name: system-architecture description: Design multi-service architectures with RAW workflows as components

System Architecture Patterns

Use this skill when the user's requirements span multiple days, require webhooks, scheduling, or long-running state.

When to Build a System (vs Single Workflow)

Single RAW Workflow

Use for:

• Data processing tasks
• API integrations (fetch → transform → save)
• Report generation
• One-time executions

Multi-System Architecture

Use when you need:

• Long-running state (track entities across days/weeks)
• Multiple entry points (webhooks, scheduled triggers)
• Event-driven coordination (webhook → workflow → callback)
• Process orchestration (multi-step processes with waiting)

Architecture Components

1. RAW Workflows (The Workers)

RAW workflows are pure functions that:

• Take inputs
• Process data (API calls, transformations, LLM operations)
• Return outputs
• Are stateless and reusable

Example:

# .raw/workflows/analyze-cv/run.py
class AnalyzeCVWorkflow(BaseWorkflow[CVParams]):
    @step("analyze")
    def analyze(self) -> dict:
        cv_text = self.params.cv_text
        score = self.llm.score_candidate(cv_text)
        return {"score": score, "qualified": score > 0.7}

2. FastAPI (The Coordinator)

FastAPI app provides:

• Webhooks for external triggers (form submissions, calendar events)
• REST API for status queries
• Database connection for state persistence
• Calls to RAW workflows via subprocess or Python imports

Example:

# api/main.py
from fastapi import FastAPI
import subprocess

app = FastAPI()

@app.post("/webhooks/cv-submission")
async def handle_cv_submission(cv: CVSubmission):
    # Save to database
    candidate = await db.candidates.create(cv)

    # Run RAW workflow for analysis
    result = subprocess.run([
        "raw", "run", "analyze-cv",
        "--cv-text", cv.text
    ], capture_output=True)

    # Update state based on result
    await update_candidate(candidate.id, result)

    return {"status": "processing"}

3. Temporal (The Orchestrator)

Temporal workflows manage:

• Multi-day processes
• Reliable scheduling (with retries)
• Durable state across service restarts
• Long-running processes with human interaction

Example:

# workflows/temporal/hiring_process.py
@workflow.defn
class HiringProcess:
    @workflow.run
    async def run(self, candidate_id: str) -> str:
        # Step 1: Analyze CV (RAW workflow)
        analysis = await workflow.execute_activity(
            run_raw_workflow,
            args=["analyze-cv", f"--id={candidate_id}"],
            start_to_close_timeout=timedelta(minutes=5)
        )

        if not analysis["qualified"]:
            await send_rejection_email(candidate_id)
            return "rejected"

        # Step 2: Schedule call
        await send_scheduling_email(candidate_id)

        # Wait for calendar event (webhook will signal workflow)
        call_scheduled = await workflow.wait_condition(
            lambda: self.call_scheduled,
            timeout=timedelta(days=7)
        )

        # Step 3: Day of call (sleep until call time)
        await workflow.sleep_until(call_scheduled.call_time)

        # Trigger Twilio call
        await make_outbound_call(candidate_id)

        # Wait for call completion webhook
        await workflow.wait_condition(lambda: self.call_completed)

        # Step 4: Generate summary (RAW workflow)
        summary = await workflow.execute_activity(
            run_raw_workflow,
            args=["generate-call-summary", f"--id={candidate_id}"],
            start_to_close_timeout=timedelta(minutes=3)
        )

        return "completed"

4. Deployment (Docker + K8s)

Production deployment requires:

• Docker Compose for local development
• Kubernetes manifests for production
• Helm charts for configuration management

Example structure:

docker-compose.yml       # Local dev environment
k8s/
  deployment.yml         # API deployment
  service.yml            # API service
  ingress.yml            # External access
  postgres.yml           # Database
  temporal.yml           # Temporal server

Common System Patterns

Pattern 1: Webhook → RAW → Database

Use case: Form submission triggers analysis and saves results

Components:

• FastAPI webhook endpoint
• RAW workflow for processing
• PostgreSQL for state

Flow:

1. User submits form → POST /webhooks/submit
2. FastAPI saves to database (status: pending)
3. FastAPI calls raw run analyze --id=123
4. RAW workflow processes and returns result
5. FastAPI updates database (status: complete)

Pattern 2: Temporal + RAW Multi-Day Process

Use case: Process spans multiple days with external events

Components:

• Temporal workflow for orchestration
• RAW workflows for data processing
• FastAPI for webhooks
• Database for state

Flow:

1. Trigger starts Temporal workflow
2. Workflow runs RAW workflow for initial processing
3. Workflow schedules reminder (sleeps until date)
4. External event triggers webhook → signals Temporal
5. Workflow continues execution
6. Workflow runs another RAW workflow for final step

Pattern 3: Scheduled Reports

Use case: Daily/weekly reports generated and sent

Components:

• Kubernetes CronJob
• RAW workflow for report generation
• Email/Slack integration

Flow:

1. CronJob triggers at scheduled time
2. Runs raw run generate-report --date=today
3. RAW workflow queries database, generates report
4. Saves to S3 and sends notification

Decision Tree

Does it need state across days? ────NO───→ Single RAW Workflow
    │
    YES
    │
    ↓
Does it need webhooks? ────NO───→ Temporal + RAW Workflows
    │
    YES
    │
    ↓
FastAPI + Temporal + RAW Workflows + Deployment

Code Generation Checklist

When generating a multi-system architecture:

• Create RAW workflows for data processing
• Create FastAPI app with webhook endpoints
• Add database models (SQLAlchemy or Pydantic)
• Create Temporal workflows for orchestration
• Add Docker Compose for local dev
• Create Kubernetes manifests for production
• Add tests for each component
• Document the architecture in README.md

Validation Gates

For multi-system architectures, add these gates:

builder:
  gates:
    default:
      - validate      # RAW workflow validation
      - dry          # RAW workflow dry run
    optional:
      api-test:
        command: "pytest api/tests/ -v"
        timeout_seconds: 60
      temporal-validate:
        command: "temporal workflow validate"
        timeout_seconds: 30
      docker-validate:
        command: "docker-compose config"
        timeout_seconds: 10

Key Principles

1. RAW workflows are pure functions - No state, no side effects beyond outputs
2. FastAPI coordinates - Handles webhooks, database, calls RAW workflows
3. Temporal orchestrates - Manages multi-day processes reliably
4. Deploy everything together - Docker Compose for dev, K8s for prod

This pattern keeps RAW workflows simple and reusable while building complex systems around them.