User repository on GitHub
Check agent file changes against your project's written constraints, using TypeSafe AI's Jev decision model
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Pi Jev Gate
Pi extension that checks file-mutating tool calls against written rules in constraints.md, asking Jev which rule a change violates and offering rework with the rule text.
- Stars
- 1
- Language
- TypeScript
- Published by
- Community
- Author
- hcl-z
- Last verified
- Oct 2026
About this project
Check file changes against your project's written constraints, at the moment they happen. Your agent has read your AGENTS.md. It still drifts — three turns into a task it writes a file that plainly breaks one of your rules, and you find out at review time, after other work has been built on top. This pi extension hooks every file-mutating tool call, asks TypeSafe AI's Jev whether the change violates any rule in your constraints.md, and interrupts when it thinks one is broken:
Quoted from the project's README. Read the full README.
Works forJev for Tool calling
Used inJev for Software & DevTools
- Languages
- TypeScript
Highlights
- Hooks write, edit, and file-changing shell commands; read-only commands like test, status, and listing are never sent.
- Each heading is one independent question judged in parallel, so the verdict names which rule broke for targeted rework.
- Defaults use a 0.6 threshold with a pinned model version; the log records probabilities alongside the button pressed.
- Escape means rework with no timeout, and the ignore list lives only in memory so it cannot weaken rules on disk.
Watch out
A fail-open constraint advisor, not a security boundary; it is weak at arithmetic, dates, indirection, and adversarial input. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
No. Every failure mode lets the change through with a warning: missing key, network trouble, timeout, rate limit, overload, or unreadable response. A guard that breaks the agent on network hiccups would be uninstalled.
Because jev-latest is a moving alias and threshold tuning against it would silently shift behavior on release day. Override the model field when you intend to move, then re-check the threshold.
Write each rule like a contract naming the thing directly, then run /jev-gate once to see how headings split into questions. The dialog shows probabilities and the log records them with your decision for tuning.
Yes in principle. Diffs are treated as data, not hostile input, and anything that can run shell commands can bypass the hook. Use it to catch honest drift early, never as access control.
constraints.md in the project root or docs/constraints.md; both are read and merged, and the filename matches case-insensitively. Each ## or ### heading is one rule and bullets under it belong to that rule, so a rule and its exceptions get judged together. A flat list of top-level bullets works too, with each bullet as its own rule.
Only the operation, the path, and the change itself. The target file's existing contents and your project structure are deliberately withheld, because Jev loses accuracy when the input carries detail the question does not need. A project with no constraints.md at all stays completely silent and never calls the API, so a global install costs nothing elsewhere.
With log: true, one JSON line per check in ~/.pi/agent/jev-gate.jsonl, carrying the tool, path, operation, change, per-constraint probabilities, violations, threshold, model, verdict, elapsed time and token counts. Skipped checks are logged too as verdict degraded, so a gap in enforcement shows up afterwards. The file sits outside any project so it cannot be committed by accident, but it does record your file changes verbatim.
More Jev resources like this
2GitHub stars
Prompt-injection scanner that combines deterministic checks with focused Jev judgments and a configurable reasoning model, exposed over CLI, SDK, HTTPS, MCP, and editor plugins.
Jev Agents & MCP#community#github
Community
revsmoke/promptrejectormcpTypeScript
revsmoke/promptrejectormcpTypeScript1GitHub stars
OpenCode plugin that asks Jev once per prompt how much reasoning is needed, then maps minimal to xhigh effort onto the model's own variants without rewriting prompts.
Jev Agents & MCP#community#github
1GitHub stars
Community Node.js MCP server exposing Jev as four read-only advisory tools for bounded probabilistic decisions, with the host model keeping final authority over every action.
Jev Agents & MCP#community#github
Community
Afloat16/jev-mcpTypeScript
Afloat16/jev-mcpTypeScript0GitHub stars
MCP done-gate for coding agents: it diffs git reality, measures conventions deterministically, asks Jev eight quality questions, and returns a fix list.
Jev Agents & MCP#community#github
Community
BilalUchiha/vibecheck-mcpTypeScript
BilalUchiha/vibecheck-mcpTypeScript1.5kGitHub stars
Open-source platform that puts ACP-compatible agents such as Claude Code, Codex, and Pi inside the team's Slack, Telegram, Discord, and GitHub workflows, with reusable Jev decisions for routing, response gating, and reviewer selection.
Jev Agents & MCP#community#jevbest
488GitHub stars
A coding agent built on pi that hands its non-code decisions — what enters the context, whether a command is safe, whether the work is done — to a small judge model at 35 decision points per turn.
Jev Agents & MCP#community#jevbest
Community
qybaihe/muTypeScript
qybaihe/muTypeScriptFrom the Jev community
Posts from builders shipping with Jev right now.