feat(opencode): dynamic model discovery for local providers (LM Studio, llama.cpp, etc.)

[dmitryryabkov]

Issue for this PR

Closes #6231

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Adds an option for dynamic model list population for OpenAI-compatible providers supporting /model API (as opposed to manually typing them in opencode.json):

• There's a new option which can be added to a "provider" in opencode.json: "dynamicModelList": true
• If the option is not set, the old behavior takes precedence (backward compatibility)
• When set to true issues an API request and retrieves the list of the models supported by the provider
• If the list is returned successfully, populates the models from the response
• In the explicit list of models is provided, it takes precedence even if the "dynamicModelList": true flag is set

This change makes working with local AI inference engines much nicer, because it eliminates the need to update opencode.json every time a new model is added, or if the engine is started with the specific model as a parameter. The model list is now completely driven by the inference engine itself.

Example configuration (opencode.json):

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "llama.cpp": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "llama.cpp (local)",
      "options": {
        "baseURL": "http://127.0.0.1:8080/v1"
      },
      "dynamicModelList": true
    },
    "lm_studio": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "LM Studio (local)",
      "options": {
        "baseURL": "http://127.0.0.1:1234/v1",
      },
      "dynamicModelList": true
    }
  }
}

will result in the following model list in OpenCode, for LM Studio with "Just-in-Time Model Loading" enabled:

or llama.cpp started without specifying a model:

The selected model will be loaded by the inference engine once the first /chat/completion request is received.

If, however, the inference engines only serving specific models, only those models will be returned and available in OpenCode model selector. For LM Studio with "Just-in-Time Model Loading" disabled:

or llama.cpp started with a specific model:

I'm aware that there's a number of opened PRs on this issue already (i.e. #15732, #13234), however I believe this is the most robust implementation because:

• it doesn't hard-code any specific provider name
• it maintains the existing behavior by default (if the new flag is not used)
• it works with any provider which supports /model endpoint (not even just a local one, i.e. Ollama Cloud also works just fine)
• adds a bunch of tests

🤖 Developed with some help from OpenCode running against a local model hosted in llama.cpp!

How did you verify your code works?

Run OpenCode with models hosted in LM Studio and llama.cpp. Tried various scenarios:

• with a single model loaded vs dynamic model loading (see the screenshots above)
• with/without auth
• with the inference engine down

Also tested it agains Ollama Cloud

Screenshots / recordings

See above

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Potential Duplicate PRs Found

1. PR feat(opencode): add dynamic configuration and context discovery for LM Studio #15732: "feat(opencode): add dynamic configuration and context discovery for LM Studio"

   • feat(opencode): add dynamic configuration and context discovery for LM Studio #15732
   • Related: This is specifically mentioned in the current PR as an existing attempt at this feature. It addresses dynamic model loading for LM Studio.

2. PR feat(opencode): add LiteLLM provider with auto model discovery #14468: "feat(opencode): add LiteLLM provider with auto model discovery"

   • feat(opencode): add LiteLLM provider with auto model discovery #14468
   • Related: Implements auto model discovery functionality, though for LiteLLM provider specifically rather than generic OpenAI-compatible providers.

3. PR feat(opencode): add auto loading models for litellm providers #13896: "feat(opencode): add auto loading models for litellm providers"

   • feat(opencode): add auto loading models for litellm providers #13896
   • Related: Similar auto-loading functionality for LiteLLM providers.

The current PR (#17670) explicitly acknowledges PRs #15732 and #13234 as prior attempts but argues that this implementation is more robust because it's provider-agnostic, maintains backward compatibility, and works with any OpenAI-compatible provider supporting the /models endpoint.