-
Notifications
You must be signed in to change notification settings - Fork 0
Home
🇬🇧 English | 🇧🇷 Português
Stop burning 50,000 frontier tokens on missing packages, network flakiness, and circular doom loops.
jev-harnessis a zero-dependency System One decision harness, test failure triage gate, and semantic guardrail for AI coding agents — powered by TypeSafe AI's Jev System One non-autoregressive decision model.
When an autonomous AI coding agent hits a test failure or compiler error, the default behavior is catastrophically wasteful:
Test fails → Agent dumps 500-line traceback → Frontier LLM (GPT-6 Astra / Claude Fable 5.1)
→ 15 seconds + $0.50–$2.50 per call → Output: "pip install scipy"
This happens 70–80% of the time for trivially mechanical failures:
| Failure Type | Root Cause | Without Jev | With Jev |
|---|---|---|---|
| Missing dependency |
ModuleNotFoundError, Cannot find module, E0463
|
💸 50,000 tokens burned ($0.50–$2.50) | ⚡ 90ms triage ($0.00004) → pip install. 0 LLM tokens
|
| Flaky transient | Network timeout, ECONNREFUSED, port busy |
💸 LLM hallucinates architectural changes | ⚡ Auto-retry once. 0 code changes |
| Doom loop | Same fix attempted 3+ times | 💸 200,000+ tokens in circular failure | 🛑 Abort gate triggers. Developer alerted |
| Trivial task | Fix typo, reformat, git status | 💸 Frontier tier burns $0.80+ reasoning | ⚡ Route to local script or fast tier |
Inspired by Daniel Kahneman's cognitive science framework:
┌────────────────────────────────────────────────────────┐
│ AI Coding Agent Loop │
└──────────────────────────┬─────────────────────────────┘
│
Command/Test Execution
│
▼
[Test / Step Output]
│
┌─────────────────────┴─────────────────────┐
▼ ▼
[PASS: Continue] [FAIL: Error Log]
│
▼
┌───────────────────────┐
│ jev-harness gate │
│ (System One — 70ms) │
└───────────┬───────────┘
│
┌───────────────────────────────────────┴──────────────────────────────────┐
▼ ▼
[skip_llm = true] [skip_llm = false]
(Env missing / Flaky / Trivial) (Deep Logic Bug)
│ │
▼ ▼
Deterministic Shell Action Dispatch Targeted Trace
(pip/npm install or fast retry) to Frontier LLM (System 2)
⚡ 0 Frontier Tokens / Instant Fix 💸 Cost Reduced by ~80%
- System 1 (Jev): Non-autoregressive, parallel, typed decisions in 70–300ms at $0.042/1M tokens
- System 2 (Frontier LLMs): Slow, deliberative, generative — used only when truly needed
| Feature | Detail |
|---|---|
| 🛡️ Zero Dependencies | Pure Python stdlib (urllib.request, dataclasses, json). No pip bloat. Instant startup < 50ms |
| 🔌 Universal MCP Server | Exposes 6 Jev tools over stdio (jev-mcp) for Cursor, Claude Desktop, Antigravity, Windsurf, Zed, OpenCode |
| 🚦 UNIX Philosophy | Semantic exit codes: 0 (safe/skip_llm), 1 (abort/logic defect), 2 (syntax error). Clean pipe composition |
| 🔄 Offline Simulation | Intelligent local heuristics (< 500µs) if offline or no API key. CI never crashes |
| 🌐 Multi-Provider | TypeSafe AI Direct, OpenCode Zen (free tier), OpenRouter adapter |
| 🧠 Astra-Jev | Per-generation reasoning effort modulation — eliminates 3-minute CoT latency on mechanical tasks |
# Python
pip install jev-harness
# Node.js / TypeScript
npm install @ismaelsoilet/jev-harness
# Rust
cargo add jev-harness# Pipe your test runner through Jev
pytest 2>&1 | jev-harness test-gate
# → exit 0 (skip_llm=true): fix deterministically, no LLM call
# → exit 1 (skip_llm=false): forward to frontier model| Page | Description |
|---|---|
| 🏗️ Architecture | Tri-runtime layout, data flow, zero-dependency contracts |
| 🚦 Gates Reference | All 5 semantic gates: inputs, outputs, exit codes, examples |
| ⚡ Astra-Jev | Dynamic reasoning effort governance for 2026 frontier models |
| 🔌 Integrations | Config snippets for every IDE and agent |
| 📦 SDK Reference | Python, TypeScript, and Rust API documentation |
| 🔁 CI/CD & Git Hooks | GitHub Actions, pre-commit, Husky integration |
| 📋 Changelog | Release notes and version history |
| Metric | Frontier Reasoning (GPT-6 Astra, Claude Fable 5.1) | Fast Tier (Gemini 3.8 Flash) | Jev System One |
|---|---|---|---|
| Input Pricing | $10.00 / 1M tokens | $0.75 / 1M tokens | $0.042 / 1M tokens |
| Output Pricing | $50.00 / 1M tokens | $3.75 / 1M tokens | $0.00 (non-autoregressive) |
| Latency | 10,000–30,000ms | 1,500–4,000ms | 70–300ms |
| Hallucination | Stochastic | Stochastic | Zero — calibrated typed bounds |
If you are an autonomous AI agent operating in this repository, you MUST read AGENTS.md before executing any task.
Powered by TypeSafe AI's Jev System One — MIT License
⚡ Jev Harness v0.1.6 | PyPI | npm | crates.io | MIT License | Powered by TypeSafe AI's Jev System One
Navigation / Navegação
Getting Started
Gates Reference
- 🚦 All 5 Gates
- Gate 1 — Test Triage
- Gate 2 — Abort Check
- Gate 3 — Model Router
- Gate 4 — Step Verify
- Gate 5 — Astra-Jev
Integrations
- 🔌 All IDEs & Agents
- Claude Code
- OpenAI Codex
- Cursor IDE
- Antigravity IDE
- Windsurf / Zed
- Pi & Oh My Pi
SDK Reference
CI/CD
Começando
Referência dos Gates
Integrações
SDK
CI/CD
v0.1.6 — MIT License