[BUG] Go service /v1/responses returns HTTP 200 but emits an incomplete SSE event stream, breaking Codex-style clients

[YiRan0]

Summary

The Go service advertises POST /v1/responses and returns valid JSON for non-streaming requests, but streaming responses emit an incomplete SSE event sequence. The stream is missing the response.output_item.added and response.content_part.added events that OpenAI Responses-API clients (e.g. the Codex CLI) rely on to parse output incrementally. In practice the client only receives response.output_text.delta → response.completed → [DONE], which does not constitute a valid Responses-API stream.

Reproduction

1. Point the Codex CLI at the Go service (in ~/.codex/config.toml):
   • base_url = https://opencode.ai/zen/go/v1
   • wire_api = responses
   • model: deepseek-v4-flash
2. Run any prompt. Codex fails to parse the stream and the turn errors out or hangs.
3. Raw SSE captured from POST /v1/responses with "stream": true:

event: response.output_text.delta
data: {"type":"response.output_text.delta","delta":"ok",...}

event: response.completed
data: {"type":"response.completed","response":{...}}

data: [DONE]

event: ping
data: {"type":"ping","cost":"0"}

Note the absence of response.created, response.in_progress, response.output_item.added, response.content_part.added, and response.output_item.done.

Expected behavior

A standard Responses-API SSE stream:

event: response.created
event: response.in_progress
event: response.output_item.added      (message item)
event: response.content_part.added     (output_text part)
event: response.output_text.delta      ...
event: response.output_item.done
event: response.completed

Impact

• Non-streaming POST /v1/responses returns a complete object and works via curl.
• Streaming is what Codex-style clients actually use; with the current event set they cannot render or complete a turn.

Related

• [FEATURE]: Support Responses API for deepseek-v4-flash on opencode-go #39829 (feature request for DeepSeek V4 Flash Responses API on Go)
• [FEATURE]: add Responses API support for Go service #23655 (Responses API support for the Go service)
• PR feat(core): support response-shaped catalog models #40011 (feat(core): support response-shaped catalog models) appears to be the fix in progress, but has not been merged yet.

[Yuxin-Qiao]

Reproduced live today against POST https://opencode.ai/zen/go/v1/responses with deepseek-v4-flash, stream: true (real opencode-go subscription key):

event: response.output_text.delta   xN
event: response.completed
data: [DONE]
event: ping

HTTP 200 with content-type: text/event-stream, but no response.created, response.in_progress, response.output_item.added, response.content_part.added or response.output_item.done. Non-streaming POST /v1/responses returns a complete JSON object as described.

Root cause: for chat-completions-backed (oa-compat) upstreams, packages/console/app/src/routes/zen/util/provider/openai.ts synthesizes the Responses-API SSE stream from chat chunks and only ever emitted response.output_text.delta, response.completed (plus function-call events). The stream converter is stateless per chunk, so the lifecycle events are never produced.

I opened a fix: #40210 (fix(console): emit complete Responses-API SSE lifecycle for converted streams), which makes the converter stateful and emits the full event sequence (created -> in_progress -> output_item.added -> content_part.added -> output_text.delta -> output_item.done -> completed).

Note on #40011: that PR changes the client-side model catalog (shape: "responses", @ai-sdk/openai) and docs only; it does not touch packages/console gateway code, so it does not fix this gateway-side stream issue on its own.