Back to Gallery
Fastapi Production Architecture Agent Rule
View Full Resolution
100% Free Access
AI Architecture Cursor AI / Claude 3.5
Category AI Agents
Best Use Case Commercial & Cinematic
AI Agents Verified Blueprint

Fastapi Production Architecture Agent Rule

Cursor rules for FastAPI services with router/service/repository boundaries, typed provider adapters, bulkhead isolation, idempotency, and domain exceptions.

Ready-to-Run Prompt
100% Free Copy
# FastAPI Production Architecture Rules
# Principles for production-ready FastAPI services.

## LAYER ARCHITECTURE (Principles A1-A8)

This codebase follows strict 4-layer architecture: Router → Service → Repository → ORM/HTTP/Storage.
Imports flow downward only. Each layer has hard boundaries you must NOT cross.

### Router rules (app/routers/**)
- Handlers are THIN: ≤10 lines of executable code per handler
- Allowed imports: fastapi, app.schemas.*, app.core.deps, app.services.*
- FORBIDDEN imports: sqlalchemy, httpx, boto3, app.models.*, app.repositories.*
- Every endpoint declares response_model= for OpenAPI fidelity
- Every protected/business endpoint requires user_id: str = Depends(get_current_user_id)
- Public endpoints (health checks, webhooks, callbacks) are exempt from auth
- Business logic lives in services. Routers parse input, call one service method, return response.

GOOD:
@router.post("/wallet/charge", response_model=WalletResponse, status_code=201)
async def charge(
req: ChargeRequest,
user_id: str = Depends(get_current_user_id),
svc: WalletUserService = Depends(get_wallet_service),
) -> WalletResponse:
wallet = await svc.charge(
user_id=user_id,
amount=req.amount,
idempotency_key=req.idempotency_key,
)
return WalletResponse.from_domain(wallet)

BAD (business logic + SQL in router):
@router.post("/wallet/charge")
async def charge(req: ChargeRequest, db: Session = Depends(get_db)):
wallet = db.query(Wallet).filter(Wallet.user_id == user_id).with_for_update().one()
...

### Service rules (app/services/**)
- FORBIDDEN imports: sqlalchemy, httpx, boto3, redis, FastAPI Request/Response/HTTPException
- Constructor injects Protocol-typed dependencies, not concrete classes
- Raise domain exceptions (InsufficientFundsError), not HTTPException

GOOD:
from app.repositories.protocols import WalletRepoProtocol
class WalletUserService:
def __init__(self, repo: WalletRepoProtocol): # Protocol, not SQLAlchemy Session
self._repo = repo

BAD:
from sqlalchemy.orm import Session
class WalletUserService:
def __init__(self, db: Session): ... # Wrong — service depends on infrastructure

### Repository rules (app/repositories/**)
- ONLY layer allowed to import sqlalchemy
- Implements Protocol from app/repositories/protocols.py
- Returns domain objects, not ORM models
- Every query scoped by user_id (multi-tenancy)

### Provider rules (app/providers/**)
- ONLY layer allowed to import httpx directly
- Returns GenerateResult | ProviderError — NEVER raw dict
- Uses per-provider httpx.AsyncClient (bulkhead pattern)

## FILE SIZE RULES (Principle A1)

| LOC | State | Action |
|----------|--------|---------------------------------------------|
| 0–399 | Green | None. |
| 400–599 | Yellow | Plan split. Add TODO(decompose) header. |
| 600+ | Red | BLOCK merge. Decompose first. |

Convert file to package when ANY is true:
- Crosses 400 LOC and next change pushes past 500
- Contains 2+ disjoint sub-domains (image vs video, user vs admin)
- Mixes HTTP handlers with worker handlers
- Has 2+ callers each importing only one symbol

Safe split pattern (atomic PR):
1. Create <file>/__init__.py (empty for now)
2. Move pieces to sub-files (a.py, b.py, c.py)
3. Re-export old public names from __init__.py
4. Run tests — must pass without changes
5. Follow-up PR to migrate callers off legacy alias

__init__.py pattern:
from .user import WalletUserService
from .admin import WalletAdminService
WalletService = WalletUserService # backwards-compat alias
__all__ = ["WalletUserService", "WalletAdminService", "WalletService"]

## EXTERNAL INTEGRATION RULES (Principles B1-B10)

### Rule 1: Anti-Corruption Layer (ACL)
Providers return GenerateResult | ProviderError, never dict.

from dataclasses import dataclass
from decimal import Decimal

@dataclass(frozen=True)
class GenerateResult:
url: str
cost_usd: Decimal
latency_ms: int
provider_request_id: str

class ProviderError(Exception):
def __init__(self, message: str, *, retryable: bool, code: str | None = None):
super().__init__(message); self.retryable = retryable; self.code = code

class ProviderTimeout(ProviderError):
def __init__(self, message: str): super().__init__(message, retryable=True, code="timeout")

