← 返回 Skills 市场

claw-code

Name: claw-code
Author: dae-xiaoe

作者 dae-xiaoe · GitHub ↗ · v1.0.0 · MIT-0

cross-platform ⚠ suspicious

142

总下载

当前安装

版本数

在 OpenClaw 中安装

/install claw-code

功能描述

TypeScript runtime port of Claude Code's AI agent harness for running commands, routing prompts, managing sessions, and auditing tool/command registries.

使用说明 (SKILL.md)

claw-code — AI Agent Harness Runtime

TypeScript port of Claude Code's harness system, compiled and ready to run.

Quick Start

cd C:\Users\dae\.openclaw\workspace\skills\claw-code\scripts
node dist/main.js \x3Ccommand> [options]

Or build from source:

cd C:\Users\dae\.openclaw\workspace\skills\claw-code\scripts
npm install
npm run build

Available Commands

Info & Audit

node dist/main.js summary                          # Render workspace summary
node dist/main.js manifest                         # Print workspace manifest
node dist/main.js parity-audit                     # Compare against archive
node dist/main.js setup-report                     # Startup/prefetch report
node dist/main.js command-graph                    # Show command graph
node dist/main.js tool-pool                        # Show tool pool
node dist/main.js bootstrap-graph                  # Show bootstrap stages

Query Registries

node dist/main.js subsystems --limit 16             # List Python modules
node dist/main.js commands --limit 20             # List mirrored commands
node dist/main.js commands --query \x3Ckeyword>      # Search commands
node dist/main.js tools --limit 20                 # List mirrored tools
node dist/main.js tools --query \x3Ckeyword>          # Search tools
node dist/main.js tools --simple-mode             # Core tools only
node dist/main.js tools --no-mcp                   # Exclude MCP tools
node dist/main.js show-command \x3Cname>               # Show one command
node dist/main.js show-tool \x3Cname>                 # Show one tool

Runtime Execution

node dist/main.js route "\x3Cprompt>" --limit 5      # Route prompt to commands/tools
node dist/main.js bootstrap "\x3Cprompt>" --limit 5   # Bootstrap a session
node dist/main.js turn-loop "\x3Cprompt>" --max-turns 3   # Run turn loop
node dist/main.js flush-transcript "\x3Cprompt>"      # Persist session
node dist/main.js load-session \x3Csession_id>        # Load saved session

Remote Modes

node dist/main.js remote-mode \x3Ctarget>
node dist/main.js ssh-mode \x3Ctarget>
node dist/main.js teleport-mode \x3Ctarget>
node dist/main.js direct-connect-mode \x3Ctarget>
node dist/main.js deep-link-mode \x3Ctarget>

Execution

node dist/main.js exec-command \x3Cname> "\x3Cprompt>"   # Execute mirrored command
node dist/main.js exec-tool \x3Cname> \x3Cpayload>       # Execute mirrored tool

Project Structure

claw-code/
├── SKILL.md              # This file
└── scripts/
    ├── dist/             # Compiled JavaScript (ready to run)
    │   └── main.js       # CLI entry point
    ├── src/              # TypeScript source
    │   └── reference_data/  # Snapshot data
    ├── package.json
    └── tsconfig.json

Architecture

Runtime Core (runtime.ts) — PortRuntime class: route_prompt, bootstrap_session, run_turn_loop
Tool/Command Registries — loaded from reference_data/*.json snapshots
Session Management — session_store.ts, history.ts
Query Engine — query_engine.ts with streaming support
Graphs — bootstrap_graph.ts, command_graph.ts
Remote Modes — remote_runtime.ts, direct_modes.ts

安全使用建议

This skill appears to be what it claims (a TypeScript/Node CLI port of an agent harness) and requests no secrets, but take these precautions before running it: - Ensure you have Node installed; SKILL.md assumes 'node' though the metadata didn't list it. - Inspect session_store.js and remote_runtime.js (both present in the bundle but truncated in the listing). They are the most likely places for filesystem or network access. Confirm they only read/write local files under the skill workspace and don't attempt to exfiltrate data or open network connections you don't want. - Because the Quick Start uses a user-specific Windows path (C:\Users\dae\...), adjust to your environment and run from a sandbox or non-critical environment first. - If you want stronger assurance, run node dist/main.js with harmless commands like 'summary' or 'command-graph' and observe behavior (and what files get created) before using any 'exec-tool', 'remote-mode', or 'ssh-mode' commands. - If you do not trust the unknown owner or need stricter isolation, run the skill inside a VM or container or decline installation.

