← all repositories
Alex314618-create/JevRev

A spine for your agent: cheap reflexes, expensive thinking

JevRev borrows evolution's trick — a cheap, fast decision layer (the spine) so your LLM's expensive reasoning (the brain) is only spent where it matters.

★706 stars TypeScript AgentsLLMOps · Eval
JevRev
Too new to rank. Velocity needs a few days of history.
not enough history yet
star history

What it does

JevRev is a TypeScript CLI (Node 20+, MIT) that wraps an LLM agent workflow in three parts. JevSift takes the several genuinely different approaches your model proposed, scores them, and issues bounded probe work orders — budgets and stop conditions included — so tokens go only to directions that earn them. JevLoop audits each round against recorded evidence (commands, tests, metrics, artifacts) and returns the next move: fix, verify, continue, replan, or human. JevLong tails the agent’s JSONL event stream on long tasks, flagging stalls, repeated failures, drift, tool problems, and budget risk. The host agent — Codex, Claude Code, OpenCode, or a CI harness — keeps the session; JevRev decides, filters, and nudges from the sidelines.

The interesting bit

The README’s JSONL case is the whole pitch in miniature: six approaches entered the sift, and the regex-based one was marginally faster on paper (180,530 vs 180,317 events/s) but caught only 65 of 184 malformed lines. JevRev picked batch-index because it met the task contract, and the loop converged in three rounds — 838 ms wall-clock, 1,628 tokens. A one-shot workflow would have shipped the regex. The repo tagline yells “changes EVERYTHING”; this quiet results table is the better argument.

Key highlights

  • Sift scores competing proposals and gives survivors real probes with budgets and stop conditions, instead of letting the model marry its first idea.
  • Loop gates every “completed” on replayable evidence — nothing that merely looks good gets rubber-stamped.
  • Long raises severity-graded alerts on long sessions; the demo ends with a high-severity failure_loop and a critical protocol alert while the loop still completes.
  • Pluggable providers: hosted Jev, local SemIf, or a replay file for fully offline runs; credentials live in environment variables, never command arguments.
  • Ships a TUI cockpit (Sessions, Kanban/Monitor, Config) that degrades to plain human/JSON output outside an interactive terminal.

Caveats

  • The activation policy — deciding whether a task deserves a Sift round at all — currently runs in shadow mode: it doesn’t call Jev and won’t start or skip Sift for you.
  • The README never explains what Jev actually is; you get an API endpoint (/v1/systemone) and a provider interface, and that’s about it.
  • The case studies are the project’s own showcases — the TidalCity before/after is illustrative, not a controlled benchmark.

Verdict

Worth a look if you run long agent sessions and want a cheap decision layer gating expensive reasoning — the JSONL case shows the mechanism earning its keep. Skip it if you wanted an autopilot: JevRev deliberately stays on the sidelines, and the host agent keeps the wheel.

Frequently asked

What is Alex314618-create/JevRev?
JevRev borrows evolution's trick — a cheap, fast decision layer (the spine) so your LLM's expensive reasoning (the brain) is only spent where it matters.
Is JevRev open source?
Yes — Alex314618-create/JevRev is open source, released under the MIT license.
What language is JevRev written in?
Alex314618-create/JevRev is primarily written in TypeScript.
How popular is JevRev?
Alex314618-create/JevRev has 706 stars on GitHub.
Where can I find JevRev?
Alex314618-create/JevRev is on GitHub at https://github.com/Alex314618-create/JevRev.

heatdrop uses Google Analytics to see which pages get read — nothing else. Your call. How we handle data.