Summary
OpenCodex can successfully work around providers whose Responses/chat streaming is unreliable by setting a per-model upstream-streaming policy to false. However, the web-search loop currently forces stream:true for each internal model iteration, bypassing that policy.
This caused GPT-5.6 Luna through OpenCode Go to keep failing in Codex Desktop even after the provider registry and main Responses normalization paths were configured for buffered upstream requests.
Environment
- OpenCodex 2.10.2
- macOS
- Codex Desktop
- Provider adapter:
openai-chat
- Provider: OpenCode Go
- Model:
gpt-5.6-luna
Reproduction
- Configure a routed model with a provider registry entry such as:
modelResponsesUpstreamStreaming: { "gpt-5.6-luna": false }
- Send a Codex Responses request with
stream:true, tools, and hosted web_search_preview enabled.
- The normal normalization path changes the upstream request to non-streaming.
src/web-search/loop.ts creates an iteration with stream: true and calls the adapter with stream: true.- The provider receives a streaming tool request and fails; OpenCodex returns the truncation error/502.
Without the hosted web-search tool, the buffered compatibility path works. With it, the loop reintroduces upstream streaming.
Expected behavior
The web-search loop should preserve the effective upstream streaming decision for the routed provider/model. A client may still receive downstream SSE synthesized from a completed buffered response, while the upstream request remains non-streaming.
Fix direction
- Do not unconditionally set
iterParsed.stream = true in the web-search loop.
- Pass the effective stream setting to
fetchResponse.
- Use
parseResponse for buffered JSON iterations and parseStream only for genuinely streaming upstream responses.
- Keep the client-facing web-search loop and SSE behavior unchanged.
This report intentionally excludes API keys, private prompts, and workspace/tool payloads.
Summary
OpenCodex can successfully work around providers whose Responses/chat streaming is unreliable by setting a per-model upstream-streaming policy to false. However, the web-search loop currently forces
stream:truefor each internal model iteration, bypassing that policy.This caused GPT-5.6 Luna through OpenCode Go to keep failing in Codex Desktop even after the provider registry and main Responses normalization paths were configured for buffered upstream requests.
Environment
openai-chatgpt-5.6-lunaReproduction
stream:true, tools, and hostedweb_search_previewenabled.src/web-search/loop.tscreates an iteration withstream: trueand calls the adapter withstream: true.Without the hosted web-search tool, the buffered compatibility path works. With it, the loop reintroduces upstream streaming.
Expected behavior
The web-search loop should preserve the effective upstream streaming decision for the routed provider/model. A client may still receive downstream SSE synthesized from a completed buffered response, while the upstream request remains non-streaming.
Fix direction
iterParsed.stream = truein the web-search loop.fetchResponse.parseResponsefor buffered JSON iterations andparseStreamonly for genuinely streaming upstream responses.This report intentionally excludes API keys, private prompts, and workspace/tool payloads.