FILTERAgents & automation
Jev Codex Plugin
James@ThryveJames
An open-source Codex plugin that utilizes Jev to reduce the amount of evidence Codex needs to process by selecting relevant source text.
I built an open source Codex plugin that uses @typesafeai Jev to reduce the amount of evidence Codex has to process. I’m looking for feedback from people working on coding agents, retrieval, or context management.
The idea is straightforward: its search and log reading tools gather candidates, sends bounded excerpts to Jev for selection, and returns relevant source text with file paths and line numbers to Codex.
If that evidence is incomplete, Codex can identify the missing fact and make a targeted follow-up selection through Jev.
Early results from 108 runs, 12 synthetic investigation tasks, three repetitions comparing stock Codex, deterministic local selection, and Jev selection:
• 39.2% fewer Codex input tokens versus stock in the task-paired analysis
• 7.9% fewer Codex input tokens versus deterministic local selection
• 32.6% fewer total Codex output tokens versus stock
• 44.8% lower estimated combined API cost versus stock, including Jev
• 31.9% less total execution time versus stock
I also just tested the recovery behavior live. Codex found an incident’s cause in a large log, recognized that the accepted configuration range was still missing, and independently made a targeted Jev search that retrieved the exact validator. Both Jev requests together took 971 ms, and Codex produced the correct answer.
These are encouraging results from a controlled synthetic suite. I’d like to see how well they carry over to real repositories and messier investigations.
Would this fit your workflow? What failure cases should I test next, and what would make the evaluation more convincing to you?
Code and benchmark records:
github.com/jcressler/jev-codex-token-saver
Related builds
