# ADR-0003: Keep Native Security-Critical Capabilities in Rust - Status: Accepted - Date: 2026-07-18 - Scope: Credentials, diagnostics, local actions, and file transactions ## Context AstrBot Desktop needs Windows Credential Manager access, bounded diagnostics, UI Automation, window capture, event-log queries, and recoverable NAS file transactions. Remote model output, screen content, and file content are untrusted. Exposing generic operating-system execution or unrestricted paths would violate the project security model. ## Options considered 1. Expose broad filesystem or command APIs to React and enforce safety in the UI. 2. Add a generic command runner and restrict it through prompts or string validation. 3. Expose only typed, bounded Tauri commands backed by Rust policy checks and Windows-native APIs. ## Decision All credentials, operating-system inspection, local policy decisions, and filesystem mutations remain in Rust. React receives only display-safe results, opaque credential references, bounded path-selection tokens, and explicit operation states. It never receives credential values or unrestricted filesystem access. The product does not include a shell plugin or any generic PowerShell, CMD, shell, registry-write, keyboard-script, or arbitrary-path execution API. Local actions follow schema validation, tool allowlisting, profile and slot validation, path-scope validation, risk classification, authorization, execution, and audit. P00 NAS validation uses plan, dry run, temporary copy, size and SHA-256 verification, and atomic destination commit. Source files are preserved; source cleanup is represented only as a non-executed plan field. ## Rationale - Makes security enforcement independent of UI state and remote prompts. - Limits the native attack surface to explicit operations. - Keeps secret values out of React, SQLite ordinary fields, logs, and snapshots. - Preserves source data during an interrupted or failed NAS operation. ## Consequences - Windows-native behavior requires Windows-specific tests and cannot be certified by Linux CI. - Unsupported or elevated targets must fail safely without privilege escalation. - New native tools require an explicit schema, allowlist entry, risk level, tests, and audit behavior. - Destructive cleanup, updater signing, and broader desktop control require separate decisions and authorization.