Cloud Functions

Sibling skills (local only)

Sibling CloudBase skills ship beside this skill. Use local relative paths such as ../auth-tool-cloudbase/SKILL.md.

If a referenced sibling skill file is missing from this environment, ask the user to install the full CloudBase plugin (or the missing skill). Do not HTTP-fetch remote skill or protocol markdown into the agent context.

Cross-cutting protocols (required before code changes or deployments):

• Change Safety Protocol: ../cloudbase-platform/references/protocols/change-safety-protocol.md
• Deployment Gate: ../cloudbase-platform/references/protocols/deployment-gate.md
• Sensitive Runtime Data Protection: ../cloudbase-platform/references/protocols/sensitive-runtime-data-protection.md

Cloud Functions Development

Activation Contract

Use this first when

• The task is to create, update, deploy, inspect, or debug a CloudBase Event Function or HTTP Function that serves application runtime logic.
• The request mentions function runtime, function logs, scf_bootstrap, function triggers, or function gateway exposure.

Read before writing code if

• You still need to decide between Event Function and HTTP Function.
• The task mentions manageFunctions, queryFunctions, manageGateway, or legacy function-tool names.
• The task might require callCloudApi as a fallback for logs or gateway setup.
• An HTTP Function will call CloudBase resources through @cloudbase/node-sdk or @cloudbase/manager-node -> read ./references/http-function-credentials.md. HTTP Functions must use explicit credentials; do not rely on the Event Function passwordless runtime path.

Exception only (do not read by default)

• Migrating an existing app that already uses classic TCP DB clients (DATABASE_URL / Prisma / mysql2 / pg / Redis) → read ./references/vpc-and-tcp-database.md via ./references.md. New business CRUD must prefer CloudBase native SDK (app.database() / app.rdb()) or MCP SQL tools instead of TCP.

Then also read

• Detailed reference routing -> ./references.md
• Auth setup or provider-related backend work -> ../auth-tool-cloudbase/SKILL.md
• CloudBase Integration Center generated WeChat Pay or Official Account functions -> ../cloudbase-wechat-integration/SKILL.md (official docs: https://docs.cloudbase.net/integration/introduce.md)
• AI in functions -> ../ai-model-nodejs/SKILL.md
• Long-lived container services or Agent runtimes -> ../cloudrun-development/SKILL.md
• Calling CloudBase official platform APIs from a client or script -> ../http-api-cloudbase/SKILL.md

Do NOT use for

• CloudRun container services.
• Web authentication UI implementation.
• Database-schema design or general data-model work.
• CloudBase official platform API clients or raw HTTP integrations that only consume platform endpoints.
• Creating Integration Center instances through guessed APIs. For WeChat Pay or Official Account generated functions, use cloudbase-wechat-integration for the business contract and this skill only for function operations.
• Tasks that the CloudBase JS SDK can handle directly — simple data reads/writes, leaderboards, file uploads, real-time queries. Reach for the matching SDK surface before writing a function: db.collection(...).get/add/update only for confirmed NoSQL collections, and app.rdb().from(...) for CloudBase PG tables. Functions add deployment complexity, CORS configuration, and HTTP gateway binding that the SDK eliminates entirely.

