X · found by @codeaholicguy
Compacts AI DevKit sessions by labeling messages, keeping only needed context, cutting 1.72M tokens to 5.9K
Jev judged this in 289 ms for $0.00010
Hoang Nguyen
@codeaholicguy
I built a session compact for AI DevKit, using Jev to label every message, and code keeps only what the next agent needs. Tested on a real session 1.72M → 5.9K tokens.
Sep 26, 2026 · 5 likesOpen on X
Found something worth sharing?
Drop the link and Jev reads it on the spot. What belongs lands in the community feed with your name on it.
Hossein Seifoory
@Hosseinberg
Wanted to try out @typesafeai Jev, got inspired by the fast-jev-compaction and built log-sum. It uses Jev to decide which lines to keep or drop. Definitely not optimized and there are better ways to do this. It was just a fun experiment.
GitHubRouting and model choice
Coordinates AI agents to complete tasks, verifies results against the original request, and exposes every step with per-call cost tracking
@mgtf
GitHubBenchmarks and evals
Reproducible benchmark measuring whether decision probes detect code regressions better than direct model judgment or execution tests
@Beldureil