### Rule 2: Per-Provider Bulkhead
Each external provider has its OWN httpx.AsyncClient with its OWN Limits. NEVER share.

GOOD:
FAL_HTTP = httpx.AsyncClient(
base_url=settings.FAL_BASE_URL,
timeout=httpx.Timeout(connect=5.0, read=60.0, write=10.0, pool=5.0),
limits=httpx.Limits(max_connections=20, max_keepalive_connections=10),
)
OPENAI_HTTP = httpx.AsyncClient(
base_url="https://api.openai.com/v1",
limits=httpx.Limits(max_connections=50, max_keepalive_connections=20),
)

BAD:
HTTP = httpx.AsyncClient() # shared across all providers — no bulkhead isolation

# Shutdown cleanup — close all provider clients in FastAPI lifespan
from contextlib import asynccontextmanager

@asynccontextmanager
async def lifespan(app):
yield # app startup
await FAL_HTTP.aclose()
await OPENAI_HTTP.aclose()

app = FastAPI(lifespan=lifespan)
# Or on_event("shutdown"):
# @app.on_event("shutdown")
# async def close_http_clients() -> None:
# await FAL_HTTP.aclose()
# await OPENAI_HTTP.aclose()

### Rule 3: Idempotency Keys
Every side-effect operation accepts an idempotency_key: UUID. Look up before retrying.

### Rule 4: Structured Logging with contextvars
Use ContextVar to thread provider, user_id, request_id through async call stacks.

import contextvars

provider_var = contextvars.ContextVar[str | None]("provider", default=None)
user_id_var = contextvars.ContextVar[str | None]("user_id", default=None)
request_id_var = contextvars.ContextVar[str | None]("request_id", default=None)

# Read in logs: provider_var.get(), user_id_var.get(), request_id_var.get()
# JSON formatter picks these up automatically via extra={} or ContextVar.get()

### Rule 5: Single-Writer Principle (Principle B10)
For safety-critical state: exactly ONE service-layer module does the writing.
Only the designated writer service for a domain may call repo.hold().
Routers and providers must NOT call repo.hold() directly.
Admin services implemented in the service layer may call repo.hold() only
if they are the designated writer for that domain.
Enforce via: code-review grep check (`grep -r "repo\.hold(" --include="*.py"`)
and unit tests that assert call-origin of repo.hold().

## ANTI-PATTERNS — REJECT ON SIGHT

1. def some_method(self, db: Session, ...) → use Protocol-typed repo
2. from app.models.user import User inside a service → return domain types from repo
3. httpx.AsyncClient() instantiated inside a function → use shared per-provider client
4. raise HTTPException(...) inside a service → raise domain exception
5. db.query(...) inside a router → move to service then repo
6. async def call(self, req) -> dict: return resp.json() → ACL violation
7. HTTP = httpx.AsyncClient() shared → bulkhead violation
8. result["vendor_field"]["nested"] in service code → ACL violation
9. logger.info(f"{user_id} did X") → use structured logging with extra={}
10. Side-effect operation without idempotency_key → double-charge risk

## DEPENDENCY INJECTION

Use FastAPI Depends() + factory functions in app/core/deps.py.
Do NOT install dependency-injector, punq, or any DI container.

def get_wallet_service(db: Session = Depends(get_db)) -> WalletUserService:
return WalletUserService(repo=SQLAlchemyWalletRepo(db))

## DOMAIN EXCEPTIONS PATTERN

Services raise domain errors. Routers map to HTTP.

# services/wallet/exceptions.py
class InsufficientFundsError(Exception): ...
class WalletNotFoundError(Exception): ...

# routers/wallet.py
try:
wallet = await svc.charge(...)
except InsufficientFundsError:
raise HTTPException(402, detail="insufficient funds")

