miaoda-app-builder

中文文档

An agent skill for DeepSeek Harness that lets an AI agent drive the Baidu Miaoda (秒哒) platform — a chat-driven full-stack application builder — through natural-language instructions.

Miaoda turns a natural-language description into a production-ready product: frontend UI, backend services, database schema, integrations, and deployable hosting. Typical outputs include websites, web apps, dashboards, SaaS products, admin panels, internal tools, browser/mini games, Weixin mini programs, native iOS/Android apps, research reports, PPTs, and data analysis documents.

Repository Layout

.
├── SKILL.md               # Skill manifest + full usage guide for the agent
├── _meta.json              # Skill registry metadata (slug, version, publish time)
└── scripts/
    └── miaoda_api.py        # Stateless CLI wrapping the Miaoda HTTP API

• SKILL.md — the source of truth. Defines when the skill should trigger, the application lifecycle (create → PRD refine → generate → publish), all CLI commands with parameters and examples, LGUI node-scoped editing, and error-handling guidance.
• scripts/miaoda_api.py — the only supported way to talk to the Miaoda platform. All operations go through this script; direct API calls are disallowed by the skill contract.
• _meta.json — package identity used by the skill registry (ownerId, slug, version, publishedAt).

Requirements

• Python 3
• requests package (pip install requests)
• A Miaoda API key, exposed as the MIAODA_API_KEY environment variable

export MIAODA_API_KEY="sk_xxxxx"

When running inside qianfan-desk without an explicit key, the CLI can fall back to a session-bound Miaoda proxy (DUMATE_SESSION_ID + desktop gateway). Outside that environment, set MIAODA_API_KEY explicitly.

Quick Start

# 1. Create a new app (PRD refinement stage)
python scripts/miaoda_api.py chat --text "创建一个待办事项管理应用"

# 2. Once the trajectory exposes a "Generate App" button, generate it
python scripts/miaoda_api.py generate-app --app-id <app_id> --context-id <context_id> --watch

# 3. Publish to production
python scripts/miaoda_api.py publish --app-id <app_id> --wait

To modify an already-generated app, keep using chat with the same appId/conversationId — do not call generate-app again.

Available Commands

Command	Purpose
list-apps	List applications for the authenticated user
app-detail	Get app details; auto-injects conversationId and appFocus (publish readiness)
get-context-id	Recover a lost conversationId from an app's trajectory
conversation-history	Human-readable summary of past chat turns
chat	Create a new app, refine its PRD, or modify an existing generated app (supports LGUI --node-json for node-scoped edits)
trajectory / fetch-trajectory	Poll or fetch conversation/generation progress events
generate-app	Trigger app generation (once, only after PRD is ready)
publish / publish-status	Deploy to production and poll release status
download-source	Download the full generated app source ZIP
list-task-files	List files in the Agent workspace task directory
download-task-file	Download one Agent workspace task file (report, image, doc, etc.)

Full parameter lists, JSON response shapes, the lifecycle state machine, and troubleshooting (IAM AccessKey validation failed, NotOpenSSLWarning, 秒点不足) are documented in SKILL.md.

Lifecycle Rules (summary)

• New app: chat → PRD refinement → generate-app (once) → publish
• Modify existing app: chat → chat → ... (no re-generation)
• General task (report / PPT / analysis, no app): chat alone; response has needGenerateApp: false
• Always pass --app-id and --context-id together when continuing an existing app — passing --app-id alone silently creates a new app (the CLI now errors on this instead).
• Treat a successful publish (SUCCESS) as terminal; do not re-publish just because downstream UI/events look incomplete.