|
| 1 | +# CLI Registry |
| 2 | + |
| 3 | +The CLI Registry is AgentOS's auto-discovery system for installed command-line tools. It scans the user's PATH for known binaries, detects versions, and exposes results to providers, extensions, and the capability discovery engine. |
| 4 | + |
| 5 | +## Overview |
| 6 | + |
| 7 | +AgentOS ships with a JSON-based registry of 54 CLI descriptors across 8 categories. At startup (or on demand), the `CLIRegistry` runs `which` + `--version` for each registered binary in parallel, producing a scan result that tells the runtime exactly what's available on the host machine. |
| 8 | + |
| 9 | +This powers: |
| 10 | +- **LLM provider auto-detection** -- `ClaudeCodeCLIBridge` and `GeminiCLIBridge` check if their binary is installed before attempting subprocess calls. |
| 11 | +- **`wunderland doctor`** -- health-check output includes detected CLIs. |
| 12 | +- **Capability discovery** -- the discovery engine indexes installed tools as capabilities agents can reference. |
| 13 | +- **cli-executor extension** -- `shell_execute` relies on the host having the right binaries. |
| 14 | + |
| 15 | +## Registry Categories |
| 16 | + |
| 17 | +The 54 bundled descriptors live in `src/sandbox/subprocess/registry/` as plain JSON files: |
| 18 | + |
| 19 | +| File | Category | Count | Examples | |
| 20 | +|------|----------|-------|----------| |
| 21 | +| `llm.json` | llm | 5 | claude, gemini, ollama, lmstudio, aichat | |
| 22 | +| `devtools.json` | devtools | 10 | git, gh, docker, docker-compose, kubectl, terraform, make, jq, yq, tmux | |
| 23 | +| `runtimes.json` | runtime | 8 | node, python3, deno, bun, ruby, go, rustc, java | |
| 24 | +| `package-managers.json` | package-manager | 7 | npm, pnpm, yarn, pip, uv, brew, cargo | |
| 25 | +| `cloud.json` | cloud | 9 | gcloud, aws, az, flyctl, vercel, netlify, railway, heroku, wrangler | |
| 26 | +| `databases.json` | database | 5 | psql, mysql, sqlite3, redis-cli, mongosh | |
| 27 | +| `media.json` | media | 5 | ffmpeg, ffprobe, magick, sox, yt-dlp | |
| 28 | +| `networking.json` | networking | 5 | curl, wget, ssh, rsync, scp | |
| 29 | + |
| 30 | +## CLIDescriptor Shape |
| 31 | + |
| 32 | +Each JSON entry conforms to the `CLIDescriptor` interface: |
| 33 | + |
| 34 | +```typescript |
| 35 | +interface CLIDescriptor { |
| 36 | + /** Binary name on PATH (e.g. 'claude', 'docker', 'ffmpeg'). */ |
| 37 | + binaryName: string; |
| 38 | + /** Human-readable display name. */ |
| 39 | + displayName: string; |
| 40 | + /** What this CLI does. */ |
| 41 | + description: string; |
| 42 | + /** Category for grouping (e.g. 'llm', 'media', 'devtools'). */ |
| 43 | + category: string; |
| 44 | + /** How to install if missing. */ |
| 45 | + installGuidance: string; |
| 46 | + /** Version flag override if not --version. */ |
| 47 | + versionFlag?: string; |
| 48 | + /** Regex to parse version from output (default: /(\d+\.\d+\.\d+)/). */ |
| 49 | + versionPattern?: RegExp; |
| 50 | +} |
| 51 | +``` |
| 52 | + |
| 53 | +Example from `cloud.json`: |
| 54 | + |
| 55 | +```json |
| 56 | +{ |
| 57 | + "binaryName": "gcloud", |
| 58 | + "displayName": "Google Cloud SDK", |
| 59 | + "description": "Google Cloud resource management", |
| 60 | + "category": "cloud", |
| 61 | + "installGuidance": "https://cloud.google.com/sdk/docs/install", |
| 62 | + "versionFlag": "--version" |
| 63 | +} |
