Zen lists muse-spark-1.3-contributor-free, but OpenCode 1.18.27 omits it and invocation returns UnknownError

Description

OpenCode Zen advertises muse-spark-1.3-contributor-free through the authenticated model endpoint, but the current OpenCode CLI both omits it from discovery and receives a server-side error when the documented model ID is invoked.

This is related to the provider-catalog work in #40524, but the failed request below needs Zen-side investigation too.

Environment

• Windows 11
• OpenCode CLI 1.18.27
• Authenticated OpenCode Zen credential
• No plugins involved
• The active Zen profile has no explicit model or models mapping

Verified facts

1. GET https://opencode.ai/zen/v1/models succeeds with the authenticated Zen credential and includes:

   • muse-spark-1.3-contributor-free
   • muse-spark-1.2-contributor-free

2. The official Zen documentation names the same model ID and labels it as available: https://dev.opencode.ai/docs/zen

3. opencode models --refresh succeeds, but omits opencode/muse-spark-1.3-contributor-free from the local picker/catalog.

4. A minimal invocation with no project or user data fails:

opencode run --model opencode/muse-spark-1.3-contributor-free "Reply with exactly: MUSE13_OK"

Result:

UnknownError
Unexpected server error. Check server logs for details.
ref: err_20d48c32

Expected behavior

A model returned by the authenticated Zen /v1/models endpoint should be discoverable without a manual local model entry and should accept a minimal request, or return a specific actionable availability/policy error.

Actual behavior

The client catalog is missing an API-advertised, workspace-enabled model, and the exact documented model ID reaches Zen but returns an opaque server error.

Requested fixes

• Reconcile the opencode provider picker with the authenticated Zen model endpoint as described in #40524.
• Investigate the Zen request identified by err_20d48c32 for muse-spark-1.3-contributor-free.
• Add an integration regression test: a model returned by Zen discovery is selectable and accepts a minimal inference request.

Start with the provider-catalog work in #40524 and reproduce opencode models --refresh and the minimal opencode run --model command. Compare the authenticated Zen /v1/models response with the local picker, then investigate request reference err_20d48c32. Done means the advertised model is selectable and the minimal request succeeds or returns an actionable availability or policy error, with an integration regression test.