A Claude Code plugin for structured, repeatable AI-assisted architecture reviews. Accepts a brief YAML file or inline description, produces a narrative review with risk table, tradeoffs, follow-up questions, and a checklist — then appends an evidence record to an audit log.
Built on the methodology of NathanMaine/architectural-design-review-agent.
- Loads an architecture brief from a YAML file or inline text.
- Validates the brief — asks for missing fields rather than guessing.
- Generates a structured review with:
- Narrative summary
- Risk table (likelihood, impact, mitigation per risk)
- Tradeoffs (gains vs. costs per key decision)
- Follow-up questions the brief does not answer
- Checklist of standard items for the architecture type
- Presents the review as readable Markdown.
- Writes the structured JSON review to an optional output file.
- Appends a JSONL evidence record (who reviewed, when, how many risks/questions).
claude mcp add memoriant-architecture-review-skillOr clone locally:
git clone https://github.com/NathanMaine/memoriant-architecture-review-skill ~/.claude/plugins/memoriant-architecture-review-skillcodex install NathanMaine/memoriant-architecture-review-skillgemini extension install NathanMaine/memoriant-architecture-review-skillReview the architecture in fixtures/my_brief.yaml
Conduct an architecture review for our new caching layer. Context: [description]
/architecture-review --brief fixtures/my_brief.yaml --out review.json
codex run architecture-review-agent --var brief=fixtures/my_brief.yamltitle: "Notifications Service v2"
context: |
Moving notification dispatch off the critical request path to reduce P99 latency.
constraints:
- Must not require a new database
- Must deploy to existing Kubernetes cluster
components:
- name: Message Queue
description: RabbitMQ for async dispatch
- name: Worker Pool
description: 4 Python consumers
key_risks:
- Queue saturation under traffic spikes
- Worker failure leaving notifications undeliveredPlain Markdown or free-form prose is also accepted.
memoriant-architecture-review-skill/
├── .claude-plugin/
│ └── plugin.json # Plugin manifest
├── skills/
│ └── architecture-review/
│ └── SKILL.md # Full methodology for Claude Code
├── agents/
│ └── architecture-review-agent.md # Autonomous agent definition
├── AGENTS.md # Codex CLI agent definitions
├── gemini-extension.json # Gemini CLI extension manifest
├── SECURITY.md # Security policy
├── README.md # This file
└── LICENSE # MIT
| Variable | Default | Description |
|---|---|---|
LLM_API_KEY |
(empty — stub mode) | API key for the LLM provider |
LLM_BASE_URL |
https://api.openai.com/v1 |
Base URL for the OpenAI-compatible API |
LLM_MODEL |
gpt-4o-mini |
Model identifier |
REVIEWER_ID |
anonymous |
Identity recorded in the evidence log |
EVIDENCE_LOG_PATH |
logs/evidence.jsonl |
Evidence log location |
## Architecture Review: Notifications Service v2
**Reviewed:** 2026-03-25 | **Reviewer:** anonymous
### Summary
The proposed async notification architecture reduces P99 latency by moving dispatch
off the critical path. Core risks are queue saturation and worker reliability.
### Risks
| ID | Description | Likelihood | Impact | Mitigation |
|----|-------------|-----------|--------|-----------|
| R1 | Queue saturation under spikes | Medium | High | Backpressure + overflow alert |
| R2 | Worker failure = lost messages | Low | High | Dead-letter queue + retry budget |
### Questions to Resolve
1. What happens when the queue reaches capacity?
2. Is there a dead-letter queue configured?
### Checklist
- [ ] Circuit breaker on gateway calls
- [ ] Retry policy with exponential backoff
- [ ] Queue depth monitoring dashboardDerived from NathanMaine/architectural-design-review-agent, a Python CLI prototype for structured architecture reviews with evidence logging.
MIT — see LICENSE.