docs-impact-localizer

作者: microsoft

使用此技能將分類器的就地判定轉換為精確、逐頁的工作計畫,供 docs-sync 面板使用。在……之後啟用。

npx skills add https://github.com/microsoft/apm --skill docs-impact-localizer

docs-impact-localizer

Single responsibility: given a list of candidate pages from the classifier, produce a per-page task brief the docs-sync panel can fan out against.

You are NOT the verdict-maker (classifier owns that). You are NOT the writer (doc-writer owns that). You are the work planner.

When to invoke

The docs-sync orchestrator invokes you ONLY when the classifier returned verdict: in_place. For no_change you don't run. For structural the architect runs first; you may run after, scoped to existing pages that need amendment.

Inputs

  • scope_pages[] from the classifier
  • The PR diff (gh pr diff $PR)
  • .apm/docs-index.yml (per-page metadata)
  • Optional: the structural architect's TOC delta (if you run after the architect on a structural verdict)

Step 1: load page contents

For each path in scope_pages[], read the file. Pages are typically 3-10 KB; total budget for this step is bounded by the candidate count (the classifier should have kept it to <= 6).

Step 2: narrow scope inside each page

For each page, identify the SPECIFIC section(s) that need to change:

  • Read the page's H2/H3 structure
  • For each diff symbol from the classifier output, find the section most directly documenting it
  • Capture line ranges: lines 120-145 not the whole page

The output is a sections_to_edit[] per page, where each entry is:

page: docs/src/content/docs/consumer/install.md
sections_to_edit:
  - section: "## From Git"
    line_range: [120, 145]
    diff_symbol: "--no-cache flag"
    edit_kind: add | modify | remove
    rationale: "the new --no-cache flag is documented nowhere; section already lists other flags so this is the natural home"

Step 3: detect cross-page conflicts

If two pages document the same symbol and the diff changes the symbol's behaviour, BOTH pages need an edit AND they must stay consistent. Flag this in the brief so the CDO synthesizer knows to cross-check coherence between the two redrafts:

cross_page_constraint:
  pages: [path1, path2]
  shared_symbol: "apm install --target"
  consistency_required: "both pages must reflect the same default value"

Step 4: emit the per-page task brief

Return JSON with this shape (one entry per page in scope_pages[]):

{
  "tasks": [
    {
      "page": "docs/src/content/docs/consumer/install.md",
      "persona_owner": "consumer",
      "promise": 1,
      "sections_to_edit": [
        {
          "section": "## From Git",
          "line_range": [120, 145],
          "diff_symbol": "--no-cache flag",
          "edit_kind": "add",
          "rationale": "..."
        }
      ],
      "verify_claims": [
        {"claim": "the flag is named --no-cache", "verify_with": "apm install --help"},
        {"claim": "the flag is documented in click.option decorator", "verify_with": "grep -n no-cache src/apm_cli/commands/install.py"}
      ]
    }
  ],
  "cross_page_constraints": [
    {"pages": [...], "shared_symbol": "...", "consistency_required": "..."}
  ],
  "estimated_panel_calls": 8
}

The verify_claims[] per page is consumed by the python-architect panelist -- it tells the verifier WHICH claims need a S7 tool-call check (run apm install --help, grep the source) rather than prose-trusting.

Output contract

Return a SINGLE JSON document matching the schema in Step 4 as the final message of your task. No prose around the JSON.

Anti-patterns

  • Selecting whole pages when one section suffices (inflates context per panelist).
  • Skipping verify_claims[] -- that's the S7 tool-bridge hook; the verifier needs it.
  • Inventing pages not in scope_pages[] -- that's the classifier's job, not yours. If you think the classifier missed a page, return an extra field localizer_concern instead of expanding scope unilaterally.

來自 microsoft 的更多技能

oss-growth
microsoft
開源增長駭客角色
agent-framework-azure-ai-py
microsoft
使用Microsoft Agent Framework Python SDK(agent-framework-azure-ai)构建Azure AI Foundry代理。适用于使用AzureAIAgentsProvider创建持久化代理、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。涵盖函数工具、结构化输出和多工具代理。
development
airunway-aks-setup
microsoft
在AKS上設定AI Runway——從裸叢集到執行模型。涵蓋叢集驗證、控制器安裝、GPU評估、供應商設定及首次部署。時機:「設定AI Runway」、「上線AKS叢集」、「安裝AI Runway」、「airunway設定」、「部署模型至AKS」、「在AKS上進行GPU推論」、「在AKS上設定KAITO」、「在AKS上執行LLM」、「在AKS上使用vLLM」、「在AKS上設定模型服務」、「AI Runway控制器」。
devops
appinsights-instrumentation
microsoft
使用Azure Application Insights檢測Web應用程式的指南。提供遙測模式、SDK設定與組態參考。適用時機:如何檢測應用程式、App Insights SDK、遙測模式、什麼是App Insights、Application Insights指南、檢測範例、APM最佳實踐。
devops
applicationinsights-web-ts
microsoft
使用Application Insights JavaScript SDK(@microsoft/applicationinsights-web)為瀏覽器/Web應用程式進行檢測。適用於真實使用者監控(RUM)——頁面檢視、點擊、AJAX/fetch依賴、例外、自訂事件,以及與後端OpenTelemetry追蹤關聯的瀏覽器端GenAI代理追蹤。涵蓋SDK載入器指令碼與npm設定、框架擴充(React、React Native、Angular)、點擊分析、遙測初始化器,以及從瀏覽器發出的代理/工具/模型span的OTel GenAI語意慣例。
devops
azure-ai-anomalydetector-java
microsoft
使用適用於 Java 的 Azure AI 異常偵測器 SDK 建置異常偵測應用程式。在實作單變量/多變量異常偵測、時間序列分析或 AI 驅動監控時使用。
development
azure-ai-language-conversations-py
microsoft
使用 azure-ai-language-conversations Python SDK 實作對話語言理解(CLU)。當使用 ConversationAnalysisClient 分析對話意圖與實體、建置 NLP 功能,或將語言理解整合至應用程式時使用。
development
azure-ai-ml-py
microsoft
Azure Machine Learning SDK v2 for Python。用於機器學習工作區、作業、模型、資料集、計算資源與管線。 觸發詞:「azure-ai-ml」、「MLClient」、「workspace」、「model registry」、「training jobs」、「datasets」。
development