dsh-llm-opencode-go

Web configuration

Open Settings → Plugins → OpenCode Go. The card stores the API key through the Harness credentials service under OPENCODE_GO_API_KEY; the host never returns the stored literal. It also edits the endpoint, the refresh interval, and the model overrides, and it can interrogate the endpoint for the ids it currently advertises.

The card is the supported editor for this provider because the Settings → Models page ships layouts for the two namespaces it owns (llm-deepseek, llm-pi-ai) and renders a deployment-added provider as a row with no editor. The Plugins page dispatches a card by settings namespace instead, which is exactly what a plugin distributed outside the harness repository needs.

Configuration

Everything except the key is a settings field, layered over the bundle row by the llm-opencode-go section of $DSH_HOME/settings.yaml:

llm-opencode-go:
  apiRoot: https://opencode.ai/zen/go
  refreshIntervalMs: 3600000        # hourly; 0 disables the timer
  refreshOnMount: true
  models:
    - id: some-new-model
      api: anthropic-messages       # openai-completions | openai-responses | anthropic-messages
      contextWindow: 262144
      maxTokens: 65536
      reasoning: true
      input: [text, image]

input is enforced: a prompt carrying an image is refused with UNSUPPORTED_CONTENT before the request when the model does not list image. reasoning: false hides the harness thinking levels for that model.

The credential is not a settings field — secrets go through the credentials service, which stores them in $DSH_HOME/.credentials.yaml:

refs:
  OPENCODE_GO_API_KEY: sk-...

The reference deliberately avoids the conventional OPENCODE_API_KEY: an exported environment value shadows the managed store, so a shell that exports another OpenCode key would silently win.

Shell export warning. If your shell exports OPENCODE_GO_API_KEY, that value wins over the store and the card's API key field renders as read-only. Unset it, or point apiKeyEnv at a reference nothing exports.