In a Kilo Blog post published May 15, 2026, Brendan O’Leary describes how Kilo Code has been adding human-in-the-loop controls after the v7 rewrite of its VS Code extension brought a new architecture, the OpenCode server, Agent Manager for parallel sessions and tighter OpenCode integration, but also drew feedback that developers wanted more control back.
The post portrays that feedback as less about disabling AI entirely than about seeing what the agent plans to do before it does it. Since the Week Three update, Kilo Code has been shipping changes that make agent behavior easier to inspect.
Among the updates already shipped, reasoning blocks now stay expanded by default, while terminal command blocks can collapse into a header bar. A context usage graph now appears in the task header, giving a turn-by-turn view of token consumption.
Review flows have also been adjusted. Patch operations now show diffs in the permission prompt before approval, edit tool diffs open in side-by-side mode by default, and Agent Manager diffs auto-expand reviewable files while keeping generated or very large files collapsed. Kilo Code also merged separate workspace and session change views into a single Changes panel with a source dropdown, warning users when snapshots are disabled for a repository instead of quietly surfacing stale data.
Shell command output has gained syntax highlighting, copy buttons and an “Open in Editor” action that reveals the full, untruncated output. The post also notes fixes to permission handling: the auto-approve toggle now persists across VS Code restarts, appears in session prompt controls and is less likely to get stuck when the same prompt shows in multiple panels or when a subagent’s permission is routed to the wrong worktree.
The update also adds per-agent tool permissions for custom agents in the VS Code agent editor, making it possible to set up a read-only reviewer or an agent that cannot use bash. On the checkpoint side, timeline, token, context and cost stats now stay in sync after restoring a checkpoint, older revert points remain available after restoring an earlier message, and the thinking level persists through compaction.
Work still underway includes step-level checkpoint restore under #8378, a cleaner treatment for diff code blocks that currently take up too much vertical space, and onboarding that would ask new users how they prefer to work before configuring the extension. The broader human-in-the-loop effort remains tracked in #8415.
Source: Kilo Blog
