Does this quietly spawn reviewer agents behind my back?

No. Your normal Pi model stays the only coding agent, with no planners, subagents, reviewers, task graphs or orchestration, and the final diff review explicitly creates no other agent. A REVIEW or FAIL verdict can trigger a same-agent follow-up in the same session, capped at three automatic diff reviews per user task.

What stops the pre-edit gate from blocking every change?

Two thresholds and an escape hatch. Hard semantic concerns block at 0.80 while purely epistemic concerns need 0.90, and a strong explicit jev_evidence result for the same blocked target, at or above 0.85 supported and 0.70 confidence and non-ambiguous, can clear an epistemic-only retry. It never overrides a hard concern, and telemetry records how each block was resolved.

What happens if the TypeSafe API is down mid-session?

Automatic checks fail open with a warning instead of making the repository uneditable. An error record is written to telemetry so a skipped check is visible, but the practical effect is that an outage removes enforcement rather than blocking your work, which is the trade the README makes deliberately.

Is the final review actually independent of the agent that wrote the code?

It is a separate typed model call over the diff, not the same agent re-reading its own context, and it reviews from a baseline commit rather than from the agent's last message. The README is explicit that PASS is post-hoc confirmation after the fact, not a guarantee, which is why the unverified_claim question targets consequential claims that still lack repository evidence.

What are the install requirements for pi-jev-code?

Pi 0.85 or newer, Node 20 or newer, and a TYPESAFE_API_KEY. Install with ./install.sh and then pi install "$(pwd)".

How do the pre-edit thresholds split between hard and soft concerns?

Hard semantic concerns such as out_of_scope, target_mismatch and requirement_ambiguity block at 0.80, while purely epistemic concerns such as insufficient_evidence block only at 0.90.

What telemetry does it write, and where?

One JSON object per line to .pi/jev-code/telemetry.jsonl, append-only across slices. Large source excerpts, diffs and tool outputs are represented by hashes and compact metadata rather than written out.

CommunityJev Agents & MCP

Pi Jev Code

A Pi extension that keeps your existing coding model as the only agent and adds Jev as a coprocessor for search reranking, a pre-edit semantic gate, and a diff review against the task's starting commit.

See highlights

Open github.com

Stars: 0
Language: TypeScript
License: MIT
Published by: Community
Author: KamilPostrozny
Last verified: Oct 2026

KamilPostrozny/pi-jev-code

User repository on GitHub

View on GitHub

Single-agent Pi coding coprocessor with Jev semantic gates, baseline-to-current diff review, and append-only observability telemetry.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

v0.5.2 fixes a v0.5.1 runtime regression where the tool-result hook referenced looksLikeSearchCommand without importing/exporting it from the extracted shell-policy module. The search classifier is now exported from src/policy.js, imported by src/index.ts, and covered by a regression test including heredoc handling. Single-agent Jev coprocessor for Pi coding sessions. Your normal Pi model remains the only coding agent; there are no planners, subagents, reviewers, task graphs, or orchestration.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification Jev for Tool calling

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Hard concerns such as out_of_scope and target_mismatch block an edit at 0.80; purely epistemic concerns need 0.90.
The gate reconstructs the target file after the proposed change and includes the cumulative diff, so splitting a write hides nothing.
Final review runs from a git baseline captured at each prompt, so mid-task commits are covered and new untracked source files are included.
Append-only telemetry at .pi/jev-code/telemetry.jsonl records every gate outcome as hashes, and links a block to what happened next.

Watch out

This is a local-path Pi install rather than a published package: Pi loads it in place, so the npm runtime dependencies have to already exist in the package directory or npm install --omit=dev has to be run. It requires Pi 0.85 or newer, Node 20 or newer and a TypeSafe key. The README is candid that the pre-edit thresholds are still heuristics you should calibrate from real session logs, and that the search parser is intentionally conservative, so use the explicit jev_rank tool for structured LSP candidate sets or unusual output formats. Automatic TypeSafe failures fail open with a warning rather than making the repository uneditable, which is a deliberate choice but does mean an outage quietly removes the gate. One earlier release shipped a broken import in the tool-result hook that the current version fixes. Telemetry stores hashes and metadata, never source excerpts or diffs. 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.

More Jev resources like this

JevLoop

Runnable agent harness that moves pick, score and yes-no loop choices to Jev or Laya in milliseconds while reserving the language model for writing, compiling DECISION.md into executable questions.