Permafrost has been released—a new plugin for Claude Code designed to optimize DeepSeek API usage by stabilizing the system prompt prefix.
What Happened
Developer jianzhichun has released Permafrost, a tool that allows "freezing" the system prompt prefix in Claude Code. This solution prevents it from changing when new tools are added or context is updated (for example, when executing a git status command). Thanks to this stabilization, requests to DeepSeek consistently hit the automatic prefix cache, which can reduce API usage costs by 64% under real-world workloads.
Context
In agentic systems, dynamic context often causes "cache busting." When data at the beginning of a prompt constantly changes, the LLM provider's caching mechanisms stop working effectively, making every request expensive. Permafrost solves this problem at the interaction level between Claude Code and the API, without requiring model retraining.
Why It Matters for the Industry
This tool demonstrates the critical importance of prompt structure management for optimizing the cost of autonomous AI agents. It is a significant step toward transitioning from a "simple context" architecture to a "managed cacheable context" architecture, which could become a standard in future agentic frameworks like LangChain or CrewAI.
Why It Matters for Users
Developers using the Claude Code and DeepSeek combination gain the ability to radically reduce operational expenses (OpEx) on APIs without losing agent performance quality. This turns potentially expensive agentic work cycles into extremely cheap ones through the efficient use of prefix caching.
Sources
Author
Look at AI, Editorial Team