adr-author

작성자: microsoft

아키텍처 결정 기록(ADR) 작성을 위한 스킬로, 캡처, 플래너로부터의 인계, 템플릿 채택 진입 모드를 지원하며 선택 가능한 Y-문장을 포함합니다…

npx skills add https://github.com/microsoft/hve-core --skill adr-author

adr-author

Overview

This skill encodes the per-phase authoring conventions for Architecture Decision Records consumed by the ADR Creator agent. It also supports direct invocation when no ADR Creator state file exists. Direct callers first run the session recovery and bootstrap protocol from adr-identity.instructions.md: resolve or create .copilot-tracking/adr-plans/{projectSlug}/state.json, confirm entryMode, projectSlug, and outputTemplate, then continue at the phase recorded in state. It supports three entry modes and two output templates and converges all of them at the Govern phase, where the final ADR file is written and lineage is updated atomically.

Entry modes (state.entryMode):

  • capture — Interactive authoring driven by user answers to Frame and Decide questions.
  • from-planner-handoff — Entry from an upstream planner (Security, RAI, SSSC) with pre-populated Frame fields. Frame still requires user confirmation before exit.
  • adopt-template — One-time setup mode that ingests a project's pre-existing ADR template and emits both the first ADR and a committed .adr-config.yml.

Output templates (state.outputTemplate):

  • y-statement — Compact Y-Statement-shaped ADR for low-stakes or reversible decisions. Compressed Frame; ASR triggers optional.
  • madr-v4 — Long-form MADR v4.0.0 ADR for architecturally significant decisions. ASR trigger evaluation required during Frame.

Entry mode and output template are independent: a from-planner-handoff session can target either y-statement or madr-v4, and a capture session can do the same. The adopt-template mode produces output shaped by the user's normalized template.

Lifecycle at a glance:

ModePhase sequenceOutput
captureFrame → Decide → GovernShaped by outputTemplate (y-statement or madr-v4)
from-planner-handoffFrame (confirm pre-populated) → Decide → GovernShaped by outputTemplate (y-statement or madr-v4)
adopt-templateIngest → Normalize → Derive Questions → Fill → GovernFirst ADR + .adr-config.yml per the BYO contract

The state machine, hard exit gates, autonomy tiers (manual, partial, full), and the canonical state.json schema are defined in adr-identity.instructions.md. This skill provides the authoring activities and artifact contracts; it does not redefine the state machine.

Frame

Activities:

  • Scope — capture the decision in one or two sentences; bound it to a single project.
  • Decision-makers — record deciders, consulted, informed (RACI-aligned). Prefer a role or team handle over a personal name, and never record personal contact details, secrets, credentials, or third-party or customer PII in any ADR field.
  • Drivers — list decision drivers (functional needs, business goals).
  • Constraints — list non-negotiables (regulatory, platform, contractual, time).
  • ASR trigger evaluation — required when state.outputTemplate == 'madr-v4'. Evaluate triggers against the rubric in adr-standards.instructions.md and record results in state.asrTriggers[]. Defer the rubric and full taxonomy to that file and to references/asr-trigger-taxonomy.md.
  • Diagram-format prompt — when state.userPreferences.diagramFormat is unset, check the architecture-diagrams root state at .copilot-tracking/architecture-diagrams/state.json; if that file provides userPreferences.diagramFormat, use it. Otherwise ask the user for ascii or mermaid, persist the answer to state.userPreferences.diagramFormat, and persist it to the architecture-diagrams root state for standalone reuse. When a caller, handoff, or existing ADR state already provides state.userPreferences.diagramFormat, treat that value as authoritative and do not ask again. Required before Frame can exit.

Hard exit gate (restated from adr-identity.instructions.md):

The Frame phase cannot advance without all of the following recorded: scope statement, deciders list, decision drivers, ASR triggers determination (when outputTemplate == 'madr-v4'), and userPreferences.diagramFormat. The user must confirm the Frame summary before advancing.

Output artifacts:

  • Frame section of the in-progress ADR draft (working draft only; not yet written to disk).
  • Updated state.json fields: scope, deciders, consulted, informed, drivers, constraints, asrTriggers (when outputTemplate == 'madr-v4'), userPreferences.diagramFormat.

