New VS Code Extension - Week Two: 188 PRs, 21 Upstream Merges, and a Lot of Bug Fixes
Here's what landed this week
Last week we shared what we were hearing and what we were shipping after the new VS Code extension went GA. This week, we’re following up with what actually landed.
The short version: v7.2.10 is out now for both the CLI and VS Code extension, and improves on every area of feedback we received last week — plus a lot more.
Progress on Last Week’s Issues
Here’s where we stand on the specific problems we called out:
Rate limiting and quota handling is fixed. Sessions no longer get stuck on rate limit errors. The extension now retries with exponential backoff, shows a countdown, and lets you cancel or switch models while waiting (#8435, #8552). We also fixed a separate issue where signing in or changing your auth state didn’t re-evaluate provider access, leaving users stuck on the free model (#8898).
Plan agent can no longer write files outside the plan directory. We hardened Plan mode permissions and propagated restrictions to sub-agents so Plan stays read-only where it should (#8417). We also strengthened Ask mode with explicit “supersedes” language so it stays read-only even when project instructions try to override it (#8320).
Progress on human-in-the-loop controls. This is a bigger effort, but we’ve shipped the first piece: inline diffs now appear directly in edit/create-file permission prompts, so you can review exactly what an agent wants to change before approving (#8617). The per-file revert buttons in Agent Manager diffs also shipped, letting you selectively discard individual changes (#7455). More work from the tracking issue is ongoing.
Config handling is more resilient. Invalid or incomplete configuration no longer silently breaks the extension. Kilo now skips broken agent/command/skill configs instead of crashing, surfaces schema validation errors with actionable messages, and adds kilo config check so you can diagnose problems (#8473). Config warnings also surface directly in the TUI and VS Code so you know something is off before it becomes a problem.
Prereleases: How We Ship Now
Starting with this release cycle, we’re using a prerelease channel. Versions 7.2.3 through 7.2.7 shipped as prereleases before being promoted to stable as 7.2.10.
If you’re on the prerelease channel from the beta period and want stable releases, switch to the stable/release channel in VS Code’s extension settings. Stable releases go through the prerelease process first.
If you have an experimental spirit, the prerelease channel gets fixes and features earlier. You’ll be running the same builds the team dogfoods internally, and your feedback on prereleases directly shapes what goes stable. We try to ship whenever we think we made a meaningful change which will improve the product, and don’t wait to bundle them. This might mean a lot of versions, especially in the pre-release channel.
21 Upstream OpenCode Releases Merged
Kilo’s CLI engine is built on OpenCode server, and this release includes a major sync — 21 upstream versions merged, from v1.2.25 through v1.4.3. Highlights from upstream include:
AI SDK v6 support with improved provider compatibility
Venice AI and Poe as new built-in providers
Event-sourced session sync for more reliable state
Full HTTP proxy support
OTLP observability export
Vertex Anthropic prompt caching
First-class PowerShell support on Windows
Many Windows stability fixes — terminal keyboard handling, path canonicalization, process management
This is one of the advantages of building on a shared open-source foundation: the broader OpenCode community’s work lands in Kilo automatically.
Codebase Indexing Is Progressing
Community contributor @shssoichiro continues work on codebase indexing (#6966), which will give agents awareness of your full project structure without manual context or large numbers of tool calls. The PR is active with an ongoing review. We really appreciate the sustained effort here — this is a complex feature and the kind of contribution that makes open source work.
