Review the diff the way a senior reviewer on this client would, then fix what's real. Findings must be specific (file:line + why + fix), and ranked so the dev fixes the important ones first.
Step 1: Get the diff and the context
git fetch origin
git diff origin/main...HEAD --stat
git diff origin/main...HEAD
Use the repo's default branch if it isn't main. Also read: the ticket/PRD (to check scope), the repo map's "How we do X here" section, and any CONTRIBUTING.md / CODING_STANDARDS.md. A documented house rule always beats the generic advice below.
Step 2: Run the tools first
Run whatever the project has configured, so you don't spend review time on things a linter catches: pre-commit run --from-ref origin/main --to-ref HEAD, or ruff/black/isort/flake8/mypy directly, plus:
python manage.py check
python manage.py makemigrations --check --dry-run
pytest <changed apps>
Report tool failures first.
Step 3: Review in three passes
Keep the passes separate in the report.
Pass 1: Security (Django/DRF)
- Every new view has explicit
permission_classes; noAllowAnyunless it's meant to be public. - Querysets scoped to the requesting user/org in
get_queryset.Model.objects.get(pk=...)with an id from the request and no ownership check = IDOR. - Serializers that accept input list
fieldsexplicitly; no"__all__"; server-owned fields (org,owner,is_staff,status, prices) areread_onlyand set in the view/service. - Raw SQL /
.extra()/RawSQLparameterized, never built with f-strings or%. mark_safe,|safe,format_htmlmisuse with user content;csrf_exempton session-authenticated views.- File uploads: size, extension and content-type checked; stored outside web root; filenames not trusted.
- Secrets or credentials in code, settings, fixtures or tests. Settings changes that weaken security (
DEBUG,ALLOWED_HOSTS = ["*"], CORS*with credentials). - Outbound requests have timeouts; URLs built from user input are validated (SSRF).
pickle,yaml.loadwithout SafeLoader,eval/execon anything external.- Sensitive data (tokens, passwords, PII) not logged.
Pass 2: Design and hidden coupling
- Signals that carry business logic: a
post_savereceiver that sends emails or updates other models hides side effects from anyone reading the calling code, and skips onbulk_*/update(). Prefer an explicit call in the service. - Logic in serializers:
create()/update()/validate()doing multi-model writes, external calls, or business rules. Move to the service layer (or wherever the house puts it). - Fat views: ORM-heavy logic, branching on business rules in the view.
- Fat models /
save()overrides with side effects (emails, API calls, other-model writes). - Side effects outside
transaction.on_commit(task queued before the row is committed → task can't find it). - Cross-app imports that create cycles, or reaching into another app's internals instead of its service API.
- Shotgun surgery: the same rule duplicated in several places; magic strings for statuses instead of the model's choices.
- N+1s introduced (quick check: new serializer fields touching relations, loops over querysets).
- Scope creep vs the ticket, and acceptance criteria not met.
Pass 3: Tests
- Each acceptance criterion has a test. Permission and cross-tenant cases covered.
- Tests assert behavior, not implementation (no mocking the code under test).
- Migrations reviewed per
django-migrationsif any.
Step 4: Report, then fix
Security
[Must fix] billing/views.py:58 InvoiceVoidView fetches Invoice.objects.get(pk=pk) with no org check → any user can void any org's invoice. Use self.get_queryset().get(pk=pk) (scoped) or get_object().
Design
[Should fix] billing/signals.py:12 post_save sends the void email; move to services.void_invoice() inside on_commit.
Tests
[Should fix] no test for member role → add test_member_cannot_void.
Nits
[Optional] ...
Levels: Must fix (security, data loss, broken behavior, failing checks), Should fix (design/test gaps a reviewer will flag), Optional. Ask the dev which to fix, fix them, re-run the tools and tests, then suggest pr-description.
Rules
- No finding without a file:line and a concrete fix.
- Don't flag style the formatter owns, or pre-existing issues outside the diff (mention those separately as "noticed, not in scope").
- Don't approve something you couldn't run; say what you couldn't verify.