sqlalchemy-annotate 0.1.0

Rails annotate_models for modern SQLAlchemy 2.x + Alembic projects.

sqlalchemy-annotate

Rails annotate_models for modern SQLAlchemy 2.x + Alembic projects.

It keeps a Schema Information comment block at the top of every model file, in sync with your actual SQLAlchemy metadata, so reviewers and editors see the table shape without opening a database or a migration.

# == Schema Information
#
# Table name: users
#
# id         : integer, primary key
# email      : varchar, not null
# created_at : timestamp
#
# Indexes
#   ix_users_email (email) UNIQUE
#
# Foreign Keys
#   profile_id -> profiles.id (ON DELETE CASCADE)
#
# == End Schema Information

class User(Base):
    __tablename__ = "users"
    ...

Why it is safe

• No database required. Everything is read from Base.metadata / __table__. A connection URL, if you pass one, is parsed only for its dialect so types compile correctly offline. No engine is ever connected.
• No regex rewriting. Files are edited through a single libcst transformer. Imports, comments, blank lines and Black formatting are preserved byte-for-byte; only the region between the markers is touched.
• Idempotent. Running generate twice produces no diff, which is what makes check reliable in CI and pre-commit.

Install

pip install sqlalchemy-annotate      # or: uv pip install sqlalchemy-annotate

Usage

sqlalchemy-annotate generate
sqlalchemy-annotate generate --models app.models --engine postgresql://localhost/db
sqlalchemy-annotate check        # exit 1 if stale (CI)
sqlalchemy-annotate remove       # strip every block
sqlalchemy-annotate generate --dry-run     # show the diff, write nothing

--models points at the dotted package path. Subpackages are imported recursively (app/models/user.py, app/models/post.py, ...), so every model registers regardless of how many files you split them across. A broken or circularly-importing module is reported as a warning and skipped, never fatal.

Configuration

Put defaults in pyproject.toml; CLI flags override them.

[tool.sqlalchemy-annotate]
models = "app.models"
engine = "postgresql+asyncpg://..."   # dialect only, never connected
include_indexes = true
include_foreign_keys = true
include_relationships = false
normalize_types = false               # true -> Rails-style (bigint, varchar(255))
sort = "definition"                   # or "alphabetical"
exclude = ["audit_*", "*_history"]

Type rendering

By default types are truthful: rendered exactly as SQLAlchemy compiles them for the resolved dialect (Mapped[int] PK -> integer, Mapped[str] -> varchar). Set normalize_types = true for Rails-style fabrication (autoincrement int PK -> bigint, unlengthed string -> varchar(255)).

The dialect is --dialect if given, else the backend of engine (async drivers collapse to their sync backend), else PostgreSQL.

Alembic / pre-commit

Run it right after a schema change:

alembic revision --autogenerate -m "add users"
alembic upgrade head
sqlalchemy-annotate generate

.pre-commit-config.yaml:

repos:
  - repo: https://github.com/abhinavs/sqlalchemy-annotate
    rev: v0.1.0
    hooks:
      - id: sqlalchemy-annotate

CI gate:

- run: sqlalchemy-annotate check

How it works

Module	Responsibility
discovery.py	import-walk the package, collect mapped classes -> files
schema.py	Table/Mapper -> ModelSchema value objects (pure)
formatter.py	ModelSchema -> comment lines (pure, pluggable renderer)
parser.py	libcst helpers: build lines, locate the existing block
writer.py	one libcst transformer: insert / replace / remove
config.py	defaults <- pyproject.toml <- CLI
cli.py	typer commands; heavy imports stay lazy

Roadmap (designed, not built)

The formatter is already a swappable Renderer behind the sqlalchemy_annotate.renderers entry-point group, leaving room for: Markdown schema export, ER diagram generation, a schema diff command, watch mode, and an IDE/VSCode integration.

Project

• Contributing - dev setup, the idempotency invariant, release steps
• Changelog
• Security policy - note: the tool imports your model package, like pytest/alembic
• Ships with PEP 561 type information (py.typed)

License

MIT (c) 2026 Abhinav Saxena