Vertex AI provider can't run two regions in the same session #28524

Observation

OpenCode loads custom providers only when their ID exactly matches a key in the CUSTOM_LOADERS dictionary in packages/opencode/src/provider/provider.ts. If you want two Vertex AI instances in one session, one pinned to europe-west4 and another pinned to the eu multi-region, declaring two provider entries with distinct IDs doesn't work. Only the canonical google-vertex ID triggers the Vertex-specific loader.

Reproduction

opencode.json:

{
  "provider": {
    "google-vertex": {
      "options": { "project": "P", "location": "europe-west4" },
      "models": { "gemini-2.5-pro": {} }
    },
    "google-vertex-eu": {
      "npm": "@ai-sdk/google-vertex",
      "options": {
        "project": "P",
        "location": "eu",
        "baseURL": "https://aiplatform.eu.rep.googleapis.com/v1/projects/P/locations/eu/publishers/google"
      },
      "models": { "gemini-3.5-flash": {} }
    }
  },
  "enabled_providers": ["google-vertex", "google-vertex-eu"]
}

Run:

opencode run --model google-vertex-eu/gemini-3.5-flash "ping"

Result:

Error: Model not found: google-vertex-eu/gemini-3.5-flash.
Error: {
  "name": "UnknownError",
  "data": { "message": "Unexpected server error. Check server logs for details." }
}

The same model on google-vertex/gemini-3.5-flash (canonical ID) works.

Where the behaviour comes from

• packages/opencode/src/provider/provider.ts:1320-1337: the custom-loader iteration walks Object.entries(custom(dep)), i.e. the canonical loader keys only. A custom-IDed entry never enters this loop.
• packages/opencode/src/provider/provider.ts:450-494: the google-vertex loader is where ADC token acquisition and the Authorization-injecting fetch handler get wired into options.fetch. Skipped for the alias.
• packages/opencode/src/provider/provider.ts:1524-1560: the fallback SDK resolver does instantiate @ai-sdk/google-vertex for the aliased ID via the bundled createVertex factory, but options.fetch is absent because the loader was skipped. Requests fail with no credentials attached.
• Several downstream call sites compare model.providerID === "google-vertex" directly (e.g. provider.ts:1431 and ~11 sites in transform.ts).

Expected behaviour

Two provider entries that share the same SDK (@ai-sdk/google-vertex) with distinct options should produce two independent SDK instances, each honouring its own project, location, and baseURL. That matches the upstream factory's intended usage: createVertex({ project, location }) returns one provider per call.

Environment

• OpenCode: built from dev branch (HEAD around 77da433e0)
• Bun 1.x
• @ai-sdk/google-vertex: 4.0.112
• Linux

Related

• @ai-sdk/google-vertex has separate upstream URL-construction bugs for the eu and global multi-region values (feat(google-vertex/anthropic): support eu/us multi-region endpoints vercel/ai#14634, @ai-sdk/google-vertex: Broken URL generated when using location: 'global' (Required for gemini-3.1-pro-preview) vercel/ai#14702). Out of scope here; the workaround is to set baseURL explicitly. This issue is only about the OpenCode-side gap that stops that workaround from being applied per-instance.
• Same code path is the cause of Cannot load custom providers from local file paths #4324 and overlaps with GitHub Copilot integration: Missing custom loader in provider.ts #3019.