Summary
Please add an OpenAI Responses API endpoint (POST /v1/responses) so that OpenAI Codex (CLI, desktop app, and IDE extensions) can use Kiro Gateway as a custom model provider.
Problem / Motivation
Codex custom providers now only support wire_api = "responses" (the OpenAI Responses API). Support for wire_api = "chat" (Chat Completions) was deprecated in early 2026 and can cause Codex to fail on startup.
Kiro Gateway currently exposes only:
/v1/chat/completions (OpenAI Chat Completions)
/v1/messages (Anthropic Messages)
Because there is no /v1/responses endpoint, Codex cannot connect to the gateway (its requests hit a non-existent route). Tools that speak Chat Completions or Anthropic Messages (Claude Code, Cursor, Cline, etc.) work fine, but Codex does not.
For reference, AWS added a Responses-API-compatible surface to Amazon Bedrock specifically so OpenAI tools like Codex could connect. An equivalent adapter in Kiro Gateway would let it serve Codex while reusing Kiro subscriptions.
Proposed solution
Add a Responses API adapter mirroring the existing OpenAI/Anthropic layers and reusing the shared Core layer:
kiro/models_responses.py - Pydantic models for the Responses request/response (input items, instructions, tools, previous_response_id, etc.)
kiro/converters_responses.py - Responses -> Kiro payload (reuse converters_core.build_kiro_payload)
kiro/streaming_responses.py - Kiro events -> Responses SSE events (response.created, response.output_text.delta, response.completed, tool-call events, ...)
kiro/routes_responses.py - POST /v1/responses (reuse verify_api_key)
- Register the router in
main.py
Additional context
- Codex + custom provider requires
wire_api = "responses".
- Streaming semantics differ from Chat Completions (event-typed SSE), so a Responses-specific formatter over
streaming_core is likely needed.
Thanks for the project!
Summary
Please add an OpenAI Responses API endpoint (
POST /v1/responses) so that OpenAI Codex (CLI, desktop app, and IDE extensions) can use Kiro Gateway as a custom model provider.Problem / Motivation
Codex custom providers now only support
wire_api = "responses"(the OpenAI Responses API). Support forwire_api = "chat"(Chat Completions) was deprecated in early 2026 and can cause Codex to fail on startup.Kiro Gateway currently exposes only:
/v1/chat/completions(OpenAI Chat Completions)/v1/messages(Anthropic Messages)Because there is no
/v1/responsesendpoint, Codex cannot connect to the gateway (its requests hit a non-existent route). Tools that speak Chat Completions or Anthropic Messages (Claude Code, Cursor, Cline, etc.) work fine, but Codex does not.For reference, AWS added a Responses-API-compatible surface to Amazon Bedrock specifically so OpenAI tools like Codex could connect. An equivalent adapter in Kiro Gateway would let it serve Codex while reusing Kiro subscriptions.
Proposed solution
Add a Responses API adapter mirroring the existing OpenAI/Anthropic layers and reusing the shared Core layer:
kiro/models_responses.py- Pydantic models for the Responses request/response (input items,instructions, tools,previous_response_id, etc.)kiro/converters_responses.py- Responses -> Kiro payload (reuseconverters_core.build_kiro_payload)kiro/streaming_responses.py- Kiro events -> Responses SSE events (response.created,response.output_text.delta,response.completed, tool-call events, ...)kiro/routes_responses.py-POST /v1/responses(reuseverify_api_key)main.pyAdditional context
wire_api = "responses".streaming_coreis likely needed.Thanks for the project!