v0.14.1

erdify 0.14.1 — a fix for django-ninja ModelSchema, and the new documentation image.

Fixed

• erdify now recognises ninja-style ModelSchema classes — those taking their
  fields from a Django model via an inner Meta (django-ninja) or Config
  (ninja-schema) — and skips them with an explanatory warning instead of
  drawing an entity with no fields. Naming ModelSchema via --base-classes
  previously produced an empty box beside the real table, which was worse than
  no support at all. Detection is narrow: only an inner Meta/Config that
  assigns model counts, so an ordinary Pydantic model with a nested config
  class and Django's own class Meta are both untouched. Resolving those
  schemas properly remains open in
  #171.
• docs/examples/frameworks.svg is well-formed XML again. Its header comment
  contained a command-line flag, and XML comments may not contain a double
  hyphen, so the file failed to parse — harmless for the site, which serves the
  PNG, but the file documented as "the editable source" would not open in a
  browser or vector editor.
• The image footer reads "and raw SQL DDL" rather than listing it as a sixth
  bullet. The headline counts five frameworks; SQL DDL is an input format, and
  the flat list made the two numbers look like they disagreed.

Changed

• The README and the documentation home page now open with a "five frameworks,
  one diagram" image — the same User/Order schema written in SQLModel,
  SQLAlchemy, Django, Pydantic and dataclasses, converging on the ERD they all
  produce. The previous image was the bare ERD under a caption claiming five
  frameworks, which the picture itself never showed. The snippets are taken
  from the runnable examples in docs/examples/, and it is drawn in the devsuit
  brand palette. docs/examples/frameworks.svg is the editable source; the
  now-unreferenced docs/examples/erd.png is removed.

---

Install: pip install erdify · Docs: https://erdify.devsuit.io/ · Full diff: v0.14.0...v0.14.1

v0.14.0

erdify 0.14.0 — ER diagrams for django-ninja schemas, and an honest comparison.

Added

• --base-classes NAME [NAME ...] (and base_classes in [tool.erdify]) names
  extra base classes to treat as Pydantic models. Pydantic detection resolves
  ancestors only across the scanned files, so a base defined in an installed
  package — ninja.Schema from django-ninja — or in an internal library that
  --include does not match was unresolvable, and every subclass of it was
  silently skipped. Both the bare (Schema) and qualified (ninja.Schema)
  forms are matched, and intermediate bases defined in scanned files still
  resolve transitively. Partially addresses
  #171; the
  ModelSchema variants, whose fields come from an inner Meta/Config
  rather than the class body, remain unsupported.

Fixed

• The comparison page and the README comparison table were wrong about two
  rows. Both claimed no alternative ships a Markdown-injection or a CI
  drift-check feature; paracelsus ships
  both, along with [tool.paracelsus] configuration and Mermaid output. It is
  now a column in the comparison table and a named entry under "when to use
  something else": it reaches models by importing them (--import-module), so
  it sees the fully-resolved metadata an AST scan cannot, and for a
  single-SQLAlchemy project that accuracy may be worth more than erdify's
  independence.

---

Install: pip install erdify · Docs: https://erdify.devsuit.io/ · Full diff: v0.13.0...v0.14.0

v0.13.0

erdify 0.13.0 — one ERD generator for every Python schema in your repository.

Warning

Breaking: a run that finds no tables now exits 1 instead of 0 and writes nothing. Pass --allow-empty (or set allow_empty = true under [tool.erdify]) to restore the previous behavior. Details below.

Changed — breaking

• A run that finds no tables now exits 1 instead of 0, and writes
  nothing. Up to 0.12.3 it warned, wrote a structurally valid but empty
  diagram, and exited 0 — so a CI job that commits the regenerated ERD
  silently replaced a good diagram with an empty one whenever --include
  stopped matching, which a rename or a moved package is enough to cause. The
  overwhelmingly likely cause of zero entities is a misconfiguration, so that
  is now an error.

  Failing before generation means an existing output file is left untouched.

  If an empty schema is a legitimate outcome for your project — everything
  filtered out by --exclude, a --sources filter that matches nothing, a
  schema mid-migration — pass --allow-empty or set allow_empty = true under
  [tool.erdify]. That restores the previous behavior exactly.

