implementing-code

作者: microsoft

以TDD工作流执行一批实现任务,基于源锚定重写确保行为保真度,依据指南进行代码转换,并完整…

npx skills add https://github.com/microsoft/github-copilot-modernization --skill implementing-code

User Input

You MUST consider the user input before proceeding (if not empty).

Output

Code files are written into the project tree. The batch report is written under the provided artifact path.

batch-report.yaml   ← structured result consumed by caller

Workflow

Step 1: Verify Prerequisites (first batch only)

Confirm required artifacts exist: plan.md (with task list and Requirement Mapping table), feature spec.

Step 2: Load Context

Read from the provided artifact path and dependency artifacts. Scan the artifact directory to find these categories:

CategoryRequiredPurposeLook for
Task breakdown✅Work items with T-IDs to implementArtifact with task IDs (T001, T002...), may be in tasks/ subdirectory or main planning artifact
Feature spec✅Requirements (REQ-XXX) and acceptance criteriaArtifact with REQ-XXX identifiers and user scenarios
Constitution✅Project principles and constraintsUsually named constitution.md in artifact root
Implementation plan✅Phased plan with architecture decisionsArtifact with phased sections and plan references
Architecture design✅API contracts, layering, package structureArtifact with design decisions
Data modelIf existsEntity definitions, mappings, FK strategyArtifact with schema/entity details
UX/UI specIf existsScreen layouts, user flows, componentsArtifact with screen specs
Knowledge graphIf existsModule dependencies and class relationshipsJSON file with nodes/edges in artifact path
CheckpointsIf existsUpstream traceabilityYAML files in checkpoints/ subdirectory
GuidelinesIf existsMigration rules and transformation patternsSearch guidelines skill

Step 3: Checklist Gate (first batch only)

If checklists/ directory exists, scan checklist files. If any incomplete items: display table and ask before proceeding.

Step 4: Project Setup (first batch only)

Create/verify ignore files. See references/ignore-patterns.md for patterns by language/tool.

Step 5: Match Tasks from Breakdown

From the task breakdown artifact, find tasks that match your current assignment:

  1. Read the task breakdown (from dependency artifacts)
  2. Match tasks by role, module/domain, or phase that align with your task description
  3. For each matched task, extract:
    • Task ID, description, acceptance criteria
    • Plan references and REQ-XXX traceability
    • Source file references (rewrite mode)
    • Dependencies and parallel markers

Unmet dependency not in current batch → report as blocked.

Step 5.5: Discover Required API Endpoints (web applications only)

Before writing any implementation code, check whether an endpoint contract test script exists in the project root:

for script in api-test.sh api-check.sh smoke.sh health-check.sh test-api.sh; do
  test -f "./$script" && echo "found: $script" && break
done

If a script is found:

  1. Read its full contents to extract every endpoint it exercises: URL paths, HTTP methods, expected status codes, and expected response shape.
  2. Treat every such endpoint as a mandatory acceptance criterion for this batch — equivalent to an explicit REQ in the feature spec. Missing even one will cause the post-build evaluation to fail.
  3. Cross-check each endpoint against the matched task list. If any script endpoint is absent from the task breakdown, add it as an implicit sub-task in your execution plan (e.g., T_API_dashboard: Implement GET /api/dashboard).
  4. Record the discovered endpoint list at the top of batch-report.yaml under required_endpoints:
    required_endpoints:
      - method: GET
        path: /api/dashboard
        source: api-test.sh
        status: pending   # updated to "implemented" when the endpoint is wired up and verified
      - method: GET
        path: /api/items
        source: api-test.sh
        status: pending
    

If no script is found:

  • Check clarification.md (if present) for any explicitly listed required API endpoints.
  • Check the feature spec for API contract sections.
  • If neither source provides an endpoint list, proceed without this step.

This discovery step ensures that evaluation scripts are treated as first-class requirements from the start of implementation — not discovered only at post-build verification when fixes are costly.

Step 6: Execute

Ordering:

  • Sequential tasks: in order
  • Parallel [P] tasks: can run together
  • Same-file tasks: must run sequentially

TDD Flow:

  1. Setup: project structure, dependencies, configuration
  2. Tests before code: contract, entity, integration tests
  3. Core: models, services, endpoints
  4. Integration: DB, middleware, logging, external services
  5. Run tests: execute module-level tests (e.g., mvn test -pl <module> -am). ALL tests must PASS. Report pass/fail/skip counts in batch report. Test failure = task NOT complete.

Constitution & Requirement Fidelity:

  • Before each task, re-read traced REQ-XXX — implement requirement intent, not just task description.
  • Verify decisions align with constitution principles.

Source-Anchored Rewrite (MANDATORY in rewrite mode): For tasks with [Source:], follow references/source-anchored-rewrite.md.

Guideline-Based Transformation: For tasks marked [GUIDELINE:skill-name]:

  1. Load each top-level guideline skill named in the plan
  2. Apply transformation rules, import changes, method mappings
  3. Use before/after examples as reference

Parallelism: Run independent tasks ([P] marked) in parallel as much as possible to maximize throughput.

