Question 1

Does it actually cut token spend, or did the benchmarks just fail?

Accepted Answer

There is no single validated savings percentage, and the README says so. The Fable diagnostic showed 34.39% accumulated input reduction and 41.46% peak, labelled a diagnostic rather than a production claim. The Sonnet mixed session showed -0.03% accumulated and -0.06% peak: no savings demonstrated.

Question 2

Will this make my agent slower on every request?

Accepted Answer

Both current trials were slower. Baseline 39.77 s became 210.54 s on the Fable study and 35.14 s became 171.99 s on the Sonnet session, each with three Jev failures. A Jev outage can also hold a request up to the 45-second evaluation deadline, and combined cost is unknown because usage receipts are incomplete.

Question 3

Does my source code get sent to TypeSafe?

Accepted Answer

Yes, when Jev is enabled. Candidate spans go to TypeSafe through the Vercel AI Gateway, including source code from Write, Edit and Bash arguments, and there is no zero-data-retention option. Without a Gateway key Yoshi forwards requests unchanged. Authentication headers pass through and are never written to traces.

Question 4

What happens when a judgment fails or the model is wrong?

Accepted Answer

The failure direction is safe. Invalid, failed and timed-out judgments keep their span, source hashes and exact offsets are rechecked before a verdict is applied, and decisions are frozen so later turns only judge new arrivals. Human instructions and assistant notes stay whole, and YOSHI_DRY_RUN=1 logs decisions without applying them.

Question 5

How do I install and start Yoshi, and can it run without Jev?

Accepted Answer

Run bun install, copy .env.example to .env, then bun start. Set AI_GATEWAY_API_KEY to enable Jev; without a Gateway key the default configuration forwards requests unchanged, and YOSHI_DRY_RUN=1 logs decisions without applying them.

Question 6

What size gate triggers pruning, and how are spans split?

Accepted Answer

The default gate starts above 50,000 estimated input tokens; below it new content passes unjudged. Above it, Yoshi resolves completed call/result pairs and splits eligible arguments, results and reports into spans of about 1,200 characters at line boundaries.

Question 7

What is Yoshi's license, and is this the final version?

Accepted Answer

It is MIT-licensed, but the README labels it an experimental proof of concept whose real version is heading into CompozyOS. There, pruning becomes automatic and judging moves between turns rather than onto the request path.

Yoshi

About this project

Highlights

Quickstart

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community