Agent mode
Two keys, two modes:
- ⌘L is ask. A one-shot AI bar over the focused pane. The answer comes back as a result block with Run / Edit / Copy actions; nothing executes until you say so.
- ⌘⇧L is agent. A full conversation with tools: read files, edit, grep, glob, run commands, call MCP servers, drive the screen.
The approval gate
Section titled “The approval gate”Every tool call passes an ApprovalPolicy before it touches your system:
| Policy | Behaviour |
|---|---|
auto |
Runs without asking (reads, greps) |
prompt |
Asks you, with the exact command shown |
deny |
Never runs |
Policies are per-tool and per-profile, so “read anything, ask before writing, never touch the network” is one profile away. Approved commands execute in the pane — visible, scrollable, and part of your block history like anything you typed yourself.
Rollback
Section titled “Rollback”Agent edits are checkpointed; ⌘Z on the agent timeline rolls the working tree back step by step. An agent can be wrong; it shouldn’t be irreversibly wrong.
Driving TUIs
Section titled “Driving TUIs”Agents can attach to interactive panes; psql, python, gdb; read the live
screen, and type into it under the same approval gate. The terminal detects
attachment prompts and routes them to the conversation.
History
Section titled “History”Every agent conversation is persisted: replay it, export it, or resume it. The history pane shows each tool call with its approval verdict; an audit trail, not a chat log.
Providers
Section titled “Providers”Anthropic, OpenAI, OpenRouter, and Ollama (local). Keys live in the OS keychain. Pick per-conversation models, or set a default in Settings → Providers.