Cloudbase

CloudBase Development Guidelines

📁 Reference Files Location

All reference documentation files are located in the references/ directory relative to this file.

File Structure:

cloudbase/
├── SKILL.md              # This file (main entry)
└── references/           # All reference documentation
    ├── auth-web/SKILL.md       # Web authentication guide
    ├── auth-wechat/SKILL.md    # WeChat authentication guide
    ├── no-sql-web-sdk/SKILL.md # NoSQL database for Web
    ├── ui-design/        # UI design guidelines
    └── ...               # Other reference docs

How to use: When this document mentions reading a reference file like references/auth-web/SKILL.md, simply read that file from the references/ subdirectory.

---

Workflow

Every CloudBase task follows this three-stage process:

1. Exploration  →  Read the matching skill completely before writing any code.
                   Search for it with searchKnowledgeBase(mode="skill"), then
                   Read the full SKILL.md content. Do not rely on search summaries.
2. Implementation
   ├── 2a. Resource preparation → Use MCP tools to prepare backend resources
   │     (enable auth providers, create database tables, configure storage,
   │      set up security rules — before writing any frontend code)
   └── 2b. Frontend implementation → Write code, install deps, start server, test
3. Close-out  →  Run cloudbase-code-review, fix errors, declare done

Key constraints:

• Stage 2a (resource preparation) must precede frontend code. Always use MCP tools.
• Stage 3 is mandatory. The close-out includes automated lint checks and manual LLM review. Do not skip it.

Activation Contract

Read this section first. The routing contract uses stable skill identifiers such as auth-tool, auth-web, and http-api, so it works across source files, generated artifacts, and local installs.

Standalone skill fallback

If the current environment only exposes a single published skill, start from the CloudBase main entry:

• CloudBase main entry: https://cnb.cool/tencent/cloud/cloudbase/cloudbase-skills/-/git/raw/main/skills/cloudbase/SKILL.md
• Sibling skill pattern: https://cnb.cool/tencent/cloud/cloudbase/cloudbase-skills/-/git/raw/main/skills/cloudbase/references//SKILL.md

When a skill body references stable sibling ids such as auth-tool, auth-web, ui-design, or web-development, replace `` with that published directory name to open the original file.

If a skill points to its own references/... files, keep following those relative paths from the current skill directory. If the environment does not support MCP directly, read cloudbase first and follow its mcporter / MCP setup guidance before using any platform-specific skill.

Global rules before action

• Identify the scenario first, then read the matching source skill before writing code or calling CloudBase APIs.
• Prefer semantic sources when maintaining the toolkit, but express runtime routing in stable skill identifiers rather than repo-only paths. Do not treat generated, mirrored, or IDE-specific artifacts as the primary knowledge source.
• Use MCP or mcporter first for CloudBase management tasks, and inspect tool schemas before execution.
• If the task includes UI, read ui-design first and output the design specification before interface code.
• If the task includes login, registration, or auth configuration, read auth-tool first and enable required providers before frontend implementation.
• Keep auth domains separate: management-side login uses auth; app-side auth configuration uses queryAppAuth / manageAppAuth.

Universal guardrails

• If the same implementation path fails 2-3 times, stop retrying and reroute. Re-check the selected platform skill, runtime, auth domain, permission model, and SDK boundary before editing more code.
• Always specify EnvId explicitly in code, configuration, and command examples when initializing CloudBase clients or manager operations. Do not rely on the current CLI-selected environment, implicit defaults, or copied local state.
• When saving MCP or tool results to a local file with a generic file-writing tool, pass text, not raw objects. For JSON output files, serialize first with JSON.stringify(result, null, 2) and write that string as the file content.
• If the file-writing tool reports that a field such as content expected a string but received an object, do not retry with the same raw object. Serialize the object first, then retry once with the serialized text, and make sure the retried call actually passes the serialized string rather than the original object.
• Keep scenario-specific pitfall lists in the matching child skills instead of expanding this entry file.
• For new frontend projects, the first deployment must use manageApps(action="createApp", ...) to deploy to an independent subdomain. manageHosting is prohibited for first-time deployments — it is only allowed for incremental updates to existing projects that were originally deployed via manageHosting.

Engineering constitution (applies to every scenario)

These rules override convenience. They are a gate before saying "done". Full rationale + replacements live in web-development (Engineering constitution section).

