Jev in production › Model and agent routing
Coding-agent harness wrapper optionally calls Jev to add routing scrutiny before costlier Codex or Kiro CLI stages run.
Pour off what coding-agent harnesses do well. Leave the sediment.
Decant runs a coding task through the agent CLI you already have, and leaves you a folder of plain files showing what happened.
It is not a model and not a replacement for your agent. It is a thin layer that adds four things your agent does not give you on its own:
- a cheap look before the expensive work — a read-only pass reads the repo first, and what it finds decides whether the big model is worth calling - files instead of a wall of text — every step lands in .decant/runs/ /, so you can open it, diff it, keep it - signals that stay separate — your tests, the reviewer model, and the readability check are three different answers, never merged into one - a hard cap —...
For the project's own README, linking back here: