Kilo Code v7.4.0 Changelog

This update focuses on making the AI assistant smarter about managing your work and easier to interact with. It matters most to developers who use the AI to juggle multiple tasks at once and want better control over their project costs and plans.

What's new

• Automatic branch naming: When the AI sets up a new work session (a "branch"), it will automatically give it a name based on the task you're doing, rather than a random default name.
• Interactive planning: When the AI finishes drafting a plan, you now get clickable buttons to either "Implement" it or "Keep refining" it, instead of having to type out a number to make a choice.
• Session cost alerts: You can now set a spending limit (a whole dollar amount) in the settings. If a chat session hits that limit, the AI will warn you before you continue. This is turned off by default, so you have to opt in.
• Account balance visibility: You can now see your account balance right in the VS Code header (where you switch accounts), and view Kilo Pass subscription details on your profile page.

Improvements

• Remembering your layout: The model picker (the menu where you choose which AI model to use) now remembers if you left it expanded or collapsed, keeping things consistent across the sidebar and the Agent Manager.
• Cleaner provider settings: In the settings menu, "Show more providers" is now easier to find, and providers you've disabled are hidden by default to reduce clutter.
• Sandbox memory: If you turn on "sandbox mode" (a safe, restricted environment for the AI to run commands in) and then fork or move your session, it now remembers to stay in sandbox mode instead of resetting to the default.
• Better plan file links: When you save a plan, the links pointing to that plan now correctly lead to the actual saved file instead of a mistakenly generated file name. If no plan was written, it gives you a clear error instead of failing silently.
• Cleaner interface: Removed the "balance chip" (a small display showing your remaining funds) from the session header to make it less cluttered.

Fixes

• Fixed slow chat loading: Opening or switching chats used to take several seconds because the app was scanning the entire message history to calculate token usage (how many words the AI processed). This scan was delayed so the chat can load instantly.
• Fixed AWS Bedrock connection: If you use AWS Bedrock (an AI service from Amazon) through your work's single sign-on (SSO) or credential system, it was previously failing to connect. This is now fixed.
• Fixed chat sync issues: If the app's background connection briefly dropped, switching between sessions would only update the code changes on the side, leaving the chat stuck on the old conversation. They are now kept in sync.
• Fixed memory leaks from deleted sessions: When you deleted a chat session, the app was accidentally holding onto unsent text, review comments, and pending images in its memory. This is now cleared out. However, if a session gets deleted while you are in the middle of sending a message, it safely restores your unsent text so you don't lose your work.
• Fixed subagent display: Results from "subagents" (smaller AI helpers) were getting cut off at the bottom when they popped up. They are now properly aligned with other tool results.
• Closed security loopholes in bash commands: Certain restricted commands that the AI is allowed to run had hidden flags (like --compress-program or --pager) that could be used to sneakily run unauthorized programs. These loopholes are now blocked.