Warmup request downgraded to model: gpt-5-mini - Claude Code
Quick fix
Turn the Copilot warmupDowngrade option off in CC Switch so requests are not silently replaced with gpt-5-mini.
Symptom
Section titled “Symptom”[Copilot] Warmup request downgraded to model: gpt-5-miniThe CC Switch Copilot optimiser enables warmup downgrading by default (`warmup_downgrade: true`) with `gpt-5-mini` as the default replacement model. When the system classifies a sub-agent request as a warmup probe, it ignores the model you explicitly selected (`claude-opus-4.8-1m`, for example) and silently substitutes before the request reaches GitHub Copilot.
The cluster also holds other Copilot proxy anomalies: deeply indented code in a streaming response being misread as the "infinite whitespace bug" and the tool call discarded (#2646), and model mapping config not applying in proxy mode (#3231). Both stem from the proxy layer intervening too much or from logic defects, but this page addresses the model being downgraded unexpectedly.
Edit the Copilot optimizer config by hand and set `warmupDowngrade` to `false` to disable the downgrade: `{"warmupDowngrade": false}`.
Affected Versions
Section titled “Affected Versions”Source Issues
Section titled “Source Issues”This page is distilled from 5 real issues
- Why are main window requests fine while sub-agent requests get downgraded?
- CC Switch classifies certain sub-agent requests as warmup probes, which triggers the default downgrade substitution. Main window requests are not classified that way.
- Is a lost tool call from writing indented code through the Copilot proxy the same cause?
- No. That comes from the infinite-whitespace detection threshold being too low (20) and catching legitimate output — a separate bug (#2646) that needs a cc-switch upgrade.
Related problems
Section titled “Related problems”This is an unofficial community wiki with no affiliation to the cc-switch authors or the project itself. Its content is compiled from the project's public GitHub issues. This site distributes no software.