feat: add plugin-litellm + provider-litellm presets (runtime model discovery)

[trick77]

What

Wires opencode to a LiteLLM proxy using runtime model discovery — the model list is pulled from the proxy at startup rather than hand-maintained in opencode.json. Two presets, matching the two config areas the plugin needs:

• plugin-litellm (@mode: append → plugin) — adds opencode-plugin-litellm (npm opencode-plugin-litellm, pinned @0.5.0). At startup the plugin queries the proxy's /v1/models and injects every model into the picker. For a proxy on localhost:4000/8000/8080 with the key in $LITELLM_API_KEY, the plugin auto-detects the URL and creates the litellm provider itself — so this preset alone is enough.
• provider-litellm (@mode: replace → provider.litellm) — for a custom/remote proxy URL. Prompts for baseURL, sets @ai-sdk/openai-compatible with apiKey: {env:LITELLM_API_KEY}, and declares no models (the plugin fills them).

Resulting config for the remote case matches the plugin README's "explicit provider" block verbatim:

{
  "plugin": ["opencode-plugin-litellm@0.5.0"],
  "provider": {
    "litellm": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "LiteLLM (proxy)",
      "options": { "baseURL": "...", "apiKey": "{env:LITELLM_API_KEY}" }
    }
  }
}

Why this shape (not a static provider block)

LiteLLM is not in models.dev, and opencode only auto-discovers models for a few hardcoded providers (gitlab, etc.) — a plain @ai-sdk/openai-compatible provider shows zero models unless each is listed by hand. The opencode-plugin-litellm plugin is the mechanism that makes "all models available"; it reads $LITELLM_API_KEY / $LITELLM_MASTER_KEY directly for discovery, which is why the key is env-based and the provider key must be exactly litellm.

Usage

# localhost proxy, key in env — plugin auto-detects everything
export LITELLM_API_KEY=sk-...
opencode-presets install plugin-litellm

# remote/custom proxy — add the provider block too (run separately; --set is one preset per call)
opencode-presets install plugin-litellm
opencode-presets install provider-litellm --set baseURL=https://litellm.example.com/v1

Restart opencode after installing a plugin preset.

Changes

• presets/plugin-litellm.conf — new
• presets/provider-litellm.conf — new
• README.md — two built-in-preset rows
• test/builtin-presets.test.ts — parseConf tests for both

Verification

• npm test → 99/99 pass (both new tests included)
• Temp-config install of both presets produces the exact documented config; validate config passes against opencode's live schema; remove plugin-litellm provider-litellm cleanly reverts.
• Not verified locally: runtime /v1/models discovery (needs a live LiteLLM proxy + opencode). Correctness rests on matching the plugin's documented config exactly.

Note

Plugin pinned to @0.5.0 rather than the README's @latest, per this repo's supply-chain stance. Bump in plugin-litellm.conf to update.