dragon-guard

To prevent AI agents from executing unsafe mutating tool calls, you can enforce interactive permission checks and mode-based policies. This ensures exploratory analysis does not cause accidental changes to files, system settings, git, or package operations.

The best way to manage agent permissions is by using three-tier modes: Dog for permission-gated actions, Puppy for read-only planning, and Dragon for full implementation. This matches your risk tolerance and intent during development workflows.

Yes, you can temporarily allow repetitive agent actions without persisting risky overrides by using session-scoped allow and deny lists. This restricts permissions to the current session, preventing permanent risky configurations.

Bash command classification for agent safety works by heuristically evaluating commands to identify potential mutations. It pairs with safe-tool whitelisting to always allow read-only inspection tools while prompting for permission on ambiguous or mutating actions.

Yes, interactive permission control supports complex prompts during active development by optionally auto-switching to Puppy Mode. It also summarizes risky calls using LLM summaries to help you review potentially unsafe mutations.

You should not use full implementation mode, such as Dragon Mode, when performing exploratory analysis or planning. Instead, use read-only Puppy Mode to prevent accidental mutations and ensure unsafe or unintended tool calls are blocked.