[Zen Go] /v1/responses returns 500 for all non-DeepSeek models (mimo/glm/ox-alpha) while /v1/chat/completions works

Zen Go gateway: /v1/responses returns 500 for every non-DeepSeek model while /v1/chat/completions works

Summary

Since ~2026-08-25 13:00 UTC+8, the Go gateway (https://opencode.ai/zen/go/v1/responses) returns 500 Internal server error for all non-DeepSeek models, while the same models keep working via https://opencode.ai/zen/go/v1/chat/completions. This breaks Codex (desktop & CLI) users on the Go subscription, because Codex speaks the Responses wire protocol exclusively.

This looks like a regression in the gateway's chat→responses adapter layer. Notably glm-5.2 was already served natively over Responses (verified 2026-08-23) and is now also 500ing, so the breakage seems to affect the whole non-deepseek responses path.

Evidence matrix (probed directly against opencode.ai, bypassing any local proxy; 2026-08-25 14:10–14:20 UTC+8)

Minimal reproduction

# 500 — responses endpoint
curl -sS -X POST https://opencode.ai/zen/go/v1/responses \
  -H "Authorization: Bearer $ZEN_API_KEY" -H 'Content-Type: application/json' \
  -d '{"model":"mimo-v2.5","input":[{"role":"user","content":[{"type":"input_text","text":"hi"}]}],"max_output_tokens":64,"stream":false,"store":false}'
# → HTTP 500 {"type":"error","error":{"type":"error","message":"Internal server error"}}

# 200 — same model, chat endpoint, same moment
curl -sS -X POST https://opencode.ai/zen/go/v1/chat/completions \
  -H "Authorization: Bearer $ZEN_API_KEY" -H 'Content-Type: application/json' \
  -d '{"model":"mimo-v2.5","messages":[{"role":"user","content":"hi"}],"max_tokens":32}'

Also reproducible with "stream":true, with string-style "input", and for every model listed above.

Impact

All Responses-wire clients (Codex family) on Go subscriptions cannot use mimo/glm/ox-alpha at all; only DeepSeek-family models remain usable. OpenCode desktop keeps working for these models because it uses the chat-completions path.

Request

Either restore the responses path for these models, or expose a per-model wire-format hint so clients can pin chat-completions. Happy to provide more traces if needed.