Permission System

HelixMind uses a 3-tier permission system to control what the AI agent can do.

The 3 Tiers

Tier	Behavior	Tools
Safe (auto)	Runs without asking	read_file, list_directory, search_files, find_files, git_status, git_log, git_diff, spiral_query, spiral_store
Ask (prompt)	Asks for permission	edit_file, write_file, git_commit, run_command (non-destructive), web_research
Dangerous (always ask)	Always requires approval	run_command with rm -rf, dd, format, DROP TABLE, curl | sh, etc.

Permission Modes

Mode	Flag	Behavior
Default	.	Safe = auto, Ask = prompt, Dangerous = prompt
Skip-Permissions	--skip-permissions or /skip-permissions	Safe + Ask = auto, Dangerous = prompt
YOLO	--yolo or /yolo	Everything auto-approved (⚠️ use with caution)

Permission Prompts

When a tool requires permission:

🔒 edit_file: src/middleware/auth.ts (line 42)
   Allow? [y/n/always]:

Response	Effect
y	Allow this one time
n	Deny this operation
always	Allow this tool type for the rest of the session

Command Classification

The run_command tool classifies every command:

Category	Examples
Auto-approve	ls, cat, echo, npm test, tsc, eslint, git status
Ask	npm install, docker, make, python, mkdir, cp, mv
Dangerous	rm -rf, dd, mkfs, chmod 777, DROP, DELETE, TRUNCATE, curl | sh

Emergency Stop

Press ESC ESC (double-tap) to:

1. Immediately abort all running operations (main + background)
2. Open the checkpoint browser
3. Choose a state to revert to

Works in all modes: including YOLO.

Tips

1. Start with default mode until comfortable
2. Skip-permissions when actively watching the AI work
3. YOLO only in safe, isolated environments (containers, test projects)
4. Use /undo immediately if an operation doesn't look right