Docs track: Current (v0.1). Versioned docs planned.

Safety Guide

terminaI is a “terminal operator”. Safety is enforced in layers:

  • Deterministic enforcement (parser/policy driven) is the line of defense.
  • Model risk assessment is used for UX (explain consequences, choose preview/confirm), not for final allow/deny.

Approval ladder (A/B/C)

TerminaI routes actions into three review levels:

  • Level A: no approval (read-only, bounded, reversible; includes low-impact git ops like git add / git commit)
  • Level B: click-to-approve after a clear plain-English explanation of ramifications
  • Level C: click-to-approve + a 6-digit PIN for extreme/irreversible actions

If parsing is uncertain or scope is unbounded, TerminaI fails closed into a higher level.

PIN

PIN is stored in settings as security.approvalPin (default: 000000). See Configuration.

Where it applies today

  • Shell tool: uses a deterministic action profile + minimum review computation; Level C requires PIN.
  • Other tools: continue to use tool-specific confirmation prompts; they should migrate to the same ladder over time.

In clients

The same confirmation semantics work across:

  • CLI (Ink TUI)
  • Desktop (Tauri)
  • Browser /ui (A2A web client)

Level C prompts for a PIN in all clients.

Architecture details

See Safety Architecture for the detailed pipeline and invariants.