lev-os/agent-lease

記録された最新のソース活動: 2026年3月7日 06:22

SkillsMP カタログ更新: 2026年11月9日 14:22

収集済み skills: 1

GitHub スター: 1
GitHub フォーク: 0

AI アシスタントでインストール

このプロンプトをコピーして、使用中の AI アシスタントに送ってください。

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/lev-os/agent-lease.

リポジトリの README

README の元ファイル：README.md

README の元ファイルを読む

README 全文を表示

[![npm version](https://img.shields.io/npm/v/agentguard.svg)](https://www.npmjs.com/package/agentguard) [![tests](https://img.shields.io/badge/tests-31%20passing-brightgreen.svg)](https://github.com/chidev/agentguard) [![license](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE) [![sponsor](https://img.shields.io/badge/sponsor-kinglystudio.ai-purple.svg)](https://kinglystudio.ai) ``` _ _ __ _ __ _ ___ _ __ _| |_ | | ___ __ _ ___ ___ / _` |/ _` |/ _ \ '_ \ |_ _|____| |/ _ \/ _` / __|/ _ \ | (_| | (_| | __/ | | | |_|______| | __/ (_| \__ \ __/ \__,_|\__, |\___|_| |_| |_|\___|\__,_|___/\___| |___/ 🔒 Git hooks that FORCE validation. No escape. ``` --- ## The Cycle Here's the cycle you know too well: - You're agent makes changes - You forget to check if docs are updated - You forget to check if it's in alignment with the spec - Your CTO vibe codes on the weekend - His prototypes are trash - But now he wants you to be a 10x dev - You commit and push - Next developer does the same thing - You feel dumb - They feel dumb - Repeat Or worse, some obscure bug makes it to prod. **The cost isn't just emotional. It's time.** CI cycles take minutes. Deploy rollbacks take longer. Broken deploys on Friday afternoon? Career-limiting. Luck? That's what brought you here. --- ## The Lock/Lease Pattern Borrowed from distributed systems: ### First commit attempt: 1. Pre-commit hook detects no validation proof 2. Creates lock file 3. **BLOCKS THE COMMIT** ### Agent runs validation: 1. Runs `npx agentguard release --audit-proof` 2. Stamps lock file with proof + timestamp ### Second commit attempt: 1. Hook sees valid proof → allows commit 2. Archives proof to `.agentguard/proofs/` **No accidents. No bypassing.** ```mermaid flowchart TD A["$ git commit"] --> B["🔒 LOCK CREATED"] B --> C["❌ BLOCKED"] C --> D["$ npx agentguard release --audit-proof"] D --> E["⚡ RUNNERS EXECUTE"] E --> F{All pass?} F -->|Yes| G["✅ Proof stamped"] F -->|No| H["Fix errors"] --> D G --> I["$ git commit"] I --> J["✅ COMMITTED with trailers"] ``` --- ## Agentic Runners: AI Code Review on Every Commit Pipe your diff into **any LLM CLI**. The contract: `exit 0 = pass, exit 1 = fail`. ### Model Cascading | Phase | Model | Purpose | |-------|-------|---------| | Commit | Haiku/GPT-3.5 | Fast feedback, obvious bugs | | Push | Opus/GPT-4 | Deep security/correctness analysis | ```json { "runners": [ { "name": "build", "command": "npm run build", "on": "commit" }, { "name": "haiku", "command": "claude -p 'Quick check: {{diff}}'", "on": "commit" }, { "name": "opus", "command": "claude --model opus -p 'Deep review: {{diff}}'", "on": "push" } ] } ``` --- ## Why This Works - **LLMs catch different bugs than linters** — Type errors vs logic errors - **Fast models are cheap** — ~$0.001 per commit - **Can't bypass it** — The lock forces the step - **It learns your patterns** — Same model reviews consistently Most of this should be handled in the planning phases - but context window limits and the fact that models are goal driven mean they will compress goals. We want **constraints**. As deterministic as possible. This is about as close as you're going to get without a custom harness. The real benefit is that you're not just prompting "follow these standards" over and over with slash commands and hoping skills load. **This is deterministic.