Skip to content

Use the AI Help window

AI Help is a separate window that can read recent context from the active terminal and help you understand or act on it. It is review-first by default.

Open AI Help

Click AI Help in the status bar. The window opens separately from the main terminal, so you can move it to another monitor, snap it beside the workspace, or keep it open while terminals continue running.

If AI Help is locked in your build, the window explains the access state instead of pretending the feature is ready.

Configure a provider

Open Settings and choose AI Help. Add an OpenAI-compatible provider with:

  • provider name
  • endpoint URL
  • API key
  • default model

ButtonsCLI can try the provider’s /v1/models endpoint when an endpoint and key are present. If model discovery works, use the model picker. If it does not, type the model ID manually.

Use Test Connection before depending on AI Help during real work.

Ask about the active terminal

AI Help can include recent output from the active tab when terminal context is enabled.

Use it for:

  • explaining error output
  • suggesting the next command
  • summarizing what a long command did
  • turning confusing logs into a short diagnosis
  • drafting a command without running it automatically

AI Help does not send full terminal history. Context is bounded by the configured line limit. Requests also include basic freshness details, such as when the terminal snapshot was captured and how recently input or output changed, so the model can tell old output from new output.

Review suggested actions

AI Help can return terminal actions with its answer. In reviewed mode, you choose whether to run them.

Actions may include:

  • typing a command
  • typing and pressing Enter
  • sending a control key such as Ctrl+C

Keep reviewed mode on when you are working in a sensitive directory, connected to a production host, or unsure what the suggested command will do.

Use Agent Mode

Agent Mode lets AI Help continue a short task in steps.

The loop is:

  1. AI Help proposes one next action.
  2. You review the action.
  3. ButtonsCLI sends it to the terminal after approval.
  4. AI Help watches for a quiet pause in terminal output.
  5. AI Help asks for the next step with the output produced by that approved action called out separately from older terminal output.

That quiet pause is only an observation, not proof that a command finished. If the terminal keeps running or never produces fresh output, Agent Mode stops auto-continuing and asks you to review the terminal before continuing.

Agent Mode availability depends on preview or entitlement state.

Understand full permissions

YOLO / full permissions skips the review dialog after you accept the warning. Use it only when the terminal context is safe and you are comfortable with AI Help sending actions immediately.

Longer task mode changes planning behavior. It encourages AI Help to carry a task farther before asking clarifying questions, but it does not by itself remove review prompts.

Use Stall Recovery

Stall Recovery watches a specific tab for signs that an agent-driven task is stuck.

Enable it per tab:

  1. Open the tab’s three-dot menu.
  2. Turn on Stall Recovery.
  3. Configure the provider and model in Settings under AI Help.

Stall Recovery can use:

  • recent terminal text
  • two screenshots compared over time, when a vision model is configured

Tabs without Stall Recovery enabled are not evaluated.

Check privacy before sending

When terminal context is enabled, AI Help sends your prompt and recent terminal output to the provider you configured. It does not automatically send clipboard contents, files, or full terminal history.

Use AI Debug to inspect the request details when you are troubleshooting provider setup or prompt behavior.