功能分析

Type: OpenClaw Skill Name: claw-code Version: 1.0.0 The skill bundle contains several high-risk indicators, most notably functions named 'startKeychainPrefetch' and 'startMdmRawRead' within 'prefetch.js' which are called during the mandatory setup phase in 'setup.js'. While currently implemented as no-op stubs, the inclusion of logic designed to target system keychains and MDM (Mobile Device Management) data is highly irregular for an AI agent harness. Additionally, 'remote_runtime.js' and 'direct_modes.js' define stubs for 'ssh-mode', 'teleport-mode', and 'direct-connect-mode', indicating an architectural intent to facilitate remote access and control.

能力标签

cryptorequires-oauth-token

能力评估

ℹ Purpose & Capability

The code and SKILL.md match the description: a TypeScript CLI/runtime for an agent harness with command/tool registries, query engine, session store, and parity/audit stubs. Minor mismatch: SKILL.md expects you to run 'node dist/main.js' but the skill metadata lists no required binaries — 'node' is implicitly required but not declared.

ℹ Instruction Scope

Runtime instructions are explicit CLI commands that operate on the included compiled code (dist/*.js). The SKILL.md tells the user to cd into a user-specific path (C:\Users\dae\.openclaw\...) and run node; it also instructs reviewing source files. The visible modules are mainly stubs/mirrors and do not access external endpoints in the shown files, but some modules that were truncated (notably remote_runtime.js and session_store.js) could perform filesystem I/O or networking — these should be inspected before execution.

✓ Install Mechanism

No install spec is provided (instruction-only with bundled code). Nothing is being downloaded at install time; all code is present in the skill bundle, which reduces install-time risk. The skill will require a Node runtime to execute its dist/*.js files but does not declare it.

✓ Credentials

The skill declares no required environment variables or credentials and the visible code does not reference secrets or external API keys. That is proportionate for a local CLI-style harness. Note: session persistence and keychain-prefetch functions are present as stubs; verify session_store.js and any 'keychain' related code to confirm they are indeed no-ops and not reading secrets.

ℹ Persistence & Privilege

The skill can persist and load sessions (saveSession/loadSession are imported). This is expected for an agent harness; the skill is not set to always:true. Confirm where sessions are written (paths) before running if you care about local disk writes. Autonomous invocation (disable-model-invocation=false) is normal and not by itself a risk.

如何使用

确保已安装 OpenClaw（本地或 Docker 部署）
在对话框中输入安装命令：/install claw-code
安装完成后，直接呼叫该 Skill 的名称或使用 /claw-code 触发
根据 Skill 的参数说明提供必要输入，即可获得结构化输出

版本历史

v1.0.0

Initial release: AI Agent Harness Runtime, TypeScript port of Claude Code's harness system

元数据

Slug claw-code

版本 1.0.0

许可证 MIT-0

累计安装 1

当前安装数 1

历史版本数 1

常见问题

claw-code 是什么？

TypeScript runtime port of Claude Code's AI agent harness for running commands, routing prompts, managing sessions, and auditing tool/command registries. 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 142 次。

如何安装 claw-code？

在 OpenClaw 或 Claude Code 对话框中运行命令「/install claw-code」即可一键安装，无需额外配置。

claw-code 是免费的吗？

是的，claw-code 完全免费，采用 MIT-0 许可证，可自由下载、安装和使用。

claw-code 支持哪些平台？

claw-code 跨平台运行，可在任意部署了 OpenClaw / Claude Code 的环境中使用（cross-platform）。

谁开发了 claw-code？

由 dae-xiaoe（@dae-xiaoe）开发并维护，当前版本 v1.0.0。