| name | agentic-sdd-architecture-review |
| description | Review an SDD implementation plan before code generation and decide whether it is approved, needs changes, or is rejected. |
| allowed-tools | Read, Write, Edit, Glob, Grep, Bash, web_search_advanced_exa, web_fetch_exa |
| risk | safe |
| source | local |
Agentic SDD Architecture Review
Use this skill as the mandatory gate between technical planning and task generation or code.
Goal
Review the plan as a production architecture gate. Do not implement code.
Required Packet
Review:
specs/<feature>/spec.md
specs/<feature>/clarify.md, if present
specs/<feature>/plan.md
specs/<feature>/data-model.md, if present
specs/<feature>/contracts/, if present
specs/<feature>/test-plan.md, if present
.specify/memory/constitution.md
- relevant ADRs
- dependency-change notes
- rollback notes
Checklist
Assess:
- scope is clear
- open questions are resolved or explicitly deferred
- plan follows existing architecture
- new abstractions are justified
- existing libraries/framework features are used before custom layers
- dependencies are justified and approved
- data ownership is clear
- migrations and rollback are documented
- sensitive data and retention rules are considered
- API compatibility and error responses are defined
- auth and authorization boundaries are explicit
- input validation and injection risks are considered
- observability and failure modes are defined
- deployment and rollback impact are understood
- tests cover the core risks
Decision
Return one of:
Approved: task breakdown may begin.
Needs changes: list blocking changes before approval.
Rejected: the plan is structurally unsafe or inconsistent with goals.
Writing Review Artifacts
Create or update specs/<feature>/review.md with:
- reviewed artifacts
- findings by severity
- approval status
- required changes
- optional improvements
- reviewer and date placeholders
If the plan is approved, update the approval status in plan.md only if the user explicitly asked you to maintain artifacts.
Stop Conditions
Do not approve when:
- a high-risk area lacks explicit human approval
- rollback is missing for migrations or operational risk
- dependencies are unjustified
- tests are absent for behavior, security, or data boundaries
- plan contradicts constitution or ADRs
- code generation would require guessing
Output
Lead with approval status, then blocking findings, non-blocking suggestions, and next phase.