Chat & composer
Messages
Your messages appear on the right; the model's replies stream on the left in a centred reading column. Markdown is rendered — headings, lists, tables, links, and code blocks with a language label and a Copy button. Math (KaTeX) and Mermaid diagrams render inline.
Hover a message for Copy / Edit actions. Editing your message and re‑sending forks the turn from that point.
Under each reply a small stat shows ⏱ time · ~tokens for that turn; hover it for the model name. It's saved with the chat, so it stays after you switch chats or reload.
Tool cards
When the agent runs a tool you see a compact card — the tool name, a duration, and a chevron. Expand it to read the full output (command result, file diff, etc.). A spinner marks a running step; green means success, red means an error.
Composer controls
- ＋ (attach) — add files, a folder, or a connector to the message. Images are OCR'd, PDFs and spreadsheets are extracted to text.
- Approval mode — how tool actions are gated (see below).
- Effort — how hard the model "thinks" before answering.
- Style — the tone/length of replies: Normal or Explanatory.
- Clean — a toggle that hides all agent actions (tool cards, including errors) and shows only the final answer. Good for a tidy read; turn it off to see what the agent actually did.
- Send / Stop — the gradient button sends; while a turn is running it becomes Stop.
Approval modes
- Auto‑run — the agent executes tools immediately. Fastest; best when you trust the task.
- Ask first — before each state‑changing tool (shell, file writes, image generation…) the agent pauses and asks for your approval.
- Plan only — the agent plans and describes what it would do, but never executes state‑changing tools. Great for reviewing an approach first.
Read‑only tools (reading files, searching, web lookups) always run; approval gates the actions that change something.
Effort
Off disables the model's hidden "thinking" — the fastest replies for simple questions. Low / Medium / High / Max turn on reasoning of increasing depth for harder tasks. If your model is a reasoning model, higher effort trades speed for quality.
Response styles
- Normal — direct, gets to the point.
- Explanatory — walks through the reasoning and context.
The files indicator
The top‑bar chip counts the files the agent touched this turn: ＋ created, ✎ edited, 👁 read. Click it to see the full list. It persists per chat so you can review what a turn changed.
Next: Agent tools.
