New VS Code Extension - Week Three: Memory, Stability, and Moving at Kilo Speed Into the Future
Where we are three weeks in, what landed this cycle, and how we’re thinking about what comes next
Three weeks ago we GA’d the completely rebuilt Kilo Code extension for VS Code. Week one was about what we were hearing and what we were shipping. Week two was about addressing the most urgent feedback and bumps.
This week is about the two other areas of frequent feedback and challenges: memory usage on Windows and session stability under sustained use. Both are materially better now than they were a week ago. Neither is 100% fixed and “done”, we can see from open GitHub issues that some of you still hit rough edges, but the experience is significantly improved especially on Windows when using Agent Manager.
Across the week we shipped 80+ Kilo PRs and merged three more upstream OpenCode releases.
Windows Memory: A Big Step Forward
This is the one we know has caused the most pain. Users on Windows reported the Kilo core process climbing into multiple GB of RAM within minutes of opening Agent Manager and staying there. A handful of you sent us heap snapshots — thank you — which helped track down root cause on some harder to reproduce issues.
The high-level story: Agent Manager was polling git status and diffs through the Kilo core subprocess, and on Windows the combination of IPC round-trips, diff payload sizes, and allocator behavior meant freed memory wasn’t being returned to the OS cleanly. In v7.2.20 we’ve restructured that path (#9046) and made the extension much more careful about what it holds in memory:
Agent Manager’s git work now runs directly in the extension host, not through the core process.
We cap how much of any single diff we’ll read into memory, so opening a very large file no longer causes a spike the allocator can’t recover from.
We also tuned the allocator on the core process itself to release memory back to the OS more promptly on Windows.
If you were running on a downgraded 5.x build because of memory issues, this is the release to come back on. If you’re still seeing unbounded growth, please keep the issues coming — the heap-snapshot command we added this cycle (#9034) makes those reports much easier to act on.
Session Stability: Fewer Interruptions
The second theme was sessions getting interrupted mid-flow — usually recoverable by sending another message or re-opening the session/extension. Most of the reports we got traced back to a handful of specific state-machine edges, and those are now meaningfully better.
The one we heard about most often was sessions ending up stuck — most visibly when VS Code was closed while a suggestion prompt was still showing, which left the session permanently marked busy and any follow-up message queued forever. Sessions now go idle correctly while waiting on a suggestion response (#9199). A related set of stuck states around the end-of-plan flow — where “Start new session” and “Continue here” didn’t reliably transition you into the handover session — also got fixed, so those buttons now move you into a new session that stays visibly busy until the handover summary lands (#9245, #9300).
Everyday chat behavior got a lot smoother too. The most common irritation was the chat view snapping back to the bottom while you were trying to read earlier context during a streaming response; that no longer happens, and scrolling back through long sessions now correctly reloads earlier history from the virtualized list (#9236, #9194). Switching between long sessions in Agent Manager — which used to briefly freeze the UI — is now near-instant, with the chat view self-healing if messages arrived while it was in the background (#8911). Smaller queue and layout fixes also landed around follow-up prompts and tool output interleaving.
Finally, a nice performance-and-stability win from the community: @IamCoder18 landed visibility-aware git polling plus resolution caching in Agent Manager’s git stats poller (#8703), meaningfully reducing the number of git subprocesses the extension spawns on repos with many worktrees.
New Capabilities This Cycle
Stability was the priority, but we still shipped meaningful new capability:
Fork sessions from any user message — both in Agent Manager (#9207) and in the sidebar (#9244). Branch at any point without losing the original.
KiloClaw chat panel in VS Code — the KiloClaw group chat experience now lives directly inside the editor (#7960).
Folder @-mentions — reference a folder with @ and include its top-level file contents as context (#9023).
Autocomplete backend prewarm — inline completions are ready on the first keystroke without having to open the Kilo sidebar first, and autocomplete state refreshes when workspace folders change (#9305).
Heap snapshots from the Command Palette — capture a snapshot of the bundled Kilo core directly from VS Code (#9034).
“Contribute on GitHub” CTA in Marketplace — a subtle footer link inviting contributions of new skills, modes, and MCP servers (#9099).
Upstream OpenCode
Three more OpenCode upstream releases merged this cycle — v1.4.4, v1.4.5, and v1.4.6 — bringing continued improvements to session sync, provider compatibility, Windows terminal handling, and the underlying AI SDK layer. Building on a shared open-source foundation continues to pay off: work from the broader OpenCode community lands in Kilo automatically.
Codebase Indexing Progress
Community contributor @shssoichiro‘s codebase indexing work (#6966) remains active. The branch is being kept current against main, review iterations are ongoing, and we’re closing in on a form we can land. This is a substantial feature and we want to get it right — thank you for the sustained effort here.
Community Update
Some numbers and names from this cycle:
80+ PRs merged on top of the upstream OpenCode work.
3 upstream OpenCode releases merged — v1.4.4, v1.4.5, and v1.4.6.
Multiple stable releases promoted to the marketplace through the period, with v7.2.20 as the current stable.
Thank you to community contributors whose work landed or continued this cycle:
@shssoichiro — continued work on codebase indexing (#6966).
@IamCoder18 — visibility-aware git polling in GitStatsPoller (#8703).
And broad thanks to every community member who filed heap snapshots, reproduction steps, Discord reports, and sustained the long-running Windows performance thread (#8030). That conversation is the reason we had the signal we needed to tackle the memory work head-on this week.
Moving at Kilo Speed Into the Future
This is the last of the regular weekly updates in this series. The core issues that we highlighted in Week 1, rate limiting, Plan/Ask strictness, human-in-the-loop controls, config resilience, and Windows memory, are either resolved or meaningfully better. We will continue to focus on smoothing out the rough edges, in the near future.
We will also be driving Kilo further towards the vision of where agentic coding is going, enabling engineering teams to ship at Kilo Speed safely and confidently, faster than ever before. We are excited about this future and believe that the new V7 is on a strong foundation to build on. Agent Manager continues to improve for those who like to run multiple agent sessions in parallel, and will only become more useful as models continue to improve and become more capable and needing less oversight. And when a particular change or workstyle requires closer agent supervision and pair programming, you can do that too. The AI landscape is evolving quickly and models keep advancing, and the tools we use need to keep pace.
To everyone who showed up over these three weeks — the issue filers, the PR authors, the Discord commenters, the prerelease testers, the heap-snapshot senders, and the folks who point to the future with feature requests — thank you. Your feedback, issues, and pull requests are genuinely what makes this community great. We value every piece of it, and we’ll keep making the extension better because of it.
See you in the release notes.
— Josh and Mark
Move at Kilo Speed.