palladius/antigravity-ruby-sdk

最近记录的来源活动: 2026年10月8日 16:11

SkillsMP 收录数据更新: 2026年11月5日 19:57

已收集 skills: 8

GitHub 星标: 2
GitHub Forks: 0

用 AI 助手安装

复制这段 Prompt，发给你正在使用的 AI 助手。

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/palladius/antigravity-ruby-sdk.

仓库 README

仓库 README 来源：README.md

查看 README 源文件

展开完整 README

The Antigravity Ruby SDK is an unofficial, community project. It is **not** an official Google product. # 💎 Antigravity Ruby SDK <img src="docs/images/logo.jpg" alt="Ruby Antigravity Logo" width="350"/> <img src="images/tui-multiturn-demo.gif" alt="Antigravity Ruby SDK Multiturn TUI Demo" width="700"/> [![Gem Version](https://badge.fury.io/rb/antigravity-sdk.svg)](https://rubygems.org/gems/antigravity-sdk) [![GitHub](https://img.shields.io/github/license/palladius/antigravity-ruby-sdk)](LICENSE) An elegant, expressive Ruby SDK for building autonomous AI agents with **Google Antigravity**. > **Gem**: [`antigravity-sdk`](https://rubygems.org/gems/antigravity-sdk) on RubyGems > **Source**: [`palladius/antigravity-ruby-sdk`](https://github.com/palladius/antigravity-ruby-sdk) on GitHub Inspired by [RubyLLM](https://rubyllm.com) and the Ruby philosophy of developer happiness: configure agents, stream responses, load skills from GitHub, analyze workspaces, and attach safety guards -- all with minimal boilerplate. ```ruby agent = Antigravity::Agent.new( skills: ["./skills/code-quality-review"], workspace: "." ) agent.ask("Review this codebase for best practices") { |chunk| print chunk.content } ``` --- ## ⚡ Zero-Install Quickstart with `rv` No `gem install` needed — run directly with [rv](https://github.com/nicholasgasior/rv): ```bash # Simple chat rv run ruby examples/04_simple_llm_chat.rb # Workspace analysis (indexes your project, asks about it) rv run ruby examples/05_workspace_analysis.rb ~/git/my-app # Code quality review with skills rv run ruby examples/06_skill_security_audit.rb . # Load SRE skills from GitHub and draft a post-mortem rv run ruby examples/07_skill_sre_postmortem.rb . ``` Or with `just`: ```bash just rv-chat # Simple LLM chat just rv-workspace # Workspace analysis just rv-skill-audit # Code review with local + inline skills just rv-skill-sre-postmortem # SRE post-mortem from GitHub skills ``` --- ## 🇫🇷 *Richard* — Interactive Console Remember `rails console` and `irb`? We got inspired by those to create ***Richard*** (pronounced *Ri-shar*, a la francaise 🇫🇷) — a full-featured interactive REPL for your Antigravity agents. <img src="demo/richard_demo.gif" alt="Richard Console Demo" width="700"/> ```bash # Launch it just rv-console # Or directly rv run ruby examples/10_console.rb ``` ### 🎬 Full Demo (v10) <a href="demos/richard-console/demo.mp4"> <img src="demos/richard-console/demo.gif" alt="Richard Console v10 Demo" width="700"/> </a> *In this demo we show the Richard console's ability to interact with a normal LLM conversation, run shell commands with `! cmd`, evaluate Ruby with `r! expr`, drop into full IRB mode with `/irb`, and -- most importantly -- interact safely with the Policy DSL to dynamically add protections at runtime: `read_file(".env")` exposes secrets, then `policy.deny(:read_file, ".env")` locks it down, and the next `read_file` raises a red `PolicyDeniedError`. Policies can only tighten, never loosen.