# yolonot

> Configurable command guardrails for AI coding agents across 24 supported harnesses.

yolonot is a single-binary Go policy gate that evaluates shell commands before your AI coding agent runs them. It combines deterministic `.yolonot` rules, SHA256 content-pinned session memory, an in-process bash AST parser (`fast_allow`), external `exec:` hooks, fast typed decision models (`jev`, `ms-decision`), and chat LLM judges (`gemini`, `claude`, `openai`, `grok`, `ollama`, `openrouter`) with optional second-opinion escalation.

## Quick Install & Setup

```bash
# Homebrew (macOS / Linux)
brew install sezaakgun/tap/yolonot && yolonot setup

# Go install
go install github.com/sezaakgun/yolonot@latest && yolonot setup

# Import existing permissions & Auto Mode hints from any harness into .yolonot
yolonot rules discover
```

## Real-World Performance (n = 5,000 Sampled Gate Evaluations)

- **Allow / Ask / Deny Breakdown:** 96.5% Allow · 2.35% Ask · 0.17% Deny (hard blocks) · 0.97% Passthrough.
- **54.2% Instant Local (<1ms, ~1.6µs AST parse):** Over half of all commands are cleared locally by `fast_allow` bash AST parser, `.yolonot` rules, and project-scoped session memory with zero API calls.
- **Layered Risk Defense:** Combines deterministic `.yolonot` `deny-cmd` rules, `fast_allow` read-only AST verification, SHA256 script content pinning, the `critical` risk invariant (`critical` never maps to `allow`), and optional second-model escalation (which caught and hardened 8 under-rated risks in our 5,000-command sample).
- **Reduced Prompt Friction:** Second-opinion escalation rescued 48.3% of overly cautious Stage 1 `ask` verdicts to `allow`, leaving only a 1.75% human-approved prompt rate (which drops to 0% when promoted via `yolonot suggest`).
- **Bring Any Judge (LLMs, RL/Decision Models, or Custom AST Tools):** Bind any remote paid, open, or local (`ollama`) LLM, purpose-built decision/RL models (`TypeSafe Jev`, `Kev-4B` [Apache 2.0 Qwen3.5 LoRA], `Cloudflare Clef-flash`, `Microsoft-Decision-1`, `OpenAI Decisions`), or custom AST binaries (`exec:dippy`).
- **260-Case Reproducible Eval Suite (`yolonot eval --all`):** Includes 176 greenfield, 70 brownfield, and 14 script-attachment test cases (`Kev-4B`: 93.2% accuracy / 100% dangerous catch; `TypeSafe Jev` & `Clef-flash`: 92.0% accuracy / 100% dangerous catch; `script-attach`: 14/14 hidden attacks caught).
- **Tunable Profiles & Fast Headless Mode (`yolonot profile use <mode>`):** Switch between `balanced` (4.9% risky missed across 3 suites, 0% on Open Eval, 1.4% asked, 0.0% blocked), `fast` (`0%` ask for unattended headless runs; allows `safe/low/moderate`, hard-denies `high/critical`), `strict` (1.6% risky missed, 0% on high/crit & OS attacks), and `paranoid` (0.0% risky missed, 100% caught across all 456 cases).
- **24 Supported Harnesses:** One `.yolonot` policy file works across all 24 agents.
- **Full Benchmark & Methodology:** https://yolonot.dev/benchmarks

## 24 Supported Harnesses

- **Native Ask Supported (18):** `claude` (Claude Code + optional `--mod`), `codex` (Codex CLI with `.codexpre` dual-phase deduplication), `opencode` (OpenCode V2 + V1 plugin), `antigravity` (Google Antigravity / `agy`), `cursor` (Cursor IDE & CLI), `copilot` (GitHub Copilot CLI), `qwen` (Qwen Code), `amp` (Amp), `grok` (Grok CLI), `deepseek` (DeepSeek CLI), `kilo` (Kilo Code), `devin` (Devin), `pi` (Pi Coding Agent), `zcode` (ZCode), `muse` (Muse), `hermes` (Hermes), `openclaw` (OpenClaw), `mirror` (Mirror).
- **Ask-less Gates where unresolved `ask` fails closed to `deny` (6):** `cline` (Cline), `kiro` (Kiro), `crush` (Charm Crush), `goose` (Block Goose), `openhands` (OpenHands), `kimi` (Kimi CLI).

## Key CLI Subcommands

- `yolonot setup [--harness <name>]`: Auto-detects installed harnesses and registers hooks.
- `yolonot rules discover`: Scans existing host permissions and Auto Mode hints across Claude Code, Cursor, Codex, Copilot, Qwen, Grok, Hermes, Devin, Kiro, ZCode, and more, exporting them to `.yolonot`.
- `yolonot check "<command>"`: Dry-runs the full evaluation pipeline for a command.
- `yolonot status --all`: Inspects active profile, judge chain, rules, and hook registration.
- `yolonot doctor` & `yolonot audit`: Verifies installation, latency, rule precedence, and host sandbox settings.
- `yolonot suggest --list`: Suggests new `.yolonot` rules and hints from decision history and human resolutions.

## Links

- Landing Page: https://yolonot.dev/
- 24 Harnesses Reference: https://yolonot.dev/harnesses
- Source Code & Documentation: https://github.com/sezaakgun/yolonot