Structured JSON Schema

Use with automated API pipelines, LangChain, or custom image generators

{
    "system_prompt": "# FastAPI Production Architecture Rules\n# Principles for production-ready FastAPI services.\n\n## LAYER ARCHITECTURE (Principles A1-A8)\n\nThis codebase follows strict 4-layer architecture: Router → Service → Repository → ORM/HTTP/Storage.\nImports flow downward only. Each layer has hard boundaries you must NOT cross.\n\n### Router rules (app/routers/**)\n- Handlers are THIN: ≤10 lines of executable code per handler\n- Allowed imports: fastapi, app.schemas.*, app.core.deps, app.services.*\n- FORBIDDEN imports: sqlalchemy, httpx, boto3, app.models.*, app.repositories.*\n- Every endpoint declares response_model= for OpenAPI fidelity\n- Every protected/business endpoint requires user_id: str = Depends(get_current_user_id)\n- Public endpoints (health checks, webhooks, callbacks) are exempt from auth\n- Business logic lives in services. Routers parse input, call one service method, return response.\n\nGOOD:\n@router.post(\"/wallet/charge\", response_model=WalletResponse, status_code=201)\nasync def charge(\n    req: ChargeRequest,\n    user_id: str = Depends(get_current_user_id),\n    svc: WalletUserService = Depends(get_wallet_service),\n) -> WalletResponse:\n    wallet = await svc.charge(\n        user_id=user_id,\n        amount=req.amount,\n        idempotency_key=req.idempotency_key,\n    )\n    return WalletResponse.from_domain(wallet)\n\nBAD (business logic + SQL in router):\n@router.post(\"/wallet/charge\")\nasync def charge(req: ChargeRequest, db: Session = Depends(get_db)):\n    wallet = db.query(Wallet).filter(Wallet.user_id == user_id).with_for_update().one()\n    ...\n\n### Service rules (app/services/**)\n- FORBIDDEN imports: sqlalchemy, httpx, boto3, redis, FastAPI Request/Response/HTTPException\n- Constructor injects Protocol-typed dependencies, not concrete classes\n- Raise domain exceptions (InsufficientFundsError), not HTTPException\n\nGOOD:\nfrom app.repositories.protocols import WalletRepoProtocol\nclass WalletUserService:\n    def __init__(self, repo: WalletRepoProtocol):  # Protocol, not SQLAlchemy Session\n        self._repo = repo\n\nBAD:\nfrom sqlalchemy.orm import Session\nclass WalletUserService:\n    def __init__(self, db: Session): ...  # Wrong — service depends on infrastructure\n\n### Repository rules (app/repositories/**)\n- ONLY layer allowed to import sqlalchemy\n- Implements Protocol from app/repositories/protocols.py\n- Returns domain objects, not ORM models\n- Every query scoped by user_id (multi-tenancy)\n\n### Provider rules (app/providers/**)\n- ONLY layer allowed to import httpx directly\n- Returns GenerateResult | ProviderError — NEVER raw dict\n- Uses per-provider httpx.AsyncClient (bulkhead pattern)\n\n## FILE SIZE RULES (Principle A1)\n\n| LOC      | State  | Action                                      |\n|----------|--------|---------------------------------------------|\n| 0–399    | Green  | None.                                       |\n| 400–599  | Yellow | Plan split. Add TODO(decompose) header.     |\n| 600+     | Red    | BLOCK merge. Decompose first.               |\n\nConvert file to package when ANY is true:\n- Crosses 400 LOC and next change pushes past 500\n- Contains 2+ disjoint sub-domains (image vs video, user vs admin)\n- Mixes HTTP handlers with worker handlers\n- Has 2+ callers each importing only one symbol\n\nSafe split pattern (atomic PR):\n1. Create <file>/__init__.py (empty for now)\n2. Move pieces to sub-files (a.py, b.py, c.py)\n3. Re-export old public names from __init__.py\n4. Run tests — must pass without changes\n5. Follow-up PR to migrate callers off legacy alias\n\n__init__.py pattern:\nfrom .user import WalletUserService\nfrom .admin import WalletAdminService\nWalletService = WalletUserService  # backwards-compat alias\n__all__ = [\"WalletUserService\", \"WalletAdminService\", \"WalletService\"]\n\n## EXTERNAL INTEGRATION RULES (Principles B1-B10)\n\n### Rule 1: Anti-Corruption Layer (ACL)\nProviders return GenerateResult | ProviderError, never dict.\n\nfrom dataclasses import dataclass\nfrom decimal import Decimal\n\n@dataclass(frozen=True)\nclass GenerateResult:\n    url: str\n    cost_usd: Decimal\n    latency_ms: int\n    provider_request_id: str\n\nclass ProviderError(Exception):\n    def __init__(self, message: str, *, retryable: bool, code: str | None = None):\n        super().