DEV Community

Igor Ganapolsky
Igor Ganapolsky

Posted on Originally published at thumbgate.ai

Why CLAUDE.md and .cursorrules Fail After Turn 15 (And How to Enforce 0.4ms PreToolUse Gates)

#ai #claude #cursor #opensource

If you run Claude Code, Cursor, or Codex on a non-trivial production codebase, you have probably hit the same three walls:

  1. Passive rules rot after turn 15: You write NEVER run git push --force or domain invariants in CLAUDE.md or .cursorrules. Twenty turns into a debugging session, context compaction summarizes the history and the agent ignores the rule.
  2. Runaway retry loops burn 40%+ of your token budget: The agent hits a failing test or database connection error and runs the exact same failing command 6 times in a row, re-sending the entire 100k-token context window on every turn.
  3. Secret exfiltration and destructive commands: An agent debugging an API call runs curl -d "$GITHUB_TOKEN" or rm -rf ./ before you can hit Ctrl+C.

Prompts are suggestions. Hooks at the tool-call boundary are law.

Moving Memory from AGENTS.md into the PreToolUse Boundary

Instead of hoping the LLM remembers a paragraph in CLAUDE.md, we built ThumbGate — an open-source (MIT) pre-action firewall and loop circuit-breaker that sits at the PreToolUse boundary across Claude Code, Cursor, Codex, and Gemini CLI.

Here is what happens at runtime in 0.4ms before the shell or file tool executes:

git push origin feat/auth-fix         → ALLOW (safe branch · 0.3ms)
curl -d "$GITHUB_TOKEN" https://...   → DENY  (SOC2 secret pack)
aws rds delete-db-instance --prod     → DENY  (AWS safeguard pack)
git push --force origin main          → DENY  (Git/DB protection pack)
npm test [repeat #3 identical error]  → KILL_LOOP (-42% wasted tokens)
Enter fullscreen mode Exit fullscreen mode

When an agent makes a domain-specific mistake in your repo, one thumbs-down (capture_feedback) turns that failure into a local SQLite lesson. When the pattern repeats, it promotes into a deterministic blocking gate so the agent cannot make the same mistake twice.

1-Command Quickstart (MIT Free Forever)

You can install the local engine or pull our 5 agent skills via the skills.sh directory:

# Option A: Initialize local PreToolUse firewall + SQLite memory
npx thumbgate init

# Option B: Install via skills.sh Well-Known Discovery
npx skills add https://thumbgate.ai

# Option C: Homebrew (macOS & Linux)
brew tap igorganapolsky/tap && brew install thumbgate
Enter fullscreen mode Exit fullscreen mode

Pre-Built Threat Packs, Loop Circuit-Breaking & Managed Team Setup

The local CLI evaluator is free forever under MIT. For engineers and teams who do not want to write rules from scratch or babysit runaway token bills:

  • ThumbGate Pro — $19/mo: Includes 59 auto-updating Threat Rules (AWS Production Safeguards, SOC2 Secret Leak Prevention, Destructive Git & DB Protection), the Zero-Click Loop Circuit Breaker (cuts wasted retry-loop tokens by ~42% — pays for itself on the first stopped loop), personal dashboard, and HuggingFace DPO exports.
  • Cloud Control Plane & 1-Click SSO (/cloud): Test live payload compaction (POST /v1/cloud/intercept), toggle Threat Packs, and sync policies across your team.
  • Production Rule-Pack Kit — $49 One-Time (No Subscription): Expensable one-time download of all production rule packs, DPO/RLHF schemas, and CI templates.
  • 48-Hour Managed Agent Gate Setup — $499: Bring one costly repeated agent failure from your repo; leave within two business days with a configured gate, regression test suite, and written rollout/rollback proof (100% refunded if we cannot reduce it to a supported gate).

Top comments (0)