Agent Requirement System

Agent is a multi-agent system that turns a natural-language business requirement into a delivery package. It clarifies the requirement, designs a solution, produces a technical design, reviews the result, and exports the files. The web application is branded Agent. The API title is Multi-Agent Requirement Delivery System.

• Accepts a requirement and runs it as an asynchronous task.
• Splits the work across planner, solution, engineer, and reviewer roles.
• Pauses for human clarification when too many questions remain open.
• Waits for human approval after review, then exports the deliverables.
• Shows failure reasons, the PRD, the technical design, and the review on the task page.

Deliverables are written to output/<task-id>/. Task state is stored in SQLite.

normalize input → plan → evaluate plan
    → human clarification, when many questions remain
    → solution → engineering → review
    → automatic repair and another review, when review fails
    → human approval → package output

During plan evaluation, answer open questions in the clarification box. That step is not final approval. The approve and reject actions appear after review has passed.

The engineer calls an OpenAI-compatible API with structured output. If a response hits the length limit, the same conversation continues and the truncated JSON is repaired. OPENAI_MAX_TOKENS sets the size of one completion, default 8192. STRUCTURED_OUTPUT_MAX_CONTINUATIONS sets how many continuation rounds are allowed, default 3. A very large request can still fail after those rounds. State the goal, core features, and constraints, and avoid asking for a complete codebase in one task.

Layout:

app/                 Backend agents, workflow, API, and settings
frontend-vue/        Vue application
data/                SQLite database and checkpoints
output/              Task deliverables
tests/               Backend tests
.env.example         Environment template
docker-compose.yml   Container compose file
Dockerfile           Backend image
start.bat            Windows launcher

In local development, open http://127.0.0.1:5173. Unauthenticated visits go to the login page.

The demo account is checked only in the browser:

• Username: admin
• Password: admin

The session is stored in sessionStorage for the current tab. Closing the tab requires another login. Use 退出 in the header to sign out.

After login, the app provides a dashboard, task creation, task history, and task detail. The create page includes templates for authentication, commerce, education, project management, medical consultation, and finance/tax. A selected template can still be edited.

Required:

• Python 3.11 or 3.12
• Node.js 20 or newer
• An OpenAI-compatible model endpoint and API key

Docker deployment also needs Docker Engine and Docker Compose. The backend image uses Python 3.11. The frontend build image uses Node.js 24.

From the project root:

cp .env.example .env
mkdir -p data output

Set at least:

OPENAI_API_KEY=your-api-key
OPENAI_API_BASE=https://api.deepseek.com
OPENAI_MODEL_NAME=deepseek-chat

Replace the base URL and model name with the provider you use. The values above are only an example of a compatible endpoint. Do not commit .env or share it.

Other settings:

Restart the backend after changing .env. The model client caches settings, so saving the file alone does not apply them.

Run these commands from the project root unless noted.

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

On Windows, a Conda environment named multi_agent_v1 can be used before pip install -r requirements.txt.

cd frontend-vue
npm ci

If the lockfile does not match package.json, npm ci fails. In that case run npm install inside frontend-vue.

Terminal 1, from the project root:

source .venv/bin/activate
python -m uvicorn app.main:app --host 0.0.0.0 --port 8000

Terminal 2:

cd frontend-vue
npm run dev -- --host 0.0.0.0 --port 5173

On Windows, run start.bat backend and start.bat frontend from the project root.

• Application: http://127.0.0.1:5173
• API docs: http://127.0.0.1:8000/docs
• Health: http://127.0.0.1:8000/health, which returns {"status":"ok"}

The Vite dev server proxies /api to http://127.0.0.1:8000 and strips the /api prefix. Start the backend before using the UI.

Create .env in the project root, then run:

docker compose up --build

On Windows, the matching command is start.bat docker.

After startup:

• Application: http://127.0.0.1:8080
• API: http://127.0.0.1:8000
• Health: http://127.0.0.1:8000/health

Nginx serves the built frontend. Browser requests to /api are forwarded to the api:8000 service. The data/ and output/ directories are mounted from the host, so tasks and deliverables remain in the project directory.

Stop the stack:

docker compose down

The login page protects frontend routes only. These API endpoints do not check the demo account.

With the virtual environment active, from the project root:

pytest tests -v

1. Sign in with admin / admin.
2. Pick a template or write the requirement, including the goal, core features, and constraints.
3. When the task is waiting for clarification, answer in the clarification area and submit.
4. After review passes, approve the result. Rejection sends the task back to solution design.
5. Read the documents on the task page, or open output/<task-id>/.
6. If a task fails, the detail page shows the failed step and the reason. Narrow the requirement and create a new task.