New VS Code Extension - Week One: What We’re Hearing and What We’re Shipping

Last week, we GA’d the completely rebuilt Kilo Code extension for VS Code — the culmination of weeks of work bringing the same portable OpenCode server-based engine from the CLI into your editor.

Before making this version the default, we wanted to get it right. We ran a pre-release which roughly 20% of our user base adopted, sending out invites for people to test, and incorporated much of the feedback. Thanks to the pre-release users, we were able to catch a number of bugs and improve a number of features.

With GA and the VSCode extension update process, for many of you the GA landed as an auto-update — you opened VS Code and the new extension was already there. An auto-upgrade rolling out to the full user base always surfaces things a smaller pre-release group doesn’t, and the past week has been no exception. Some of you hit real bugs. Some of you found that workflows you relied on aren’t yet where they need to be.

Thank you. To everyone who filed issues, opened PRs, left comments, and hopped into Discord — you’re making this extension better, and your feedback and participation is invaluable. This is an open-source project, and the community showing up to not just report problems but also fix them is the best kind of collaboration. (And remember, you get $100 in credits for finding and fixing a bug!)

The team is working hard to address this feedback, and we wanted to share an update on what the team is focused on this week.

Fixing Rate Limiting and Quota Handling

Sessions are getting stuck on rate limit errors instead of retrying intelligently and allowing for cancellation. We’re implementing improved backoff and the ability to cancel future attempts — so Kilo recovers on its own gracefully or allows you to switch models and keep working. (#8333)

Tightening Plan and Ask Agents

Plan and Ask agents should never modify your files without asking. We’ve identified cases where they weren’t strict enough about preventing write operations, and we’re tightening that down. We’re also improving the handoff between Plan/Ask and other agents, so that it is smoother like ensuring the agent dropdown is updated in the UI. (PR’s 1 and 2)

Better Human-in-the-Loop Controls

This came through loud and clear. Many of you treat Kilo as a close collaborator, not an autonomous agent — you want to review changes before they land and stay firmly in control of the agents activities. We’ve opened a tracking issue covering diff review before approval, potential checkpoint improvements, permission flow challenges, per-agent tool controls, and an onboarding flow that configures the extension to match how you like to work. (#8415)

Hardening Config Handling

Invalid or incomplete configuration can leave the extension in a broken state with no indication of what’s wrong — agent selector missing, model selector unresponsive, zero error messages. We’re adding startup validation and actionable error surfacing so bad configs get caught and explained, not silently swallowed. (#8388)

Performance on Windows and large repositories

Some of you working on large monorepos or on Windows have noticed performance regressions compared to the previous extension, in particular on memory utilization. We have shipped one initial improvement on Monday and will continue to focus on this for the remainder of the week until memory usage and performance is where they should be. We know this one is frustrating. (#8030)

Switching back is an option

If you’d prefer to use the prior version of the extension while these issues are resolved, you are able to downgrade to version 5.x through most marketplaces in your editor. Click on the Kilo extension in the extension view, then click on the gear, select “Install Specific Version”, and select the latest 5.x version.

Follow Along

You can track all of this work on our project board, and the issues and PR’s within the Kilo Code repository.

Moving at Kilo Speed

One of the core reasons we rebuilt the extension on a portable, open-source foundation was to move faster — not just this week, but for every week that follows. The old architecture had us maintaining two different code bases, and after comparing them against each other, we were convinced that the CLI agentic flow is much better. Now that we’re on a single engine shared with the CLI, fixes ship everywhere at once and the feedback-to-fix loop gets shorter with every release.

Keep the feedback coming — on GitHub, in Discord in the #vscode channel, wherever works for you. We’re reading all of it, and we’re building as fast as we can.

Move at Kilo Speed.