feat(opencode): add local server provider with auto model discovery

[hmblair]

Issue for this PR

Closes #6231

Type of change

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

What does this PR do?

Adds a local provider to CUSTOM_LOADERS that auto-discovers models from any OpenAI-compatible /v1/models endpoint at startup. When configured with a baseURL, it fetches the model list, registers each model, and returns autoload: true. If the endpoint is unreachable or returns nothing, it returns autoload: false and the provider is silently skipped.

Also fixes the CUSTOM_LOADERS loop to allow custom loaders that don't have a models.dev entry — previously they were silently skipped because database[providerID] returned undefined and the loop hit continue.

Usage:

{
  "provider": {
    "local": {
      "options": {
        "baseURL": "http://localhost:11434/v1",
        "apiKey": "optional-key"
      }
    }
  }
}

This is an alternative to #17670 that solves the same problem in ~50 lines with no new config surface.

How did you verify your code works?

Three tests covering:

• Successful discovery from a mocked /models endpoint
• Graceful handling when the endpoint is unreachable
• Graceful handling when baseURL is not configured

Tested locally against llama-server and LM Studio.

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:

Potential Duplicate Found:

• PR feat(opencode): dynamic model discovery for local providers (LM Studio, llama.cpp, etc.) #17670: "feat(opencode): dynamic model discovery for local providers (LM Studio, llama.cpp, etc.)"
  • feat(opencode): dynamic model discovery for local providers (LM Studio, llama.cpp, etc.) #17670
  • Why related: The PR description explicitly mentions this PR (feat(opencode): dynamic model discovery for local providers (LM Studio, llama.cpp, etc.) #17670) as an alternative solution to the same problem. This PR takes a different, more concise approach (~50 lines with no new config surface) compared to that alternative.