Responses API SSE stream starts with a synthetic chatcmpl-dummy frame causing TypeValidationError #30916

Description

Summary

When using the OpenAI Responses API (sdk.responses()) with models served through certain OpenAI-compatible gateways, the SSE stream begins with a synthetic chatcmpl-dummy frame in Chat Completions format before the real Responses API events. This frame is not a valid Responses API event and causes a TypeValidationError that terminates the agent turn.

Environment

• OpenCode version: 1.16.0 (dev), 1.15.13 (release)
• Provider: OpenAI (@ai-sdk/openai)
• Model: gpt-5.5 (via OpenAI-compatible gateway)
• Runtime: ai-sdk

Reproducing

1. Verification command

Set up environment variables for your OpenAI-compatible provider:

export OPENAI_BASE_URL="<your-endpoint>/v1"
export OPENAI_API_KEY="<your-api-key>"

Run opencode with the Responses API (default) and Chat Completions API:

# Test with gpt-5.5 (Responses API — fails)
opencode run "say hi" --format json --model openai/gpt-5.5 --dangerously-skip-permissions

2. gpt-5.5 fails with TypeValidationError

The stream errors immediately:

{
  "type": "error",
  "error": {
    "name": "UnknownError",
    "data": {
      "message": "Type validation failed: Value: {\"id\":\"chatcmpl-dummy\",\"object\":\"chat.completion.chunk\",\"created\":1780651633,\"model\":\"gpt-5.5\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\"}}]}.\nError message: [\n  {\n    \"code\": \"invalid_union\",\n    \"errors\": [\n      [\n        {\n          \"code\": \"invalid_value\",\n          \"values\": [\"response.output_text.delta\"],\n          \"path\": [\"type\"],\n          \"message\": \"Invalid input: expected \\\"response.output_text.delta\\\"\"\n        },\n        ...\n      ],\n      ...\n    ],\n    \"message\": \"Invalid input\"\n  }\n]"
    }
  }
}

The error shows chatcmpl-dummy — a Chat Completions format frame with "content":"" — being passed to the Responses API event validator, which rejects it because it does not match any expected event type (response.output_text.delta, response.created, etc.).

3. deepseek-v4-pro works fine

The same setup works correctly with DeepSeek's native API (which does not go through the same gateway):

opencode run "say hi" --format json --model deepseek/deepseek-v4-pro --dangerously-skip-permissions

Expected output:

{"type":"step_start","part":{"type":"step-start"}}
{"type":"text","part":{"type":"text","text":"hi"}}
{"type":"step_finish","part":{"type":"step-finish"}}

4. Root cause confirmed via curl

Probing the Responses API endpoint directly reveals the synthetic frame:

curl -N -s "${OPENAI_BASE_URL}/responses" \
  -H "Authorization: Bearer ${OPENAI_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-5.5","input":"hello","stream":true}'

The very first SSE event is:

data: {"id":"chatcmpl-dummy","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"role":"assistant","content":""}}]}

Note the content: "" — this frame carries no actual content. It is a bare data: frame (no event: prefix) with object: "chat.completion.chunk" — it is not a valid Responses API event. The subsequent frames are proper Responses API events (event: response.created, event: response.output_text.delta, etc.).

The same test against DeepSeek's API shows no such frame:

curl -N -s "${DEEPSEEK_BASE_URL}/responses" \
  -H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-v4-pro","input":"hello","stream":true}'

Output starts directly with a proper Responses API event — no chatcmpl-dummy prefix.

5. Suggested fix

The chatcmpl-dummy frame has a hardcoded ID ("chatcmpl-dummy"). The fix is to filter it out at the HTTP response level in the AI SDK's custom fetch wrapper, before the stream reaches the event validator.

Location: packages/opencode/src/provider/provider.ts — in the custom fetch wrapper inside resolveSDK().

Approach: When the response is an SSE stream from a /responses endpoint, pipe the body through a TransformStream that parses each SSE event's data: line as JSON and drops any event whose id equals "chatcmpl-dummy".

This approach:

• Precisely targets only the synthetic chatcmpl-dummy frame
• Does not affect normal Chat Completions or Responses API streams
• Requires no user configuration changes
• Is transparent to all other providers

Plugins

No response

OpenCode version

No response

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response