Decide

Activities:

  • Option enumeration — at least two considered options. A single-option ADR is rejected at this gate.
  • Evaluation criteria — score each option against the drivers and constraints captured in Frame.
  • Decision outcome selection — name the chosen option and articulate the rationale.
  • Consequences — document positive, negative, and neutral consequences. Negative consequences are not optional; an ADR with no documented downside is rejected at this gate.

Y-Statement assembly (when state.outputTemplate == 'y-statement'):

  • Compose the chosen option using the verbatim six-slot formula in templates/y-statement.md:

    In the context of (USE CASE), facing (CONCERN), we decided for (OPTION) and against (ALTERNATIVES), to achieve (QUALITY), accepting (DOWNSIDE).

  • The Y-Statement is the entire Decide output for the y-statement template. The full MADR options table is omitted.

Hard exit gate:

The Decide phase cannot advance without at least two considered options, the chosen option, and the decision rationale recorded. The user must confirm the Decide summary before advancing.

Govern

This phase converges all three entry modes and is the only phase that writes ADR files to disk.

Activities:

  1. MADR v4 frontmatter assembly — render the ADR frontmatter from templates/madr-v4.md. The template is reproduced verbatim from MADR v4.0.0 (CC0); see references/standards-excerpts.md for attribution. Merge templates/madr-v4-frontmatter-overlay.md on top to inject hve-core extension fields (id, deciders, tags, supersedes, superseded-by, related, asr_triggers) without modifying the verbatim upstream template (GP-17).
  2. Diagram render — based on state.userPreferences.diagramFormat, embed the diagram body from either templates/diagram-ascii.md or templates/diagram-mermaid.md. Skill callers do not branch on platform; the template selection is purely data-driven. When the diagram is derived from infrastructure source files (Terraform, Bicep, ARM), invoke the architecture-diagrams skill with the authoritative format recorded in state.userPreferences.diagramFormat, and embed that output in place of the scaffold fragment. Standalone diagram generation uses the architecture-diagrams root state contract rather than ADR state.
  3. Lineage validation — apply the six supersession rules summarized below; full text is in references/lineage-rules.md.
  4. Frontmatter validation — invoke scripts/validate_frontmatter.py against the staged ADR file. The script returns a non-zero exit code on schema or enum violations and is the single authority for frontmatter shape.
  5. Lineage allocator — invoke scripts/update_lineage.py to mutate .adr-config.yml. The allocator is the only writer of last_decision_id. Manual edits to last_decision_id are forbidden.
  6. Final write — write the ADR to docs/planning/adrs/{NNNN}-{slug}.md. The path is derived from the allocator-issued NNNN and the slugified ADR title.
  7. Handoff trigger — emit the dual-format (ADO + GitHub) work items per adr-handoff.instructions.md. This skill stops at handoff emission; routing is the agent's responsibility.

Govern uses the autonomy and disclaimer banners required by adr-identity.instructions.md and shared/disclaimer-language.instructions.md. Do not duplicate those texts here; load them at runtime.

Supersession Lineage Rules (Summary)

Brief enumeration. Full normative text and edge cases live in references/lineage-rules.md.

  1. supersedes and superseded-by are each a scalar string or null.
  2. Single-parent supersession — a given ADR has at most one superseded-by.
  3. Status transition — the superseding ADR's status becomes accepted; the superseded ADR's status becomes superseded.
  4. Lineage updates are atomic — both ADR files MUST be updated in the same Govern phase.
  5. The lineage allocator (scripts/update_lineage.py) is the single writer of last_decision_id in .adr-config.yml; manual edits are forbidden.

Status Taxonomy

ADR status is one of six closed values. Full definitions and lifecycle transitions live in adr-standards.instructions.md.

  • proposed — under active drafting; not yet decided.
  • accepted — decision adopted; current authority for the scope.
  • rejected — considered and declined; retained for historical context.
  • deprecated — no longer recommended but not yet replaced.
  • superseded — replaced by a newer ADR via the lineage rules below.
  • withdrawn — proposal withdrawn before decision.

ASR Trigger Catalog (Summary)

ASR (Architecturally Significant Requirement) triggers are evaluated only when state.outputTemplate == 'madr-v4'. The closed enum has eight values:

  • cost
  • performance
  • security
  • compliance
  • availability
  • scalability
  • maintainability
  • evolvability

The trigger rubric, evaluation prompts, and example mappings are defined in adr-standards.instructions.md and elaborated in references/asr-trigger-taxonomy.md. Do not introduce trigger values outside the closed enum.

