Anthropic Messages

Anthropic Messages

ProxAI uses the protocol name anthropic_messages for the Anthropic Messages API.

Protocol value

anthropic_messages

Request path

/v1/messages

Main code areas

src/protocol/anthropic/messages/wiresrc/provider/anthropic_messagessrc/ingress/anthropic_messagessrc/translation/anthropic_messages

Conversion targets

pass-through selfopenai_responses

Response shape

A non-streaming Anthropic Messages response is a single top-level Message envelope:

{
  "id": "msg_...",
  "type": "message",
  "role": "assistant",
  "model": "...",
  "content": [
    { "type": "text", "text": "..." }
  ],
  "stop_reason": "end_turn",
  "usage": {}
}

Multiple output pieces live in content: Vec<ContentBlock>, not in multiple top-level messages.

Stop reason

stop_reason explains why generation stopped:

• end_turn — natural end of the assistant turn
• tool_use — model requested tool execution
• max_tokens — output was truncated by token limit
• stop_sequence — configured stop sequence matched
• pause_turn — turn paused
• refusal — model refused

Wire structs use Option<T> for nullable/missing upstream fields, including compatibility providers and streaming start events.

Streaming

Streaming is event-based: message_start, content block events, message_delta, and message_stop. The final stop reason is usually carried by message_delta.delta.stop_reason.

For cross-protocol streaming details, see Protocol Conversion.