alibabacloud-cadt-arch-draw

CADT Architecture Diagram Drawing

Submit a product/resource description → CADT AI agent creates the architecture (nodes, connections, validation, layout) and returns a Markdown summary.

Prerequisites

aliyun version              # >= 3.3.3
aliyun configure list       # valid profile exists

If not met, see references/related-commands.md § Installation.

Security: NEVER read/echo AK/SK values, NEVER ask user to input credentials in conversation.

---

Flow

Step 1: Submit Drawing Request (SendMessage)

aliyun bpstudio execute-operation-sync \
  --service-type ai_agent \
  --operation SendMessage \
  --region cn-hangzhou \
  --user-agent "AlibabaCloud-Agent-Skills/alibabacloud-cadt-arch-draw/{SESSION_ID}" \
  --attributes '{
    "prompt": "一台 Web ECS + 一台 MySQL RDS + 一台 SLB，部署在同一个 VPC 内"
  }'

• prompt (required): product/resource list — be specific about products, quantities, and relationships
• sessionId (optional): omit for new session; pass to iterate on existing architecture
• scene (optional): reserved for future drawing modes — currently no available values, omit

Session reuse (MANDATORY): If the user provides a sessionId (e.g. "sessionId 是 1001", "在上次那个架构基础上改", "继续修改刚才的图"), you MUST pass that exact sessionId in this SendMessage call to iterate on the existing CADT session. NEVER start a fresh session and NEVER regenerate the architecture locally. A new execution environment does NOT lose the session — the session lives on the CADT backend, keyed by sessionId; just pass it through the CLI.

Response:

{
  "Code": 200,
  "Data": {
    "Status": "SUCCESS",
    "Arguments": { "triggered": true, "sessionId": 1001, "requestId": "xxx-xxx" }
  }
}

Architecture generation typically takes 30–120 seconds depending on complexity.

Step 2: Poll for Result (ListMessage)

aliyun bpstudio execute-operation-sync \
  --service-type ai_agent \
  --operation ListMessage \
  --region cn-hangzhou \
  --user-agent "AlibabaCloud-Agent-Skills/alibabacloud-cadt-arch-draw/{SESSION_ID}" \
  --attributes '{"sessionId": 1001}'

Response structure:

{
  "Data": {
    "Status": "SUCCESS",
    "Arguments": {
      "hasMore": false,
      "hasNext": false,
      "lastMessageId": 5002,
      "messageList": [
        {
          "messageId": 5001,
          "sessionId": 1001,
          "senderType": "customer",
          "senderId": "120012345678****",
          "content": "...",
          "status": "done",
          "messageSendTime": 1700000000000
        },
        {
          "messageId": 5002,
          "sessionId": 1001,
          "senderType": "robot",
          "senderId": "robot",
          "content": "...Markdown architecture summary...",
          "status": "done",
          "prompt": "...",
          "messageSendTime": 1700000090000
        }
      ]
    }
  }
}

Polling logic:

• Find the last message where senderType == "robot"
• If its status == "ongoing" → wait 5s, poll again
• If its status == "done" → proceed to Step 3
• If messageList is empty / has no robot message → keep polling with the same sessionId (this alone is NOT a reason to switch sessions)
• Max poll: 30 times (~150s), then timeout → stop and report to the user

On timeout or persistently empty response: NEVER create a new session as a workaround. Report to the user that session {sessionId} returned no response (it may not exist, may have expired, or may belong to another environment) and ask how to proceed. Creating a new session without the user's explicit consent violates Hard Rule #6.

Multi-turn: The robot may ask clarifying questions (e.g., "部署在哪个地域?") instead of directly drawing. In that case, status == "done" but content is a question — present it to the user, collect their reply, and call SendMessage again with the same sessionId.

Invalid input: Nonexistent products/regions (e.g. a made-up product name or a fake region like cn-nowhere) do NOT return FAILURE — the robot replies with a clarifying question (Case B) pointing out what it cannot recognize. Relay that clarification faithfully; NEVER fabricate a successful architecture.

Step 3: Deliver Result

When the robot's last message has status == "done", check its content:

Case A — Architecture complete (content contains resource table / topology / validation):

1. Present the robot's Markdown content to the user
2. Inform user the architecture has been created in CADT and can be viewed/edited at CADT Console
3. Offer to continue modifying (reuse the same sessionId)

Case B — Clarifying question (content is a question asking for more info):

1. Present the question to the user
2. Collect user's reply
3. Call SendMessage again with the same sessionId and the reply as prompt
4. Go back to Step 2

---

Prompt Tips

Error Handling

Retry policy: Retry is a conditional edge-case that applies only to the Failed to invoke backend-unavailable FAILURE. A SUCCESS response or any business-error FAILURE MUST NOT trigger any retry.

Observability

All aliyun CLI calls in this skill MUST include --user-agent for tracing.

UA template (placeholder form):

--user-agent AlibabaCloud-Agent-Skills/alibabacloud-cadt-arch-draw/{SESSION_ID}

UA actual example (runtime form):

--user-agent AlibabaCloud-Agent-Skills/alibabacloud-cadt-arch-draw/a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6

Session-id rules:

• If env var SKILL_SESSION_ID is set, use its value directly
• Otherwise, generate a 32-char lowercase hex string (uuid4().hex) once at skill session start
• Reuse the same session-id across all CLI calls within one user interaction session

---

Hard Rules

1. All calls via aliyun bpstudio execute-operation-sync — no raw HTTP
2. uid is auto-injected by gateway — never pass in attributes
3. Poll by checking robot message status, not a top-level ongoing field
4. Present the robot's Markdown content faithfully — do not fabricate results
5. NEVER generate architecture diagrams locally (no local .md/.html/image files, no drawing from scratch). The architecture is ALWAYS produced by the CADT backend via aliyun CLI — your only job is to call the API and relay its output.
6. When the user provides a sessionId, you MUST reuse it by passing it to SendMessage — never create a new session or start over. A new/restarted execution environment does NOT invalidate the session; it persists on the CADT backend keyed by sessionId. Even on polling timeout or an empty response, do NOT switch to a new session — stop and report to the user; only the user may authorize a new session.
7. This skill is architecture-creation only — no resource provisioning or deployment
8. Every aliyun call MUST carry --user-agent as specified in Observability section