What Else Shipped
Across 7.2.1 through 7.2.10, we merged 188 Kilo PRs on top of the upstream work. Here are some highlights beyond the items above:
New Capabilities
Allow Everything mode — a session-scoped or global permission toggle for users who trust their agents fully, with a command-palette off switch (#7764)
Local Recall — agents can search and read past sessions from sibling worktrees and the current repo (#8191)
@terminal mention support — reference terminal output directly in chat and Agent Manager (#8894)
Run Scripts — execute project-specific .kilo/run-script from Agent Manager worktrees with Run/Stop controls and Cmd/Ctrl+E (#8730)
Agent Manager sections — organize worktrees into custom groups with drag-and-drop, color, and collapse (#8225)
Agent Manager PR sync — PR status is shown with approved badge checkmarks (#8655, #8657)
Custom commit message prompts — override the system prompt used for AI commit messages (#8974)
kilo help --all — full CLI reference in markdown or text (#8504)
Network resilience — sessions pause on network failures and auto-resume; kilo run reconnects automatically (#7578)
Fixes
Fuzzy model search with provider and acronym matching (#8567)
@ file mentions now prioritize open tabs and the active editor (#8727)
Token usage breakdown visible in the TUI sidebar (#8357)
Azure provider config improvements with xhigh reasoning for GPT-5.4+ (#8566)
OpenAI-compatible providers now default image input to enabled (#8729)
Fixed Bedrock Claude failures on empty text blocks (#6757)
Fixed SSE memory leaks on Windows (#8952)
Fixed git process orphaning on Windows (#8572)
Fixed prompt/question docks stealing editor focus (#8872)
Subagents can no longer spawn additional subagents (#8644)
Thank You
To everyone who filed issues, opened PRs, jumped into Discord, and tested prereleases this week — thank you. The speed we’re moving at is only possible because you’re showing up, reporting what’s broken, and in many cases fixing it yourselves.
Special thanks to the contributors whose PRs landed in this release cycle, including @shssoichiro, @maphew, @jacksonkasi1, @justincqz, @thomasboom, @ScuttleBot, @Githubguy132010, @aravhawk, @Ashwinhegde19, @IamCoder18, @gmjosund, @ricardofiorani, @0xCybin, @grandmaster451, @shantoislamdev, and @Varuu-0.
Keep the feedback coming — on GitHub, in Discord, wherever works. We’re reading all of it.



bring back the architect mode 😔
First, I'm sorry for the long post, but I'm taking this opportunity to state my issue plainly in the hopes it might make a difference.
Full file rollback is a great step, as is diff approval. Telling the agent to re-plan the updates to a single file that got rolled back due to an implementation error takes longer to get right than manually rewriting the file. And, the approval system requires me to sit in front of it and monitor everything it does, every step of the way, in real time. Worse, I have yet to successfully convert on any task using this workflow unless I jump in and rewrite the file that was rolled back manually, or provide corrections. This goes for my personal account as well as the shop account.
And I'm not the only one. There are 12 other developers in this shop facing the exact same issues and experiencing the exact same problems.
If I have to babysit my AI and give up my ability to work in parallel, the value of an agentic programming assistant (provided by 5.x) is completely removed.
We all work on one project, and the project is built incrementally, and intentionally. The new sub-agent system seems to be optimizing for writing an entire codebase instead of implementing a single feature at a time. I appreciate wanting to support vibe coders in this way, but that's not why my shop was using kilo. It feels like 7.x dropped support for one group of developers in order to add support for the vibe code crowd. (I'm not throwing shade at the vibe coding crowd btw! I'm just saying there are other developers out here doing work using kilo)
Nothing has been more disruptive at my company than kilo's move from 5.x to 7.x, and since management believes legacy will be sunset soon, We don't have any other options, we're about to have a 3-day meeting to redevelop every single one of our workflows to adopt Claude Code.
Kilo has been the absolute best VS code plugin for me, because it fits naturally within the existing workflows of my team, and me personally. The CLI tool was excellent, but it never met those needs. The reliance on GitHub in order to enable cloud dev tools is not anything my local-dev shop will ever use (We're an on-premise outfit for various, unchangable reasons)
I know that adding new features is important to maintain market relevance, and I know that supporting the fastest growing segment of potential users is important for financial security. But, I don't think the full impact of certain choices was necessarily understood.