Common mistakes / gotchas

• Picking the wrong function type and trying to compensate later.
• Confusing official CloudBase API client work with building your own HTTP function.
• Mixing Event Function code shape (exports.main(event, context)) with HTTP Function code shape (req / res on port 9000).
• Treating HTTP Access as the implementation model for HTTP Functions. HTTP Access is a gateway configuration for Event Functions, not the HTTP Function runtime model.
• Assuming db.collection("name").add(...) will create a missing document-database collection automatically. Collection creation is a separate management step.
• Forgetting that runtime cannot be changed after creation.
• Using cloud functions as the first answer for Web login.
• Forgetting that HTTP Functions must ship scf_bootstrap, listen on port 9000, and include dependencies.
• Assuming an HTTP Function can use CloudBase SDKs without explicit credentials. The default temporary credential path is not reliable for HTTP Functions and credential rotation can break a running service. Use a CloudBase server API Key or Tencent Cloud key pair for @cloudbase/node-sdk; use a Tencent Cloud key pair for @cloudbase/manager-node. See references/http-function-credentials.md.
• Forgetting to configure function security rules after creating an HTTP Function. Default rules reject anonymous callers with EXCEED_AUTHORITY. Note: anonymous login is disabled by default for new environments — if the function needs public access without authentication, configure the security rule to allow all callers rather than relying on anonymous login.
• Mismatching the scf_bootstrap Node.js binary path with the function runtime (e.g. using /var/lang/node18/bin/node but setting runtime: "Nodejs16.13").
• For Custom Image HTTP Functions: forgetting that TCR, the CloudApp build, and SCF must be in the same region; using :latest instead of a unique tag; or confusing the request-driven port-9000 image model with a long-lived CloudRun container that listens on the injected PORT.
• Assuming MCP covers the whole image pipeline. manageFunctions covers SCF image deploy (Stage B) via runtime: "CustomImage" + imageConfig, but the CloudApp custom build → TCR push (Stage A) is a raw Tencent Cloud API path — confirm action names and parameters from official docs before any callCloudApi fallback.
• Making code or configuration changes without first following the Change Safety Protocol (cloudbase-platform/references/protocols/change-safety-protocol.md).
• Exposing functions publicly or deploying without first completing the checks in cloudbase-platform/references/protocols/deployment-gate.md.
• Returning req.headers, process.env, event, or context wholesale — gateways may inject x-cloudbase-context (base64 temporary credentials). Never echo that header or dump credential env vars to clients. Follow ../cloudbase-platform/references/protocols/sensitive-runtime-data-protection.md.
• Using a bare layer name (e.g. common) across environments. SCF LayerName is an account-scoped shared namespace: same name → shared version sequence. Create new layers with fixed format {layerName}_{当前envId} (e.g. common_cloud1-d9ghadgak3edf6b36). Pass the full name as layerName — do not invent automatic suffixes. Treat MCP layer warnings as soft advisories (operation still succeeds). Details: ./references/operations-and-config.md.
• Long-running MCP image deployments must complete the full workflow: When using manageFunctions with deployFunction for a real cloud or local deployment, prefer wait=false to avoid blocking a single Tool Call for an extended period. If the tool returns a taskId, do not end the workflow, report success, or ask the user to wait while the status is running. Automatically call queryFunctions(action="getFunctionDeployStatus", taskId="...") and continue polling according to the reported progress until the status becomes succeeded or failed. Only after reaching a reasonable polling limit may you report that the deployment is still in progress; include the taskId, current stage, and latest progress. On success, report the image URI or build ID, function status, and Gateway URL. On failure, report the failed stage, error code, request ID, and diagnostic guidance. If the status is expired, explain that the local task record exceeded its retention window; the cloud deployment may still be running, so call getFunctionDetail to confirm the actual cloud-side status instead of treating it as a failure.

Minimal checklist

• Read Cloud Functions Execution Checklist before deployment or runtime changes.
• Decide whether the task is Event Function, HTTP Function, or actually CloudRun.
• Pick the detailed reference file in references.md before writing implementation code.

MCP image deployment with polling

For real cloud or local custom-image deployments, prefer:

{
  "action": "deployFunction",
  "dryRun": false,
  "confirm": true,
  "wait": false,
  "deployConfig": {}
}

The wait field controls whether the current MCP Tool call waits for the complete deployment:

• wait=true: wait for the manager deployment to reach a terminal result and return it.
• wait=false: return a taskId promptly while the deployment continues in the MCP background.

When wait=false returns a taskId, the deployment workflow is not complete. Automatically call queryFunctions with action="getFunctionDeployStatus" and that taskId; continue while the status is running, then stop only at succeeded or failed. Wait about 5 seconds before the first follow-up query and use the returned progress/nextActions to continue without aggressive polling. Do not tell the user to ask again or imply success before a terminal status is returned. An expired status means the task exceeded the maximum retention window and was force-terminated locally — the cloud deployment may still be in progress, so confirm the real state with getFunctionDetail instead of reporting failure.

