rift
Rust terminal coding agent for local models whose swarm mode can use Jev to referee parallel attempts and pick a typed winner.
Overview
Rift is a fast terminal coding agent written in Rust and shipped as one binary of about 14 MB. It targets local models first (Ollama's native API and OpenAI-compatible servers like vLLM, LM Studio and llama.cpp), with Anthropic, OpenAI and OpenRouter optional. It addresses common failure modes of local agents: streaming that fights scrolling, context blowout (explicit num_ctx, truncation detection, two-stage compaction) and fragile tool calling. The README reports beating opencode on a 50-task suite (44 vs 42 solved, 57% fewer prompt tokens, 3.4x faster).
The Jev role is in WarpDrive, which races several models on a task in isolated git worktrees. A judge picks the winner to merge, and with --judge jev a System One model referees, returning the winner typed with a confidence instead of a generative model writing a verdict. The list describes this as an optional TypeSafe decision client for bounded Jev judgments.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which parallel attempt (worktree) should win and be merged? | — |
How it works
- rift swarm runs the task with several models in separate git worktrees
- Collect each attempt's result
- Jev judges and returns the winning attempt with confidence
- rift merge brings in the winner
Strengths
- Typed, confidence-scored referee instead of a free-text LLM verdict
- Local-first agent with strong reported efficiency
Caveats
- Jev is an optional judge in one feature
- Benchmarks are the author's own
Getting started
brew tap exYze/tap && brew install rift, then rift swarm "task" --models a,b --judge jev.At a glance
- Type
- Library
- Status
- beta
- Language
- Rust
- Runtime
- Terminal app (TUI/CLI)
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
An optional TypeSafe decision client in the Rust coding terminal Rift for bounded Jev judgments.
logicrw/awesome-jev-projects
Frequently asked questions
What does Jev decide in Rift?
Which of several parallel model attempts at a task should win.
Does Rift need Jev?
No; Jev is optional and an LLM can also judge.
Which models does Rift target?
Local models via Ollama or OpenAI-compatible servers, with cloud providers optional.
Related projects
Node/Bun CLI for asking Jev boolean, choice and score questions over text or JSON, with direct or Vercel AI Gateway providers.
CLIbetaTypeScriptDeclarative compiler that turns state, decisions and labeled examples into a measured, frozen JSON program of typed Jev questions.
LibraryprototypePythonVercel Labs terminal CLI for AI tasks that can use Jev as the evaluation model for its evaluate command.
CLITypeScriptTool that recomputes Jev's confidence identities on outputs published in public repositories to spot answers Jev never actually produced.
ServiceresearchPythonTypeSafe's official drop-in replacement for the SDK's system_one call backed by OpenAI, Anthropic or Gemini, for comparing Jev with LLMs.
LibraryproductionPythonHack Club's lightweight LLM proxy, listed for adding a Jev forwarding endpoint behind its auth, limits and usage logging.
MCP serverproductionTypeScript