Added

• New documentation page Comparison — erdify next to eralchemy,
  sqlalchemy-schemadisplay, erdantic, django-extensions graph_models and
  DBML/dbdiagram, on the axes that actually differ (live connection, importing
  your code, frameworks covered, runtime dependencies, output formats, CI drift
  gate), including where each alternative is the better choice.
• New documentation page Python Parsing Limitations — the Python-side
  counterpart to the SQL frontend's "Deferred / Not Supported" table: what the
  AST parser cannot see (runtime-constructed models, non-literal
  __tablename__, __table_args__, composite foreign keys, foreign keys inside
  sa_column, unannotated relationship() assignments) and what to do instead.
• New documentation page Troubleshooting — symptom-first fixes for models
  not being found, too many entities, missing relationships and missing keys on
  Pydantic/dataclass models.
• Every documentation page now carries its own description frontmatter, so
  each page gets its own search-engine snippet and social-card subtitle instead
  of repeating site_description.
• --allow-empty (and allow_empty in [tool.erdify]) opts back in to the
  pre-0.13.0 behavior for an empty result: warn, write the empty diagram, exit
  0. See the breaking change below.
• erdify now ships .pre-commit-hooks.yaml, so it can be used as a pre-commit
  repository instead of a repo: local / language: system hook that requires
  erdify to be installed in the consumer's environment first. Two ids are
  provided: erdify regenerates the diagram (pre-commit fails the commit when
  the file changed), and erdify-check never writes and exits non-zero when the
  committed diagram has drifted. --check is part of the latter's entry point,
  so overriding args cannot drop it. SQL DDL projects add
  additional_dependencies: ['sqlglot>=25'] rather than using a separate id.
• The documentation site now renders a social card (og:image) for every page,
  so links shared on chat, social and link-preview surfaces show a real card
  instead of a bare text entry. This enables mkdocs-material's social plugin,
  which pulls in the mkdocs-material[imaging] dependencies and their system
  libraries in the docs workflow.
• The changelog is published as a page on the documentation site. It includes
  the repository's CHANGELOG.md verbatim via pymdownx.snippets, so there is
  still a single source of truth.
• Every documentation page now has an edit link back to its source on GitHub
  (content.action.edit; edit_uri was already configured but Material never
  rendered the button without the feature) and previous/next navigation in the
  footer (navigation.footer).

Changed

• The "No tables found" message now ends with a link to the troubleshooting
  page, which did not exist when the message was written.

• The Datadog coverage upload is now also skipped for Dependabot. The existing
  fork check covers pull requests from forks, but a Dependabot pull request's
  head branch lives in this repository, and GitHub runs those with Dependabot
  secrets rather than Actions secrets — so the step would have run with an
  empty key and failed the required Tests passed check on every dependency
  bump. The condition matches on the actor: secrets cannot be referenced in an
  if: conditional, and the documented job-level-env workaround would put the
  API key in the environment of every step in the job.

• SECURITY.md scopes its absolute claims to the core install. "Does not
  execute / import / connect / send" holds for the stdlib-only core; the
  erdify[sql] extra runs third-party code (sqlglot) in your process and is now
  described as its own trust boundary. Added: --output/--inject do write to
  the paths you name, and the rendering step (PlantUML server, Mermaid, browser)
  is outside erdify's control. The response times are stated as what a small
  team aims for rather than a service-level guarantee.

• docs/usage/ci.md leads with the --check drift gate and presents the
  auto-commit workflow as the alternative, with its trade-off stated. The
  workflow examples keep floating action tags for readability, and a note now
  says so explicitly and tells readers to pin to a SHA in their own repository —
  as erdify's own workflows do.

• Reduced decorative emoji in SECURITY.md and CONTRIBUTING.md, which people
  reach when something is wrong or when they are trying to get set up.

• The "No tables found" warning is now actionable: it names the active
  --include patterns, reports how many .py/.sql files were scanned and how
  many matched, and distinguishes "nothing matched the patterns" (the usual
  cause — --include defaults to models.py, so models in e.g. schema.py are
  never seen) from "files matched but held no recognized models".

• The README now answers "why erdify and not X" right after the feature list,
  with a condensed comparison table and a link to the full one.

• The README shows its example ERD as a live Mermaid diagram injected with
  erdify --inject instead of a committed PNG — the format the README itself
  advertises as rendering natively on GitHub. A readme-erd pre-commit hook
  runs the same command with --check, so the embedded diagram cannot go stale
  (and it dogfoods both flags on the project's own front page).

• The README badge row is down from ten to six: PyPI version, Python versions,
  License, Tests, Coverage and a new docs badge. The Linting, Ruff, mypy and
  download badges added rows without adding information a reader acts on.

• The PyPI project page now links to the documentation site
  (https://erdify.devsuit.io/) instead of the README anchor on GitHub, and
  gains Changelog and Issues links.

• [tool.ruff] target-version is py311, matching requires-python = ">=3.11"
  instead of contradicting it with py310.

• The documentation home page now leads with what erdify is, the
  framework-comparison diagram and a three-line install-and-run, instead of
  restating the sidebar. The Comparison, Python Parsing Limitations and
  Troubleshooting pages are in the nav.

• docs/frameworks/index.md loads its second example image from the repository
  instead of raw.githubusercontent.com, matching the first one on the page.

• The Mermaid section of docs/usage/output-formats.md shows the .mmd source
  and the rendered diagram in side-by-side tabs; the rendered-only fence never
  showed readers what the file actually contains.

