ronail’s jev-cost-guard puts Jev behind MCP, the standard for connecting agents to tools, to screen material before the main model reads it. It also labels tool results with a success probability.

The portable package supplies skills and an on-demand server. Actual context savings depend on the calling agent choosing what to retain.

Rank first, then read the evidence

  • jev_find applies a local keyword filter, then asks Jev to rank surviving files for relevance.
  • Files without matching words never reach that second stage; related code can disappear before semantic ranking begins.
  • jev_screen asks separate questions about instructions aimed at the assistant and relevance to its task. It returns admission labels and paths.
  • The default reader examines each file’s first 8,000 bytes. A verdict about that prefix cannot clear an unseen tail.

A success signal loses the explanation

  • Despite its name, jev_summarize returns a yes/no probability, with an explanatory wrapper; it creates no prose summary.
  • The author reports 24 billed output tokens across four synthetic success-result sizes. Input usage grows with the body; those measurements were not reproduced here.
  • All-positive examples cannot establish calibration, which means whether predicted probabilities match observed outcomes.
  • Replacing a diagnostic log with a success score can remove the detail needed to investigate its failure.

Missing fields and costs need explicit policy

  • An offline fixture with a missing noul field produced noul=None -> likely-success.
  • A second fixture supplied a low injection score but omitted requested relevance. The screen returned ADMIT; a missing injection answer returned UNSCORED.
  • The relay path enables paid fallback by default when a real key is configured and the free model returns certain errors. Set JEV_FALLBACK_TO_PAID=0 to disable it.
  • The separate tool-result hook prints a verdict and exits successfully. It leaves enforcement and context retention to the caller.

Verification: nine offline checks passed, including the permissive-response cases and stdio protocol discovery. No live Jev requests, token-saving benchmark or security audit was performed.

Typed output is not a correct decision. Keep source text recoverable, validate every required field and treat model admission as a triage suggestion.