Kilo Code adds more human oversight to its v7 extension

Kilo Code says its v7 rewrite of the company’s VS Code extension shipped with a new architecture, a new UI, an Agent Manager for parallel sessions, and tighter OpenCode integration - but also with too little oversight for many developers.

In a post titled “We are so back: Human in the Loop in Kilo Code,” the company said the loudest early feedback was not that users wanted to disable AI agents, but that they wanted more control over what the agents were about to do. The request was for visibility and review, so developers could verify an agent’s work before it was applied.

Kilo Code said it has been shipping fixes since its Week Three update. One of the first changes was to keep reasoning blocks expanded by default, so users can see why an agent proposed a change instead of skipping past hidden rationale. Terminal command blocks can now collapse to a header bar, and the task header includes a context usage graph that shows token consumption turn by turn.

The company said these changes are meant to make agent work easier to inspect. Instead of treating an agent’s output as something to accept on trust, the interface is being tuned so developers can read what happened and decide whether to approve it.

Review workflows have also been tightened. Patch operations now show diffs in the permission prompt before approval, and edit tool diffs open in side-by-side mode by default. In the Agent Manager, diffs auto-expand reviewable files while generated or very large files stay collapsed.

Kilo Code also combined two separate views into one unified Changes panel. The panel now uses a source dropdown instead of splitting workspace changes and session changes into different places. The sidebar badge shows session additions and deletions, and if snapshots are disabled for a repository, the UI warns the user rather than quietly showing stale data.

Shell command output got a set of readability fixes too. According to Kilo Code, output now includes syntax highlighting, copy buttons, and an “Open in Editor” action that shows the full, untruncated output.

Permission handling was another focus. The auto-approve toggle now survives VS Code restarts and appears in the session prompt controls, so users can see which mode they are in. Kilo Code also said it fixed cases where permissions could get stuck when multiple panels displayed the same prompt, or when a subagent’s permission request was routed to the wrong worktree. Those cases now produce clear errors instead of leaving buttons disabled.

Custom agents can now be assigned per-agent tool permissions in the VS Code agent editor. Kilo Code said that lets users configure, for example, a read-only reviewer or an agent that cannot use bash.

Checkpoint recovery got attention as well. Kilo Code said timeline, token, context, and cost statistics now stay in sync after restoring a checkpoint, and older revert points remain available after restoring an earlier message. The thinking level also persists through compaction, and completed todo items now show compact update cards so long task lists do not overwhelm the review surface.

Not everything is done yet. Kilo Code said the core request - accurate, focused diffs before approval - is still tracked in issue #8264. Step-level checkpoint restore, which would let users recover to a specific agent step within a turn, is also still in development under issue #8378.

The company said the current direction is about balancing capability with control. It described “agentic engineering” as using AI for parts of software authoring without outsourcing the thinking behind software engineering itself. The goal, according to Kilo Code, is to make oversight easy: review before approve, diff before commit, and visible reasoning instead of hidden reasoning.