Responses API

Used by Codex-style clients. Both sync and streaming are supported and verified:

curl {base_url}/r/{router}/v1/responses \
  -H "Authorization: Bearer sk-dodo-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "default", "input": "Reply with exactly the word: pong"}'

# Typed input items

Responses usage includes reported input_tokens_details (including cached tokens) and output_tokens_details (including reasoning tokens). These details survive provider normalization and the final response.completed event as well as synchronous responses. Reported zero remains zero; missing details are not invented.

Assistant message content arrays also retain their output blocks on resumed sessions; arrays are not wrapped inside a string-valued text field.

The client's full reasoning object is preserved on Responses upstreams, including context, summary and effort. A routing step set to provider default does not inject an effort or context value. Client-supplied reasoning still passes through; if both client and step omit it, the provider decides. Explicit parallel_tool_calls values also survive the round-trip, including false, which Codex Responses-Lite requires.

On Responses-format upstream routes (including OpenAI Codex), non-message input items retain their type, payload and order. This includes Codex additional_tools, reasoning items, function calls and function-call outputs. An item's role does not make it a message: additional_tools can have a developer role without a content field, and DodoRouter must not fabricate content: null for it. Ordinary messages still use the message conversion path.

This preserves the request representation; the selected upstream must still support the item types. It does not add translations of these native items for Chat Completions, Anthropic or Gemini fallback routes.

# Function tools on Anthropic routes

Responses-style function tools with top-level name, description and parameters are translated when a request is routed or falls back to Anthropic. The schema becomes input_schema, and any tool-level cache_control is retained. Bowser and other Responses clients can keep using /v1/responses; no switch to /v1/messages is required for this conversion.

This covers function definitions, not Responses built-in tools or native function-call history items. Those have separate cross-format limitations as noted above.

# Streaming event sequence

With "stream": true, DodoRouter emits the full item lifecycle a strict Responses-API client expects:

response.created → response.output_item.added → response.content_part.added → one or more response.output_text.delta → response.output_text.done → response.content_part.done → response.output_item.done → response.completed.

Verified with a real Codex CLI streaming run — text renders live in the terminal. An earlier version of this endpoint skipped the output_item.added / content_part.added events, which caused strict clients to log OutputTextDelta without active item and stall; that gap is fixed.