| 64 | +``` |
| 65 | + |
| 66 | +## CLIRegistry API |
| 67 | + |
| 68 | +```typescript |
| 69 | +import { CLIRegistry, WELL_KNOWN_CLIS } from '@framers/agentos/sandbox/subprocess'; |
| 70 | +``` |
| 71 | + |
| 72 | +### Constructor |
| 73 | + |
| 74 | +```typescript |
| 75 | +const registry = new CLIRegistry(); // loads bundled JSON descriptors |
| 76 | +const empty = new CLIRegistry(false); // starts empty (no defaults) |
| 77 | +``` |
| 78 | + |
| 79 | +### Methods |
| 80 | + |
| 81 | +| Method | Returns | Description | |
| 82 | +|--------|---------|-------------| |
| 83 | +| `register(descriptor)` | `void` | Register a single CLI descriptor. Overwrites existing entry for same `binaryName`. | |
| 84 | +| `registerAll(descriptors)` | `void` | Register multiple descriptors at once. | |
| 85 | +| `unregister(binaryName)` | `boolean` | Remove a descriptor by binary name. | |
| 86 | +| `scan()` | `Promise<CLIScanResult[]>` | Scan PATH for all registered CLIs (parallel `which` + `--version`). | |
| 87 | +| `check(binaryName)` | `Promise<CLIScanResult>` | Check a single binary by name. | |
| 88 | +| `list()` | `CLIDescriptor[]` | Get all registered descriptors (installed status unknown). | |
| 89 | +| `installed()` | `Promise<CLIScanResult[]>` | Get only CLIs that are installed. | |
| 90 | +| `byCategory(category)` | `Promise<CLIScanResult[]>` | Get CLIs by category (scans first). | |
| 91 | +| `categories()` | `string[]` | Get all unique categories. | |
| 92 | +| `has(binaryName)` | `boolean` | Check if a binary is registered (not whether installed). | |
| 93 | +| `get(binaryName)` | `CLIDescriptor \| undefined` | Get a descriptor by binary name. | |
| 94 | +| `size` | `number` | Total number of registered descriptors. | |
| 95 | + |
| 96 | +### CLIScanResult |
| 97 | + |
| 98 | +The result from `scan()` or `check()` extends `CLIDescriptor`: |
| 99 | + |
| 100 | +```typescript |
| 101 | +interface CLIScanResult extends CLIDescriptor { |
| 102 | + installed: boolean; // whether the binary was found on PATH |
| 103 | + binaryPath?: string; // resolved absolute path (e.g. /usr/local/bin/node) |
| 104 | + version?: string; // parsed version string (e.g. "22.4.0") |
| 105 | +} |
| 106 | +``` |
| 107 | + |
| 108 | +## Adding Custom CLIs |
| 109 | + |
| 110 | +### Option 1: Edit JSON (permanent) |
| 111 | + |
| 112 | +Add a new entry to an existing category file, or create a new `*.json` file in `src/sandbox/subprocess/registry/`: |
| 113 | + |
| 114 | +```json |
| 115 | +[ |
| 116 | + { |
| 117 | + "binaryName": "my-tool", |
| 118 | + "displayName": "My Tool", |
| 119 | + "description": "Internal deployment CLI", |
| 120 | + "category": "devtools", |
| 121 | + "installGuidance": "brew install my-tool" |
| 122 | + } |
| 123 | +] |
| 124 | +``` |
| 125 | + |
| 126 | +### Option 2: Register at runtime (dynamic) |
| 127 | + |
| 128 | +```typescript |
| 129 | +const registry = new CLIRegistry(); |
| 130 | + |
| 131 | +registry.register({ |
| 132 | + binaryName: 'my-tool', |
| 133 | + displayName: 'My Tool', |
| 134 | + description: 'Internal deployment CLI', |
| 135 | + category: 'devtools', |
| 136 | + installGuidance: 'brew install my-tool', |
| 137 | +}); |
| 138 | + |
| 139 | +const result = await registry.check('my-tool'); |
| 140 | +if (result.installed) { |
| 141 | + console.log(`my-tool v${result.version} at ${result.binaryPath}`); |
| 142 | +} |
