Sandboxes tool output, persists session memory, enforces routing. Only the first verb has a number attached, and 98 percent is a big enough figure that a reader will want to know what it was measured against.
Three separate repos on this board work on the window rather than the model, which is what happens once everyone has the same model and different amounts of room to use it.
Naming both means the tool sits between the agent and its tools rather than inside the prompt, which is where an interception layer has to live to catch every call.
No platform list, no measured workload, no memory format, no token counts. The trending description gives two numbers and no method.
Concept mock, not affiliated with the project. The GitHub Trending description on 8 September 2026 reads "Context window optimization for AI coding agents. Sandboxes tool output (98% reduction), persists session memory, and enforces routing across 17 platforms via MCP + hooks." The two bars are drawn directly from the quoted 98 percent figure and represent nothing else. No workload, tool, or baseline is shown, because the description names none.