feature: Ability to select model for prompts

[nickpoorman]

Did you check the docs?

• I have read all the docs

Is your feature request related to a problem? Please describe.

First off this plugin is awesome and the ability to configure prompts is crazy powerful!

For certain prompts, I need to use specific models.

Describe the solution you'd like

I'm fairly certain the opencode API allows us to specify a model when sending chat messages. It would be really awesome if I could specify the model with the prompts in the prompts config like this. That way, I can automatically use the correct model for each prompt.

prompts = {
    ---@class opencode.Prompt
    ---@field description string Description of the prompt. Shown in selection menu.
    ---@field prompt string The prompt to send to `opencode`, with placeholders for context like `@cursor`, `@buffer`, etc.
    explain = {
      description = "Explain code near cursor",
      prompt = "Explain @cursor and its context",
      model = "anthropic/claude-opus-4-1-20250805`
    },
    fix = {
      description = "Fix diagnostics",
      prompt = "Fix these @diagnostics",
      model = "anthropic/claude-sonnet-4-20250514`
    },
    --- ...

Describe alternatives you've considered

Create an opencode agent and then @agent. But that requires me to configure agents and these prompts are more tied to my nvim workflow so it's great to be able to keep them with my nvim config.

Additional context

No response

[nickjvandyke]

Thanks for the kind words and suggestion!

You are correct that opencode's API allows passing a model when POSTing a message. However the plugin uses the TUI endpoints (append prompt to TUI + send prompt in TUI) because then the TUI behaves the same as if you actually entered the prompt in it. Whereas when POSTing a message to a session, there's no way for the plugin to query nor modify the TUI's selected session (last I checked). I started with that approach and it made for an inconsistent and confusing UX.

So, I think @agent may be the best you can do, unless opencode extends the API as needed!

[nickjvandyke]

these prompts are more tied to my nvim workflow so it's great to be able to keep them with my nvim config.

You could use this to point to an agent config file only when launching from the plugin? https://opencode.ai/docs/config/#custom-path Configure it as opts.terminal.env.OPENCODE_CONFIG = ...

I think I saw a release with OPENCODE_CONFIG_CONTENT which might even let you inline it without a file? But the docs don't seem to mention it.

[nickjvandyke]

Closing because it's not possible currently, but if opencode's API allows it in the future, feel free to re-open!

[mortenhaahr]

Hi @nickjvandyke.
I know you have done a lot of work recently with the addition of servers. Does this change anything in terms of implementing @nickpoorman's feature request?

I would also very much like this feature. It seems that token prices are increasing, and having the option of dynamically changing the model for a session would be incredible. Specifically, what I am hoping to configure is that with one vim keybinding "ai" I am using a cheap and fast model, but for "AI" I use a more detailed model.

[nickjvandyke]

@mortenhaahr Unfortunately no 😞 But imo the @agent approach is still pretty good. I would create two agents in your opencode config, say fast and slow with appropriate models set for each, and then bind your keybindings to pre-fill accordingly. i.e. ai calls require("opencode").ask("@fast ")

[mortenhaahr]

Thanks a lot for the quick reply. I think your suggested approach is exactly what I need!

[mortenhaahr]

@nickjvandyke I tried this approach, but it doesn’t quite work. Likely due to OpenCode quirks rather than the plugin itself. Sharing here for visibility.

Test setup: ran opencode serve and used the web UI. I created an agent configured with a cheap model that always returns a fixed string, while the chat panel used an expensive model.

When triggering the agent via the prompt "@fast hello", the request was sent using the expensive model and returned a generic response (“Hello! How can I help you today?”). The usage page confirmed it used the expensive model.

However, when I manually selected the agent in the chat panel, it correctly switched to the cheap model and used that for the request.

In other words, using @... in the prompt does not guarantee that the agent’s configured model will be used.

Are there any other approaches that could work?

[nickjvandyke]

Thanks for testing! Darn. I wonder if that's a regression in opencode. It used to pick up both @<file> and @<subagent> references in prompts appended to the TUI. In v1.0.0, the former broke, but I thought it still applied the latter. Guess not 😕

Unfortunately I'm not aware of any further workarounds. The message API supports a model parameter but that defeats the purpose of leveraging OpenCode's TUI.

[mortenhaahr]

That is a shame. Thanks for looking into this and for the detailed explanation.

[nickjvandyke]

Np, thanks for your help! Feel free to open a feature request (or even PR) in https://github.com/anomalyco/opencode. I would really love to have it (and the files for speed) too.

Guess I have to remove the subagent references from the docs for now 😞

[mortenhaahr]

Done: anomalyco/opencode#28809