drift-guard

<div align="center">

[](https://github.com/JSLEEKR/drift-guard/stargazers) [](LICENSE) [](https://typescriptlang.org) [](#)

<br/>

Detect quality drift before your AI agent ships degraded code

Promise Extraction + Rule Engine + LLM Evaluation + Context Preservation

Quick Start | How It Works | MCP Tools

</div>

---

AI coding agents silently degrade during long sessions. Context gets compressed, early instructions get forgotten, and quality drops — without anyone noticing. The agent keeps responding confidently, never warning that it has lost track of your design decisions, style guides, or quality standards.

• Context compression -- as conversations grow, early instructions vanish. Your architecture decisions, style rules, and process requirements silently disappear
• Usage exhaustion -- when API quota runs low, subagents are automatically reduced. Steps get skipped, outputs get shorter, quality drops
• No warning system -- the agent doesn't know what it forgot, so it can't warn you. It confidently ships degraded work

drift-guard acts as an independent memory and quality guardian. It extracts your project's "promises" from CLAUDE.md, memory files, and specs — then continuously monitors whether those promises are being kept.

---

drift-guard operates as a three-layer quality pipeline that runs alongside your AI coding agent:

                         ┌─────────────────────────────────┐