AgentSkillsCN

build-application

从零开始构建全新应用。由编排器负责审批方案——各子代理将直接进入执行环节。同时包含部署后的强制性验证流程。

SKILL.md
--- frontmatter
name: build-application
description: Build a new application from scratch. The orchestrator handles plan approval — subagents proceed directly with execution. Includes mandatory post-deploy verification.
allowed-tools: execute, create_database, build_and_push

Build Application — Full Lifecycle

When dispatched to BUILD or SCAFFOLD a new application, follow this workflow. The orchestrator has already obtained user approval for the plan.

Step 1: Check for approved plan context

Look at the top of your prompt for an --- APPROVED PLAN --- section. If present, the user has already approved the plan — proceed directly. If absent, infer the plan from your task description and proceed.

Do NOT call ask_human. The orchestrator handles all user interaction.

Step 2: Gather requirements (internal — no tool calls)

From your task prompt, determine:

  • •App name, URL, features
  • •Stack: backend, frontend, database
  • •Container port

Step 3: Scaffold the code

Create all necessary source files in /workspace/<app-name>/:

  • •Application code (backend + frontend)
  • •Dockerfile
  • •Requirements/dependencies file
  • •Database migration or init logic

Use write_file to generate each file. Choose sensible defaults for anything not specified in the task prompt.

Step 4: Provision database (if needed)

If the app needs a database, call create_database with the app name. This provisions a dedicated PostgreSQL container on the NAS. Capture the returned connection string for use in deployment.

Do NOT inspect the NAS manually for existing databases. Always use create_database — it handles deduplication internally.

Step 5: Build and push Docker image

Use build_and_push with:

  • •context_path: the workspace directory (e.g., /workspace/todo-app)
  • •image_name: the app name (e.g., todo-app)
  • •app_port: the port the app listens on

If the build fails, read the error, fix the code, and retry.

Step 6: Deploy to Kubernetes

Create K8s manifests (Deployment, Service, Ingress) and apply with kubectl. Wire in the database connection string from Step 4 as environment variables. Use localhost:32000/<image>:<tag> for the image reference.

Step 7: Verify EVERYTHING works (MANDATORY — do NOT skip)

After deployment, run verification tests:

  1. •Pod health: all pods Running, 0 restarts
  2. •Database: tables exist, connectivity works
  3. •API smoke tests: call main endpoints (GET list, POST create, GET verify)
  4. •Frontend: returns HTML (if applicable)
  5. •Ingress: external URL responds with 200

Only report success if ALL checks pass. If any fails, fix and re-verify. Follow the verify-deployment skill for the full checklist.

NEVER

  • •Do NOT call ask_human — the orchestrator handles user interaction
  • •Do NOT ask questions in your response text — the user cannot reply
  • •Do NOT declare success without running verification tests
  • •Do NOT skip database migration/table creation before verifying the API
  • •Do NOT manually inspect the NAS for databases — use create_database
  • •Do NOT use docker build/push via execute — use build_and_push