Missing 'response.content_part.added' SSE event in Responses API streaming - breaks Codex CLI compatibility

[jerryan999]

Description

When using the OpenAI Responses API (/v1/responses) with stream: true, Sub2API emits SSE events in this order:

event: response.created
event: response.output_item.added
event: response.output_text.delta    ← content starts here
event: response.output_text.done
event: response.output_item.done
event: response.completed

The response.content_part.added event is missing after response.output_item.added and before response.output_text.delta.

Expected SSE Event Sequence

Per the OpenAI Responses API streaming specification, the correct sequence should be:

event: response.created
event: response.output_item.added
event: response.content_part.added   ← MISSING in Sub2API
event: response.output_text.delta
event: response.output_text.done
event: response.content_part.done    ← MISSING in Sub2API
event: response.output_item.done
event: response.completed

Impact

Clients that strictly follow the OpenAI Responses API event model (such as OpenAI Codex CLI) fail to process streaming responses. Codex CLI logs:

ERROR codex_core::util: OutputTextDelta without active item

It discards all output_text.delta content because there is no active content part to associate it with. The HTTP response returns 200 OK, the SSE stream is well-formed, but the text content is never delivered to the user.

Reproduction

curl -s -N -X POST http://localhost:8080/v1/responses \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <your-key>" \
  -d '{"model":"deepseek-v4-flash","input":"hello","stream":true}' \
  | grep "^event:"

Expected output includes response.content_part.added; actual output jumps directly from response.output_item.added to response.output_text.delta.

Environment

• Sub2API version: weishaw/sub2api:0.1.132 (latest as of 2026-05-29)
• Client: OpenAI Codex CLI v0.135.0

Suggested Fix

Before emitting response.output_text.delta, add:

event: response.content_part.added
data: {"type":"response.content_part.added","part":{"type":"output_text","text":""},"item_id":"<item_id>"}

And after response.output_text.done, emit:

event: response.content_part.done
data: {"type":"response.content_part.done","part":{"type":"output_text","text":"<full_text>"},"item_id":"<item_id>"}

[visa2]

Description

When an OpenAI API Key account has openai_responses_mode: "force_chat_completions", Sub2API converts the request to Chat Completions and sends it upstream. If the upstream model returns
reasoning_content in its SSE chunks (all DeepSeek models do: v4-pro, v4-flash, etc.), the SSE event ordering is wrong — response.reasoning_summary_text.delta is emitted before
response.output_item.added, causing Codex CLI to silently discard the entire stream.

Expected SSE Event Sequence

response.created
response.output_item.added ← item registered first
response.reasoning_summary_text.delta ← delta references existing item
response.reasoning_summary_text.delta
...
response.output_text.delta
response.output_text.delta
...
response.output_text.done
response.output_item.done
response.completed

Actual SSE Event Sequence

response.created
response.reasoning_summary_text.delta ← BAD: references output_index=0, but item not created yet
response.reasoning_summary_text.delta
... (dozens of reasoning deltas)
response.output_item.added ← item created too late
response.output_text.delta
...
response.output_text.done
response.output_item.done
response.completed

Root Cause

In ChatCompletionsChunkToResponsesEvents (chatcompletions_responses_bridge.go), the reasoning_content branch does not call ensureChatToResponsesMessageItem() before emitting the delta
event, while the content branch does. Since all DeepSeek models send reasoning_content chunks first (the thinking phase), the delta events arrive at Codex CLI referencing output_index=0
before any item at that index has been registered.

Codex CLI logs the error and drops the response:

ERROR codex_core::util: delta without active item

Reproduction

• Account platform: openai, type: apikey
• Account extra: "openai_responses_mode": "force_chat_completions" (or probe detected upstream does not support /v1/responses)
• Model: deepseek-v4-pro or deepseek-v4-flash
• Stream: true

Suggested Fix

Add ensureChatToResponsesMessageItem(state) before emitting reasoning_content delta, same as the content branch does:

// Before
if choice.Delta.ReasoningContent != nil {
_, _ = state.Reasoning.WriteString(*choice.Delta.ReasoningContent)
events = append(events, chatToResponsesEvent(state, "response.reasoning_summary_text.delta", ...))
}

// After
if choice.Delta.ReasoningContent != nil {
events = append(events, ensureChatToResponsesMessageItem(state)...) // <-- add this
_, _ = state.Reasoning.WriteString(*choice.Delta.ReasoningContent)
events = append(events, chatToResponsesEvent(state, "response.reasoning_summary_text.delta", ...))
}

And filter empty reasoning_content (DeepSeek sends "reasoning_content":"" in the first chunk) to avoid emitting a meaningless delta:

if choice.Delta.ReasoningContent != nil && *choice.Delta.ReasoningContent != "" {

Related

This issue coexists with #2875 (missing content_part.added) in the same code path. Both need to be fixed for Codex CLI to work correctly with force_chat_completions accounts using
DeepSeek or any reasoning model.

[jerryan999]

I have changed this.

The force_chat_completions chat-to-responses bridge now creates the message item before emitting response.reasoning_summary_text.delta, and empty reasoning_content deltas are filtered out. I also added a regression test covering reasoning-first streaming ordering. @visa2