Adapters

An adapter translates between opencodex’s internal request/response model and one provider wire format. Every adapter implements the ProviderAdapter interface (src/adapters/base.ts):

interface ProviderAdapter {
  name: string;
  buildRequest(parsed, incoming?): { url; method; headers; body };
  parseStream(response): AsyncGenerator<AdapterEvent>;
  parseResponse?(response): Promise<AdapterEvent[]>;   // non-streaming
  passthrough?: true;                                   // pipe raw, skip translation
}

buildRequest lowers an OcxParsedRequest into an upstream HTTP request; parseStream / parseResponse lift the provider’s reply back into internal AdapterEvents, which bridge.ts turns into Responses SSE.

openai-chat

Targets: OpenAI Chat Completions (POST {baseUrl}/chat/completions) and every compatible provider — xAI, Kimi, DeepSeek, GLM, Groq, OpenRouter, Ollama (local & cloud), and more. Auth: key (Bearer).

• Converts internal messages to OpenAI roles; maps tools to {type:"function", function:{…}} and tool_choice (auto/none/required or a named function).
• Rewrites Codex’s GPT-5 identity prompt to a model-agnostic intro so routed models don’t claim to be OpenAI.
• Clamps reasoning_effort to what most providers accept (minimal→low, xhigh/max→high), and omits it entirely for ids in provider.noReasoningModels.
• Streams delta.content (text), delta.reasoning_content (thinking), and delta.tool_calls[]; collects usage.

openai-responses

Targets: the OpenAI Responses API. passthrough: true — forwards the raw request body and streams the response back untranslated. Auth: forward (relay the caller’s headers) or key.

• forward URL → {baseUrl}/responses; key URL → {baseUrl}/v1/responses.
• In forward mode only a safe header allowlist is relayed (FORWARD_HEADERS): authorization, ChatGPT account id, and the OpenAI beta/originator/session headers. This is the ChatGPT-login path that also powers the sidecars.

anthropic

Targets: Anthropic Messages (/v1/messages). Auth: key (x-api-key) or oauth (Bearer + anthropic-beta, for Claude Pro/Max).

• Converts messages to Anthropic content blocks (text, base64 image, tool_use, thinking).
• Extended thinking math: Anthropic requires max_tokens > thinking.budget_tokens. The adapter maps reasoning effort to a budget (minimal 1024 … max 32000), then computes a safe max_tokens with output headroom, and drops temperature/top_p when thinking is enabled (Anthropic forbids them there).
• Always sends anthropic-version: 2023-06-01. Streams content_block_delta (text_delta, thinking_delta, input_json_delta).

google

Targets: Google Gemini (/v1beta/models/{model}:streamGenerateContent). Auth: key (x-goog-api-key).

• System prompt → systemInstruction; messages → contents[] (assistant → model); tools → functionDeclarations. Data-URL images → inline_data.
• No native reasoning; tool-call ids are synthesized (Gemini doesn’t return them).

azure

Targets: Azure OpenAI. Wraps openai-responses (so also passthrough: true). Auth: key via the api-key header (not Bearer).

• Delegates request building to the Responses passthrough, then swaps Authorization for api-key and appends an api-version query param (default 2025-04-01-preview).

Image utilities (image.ts)

Shared helpers used by the vision-aware adapters:

• parseDataUrl(url) — split a data:<type>;base64,<data> URL into { mediaType, base64 } for Anthropic/Google image blocks.
• contentPartsToText(content) — flatten content parts to text for text-only tool messages (an undescribed image becomes a short [image] marker, never a token-exploding base64 blob).