Kilo Code v7.4.21 landed on August 11, 2026, and it is the kind of release that changes how you use the tool every day — not just what it can do in theory. The all-in-one agentic engineering platform (26,000+ stars, built on OpenCode’s core) rewrote the Agent Manager’s keyboard navigation, made model search actually understand which provider you use, and gave you real control over worktrees in multi-project setups.
If you have never heard of Kilo Code: it is a free, open-source VS Code extension that turns your editor into an agentic engineering platform. You get an Agent Manager sidebar, multi-model support (Claude, GPT, Gemini, local models via Ollama), worktree-based parallel sessions, and a skill system — all without leaving VS Code. Think Cursor’s agent features, but open, provider-agnostic, and you own your data.
What’s New in v7.4.22 (Released 3 Hours Ago)
v7.4.22 landed August 13, 2026 at 22:26 UTC — a dense patch that adds daily-driver polish and a massive upstream sync.
Clickable file references in agent responses
The agent’s inline code spans (like src/utils/helpers.ts:42) are now validated against your workspace filesystem. Real files become clickable links that open at the referenced line. Dead paths stay as plain code with a “File not found” warning on click. This turns agent responses into navigable maps of your codebase.
Agent Manager PR comment actions
You can now resolve/unresolve review threads, jump to the comments section, and scroll-to-top in PR diff view — all from the Agent Manager sidebar. No more context-switching to the browser for code review hygiene.
Model reasoning variants surfaced in chat
The model selector now shows and lets you pick each model’s default reasoning variant (low/medium/high) directly in chat and Agent Manager. No more digging through provider configs to toggle thinking effort.
AWS & GCP credentials for Bedrock / Vertex AI
Structured AWS access keys and Google Cloud service-account JSON are now supported when connecting Bedrock and Vertex AI providers in VS Code. Enterprise auth just got simpler.
Subagent permission fixes
A read-only or delegating agent’s readOnlyBash allowlist no longer projects onto a writable subagent as a bash ceiling. A delegated subagent can now run its own allowed commands (e.g., git status). Edit, notebook, and MCP denials still inherit as hard ceilings — security preserved, flexibility restored.
Git changes chip persists across tab switches
The Changes chip and Git changes stay visible when switching tabs in multi-repository workspaces. Your diff context doesn’t vanish when you jump between projects.
Massive upstream sync: OpenCode v1.17.13 → v1.18.13
13 upstream releases adopted in one shot, including:
- Model compatibility: GPT-5.6 via Azure AI, OpenAI pro reasoning mode, Cerebras reasoning replay, xAI prompt cache routing, Meta Muse Spark 1.1 system prompts
- MCP reliability: paginated tool catalogs fixed, output schema validation fixed, code mode MCP adapter added
- TUI enhancements: spinner registration, CLI env var forwarding, spinner rendering across surfaces
- Provider fixes: GitHub Copilot model routing honors chat/responses endpoints, Cerebras low reasoning effort preserved, Z.ai context-window overflow classification improved
- Config resilience: unavailable config directories handled gracefully, response storage disabled by default for xAI Responses
The Agent Manager finally respects your fingers (v7.4.21)
The biggest daily win is the keyboard shortcut overhaul. v7.4.21 fixes the Agent Manager sidebar so shortcuts behave the way muscle memory expects:
Cmd+/(orCtrl+/) now focuses a visible embedded terminal before hiding it — no more guessing which terminal you’re toggling.Cmd+Shift+Tcreates a side terminal only while that terminal area has focus — it stops stealing focus from your editor.Cmd+Shift+[/]switch terminal tabs cleanly.Cmd+Shift+Mfocuses the Agent Manager prompt instead of opening VS Code Problems (which was a constant annoyance).Cmd+Whides the last side terminal instead of killing its shell — your long-running process survives a quick pane close.
These sound small. They are not. If you live in the Agent Manager sidebar — spawning agents, watching terminals, switching contexts — the old shortcuts added friction to every session. The new ones get out of your way.
The sidebar also got visual polish: keyboard shortcut badges now appear on the right edge of rows only while hovered or holding the jump modifier, worktree titles get more room (no longer reserving space for hidden actions), and project names align with row icons. Search results now show which worktree a session belongs to.
Model search that knows your providers
Kilo Code supports dozens of model providers — Anthropic, OpenAI, Google, OpenRouter, Ollama, LM Studio, custom endpoints. The old model search treated them all as a flat list. v7.4.21 adds provider-aware results and personalized usage suggestions.
