GitHub Copilot CLI v1.0.74 Changelog

This update is all about making the command-line AI helper smarter, more flexible, and less frustrating. It matters most to developers and power users who use this tool to write code, manage background AI tasks, or connect external AI plugins to their workflow.

What's new

• New AI model support: You can now use gemini-3.6-flash (a fast, new Google AI model).
• Plan mode model picker: There is a new "plan mode" where the AI thinks before acting. You can now use /model plan to choose a specific AI model just for planning. When you leave plan mode, it goes back to your normal model.
• Sandbox opt-in: The first time you run the tool, it will show a welcome screen asking if you want to use a "sandbox" (a safe, isolated space that stops the AI from accidentally messing up your computer).
• Plugin and MCP support: You can now add external tools using "Open Plugin Spec v1" and mcp.json files (these are standard ways to connect outside AI tools to your helper).
• Interactive shell shortcut: Typing the $ shortcut lets you jump into a normal computer shell right away, even if the AI is still working in the background.

Improvements

• Better IDE connection: If you are using this inside a code editor (IDE), the connection won't drop when the tool reloads external servers or changes folders.
• Smarter task timelines: When you open /tasks to see what background AI agents are doing, it now shows every prompt and response in the exact right order. It also tells you if a prompt came from the main AI or a smaller "subagent" (a helper AI).
• Safer planning: In plan mode, the AI can now create planning files inside your current session folder, but it is still blocked from changing important files outside that folder.
• Bigger uploads: The size limit for sending data to the AI has been increased, and if you send an image that is too big, the tool automatically shrinks it so the process doesn't freeze.
• Better search: When you search for an old session, it will now find it even if the spacing in the title doesn't match exactly.
• Steering updates: "Steering" (a feature that lets you nudge the AI while it runs) can now interrupt what the AI is doing without stopping the actual command running on your computer.
• Settings tabs: The settings menu now shows a "Tab" hint at the bottom to help you switch between different setting scopes.

Fixes

• Question mark bug: Typing ? while the search bar was open used to trigger a help menu. Now, it correctly types a ? into your search text.
• Broken passwords: When adding external tools, the setup wizard used to delete = signs from passwords and tokens (which are often needed for security). It now saves them correctly.
• Wasted uploads: If a remote server permanently rejected an upload, the tool used to keep trying forever. It now stops retrying when it gets a permanent error.
• Pop-up leaking: If you had multiple sessions running, a pop-up dialog from one session used to "leak" into the other one. Now, pop-ups stay in their correct session and reopen when you switch back.
• Skill flags: The tool now fully respects the "disable-model-invocation" flag, meaning if a specific skill is told not to use the AI model, it won't.
• Code symbol warnings: If a language server (a background helper that reads code) reports a different piece of code than you asked for, the tool will now warn you.

Breaking changes

No breaking changes in this release.