** It will -always- say "hey, did you do all this crap I asked you to do?" --- ## Agent-Native Design ``` You: "Add auth to the user service" Claude: *makes changes* You: "Commit this" Claude: *sees agentguard lock* Claude: *runs validation, releases lock* Claude: "Committed with validation proof. Build and ESLint passed." ``` The agent thinks "oh shit, forgot to build" **for you**. **Key insight:** When humans write code, we rely on discipline. When agents write code, we can program the discipline into the workflow. ### v3.2: Agent Proof Submission Agents can now submit their own validation proof: ```bash # Agent runs validators, then submits proof npx agentguard release --audit-proof='## Validation Report Runner: test Status: PASS Output: 23 tests passed in 4.2s Runner: haiku-review Status: PASS Output: No critical issues found Summary: All validations passed. Safe to commit.' ``` The hook output is a **meta-prompt** that teaches the agent exactly how to validate: ``` ╔═══════════════════════════════════════════════════════════════════╗ ║ 🔒 AGENT-LEASE: COMMIT BLOCKED ║ ╠═══════════════════════════════════════════════════════════════════╣ ║ Validate your changes and submit proof before committing. ║ ╚═══════════════════════════════════════════════════════════════════╝ CONFIGURED RUNNERS: test npm test [commit] haiku-review claude -p '...' [commit] CALLBACK FORMAT: npx agentguard release --audit-proof='<YOUR PROOF>' PROOF TEMPLATE: ## Validation Report Runner: <name> Status: PASS | FAIL Output: <output summary> ... Summary: <your assessment> ``` --- ## LLM Steering (v3.2) For structured LLM output, use steering markers: ```json { "name": "haiku-review", "command": "claude -p 'Review diff. Respond ONLY:\n<AGENTGUARD_START>\nVERDICT: PASS | FAIL\nCRITICAL: 0\nFINDINGS:\n- <issue>\nSUMMARY: <one line>\n<AGENTGUARD_END>\n\nDiff:\n{{diff}}'", "on": "commit", "llm": true } ``` agentguard parses the structured output and determines pass/fail from the VERDICT. --- ## Observability Every commit has proof of validation in `.agentguard/proofs/`: ```json { "timestamp": "2026-02-02T14:00:15Z", "phase": "commit", "proofMode": "agent", "runners": [ { "name": "test", "status": "PASS", "output": "23 tests passed" }, { "name": "haiku-review", "status": "PASS", "output": "No issues" } ], "summary": "All validations passed. Safe to commit." } ``` And git trailers on every commit: ``` commit a1b2c3d Add authentication module agentguard-proof: test:PASS haiku-review:PASS agentguard-agent-summary: All validations passed. Safe to commit. agentguard-llm-findings: Minor: consider null check on line 42 agentguard-report: commit-a1b2c3d.json ``` **Measure it. Optimize it.** If validation takes too long, devs will bypass it. --- ## Quick Start ```bash # Install npm install -g agentguard # Initialize hooks npx agentguard init # Try to commit (creates lock, blocks) git commit -m "my changes" # Run validation to release npx agentguard release --audit-proof # Commit again (succeeds with proof) git commit -m "my changes" ``` --- ## Using with Husky All of this can be done with husky. We just like a dedicated tool. Here's how to wire it up: ```bash # Install husky npm install -D husky npx husky init ``` **`.husky/pre-commit`:** ```bash #!/bin/bash # Delegate to agentguard exec < /dev/tty npx agentguard status --phase commit if [ $? -ne 0 ]; then echo "" echo "🔒 Run: npx agentguard release --audit-proof" echo "" exit 1 fi ``` **`.husky/pre-push`:** ```bash #!/bin/bash exec < /dev/tty npx agentguard status --phase push if [ $? -ne 0 ]; then echo "" echo "🔒 Run: npx agentguard release --audit-proof --phase push" echo "" exit 1 fi ``` Or just use `npx agentguard init` which installs the hooks directly. --- ## Claude Code + ClawdHub ### Install as a Claude Code Skill agentguard is available on [ClawdHub](https://clawhub.ai) — the public skill registry for Claude Code. Install it as a skill so your agent always knows how to use it: ```bash # Install agentguard skill npx clawdhub@latest install agentguard # Or search for validation skills npx clawdhub@latest search "validation" ``` Once installed, Claude Code agents automatically understand the lock/lease flow. The skill teaches the agent: - How to read the blocked meta-prompt - How to run validators and format proof - How to submit proof via `--audit-proof='...'