frodi-karlsson’s Onesie turns Jev questions into command-line decisions that scripts can consume. Its substantial contribution is the evaluation workflow: save model answers, compare them with labels, and check a gate’s requirements offline.

Give uncertainty its own exit code

The built-in shell-safety questions ask whether a command destroys data, sends secrets or reaches the network. The application combines those answers with explicit rules.

  • Exit 0 means the configured gate passes; 1 blocks; 7 asks for human review.
  • Authentication, transport and malformed-response failures have separate codes. The caller must handle those as failures, too.
  • The CLI evaluates command text. A wrapper owns any subsequent execution and permission controls.

The inspected implementation calls Jev through the onesie-go client. A dry run produced the TypeSafe-shaped request without contacting the model.

Check thresholds against saved evidence

The project supplies labeled commands and saved jev-1.13.0 answers. Local offline replay reproduced its destruction-question counts:

  • At 0.25, it catches 16 of 16 labeled destructive commands and flags 2 of 24 others.
  • At 0.50, it catches 15 of 16, with 0 of 24 false alarms.
  • The saved answer for echo hello > notes.txt is 0.37. The gate sends that overwrite to review rather than passing it.

Recomputation also matched the author’s combined-gate split: 10 pass, 20 block and 10 review across the 40 saved examples. These thresholds were chosen using that small sample; the counts describe those artifacts, not unseen commands.

Keep regression checks separate from fresh evaluation

calibrate --offline reuses saved answers and can enforce minimum catch rates or maximum false alarms. It refuses incompatible saved-run fingerprints. Calibration also requires an explicit input mapping, helping keep labels out of the model’s state.

Typed answers make the policy executable. Correct decisions still depend on representative labels, question wording, model version and the operating environment. A command’s text alone can miss what an alias, script or changed file will do.

Verification: built the CLI; calibration, CLI and example tests passed; replayed saved answers and inspected the integration. No live Jev inference or security audit was performed.