Using the AI panel¶

The AI panel sits on the right of the workspace and works on the session in the active tab. Ask a question, read what comes back, then approve or dismiss the command it proposes. It assumes you have configured at least one AI Provider and turned Enable AI on for the connection.

Asking a question¶

Type into the composer at the bottom of the panel and send. The question is answered against the recent scrollback of the session you are in, so you rarely need to paste anything — "why did that fail?" after a failed command is usually the whole question.

The four quick actions do the same thing with the question pre-framed:

Quick action	Use it for
Analyze Error	Something just failed and you want the cause
Explain Command	You are about to run something you did not write
Review Config	A configuration file or block on screen
Optimize	Asking whether there is a better way to do this

They appear as cards on an empty panel and in the quick-actions menu beside the composer once a conversation has started. The same menu holds Upload from computer and Browse remote files for attaching a file to the question — attachments go through redaction before they are attached, and OpsPilot tells you how many items it scrubbed.

The reply streams in as it is generated. While a turn is in flight and you have not started typing, the composer's send button becomes a stop button, which cancels the model's response — or, if a command's output is being awaited, stops waiting for that. Typing wins over stopping: as soon as there is something to send, the button goes back to Send.

Choosing the provider for this session¶

The provider dropdown at the top of the panel sets which configured provider answers this session's questions. It defaults to whichever provider is active in Settings → AI Providers.

The setting is per session rather than global: a fast local model on the tab where you are reading logs, a larger cloud model on the tab where you are stuck. Switching does not restart the conversation.

What a reply looks like¶

A reply is an explanation, plus — usually — one proposed next command rather than a multi-step plan. OpsPilot asks the model for a single next action, runs it if you approve, and then asks again with the real result.

Each proposed command arrives as a card with:

the exact command text you are about to run, verbatim
a risk badge — Read-only, Low risk or High risk
the model's stated reason for running it

The tier is not the model's word alone. Your Command Safety profile is applied on top of the model's own self-assessment, and a pattern match can promote a command to High risk. Nothing can demote one that was already flagged. When the gate is triggered by a pattern, the card names the pattern that matched, so you can tell a model's caution apart from your own policy.

The redaction badge¶

If a 🔒 badge with a count appears on a turn, redaction fired before anything was sent. The number is how many items were replaced; hover it to see which categories they came from.

The badge is an indicator, not a control. Turning it off in settings hides the indicator only; redaction still happens on every turn either way.

Approving or dismissing¶

The card has two buttons:

approve & run — the command goes to the real shell in the real session, and you watch it run in the terminal.
dismiss — nothing runs. The card closes and the conversation continues.

A High risk command adds a step: a typed justification before the confirm & run button will do anything. The button stays disabled until you have typed at least ten characters, and the reason stays with that step in the session transcript as the record of why the destructive action was approved. That transcript is in memory only, so capture it yourself if you need the record kept — see Approvals & auto-run.

Once a command runs, its output is redacted and fed back as the next turn, so the model keeps working through the problem without you re-asking after every step. It stops when it has an answer, or when it needs a decision only you can make.

Approval is the only path to execution

There is no setting, provider, prompt or assistant that lets a model run a command without a human approval event. Read-only commands can be pre-authorised as a class with Auto-run safe commands under Settings → Security → Command Safety — off by default — which is still a human authorising that class in advance rather than the model executing on its own.

When a proposal came from elsewhere¶

If a turn originated from a connected assistant rather than the panel, the card carries a badge naming the assistant it came from, and the panel shows which assistant is involved. Those proposals use the same cards, the same tiers and the same buttons — see AI Assistants (MCP).