| 143 | +``` |
| 144 | + |
| 145 | +### Option 3: Full scan with custom CLIs |
| 146 | + |
| 147 | +```typescript |
| 148 | +const registry = new CLIRegistry(); |
| 149 | + |
| 150 | +// Add several custom CLIs |
| 151 | +registry.registerAll([ |
| 152 | + { binaryName: 'tsc', displayName: 'TypeScript', description: 'TS compiler', category: 'devtools', installGuidance: 'npm i -g typescript' }, |
| 153 | + { binaryName: 'eslint', displayName: 'ESLint', description: 'JS linter', category: 'devtools', installGuidance: 'npm i -g eslint' }, |
| 154 | +]); |
| 155 | + |
| 156 | +// Scan everything (bundled + custom) |
| 157 | +const results = await registry.scan(); |
| 158 | +for (const r of results) { |
| 159 | + const status = r.installed ? `v${r.version}` : 'not installed'; |
| 160 | + console.log(`${r.displayName.padEnd(24)} ${status}`); |
| 161 | +} |
| 162 | + |
| 163 | +// Filter by category |
| 164 | +const llmClis = await registry.byCategory('llm'); |
| 165 | +console.log(`LLM CLIs found: ${llmClis.filter(c => c.installed).length}/${llmClis.length}`); |
| 166 | +``` |
| 167 | + |
| 168 | +## Integration with CLISubprocessBridge |
| 169 | + |
| 170 | +The `CLISubprocessBridge` is an abstract base class for managing CLI subprocesses. It handles spawning, stdin piping, NDJSON stream parsing, timeouts, and abort signals. Subclasses implement CLI-specific flag assembly and error classification. |
| 171 | + |
| 172 | +Two production bridges extend it: |
| 173 | + |
| 174 | +| Bridge | Binary | Purpose | |
| 175 | +|--------|--------|---------| |
| 176 | +| `ClaudeCodeCLIBridge` | `claude` | Anthropic Claude via Max subscription (no API key needed) | |
| 177 | +| `GeminiCLIBridge` | `gemini` | Google Gemini via Google account login (no API key needed) | |
| 178 | + |
| 179 | +Both bridges use `checkBinaryInstalled()` (which internally runs `which` + `--version`) before attempting LLM calls, and fall back gracefully when the binary is missing. |
| 180 | + |
| 181 | +### Creating a custom bridge |
| 182 | + |
| 183 | +```typescript |
| 184 | +import { CLISubprocessBridge } from '@framers/agentos/sandbox/subprocess'; |
| 185 | +import { CLISubprocessError, CLI_ERROR } from '@framers/agentos/sandbox/subprocess'; |
| 186 | + |
| 187 | +class MyToolBridge extends CLISubprocessBridge { |
| 188 | + protected readonly binaryName = 'mytool'; |
| 189 | + |
| 190 | + protected buildArgs(options, format) { |
| 191 | + return ['--prompt', options.prompt, '--format', format]; |
| 192 | + } |
| 193 | + |
| 194 | + protected classifyError(error) { |
| 195 | + if (error.code === 'ENOENT') { |
| 196 | + return new CLISubprocessError( |
| 197 | + 'mytool not found', |
| 198 | + CLI_ERROR.BINARY_NOT_FOUND, |
| 199 | + 'mytool', |
| 200 | + 'Install: brew install mytool', |
| 201 | + false, |
| 202 | + ); |
| 203 | + } |
| 204 | + return new CLISubprocessError( |
| 205 | + error.message, |
| 206 | + CLI_ERROR.CRASHED, |
| 207 | + 'mytool', |
| 208 | + 'Check mytool logs', |
| 209 | + true, |
| 210 | + ); |
| 211 | + } |
| 212 | + |
| 213 | + protected parseStreamEvent(raw) { |
| 214 | + if (raw.text) return { type: 'text_delta', text: raw.text }; |
| 215 | + if (raw.done) return { type: 'result', result: raw.output }; |
| 216 | + return null; |
| 217 | + } |
| 218 | +} |
| 219 | +``` |
| 220 | + |
| 221 | +## Integration with cli-executor Extension |
| 222 | + |
