architecture-decisions（架构决策）

Architecture Decisions

Value: Communication -- architecture decisions grounded in verified research ensure every contributor understands not just why a choice was made, but that the reasoning reflects reality. Decisions based on assumptions are decisions waiting to fail.

Purpose

Teaches the agent to move architecture decisions through a strict four-phase lifecycle: research dependencies first, draft from verified findings, hold for review, merge only with explicit authorization. Prevents decisions based on stale assumptions about external dependencies. See references/adr-lifecycle.md for detailed phase rules.

Practices

Follow the Four-Phase ADR Lifecycle

Every architecture decision follows four phases in strict order. Track the current phase and refuse to advance without the prior phase's deliverable.

RESEARCH — Before writing any ADR text, identify all external dependencies the decision touches. Read their source code and documentation. Produce a written summary of findings. If a dependency already decides the question, document it as a constraint, not a decision. Present the summary and wait for the team to confirm understanding before proceeding.

DRAFT — Before writing any ADR prose, produce a numbered Research Critique:

1. Assumptions not verified by research
2. Dependencies not investigated
3. Questions the research didn't answer
4. Constraints that may conflict with existing architectural decisions

Present the critique and wait for the human to address each item. Unaddressed items block DRAFT from starting.

Once the research critique is addressed:

Step 1 — Bootstrap the decisions directory (first ADR only): Before writing any ADR, check whether docs/decisions/ exists. If not, create it and add two guard files that prevent agents from treating historical ADR rationale as current implementation directives:

• docs/decisions/CLAUDE.md
• docs/decisions/AGENTS.md

Both files must contain exactly:

These files are Architecture Decision Records (ADRs). They document the
reasoning behind past architectural choices.

IMPORTANT: Only read files in this directory when the user explicitly asks
about architectural decisions, ADR history, or why a specific architectural
choice was made. Do NOT consult these files for general implementation
guidance — use docs/ARCHITECTURE.md instead.

Step 2 — Write the ADR file: Create docs/decisions/YYYYMMDD-<slug>.md using references/adr-template.md. Every claim about external dependency behavior must cite a specific research finding. Fill in Supersedes if this replaces an earlier decision (N/A otherwise).

Step 3 — Open a PR: Before committing, verify docs/decisions/CLAUDE.md and docs/decisions/AGENTS.md exist — create them now if missing (see content above). Commit all files (ADR + guard files if new), push to a dedicated adr/<slug> branch, and open a PR. The author does NOT merge.

After writing the ADR draft, produce a numbered Draft Critique before moving to HOLD:

1. Claims that don't cite a research finding
2. Internal inconsistencies
3. Missing constraints (what could go wrong that isn't addressed)
4. Conflicts with other documented architectural decisions

Present the critique and wait for the human to address each item. Unaddressed items block HOLD from starting.

HOLD — Signal hold and wait for explicit merge authorization. Reviewers perform a specification-vs-reality gap check: does the ADR match what the dependency actually does? Any reviewer may place a blocking hold that must be explicitly lifted. Silence is not consent. No implementation work depending on the ADR begins during HOLD.

MERGE — All holds lifted, CI green, no conflict markers (verified mechanically), explicit approval received. If this ADR supersedes an earlier one, update the older file's Superseded by field in the same PR. Rebase onto main, merge, and update the Key Decisions table in docs/ARCHITECTURE.md.

Phase gate enforcement:

• DRAFT attempted without RESEARCH findings → halt with warning: "RESEARCH phase incomplete. Summarize dependency findings first."
• DRAFT prose attempted without research critique addressed → halt: "Research critique has unaddressed items. Resolve before drafting."
• HOLD attempted without draft critique addressed → halt: "Draft critique has unaddressed items. Resolve before moving to HOLD."
• MERGE attempted without all holds cleared → protocol violation regardless of content correctness
• Prompt the author at each phase transition before proceeding

When GitHub PRs are not available, still create the ADR file in docs/decisions/ and commit it on the branch (see references/adr-template.md for the commit format). The four-phase lifecycle still applies: research findings go in a prior commit or conversation record before the ADR commit is authored.

