菜单
Initialize a new agentic-usability benchmark pipeline project. Use when setting up a new SDK benchmark, creating a config.json, or starting a new evaluation project.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/PSPDFKit-labs/agentic-usability --skill init命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
Launch an interactive shell inside a microsandbox for debugging. Supports bare mode, executor setup, or judge setup with optional test case scaffolding.
Run the full evaluation pipeline (execute, judge, report) for an SDK usability benchmark. Use when running a complete benchmark end-to-end, resuming an interrupted pipeline, or checking pipeline status.
Execute benchmark test cases in sandboxed environments with AI agents. Spins up microsandbox containers for each test case and extracts solutions.
正在显示 SKILL.md
| name | init |
| description | Initialize a new agentic-usability benchmark pipeline project. Use when setting up a new SDK benchmark, creating a config.json, or starting a new evaluation project. |
| argument-hint | [project-directory] |
| disable-model-invocation | true |
| allowed-tools | Bash(agentic-usability *) Write Read Glob |
Set up a new agentic-usability benchmark pipeline in the given project directory.
echo "Project directory: $ARGUMENTS"
You have two approaches:
Run agentic-usability init -p $ARGUMENTS for a step-by-step interactive setup.
If the user has described their SDK, create config.json directly. This is faster and allows you to tailor the config to their exact setup.
After init, the project should have:
<project>/
config.json # Configuration (you create this)
suite.json # Test suite (created by generate)
results/ # Run results (created by eval/execute)
cache/repos/ # Git repo cache (created automatically)
{
"privateInfo": [],
"publicInfo": [],
"agents": {},
"targets": [],
"sandbox": {}
}
privateInfo (required, non-empty array)SDK source code and internal docs. Visible to generator and judge, never to executor. Each entry is a SourceConfig with a type discriminator:
Local source — filesystem path:
{ "type": "local", "path": "./src", "subpath": "packages/core", "additionalContext": "Focus on the Builder API" }
Fields: path (required), subpath (optional), additionalContext (optional)
Git source — clone a repository:
{ "type": "git", "url": "https://github.com/org/sdk.git", "branch": "main", "subpath": "src", "sparse": ["src/api"], "additionalContext": "..." }
Fields: url (required), branch, subpath, sparse (sparse checkout paths), additionalContext (all optional)
URL source — fetch documentation:
{ "type": "url", "url": "https://internal-docs.example.com/api-ref", "additionalContext": "..." }
Fields: url (required), additionalContext (optional)
Package source — metadata about the SDK package:
{ "type": "package", "name": "@example/sdk", "installCommand": "npm install @example/sdk", "language": "typescript", "additionalContext": "..." }
Fields: name (required), installCommand, language, additionalContext (all optional)
publicInfo (optional array)Public docs and package info visible to both executor and judge. Same SourceConfig types as above. Typically includes:
package source so executors know what to installurl source for public documentationagents (optional object)| Role | Type | Runs in sandbox? | Secret required? |
|---|---|---|---|
generator | AgentConfig | No (host) | No |
executor | SandboxAgentConfig | Yes | Yes |
judge | SandboxAgentConfig | Yes | Yes |
insights | AgentConfig | No (host) | No |
AgentConfig fields (generator, insights):
command (required): "claude", "codex", "gemini", or custom CLI namesystemPrompt (optional): supports {{packageName}} and {{docsUrl}} placeholdersSandboxAgentConfig — extends AgentConfig with required secret:
{
"command": "claude",
"secret": { "value": "$ANTHROPIC_API_KEY" }
}
AgentSecretConfig fields:
value (required): raw API key or "$ENV_VAR" referenceenvVar: env var name for key inside sandbox — auto-detected for known agentsbaseUrl: API base URL — auto-detected for known agentsbaseUrlEnvVar: env var for base URL override — auto-detected for known agentsKnown agent defaults (auto-filled, user only needs value):
| command | envVar | baseUrl | baseUrlEnvVar |
|---|---|---|---|
claude | ANTHROPIC_API_KEY | https://api.anthropic.com | ANTHROPIC_BASE_URL |
codex | CODEX_API_KEY | https://api.openai.com/v1 | OPENAI_BASE_URL |
gemini | GEMINI_API_KEY | https://generativelanguage.googleapis.com | GEMINI_API_BASE_URL |
Custom agents must explicitly set envVar and baseUrl in the secret.
targets (required, non-empty array)Docker images for sandboxed execution:
{ "name": "node-20", "image": "node:20-slim", "timeout": 1200, "additionalContext": "Node.js 20 with npm" }
Fields: name (required), image (required), timeout (optional, seconds), additionalContext (optional, included in generator prompt)
sandbox (required object, can be {}){
"concurrency": 3,
"defaultTimeout": 600,
"memoryMib": 2048,
"cpus": 2,
"secrets": {
"EXTRA_API_KEY": {
"value": "$EXTRA_KEY",
"allowHosts": ["api.extra-service.com"],
"allowHostPatterns": ["*.extra-service.com"]
}
},
"env": {
"LICENSE_KEY": "$MY_LICENSE_KEY"
}
}
secrets: TLS-injected secrets that never enter the VM. Each needs value and non-empty allowHosts.env: Plain env vars passed directly into sandbox. Values can use $VAR to reference host env.workspace (optional){ "template": "./workspace-template", "setupScript": "./setup.sh" }
For the full schema with all validation rules, see config-schema.md.
{
"privateInfo": [
{ "type": "local", "path": "./sdk-source", "additionalContext": "Main SDK source code" }
],
"publicInfo": [
{ "type": "package", "name": "my-sdk", "installCommand": "npm install my-sdk", "language": "typescript" },
{ "type": "url", "url": "https://docs.my-sdk.io/getting-started" }
],
"agents": {
"generator": { "command": "claude" },
"executor": { "command": "claude", "secret": { "value": "$ANTHROPIC_API_KEY" } },
"judge": { "command": "claude", "secret": { "value": "$ANTHROPIC_API_KEY" } }
},
"targets": [
{ "name": "node-20", "image": "node:20-slim", "timeout": 1200 }
],
"sandbox": {
"concurrency": 3,
"defaultTimeout": 600
}
}
After creating config.json, run agentic-usability generate -p <project> to create the test suite.