alibaba-expert

name	alibaba-expert
description	Alibaba Cloud expert guardrails — RAM least-privilege, data boundaries, residency (mainland/international split), cost, and official-source validation. Trigger when designing or reviewing any Alibaba Cloud workload, especially agents or LLMs (Model Studio/Bailian, Qwen, AgentScope).

Alibaba Cloud Expert

Same discipline, different control plane: RAM and ActionTrail are the guardrails; Model Studio and AgentScope are the agent surface. The mainland/international split is a design input, not an afterthought.

This skill enforces the discipline that makes Alibaba Cloud workloads production-safe: identity, data boundaries, cost controls, and residency. It is not a feature tour — it is a checklist of the things that cause incidents and compliance failures when skipped.

When to use

• Designing any Alibaba Cloud infrastructure (new or modified)
• Before deploying agents or LLM workloads (Model Studio / Bailian, DashScope, PAI, AgentScope)
• When reviewing a Terraform plan or architecture for an Alibaba Cloud workload
• When a system touches regulated data or spans mainland-China and international regions

Procedure

1. Identity and RAM — verify least-privilege for every principal:

   • RAM roles with STS temporary credentials over long-lived AccessKey pairs. No AccessKeys in code or config.
   • Policies scoped to the specific function; resource groups used to bound blast radius.
   • ActionTrail enabled on all account activity and shipped to a central audit store.

2. Data boundaries — for every data store:

   • Classification recorded (public / internal / confidential / regulated — note PIPL and Data Security Law categories where in scope).
   • OSS bucket policies explicit; public access blocked; KMS CMK where required.
   • Cross-account sharing explicit and documented.

3. Data residency — for each resource:

   • Mainland-China vs international accounts are separate control planes — decide which side the workload and its data live on; regulated mainland data stays in mainland regions.
   • For Model Studio / DashScope calls: confirm the endpoint region matches the residency requirement.

4. Cost controls — for every LLM, compute, or storage resource:

   • Budget alerts configured (User Center budgets) at staged thresholds.
   • DashScope token quotas and rate limits set; unbounded agent loops are unbounded spend.
   • Auto-scaling maximums bounded.

5. Network and egress — confirm:

   • PrivateLink / PrivateZone used where public endpoints are avoidable.
   • Security groups default-deny; egress costed for cross-border traffic (cross-border egress is materially expensive).

6. Observability — confirm:

   • CloudMonitor dashboards and alerts on error rate, latency, and cost.
   • SLS (Log Service) retention configured for audit — evidence before go-live.

7. Run the Adversarial Gate — common Alibaba Cloud failure modes: long-lived AccessKeys, public OSS buckets, mainland/international residency confusion, uncapped DashScope spend, missing ActionTrail.

Official sources — validate before you assert

• Documentation: alibabacloud.com/help (international) · help.aliyun.com (mainland)
• Live docs via MCP: official MCP coverage is thinner here than the other vendors — verify what exists before pinning; spring-ai-alibaba ships MCP integrations (github.com/alibaba/spring-ai-alibaba).
• GitHub, foundations: github.com/aliyun (incl. terraform-provider-alicloud and landing-zone modules)
• GitHub, agent examples: github.com/modelscope/agentscope · github.com/alibaba/spring-ai-alibaba · github.com/QwenLM
• Rule: every service/API claim cites an official doc. Service names, quotas, and model IDs are dated facts — stale until re-verified against the source.

Outputs

• Alibaba Cloud guardrail checklist (pass/fail per item)
• RAM matrix: principal | role | scope | justification
• Data classification and boundary map (incl. residency side)
• Budget alert confirmation
• Open findings for human review

Guardrails

• No long-lived AccessKeys in workloads. STS or nothing.
• Residency is a constraint, not a preference. The mainland/international split is decided in design, recorded in the ADR.
• Budget alerts are not optional. An unmonitored LLM workload will produce a surprise invoice.
• ActionTrail is evidence. Enable it before go-live, not after an incident.