Overview & result
jev-superpowers builds on the obra/superpowers methodology for AI coding agents and adds System One typed decisions. According to the README, architectural forks, dependency selection, shell commands and completion checks are gated by those judgments, using either TypeSafe's hosted Jev or the open-weight Laya model running offline.
From the original post
Systematic software development framework for AI coding agents upgraded with TypeSafe Jev System One typed decisions
How Jev fits in the loop
- It starts from the obra/superpowers methodology for coding agents.
- Typed System One decisions gate architectural choices, dependency selection, shell commands and completion checks.
- A package-vetting step checks that suggested libraries are real before they are used.
- The decisions run on TypeSafe's hosted Jev or, offline, on the open-weight Laya model.
How to reproduce
- Inspect the original showcase and source material at https://github.com/AkashPriyadarshii/jev-superpowers
- Verify the bounded prompt and input candidate schema configured for Jev
- Benchmark decision latency and classification accuracy against baseline models
Why this build matters
It targets specific agent failures the author names: invented packages, guessed architecture, unchecked destructive commands and premature "done" claims.
Limitations
- The README calls its benchmark table design targets measured on the maintainer's setup and asks readers to reproduce them before quoting.
- It depends on the upstream superpowers workflow.
- Cloud and local backends are different models, so results may differ between them.