Adopt-Template Lifecycle

Five-step pointer. Full lifecycle, including GP-13 (the .adr-config.yml schema and the 2-layer config resolution), lives in adr-byo-template.instructions.md.

  1. Ingest — accept the user's existing ADR template file or template directory.
  2. Normalize — invoke scripts/normalize_template.py to convert the template into the canonical ADR frontmatter and section structure.
  3. Derive Questions — generate the Frame and Decide question set from the normalized template's required fields.
  4. Fill — execute the derived questions to populate the first ADR.
  5. Govern — run the standard Govern phase. When lineage_fields are absent from the adopted template, the agent MUST warn the user and require an explicit confirmation before writing; this is the warn-and-confirm Govern behavior referenced in adr-byo-template.instructions.md.

Templates

  • templates/madr-v4.md — MADR v4.0.0 ADR template (verbatim, CC0). Used by Govern frontmatter assembly when state.outputTemplate == 'madr-v4'.
  • templates/y-statement.md — Six-slot Y-Statement formula for Decide assembly when state.outputTemplate == 'y-statement'.
  • templates/diagram-ascii.md — ASCII diagram block, selected when state.userPreferences.diagramFormat == "ascii".
  • templates/diagram-mermaid.md — Mermaid diagram block, selected when state.userPreferences.diagramFormat == "mermaid".

References

  • references/standards-excerpts.md — MADR v4.0.0 verbatim text and CC0 attribution; Y-Statement attribution; status taxonomy.
  • references/lineage-rules.md — Full text of the six supersession rules with edge cases and worked examples.
  • references/asr-trigger-taxonomy.md — Full ASR trigger taxonomy, rubric prompts, and examples for each of the eight enum values.

Scripts

  • scripts/render_template.py — Renders a template from templates/ against a Frame+Decide payload to produce an in-memory ADR draft. Path-traversal guarded: refuses any output path outside docs/planning/adrs/.
  • scripts/validate_frontmatter.py — Validates ADR frontmatter against the MADR v4 schema and the closed enums. Returns non-zero on violation. Path-traversal guarded against the same root.
  • scripts/update_lineage.py — Single writer of last_decision_id in .adr-config.yml. Mutates predecessor ADRs' superseded-by atomically with the new ADR's supersedes. Path-traversal guarded.
  • scripts/normalize_template.py — Converts a user-supplied ADR template into the canonical structure used by templates/madr-v4.md. Used only by the adopt-template lifecycle. Path-traversal guarded.
  • scripts/scan_sensitive_content.py: deterministic disclosure-risk scanner accepting file paths or stdin and emitting masked JSON findings. Base behavior detects high-confidence PII. --public adds internal URL detection. --data independently adds structured sensitive-column, connection, credential, sample-row, and international identifier detection. --denylist <path> independently adds case-insensitive literal customer-term detection from a UTF-8 file, resolved through the same allow-root guard as scanned paths and capped by file size, term count, and minimum term length. --allow-root <dir> adds a directory to that guard. The flags combine by union without changing no-flag ADR behavior. Required before durable ADR writes and external or handoff emission.

Scanner report contract

The scanner prints one JSON object to standard output on every terminal path, including error paths, so callers can assert on fields rather than parse prose. The object carries schema_version (scan-sensitive-content-v1), status (completed or error), modes (booleans for public, data, and denylist), denylist_rule_count, findings, and summary with high, warn, and total counts. On status error it also carries error with a stable code and a message. Exit codes are 0 for a completed scan with no high-confidence finding, 1 for a completed scan with at least one high-confidence finding, and 2 for an error. A caller that requires denylist enforcement asserts modes.denylist is true and denylist_rule_count is greater than zero, so an empty or unloaded denylist cannot silently disable the gate.

All scripts treat their working directory as untrusted input. The ADR authoring scripts reject paths that resolve outside the project ADR root. scan_sensitive_content.py scans content that is not an ADR, so it rejects paths that resolve outside its allow roots: this skill's directory, the repository root, the parent directory of each path passed on the command line, and any directory supplied through --allow-root.

