Selected custom endpoint can silently auto-route to another paid model
- Copilot Chat Extension Version: 0.66.0 (GitHub Copilot extension: 1.388.0)
- VS Code Version: 1.138.0 (
7debcd0e2acdea1c52de81bf9ee1620444407dda, x64) - OS Version: Linux x64 under WSL2
- Feature (e.g. agent/edit/ask mode): Agent mode with a custom endpoint
- Selected model (e.g. GPT 4.1, Claude 3.7 Sonnet): Custom endpoint model
azure-gpt-56-sol - Logs: Attached. The available logs contain session metadata and the public model catalog, but do not appear to record the selected custom endpoint or automatic-routing decision.
Steps to Reproduce:
- Select the custom endpoint model
azure-gpt-56-soland use it in an existing chat. - Intermittently, a later request reports
Auto routed to ..., although the custom endpoint still appears selected.
I do not currently have reliable reproduction steps or enough evidence to identify the trigger. It may coincide with degraded connectivity or an unavailable custom endpoint, but this is unconfirmed.
Actual behavior
Copilot Chat can automatically route a request to a different GitHub-hosted model while azure-gpt-56-sol remains visibly selected.
Observed routing targets include Claude Opus 5.5 and GPT-6 Luna. Therefore, the problem does not appear specific to one destination model.
The automatic-routing notice is easy to overlook. This behavior can consume premium or paid requests even though the user did not knowingly select the model that processed the request. It also leaves the selected model shown by the UI inconsistent with the model that handled the request.
Expected behavior
When a specific custom endpoint is selected, no request should be sent to another model unless the user has explicitly enabled or approved automatic routing.
If the selected model cannot handle the request, Copilot Chat should stop and clearly report the problem rather than silently falling back. The displayed selection must match the model that processes the request.
Please provide a setting or policy that completely disables automatic model routing and fallback, especially fallback that can consume premium requests.
Frequency
Intermittent, but observed multiple times.
Attachments
Source: microsoft/vscode