alibabacloud-mcp-core-script-generate

name	alibabacloud-mcp-core-script-generate
description	Generate Alibaba Cloud MCP Core RunScript-compatible Python scripts from natural-language cloud operation requests. Use when the user explicitly asks for RunScript scripts, MCP Core scripts, or sandbox-compatible cloud automation code. For standard SDK Python code with imports and credentials, use `alibabacloud-sdk-usage` instead.
license	Apache-2.0
metadata	{"domain":"aliyun-runscript","owner":"sdk-team","contact":"[e-mail]","triggers":"RunScript, RunScript脚本, MCP脚本, 沙箱脚本, generate RunScript, MCP Core script, script recommend\n"}
allowed-tools	mcp__plugin_alibabacloud-core_alibabacloud-core__AlibabaCloud___SearchApis,mcp__plugin_alibabacloud-core_alibabacloud-core__AlibabaCloud___GetApiDefinition,mcp__plugin_alibabacloud-core_alibabacloud-core__AlibabaCloud___ListApis,mcp__plugin_alibabacloud-core_alibabacloud-core__AlibabaCloud___ListProducts

Alibaba Cloud Script Generator

Turn a natural-language Alibaba Cloud request into one RunScript-compatible Python script. The script must use await call_cli(...), assign the final value to result, and pass check_sandbox.py validation before output. Do not execute the script unless the user asks.

Scope Check Before You Start

• RunScript sandbox scripts (this skill): generates call_cli()-based Python for the RunScript sandbox — no SDK imports, no credentials, no local execution.
• SDK project code → alibabacloud-sdk-usage: generates typed/generic SDK code with imports, credential providers, and dependency management for user projects.
• Operational patterns (batch ops, audits, rotations) → alibabacloud-find-skills first.

Workflow

1. Split the request into atomic cloud operations. For each operation, you must call at least one of the following methods to discover API parameters before writing any code:

   Method A — MCP tools (preferred): call AlibabaCloud___SearchApis with a natural language description, then AlibabaCloud___GetApiDefinition with the confirmed product, action, and version. Batch search calls in one parallel round.

   Method B — HTTP metadata (when MCP tools return errors): The endpoint https://next.api.aliyun.com/meta/v1/products/{product}/versions/{version}/api-docs.json contains all API definitions for a product. If WebFetch cannot extract the target API (response too large), download the full JSON with curl and extract the target API's parameters array locally — use any method that works (e.g. python3 -c, grep, jq).

   You must actually execute one of these methods — reading about them is not sufficient. Only if your tool call returns an error may you fall back to best-known parameter names, adding: # WARNING: API parameters not verified — confirm before use.

2. Generate one Python script body following the Sandbox Contract and Script Patterns below. Only whitelisted modules may be imported. When multiple tool calls are needed, batch them in parallel. Never repeat the same tool call with the same arguments.

3. Write to /tmp/aliyun-runscript.py and validate with the local sandbox checker (<SKILL_DIR>/scripts/check_sandbox.py):

   cat > /tmp/aliyun-runscript.py <<'PYEOF'
   <script body here>
   PYEOF
   python3 <SKILL_DIR>/scripts/check_sandbox.py /tmp/aliyun-runscript.py
   

4. If validation fails, read the → fix suggestion for each violation, fix ONLY the listed issues, and re-validate. Maximum 3 rounds. If violations persist, show them to the user.

5. After validation passes, output only the Python script body — no Markdown fences, headings, or explanatory text unless the user asks.

Sandbox Contract

All rules below are enforced by scripts/check_sandbox.py (local pre-check) and remote validation. See references/runscript-contract.md for full definitions with examples.

