pi-jev (y0usaf)
TypeSafe Jev as a decision layer for the Pi coding agent: a measured tool-call gate plus jev_ask for typed, calibrated answers.
Overview
pi-jev (github.com/y0usaf/pi-jev, npm @y0usaf/pi-jev) is Sami Ansari's Pi extension that batches four gate nouls and an impact Score on bash, write, and edit calls, judges bash output for leaks and failure class, and exposes a jev_ask tool for typed questions the model requests. Shadow mode is default: flagged calls notify without blocking until you switch to enforce. README documents fail-open behavior on missing keys, timeouts, 429s, and malformed responses so tool calls proceed when Jev is unavailable.
Problem: Pi agents can run destructive bash, exfiltrate secrets, or wander off-scope before you notice, and regex guards miss paraphrased risk.
Built for: Pi coding agent users who want measured Noul and Score gates on bash, write, and edit plus optional output judging on bash results.
First indexed on Jev Directory: 2026-09-23
Creator and team
- Name
- Sami Ansari
- Handle
- @realy0usaf
How Jev is used
- Role in the product flow
- Pre-tool gate on bash, write, and edit; post-tool output judge on bash; optional jev_ask for model-initiated judgments
- Primitives
- ChoiceScoreNoul
- State in
- Working directory, tool name, truncated arguments, last user message slice, and bash stdout snippets per README privacy section.
- Decision out
- Gate: destructive, exfiltration, beyond_scope nouls plus impact Score against documented blockOn thresholds. Output: leaks_secret noul and failure_class Choice with appended advice strings.
- pi install npm:@y0usaf/pi-jev loads extension with shadow gate default
- Four gate questions plus impact Score in one request (~300 ms README)
- tool_result hook runs leak and failure_class questions on bash output
- jev_ask accepts structured state plus question arrays for custom nouls and Choices
y0usaf treats Jev as a measured safety layer rather than an auto-approve toy: thresholds in blockOn come from README calibration tables on git status, rm -rf, curl exfil, and scoped edits, not guessed round numbers. Output judging catches secrets echoed after bash even when the gate cleared intent. Fail-open is honest product design for local dev, but operators should read What leaves the machine before pointing at production repos. Other Pi packages exist; this listing tracks y0usaf/pi-jev only. jomatsu-pi-jev-auto-mode focuses on auto-approve patterns; pi-warden supervises broader rule sets.
Sourced performance claims
- Gate batches four nouls and impact Score in one request for about 300 ms instead of four round trips per README.Source: github.com/y0usaf/pi-jev README
- Shadow mode is default; enforce mode prompts unless headless runs where gate.blockWithoutUI applies per README.Source: github.com/y0usaf/pi-jev README
- Every error path fails open: missing key, timeout, 429, or malformed response lets the tool call proceed with throttled error reporting.Source: github.com/y0usaf/pi-jev README
- Public GitHub repo y0usaf/pi-jev had 141 stars and 9 forks when this listing was drafted.Source: GitHub API September 2026
Features and stack
Features
- Gate on bash, write, edit with shadow default
- Output judge on bash with CLASS_ADVICE table
- jev_ask tool for typed Choice, Score, and Noul arrays
- Config at ~/.pi/agent/pi-jev.json with project overrides
- Slash commands /jev, /jev mode, /jev last, /jev check
Stack
- TypeScript Pi extension
- TypeSafe System One
- npm @y0usaf/pi-jev
Pricing: Open source extension; TypeSafe API key via env, config, or apiKeyFile per README.
Links
FAQ
- Does pi-jev block tools when Jev errors?
- No. README states fail-open behavior on errors so development is not bricked by a dead endpoint.
- What is shadow versus enforce?
- Shadow notifies on flagged calls. Enforce asks for confirmation before running flagged tools when UI is available.
- Is this the same as jomatsu pi-jev-auto-mode?
- No. jomatsu focuses on auto-approve thresholds for tool calls. y0usaf/pi-jev documents destructive, exfiltration, scope, and impact scoring with output judging.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- Jev use cases
The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.