GitHub Copilot provider ([ai.github]) fails with `github_models_retirement_brownout` — appears to hit retired GitHub Models endpoint instead of Copilot Chat API

[Giansass]

Describe the bug

Description

The [ai.github] AI provider (documented as "GitHub Copilot" in the LLM Providers guide) fails on every request with an error referencing "GitHub Models" retirement — regardless of which model ID is configured.

GitHub Models (the separate marketplace product at models.github.ai) was fully retired on July 30, 2026 (see GitHub's changelog). GitHub Copilot Chat (api.githubcopilot.com) is a different, still-active product and continues to work fine in other clients (VS Code, OpenCode, Copilot CLI) with the same OAuth token.

This suggests the [ai.github] provider in marimo is calling the retired GitHub Models endpoint rather than the Copilot Chat Completions API (https://api.githubcopilot.com), despite being documented and labeled as "GitHub Copilot."

Steps to reproduce

1. Obtain a GitHub Copilot OAuth token (gho_...), e.g. via gh auth login (web browser flow) + gh auth token. Confirmed working for Copilot Chat via other clients (e.g. OpenCode) with the same token.
2. Configure marimo.toml:

   [ai.models]
   chat_model = "github/claude-sonnet-5"

   [ai.github]
   api_key = "gho_..."

3. Open the AI chat panel in the marimo editor and send any message.

Expected behavior

The request is sent to https://api.githubcopilot.com/chat/completions and returns a completion from Claude Sonnet 5 (assuming it's enabled on the account/plan), matching the behavior of other Copilot Chat clients using the same token.

Actual behavior

Every request fails with:

status_code: 410, model_name: claude-sonnet-5, body: {'code': 'github_models_retirement_brownout', 'message': 'GitHub Models is temporarily unavailable as part of a scheduled retirement brownout.'}

This happens regardless of the model ID used (tested with both a GitHub Models-style nested ID like deepseek/deepseek-r1-0528 and a Copilot-style flat ID like claude-sonnet-5), which suggests the base URL/endpoint is hardcoded rather than derived from the model prefix.

Environment

• marimo version: 0.24.0
• OS: macOS (26.4.1)
• Python version: 3.12.13

Suggested fix

Update the [ai.github] provider implementation to target https://api.githubcopilot.com (the Copilot Chat Completions API) instead of the retired GitHub Models endpoint (models.github.ai / models.inference.ai.azure.com), so the documented "GitHub Copilot" provider actually reaches Copilot rather than the now-defunct GitHub Models service.

Will you submit a PR?

• Yes

Environment

• marimo version: 0.24.0
• OS: macOS (26.4.1)
• Python version: 3.12.13

Code to reproduce

1. Obtain a GitHub Copilot OAuth token (gho_...), e.g. via gh auth login (web browser flow) + gh auth token. Confirmed working for Copilot Chat via other clients (e.g. OpenCode) with the same token.
2. Configure marimo.toml:

   [ai.models]
   chat_model = "github/claude-sonnet-5"

   [ai.github]
   api_key = "gho_..."

3. Open the AI chat panel in the marimo editor and send any message.

[Light2Dark]

Thank you. This API was used before https://api.githubcopilot.com/, but it had issues with auth and enterprise users. We will look into a possible migration or alternatives.