We are so back: Human in the Loop in Kilo Code

The v7 rewrite of Kilo Code’s VS Code extension was ambitious. We rebuilt everything— new architecture based on OpenCode server, new UI, Agent Manager for parallel sessions, tighter OpenCode integration. It shipped fast, and it shipped rough.

One piece of feedback came through louder than the rest: developers wanted more control back.

Not “turn off the AI” control. More like “show me what you’re about to do before you do it” control. The kind of oversight that lets you trust an agent because you can verify its work, not because you’re hoping it got things right.

We’ve been shipping fixes steadily since the Week Three update. Here’s where things stand.

What’s Already Shipped

See the Agent’s Thinking

The simplest ask was visibility. When an agent proposes changes, you want to understand why — not scroll past walls of collapsed reasoning to find the diff.

Reasoning blocks now stay expanded by default. Terminal command blocks can collapse to just a header bar so they don’t dominate the view. The context usage graph landed in the task header so you can see token consumption turn by turn.

These sound minor but in practice, they’re the difference between scrolling past an agent’s work and actually reading it.

Better Diffs Before You Approve

The v7 extension initially showed diffs after approval. But our community made it clear many desire to review changes before committing to them.

Patch operations now show diffs in the permission prompt before you approve. Edit tool diffs open in side-by-side mode by default. Agent Manager diffs auto-expand reviewable files while keeping generated or huge files collapsed.

The unified Changes panel was a bigger lift. Instead of two separate views for “workspace changes” and “session changes,” there’s now a single panel with a source dropdown. The sidebar badge shows session additions and deletions. If you’ve disabled snapshots for a repo, it warns you instead of silently showing stale data.

Shell command output got syntax highlighting, copy buttons, and an “Open in Editor” action that shows full untruncated output. You can actually read what the agent ran.

Permission Flow Fixes

The auto-approve toggle now persists across VS Code restarts. It shows in the session prompt controls so you know what mode you’re in. The settings UI got readable section titles in light themes.

A more subtle fix: permissions were getting stuck when multiple panels showed the same prompt, or when a subagent’s permission got routed to the wrong worktree. Those now surface clear errors instead of leaving buttons permanently disabled.

Custom agents can have per-agent tool permissions configured in the VS Code agent editor. If you want a read-only reviewer or an agent that can’t touch bash, you can set that up.

Checkpoint Improvements

Session recovery is more reliable. Timeline, token, context, and cost stats stay in sync after restoring a checkpoint. Older revert points stay available after you restore an earlier message. The thinking level preserves through compaction.

Todo items show compact update cards when checked off in long lists, so the review surface doesn’t get cluttered.

What’s Still Coming

The core ask — see exactly what’s changing before approving it with accurate, focused diffs — is tracked in #8415. It’s not fully resolved yet, but we’ve been shipping some of the pieces like patch diffs in permission prompts, and the unified Changes panel, but the full long tail of making this exactly the experience we want is still being worked on.

Step-level checkpoint restore (#8378) would let you see which files changed at each agent step and restore to a specific step within a turn — not just at the user-prompt level. Useful for recovering from rate limits or interrupted runs mid-turn.

The diff code blocks still take too much vertical space. Long edit blocks push reasoning off-screen. The fix (scrollable containers with a higher threshold before expanding) is agreed on but not shipped.

And for new users, there’s an idea to ask how they prefer to work — more hands-on or more agentic — and configure the extension accordingly from the start.

The Philosophy

When I joined Kilo, I heard “agentic engineering” tossed around a lot. It can sound like “let the AI do everything.” That’s not what it means - you can outsource some of software authoring to AI but you cannot outsource the thinking behind software engineering.

The best agents are the ones you can trust because you understand what they’re doing. The tooling makes oversight easy — that’s the point. Review before approve. Diff before commit. Reasoning visible, not hidden.

The v7 rewrite pushed hard on capability — Agent Manager, parallel sessions, tighter model integration. Some of that came at the cost of visibility and control. The work over the past few weeks has been about rebalancing. The tracking issue for human-in-the-loop improvements (#8415) has five active areas, and not all of them are closed. But the direction is clear, and the shipped improvements are real.

If you downgraded to 5.x because v7 felt too hands-off, this is a good time to check back. The experience is meaningfully different now.

---

Following along? The HITL tracking issue has the full list. And if you hit rough edges, file an issue — the heap-snapshot command we added makes reports actionable.