Adapters and Translation

Overview

Adapters are the protocol boundary. The Anthropic adapter preserves the Anthropic Messages shape and streams upstream bytes back to Claude Code. The Responses adapter rebuilds the request for OpenAI’s Responses API, sends provider-specific credentials, and converts Responses SSE events into Anthropic SSE events through a state machine src/adapters/anthropic.rs:31-104 src/adapters/responses.rs:34-213 src/model/responses_request.rs:4-280 src/model/responses.rs:45-378.

Adapter	Provider kinds	Mutation level	Streaming behavior	Source
`AnthropicAdapter`	`kind = "anthropic"`	Header credential swap only for API-key providers; body is unchanged	Relays upstream `bytes_stream()`	src/adapters/anthropic.rs:31-104
`ResponsesAdapter`	`kind = "responses"`	Full request rebuild into Responses shape	Parses upstream SSE and emits Anthropic SSE	src/adapters/responses.rs:34-213
`translate_request`	Responses providers	Maps system/messages/tools/effort into Responses request	Always requests upstream streaming	src/model/responses_request.rs:4-280
`ResponsesFlavor`	Responses providers	Gates per-backend dialect quirks (OpenAI / ChatGPT-Codex / xAI) table-driven from auth mode + base-URL host	xAI: reasoning opt-in, no `text` object, no `OpenAI-Beta` header; ChatGPT: no `max_output_tokens`	src/config.rs:167-186 src/config.rs:489-505
`AnthropicSseMachine`	Responses providers	Converts response events into content blocks and final JSON	Supports stream and non-stream output	src/model/responses.rs:62-113

Adapter Dispatch

flowchart TB
    Route[Route adapter] --> Choice{AdapterKind}
    Choice -->|Anthropic| A[AnthropicAdapter.forward]
    Choice -->|Responses| R[ResponsesAdapter.forward]
    A --> AU[upstream_url + filtered headers]
    R --> TR[translate_request]
    R --> RB[request_builder + responses_url]
    R --> SM[AnthropicSseMachine]
    classDef dark fill:#2d333b,stroke:#6d5dfc,color:#e6edf3;
    class Route,Choice,A,R,AU,TR,RB,SM dark;
    linkStyle default stroke:#8b949e;

Request Translation

flowchart LR
    AM[Anthropic Messages JSON] --> System[system to instructions]
    AM --> Messages[messages to input items]
    AM --> Tools[tools to function tools]
    AM --> Choice[tool_choice mapping]
    AM --> Effort[effort mapping]
    System --> OR[OpenAI Responses request]
    Messages --> OR
    Tools --> OR
    Choice --> OR
    Effort --> OR
    OR --> Stream[stream true and store false]
    classDef dark fill:#2d333b,stroke:#6d5dfc,color:#e6edf3;
    class AM,System,Messages,Tools,Choice,Effort,OR,Stream dark;
    linkStyle default stroke:#8b949e;

Streaming Conversion

sequenceDiagram
    autonumber
    participant Up as Responses upstream
    participant Parser as SseParser
    participant Machine as AnthropicSseMachine
    participant CC as Claude Code
    Up-->>Parser: response.created
    Parser->>Machine: ResponseEvent
    Machine-->>CC: message_start + ping
    Up-->>Parser: response.output_item.added
    Machine-->>CC: content_block_start
    Up-->>Parser: response.output_text.delta
    Machine-->>CC: content_block_delta
    Up-->>Parser: response.completed
    Machine-->>CC: message_delta + message_stop

SSE State Machine

stateDiagram-v2
    [*] --> Idle
    Idle --> Started: response.created
    Started --> TextOpen: message output item
    Started --> ToolOpen: function_call output item
    TextOpen --> TextOpen: output_text.delta
    ToolOpen --> ToolOpen: function_call_arguments.delta
    TextOpen --> Started: output_text.done
    ToolOpen --> Started: arguments.done
    Started --> Completed: response.completed
    Started --> Error: response.failed or error
    Completed --> [*]
    Error --> [*]

Translation Coverage

Anthropic concept	Responses concept	Implementation	Test coverage	Source
`system` string/blocks	`instructions`	`instructions()`	Plain-text request test	src/model/responses_request.rs:31-45 tests/responses_translate.rs:25-50
User text blocks	`input_text`	`text_part()`	Multi-turn role test	src/model/responses_request.rs:113-123 tests/responses_translate.rs:52-71
Images	Data URL `input_image`	`image_part()`	Image translation test	src/model/responses_request.rs:126-138 tests/responses_translate.rs:98-119
Tool use	`function_call`	`tool_use_item()`	Tool call ID test	src/model/responses_request.rs:140-148 tests/responses_translate.rs:73-96
Reasoning effort	`reasoning.effort`	`effort()`	Thinking and override test	src/model/responses_request.rs:254-280 tests/responses_translate.rs:164-178

Error Mapping

Upstream errors on the responses path are re-shaped into the Anthropic error envelope by map_error_value, with the error.type derived from the upstream status (401 → authentication_error, 429 → rate_limit_error, 400 → invalid_request_error, other → api_error) and the upstream message preserved (src/model/responses.rs:532-563).

Context-overflow errors are the one case where the message itself is rewritten: Claude Code’s automatic compact-and-retry matches the literal phrase prompt is too long and parses N tokens > M maximum to size the retry, so upstream phrasings (context_length_exceeded, “maximum context length is N tokens”, “prompt token count of N exceeds the limit of M”) are detected and rewritten to that shape, keeping the token counts when the upstream message carries them (src/model/responses.rs:566-616, tests/responses_translate.rs:368-417).

flowchart LR
    Upstream[Upstream error body] --> Detect{context overflow?}
    Detect -->|no| Preserve[original message]
    Detect -->|yes| Rewrite["prompt is too long: N tokens > M maximum"]
    Preserve --> Envelope[Anthropic error envelope]
    Rewrite --> Envelope
    classDef dark fill:#2d333b,stroke:#6d5dfc,color:#e6edf3;
    class Upstream,Detect,Preserve,Rewrite,Envelope dark;
    linkStyle default stroke:#8b949e;

Page	Relationship
Architecture	Shows adapter layer in context
Routing and Configuration	Explains how a provider selects an adapter
Authentication	Explains credentials adapters consume
Testing and Quality	Shows tests that lock translation behavior