Question 1

Will installing this change how every Codex message is routed?

Accepted Answer

No. The README is explicit that installing does not auto-switch the main desktop session model and does not intercept every message, and there is no background service or automatic hook. In desktop mode Codex calls the selector for a suitable independent subtask and dispatches through host-native tools. Routing every turn on the macOS CLI is a separate upstream runtime you install per the guide.

Question 2

Does it actually save tokens, or is that just the pitch?

Accepted Answer

The project explicitly declines to promise fixed quota savings or task success rates. What it commits to is auditability: the router returns a decision ID, confidence and a local audit record, and the flow reports the model selection separately from the execution evidence, including a check of which model actually executed the subtask.

Question 3

What happens if the host cannot take an explicit model and effort?

Accepted Answer

The selection becomes advice. The routing script returns structured parameters and never creates agents itself, and the host must support passing model and reasoning effort explicitly for the chosen configuration to be adopted at all. If the result is invalid, under the confidence bar or deferred, the flow stops dispatching and the main agent continues analysing instead.

Question 4

Is it safe to install, and to install again over an existing copy?

Accepted Answer

Yes, with one caveat. The installer does not go online, does not read keys, and does not modify Hooks, AGENTS.md or existing model configuration; it copies eight files, returns already_installed when the content matches, and stops instead of overwriting when it differs. Back up the existing skill directory and your local edits before updating. Keys go in environment variables the Codex process can read, never in the skill or the repository.

Question 5

How does it discover candidate models?

Accepted Answer

Automatic model discovery is on by default: every routing run reads the local Codex model and reasoning-tier list, so a new model needs no manual configuration, and it needs a logged-in Codex supporting app-server model/list. You can inspect candidates with python route.py --config router.example.json --list-models, which calls no model at all. The README warns the catalogue may come from a client cache, so it does not prove your account's reasoning entitlements.

Question 6

What has been verified, and what has not?

Accepted Answer

v1.1.0 read 7 models and 40 model/tier combinations locally, passed 26 Windows tests with one skipped on link permissions, and made real Jev calls, though this round's full example fell under the 0.65 confidence bar and correctly stopped dispatching rather than counting as success. The macOS upstream CLI login and per-turn routing chain is still unverified and is not covered by cross-platform offline tests. Offline regression tests need no API key, and each release ships a .zip.sha256.

Question 7

How do I uninstall it?

Accepted Answer

Back up any personal edits, then remove only the jev-codex-router folder from your Skills directory, at $CODEX_HOME/skills or ~/.codex/skills when CODEX_HOME is unset. Config and audit records in the working directory survive that, as does a separately installed upstream runtime, so uninstalling the skill does not tear down the macOS CLI chain.

Jev Codex Router Skill

Highlights

Questions about this entry

More Jev resources like this

From the Jev community