dsh-agent-eval

DSH 插件：Agent 能力评估框架 — 定义任务、跑 benchmark、评分、跟踪改进。

为什么需要

换了 model？改了 prompt？加了新 skill？怎么知道 Agent 变强了还是变弱了？

本插件提供一个自托管的 eval 框架：定义一组任务 + 预期结果，headless 跑完自动评分，对比前后差异。

工具

快速开始

1. 定义 eval 任务

创建 eval/tasks/basic.json：

[
  {
    "id": "hello-world",
    "name": "Create hello world script",
    "prompt": "Create a file hello.js that prints 'Hello, World!' to stdout",
    "expectedOutcome": { "type": "file_contains", "path": "hello.js", "content": "Hello, World!" },
    "tags": ["basic"]
  },
  {
    "id": "fix-bug",
    "name": "Fix the add function",
    "prompt": "add.js has a bug: it subtracts instead of adding. Fix it.",
    "expectedOutcome": { "type": "command_succeeds", "command": "node -e \"if(require('./add.js').add(2,3)!==5) throw 'FAIL'\"" },
    "tags": ["basic"]
  }
]

2. 运行 eval

Agent: → eval_run({ suite: "basic", model: "deepseek-v4-flash" })

结果：
{
  "score": "75%",
  "passed": "3/4",
  "results": [
    { "task": "Create hello world", "passed": "✓", "duration": "8s" },
    { "task": "Fix the add function", "passed": "✓", "duration": "12s" },
    { "task": "Find TODO comments", "passed": "✓", "duration": "5s" },
    { "task": "Run tests and report", "passed": "✗", "details": "npm not installed" }
  ]
}

3. 换模型后对比

Agent: → eval_compare({ baseline: "eval/results/basic-old.json", current: "eval/results/basic-new.json" })

{
  "verdict": "improved",
  "scoreDelta": "+25%",
  "improvements": ["Fix the add function: was FAIL, now PASS"],
  "regressions": []
}

支持的预期结果类型

目录结构

eval/
├── tasks/           # eval 任务定义（JSON）
│   ├── basic.json
│   ├── coding.json
│   └── testing.json
└── results/         # 运行结果（自动生成）
    ├── basic-1692000000.json
    └── basic-1692100000.json

配置

- insert:
    - id: agent-eval
      name: dsh-agent-eval
      config:
        fixturesDir: ./eval/tasks
        resultsDir: ./eval/results
        defaultTimeout: 120000    # 每个任务最长 2 分钟

典型用途

1. Model 选型：同一套任务，跑 Flash vs Pro，看谁分数高
2. Prompt 调优：改完 system prompt 后跑 eval 确认没回归
3. Skill 验证：加了新 skill 后跑 eval 看是否提升相关任务分数
4. CI 集成：每次 prompt/config 变更后自动跑 eval，分数下降则阻断

License

MIT