Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

jev-laya-dsh

Local Laya decision layer for DeepSeek Harness: routes prompts, gates risky tool calls and swaps models, all offline at no token cost.

ProjectbetaMIT
View on GitHub ↗

Overview

jev-laya-dsh adds a local System 1 decision layer in front of DeepSeek Harness agents, following a User to Jev(Laya) to Agent pattern. Decisions come from Laya, an open, Jev-compatible calibrated model using a multilingual mmBERT encoder, running fully locally with no token cost. A router classifies each prompt (task type, whether a code model is needed, urgency) in one forward pass and appends an advisory notice; an optional gate runs a three-probe vote (destructive, irreversible, outside workspace) on bash and pwsh calls and escalates to user confirmation; and an optional per-tier model swap pins hard turns to a stronger model.

Serving keeps the ONNX runtime warm (about 2.5 s cold start, 0.3 s warm decisions) while unloading weights when idle or under memory pressure. Every layer fails open to no routing rather than blocking a turn. Documentation is in English, Chinese and Vietnamese.

What the decision model decides

TypeQuestionAnswers
ChoiceWhat type of task is this prompt (chat, code, architecture)?chat · code · architecture
Yes / noIs this command destructive, irreversible or outside the workspace?—

How it works

  1. Router classifies each prompt locally with Laya
  2. Gate votes on risky shell calls and escalates to the user
  3. Optionally swap to a stronger model for hard turns
  4. Fail open on any error

Strengths

  • Fully local with no token cost
  • Fail-open design
  • Multilingual encoder

Caveats

  • Laya accuracy may be lower than hosted Jev
  • DeepSeek Harness specific

At a glance

Type
Project
Status
beta
Runtime
DeepSeek Harness plugin with local ONNX runtime
Has code
yes
License
MIT

Use cases and integrations

How curated lists describe it

  • User → Jev(Laya) → Agent — a local System-1 decision layer for DeepSeek Harness agents. Laya (open, Jev-compatible) routes prompts, gates risky tool calls and swaps models in milliseconds. 100% local, $0/token. MIT.

    hellogumbo/awesome-jev

Frequently asked questions

What decides?

Laya, an open Jev-compatible model running locally.

Does it cost tokens?

No, decisions run locally.

What happens on errors?

Every layer fails open to normal behaviour.

Related projects

  • Keel

    Local-first macOS coding workspace that runs your agents and lets a local Laya or hosted Jev selector pick a route the host validates.

    ServicebetaRust
  • cursor-jev

    Cursor plugin whose hooks ask Jev whether subagent spawns, file reads, writes and shell commands are worth taking.

    PluginbetaJavaScript
  • pi-jev-governor

    Decision governor for Pi routing plan strategy, model tier, tool permissions and outcome (finish, retry, replan) through Jev.

    ProjectbetaTypeScript
  • s1-auto-router

    Claude Code plugin where a local Decider model on your Mac picks the model and effort level for each subagent task.

    PluginbetaPython
  • TriRouter

    Prompt router for Claude Code, Codex and Antigravity that picks model, effort, skill and language per request via Jev.

    ProjectbetaPython
  • Jev-chooses-a-LLM

    Jev Router for Cursor: Jev picks a COST, BALANCED or INTELLIGENCE tier per task, with soft routing in chat and hard routing via the SDK.

    ProjectprototypeTypeScript