verify-samples-tool

作者: microsoft

如何使用 verify-samples 工具来运行、验证和管理 Agent Framework 仓库中的示例定义。在添加、更新或运行…时使用此工具。

npx skills add https://github.com/microsoft/agent-framework --skill verify-samples-tool

verify-samples Tool

The verify-samples project (dotnet/eng/verify-samples/) is an automated tool that runs sample projects and verifies their output using deterministic checks and AI-powered verification.

Running verify-samples

Important: By default, samples must be pre-built before running verify-samples. Build the solution first, or pass --build to build samples during the run:

cd dotnet
dotnet build agent-framework-dotnet.slnx -f net10.0

Then run verify-samples:

# Run all samples across all categories
dotnet run --project eng/verify-samples -- --log results.log --csv results.csv

# Run a specific category
dotnet run --project eng/verify-samples -- --category 02-agents --log results.log

# Run specific samples by name
dotnet run --project eng/verify-samples -- Agent_Step02_StructuredOutput Agent_Step09_AsFunctionTool

# Control parallelism (default 8)
dotnet run --project eng/verify-samples -- --parallel 8 --log results.log

# Build samples during run (skips the need for a prior build step)
# This may cause build conflicts as multiple samples are built in parallel, so use with caution
dotnet run --project eng/verify-samples -- --build --log results.log

# Combine options
dotnet run --project eng/verify-samples -- --category 03-workflows --parallel 4 --log results.log --csv results.csv --md results.md

Required Environment Variables

The tool itself needs:

  • AZURE_OPENAI_ENDPOINT — for the AI verification agent
  • AZURE_OPENAI_DEPLOYMENT_NAME (optional, defaults to gpt-5-mini)

Individual samples require their own env vars (e.g., AZURE_AI_PROJECT_ENDPOINT). The tool automatically checks and skips samples with missing env vars.

Output Files

  • --log results.log — detailed per-sample log with stdout/stderr, AI reasoning, and a summary
  • --csv results.csv — tabular summary with Sample, ProjectPath, Status, FailedChecks, and Failures columns
  • --md results.md — Markdown summary with results table and collapsible failure details (suitable for GitHub PR comments)

Sample Categories

Definitions are in the dotnet/eng/verify-samples/ directory:

CategoryConfig FileRegistered Key
01-get-startedGetStartedSamples.cs01-get-started
02-agentsAgentsSamples.cs02-agents
03-workflowsWorkflowSamples.cs03-workflows

Categories are registered in VerifyOptions.cs in the s_sampleSets dictionary.

SampleDefinition Properties

Each sample is defined as a SampleDefinition in the appropriate config file. Key properties:

new SampleDefinition
{
    // Required: Display name for the sample
    Name = "Agent_Step02_StructuredOutput",

    // Required: Relative path from dotnet/ to the sample project directory
    ProjectPath = "samples/02-agents/Agents/Agent_Step02_StructuredOutput",

    // Environment variables the sample requires (throws if missing)
    RequiredEnvironmentVariables = ["AZURE_OPENAI_ENDPOINT"],

    // Environment variables with defaults that would prompt on console if unset
    OptionalEnvironmentVariables = ["AZURE_OPENAI_DEPLOYMENT_NAME"],

    // Skip this sample with a reason (for structural issues only)
    SkipReason = null, // or "Requires external service X."

