ROUTEDevtools & code
Qualixar Jev Control for Codex
varun bhardwaj@varunPbhardwaj
Qualixar Jev Control for Codex is a system that classifies tasks before Codex expands context, offering modes like SKIP, SUGGEST, REQUIRE, and BLOCK to manage tool and context usage.
TakeThat means you can make narrow semantic decisions without generating free-form text and then trying to parse it back into routing logic.
Pattern↑ Typed decisions for agent routing and selection
I got early access to TypeSafe’s Jev with live API access.
What interested me was not “another model inside Codex.”
It was a more important question:
Why should a premium coding model spend its reasoning budget deciding which files, tests, tools, skills, or sources deserve attention in the first place?
Jev is TypeSafe’s System One model:
state + atomic typed questions → structured decisions.
Choice. Score. Noul.
That means you can make narrow semantic decisions without generating free-form text and then trying to parse it back into routing logic.
Today we’re releasing Qualixar Jev Control for Codex v1.1.1.
The key addition is Policy Mode — a local decision control plane that classifies each task before Codex expands context:
SKIP → Codex continues normally
SUGGEST → a bounded Jev workflow fits
REQUIRE → optional enforce mode requires the matching live decision before governed tools
BLOCK → sensitive material stays local
A few important boundaries:
Policy Mode defaults to assist
It does not call TypeSafe on every turn
The local classifier does not persist prompt text
The workbench now exposes 7 Codex tools + 20 bounded workflows for:
routing
file ranking
context selection
claim checks
completion review
test selection
research ranking
memory admission
For this launch, our verified live path is TypeSafe direct / jev-1.13.0:
all 20 nominal workflows completed
recorded outcomes: 15 RECOMMEND / 5 REVIEW
the 60 offline fixture contracts remain a separate evidence class
The bigger architectural point is simple:
decide what deserves context before paying to load it.
TypeSafe already documents intent routing in front of expensive LLM handlers and shows strong batching economics for shared-state questions.
Qualixar has not published a universal Codex token-saving percentage yet. That requires an on/off Codex-side benchmark, and that is the next measurement.
If you are building serious coding-agent workflows, this is the direction I believe matters:
Jev decides. Codex builds.
Open source repo:
lnkd.in/g989cbxY
Full architecture article:
lnkd.in/gcvEdHra
#Codex #AIEngineering #AIAgents #DeveloperTools #OpenSource #LLM #AIReliabilityEngineering
Related builds
