Docs › Get started
The chat panel
Agent, Plan, Ask and Debug modes, AstraOne, and the context meter.
The panel is where everything happens. The composer sits at the bottom; everything below the text box is a control you can click.
Four modes
| Mode | What it does | Can it change files? |
|---|---|---|
| Ask | Answers questions about your code. Can run a short list of read-only shell commands. | No |
| Agent | Does the work: reads, edits, runs commands, runs your tests. | Yes, with approval |
| Plan | Writes a plan document first, so you can agree the approach before any code changes. | No |
| Debug | Reproduces and measures the bug before fixing it, and removes any logging it added. | Yes, with approval |
Switch with the mode control under the composer. The conversation is shared across all four: changing mode does not start a new chat.
AstraOne
You do not pick a model. AstraOne does: it runs the model measured to finish real repository work, and when the run's checks fail it brings in a stronger model and keeps going. The composer shows AstraOne where other editors show a model list.
If you bring your own provider key, a model picker appears and you can choose one yourself, along with its effort: how much it thinks before answering. AstraOne sets effort itself. See Models.
What an answer looks like
An agent run ends with the same three parts every time, so you can see at a glance what was done and what is left for you:
- Changed: each file or area that changed, and why.
- Verified: every check it ran after its last edit and the result, and anything it could not verify.
- You need to do: the steps that are yours, such as an apply against production, a secret to create or a DNS change, each with the exact command. Every command has Copy and Run it yourself beside it. When nothing is left, it says so.
Under the answer, the run card lists the files changed with Keep, Undo and Review, and a Needs you box for what the editor itself knows is outstanding: an infrastructure plan made and not applied, a command still running in the background, a changed file the answer never mentioned, or a file the answer claims to have changed that it did not.
Running a command yourself
Start a message with ! to run it as a shell command instead of sending it to the model: !npm test. It runs in your workspace, its output lands in the chat, and the agent sees it on your next message, so you can show it something without asking it to run anything.
The context meter
The ring beside the composer shows how full the model's context window is. You do not usually need to act on it: AstraCode clears finished tool results at half the window and compacts the conversation at 70%, both automatically. See Context.
All documentation
Get started
Agent
- Agent overview
- Approvals and permissions
- Plan mode
- What the agent can do
- Running commands
- Browser tools
- Checkpoints and undo
- Infrastructure as code
- Running several agents at once