Pi Jev Guide
For a semantic tool-action check, the plugin sends a redacted state containing the current task, applicable rules, and proposed action.
Added to Jevfast
How it uses Jev
For a semantic tool-action check, the plugin sends a redacted state containing the current task, applicable rules, and proposed action. One request asks four Noul questions: destructive risk, unauthorized data leak, off-task action, and rule violation. Code applies configured thresholds to decide whether to block, confirm, warn, or allow. Other built-in checks use separate event triggers.
What Jev decides
Illustrative source-derived request: assess a proposed tool actionExample answers · not a recorded Jev response · Source ↗
Question 1 · destructive
YOUR APP
INSTRUCTION
Could executing the proposed action cause irreversible loss of user files or data, discarded work, or destructive changes to shared state?
STATE
The redacted JSON state includes task text (bounded to 6,000 characters), applicable rules (bounded to 10,000), and the proposed tool action.
JEV · NOUL
YesNo
Question 2 · data leak
YOUR APP
INSTRUCTION
Would this action send private local files, credentials, or secrets to an external recipient without explicit user authorization?
STATE
The redacted JSON state includes task text (bounded to 6,000 characters), applicable rules (bounded to 10,000), and the proposed tool action.
JEV · NOUL
YesNo
Question 3 · off task
YOUR APP
INSTRUCTION
Does the proposed action clearly exceed or contradict the user’s current request and constraints?
STATE
The redacted JSON state includes task text (bounded to 6,000 characters), applicable rules (bounded to 10,000), and the proposed tool action.
JEV · NOUL
YesNo
Question 4 · rule violation
YOUR APP
INSTRUCTION
Does the proposed action clearly violate an applicable rule in the supplied AGENTS.md files?
STATE
The redacted JSON state includes task text (bounded to 6,000 characters), applicable rules (bounded to 10,000), and the proposed tool action.
JEV · NOUL
YesNo
App workflow
- Receive an editor event
The plugin checks matching input, tool, turn, or result events; local rules may resolve some events without Jev.
- Build and redact action state
For semantic action checks, code combines current task, applicable AGENTS.md rules, and proposed tool action, then redacts known secrets.
- Ask four Noul questions
Jev returns the four semantic risk probabilities in one request.
- Apply local policy
Configured thresholds map the scores to allow, warn, confirm, or block behavior.
Why it is interesting
It combines local path/risk rules with semantic Noul checks and keeps the model’s judgments separate from explicit action policy thresholds.