You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Go endpoint: kimi-k3 rejects 'n' parameter (400: n: Input should be None) #47568
The OpenCode Go endpoint (https://opencode.ai/zen/go/v1/chat/completions) returns HTTP 400 for kimi-k3 whenever the request includes the n parameter — even the OpenAI-default n: 1:
{
"error": {
"type": "server_error",
"message": "Error from provider (Console Go): Upstream request failed: n: Input should be None"
}
}
-> 400 with the error above. The same request withoutn succeeds.
Scope
Affects kimi-k3 with any reasoning_effort value (low/medium/high/max) and also with no effort parameter at all.
Other Go models tested (kimi-k2.6, kimi-k2.7-code, glm-5.3, minimax-m3, mimo-v2.5) accept n: 1 fine — the bug is specific to kimi-k3's upstream validation.
Impact
Any OpenAI-compatible client that always sends n: 1 (VS Code Copilot Chat custom models, several SDKs) cannot use kimi-k3 on the Go plan at all. Presumably the upstream model's pydantic schema declares n: None = None and rejects any provided value instead of ignoring n=1.
This issue might be a duplicate of or related to existing issues about the Go gateway forwarding parameters that strict upstream validators reject. Please check:
The n parameter issue you describe is a distinct specific case (the upstream kimi-k3 model declares n: None and rejects any provided value, including the OpenAI-default n=1), but it falls into the same general class of the gateway not sanitizing/stripping incompatible parameters per-model before forwarding.
Thanks for the clear report and the curl reproduction. Testing this needs a real OpenCode Go API key so the request reaches the kimi-k3 upstream, and our automated environment doesn't have one. Without a key, the endpoint rejects the request before it gets as far as checking n. The problem is in the hosted Go gateway (or the upstream behind it) rather than in the OpenCode app, which doesn't send n itself, so a maintainer with Go access will need to confirm it.
Activity
github-actions commentedon Sep 6, 2026
This issue might be a duplicate of or related to existing issues about the Go gateway forwarding parameters that strict upstream validators reject. Please check:
mcp,system) are forwarded verbatim to strict provider validators — same root cause class (gateway not stripping provider-incompatible fields before forwarding).The
nparameter issue you describe is a distinct specific case (the upstream kimi-k3 model declaresn: Noneand rejects any provided value, including the OpenAI-defaultn=1), but it falls into the same general class of the gateway not sanitizing/stripping incompatible parameters per-model before forwarding.opencodejr commentedon Oct 28, 2026
Thanks for the clear report and the curl reproduction. Testing this needs a real OpenCode Go API key so the request reaches the kimi-k3 upstream, and our automated environment doesn't have one. Without a key, the endpoint rejects the request before it gets as far as checking
n. The problem is in the hosted Go gateway (or the upstream behind it) rather than in the OpenCode app, which doesn't sendnitself, so a maintainer with Go access will need to confirm it.Related reports:
/zen/go/v1/responses):deepseek-flashnever returnsphaseon assistant messages, which breaks Codex desktop's turn collapsing #57372: another behaviour bug on the Go endpoint