opencode go 套餐中部分模型使用报错【API key is invalid】问题临时解决方案

[simplifyOurLife]

我理解问题原因是deepseek harness对于opencode的go套餐可用模型还是停留在之前的可用模型列表，所以目前在添加模型的时候需要选择自定义，后续待官方同步后再配置

然后填好API Key，在API地址处填写：https://opencode.ai/zen/go/v1。就可以点击获取最新的可用模型列表自行裁剪了

[yuhengShii]

opencode 那边有个选项 勾选 启用部署在中国的模型 就可以了

[simplifyOurLife]

这个一直是勾选的，如果可以的话我就不这么折腾了...

[denial123789]

There is a second boundary worth preserving here: a successful Go model listing does not make every returned model callable through one DSH custom route.

The current official OpenCode Go table mixes three endpoints and protocols under the same https://opencode.ai/zen/go/v1 prefix:

• DeepSeek V4 models: /chat/completions → openai-completions
• GPT 5.6 Luna: /responses → openai-responses
• Qwen and MiniMax entries: /messages → anthropic-messages

In DSH rc.8, Fetch available models reads ids, optional names, and optional capacity fields from GET /models. It does not import a per-model protocol. A custom provider route has one route-level api, while its models entries cannot each name a different protocol. Adopting the full mixed listing into one route can therefore make only one protocol family work and leave the others looking like credential failures.

A safer setup is three protocol-homogeneous DSH provider routes sharing the same credential reference and Go base prefix. The Messages route must be entered manually because rc.8 intentionally limits network discovery to OpenAI Completions and Responses listings.

The reported account option for models deployed in China is a separate entitlement and data-residency decision. It can change model availability, but it should not be treated as a universal fix for a wrong request path.

Full rc.8 routing examples and evidence matrix: canonical guide · visual map.

Disclosure: I maintain the independent SandBase DeepSeek Harness Handbook.

[simplifyOurLife]

Understood. Then performing the configuration multiple times according to the protocol should be a better temporary solution.

Provider	API protocol	Model
opencode-go-chat	openai-completions	DeepSeek V4 Pro/Flash、GLM 5.x、Kimi K3/K2.x、MiMo 2.5/Pro、Hy3
opencode-go-responses	openai-responses	GPT 5.6 Luna、Grok 4.5、Muse Spark
opencode-go-messages	anthropic-messages	MiniMax M3/M2.7/M2.5、Qwen 3.8/3.7/3.6

[weijiafu14]

你最后整理的“三条协议同质 route”方案是正确边界：Go 目录混合 Chat Completions、Responses 和 Anthropic Messages，而一个 route 只能有一个 api。不要把 /models 返回的全部模型塞进同一 route；应共享凭证并按协议拆成 chat/responses/messages，且每条协议各跑一次真实生成。

[simplifyOurLife]

感谢大佬的确认