A coding agent and personal assistant built on the Pi coding agent. Jev checks every tool call, turn and voice transcript, and code decides what happens next: allow, ask or block an action, which model tier to use, and whether a "done" was actually verified.
Jev returns a structured decision for the local program; consult the source for the exact decision policy.
Adds structured choices or scores to the workflow; performance and cost benefits have not been independently verified.
The cited README and TypeSafe client were checked for protected-path precedence and numerical Choice validation. This site did not run a Reflex session or tool. Source review does not measure injection resistance or establish that a high-confidence answer grants permission.
Study a coding harness that uses judgments for action review, routing and progress monitoring.
Material analysed：
The README states protected paths ask regardless of model output and a model can add a question/block but cannot remove one imposed by code. This hierarchy is the key control contract.
Inspect the supporting source ↗client.ts checks Choice labels, candidate probabilities, numeric validity and the selected option's position before treating it as valid. It does not turn confidence into a permission.
Inspect the supporting source ↗Useful for examining a harness-wide decision layer. Compare cautious/balanced/bold policies and protected-path handling before assuming every tool gets the same behavior.
Inspect isValidChoice and README action-gate behavior together. Test a benign-looking proposed action against a protected path with a fake high-confidence safe response.