python-feature-lifecycle

作者: microsoft

Agent Framework 在兩個不同層級使用生命週期:

npx skills add https://github.com/microsoft/agent-framework --skill python-feature-lifecycle

Python Feature Lifecycle

Two lifecycle levels

Agent Framework uses lifecycle at two different levels:

  1. Package lifecycle — the maturity of the package as a whole
  2. Feature lifecycle — the maturity of a specific API or feature inside that package

These are related, but they are not the same thing.

  • The package stage is the default for everything in the package.
  • Feature-stage decorators are only for exceptions when a feature is behind the package's default stage.
  • Do not decorate every class or function just because the package is experimental or release candidate.

Important default

If a package is still in beta / experimental preview, all public APIs in that package are experimental by default.

  • Do not add @experimental(...) everywhere in that package.
  • The package stage already communicates that default.

Once a package moves forward, you can keep individual features behind:

  • If a package moves to release candidate, a feature may remain experimental
  • If a package moves to released / GA, a feature may remain experimental or release candidate

That is the main use case for feature-stage decorators.

The four stages

1. Experimental

Use for features that are still unstable and may change or be removed without notice.

Feature-level code pattern:

from ._feature_stage import ExperimentalFeature, experimental


@experimental(feature_id=ExperimentalFeature.MY_FEATURE)
class MyFeature:
    ...

Behavior:

  • Adds an experimental warning block to the docstring
  • Records feature metadata on the decorated object
  • Emits a runtime warning the first time the feature is used (once per feature by default)

Enum setup:

  • Add an all-caps member to ExperimentalFeature
  • Reuse the same feature ID across all APIs that belong to the same conceptual feature

2. Release candidate

Use for features that are nearly stable but may still receive small refinements before GA.

Feature-level code pattern:

from ._feature_stage import ReleaseCandidateFeature, release_candidate


@release_candidate(feature_id=ReleaseCandidateFeature.MY_FEATURE)
class MyFeature:
    ...

Behavior:

  • Adds a release-candidate note to the docstring
  • Records feature metadata on the decorated object
  • Does not emit the experimental warning

Enum setup:

  • Add an all-caps member to ReleaseCandidateFeature

3. Released

Use for stable GA APIs.

Code pattern:

  • No feature-stage decorator
  • No entry in ExperimentalFeature
  • No entry in ReleaseCandidateFeature

If a feature is fully released, remove any stage-specific feature annotation.

4. Deprecated

Use for APIs that still exist but should not be used for new code.

Code pattern:

import sys

if sys.version_info >= (3, 13):
    from warnings import deprecated  # type: ignore # pragma: no cover
else:
    from typing_extensions import deprecated  # type: ignore # pragma: no cover


@deprecated("MyOldFeature is deprecated. Use MyNewFeature instead.")
class MyOldFeature:
    ...

Behavior:

  • Uses the repository's version-conditional deprecation import pattern
  • Should describe what to use instead

Deprecated APIs should not also carry feature-stage decorators.

Expected decorators by stage

Feature stageExpected annotation
Experimental@experimental(feature_id=ExperimentalFeature.X)
Release candidate@release_candidate(feature_id=ReleaseCandidateFeature.X)
ReleasedNo feature-stage decorator
Deprecated@deprecated("...")

Feature enums

The feature enums are the inventory of currently staged features:

  • ExperimentalFeature
  • ReleaseCandidateFeature

Guidance:

  • Use one enum member per conceptual feature, not per class
  • Ideally, an ADR already defines the overall feature boundary and therefore the feature ID that staged APIs for that feature should reuse
  • Keep feature IDs all caps
  • Reuse the same member across related APIs for the same feature
  • Remove enum members when the feature no longer belongs to that stage
  • Treat these enums as current-stage inventories, not as a stable consumer introspection API

Minimal consumer guidance:

  • Treat __feature_stage__ and __feature_id__ as optional staged metadata, not as stable contracts
  • Use getattr(obj, "__feature_stage__", None) and getattr(obj, "__feature_id__", None) rather than direct attribute access
  • Treat missing metadata as "no explicit feature-stage annotation"
  • For warning filters while a feature is staged, match the literal feature ID string
  • Do not rely on ExperimentalFeature.X, ReleaseCandidateFeature.X, or the continued presence of __feature_id__ after a feature moves stages or is released

For consumers, the enums are also re-exported from agent_framework.

For internal implementation code inside agent_framework, continue to import the enums and decorators from ._feature_stage.

Package stage vs feature stage

Use the following rules:

Package is experimental / beta

  • All public APIs are experimental by default
  • Do not add feature-stage decorators just to restate that
  • Only introduce feature-level annotations later if the package advances first

Package is release candidate

  • All public APIs are RC by default
  • Do not decorate everything
  • Add @experimental(...) only for features that are intentionally still behind the package

Package is released / GA

  • All public APIs are released by default
  • Add @experimental(...) or @release_candidate(...) only for features still being held back

Moving a feature from one stage to the next

Experimental -> Release candidate

  1. Move the feature ID from ExperimentalFeature to ReleaseCandidateFeature
  2. Replace @experimental(...) with @release_candidate(...)
  3. Update any tests or docs that mention the old stage

Experimental -> Released

  1. Remove @experimental(...)
  2. Remove the feature from ExperimentalFeature
  3. Do not add a replacement feature-stage decorator

Release candidate -> Released

  1. Remove @release_candidate(...)
  2. Remove the feature from ReleaseCandidateFeature
  3. Leave the API undecorated

Any stage -> Deprecated

  1. Remove any feature-stage decorator
  2. Remove the feature from the stage enum
  3. Add @deprecated("...")
  4. Update docs/tests to reflect the replacement path

Promotion guidance

Features do not have to pass through every stage.

  • It is usually a good idea to move features in order when that reflects reality
  • But it is completely acceptable to go experimental -> released
  • Do not force a feature through release candidate if there is no real RC period

Likewise, when a package advances, do not automatically move every feature with it.

  • Promote features based on actual readiness
  • Keep lagging features explicitly marked only when they are behind the package default

Practical rules of thumb

  • Package default first, feature exceptions second
  • Do not decorate everything in preview packages
  • Do not double-annotate members of an already-staged class
  • Use enums only for currently staged features
  • Do not treat stage enums as a compatibility contract
  • Treat __feature_stage__ and __feature_id__ as optional metadata; use getattr
  • Remove stage annotations once a feature is released or deprecated

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