* **What you get:** | Feature | Syntax | Description | |---------|--------|-------------| | 💬 Chat | Just type | Multi-turn conversation with thinking/tools display | | ⚡ Shell | `! pwd` | Execute shell commands inline | | 💎 Ruby eval | `r! 2+2` | Single-line Ruby evaluation | | 💊 IRB mode | `/irb` | Persistent Ruby sub-REPL (*"So you chose the RED pill, Neo!"*) | | 🛡️ Policy | `/policy` | Show active safety policy (auto-allow, confirm, deny) | | 🤔 Thinking | `/think` or `Ctrl-O` | Toggle thinking expansion | **Matrix Mode** (`/irb`): Full Ruby introspection with smart setters that propagate to the agent: ```ruby 💎irb> config # full session state (api_key masked!) 💎irb> cd '~/git/my-project' # change workspace (validated) 💎irb> set_policy :turbo # switch safety policy 💎irb> @agent.class # => Antigravity::Agent 💎irb> exit # "Welcome... to the real world, Neo. 🕶️" ``` --- ## 📚 Agent Skills Skills are reusable instruction sets (SKILL.md files) that teach agents new capabilities. Load them from local folders, GitHub repos, or define them inline: ```ruby agent = Antigravity::Agent.new( # Mix local and remote skills in the constructor skills: [ "./skills/code-quality-review", # Local "https://github.com/gemini-cli-extensions/sre", # GitHub (auto-clones all 16 skills!) ] ) # Add a specific skill from a GitHub repo agent.add_skill("https://github.com/gemini-cli-extensions/sre", skill_name: "skills/postmortem-generator") # Define a skill inline (no file needed) agent.add_inline_skill( name: "emoji-formatter", description: "Formats output with emoji severity markers", instructions: "Use: CRITICAL: 🚨, HIGH: 🔴, MEDIUM: 🟡, LOW: 🔵, PASSED: ✅" ) # Discover skills in a folder Agent.list_skills("~/git/skillume/sre-extension/") # => ["/path/to/anomaly-detection", "/path/to/cloud-logging", ...] ``` ### 🌐 Installing Skills via CLI (`npx skills`) You can also install and manage skills using standard CLI tools: ```bash # Add the SDK skill to your workspace npx skills add palladius/antigravity-ruby-sdk/skills/using-antigravity-ruby-sdk ``` Here, for example, we are using an inline skill for custom severity emojis (`severity-emoji`) alongside a local code quality audit skill (`code-quality-review`): <img src="images/tui-security-audit-cropped.gif" alt="Code Quality & Security Audit Skill Demo" width="850"/> --- ## 📂 Workspace Analysis Point an agent at a directory — it indexes the files and uses built-in tools (`list_dir`, `view_file`, `grep_search`) to explore: ```ruby agent = Antigravity::Agent.new(workspace: "~/git/my-project") agent.connect! agent.ask("What tech stack does this project use?") { |c| print c.content } agent.close! ``` --- ## 🪵 Automagic Logging Dual-output logging out of the box: - `log/antigravity.jsonl` — structured telemetry (request/response sizes, tool calls) - `log/antigravity.log` — compact human-readable one-liners - Auto-attaches `Rails.logger` in Rails apps ```ruby agent = Antigravity::Agent.new # Logging just works! # => 🪵 Logging to log/antigravity.jsonl ``` --- ## 🛡️ Guards & Sidecars ```ruby agent = Antigravity::Agent.new do |a| a.system_instruction = "You are a helpful Ruby assistant." a.attach_sidecar(Antigravity::Sidecar::AuditLogger.new("log/audit.jsonl")) a.before_tool_call(&Antigravity::Guards::FileProtection.new) a.after_tool_call(&Antigravity::Guards::SecretMasker.new) end ``` --- ## 🔒 Declarative Policy DSL Control what your agent can and can't do with a beautiful, Rails-like DSL: ```ruby agent = Antigravity::Agent.new(policy: :default) # Use a preset agent = Antigravity::Agent.new(policy: :cautious) # Locked down for prod agent = Antigravity::Agent.new(policy: :turbo) # Wide open for dev agent = Antigravity::Agent.new(policy: :auto) # Picks from RAILS_ENV! ``` Or define a custom policy: ```ruby policy = Antigravity::Policy.define do deny_all allow :view_file allow :grep_search allow :run_command, when: cmd('echo', 'git status', 'bundle exec rspec') allow :write_to_file deny :write_to_file, when: path('.env', '*.key', '*.pem') deny :run_command, when: cmd('rm', 'git reset --hard') end agent = Antigravity::Agent.new(policy: policy) ``` ### ⚠️ Order does NOT matter! The DSL is **declarative** — like SQL, not like a script. Rules are resolved by **precedence**, not by insertion order. These two policies behave **identically**: ```ruby # Order A # Order B Policy.define do Policy.define do allow :run_command deny :run_command, when: cmd('rm') deny :run_command, allow :run_command when: cmd('rm') end end ``` Precedence (highest wins): 1. **Tool specificity**: `deny :run_command` beats `deny_all` 2. **Condition specificity**: `deny :run_command, when: cmd('rm')` beats `deny :run_command` 3. **Restrictiveness**: `deny` beats `confirm` beats `allow` ### 📋 Presets | Preset | Shell | Writes | `rm` | `git reset --hard` | Best for | |--------|-------|--------|------|---------------------|----------| | 🔒 `:cautious` | Safe only (`echo`, `pwd`) | Confirm | ❌ Deny | ❌ Deny | Production | | ⚖️ `:default` | Allow | Allow | ⚠️ Confirm | ⚠️ Confirm | Day-to-day dev | | 🚀 `:turbo` | Allow | Allow | Allow | ⚠️ Confirm | Rapid prototyping | | 🧪 `:test` | Allow | Allow | ⚠️ Confirm | ❌ Deny | CI / test suites | | 🔮 `:auto` | — | — | — | — | Reads `RAILS_ENV` | ### 📂 Sandbox directories `scratch/` and `out/` are **always writable**, even in `:cautious` / production. Use them as throwaway output dirs: ```ruby # In production — this works! agent.hooks.run_pre_tool(:write_to_file, path: 'scratch/debug.log', content: '...') # => { allowed: true } # But this is blocked: agent.hooks.run_pre_tool(:write_to_file, path: 'app.rb', content: '...') # => { allowed: false, reason: "Denied by policy" } ``` ### 🔮 Auto-mapping from RAILS_ENV `policy: :auto` reads `ANTIGRAVITY_ENV` → `RAILS_ENV` → `RACK_ENV`: | Environment | Preset | |-------------|--------| | `development` / `dev` | 🚀 `:turbo` | | `test` | 🧪 `:test` | | `staging` | ⚖️ `:default` | | `production` / `prod` | 🔒 `:cautious` | | *(unset)* | ⚖️ `:default` | See [`lib/antigravity/policy.rb`](lib/antigravity/policy.rb) and [`lib/antigravity/policy/constants.rb`](lib/antigravity/policy/constants.rb) for the full implementation. --- ## 🎈 *Jevity* — JEV-powered levity for your Ruby Antigravity agents **Jevity** (JEV + *levity*, the opposite of gravity) is a terminal assistant built on this SDK, since **v0.7.0**. You run it as **`rujev`** (Ru-by + Jev; `bin/jevity` works too). **JEV (System One)** makes fast decisions (~200–300ms each); **Gemini** does the talking: 1. 🚀 **Route**: JEV classifies your prompt as *simple* → fast model (`gemini-3.5-flash-lite`, ~0.8s) or *complex* → HIGH-thinking model (`gemini-3.8-flash-high`). 