Bug: non-streaming Responses-to-Chat conversion returns HTTP 400 "completed"

[beilo]

Environment

• Aether: v0.7.12
• Image revision: 06f5d3c8c0fec4599767749f6b321ec640ebb3da
• Client surface: OpenAI-compatible API
• Provider endpoint format: openai:responses

Problem

When Aether converts a successful non-streaming Responses API result to an OpenAI Chat Completions response, it appears to treat the top-level Responses status "completed" as an error. The outer /v1/chat/completions request then returns HTTP 400:

{
  "error": {
    "message": "completed",
    "type": "invalid_request_error",
    "code": "completed"
  }
}

The same model/provider succeeds through /v1/responses, and Chat Completions succeeds when streaming is enabled.

Minimal reproduction

# Fails
curl "$AETHER/v1/chat/completions" \
  -H "Authorization: Bearer $KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MODEL",
    "messages": [{"role": "user", "content": "Reply OK"}],
    "stream": false,
    "max_tokens": 32
  }'

# Succeeds
curl "$AETHER/v1/chat/completions" \
  -H "Authorization: Bearer $KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MODEL",
    "messages": [{"role": "user", "content": "Reply OK"}],
    "stream": true,
    "max_tokens": 32
  }'

# Also succeeds
curl "$AETHER/v1/responses" \
  -H "Authorization: Bearer $KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MODEL",
    "input": "Reply OK",
    "max_output_tokens": 32
  }'

Observed matrix:

Request	Result
/v1/chat/completions, stream:false	HTTP 400, completed error
/v1/chat/completions, stream:true	HTTP 200, normal SSE ending with finish_reason:"stop"
/v1/responses, non-streaming	HTTP 200, normal response with status:"completed"

Aether access logs for one request show the inner Responses request completing with 200, followed by the outer Chat Completions request returning 400 under the same trace ID:

POST /v1/responses         execution_runtime_sync status_code=200
POST /v1/chat/completions  execution_runtime_sync status_code=400

This reproduces with both text-only and image input, so it does not appear specific to multimodal request handling.

Expected behavior

A Responses payload with HTTP 200 and top-level status: "completed", no non-null error, and no type: "error" should be converted to a successful Chat Completion response.

Suspected cause

The sync/finalize response conversion appears to use the successful top-level status value as error text/code. completed should be recognized as a successful terminal status, not classified as invalid_request_error.

Likely relevant paths:

• apps/aether-gateway/src/execution_runtime/submission.rs
• Responses-to-Chat non-streaming compilation/finalize code in crates/aether-ai/formats

A regression test covering HTTP 200 + status:"completed" for openai:responses -> openai:chat with stream:false would catch this.