Agent feedback learning with error blocking

★ 7.0 · vibe-coding

thumbgate is a Claude Code skill that adds pre-action safety gates to AI agent workflows, intercepting tool calls via the PreToolUse hook and blocking known-bad patterns before they execute. It captures thumbs-up/down feedback with structured context, then auto-promotes repeated failures into prevention rules after a configurable threshold (default: 2 occurrences), using LanceDB vector similarity and Thompson Sampling confidence scores to evaluate risky actions. Five MCP tools are included: capture_feedback, search_lessons, get_prevention_rules, session_primer, and export_dpo — the last one exports production feedback as DPO training pairs for fine-tuning. Compatible with any MCP-supporting agent including Cursor, Codex, Gemini CLI, Amp, and OpenCode, making it practical for teams who need guardrails against recurring agent mistakes.