Optional domain skill, not required by the core module. Every rule below is important to follow. None of it can be skipped. Architecture artifacts support SDD; they do not replace requirements or ADR authority.

• Skill name:

  ai-sdlc-architecture

• Primary audience: Architecture, Dev
• Supporting audience: QA, Security, Delivery, BA
• Audience tags: Architecture, Dev, QA, Security
• SDLC stage: Design and implementation planning
• Purpose: Preserve traceable architecture boundaries, decisions, and risks.
• Output:

  architecture.md

  and

  _ai_sdlc/architecture.toon

• Implementation feature root.
• Architecture input using

  ai-sdlc-architecture-input/v1

  .
• Requirement, acceptance, risk, or decision trace targets.

• Ask when system boundary, quality attribute, authority, or irreversible choice is ambiguous.
• Separate constraints from decisions and decisions from implementation tasks.
• Record alternatives and consequences for every material decision.
• Do not invent infrastructure, data classification, or production topology.

• Support

  --quick-flow

  and

  --full-flow

  ; full flow takes precedence.
• Quick flow permits a bounded architecture slice with explicit gaps.
• Full flow requires at least one decision, risk, and validation check.
• Both modes require trace targets for constraints, interfaces, decisions, and risks.

• Return design scope, decision/risk counts, blockers, validation status, and output paths directly in the Codex response.
• Before the final response, emit

  ai-sdlc-handoff/v1

  with

  result

  ,

  blockers

  ,

  next_required

  , and

  next_optional

  ; every action includes

  reason

  ,

  command

  , and

  expected_artifact

  .
• Do not create

  summary.txt

  ,

  *-summary.txt

  , or untraced diagrams.
• Keep Markdown authoritative for detail and TOON bounded for routing.

• Write

  <feature-root>/architecture.md

  .
• Write

  <feature-root>/_ai_sdlc/architecture.toon

  .
• Keep ADRs or decision-log entries separate when organizational authority requires them; link their IDs from architecture decisions.
• Do not write into refinement unless architecture work is explicitly upstream.

• Read

  <feature-root>/_ai_sdlc/state.toon

  before architecture work.
• Architecture is an optional design utility and does not add a core lifecycle stage.

• --state-check

  is read-only;

  --begin-state

  and

  --complete-state

  are rejected.
• Route accepted design changes back through SDD and change-impact recovery.

• Markdown starts with

  artifact_metadata

  using schema

  ai-sdlc-architecture-metadata/v1

  .
• Include

  metatags

  for

  ai-sdlc

  ,

  architecture

  ,

  design

  , and

  traceable

  .
• Record feature, workspace, flow mode, state file, trace IDs, and status.

• Read

  specs/_ai_sdlc/specs-index.toon

  and feature state before broad reads.
• Refresh

  specs/specs-index.md

  only after a durable architecture write.
• Do not alter

  specs-refiniment/_ai_sdlc/specs-index.toon

  or

  specs-refiniment/specs-index.md

  for implementation-owned architecture.

• Read

  references/architecture-contract.md

  for the input schema and design gates.
• Use

  scripts/architecture.py

  to validate and route canonical outputs.

• Capture design context and constraints before components.
• Trace interfaces and decisions to durable requirement or decision IDs.
• Give risks an owner and mitigation.
• Provide executable or inspectable validation evidence.

1. Read requirements, decisions, project context, state, and relevant quality findings.
2. Define boundaries, constraints, components, and interfaces.
3. Compare alternatives and record decisions plus consequences.
4. Identify architecture risks, owners, and mitigations.
5. Define validation checks for the design claims.
6. Finalize canonical outputs with the deterministic script.
7. Route implementation changes through SDD tasks and validation.

ai-sdlc-architecture/v1

• Pass when boundaries are explicit and every material claim has traceability.
• Full flow fails without decisions, risks, validation, alternatives, consequences, risk ownership, or mitigation.

• A local reversible patch may use quick flow and record no new decision.
• External systems remain components with explicit unverified interface assumptions.
• Existing ADRs are referenced, not duplicated.
• Diagram generation is optional; structured contracts remain authoritative.

• Do not approve architecture on behalf of owners.
• Do not provision infrastructure or change production systems.
• Do not replace SDD requirements, tests, tasks, or decisions.
• Do not hide unresolved architecture risks.
• Use

  $ai-sdlc-change-impact

  when an accepted design changes downstream work.