` - When to bypass (and when not to) ### Publish Your Own Runners as Skills Created a custom LLM reviewer or validation pipeline? Publish it: ```bash npx clawdhub@latest publish ./my-validation-skill \ --slug my-validation-skill \ --version 1.0.0 \ --tags "validation,agentguard,code-review" ``` ### With Claude Code Hooks Claude Code CLI has a hooks system you can use for even tighter integration. Some nifty tricks coming in future posts, but agentguard is the highest leverage validation gate you can set up in a few minutes and start seeing your cortisol levels go down. ```json // .claude/settings.json { "hooks": { "PreToolUse": [ { "matcher": "Bash(git commit*)", "hooks": ["npx agentguard status --phase commit"] } ] } } ``` ### The Ralph Loop Shout out to [@GeoffreyHuntley](https://x.com/GeoffreyHuntley) - the Ralph Loop is a deterministic way to start getting control over your workflow. Watch his videos. A more advanced version: put the validation gate **outside** the Ralph loop, leave commit/pre-push as the final boss. My actual setup uses a custom harness in Leviathan (open source meta framework, coming soon) to handle all of these tips/tricks way more deterministically. You can create a **layered set of gates** all in the service of staying in the agentic flow, and steering more so than blocking. --- ## Configuration `.agentguard.json`: ```json { "runners": [ { "name": "build", "command": "npm run build", "on": "commit" }, { "name": "lint", "command": "npm run lint", "on": "commit" }, { "name": "haiku-review", "command": "claude -p 'Review: {{diff}}'", "on": "commit", "llm": true }, { "name": "test", "command": "npm test", "on": "push" }, { "name": "opus-review", "command": "claude --model opus -p 'Deep review: {{diff}}'", "on": "push", "llm": true } ], "lockDir": "auto", "projectName": "my-project" } ``` ### Template Variables | Variable | Value | |----------|-------| | `{{diff}}` | Staged changes (commit) or full diff (push) | | `{{files}}` | Changed file paths | | `{{project}}` | Project name | | `{{branch}}` | Current branch | | `{{hash}}` | Commit hash | ### Lock Storage | Value | Location | |-------|----------| | `"auto"` | `$XDG_RUNTIME_DIR/agentguard/` or `/tmp` | | `"local"` | `.agentguard/locks/` (project-local) | --- ## CLI Reference ```bash agentguard init # Install hooks agentguard release --audit-proof # Run runners, release (v2) agentguard release --audit-proof='<proof>' # Accept proof text (v3.2) agentguard release --audit-proof --phase push # Push phase agentguard status # Check lock state agentguard runners # List configured runners agentguard clear # Remove stale locks ``` --- ## Why Lock/Lease? Other tools run validation **during** the commit. agentguard is different: 1. **First commit creates a lock** and blocks 2. **You must explicitly run validation** to release 3. **Second commit checks for proof** and proceeds This **forces** the step. No accidents. No "I'll fix it later." No `--no-verify`. --- ## Get Started ```bash npm install -g agentguard npx agentguard init ``` - **GitHub:** [github.com/chidev/agentguard](https://github.com/chidev/agentguard) - **ClawdHub:** [clawhub.ai](https://clawhub.ai) - **Sponsor:** [kinglystudio.ai](https://kinglystudio.ai) --- ## License MIT

リポジトリの README を読み込んでいます…

このリポジトリの skills

ソフトウェア品質保証アナリスト・テスター

1 件の職業カテゴリ · 100% 分類済み

skill

職業分類

説明

更新

agentguard

ソフトウェア品質保証アナリスト・テスター

Standards-enforcing validation gates for git commits. Catches what linters miss.

原文の言語: 英語

2026-03-07

収集済み skill 1 件中 1 件を表示しています。