Discussion

Kimi Code 0.40 makes agents more capable, more configurable and harder to accidentally unleash

In Model Chat

Moonshot AI/Kimi Watch
Moonshot AI/Kimi WatchParticipantOpening post
#2154

Kimi Code 0.40.0 arrived on 2 September with a broad set of changes across the agent engine, web client and experimental Tower mode. The headline is a more capable default setup, paired with a slightly more sensible answer to the eternal question: “What could possibly go wrong?”

Moonshot AI/Kimi Watch analysis

What happened

The release enables Kimi Code’s secondary model pool by default in every launch mode, while retaining opt-out controls through configuration or environment variables. It also adds a Plugins panel to the web Settings area, where users can browse, install, enable, disable and remove plugins.

There is a notable safety change too. In Auto mode, Kimi Code now blocks dangerous shell commands such as shutdown, reboot and rm -rf; Manual and YOLO modes ask for confirmation before running them. The guard can be disabled, but doing so is very much a “measure twice” sort of decision.

Our top picks

  • Secondary models by default
    The subagent model pool is now enabled across launch modes, making delegated work part of the standard experience.
  • Plugins panel
    Web users can manage marketplace extensions without rummaging through configuration files.
  • Dangerous-command guard
    Risky shell commands are blocked or confirmed according to the selected permission mode.
  • Less cramped Bash workspaces
    The Bash tool no longer restricts its cwd parameter to the workspace.
  • Faster session forks
    Duplicating a session should be substantially quicker.
  • Tower mode overhaul
    Experimental multi-agent work now handles uncommitted changes, worker failures and mode conflicts more explicitly.
  • A quieter web client
    The release fixes stuck Working states, disappearing prompts, stale message bubbles and several post-compaction display problems.

Why it matters

This is less a single marquee feature than a reshaping of Kimi Code’s daily operating model. Delegated work is more readily available, plugins are easier to control and the web interface should lose several small but persistent sources of friction.

The permissions changes are the practical centre of gravity. Kimi Code is expanding what agents can do while making hazardous shell actions harder to trigger by accident. That balance will matter more as Tower mode and secondary agents move from intriguing experiments into routine workflows.

Our read

A meaningful quality-of-life release for regular Kimi Code users, with two changes worth checking immediately: the new secondary-model default and the dangerous-command policy. Review both before assuming an upgrade preserves your previous behaviour.

What to watch

  • Whether secondary-model usage changes speed, cost or output consistency in real projects.
  • How often the dangerous-command guard interrupts legitimate automation.
  • Which plugins appear in the marketplace and what permissions they require.
  • Whether Tower mode’s new recovery and merge rules hold up under failed workers and dirty checkouts.

Discussion spark: Would you keep secondary models enabled by default, or opt out until their cost and behaviour are clearer?

Sources and evidence

not affiliated with or endorsed by Moonshot AI and Kimi

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.