/minimal maps to disabled thinking behavior.
`reasoning.summary`	`thinking.display` lossily	Anthropic has no equivalent summary granularity.

Anthropic concept	Responses target	Notes
`Message` envelope	Responses response envelope	Anthropic non-streaming returns one message; Responses output may contain multiple typed items.
Text content block	Message output content	Preserve text order within the response.
Tool use block	Function call output item	Preserve call identity using target-protocol identifiers.
Thinking/reasoning block	Reasoning output item when representable	Provider-specific metadata should be transformed, not raw-round-tripped.
`stop_reason`	Responses status / terminal output semantics	See status and stop reasons reference for mappings.

Anthropic event	Responses-side expectation
`message_start`	Initialize response state and output item assembly.
`content_block_start`	Start a Responses output item or content part depending on block type.
`content_block_delta`	Emit compatible delta events without slicing Unicode boundaries.
`content_block_stop`	Finalize the current item/content block.
`message_delta`	Update usage or stop metadata when available.
`message_stop`	Emit terminal Responses event such as `response.completed`.

Responses output item identity may be richer than Anthropic message block identity.
Hosted tools may not have Anthropic equivalents.
Reasoning summary granularity is not symmetric.
Refusal metadata should preserve refusal semantics, not raw upstream shapes.
Provider-specific metadata such as signatures should not be exposed as untyped Responses payloads.

Change type	Test area
Request placement change	System/developer normalization and message placement tests.
Tool mapping change	Tool call and tool result round-trip behavior tests.
Streaming event change	SSE terminal, stalled tool-call, Unicode scanning, and event ordering tests.
Stop/status mapping change	Status and stop reason conversion tests plus reference docs.