    // Deterministic checks: substrings that must appear in stdout
    MustContain = ["=== Section Header ==="],

    // Substrings that must NOT appear in stdout
    MustNotContain = [],

    // If true, only MustContain checks are used (no AI verification)
    IsDeterministic = false,

    // AI verification: natural-language descriptions of expected output
    // Each entry describes one aspect to verify independently
    ExpectedOutputDescription =
    [
        "The output should show structured person information with Name, Age, and Occupation fields.",
        "The output should not contain error messages or stack traces.",
    ],

    // Stdin inputs to feed to the sample (for interactive samples)
    Inputs = ["Y", "Y", "Y"],

    // Delay between stdin inputs in ms (default 2000, increase for LLM calls between inputs)
    InputDelayMs = 3000,
}

How to Add a New Sample Definition

  1. Check the sample's Program.cs to understand:

    • What environment variables it reads (look for GetEnvironmentVariable)
    • Whether it needs stdin input (look for Console.ReadLine, Application.GetInput)
    • Whether it has an external loop (look for EXIT patterns in YAML workflows)
    • What output it produces (section headers, markers, expected behavior)
    • Whether it exits on its own or runs as a server
  2. Choose the right verification strategy:

    • Deterministic (IsDeterministic = true): Use MustContain for samples with fixed output strings. No AI verification.
    • AI-verified (default): Use ExpectedOutputDescription with semantic descriptions. Write expectations that are flexible enough for non-deterministic LLM output.
    • Both: Use MustContain for fixed markers AND ExpectedOutputDescription for LLM-generated content.
  3. Set SkipReason only for structural issues:

    • Web servers that don't exit
    • Multi-process client/server architectures
    • Samples requiring external infrastructure (MCP servers you can't reach, Docker, etc.)
    • Do NOT skip for missing env vars — the tool checks those dynamically.
  4. For interactive samples, provide Inputs:

    • Samples using Application.GetInput(args) need one initial input
    • Samples with Console.ReadLine() approval loops need "Y" inputs
    • YAML workflows with externalLoop need "EXIT" as the last input
    • Set InputDelayMs to 3000-8000ms for samples with LLM calls between inputs
  5. Add the definition to the appropriate config file (e.g., AgentsSamples.cs) in the All list.

  6. Register new categories (if needed) in VerifyOptions.cs s_sampleSets dictionary.

Writing Good ExpectedOutputDescription

  • Write descriptions that are semantically flexible — LLM output varies between runs
  • Each array entry should describe one independent aspect to verify
  • Always include "The output should not contain error messages or stack traces." as the last entry
  • Avoid exact wording expectations — use "should mention", "should contain information about", "should show"
  • Bad: "The output should say 'The weather in Amsterdam is cloudy with a high of 15°C'"
  • Good: "The output should contain weather information about Amsterdam mentioning cloudy weather with a high of 15°C."

Example: Simple LLM Sample

new SampleDefinition
{
    Name = "Agent_With_AzureOpenAIChatCompletion",
    ProjectPath = "samples/02-agents/AgentProviders/azure/Agent_With_AzureOpenAIChatCompletion",
    RequiredEnvironmentVariables = ["AZURE_OPENAI_ENDPOINT"],
    OptionalEnvironmentVariables = ["AZURE_OPENAI_DEPLOYMENT_NAME"],
    ExpectedOutputDescription =
    [
        "The output should contain a joke about a pirate.",
        "The output should not contain error messages or stack traces.",
    ],
},

Example: Deterministic Sample

new SampleDefinition
{
    Name = "Workflow_Visualization",
    ProjectPath = "samples/03-workflows/Visualization",
    IsDeterministic = true,
    MustContain = ["Generating workflow visualization...", "Mermaid string:", "DiGraph string:"],
    ExpectedOutputDescription = ["The output should show workflow visualization in Mermaid and DiGraph formats."],
},

Example: Interactive Sample with Approval Loop

new SampleDefinition
{
    Name = "FoundryAgent_Hosted_MCP",
    ProjectPath = "samples/02-agents/ModelContextProtocol/FoundryAgent_Hosted_MCP",
    RequiredEnvironmentVariables = ["AZURE_AI_PROJECT_ENDPOINT"],
    OptionalEnvironmentVariables = ["AZURE_AI_MODEL_DEPLOYMENT_NAME"],
    Inputs = ["Y", "Y", "Y", "Y", "Y"],
    InputDelayMs = 5000,
    ExpectedOutputDescription = ["The output should show an agent using the Microsoft Learn MCP tool with approval prompts."],
},