• Prepare backend resources via MCP before writing frontend code. Auth providers, database tables, storage domains, and security rules must be set up through MCP tools before writing any frontend code that depends on them. Writing frontend code against non-existent resources will cause grader failures. This applies to every scenario — auth, database, storage, functions, CloudRun.
• Do NOT use any to bypass type errors. Not : any, not as any, not @ts-ignore, not @ts-nocheck. Use unknown + a type guard, a precise interface, or declare module augmentation instead. any propagates silently and defeats the compile-time safety net.
• Self-verify before claiming done. Static layer (tsc --noEmit / lint / project build / unit tests) and runtime layer (use agent-browser to exercise user-visible flows when the change touches routing, rendering, forms, auth, or async UI). "It should work" without evidence is not acceptable. If a layer cannot be run locally, name the gap explicitly.
• Do not paper over failures. No empty try/catch to silence bugs, no skipping / deleting failing tests to make CI green, "it compiles" is not "it works".
• ai.createModel(...) / wx.cloud.extend.AI.createModel(provider) argument is a GroupName, not a vendor / model id. Only three legal shapes: "cloudbase" (default, TokenHub-backed managed pool), "hunyuan-exp" (only if DescribeAIModels returns it, mainly Mini Program Growth Plan), or "custom-" (user-defined via CreateAIModel, must start with custom-). The concrete model id (deepseek-v4-flash, hunyuan-2.0-instruct-20251111, kimi-k2.6, …) goes into the model field of generateText / streamText, never into createModel(...). See ai-model-web / ai-model-nodejs / ai-model-wechat for the full STOP card.
• Low-capability-model STOP card: Do not paste legacy CloudBase NoSQL or old-auth snippets from memory. If the task says PostgreSQL, CloudBase PG, PG mode, app.rdb(), queryPgDatabase, managePgDatabase, PostgREST, or RLS, route to postgresql-development and write app.rdb().from(...) / queryPgDatabase / managePgDatabase; do not write app.database(), db.collection(...), .where(), .orderBy(), app.uploadFile(), queryMysqlDatabase, or manageMysqlDatabase for that PG data path. For Web auth guards, use auth.getSession() and require data.session; do not use deprecated getLoginState() or auth.getUser() as proof of login.

High-priority routing

| Scenario | Read first | Then read | Do NOT route to first | Must check before action | |----------|------------|-----------|------------------------|--------------------------| | Web login / registration / auth UI | auth-tool | auth-web, web-development | cloud-functions, http-api | Provider status and publishable key | | WeChat mini program + CloudBase | miniprogram-development | auth-wechat, no-sql-wx-mp-sdk | auth-web, web-development | Whether the project really uses CloudBase / wx.cloud | | Native App / Flutter / React Native | http-api | auth-tool, relational-database-tool | auth-web, no-sql-web-sdk, web-development | SDK boundary, OpenAPI, auth method | | Web projects + NoSQL Database | web-development | no-sql-web-sdk, auth-web | relational-database-tool, http-api | Login state and database access permission model | | CloudBase PostgreSQL / PG | postgresql-development | auth-tool, auth-web, web-development, miniprogram-development, cloud-storage-web, http-api | relational-database-tool, no-sql-web-sdk | PG schema, usernamePassword login, backend/RLS permission model | | MySQL Database (relational) | relational-database-tool | relational-database-web, http-api | no-sql-web-sdk, web-development | Distinguish MCP management vs app code access | | Cloud Functions | cloud-functions | auth-tool, ai-model-nodejs | cloudrun-development, auth-web | Event vs HTTP function, runtime, scf_bootstrap | | CloudRun backend | cloudrun-development | auth-tool, relational-database-tool | cloud-functions | Container boundary, Dockerfile, CORS | | AI Agent (智能体开发) | cloudbase-agent | cloud-functions, cloudrun-development | cloud-functions, cloudrun-development | AG-UI protocol, scf_bootstrap, SSE streaming | | UI generation | ui-design | web-development, miniprogram-development | cloud-functions | Design specification first | | AI Model (Web) | web-development | ai-model-web, ui-design | ai-model-wechat, http-api | Platform and streaming interaction mode | | AI model call (大模型调用 / 文本生成 / 图片生成 / 流式对话) | ai-model-web | ai-model-nodejs, ai-model-wechat | cloudbase-agent, cloud-functions, cloudrun-development | 先跑「调用前必须的资格检查」：DescribeActivityInfo（小程序成长计划） + DescribeEnvPostpayPackage（Token Credits 资源包） | | Resource health inspection / troubleshooting | ops-inspector | cloud-functions, cloudrun-development | ui-design, spec-workflow | CLS enabled, time range for logs | | Spec workflow / architecture design | spec-workflow | cloudbase | web-development, cloud-functions | Requirements, design, tasks confirmed |

Activation triggers (derived from references/activation-map.yaml)

