MEDIA PREVIEW
Overview & result
jev-classifier brings Jev tool selection to Codex and Claude Code; the author says it reduces input-token usage substantially by selecting tools before the main coding-agent call.
From the original post
I built jev-classifier, an open-source tool that brings JEV tool selection to coding agents like Codex and Claude Code.
In my benchmarks, input-token usage per task dropped ~88% on Claude Fable 5.1 and ~85% on GPT-6 Astra.
https://github.com/felpsdev/jev-classifier
How Jev fits in the loop
- The coding agent receives a user task and a catalog of tools or skills.
- Jev chooses the relevant bounded tool or skill.
- The coding agent executes with the selected context and handles failures or fallbacks.
How to reproduce
- Read the repository benchmark instructions and record the exact model, tool catalog, and task set.
- Run the full-tool-schema baseline and the Jev-filtered condition on the same tasks.
- Report token usage, selected-tool accuracy, latency, and false-selection cases separately.
Why this build matters
It addresses a practical scaling problem for coding agents: large tool registries consume context even when only one tool is relevant.
Limitations
- The reviewed post does not expose the benchmark task set or the decision rubric.
- Token savings can hide missed tools, so selection accuracy and fallback behavior must be measured.
- The reported percentages are author-reported, not JevForAgents measurements.