Desktop: Vertex Anthropic models resolve to wrong region + invalid @default version in new worktrees

[lenrt]

Description

In opencode Desktop, the model selection dropdown shows a different (incorrect) set of Vertex Anthropic models in newly created git worktrees compared to existing/older worktrees of the same project.

In older worktrees, the dropdown correctly lists region-appropriate Vertex Anthropic models (e.g. Claude 4.7 Opus under "Vertex Anthropic (EU Multi-Region)" and Claude Sonnet 4.6 under "Vertex Anthropic (europe-west1)").
In newly created worktrees, the dropdown instead lists models that are not configured the same way (e.g. Claude Opus 4.8 and Claude Sonnet 4.6 under a generic "Vertex (Anthropic)" group), and selecting them fails.

When Claude Opus 4.8 is selected in a new worktree, the request fails with:
Publisher Model projects/devtools-288204/locations/europe-west1/publishers/anthropic/models/claude-opus-4-8@default was not found or your project does not have access to it. Please ensure you are using a valid model version.
Two things appear wrong in this resolved model path:

An invalid version suffix @default is appended to the model ID. Vertex expects either a concrete version or no suffix; @default is not a valid Vertex model-version identifier.
The model is resolved against the regional endpoint locations/europe-west1, where this model is not available for the project, instead of the multi-region / correct endpoint used successfully in the older worktree.

Plugins

none

OpenCode version

v1.17.7

Steps to reproduce

1. Have a project configured with Vertex Anthropic models via ADC (gcloud auth application-default login), using EU endpoints (EU Multi-Region + europe-west1).
2. Open an existing/older worktree → model dropdown shows correct models, selection works.
3. Create a new git worktree of the same project and open it in opencode Desktop.
4. Open the model dropdown → observe a different model set.
5. Select a Vertex Anthropic model → request fails with the @default / wrong-region error above.

Screenshot and/or share link

Operating System

No response

Terminal

No response

[github-actions]

This issue might be a duplicate of or closely related to existing issues. Please check:

• Worktrees resolve to different project IDs due to per-worktree cache lookup #16995: Worktrees resolve to different project IDs due to per-worktree cache lookup — the different model sets appearing in new vs old worktrees may be a symptom of this same root cause (new worktrees getting a different project identity and thus a different config context).
• Vertex AI provider can't run two regions in the same session #28524: Vertex AI provider can't run two regions in the same session — covers the same region-resolution and multi-region Vertex provider limitations that cause the @default suffix and wrong-region errors you're seeing.

[ghchinoy]

Adding a related data point from the TUI (not Desktop) that corroborates the wrong-region routing and adds more detail on the env var fallback chain that causes it.

Environment

• OpenCode 1.17.14 (TUI, Homebrew install)
• macOS, google-vertex provider (not google-vertex-anthropic)
• VERTEX_LOCATION=global set in shell
• Google Cloud project: standard ADC via gcloud auth application-default login

What happened

claude-opus-4-8@default failed on every attempt. The OpenCode UI reported
"The response was blocked by the provider's content filter." Logs showed the request going to google-vertex with no error written — the session silently fell back to claude-sonnet-4-6@default.

Direct API test confirmed the actual failure:

POST https://us-east5-aiplatform.googleapis.com/v1/projects/.../locations/us-east5/publishers/anthropic/models/claude-opus-4-8@default:streamRawPredict
→ HTTP 404 NOT_FOUND

POST https://aiplatform.googleapis.com/v1/projects/.../locations/global/publishers/anthropic/models/claude-opus-4-8@default:streamRawPredict
→ HTTP 200  (model responds normally)

Opus 4.8 is only available on the global endpoint. Regional endpoints return 404.

Root cause: env var fallback chain

OpenCode's google-vertex provider (Gr() function in the bundle) resolves location via this fallback chain:

o.location → GOOGLE_VERTEX_LOCATION → GOOGLE_CLOUD_LOCATION → VERTEX_LOCATION → "us-central1"

My shell had VERTEX_LOCATION=global — which is the third fallback — but GOOGLE_VERTEX_LOCATION was not set. The provider silently fell through the chain andlanded on us-central1 (the hardcoded default), not global.

This is a subtle trap: VERTEX_LOCATION is a common env var name for GCP users (it's what the Vertex AI Python SDK and gcloud docs suggest), but OpenCode's canonical var is GOOGLE_VERTEX_LOCATION. If a user has only VERTEX_LOCATION set, they get us-central1 silently.

Fix that worked

# ~/.zshrc
export GOOGLE_VERTEX_LOCATION=global
export GOOGLE_VERTEX_PROJECT=my-project-id

After adding these, claude-opus-4-8@default routes to the global endpoint and responds correctly.

Suggestions

1. Document GOOGLE_VERTEX_LOCATION as the canonical env var — or check VERTEX_LOCATION earlier in the chain (before the us-central1 default), since it's the more commonly set name.

2. Surface the actual endpoint URL in error logs when a Vertex request fails, so users can tell whether routing went to global vs a regional endpoint without needing to instrument raw API calls.

3. Consider defaulting google-vertex to global rather than us-central1, since global improves availability (per Vertex docs) and newer Anthropic models are often global-only.

Note: google-vertex-anthropic already defaults to "global" — the inconsistency between the two providers is confusing when both appear in the model picker for the same model.