OpenCode supports provider passthrough controls

OpenCode’s documentation now spells out a configuration pattern many enterprise AI teams have been trying to balance: one stable agent abstraction, with explicit room for provider-specific controls. The company’s agents page says that “any other options” included in an agent configuration are passed through directly to the provider as model options. (opencode.ai) That means a team can keep a standard top-level agent definition — prompt, model choice, tool access and permissions — while still attaching knobs that only exist on a given provider. (platform.openai.com) In OpenAI’s case, that includes reasoning controls for reasoning-capable models. (opencode.ai) OpenAI’s documentation says the `reasoning.effort` parameter guides how many reasoning tokens a model generates before responding, with lower settings favoring speed and economy and higher settings favoring more complete reasoning. ### Where does OpenCode say this works? OpenCode’s agents documentation says provider-specific options can be added directly inside agent configuration, and those values are forwarded as model options to the underlying provider. (opencode.ai) The page describes those fields as “model and provider-specific” and tells users to check provider documentation for supported parameters. OpenCode’s broader docs show the same pattern at the model layer. The models page includes examples for OpenAI and Anthropic settings, including `reasoningEffort`, `textVerbosity`, `reasoningSummary`, and Anthropic “thinking” budgets, and says agent config can override global options. (platform.openai.com) ### What is the concrete OpenAI control here? OpenAI’s API documentation identifies reasoning effort as a supported control on reasoning models. The company says `reasoning.effort` can be set to levels including low, medium and high, and that lower effort can reduce latency and token use while higher effort can produce more complete reasoning. (opencode.ai) OpenAI’s model documentation also describes GPT-5-family models as offering configurable reasoning effort, and newer pages describe related controls such as verbosity and reasoning summaries. (opencode.ai) ### Why does this matter for teams using multiple providers? Federated engineering groups often want two things at once: portability across providers and access to provider-native features. OpenCode’s approach preserves a common agent interface while leaving a documented escape hatch for settings that are not universal. (platform.openai.com) The docs do not present the passthrough as automatic model normalization; they present it as an explicit handoff to the provider’s own option surface. That design also makes changes easier to audit than ad hoc per-team forks. Because the option is declared in configuration rather than hidden in custom wrappers, the setting can be inspected alongside the rest of the agent definition. (platform.openai.com) OpenCode’s published docs stop short of promising a specific run-metadata schema for every passthrough field, but they do show a typed configuration path rather than an undocumented override. ### Does this replace the need for provider docs? OpenCode’s documentation says no. The agents page tells users to consult provider documentation for available parameters, and the CLI docs point users to `opencode models` to see exact provider and model names available in their setup. (opencode.ai) The practical next step is in those same docs: teams can define or modify agents in configuration, verify supported provider parameters against OpenAI or other vendor documentation, and inspect available models through OpenCode’s CLI and model pages. (opencode.ai)