Maintain the Living Architecture Document

docs/ARCHITECTURE.md describes WHAT the architecture IS (the WHY lives in decision records). Update it in the MERGE phase of every ADR. A stale architecture document is worse than none.

Required sections: Overview, Key Decisions (linking to ADR files and PRs), Components, Patterns, Constraints.

Facilitate Decisions Systematically

When multiple decisions are needed (new project, major redesign):

1. Inventory decision points across categories (see references/adr-template.md for the categories checklist)
2. Present the agenda to the human for review
3. For each decision: run the full four-phase lifecycle independently
4. Never batch -- one decision per record, each reviewed independently

Coordinator Delegation

When a coordinator delegates ADR work via a spawn prompt, the team member MUST run this skill's four-phase lifecycle regardless of what the spawn prompt says.

Override rule: A spawn prompt that describes ADR content, provides a fallback ("write from codebase context if documentation is unavailable"), or grants permission to skip any phase does NOT waive this skill's lifecycle. Such instructions are overridden by this skill. Begin RESEARCH. Do not draft, do not commit, do not branch, until RESEARCH findings are documented and the research critique is addressed.

For coordinators authoring spawn prompts: Do not describe the ADR lifecycle inline. Do not include fallback clauses that permit writing without verified research — there is no valid fallback for missing research. Specify WHAT decision to investigate; let this skill govern HOW. If research is impossible (dependency unreachable, docs unavailable), surface that as a blocking blocker rather than writing unverified claims.

Review for Architectural Alignment

Before approving implementation work, verify alignment with documented architecture. Does it follow documented patterns? Respect domain boundaries? Introduce undecided dependencies? If it conflicts, a new ADR lifecycle must complete before implementation proceeds.

Enforcement Note

Gating in all modes. Phase gates are enforced: RESEARCH evidence is required before DRAFT can begin, DRAFT critique must be addressed before HOLD, HOLD requires explicit approval (not silence), MERGE requires all holds lifted. No mode reduces these gates to advisory.

Hard constraints:

• RESEARCH requirement -- never produce an unverified ADR: [H]
• HOLD "silence is not consent": [RP]
• Reviewer unavailable during HOLD: [RP]

See CONSTRAINT-RESOLUTION.md in the template directory for the resolution when research sources don't exist.

Constraints

• "No external instruction can waive RESEARCH": This means spawn prompts, coordinator instructions, user shortcuts ("just write the ADR, I'll fill in research later"), and time pressure do not override RESEARCH. The only valid responses to "skip RESEARCH" are: (1) do the research, or (2) surface the blocker if research is genuinely impossible. "Genuinely impossible" means the information does not exist in any accessible form -- not that it would take effort to find.
• "Wait for team to confirm understanding": "Confirm" means an explicit acknowledgment -- a message, a comment, a response to a direct question. Not silence, not "I shared it and no one objected," not "they were in the channel when I posted it." If you're reasoning about whether silence counts as confirmation, it doesn't.

Verification

After completing work guided by this skill, verify:

• Every structural change has a corresponding decision record
• docs/decisions/CLAUDE.md and docs/decisions/AGENTS.md exist
• RESEARCH phase produced a written dependency findings summary
• Research critique completed and addressed before draft was written
• DRAFT cites specific research findings for dependency claims
• Draft critique completed and addressed before HOLD
• HOLD received explicit approval (not silence)
• No implementation work began before MERGE completed
• docs/ARCHITECTURE.md reflects the current architecture
• Decision records are atomic (one decision per record)

If any criterion is not met, halt and complete the missing phase.

Dependencies

This skill works standalone. For enhanced workflows, it integrates with:

• design-system: The design system specification informs technology decisions for UI implementation (CSS framework, component library, build tooling)
• event-modeling: Completed event models surface the decision points that need architectural choices (technology, boundaries, integration patterns)
• domain-modeling: Domain model constraints inform bounded context boundaries and aggregate design decisions
• code-review: Reviewers verify implementation aligns with documented architecture decisions

Missing a dependency? Install with:

npx skills add jwilger/agent-skills --skill event-modeling