• anyOS @Tyriar
• anyOS @kycutler
• anyOS @deepak1556

Complexity: 4

Create Issue

Thinking tokens are now very widely available in GPT-5, GPT-5-Codex, GPT-5-mini, and Gemini 2.5 Pro.

Settings

chat.agent.thinkingStyle: collapsed, collapsedPreview, fixedScrolling (default)
chat.agent.thinking.collapsedTools: none, all, readOnly (default)

These should be default on already:
github.copilot.chat.responsesApiReasoningEffort: default or high
github.copilot.chat.responsesApiReasoningSummary: detailed
github.copilot.chat.useResponsesApi: true

Test Steps

1. Set thinkingStyle to fixedScrolling and the collapsedTools to readOnly. These are the current defaults in Insiders.
2. Ask chat a few questions. You can prompt it additionally with "Think really hard" or "Ultra think" or "make sure to think before replying".
3. Try it our with tool calls, code base search questions, from a new workspace, and in general in your day to day.
   ❗ Make sure to try the built in GPT-5, GPT-5-Codex, GPT-5-mini, and Gemini 2.5 Pro ❗
4. Verify reasoning parts by setting your Log Level to Trace, and go to Output -> Github Copilot Chat. If you're not seeing reasoning, check to see if the model actually gave us reasoning. Keywords to look for are "reasoning" or "reasoning_summary_text`
5. Repeat steps with none and all for the collapsedTool styles.

Things to Note

• Codex is the model that benefits the most from all and readOnly modes since there is interweaving thinking and tool calls.
• Let me know if you notice any unexpected UI bugs or issues!