• Web login / registration / auth UI — CloudBase Web 登录, Web 注册, auth login page, publishable key, 短信登录, 邮箱登录
• WeChat mini program + CloudBase — 小程序 云开发, wx.cloud, mini program cloudbase, OPENID, 小程序数据库
• Native App / Flutter / React Native — Android CloudBase, iOS CloudBase, Flutter CloudBase, React Native CloudBase, 原生 App 接入
• Web projects + NoSQL Database — Web 文档数据库, CloudBase collection, 前端查库, NoSQL Web SDK
• CloudBase PostgreSQL / PG — CloudBase PG, PostgreSQL, Postgres, PG 模式, JS SDK v3 PostgreSQL, app.rdb(), queryPgDatabase, managePgDatabase, mysqldb OpenAPI, PostgREST, RLS, service_role, auth schema, storage schema, pgvector
• MySQL Database (relational) — MySQL 建表, executeWriteSQL, security rule, CloudBase 关系型数据库管理
• Cloud Functions — 创建云函数, HTTP 云函数, getFunctionLogs, scf_bootstrap, runtime
• CloudRun backend — CloudRun 部署, 云托管, container backend, Dockerfile
• AI Agent (智能体开发) — AI Agent, 智能体, 智能体开发, AG-UI protocol, LangGraph, LangChain, CrewAI, streaming agent, agent UI
• UI generation — 设计页面, 登录页 UI, frontend interface, 组件样式, prototype
• AI Model (Web) — Web AI 对话, CloudBase AI 流式输出, Web 集成模型
• AI model call (大模型调用 / 文本生成 / 图片生成 / 流式对话) — 大模型调用, AI 模型调用, generateText, streamText, generateImage, 文本生成, 图片生成, 流式对话, hunyuan-exp, deepseek-v4-flash, Token Credits 资源包, 小程序成长计划, aiminiprograminspire_plan, callCloudApi AI 模型, CreateAIModel
• Resource health inspection / troubleshooting — 巡检, 诊断, health check, 资源健康, 异常日志, error inspection, troubleshooting, 错误排查
• Spec workflow / architecture design — 需求文档, 技术方案, tasks.md, Spec 工作流

Routing reminders

• Web auth failures are usually caused by skipping provider configuration, not by missing frontend code snippets.
• Native App failures are usually caused by reading Web SDK paths, not by missing HTTP API knowledge.
• Mini program failures are usually caused by treating wx.cloud like Web auth or Web SDK.
• CloudBase PG failures are usually caused by falling back to MySQL/NoSQL routing, skipping username-password auth readiness, guessing raw HTTP paths instead of using JS SDK v3 app.rdb() / documented mysqldb OpenAPI, or leaving backend/RLS permissions as frontend-only checks.
• AI 大模型调用失败通常是资源包未开通或小程序成长计划未报名，不是 SDK 用错；先跑 DescribeEnvPostpayPackage / DescribeActivityInfo 资格检查，再去改代码。小程序端优先判成长计划，Web / Node.js 端优先判 Token Credits 资源包。

Web SDK quick reminder

• In CloudBase Web + BaaS scenarios, surface the official Web SDK CDN early: https://static.cloudbase.net/cloudbase-js-sdk/latest/cloudbase.full.js
• For React, Vue, Vite, Webpack, and other modern frontend projects, prefer npm install @cloudbase/js-sdk
• For static HTML, no-build demos, README snippets, or low-friction prototypes, the CDN form is acceptable
• Read web-development first for Web SDK integration, then auth-web when login or session handling is involved

⚠️ Prerequisite: MCP Must Be Configured

CloudBase MCP (Model Context Protocol) is REQUIRED before using any CloudBase capabilities. Without MCP, you cannot manage environments, deploy functions, operate databases, or perform any CloudBase management tasks.

> ⚠️ Do not skip this section. If MCP is not configured, every later step (env query, deploy, database operations, function updates) will fail. Always verify MCP availability first with npx mcporter list | grep cloudbase or the IDE's MCP panel before calling any CloudBase tool.

Approach A: IDE Native MCP

Configure CloudBase MCP via your IDE's MCP settings. For detailed config examples (Cursor, Claude Code, Windsurf, Cline, etc.), see references/mcp-setup.md.

Approach B: mcporter CLI (fallback for IDEs without native MCP)

When your IDE does not support native MCP, you MUST configure CloudBase via mcporter. Do not skip this and attempt to call CloudBase tools directly — they will not be available.

Step 1 — Check: npx mcporter list | grep cloudbase

Step 2 — Configure (if not found): create config/mcporter.json in the project root with this minimal content:

{
  "mcpServers": {
    "cloudbase": {
      "command": "npx",
      "args": ["@cloudbase/cloudbase-mcp@latest"],
      "description": "CloudBase MCP",
      "lifecycle": "keep-alive"
    }
  }
}

Step 3 — Verify: npx mcporter describe cloudbase --all-parameters

For the full command reference (list, call, schema, app-auth flows, etc.), see references/mcp-setup.md.

Important Rules

• When managing or deploying CloudBase, you MUST use MCP and MUST understand tool details first. Before calling any CloudBase tool, run npx mcporter describe cloudbase --all-parameters (or ToolSearch in IDE) to inspect available tools and their parameters.
• You do not need to hard-code Secret ID / Secret Key / Env ID in the config. CloudBase MCP supports device-code based login via the auth tool, so credentials can be obtained interactively instead of being stored in config.
• When the environment identifier in the conversation is an alias, nickname, or other short form, do not pass it directly to auth.set_env, SDK init, console URLs, or generated config files. First resolve it to the canonical full EnvId with envQuery(action=list, alias=..., aliasExact=true). If multiple environments match or no exact alias exists, stop and clarify with the user.

Pricin

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: TencentCloudBase
• Source: TencentCloudBase/cloudbase-skills
• License: MIT

Install and usage instructions live in the source repository linked above.