If a reasonable polling limit is reached, report only that the task is still running, including the taskId, current status, current stage, and latest progress. For a terminal result, report the deployment strategy, action, image URI/digest, build ID, function status, Gateway URL, or the failed stage, error code, request ID, and diagnostic next step.

Personal-tier TCR credentials — never put the password in tool arguments

Personal-tier image builds (imageConfig.imageType="personal" with local / cloud) need a TCR push credential. Read it from the MCP process environment, not from tool arguments:

• Leave func.imageConfig.build.registryCredential out of the request when TCB_TCR_USERNAME and TCB_TCR_PASSWORD are set in the MCP server env block — the MCP fills them in automatically, the same way TENCENTCLOUD_SECRETID works.
• Never ask the user to paste the password into chat, and never write it into tool arguments. Anything placed in arguments enters the model context and the tool-call history.
• If deployment fails with CLOUD_REGISTRY_CREDENTIAL_MISSING or CLOUD_REGISTRY_CREDENTIAL_INVALID, instruct the user to add these two variables to the env block of their MCP configuration and restart the MCP server. Do not work around it by passing the credential inline.
• The username is the Tencent Cloud account UIN and is not itself a secret; it may be passed explicitly if needed. Explicit arguments take precedence per field, so username-in-argument plus password-from-environment is a valid combination.

Know when that environment channel does not exist. It works only for a local stdio MCP server whose client configuration exposes a custom env block. Some GUI clients do not inherit shell exports, and IDE-embedded MCP servers usually inject credentials from a hard-coded allowlist (often only TENCENTCLOUD_*), leaving the user no way to set arbitrary variables. Telling those users to "set it in the MCP env block" is an instruction they cannot act on. Route them to an enterprise registry (imageType="enterprise", which mints a short-lived TCR token instead of using a fixed password) or to buildStrategy="image" with an already-pushed image.

Enterprise-tier builds require a login state with CAM permission

cloud / local builds against an enterprise registry mint a TCR token through CAM (as does autoGrant). Environment-level API Keys and OAuth-issued STS credentials carry no CAM policy, so those calls fail with UnauthorizedOperation. The MCP probes the login state before starting a real enterprise build and refuses up front rather than failing midway; treat that error as a routing signal, not a retryable fault:

• Sign in with an account-level TENCENTCLOUD_SECRETID / TENCENTCLOUD_SECRETKEY pair, or
• Switch to buildStrategy="image" and deploy an image that was pushed elsewhere, or
• Use a personal-tier registry — its static password goes straight to docker login without touching CAM, which makes it the one build path that does work for API Key users.

Writing mode at a glance

• If the request is for SDK calls, timers, or event-driven workflows, write an Event Function with exports.main = async (event, context) => {}.
• If the request is for REST APIs, browser-facing endpoints, SSE, or WebSocket, write an HTTP Function with req / res on port 9000.
• For Node.js HTTP Functions, default to the native http module unless the user explicitly asks for Express, Koa, NestJS, or another framework.
• If the HTTP Function needs custom system libraries or an arbitrary runtime but should still be SCF request-driven and scale to zero, deploy it as a Custom Image HTTP Function (Runtime: CustomImage) from a TCR image. The container still listens on the fixed port 9000. See ./references/http-functions-custom-image.md. This is distinct from a CloudRun container, which listens on the injected PORT and runs long-lived.
• 有 Dockerfile 的 HTTP 无状态服务可优先考虑 HTTP 云函数，不必上云托管 — a Dockerfile alone does not mean CloudRun. If the service is stateless, request-driven HTTP without long connections / custom runtime / VPC database access, prefer an HTTP Function (or Custom Image HTTP Function) — faster to deploy, cheaper, and no CloudRun environment initialization needed. Route to CloudRun (../cloudrun-development/SKILL.md) only for WebSocket/SSE long connections, stable independent processes, custom system dependencies, or VPC DB access.
• If the user mentions HTTP access for an existing Event Function, keep the Event Function code shape and add gateway access separately.