Error Handling:

  • Halt on non-parallel task failure
  • For [P] tasks: continue successful ones, report failures

Step 6.5: JS/TS Scaffolding Validation Gate (JS/TS projects only)

After any scaffolding step that generates or modifies a package.json, execute this gate before proceeding to the next task. This gate is MANDATORY for all JavaScript and TypeScript projects.

6.5.1 — Verify required npm scripts

Check that package.json contains BOTH a build script AND a test script:

node -e "
  const pkg = require('./package.json');
  const missing = ['build','test'].filter(s => !pkg.scripts || !pkg.scripts[s]);
  if (missing.length) { console.error('MISSING scripts:', missing.join(', ')); process.exit(1); }
  console.log('scripts OK: build=' + pkg.scripts.build + ', test=' + pkg.scripts.test);
"

If either script is missing, inject it immediately — do NOT defer:

FrameworkMissing test scriptMissing build script
Angular (@angular/core in deps)"test": "ng test --watch=false --browsers=ChromeHeadless""build": "ng build"
React / Vite"test": "vitest run" or "test": "react-scripts test --watchAll=false""build": "vite build"
React / CRA"test": "react-scripts test --watchAll=false --ci""build": "react-scripts build"
Vue / Vite"test": "vitest run""build": "vite build"
Next.js"test": "jest --ci""build": "next build"
NestJS / Node"test": "jest --ci""build": "nest build"
Generic TS"test": "jest --ci""build": "tsc"

After injecting, confirm the script was written and re-verify with the check above.

6.5.2 — Run the test script

After confirming both scripts exist, execute:

npm test

(or yarn test / pnpm test if the project uses those package managers)

  • Exit code 0: gate passes — proceed.
  • Exit code != 0: enter a remediation loop (max 3 iterations):
    1. Read the failure output to identify the root cause (missing browser binary, missing test files, misconfigured jest config, etc.)
    2. Fix the cause (install missing dev dependency, create a minimal placeholder test, fix config)
    3. Re-run npm test
    4. If still failing after 3 iterations, record the failure in batch-report.yaml under warnings with severity HIGH and continue — do NOT block the entire batch:
      warnings:
        - severity: HIGH
          message: "npm test failed after 3 remediation attempts — <last error summary>"
      

Why this gate exists: Eval harnesses run npm test unconditionally. A scaffolded project without a test script causes an immediate fatal error (npm error Missing script: "test") that masks all other results. Catching this at scaffolding time costs ~5 seconds; missing it causes total batch failure.

Step 6.6: API Endpoint Verification (web-application backends only)

For any batch that produces or modifies a web-application backend (Spring Boot, Express, NestJS, FastAPI, Django, ASP.NET Core, Go HTTP servers, and similar), verify endpoints respond correctly at runtime before the batch is considered complete. A passing build proves only compilation, not that routes are wired up.

6.6.1 — Run the discovered endpoint contract (if any)

If Step 5.5 found an endpoint contract script (api-test.sh, api-check.sh, …), start the application if not already running, then execute it:

bash ./api-test.sh   # or whichever script Step 5.5 recorded
echo "api-test exit code: $?"
  • Exit code 0 → all endpoint assertions pass → proceed.
  • Exit code != 0 → enter the fix loop (max 3 iterations):
    1. Read the output to identify which endpoints returned errors (missing route, wrong status code, unexpected body).
    2. Implement or correct the backend endpoint(s): add the missing controller/handler, fix route mapping, return the expected response shape.
    3. Re-run the script. Repeat until exit code is 0 or 3 iterations are exhausted.
    4. If still failing after 3 iterations, record the failure in batch-report.yaml under warnings with severity HIGH and escalate via [notify:coordinator] — do NOT mark the affected endpoints implemented.

6.6.2 — Manual probe (no script found)

If no endpoint contract script exists, probe every REST endpoint the batch implemented after starting the application:

curl -sf -o /dev/null -w "%{http_code}" http://localhost:<PORT><endpoint>

Every implemented endpoint MUST return a 2xx status code. A 404 or 500 means the route is not wired correctly — fix before completing the batch. Record probe results in the batch report under endpoint_probes:

endpoint_probes:
  - method: GET
    path: /api/dashboard
    status: 200
    result: PASS

Step 7: Write Checkpoint

After ALL tasks in this batch complete, write checkpoints/tasks-to-impl.yaml using templates/tasks-to-impl-checkpoint-template.yaml. This is REQUIRED — completeness gate reads it to verify traceability.

Step 8: Report

Generate batch result report per references/batch-report-format.md (YAML format).

If required_endpoints was populated in Step 5.5, update each entry's status to "implemented" once the corresponding endpoint is wired up and verified to respond correctly. Any entry still "pending" at report time is a gap — record it under warnings with severity HIGH:

warnings:
  - severity: HIGH
    message: "Endpoint GET /api/dashboard was required by api-test.sh but not implemented in this batch"

Resources

References

  • references/source-anchored-rewrite.md — Rewrite-mode behavioral fidelity process
  • references/ignore-patterns.md — Ignore file patterns by language/tool
  • references/batch-report-format.md — Required YAML output format

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