DEV Community

Igor Ganapolsky
Igor Ganapolsky

Posted on

I Built a Feedback-to-Enforcement Pipeline That Stops Claude Code From Repeating Mistakes

#ai #devtools #claudecode #webdev

If you use Claude Code, Cursor, or any AI coding agent daily, you've probably noticed three recurring problems:

  1. Session amnesia — Every new session starts from zero. Claude doesn't remember the architectural decisions you made yesterday.
  2. Hallucinated completions — Claude says "Done, all tests passing" when nothing actually passes.
  3. Repeated mistakes — You fix the same bug three times this week because the agent keeps making the same error.

I got frustrated enough to build a solution: thumbgate.

What it is

An MCP server that adds a reliability enforcement layer to AI coding agents. It's not just memory storage — it's a feedback-to-enforcement pipeline that learns from mistakes and blocks them from recurring.

The four tools

1. capture_feedback

When Claude does something wrong, you capture it:

# Captured automatically via MCP tool call
capture_feedback(
  feedback: "down",
  context: "Claude force-pushed without asking",
  whatWentWrong: "Overwrote teammate's commits",
  tags: ["git", "destructive"]
)
Enter fullscreen mode Exit fullscreen mode

2. prevention_rules

After repeated failures, the system auto-generates rules:

- Never force-push without explicit user confirmation
- Always run tests before claiming completion
- Check all 100+ occurrences when updating pricing strings, not just 3
Enter fullscreen mode Exit fullscreen mode

These rules persist outside the context window. They survive compaction.

3. satisfy_gate

Pre-action checkpoints that force the agent to prove conditions are met:

Gate: "CI green on current commit"
Status: BLOCKED — last CI run failed
Action: Agent cannot claim "done" until gate passes
Enter fullscreen mode Exit fullscreen mode

This kills the "hallucinated completion" pattern.

4. construct_context_pack

Bounded retrieval of relevant history for the current task. Instead of dumping everything into context, it selects what matters — prevention rules, recent feedback, task-specific decisions.

How it works under the hood

Each piece of feedback gets a reliability score via Thompson Sampling (beta-binomial posterior). Noisy or one-off signals don't immediately become rules. Only patterns that recur above a confidence threshold get promoted to prevention rules.

The gate engine uses a default-deny model for high-risk actions. The agent must pass through checkpoint validation before executing anything flagged by prior failures.

Install 

npx thumbgate serve
Enter fullscreen mode Exit fullscreen mode

Or add to your MCP config:

{
  "mcpServers": {
    "memory-gateway": {
      "command": "npx",
      "args": ["-y", "thumbgate", "serve"]
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

Works with Claude Code, Cursor, Codex, Gemini CLI, and any MCP client.

Numbers

  • Full regression suite and coverage reporting in CI
  • MCP tool surface for capture, recall, search, and gates
  • MIT licensed, free OSS core
  • $19/mo or $149/yr Pro tier for dashboard and analytics

What I'd love feedback on

  • The Thompson Sampling approach for reliability scoring — anyone tried different bandit strategies?
  • How do you handle the compaction problem in your workflow?
  • What's your current workaround for session amnesia?

Repo: github.com/IgorGanapolsky/ThumbGate

Disclosure: I'm the creator of this project.

Top comments (0)