The previous sol stayed context around 100m but now it cut down to 272k， after compressing the context, Codex has about 80k context, and only 190k is available for use, which causes a situation during development where a single conversation task needs to compress the context multiple times.Compared to web-based conversations, the context in the CLI is simply not durable. The core reasons are: 1. Long logs in run commands fill the context. 2. Tool chain calls. 3. File read and edit, especially ai don't understand the project architecture and need to browse through files to read.So I developed a Mac app based on Codex that provides a Codex plugin, including MCP and a visual canvas. It solves three out of four problems: 1. Using MCP to call commands only returns the core content. 2. As the project develops, it automatically keeps up with the project structure, so AI can quickly catch up in new conversations. 3. Quickly locate the exact code position through contextos, reducing file read/write I/O. Although it doesn’t reduce 99% of the context as exaggeratedly claimed, in practice it lowers the frequency of context compression by about 60%.