safe-autonomous-mode

To prevent unsafe git operations and shell commands, you need per-action risk classification with conservative default-block policies. This evaluates every proposed action with a binary allow/block decision and structured reasoning before execution during unattended workflows.

Per-action risk classification for autonomous codebase operations evaluates every proposed file change, shell command, and network access with a structured verdict. It provides reasoning and a binary allow/block decision to enforce safe allow/deny policies during unattended workflows.

Configuring allow/deny rules for unattended code maintenance agents involves maintaining conservative defaults for safe behavior while permitting project-specific overrides. This allows read-only tasks, builds, and non-destructive edits while blocking destructive commands and database mutations.

You can run autonomous agents safely without wasting tokens by using pacing and focus-aware autonomy. This provides tick-based execution, adaptive sleep durations, and different behavior thresholds when the user is focused versus away to avoid unnecessary prompts.

The best way to block destructive shell commands in semi-autonomous workflows is applying rule-driven allow/deny policy composition with conservative default-block behavior. This requires per-action risk classification to request confirmation for pushes, destructive commands, or database mutations.

You should not use default-block behavior for autonomous operations when performing trusted, non-destructive project tasks like read-only operations, builds, and safe edits. Project-specific overrides can be configured to permit these actions while maintaining guardrails for high-risk changes.