2. 🤖 **Answer**: Gemini streams 🤔 thinking and 🤖 answer live; if it needs a shell command, it proposes one. 3. 🛡️ **Guard**: JEV scores every command before it runs: ✅ ≥80% runs, 🚫 <40% blocked, ⚠️ in between asks you `[y/N]`. ![Jevity (rujev) demo](demos/jevity/jevity_demo.gif) > 🎬 Full-quality video: [`demos/jevity/jevity_demo.mp4`](demos/jevity/jevity_demo.mp4) — re-record it with `just demo-jevity` (needs `vhs` + `ttyd` + `ffmpeg`; runs in a `/tmp` sandbox with a fake `.env`). ### Quickstart ```bash export GEMINI_API_KEY=... # Gemini (answers) export JEV_API_KEY_RUJEV=... # JEV System One (routing + guardrail) bin/rujev # interactive REPL (multi-turn, /reset clears memory) bin/rujev 'what is in here?' # one-shot bin/rujev --folder ~/git/myrepo 'git status' # run in another folder bin/rujev -M gemini-3.8-flash-high '...' # force a model, skip JEV routing bin/rujev guard 'rm -rf /' # safety verdict only, never executes bin/rujev --yolo '...' # auto-approve the ⚠️ "unsure" zone (🚫 still blocks) ``` ### ⚙️ Configuration No code changes needed: edit the **config block at the top of the [`justfile`](justfile)**, or set env vars / `.env` (documented in [`.env.dist`](.env.dist)). Precedence: `--model` > ENV / `.env` / justfile > [`config/jevity.yml`](config/jevity.yml) > code default. | Variable | Default | What it does | |---|---|---| | `JEV_API_KEY_RUJEV` | *(required)* | JEV key. Looked up in ENV, then `$GIC/.env`, then `./.env` | | `GEMINI_API_KEY` | *(required)* | Gemini key (sent as a header, never in the URL) | | `JEV_FAST_MODEL` | `gemini-3.5-flash-lite` | Model for *simple* prompts | | `JEV_SMART_MODEL` | `gemini-3.8-flash-high` | Model for *complex* prompts (`-low/-medium/-high` suffix = thinking level) | | `JEV_FALLBACK_MODEL` | `gemini-3.1-flash-lite,gemini-2.5-flash-lite` | Failover chain; a failed model is skipped for the rest of the session | | `JEV_FIRST_BYTE_TIMEOUT` | `6` | Seconds to wait for the first streamed token before failing over | | `JEV_GEMINI_TIMEOUT` | `10` | Max seconds between streamed chunks | | `JEV_THINKING_TIMEOUT` | `30` | Same, for medium/high thinking models (thinking chunks count as activity) | | `JEV_ALLOW_THRESHOLD` / `JEV_DENY_THRESHOLD` | `0.80` / `0.40` | Guardrail zones: auto-run / auto-block | Not sure which Gemini model is fastest today? `just benchmark-gemini-testspeed` measures 1st token / total per model and tells you what to put in `JEV_FAST_MODEL`. Network trouble shows up as a red 🔌 line; if JEV itself is unreachable, nothing is executed (fails closed). --- ## 📊 Feature Parity with [Python SDK](https://github.com/google-antigravity/antigravity-sdk-python) > Full matrix: [`docs/FEATURE_PARITY.md`](docs/FEATURE_PARITY.md) | Epic: [GHI #20](https://github.com/palladius/antigravity-ruby-sdk/issues/20) | Feature | Status | Notes | |---------|--------|-------| | Agent lifecycle (`connect!`, `close!