Category	Rule
Imports	Only whitelisted modules may be imported: asyncio, collections, csv, dataclasses, datetime, decimal, enum, fractions, functools, itertools, json, math, re, statistics, string, time, typing, uuid. Do NOT import random, os, subprocess, requests, or any module not in this list. call_cli is pre-injected — do NOT import or define it.
API calls	Use ONLY call_cli(product, version, action, params). Pre-injected. No SDK clients, no HTTP requests, no subprocess. Parameter values must be plain strings/numbers — do NOT pass aliyun ... CLI command strings as parameter values.
Param values	Parameter values must NOT contain lowercase aliyun (triggers BLK-4001). E.g. use "1.28.3" not "1.28.3-aliyun.1". Query APIs to fetch exact values when unsure.
Output	Assign final data to result (dict or list). No print().
Concurrency	Parallel API calls must use asyncio.gather(return_exceptions=True). Thread/process-based concurrency (threading, multiprocessing, concurrent.futures) is not in the whitelist and will fail validation.
Forbidden	os, subprocess, socket, requests, eval, exec, compile, getattr, setattr, globals, input, breakpoint, __import__, threading, multiprocessing, concurrent.futures, dunder chains.
Blocked APIs	Credential-returning APIs (ram.ListAccessKeys, sts.AssumeRole, kms.GetSecretValue). CLI meta products (configure, plugin, ossutil).
Structure	call_cli must be reachable from module-level execution. Do NOT wrap all calls in an uninvoked async def — either write await call_cli(...) at top level, or define a function and call it: async def main(): ... \n await main().
Numbers	Do NOT use leading zeros in numeric literals (e.g., 01, 010). Write 1, 10 instead. IP addresses and CIDR blocks must be strings: "10.0.0.0/8", not bare numbers.
Strings	Use f-strings or .format() only with constant format strings. Do NOT build call_cli arguments dynamically via str.format(variable) or % formatting with non-constant args.
Sleep	time.sleep() argument MUST be ≤ 30. For longer waits, use a loop: for _ in range(N): time.sleep(30).
Output format	Output ONLY Python code. Do NOT output prose, explanations, or markdown before or after the code. If you must add context, use Python comments inside the script — but avoid comments unless absolutely necessary.
Other	Write/delete/update ops execute directly — the RunScript runtime intercepts write operations and presents them to the user for approval (HITL) before execution. The script itself should not add confirmation prompts.

Script Patterns

Sequential (dependent calls):

vpc = await call_cli(product="Vpc", version="2016-04-28", action="CreateVpc", params={"RegionId": region_id, "CidrBlock": "10.0.0.0/8"})
vsw = await call_cli(product="Vpc", version="2016-04-28", action="CreateVSwitch", params={"VpcId": vpc["VpcId"], "CidrBlock": "10.0.0.0/16", "ZoneId": "cn-hangzhou-a"})
result = {"VpcId": vpc["VpcId"], "VSwitchId": vsw["VSwitchId"]}

Pagination (list tasks):

items, page = [], 1
while True:
    resp = await call_cli(product="Ecs", version="2014-05-26", action="DescribeInstances", params={"RegionId": region_id, "PageNumber": page, "PageSize": 100})
    batch = resp.get("Instances", {}).get("Instance", [])
    items.extend(batch)
    if len(batch) < 100:
        break
    page += 1
result = items

Parallel (independent calls):

responses = await asyncio.gather(*[
    call_cli(product="Ecs", version="2014-05-26", action="DescribeInstances", params={"RegionId": rid})
    for rid in region_ids
], return_exceptions=True)
result = {rid: r if isinstance(r, dict) else {"error": str(r)} for rid, r in zip(region_ids, responses)}

Guardrails

• Do not execute the generated script unless the user explicitly asks.
• Do not use any MCP server except the local alibabacloud-core server for API discovery.
• Do not generate boilerplate, redundant error handling, or unused imports.
• Always validate before output. Do not skip validation.
• If the request is ambiguous but not dangerous, use placeholders instead of asking. Use the pre-injected runtime variable (e.g. region_id) or a string placeholder like "<your-vpc-name>".

References

• references/runscript-contract.md — full rule definitions and additional patterns. Read only when validation fails repeatedly.