ThumbGate: Pre-action safety gates for Forge agents

[IgorGanapolsky]

Hey Forge team and community!

I built ThumbGate — pre-action gates for AI coding agents. It captures thumbs-down feedback and turns it into prevention rules that physically block agents from repeating mistakes.

The problem ThumbGate solves for Forge users

When using Forge (or any AI agent), the agent might:

• Force-push to main
• Commit .env files to public repos
• Delete production branches
• Run destructive shell commands

You correct it once... but next session, it does the same thing again. Instructions in AGENTS.md are suggestions — the agent can ignore them after context drift.

How ThumbGate works with Forge

ThumbGate ships as a Forge skill that intercepts tool calls before execution:

npx thumbgate init --agent=forge

This creates a .forge/skills/thumbgate/SKILL.md that:

1. Intercepts tool calls (Bash, file writes, git operations)
2. Checks them against prevention rules generated from your feedback
3. Blocks known-bad patterns before execution
4. Learns from every thumbs-down — rules compound across sessions

Example

Session 1: Agent tries git push --force origin main
You: 👎 "Never force-push to main"
→ ThumbGate generates a prevention rule

Session 2: Agent tries git push --force origin main again
→ ThumbGate blocks it. Agent uses git push instead.

Quick start

npx thumbgate init --agent=forge

Free, local-first, open source. No data leaves your machine.

Would love feedback from Forge users on this integration. What patterns do your agents repeat that you wish they would stop?

🔗 GitHub | npm