pryti-contract

Your backend keeps a list of what it does. And it can't lie about it.

Companion library to pryti-semantic-reviewer. Pryti reads code from outside and guesses. This runs inside your app and knows.

What it does

While your app starts, it records:

• every route (path, method, auth)
• every model (fields, types, constraints)
• every outside call (HTTP, email, payments)
• every background job
• how much of that it actually knows vs. couldn't figure out

Then you can ask your app one question and get an answer:

"What does this app actually do?"

Why it matters

AI opens a PR with 400 changed lines. You don't read them. You read this:

RISKY (4)
  POST /orders         auth: user -> public
  POST /orders         new effect: net:analytics.example.com
  shop.Customer.email  unique: True -> False
  shop.Customer.name   field removed (data loss)

Four lines instead of four hundred.

Install

pip install pryti-contract

Use it — three levels

Level 1: nothing to write

Works on an existing Django project with zero code changes.

pryti-contract export --settings myproject.settings -o contract.json

It reads Django's real router and real model registry. Routes built in loops, DRF routers, mixins — all found, because the app already resolved them at startup.

Coverage will be partial. That's reported, not hidden:

wrote contract.json: 9 routes, 6 models, 0 jobs, auth known on 5/9

Level 2: declare what matters

from pryti_contract import contract

@contract.route("POST /orders", auth="user")
@contract.effects("net:api.stripe.com")
def create_order(request):
    ...

Eight decorators total, max. AI writes these correctly first try, because they look like every Django decorator it has ever seen.

Level 3: enforce it

# settings.py
MIDDLEWARE = ["pryti_contract.middleware.ContractMiddleware", ...]

# conftest.py or apps.py
from pryti_contract import guard
guard.install(mode="error")   # off | record | warn | error

Now an undeclared call fails loudly:

shop.views.leaky_order performed undeclared effect 'net:api.stripe.com'.
Declared: none. Add @contract.effects('net:api.stripe.com') or remove the call.

This is the part no linter can do. The bad code doesn't ship.

Don't know what to declare? Run your tests in record mode and let it tell you:

guard.install(mode="record")
# ... run test suite ...
json.dump(guard.suggestions(), open("observed.json", "w"))

pryti-contract suggest observed.json     # prints the decorators to paste in

In CI

- run: pryti-contract export --settings myproject.settings -o head.json
- run: git checkout ${{ github.base_ref }}
- run: pryti-contract export --settings myproject.settings -o base.json
- run: pryti-contract diff base.json head.json --markdown --fail-on risky

--fail-on risky exits 1 on auth weakening, new outside calls, dropped fields, relaxed uniqueness, or changed relations.

How the effect guard works

One hook at the socket layer, not per-library. requests, httpx, urllib, boto3, stripe — all covered by the same code, including libraries that don't exist yet.

• socket.getaddrinfo is checked before it runs, so a blocked call never leaves the process
• socket.socket.connect catches direct-IP connections
• smtplib.SMTP.sendmail catches email
• localhost and unix sockets are never effects, so your database doesn't trip it

Patterns support wildcards: net:*.stripe.com, net:*, email.

What it does not catch

Worth being blunt about.

• Business logic. If AI changes a discount from 10% to 90%, the contract is identical. Routes same, models same, effects same. Tests catch that; this doesn't.
• Anything the router never sees. Dead code, unmounted views.
• Dynamic hostnames are caught at runtime, not at export time. The contract records what you declared; the guard records what actually happened.

Structural mistakes: this. Logic mistakes: your tests. You need both.

Try it

git clone ... && cd pryti-contract
pip install -e ".[dev]"
pytest
cd examples/demo && pryti-contract export --settings settings --root . -o /tmp/base.json

MIT.