Source Attribution

  • templates/madr-v4.md — reproduced byte-identical from MADR v4.0.0 (tag 4.0.0, file template/adr-template.md), released under CC0-1.0. CC0 does not require attribution; it is recorded here for transparency. Upstream typographical anomalies (for example, the unbalanced quotation in the status: placeholder) are preserved intentionally to keep the file diff-clean against the upstream release.

Mandatory Load Directives

The ADR Creator agent enforces a phase→section load contract per adr-identity.instructions.md. Each phase MUST load its section of this skill before executing phase work, and MUST append the section anchor to state.phaseSkillsLoaded:

PhaseSection anchorRequired phaseSkillsLoaded entry
Frame#frameadr-author#frame
Decide#decideadr-author#decide
Govern#governadr-author#govern

The agent loads sections via read_file against this skill file and records the entry in state.phaseSkillsLoaded before any phase work executes. Re-entering a previously loaded phase does not require reloading; the agent checks phaseSkillsLoaded first.

microsoft의 다른 스킬

oss-growth
microsoft
OSS 성장 해커 페르소나
agent-framework-azure-ai-py
microsoft
Microsoft Agent Framework Python SDK(agent-framework-azure-ai)를 사용하여 Azure AI Foundry 에이전트를 구축합니다. AzureAIAgentsProvider로 지속적 에이전트를 만들 때, 호스팅 도구(코드 인터프리터, 파일 검색, 웹 검색)를 사용할 때, MCP 서버를 통합할 때, 대화 스레드를 관리할 때, 또는 스트리밍 응답을 구현할 때 사용합니다. 함수 도구, 구조화된 출력, 다중 도구 에이전트를 다룹니다.
development
airunway-aks-setup
microsoft
AKS에서 AI Runway 설정 — 빈 클러스터에서 실행 중인 모델까지. 클러스터 검증, 컨트롤러 설치, GPU 평가, 공급자 설정, 첫 배포를 다룹니다. 시기: "AI Runway 설정", "AKS 클러스터 온보딩", "AI Runway 설치", "airunway 설정", "AKS에 모델 배포", "AKS에서 GPU 추론", "AKS에서 KAITO 설정", "AKS에서 LLM 실행", "AKS에서 vLLM", "AKS에서 모델 서빙 설정", "AI Runway 컨트롤러".
devops
appinsights-instrumentation
microsoft
Azure Application Insights로 웹앱을 계측하기 위한 지침입니다. 원격 분석 패턴, SDK 설정, 구성 참조를 제공합니다. WHEN: 앱 계측 방법, App Insights SDK, 원격 분석 패턴, App Insights란 무엇인가, Application Insights 지침, 계측 예시, APM 모범 사례.
devops
applicationinsights-web-ts
microsoft
브라우저/웹 앱을 Application Insights JavaScript SDK(@microsoft/applicationinsights-web)로 계측합니다. Real User Monitoring(RUM) — 페이지 뷰, 클릭, AJAX/fetch 종속성, 예외, 사용자 지정 이벤트, 백엔드 OpenTelemetry 트레이스와 상관관계가 있는 브라우저 측 GenAI 에이전트 트레이스에 사용합니다. SDK Loader Script 및 npm 설정, 프레임워크 확장(React, React Native, Angular), Click Analytics, 텔레메트리 이니셜라이저, 브라우저에서 생성된 에이전트/도구/모델 스팬에 대한 OTel GenAI 의미론적 규칙을 다룹니다.
devops
azure-ai-anomalydetector-java
microsoft
Azure AI Anomaly Detector SDK for Java로 이상 탐지 애플리케이션을 구축하세요. 단변량/다변량 이상 탐지, 시계열 분석 또는 AI 기반 모니터링을 구현할 때 사용하세요.
development
azure-ai-language-conversations-py
microsoft
azure-ai-language-conversations Python SDK를 사용하여 대화형 언어 이해(CLU)를 구현합니다. ConversationAnalysisClient로 대화 의도와 엔터티를 분석하거나, NLP 기능을 구축하거나, 애플리케이션에 언어 이해를 통합할 때 사용합니다.
development
azure-ai-ml-py
microsoft
Azure Machine Learning SDK v2 for Python. ML 작업 영역, 작업, 모델, 데이터 세트, 컴퓨팅 및 파이프라인에 사용합니다. 트리거: "azure-ai-ml", "MLClient", "workspace", "model registry", "training jobs", "datasets".
development