Skip to content

AI Assistant

Se.Cherkasov edited this page Aug 23, 2026 · 1 revision

AI assistant

English · Русский

The assistant turns a description into a shell command. It is treated as an untrusted source: nothing it produces runs without your confirmation.

Where it appears

  • Desktop — a panel beside the session, Ctrl+Shift+/ (⌘/).
  • Android — an input form opened from the session; there is no docked assistant panel on a phone by design.
  • Settings → AI → Quick chat — a scratch conversation for testing the connection.

Providers

Provider What it means
On this device a GGUF model (Qwen, Phi) run locally through llama.cpp. Nothing leaves the machine. Downloaded from Settings → AI, resumable, integrity-checked.
My API key (BYOK) your own OpenAI-compatible key and endpoint. Traffic goes from the app to the endpoint you set, with no server of ours in between; the provider bills you directly.
Off the assistant is hidden and nothing is sent.

Per-host policy

Each host carries its own policy, and it decides who may receive a request:

Policy Behaviour
Strict (default for new hosts) local model only
Balanced cloud, with obvious secrets stripped from the prompt — private keys, tokens, password=…
Permissive cloud without redaction, for systems that hold nothing sensitive
Off the assistant is hidden on that host

Redaction is pattern matching. It removes what it recognizes and guarantees nothing — a Permissive host sends what you typed, verbatim.

What is sent

The request text and a fixed system prompt. Terminal output, the host list and vault records are not part of a request.

Two settings make the boundary visible: Preview prompt before sending shows the outgoing text so you can edit or cancel it, and Sanitize secrets before sending controls redaction for external providers.

Running what it suggests

A suggested command appears in a card that quotes it exactly. Running it takes an explicit confirmation; on a #prod host, or when the production guard scores it as risky, it takes a second one. A reply that is not a command says so instead of offering a run button.

When it refuses

Message Reason
Add an API key in AI settings first BYOK is selected but no key is saved
Strict policy: download the on-device model… the host is Strict and no local model is present
AI is turned off in AI settings the provider is Off

Provider failures are reported as what they are — invalid key, rate limit, network error, an unexpected response, or the local engine stopping — not as an empty answer.

Clone this wiki locally