Does Jev Save 90% of Codex Tokens? Skill and Router Limits

What the TypeSafe skill installs#

The official command is:

1
npx skills add typesafe-ai/skills --skill typesafe-ai

The installed skill gives an agent instructions for reading TypeSafe documentation, choosing Choice, Score, or Noul, designing criteria, calling the API, and interpreting probabilities. A skill is context and procedure for the agent. It is not a transparent network router.

This distinction matters because Codex must still load the task, repository instructions, and skill content before it decides to call Jev. Delegating a small classification after that point may reduce later reasoning, but it does not refund the input already processed by Codex.

Four integration patterns with different cost effects#

The fourth pattern is the only one that can keep an easy turn away from the stronger model before that model sees the prompt. It is also the pattern with the highest evaluation burden.

What the 237-turn backtest reports#

The community jev-codex-router project replayed 237 historical turns from seven days of one user’s Codex activity. Its published simulation estimated a baseline cost of about $871 and a routed cost of about$349, a reduction of 59.9%.

That result is useful as a hypothesis, not proof of equivalent production performance. The repository documents several limitations:

• It is a retrospective simulation, not live routed traffic.
• The calculation assumes constant token volume for a turn after changing models.
• It does not model cache invalidation caused by switching models.
• The workload belongs to one user and one historical period.
• It does not run the chosen model and compare correctness, review time, or accepted patches.
• Current routing policy behavior was not independently measured here.

The reported figure also does not establish a general 90% reduction. Different model prices, task mixes, cache behavior, and failure rates can move the result in either direction.

Measure the whole workflow, not one API call#

A useful evaluation records more than Jev’s low input price:

1
total cost
2
= router decision
3
+ selected model turn
4
+ cache misses caused by switching
5
+ retries and escalations
6
+ validation and review work

Track at least:

• Cost per completed and accepted task.
• Percentage of turns escalated to the stronger model.
• False-easy routes that later require rework.
• Test, review, and task-completion outcomes by route.
• End-to-end latency, including fallback.
• Cache-read and cache-write changes after model switches.

The existing Copilot per-model token report guide shows why input, output, and cache columns need separate interpretation. A lower raw token total does not prove a better development result.

A lower-risk way to start#

Use Jev first for decisions whose mistakes are cheap and observable: labeling incoming issues, ranking search candidates, selecting a read-only tool, or deciding whether a result needs manual review. Keep build failures, security-sensitive changes, migrations, and ambiguous multi-file work on the stronger path until the router has evidence for those categories.

For each route, log the Jev model version, probabilities, selected path, eventual outcome, and fallback reason. Do not log secrets or full repository context merely to make the router auditable.

The Jev capability guide explains which decisions match the model. The TypeScript API example shows how to keep thresholds and actions in code. If a managed API is unsuitable, Laya provides a local model path, with different hardware and accuracy trade-offs.

FAQ#