Overview & result
An agent skill that sends closed judgments to Jev and acts on the probabilities while the main coding model keeps writing code. Jev answers yes/no, choice and score questions, which the author says take about 250 ms, so the agent guesses less and interrupts the user only when a person must decide. Without an API key the bundled script fails open and the agent continues as usual.
From the original post
Agent Skill: send closed coding-agent judgments to TypeSafe Jev
How Jev fits in the loop
- Before asking the user a closed question, the agent checks with Jev whether the conversation already determines the answer.
- Before a command that may destroy data, leak secrets or reach beyond the task, it stops and confirms when any dimension crosses its threshold.
- After a command fails, it classifies the failure and checks the output for secrets before retrying.
- The bundled script redacts credentials, truncates long strings and never blocks the agent on a network error.
How to reproduce
- Inspect the original showcase and source material at https://github.com/HyunjunJeon/jev-judgment
- 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 states that Jev's confidence is evidence about the context, not permission to act: deletions outside the request, breaking changes, new dependencies and paid actions always go to the user.
Limitations
- Without a key the script fails open and the agent continues as if the skill were not installed.
- Requires python3 3.9+ and network access to api.typesafe.ai.
- The roughly 250 ms figure is the author's; no accuracy evaluation is given.