JevRev
LLM plus Jev alloy: JevSift, JevLoop, and JevLong over a CLI JSON protocol so the host agent keeps the session.
Overview
JevRev (github.com/Alex314618-create/JevRev, npm package jevrev 0.2.0, MIT, Node 20+) is an LLM plus Jev alloy with three parts: JevSift, JevLoop, and JevLong. The LLM proposes and executes. Jev filters paths, scores rounds, and raises reminders. The CLI speaks JSON and leaves the session in the host agent's hands. JevSift ranks 2 to 12 candidate cards and emits work orders only for survivors. JevLoop audits recorded commands, tests, metrics, and artifacts, then returns a next action such as fix, verify, continue, replan, or human. JevLong tails JSONL for stalls, repeated failures, drift, tool trouble, and budget risk. Providers are hosted Jev (POST /v1/systemone, key JEVREV_JEV_API_KEY or TYPESAFE_API_KEY), local SemIf, or a replay file. jevrev activation runs a shadow cost policy and does not call Jev. README case Tidal City compares the same ChatGPT-6-Sol-Ultra prompt with and without the three parts: the JevRev city has named buildings, districts, time of day, camera switches, and WASD walk or swim. README case JSONL asks a Node 20 ingester to process 25,000 production-shaped events (about 2.3 MB, 184 malformed lines, 258 duplicate IDs) with isolation, first-seen order, dedupe, and at least 2x a conservative baseline. Public GitHub counts were 670 stars and 28 forks when this listing was drafted.
Problem: A host agent that tries every idea in prose burns the token budget on weak branches, then declares victory when the demo merely looks finished.
Built for: Codex, Claude Code, OpenCode, and custom harness operators who want a CLI JSON protocol for approach sifting, round audits, and long-session watch, while the host agent keeps the session and writes the code.
First indexed on Jev Directory: 2026-09-29
Creator and team
- Name
- renard
- Handle
- @Alex314618-create
How Jev is used
- Role in the product flow
- Sideline decision layer: sift candidate approaches, audit each round against evidence, and watch long sessions, while the host LLM still owns execution
- Primitives
- ChoiceScoreNoul
- State in
- Protocol v1 JSON: task goal, constraints, success checks, and 2 to 12 candidate cards for Sift. Loop specs freeze criteria and protected surfaces. Long reads JSONL events the agent already wrote.
- Decision out
- Sift statuses keep, review, or reject, plus next_action implement, ask_human, revise_candidates, or relax_constraints. Decide returns winner, merge, probe_more, no_winner, or human_review. Loop returns the next round action only after facts are checked.
- npm install and npm run build, then node scripts/install-skill.mjs for codex, claude, agents, or dsh
- jevrev sift --provider jev on a rank request; strict survivors become bounded work orders
- Host agent implements only selected approaches and records a jevrev.evidence-bundle
- jevrev decide or jevrev loop audits the round; completion requires replayable evidence
- jevrev long watch and long status on .jevrev/long for stalls and protocol alerts
JevRev is the three-part alloy with a host-owned session. thruwire-foreman supervises a software factory above Codex or OpenCode workers and steers continue, stop, or finish. tianyucodings-jevharness lets an LLM author a harness, freezes the Python, and evolves it with GEPA. samuelfaj-distill routes consults inside a Rust TUI and keeps tool permissions in Distill. miuuyy-astra-ares sets reasoning effort inside a patched Codex binary. JevRev instead exposes jevrev sift, decide, loop, and long as JSON contracts (@typesafe-ai/sdk on the hosted path). Reason codes such as GOAL_MISMATCH, CONSTRAINT_RISK, and BUDGET_CUTOFF are policy rules, not a chain of thought. Hard-constraint failures can reject a card before the confidence route. Replay files must include candidate_order or JevRev refuses to apply answers to the wrong card. Builder notes are omitted from Decide judge state. The TUI (run jevrev) lists registered host sessions and shows Kanban evidence. Loop and Long switches persist per session and do not start an observer or steer the host. Cite Alex314618-create/JevRev README and docs/PROTOCOL.md.
Sourced performance claims
- JSONL sift: batch-index 0.9243 kept, regex-shortcut 0.9240 kept, state-scan 0.7965 cut by budget, baseline-parse 0.6100 rejected, worker-shards 0.6043 review, external-index 0.4931 rejected. regex-shortcut failed correctness (expected 184 malformed, actual 65). batch-index passed and was the winner. Mean throughput about 180,318 events/s versus about 180,531 for the failing regex path.Source: github.com/Alex314618-create/JevRev README JSONL event ingestion
- JevLoop on that case: round 1 returned fix_regression, round 2 verify, round 3 completed. Final state completed at round 3, 838 ms wall-clock, 1,628 tokens in replay provider accounting.Source: github.com/Alex314618-create/JevRev README JSONL event ingestion
- JevLong on the same case accepted 13 events on the first import and 0 on the duplicate import (13 duplicates). It reported a high-severity failure_loop alert and a critical protocol alert. Loop reached completed while Long progress_index stayed 0.Source: github.com/Alex314618-create/JevRev README JSONL event ingestion
- docs/PROTOCOL.md default-v1 thresholds include goal_fit 0.34, constraint_fit 0.55, feasibility 0.34, validation_quality 0.25, execution_value 0.45, confidence 0.2, duplicate 0.75. Example policy profile jev/jev-1.13.0.Source: github.com/Alex314618-create/JevRev docs/PROTOCOL.md
- Public GitHub repo Alex314618-create/JevRev had 670 stars and 28 forks when this listing was drafted. package.json version is 0.2.0.Source: GitHub API 2026-09-29 and package.json
Features and stack
Features
- JevSift, JevLoop, and JevLong as separate CLI commands with JSON envelopes
- Skill installer for Codex, Claude Code, agents, and dsh hosts
- Hosted Jev, local SemIf (default 127.0.0.1:4878), and offline replay
- Shadow activation policy that compares wrong-path cost with bounded probes and does not call Jev
- Interactive cockpit for registered sessions, evidence, and monitor switches
- README cases: Tidal City build comparison and a 25,000-event JSONL contract
Stack
- Node.js 20+
- TypeScript
- @typesafe-ai/sdk
- Optional local SemIf
Pricing: MIT package. Hosted Jev bills per System One call. SemIf and replay runs stay on your machine. README JSONL token figure is replay-provider accounting for that case, not a price quote.
Links
FAQ
- Does JevRev take over the coding session?
- README and docs/AUTHORITY.md keep execution with the host agent. JevRev returns JSON decisions, work orders, and alerts. The TUI monitor switches do not steer the host.
- What did the 25,000-event case actually prove?
- README: regex-shortcut and batch-index both scored about 0.92 and both were fast (about 180k events/s). regex missed malformed lines (65 versus 184). batch-index passed the contract and was selected. That is a project case, not an independent lab certificate.
- Does jevrev activation call TypeSafe?
- No. docs/PROTOCOL.md says activation always returns shadow true. It does not call Jev, start work, or change the host's next action.
- How is this different from Foreman or JevHarness?
- Foreman supervises factory evidence above coding workers. JevHarness freezes an LLM-written harness and evolves it. JevRev is the Sift, Loop, and Long CLI that any host can call without handing over the session.
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.