__init__(message); self.retryable = retryable; self.code = code\n\nclass ProviderTimeout(ProviderError):\n    def __init__(self, message: str): super().__init__(message, retryable=True, code=\"timeout\")\n\n### Rule 2: Per-Provider Bulkhead\nEach external provider has its OWN httpx.AsyncClient with its OWN Limits. NEVER share.\n\nGOOD:\nFAL_HTTP = httpx.AsyncClient(\n    base_url=settings.FAL_BASE_URL,\n    timeout=httpx.Timeout(connect=5.0, read=60.0, write=10.0, pool=5.0),\n    limits=httpx.Limits(max_connections=20, max_keepalive_connections=10),\n)\nOPENAI_HTTP = httpx.AsyncClient(\n    base_url=\"https://api.openai.com/v1\",\n    limits=httpx.Limits(max_connections=50, max_keepalive_connections=20),\n)\n\nBAD:\nHTTP = httpx.AsyncClient()  # shared across all providers — no bulkhead isolation\n\n# Shutdown cleanup — close all provider clients in FastAPI lifespan\nfrom contextlib import asynccontextmanager\n\n@asynccontextmanager\nasync def lifespan(app):\n    yield  # app startup\n    await FAL_HTTP.aclose()\n    await OPENAI_HTTP.aclose()\n\napp = FastAPI(lifespan=lifespan)\n# Or on_event(\"shutdown\"):\n# @app.on_event(\"shutdown\")\n# async def close_http_clients() -> None:\n#     await FAL_HTTP.aclose()\n#     await OPENAI_HTTP.aclose()\n\n### Rule 3: Idempotency Keys\nEvery side-effect operation accepts an idempotency_key: UUID. Look up before retrying.\n\n### Rule 4: Structured Logging with contextvars\nUse ContextVar to thread provider, user_id, request_id through async call stacks.\n\nimport contextvars\n\nprovider_var   = contextvars.ContextVar[str | None](\"provider\", default=None)\nuser_id_var    = contextvars.ContextVar[str | None](\"user_id\", default=None)\nrequest_id_var = contextvars.ContextVar[str | None](\"request_id\", default=None)\n\n# Read in logs: provider_var.get(), user_id_var.get(), request_id_var.get()\n# JSON formatter picks these up automatically via extra={} or ContextVar.get()\n\n### Rule 5: Single-Writer Principle (Principle B10)\nFor safety-critical state: exactly ONE service-layer module does the writing.\nOnly the designated writer service for a domain may call repo.hold().\nRouters and providers must NOT call repo.hold() directly.\nAdmin services implemented in the service layer may call repo.hold() only\nif they are the designated writer for that domain.\nEnforce via: code-review grep check (`grep -r \"repo\\.hold(\" --include=\"*.py\"`)\nand unit tests that assert call-origin of repo.hold().\n\n## ANTI-PATTERNS — REJECT ON SIGHT\n\n1. def some_method(self, db: Session, ...) → use Protocol-typed repo\n2. from app.models.user import User inside a service → return domain types from repo\n3. httpx.AsyncClient() instantiated inside a function → use shared per-provider client\n4. raise HTTPException(...) inside a service → raise domain exception\n5. db.query(...) inside a router → move to service then repo\n6. async def call(self, req) -> dict: return resp.json() → ACL violation\n7. HTTP = httpx.AsyncClient() shared → bulkhead violation\n8. result[\"vendor_field\"][\"nested\"] in service code → ACL violation\n9. logger.info(f\"{user_id} did X\") → use structured logging with extra={}\n10. Side-effect operation without idempotency_key → double-charge risk\n\n## DEPENDENCY INJECTION\n\nUse FastAPI Depends() + factory functions in app/core/deps.py.\nDo NOT install dependency-injector, punq, or any DI container.\n\ndef get_wallet_service(db: Session = Depends(get_db)) -> WalletUserService:\n    return WalletUserService(repo=SQLAlchemyWalletRepo(db))\n\n## DOMAIN EXCEPTIONS PATTERN\n\nServices raise domain errors. Routers map to HTTP.\n\n# services/wallet/exceptions.py\nclass InsufficientFundsError(Exception): ...\nclass WalletNotFoundError(Exception): ...\n\n# routers/wallet.py\ntry:\n    wallet = await svc.charge(...)\nexcept InsufficientFundsError:\n    raise HTTPException(402, detail=\"insufficient funds\")",
    "prompt_type": "agent_rule",
    "framework": "cursor",
    "globs": "**/*",
    "compatible_models": [
        "Claude 3.5 Sonnet",
        "GPT-4o",
        "Cursor AI",
        "Gemini 2.5 Flash"
    ],
    "download_filename": "fastapi-production-architecture.cursorrules",
    "tags": [
        "cursor",
        "cursorrules",
        "agent",
        "coding",
        "fastapi"
    ]
}
Internal Discovery

More AI Agents Prompts

View All →