403 "This model is not available in your region" with opencode-go relay provider for gpt-5.6-luna #40587

Description

When using the built-in opencode-go provider to call gpt-5.6-luna, every request fails with:

AI_APICallError: Upstream request failed: [403] This model is not available in your region.

The error is returned by the opencode-go relay and transparently passed through to the client. It is consistent across ~1 hour of retries, on both the title (small) and build (non-small) agents. Other models on the same provider (e.g. kimi-k3) work fine, which suggests gpt-5.6-luna specifically is geo-restricted at the upstream OpenAI side, and the opencode-go relay's egress IP falls into a disallowed region.

Environment:

• OpenCode Desktop 1.18.13 (packaged, Windows x64)
• Node 24.15.0, Electron 42.3.3
• OS: Windows 11 (10.0.26200)
• ~/.config/opencode/opencode.jsonc has disabled_providers: ["opencodego"] — id mismatch (actual id is opencode-go with hyphen), so disabling does not take effect
• Auth: only an opencode-go API key in ~/.local/share/opencode/auth.json (no official opencode account token); official opencode provider errors with Insufficient balance / No payment method

Steps to reproduce:

1. Start OpenCode Desktop 1.18.13 on Windows.
2. Select opencode-go as provider and gpt-5.6-luna as model.
3. Send any message.

What I've tried:

• Routed desktop through an overseas HTTP proxy by setting NODE_USE_ENV_PROXY=1 and HTTPS_PROXY=http://127.0.0.1:7890, then relaunching OpenCode.exe. Still 403. This indicates the geo block is on opencode-go's upstream egress, not the client's egress — so no client-side network change can fix it.
• Confirming the model name resolves by switching the same provider to kimi-k3, which works.

Questions / Request:

1. Is there an officially supported way to point the opencode-go provider at a different upstream endpoint (a relay whose egress is in an allowed region) via provider config in opencode.json? Please share a minimal config example.
2. Alternatively, can I register a custom OpenAI-compatible provider in opencode.json that maps a model id to gpt-5.6-luna and uses my own key/baseURL? What is the minimal provider block schema?
3. Side note: disabled_providers in opencode.jsonc does not match opencode-go when written as opencodego. Worth normalizing the id or documenting it.
4. Could the client surface "region not supported" more clearly (e.g. hide/disable region-locked models in the model picker) instead of failing on every message?

Sample log (server-1/opencode.log):
timestamp=2026-08-05T04:37:32.149Z level=INFO message="llm runtime selected" llm.runtime=ai-sdk llm.provider=opencode-go llm.model=gpt-5.6-luna
timestamp=2026-08-05T04:37:34.144Z level=ERROR message="stream error" providerID=opencode-go modelID=gpt-5.6-luna error.error="AI_APICallError: Upstream request failed: [403] This model is not available in your region."
stack: AI_APICallError -> postToApi19 -> OpenAIResponsesLanguageModel5.doStream -> _retryWithExponentialBackoff -> streamStep

Happy to provide more log excerpts on request. I deliberately did not attach the API key.

Plugins

No response

OpenCode version

1.18.13 (Desktop, packaged, Windows x64)

Steps to reproduce

1. Start OpenCode Desktop 1.18.13 on Windows.
2. In ~/.local/share/opencode/auth.json, have only an opencode-go API key (no official opencode account token).
3. Select opencode-go as provider and gpt-5.6-luna as model.
4. Send any message in a session.

Expected: model responds.
Actual: every request fails within ~1-2s with
"AI_APICallError: Upstream request failed: [403] This model is not available in your region."

Consistent across ~1 hour of retries, on both the title (small) and build (non-small) agents. Switching the same provider to kimi-k3 works fine, so the issue is specific to gpt-5.6-luna.

Screenshot and/or share link

N/A — unable to run /share because every message in the session fails with the 403 error before a share can be generated. Full server log is at ~/.local/share/opencode/log/opencode.log and desktop logs at ~/AppData/Roaming/ai.opencode.desktop/logs/; happy to attach specific excerpts on request.

Operating System

Windows 11 (10.0.26200, x64)

Terminal

PowerShell 5.1 (C:\WINDOWS\System32\WindowsPowerShell\v1.0\powershell.EXE) — but the issue reproduces in the OpenCode Desktop GUI, not the CLI.