What this means in practice: when you open the model selector, models from your most-used providers surface first. If you primarily use Anthropic and Ollama, those appear at the top with relevance scoring that accounts for your history. Custom providers now show their underlying models in the selector when the catalog includes them, and broad reasoning-effort options are automatically exposed for custom models.
For beginners, this removes the “which model do I pick?” paralysis. The selector guides you toward what works for your setup, not a generic list.
Multi-project worktrees: you choose the repo
Worktrees let you run parallel agent sessions on different branches of the same repository — or across different repositories entirely. v7.4.21 adds a critical missing piece: Agent Manager users can now choose the repository when creating or importing a worktree in multi-project mode.
Previously, worktree creation assumed a single project context. Now, if you have multiple projects open in VS Code (a monorepo setup, or separate frontend/backend repos), the worktree dialog lets you pick which repository the new worktree belongs to. This is the difference between “it works for my simple repo” and “it works for my actual job.”
The release also remembers your worktree dialog selections — model, variant, mode, and sandbox — so reopening the dialog doesn’t reset everything. Background Git polling load is reduced across large worktree sets, and VS Code sessions no longer start unused file watchers (indexing defers until search is used).
Nested slash commands, review modes, and pre-session model browsing
Three more features beginners will hit quickly:
-
Nested slash command suggestions for
/review— type/reviewand you now getstaged,unpushed, andquickmodes as sub-suggestions. No more memorizing flags. -
Model catalog browsing before a session starts — you can now explore an instance’s full model catalog before launching a session. This lets you verify the model you want is available and configured correctly, instead of discovering a missing API key three turns in.
-
Slash commands in worktree prompts — model, agent, variant, and sandbox slash commands now work inside Agent Manager worktree prompts. You can
/modelor/agentswitch mid-setup without leaving the flow.
The quiet reliability fixes that matter
Under the headline features, v7.4.21 ships a wave of fixes that prevent the papercuts that make you switch tools:
- Sidebar navigation bar no longer disappears in Cursor when the Kilo Code view is docked in the Secondary Side Bar. Cursor now renders the navigation inside the webview itself.
- Flickering and sticky scrolling fixed when scrolling up in Agent Manager and chat sessions.
- Negative pricing entries ignored from model catalogs; unpriced models handled gracefully in UI price formatting.
- Terminal input accepted immediately while the Agent Manager shell starts — no more typing into a void.
- Custom provider edit dialog layout improved with bulk toggle buttons for reasoning and image modalities across all models.
- Worktree sessions start faster by prefetching base branches, reducing file-watcher load, and overlapping independent multi-session setup.
- Speech-to-text audio compressed to AAC across macOS, Linux, and Windows — prevents payload size errors on long recordings.
- Agent Manager inspector width persists and shares between terminal and diff viewer.
- Extension-managed GitHub CLI commands no longer open transient Windows Terminal windows on Windows.
What breaks (almost nothing)
This is a patch/minor release (7.4.21 → 7.4.22), so breaking changes are minimal. The only config cleanup needed: if you had custom keyboard shortcuts that conflicted with the Cmd+Shift+M / Cmd+W bindings from v7.4.21, you may want to rebind them. The old resolve tool removal and hub tool merge from earlier v7.x versions are already settled.
How to update
Kilo Code updates through the VS Code Extensions marketplace. Open the Extensions view (Cmd+Shift+X), find “Kilo Code,” and click Update — or enable auto-update. The extension will reload and you’re on v7.4.22.
To verify: open the Agent Manager sidebar, press Cmd+Shift+M — it should focus the prompt, not open Problems. Press Cmd+/ on a visible terminal — it should focus, not hide. Hover a file reference in an agent response — it should become a clickable link.
The bottom line
v7.4.21 made Kilo Code feel finished as a daily driver. v7.4.22 makes it feel sharp. Clickable file refs turn agent output into navigation. PR actions keep review in the sidebar. Reasoning variants put model control where you need it. The 13-release upstream sync brings GPT-5.6, xAI caching, Meta Muse Spark, and a pile of MCP/TUI fixes.
If you have been waiting for an open-source Cursor alternative that doesn’t lock you into one model provider, this is the update to try. Install Kilo Code, open a repo, spawn an agent from the sidebar, and watch it work — with shortcuts that finally make sense, and file references you can actually click.
Want to compare AI models without switching tabs? aiFiesta gives you GPT, Claude, Gemini, Grok, DeepSeek, and Perplexity in one place for $12/mo.