| 223 | +The `cli-executor` extension pack (`@framers/agentos-ext-cli-executor`) provides tools that let agents execute arbitrary shell commands on the host. While it does not import `CLIRegistry` directly, the two systems are complementary: |
| 224 | + |
| 225 | +- **CLIRegistry** answers "what binaries exist?" -- discovery and detection. |
| 226 | +- **cli-executor** answers "can the agent run this command?" -- execution with security guardrails. |
| 227 | + |
| 228 | +When the wunderland runtime loads the cli-executor extension, it configures filesystem roots, security checks, and the `dangerouslySkipSecurityChecks` flag based on the active security tier. See the [Wunderland CLI Tools doc](../../../wunderland/docs/features/CLI_TOOLS.md) for details. |
| 229 | + |
| 230 | +## Security Considerations |
| 231 | + |
| 232 | +The CLI Registry itself is read-only and does not execute commands beyond `which` and `--version`. However, downstream consumers should respect the active security tier: |
| 233 | + |
| 234 | +| Security Tier | CLI Execution | File Writes | External APIs | |
| 235 | +|--------------|---------------|-------------|---------------| |
| 236 | +| `dangerous` | Allowed | Allowed | Allowed | |
| 237 | +| `permissive` | Allowed | Allowed | Allowed | |
| 238 | +| `balanced` | Allowed | Blocked | Allowed | |
| 239 | +| `strict` | Blocked | Blocked | Allowed | |
| 240 | +| `paranoid` | Blocked | Blocked | Blocked | |
| 241 | + |
| 242 | +The `balanced` tier is the recommended default. It permits CLI execution but blocks file writes unless the agent requests folder access through the HITL approval flow. |
| 243 | + |
| 244 | +## Error Handling |
| 245 | + |
| 246 | +The `CLISubprocessError` class provides structured errors with actionable guidance: |
| 247 | + |
| 248 | +```typescript |
| 249 | +import { CLISubprocessError, CLI_ERROR } from '@framers/agentos/sandbox/subprocess'; |
| 250 | + |
| 251 | +// Common error codes: |
| 252 | +CLI_ERROR.BINARY_NOT_FOUND // Binary not found on PATH |
| 253 | +CLI_ERROR.NOT_AUTHENTICATED // Binary installed but not logged in |
| 254 | +CLI_ERROR.VERSION_OUTDATED // Version too old for required features |
| 255 | +CLI_ERROR.SPAWN_FAILED // Process failed to start |
| 256 | +CLI_ERROR.TIMEOUT // Process exceeded timeout |
| 257 | +CLI_ERROR.CRASHED // Non-zero exit code |
| 258 | +CLI_ERROR.RATE_LIMITED // Rate limit / quota exceeded |
| 259 | +CLI_ERROR.PERMISSION_DENIED // EACCES |
| 260 | +CLI_ERROR.CONTEXT_TOO_LONG // Input too long for the CLI |
| 261 | +``` |
| 262 | + |
| 263 | +Each error carries a `guidance` string with human-readable fix instructions and a `recoverable` flag indicating whether retry/fallback is appropriate. |
| 264 | + |
| 265 | +## Exports |
| 266 | + |
| 267 | +Everything is exported from the barrel at `@framers/agentos/sandbox/subprocess`: |
| 268 | + |
| 269 | +```typescript |
| 270 | +export { CLISubprocessBridge } from './CLISubprocessBridge'; |
| 271 | +export { CLIRegistry, WELL_KNOWN_CLIS } from './CLIRegistry'; |
| 272 | +export { CLISubprocessError, CLI_ERROR } from './errors'; |
| 273 | +export type { |
| 274 | + BridgeOptions, |
| 275 | + BridgeResult, |
| 276 | + StreamEvent, |
| 277 | + OutputFormat, |
| 278 | + InstallCheckResult, |
| 279 | + CLIDescriptor, |
| 280 | + CLIScanResult, |
| 281 | +} from './types'; |
| 282 | +``` |
0 commit comments