Chat, Agent, and Plan
The three modes, how much thinking to buy, and how to steer or stop a turn that is already running.
Mode decides how much Kosmo is allowed to do on its own. The control is in the composer, next to the model picker.
The three modes
Chat talks. It reads the project and answers, and it changes nothing.
Agent executes. It reads, writes, edits, and runs commands to finish what you asked, reporting as it goes. This is the default.
Plan designs first. It creates a versioned plan linked to the evidence it used, then waits. Approval applies to that exact revision and hands it to Agent. If consequential evidence is missing, stale, or changed before execution, Kosmo blocks the affected action or asks you to revalidate it.
Reach for Plan when the task is large enough that a wrong turn is expensive, and for Agent when it is not.
Note: Plan and Agent need a model that supports tools. On a chat-only model the mode control says so.
Thinking effort
Separately from mode, you can choose how long a model thinks before it answers, from the chat toolbar or Settings → AI. The setting is per model and applies to chat, the floating assistant, and the dashboard chat widget.
| Level | Use it for |
|---|---|
| Auto | Let the model decide. A good default |
| Off | Answer immediately, no thinking |
| Minimal / Low | Short questions where the answer is lookup, not reasoning |
| Medium | A balanced think for most questions |
| High / Very high | Hard problems worth the wait |
| Maximum | The deepest thinking the model offers |
Deeper thinking is slower and spends more credits. Not every model offers the full range: some cap out partway, some always think and have nothing to adjust, and some do not support it at all. The control tells you which case you are in.
While it is working
You do not have to wait in silence. Keep typing, and what you send waits its turn by default. If it matters now, push it into the work in progress with Send to this turn and the agent reads your correction between tool calls instead of finishing the wrong thing first.
Stop ends the turn. Partial work is kept: the answer so far, the reasoning, and the tool timeline are checkpointed, so a stopped or interrupted run shows up in the conversation as interrupted rather than vanishing.
Permission for commands
In Agent and Plan mode, anything that runs code on your machine asks first. Review the arguments, then choose Allow once, Always allow, Always allow for the whole family of commands, or Deny.
Some commands can never be remembered and ask every single time: destructive filesystem operations such as rm -rf, dd, mkfs, and shred; privilege escalation such as sudo, su, and doas; and persistent system changes such as chmod -R, chown -R, systemctl, and kill.
When the conversation gets long
Long conversations compact automatically. You can also compact on demand from the context control, which summarizes the earlier part of the conversation and frees room for the rest. The context has to be at least 10% full before there is anything worth compacting.