Home / Guide

How to Stop AI Coding Agents From Repeating Mistakes

The complete guide to pre-action gates, feedback capture, history-aware lesson distillation, and automatic prevention rules.

The Problem

Your AI coding agent force-pushes to main. You correct it. Next session, it force-pushes again. You add a rule to CLAUDE.md. It ignores it. You lose an afternoon reverting.

This happens because prompt rules are suggestions. The agent can read them, forget them, or override them. There is no enforcement at the tool-call level.

The Fix: Pre-Action Gates

ThumbGate adds an enforcement layer between your agent and its tools. When the agent tries to execute a tool call, a PreToolUse hook fires before the action runs. The hook checks the call against known failure patterns. If it matches a gate, the action is blocked.

Agent: git push --force origin main
Result: Force-pushed. You lose 3 commits. Again.

Agent: git push --force origin main
[gate] Blocked: no-force-push (confidence: 0.94)
Agent: git push origin feature-branch
[gate] Passed

Install (One Command)

# Auto-detect your agent and configure hooks
npx thumbgate init

# Or specify your agent directly
npx thumbgate init --agent claude-code
npx thumbgate init --agent codex
npx thumbgate init --agent cursor
npx thumbgate init --agent gemini

How It Works

1. You give feedback

When your agent makes a mistake, tell it. ThumbGate captures the feedback as structured data with context, tags, and domain. If the signal is vague, it can reuse the recent conversation window and failed tool call to propose a better lesson instead of discarding the feedback.

# Your agent force-pushed. You say:
"thumbs down — force-pushed to main, lost commits"

# ThumbGate captures:
{
  signal: "negative",
  context: "force-pushed to main, lost commits",
  tags: ["git", "force-push", "destructive"],
  domain: "version-control"
}

2. Feedback auto-promotes to prevention rules

After repeated failures with the same pattern, ThumbGate generates a prevention rule automatically. No manual rule writing needed.

3. Rules become gates

Prevention rules are enforced as pre-action gates. The gate fires at the PreToolUse hook level — inside the agent's runtime, before the tool call executes.

4. Gates adapt via Thompson Sampling

Gates that block too aggressively (high false-positive rate) get their confidence reduced automatically. Gates that catch real mistakes get reinforced. This is Bayesian multi-armed bandit optimization, not static rules.

Memory That Persists Across Sessions

ThumbGate stores feedback in a local SQLite database with FTS5 full-text indexing. Lookups are sub-millisecond even at tens of thousands of entries. Old entries that contradict newer evidence are auto-pruned via Bayesian belief decay.

• recall — injects relevant context at session start
• search_lessons — finds promoted lessons with corrective actions
• retrieve_lessons — surfaces lessons for the tool or action you are about to run
• session_handoff — preserves continuity across sessions

History-Aware Feedback Sessions

ThumbGate supports linked feedback sessions for the messy reality of AI debugging. You can start one correction thread, append more context, and finalize it once the lesson is clear.

• open_feedback_session starts a linked correction thread.
• append_feedback_context adds later notes, failed tool output, or user corrections to the same thread.
• finalize_feedback_session promotes the combined evidence into one reusable lesson.
• reflect_on_feedback proposes a reusable rule from the same transcript when the failure pattern is obvious.

Pre-Action Gates vs Prompt Rules

ThumbGate vs Alternatives

Common Scenarios

Agent keeps deleting files

Give a thumbs-down: "deleted production config file." After 2-3 occurrences, ThumbGate generates a gate that blocks rm commands targeting config files.

Agent ignores test failures

Give a thumbs-down: "committed code with failing tests." ThumbGate learns the pattern and gates future commits when test results show failures.

Agent uses wrong API endpoint

Give a thumbs-down: "called staging API in production code." The gate blocks tool calls that reference staging URLs in production contexts.

Get Started

npx thumbgate init

One command. Works with Claude Code, Cursor, Codex, Gemini, Amp, and OpenCode. Claude Code can also call Codex for review, adversarial review, and second-pass handoffs through the repo-local bridge plugin.

Free keeps local enforcement with 5 daily feedback captures, 10 lesson searches, unlimited recall, blocking, and history-aware lesson distillation. Pro adds a personal local dashboard and DPO export. Team rollout adds the hosted shared lesson DB, org dashboard, and generated review views. Founder $49 one-time link stays preserved.