feat(opencode): add dynamic model fetching for OpenAI-compatible provider

[shivamashtikar]

Issue for this PR

Closes #12814

Type of change

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

What does this PR do?

Adds a fetchModels config option for providers that dynamically fetches available models from the provider's API at startup, eliminating the need to manually list every model in opencode.json.

How it works:

1. When fetchModels: true is set on a provider config, the provider initialization calls the API to discover models.
2. It first tries the LiteLLM-specific /model/info endpoint, which returns rich metadata (context/output limits, costs, vision/reasoning/toolcall capabilities).
3. If /model/info is unavailable (non-LiteLLM provider), it falls back to the standard OpenAI-compatible /models endpoint with sensible defaults.
4. Manually configured models in opencode.json always override fetched ones, so users can still fine-tune specific models.

Changes:

• config.ts: Added fetchModels boolean option to the provider config schema.
• provider.ts: Added fetchModelInfo, fetchModelList, and fetchModels functions. Integrated into state() after config processing so manual models take precedence.

Example config:

{
  "provider": {
    "litellm": {
      "npm": "@ai-sdk/openai-compatible",
      "fetchModels": true,
      "options": {
        "baseURL": "https://my-litellm-proxy.com",
        "apiKey": "{env:API_KEY}"
      }
    }
  }
}

Works for any OpenAI-compatible provider (LiteLLM, Ollama, vLLM, etc).

How did you verify your code works?

Ran bun run --conditions=browser ./src/index.ts models litellm against a LiteLLM proxy. Verified:

• All 16 models auto-populated from /model/info endpoint
• Context limits, output limits, costs, and capabilities (vision, reasoning, PDF, toolcall) correctly mapped
• Models with /model/info data get real limits (e.g., claude-opus-4-5: 1M context, 128K output) instead of defaults
• Fallback to /models works when /model/info is unavailable
• Manual model overrides in config take precedence over fetched data
• Built binary with --single flag and confirmed it works

Screenshots / recordings

N/A - no UI changes

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:

Found a potential duplicate:

PR #13896 - feat(opencode): add auto loading models for litellm providers

• feat(opencode): add auto loading models for litellm providers #13896

Why it's related: This PR appears to implement very similar functionality - automatic model loading for LiteLLM providers. Given that PR #14277 also adds dynamic model fetching for OpenAI-compatible providers (including LiteLLM), these PRs may be addressing the same or overlapping feature requests. You should verify whether PR #13896 was already merged or closed and check if #14277 extends/improves upon it.

[shivamashtikar]

this PR provides compatibility with any open-ai compatible endpoints along with litellm as opposed to #13896 which only handles for litellm

[PratikNarola1]

LGTM

[8x22b]

PLEASE MERGE IT

[espetro]

up! we're looking forward to it!