`, block) | ✅ | + auto-connect on first `ask` | | Streaming responses | ✅ | Token-by-token via block | | Custom tools (declarative + dynamic) | ✅ | `Tool` DSL + `Tool::Dynamic` | | Agent Skills (local + GitHub + inline) | ✅ | Ruby-only: GitHub auto-clone, inline skills | | Workspace analysis | ✅ | Built-in file tools | | Guards (FileProtection, SecretMasker) | ✅ | Ruby-only feature | | Sidecars (AuditLogger, VulnScanner) | ✅ | Ruby-only feature | | Hooks (pre/post prompt, tool) | ✅ | + generic event system | | Logging (JSONL + .log) | ✅ | Auto-attach | | Declarative Policies | ✅ | [#21](https://github.com/palladius/antigravity-ruby-sdk/issues/21) — DSL, 5 presets, sandbox dirs | | MCP Servers (Stdio + HTTP) | ❌ | Planned P0 | | Multimodal Input (Image, Audio, Doc) | ❌ | Planned P1 | | Structured Output (JSON Schema) | ❌ | Planned P1 | | Stateful ToolContext | ❌ | Planned P1 | | Session Persistence (save/resume) | ❌ | Planned P1 | | Multi-Agent / Subagents | ❌ | Planned P2 | | Triggers (background tasks) | ❌ | Planned P2 | | Vertex AI backend | ❌ | Planned P1 | | Response Cancellation | ❌ | Planned P1 | | Budget Limits | ❌ | Planned P1 | | OpenTelemetry | ❌ | Planned P2 | | LiteRT / Ollama backends | ❌ | Planned P3 | **Overall: ~40% parity** | 10 Ruby-only features | [Convergence plan](docs/FEATURE_PARITY.md#prioritized-gap-closure-plan) ## 🧪 Development ```bash just test # 76 unit specs (fast, no harness needed) just integration # Integration tests (requires GEMINI_API_KEY) just rv-examples # Run all rv examples ``` ## 🤖 Telegram Integration Chat with your Antigravity agent on Telegram — text and voice messages with automatic transcription! 1. Create a bot with [@BotFather](https://t.me/BotFather) 2. Add these to your `.env`: ```bash TELEGRAM_BOT_TOKEN=your-token-from-botfather TELEGRAM_CHAT_ID=your-chat-id # Optional: enables startup greeting TELEGRAM_SKILLS=./skills/my-skill # Optional: comma-separated skill paths/URLs ``` 3. Run: ```bash just rv-skill-telegram ``` Commands: `/start` `/skills` `/stop` — supports voice messages with 🇮🇹🇬🇧🇪🇸 language detection. See `.env.dist` for all available options. ![Telegram integration](image.png) ![Telegram voice transcription](image-1.png) ## 📦 Publishing to gem.coop This gem can be published to the co-op RubyGems alternative [gem.coop](https://gem.coop/). To push a release, obtain an API key from `gem.coop` and pass it to the push/release commands: ```bash # Push a specific gem file GEM_HOST_API_KEY=your_api_key_here gem push antigravity-sdk-VERSION.gem --host https://gem.coop/@palladius # Or use rake release to tag and push automatically GEM_HOST_API_KEY=your_api_key_here RUBYGEMS_HOST=https://gem.coop/@palladius rake release ``` ## 🔗 Related Projects | Project | Language | Link | |---------|----------|------| | **Antigravity Python SDK** (official) | Python | [google-antigravity/antigravity-sdk-python](https://github.com/google-antigravity/antigravity-sdk-python) | | **Antigravity Java SDK** (unofficial) | Java | [glaforge/antigravity-java-sdk](https://github.com/glaforge/antigravity-java-sdk) | | **Antigravity Ruby SDK** (this repo) | Ruby | [palladius/antigravity-ruby-sdk](https://github.com/palladius/antigravity-ruby-sdk) | | **antigravity-sdk gem** | RubyGems | [rubygems.org/gems/antigravity-sdk](https://rubygems.org/gems/antigravity-sdk) | | **antigravity-sdk gem (co-op)** | gem.coop | [gem.coop/@palladius/antigravity-sdk](https://gem.coop/@palladius/antigravity-sdk) | --- ## 📄 License Apache 2.0 - see `LICENSE` for details. *The Antigravity Ruby SDK is an unofficial, community project. It is **not** an official Google product.*

正在加载仓库 README…