Example: Declarative Workflow with External Loop

new SampleDefinition
{
    Name = "Workflow_Declarative_FunctionTools",
    ProjectPath = "samples/03-workflows/Declarative/FunctionTools",
    RequiredEnvironmentVariables = ["AZURE_AI_PROJECT_ENDPOINT"],
    OptionalEnvironmentVariables = ["AZURE_AI_MODEL_DEPLOYMENT_NAME"],
    Inputs = ["What are today's specials?", "EXIT"],
    InputDelayMs = 8000,
    ExpectedOutputDescription = ["The output should show a workflow calling function tools to answer a question about restaurant specials."],
},

Example: Skipped Sample

new SampleDefinition
{
    Name = "Agent_MCP_Server",
    ProjectPath = "samples/02-agents/ModelContextProtocol/Agent_MCP_Server",
    RequiredEnvironmentVariables = ["AZURE_OPENAI_ENDPOINT"],
    OptionalEnvironmentVariables = ["AZURE_OPENAI_DEPLOYMENT_NAME"],
    SkipReason = "Runs as an MCP stdio server that does not exit on its own.",
},

来自 microsoft 的更多技能

oss-growth
microsoft
OSS增长黑客角色
agent-framework-azure-ai-py
microsoft
使用Microsoft Agent Framework Python SDK(agent-framework-azure-ai)构建Azure AI Foundry代理。在创建使用AzureAIAgentsProvider的持久化代理、使用托管工具(代码解释器、文件搜索、网络搜索)、集成MCP服务器、管理对话线程或实现流式响应时使用。涵盖函数工具、结构化输出和多工具代理。
development
airunway-aks-setup
microsoft
在AKS上设置AI Runway——从裸集群到运行模型。涵盖集群验证、控制器安装、GPU评估、提供商设置和首次部署。适用场景:“设置AI Runway”、“接入AKS集群”、“安装AI Runway”、“airunway设置”、“将模型部署到AKS”、“在AKS上进行GPU推理”、“在AKS上配置KAITO”、“在AKS上运行LLM”、“在AKS上使用vLLM”、“在AKS上设置模型服务”、“AI Runway控制器”。
devops
appinsights-instrumentation
microsoft
使用Azure Application Insights对Web应用进行插桩的指南。提供遥测模式、SDK设置和配置参考。适用场景:如何对应用进行插桩、App Insights SDK、遥测模式、什么是App Insights、Application Insights指南、插桩示例、APM最佳实践。
devops
applicationinsights-web-ts
microsoft
使用Application Insights JavaScript SDK(@microsoft/applicationinsights-web)为浏览器/Web应用添加检测。用于真实用户监控(RUM)——页面视图、点击、AJAX/fetch依赖项、异常、自定义事件,以及与后端OpenTelemetry追踪关联的浏览器端GenAI代理追踪。涵盖SDK加载器脚本和npm设置、框架扩展(React、React Native、Angular)、点击分析、遥测初始化器,以及从浏览器发出的代理/工具/模型跨度所遵循的OTel GenAI语义约定。
devops
azure-ai-anomalydetector-java
microsoft
使用适用于 Java 的 Azure AI 异常检测器 SDK 构建异常检测应用程序。在实现单变量/多变量异常检测、时间序列分析或 AI 驱动的监控时使用。
development
azure-ai-language-conversations-py
microsoft
使用azure-ai-language-conversations Python SDK实现对话语言理解(CLU)。当使用ConversationAnalysisClient分析对话意图和实体、构建NLP功能或将语言理解集成到应用程序中时使用。
development
azure-ai-ml-py
microsoft
Azure Machine Learning SDK v2 for Python。用于机器学习工作区、作业、模型、数据集、计算资源和管道。 触发词:“azure-ai-ml”、“MLClient”、“工作区”、“模型注册表”、“训练作业”、“数据集”。
development