• The required Tests passed check now fails for a pull request from the
  machine-written badges branch into main (the only base the test workflow
  runs for). GitHub cannot bar a branch from being a pull-request source, and
  such a pull request would otherwise look mergeable while replacing the README
  on main with the branch's own.

• Test coverage is now uploaded to Datadog Code Coverage from the Linux /
  Python 3.13 matrix leg, with code-coverage.datadog.yaml defining the
  service mapping and 90% total/patch PR gates (matching the existing
  fail_under), and service.datadog.yaml registering erdify in the Datadog
  Software Catalog. The upload is skipped on pull requests from forks, which
  have no access to repository secrets; the workflow deliberately stays on the
  pull_request trigger.

• The README now carries a coverage badge. Datadog exposes no public badge
  endpoint, so pushes to main publish the percentage as a shields.io endpoint
  payload on the orphan badges branch of this repository, committed by
  github-actions[bot] with the workflow's built-in GITHUB_TOKEN. Pull
  requests never update it. Keeping the payload in the repository rather than in
  a Gist or a third-party badge service means no personal credential is involved
  and the badge survives any contributor leaving.

• CI now resolves uv to latest-known instead of latest, so the version
  installed in every workflow (including the PyPI publish job) is one whose
  checksum ships with the pinned setup-uv release. New uv versions reach CI
  only through a setup-uv bump, which arrives as a reviewed, cooldown-gated
  Dependabot PR.

• Dependabot's github-actions version-update group now includes major, so
  major action bumps arrive as a single grouped PR instead of one PR each. The
  uv entry deliberately keeps majors separate, and security updates are
  unaffected in both.

---

Install: pip install erdify · Docs: https://erdify.devsuit.io/ · Full diff: v0.12.3...v0.13.0

v0.12.3

Maintenance patch. No runtime or user-facing change; the core install stays dependency-free.

Security

• Updated two transitive development dependencies carrying open advisories, clearing all six Dependabot alerts (4 high, 2 moderate):

  • sqlparse 0.5.5 → 0.6.0 — GHSA-prg7-hcfm-mfcr (ReDoS via dollar-quoted SQL literals), GHSA-pwgv-4x5q-6m9f, GHSA-f2ff-p2ww-7p4p, GHSA-3496-9g83-7v6x
  • pymdown-extensions 10.21.3 → 11.0.1 — GHSA-gm37-52c6-37mw, GHSA-9xwg-3r6f-jcx2

  No previous release was affected. erdify declares no runtime dependencies, and both packages are reached only through the development and documentation groups (sqlparse via django, pymdown-extensions via mkdocs-material). This hardens the development and CI toolchain only.

Changed

• Dependabot now holds freshly published versions back for 7 days (cooldown.default-days: 7 on every updates entry) before proposing a version update, giving the ecosystem and security researchers time to catch a compromised or broken release. Security updates are unaffected — cooldown applies to version updates only.
• Bumped CI actions to their latest majors: astral-sh/setup-uv 9.0.0 → 10.0.1, actions/configure-pages 5.0.0 → 6.0.0, actions/upload-pages-artifact 4.0.0 → 5.0.0, and actions/deploy-pages 4.0.5 → 5.0.0. Repository infrastructure only.

Full Changelog: v0.12.2...v0.12.3

v0.12.2

Maintenance patch. No runtime or user-facing change; the core install stays dependency-free.

Changed

• Bumped development dependencies (version-updates group): sqlglot (30.15.0 → 30.17.0), ruff (0.16.2 → 0.16.3), mypy (2.3.0 → 2.3.1), and sqlalchemy (2.0.51 → 2.0.52).
• The documentation site is now deployed through the GitHub Pages Actions workflow instead of pushing to the gh-pages branch. Repository infrastructure only — the published site is unchanged.

Full Changelog: v0.12.1...v0.12.2

v0.12.1

Dev-deps-only patch. No runtime or user-facing change; the core install stays dependency-free.

Changed

• Bumped development dependencies: ruff (0.16.0 → 0.16.2), sqlglot (30.14.0 → 30.15.0), and django (5.2.16 → 5.2.17).

Full Changelog: v0.12.0...v0.12.1

v0.12.0

Added

• SQL frontend: UNIQUE single-column foreign keys now render as 1:1 (instead of N:1). Drawn child ... parent: |o--|| when the FK column is NOT NULL, |o--o| when nullable — the child end is |o because a UNIQUE FK caps a parent at zero-or-one child. Both column-level UNIQUE and single-column table-level UNIQUE (col) (anonymous or named CONSTRAINT ... UNIQUE (col)) are recognized; composite UNIQUE (a, b) is out of scope. A unique non-primary-key column is also marked UK in the entity block, and --format json gains a unique boolean per field. Other frontends are unaffected (#97, #129).

Full changelog: v0.11.5...v0.12.0

v0.11.5

Maintenance patch release. Only development dependencies and build tooling
changed since v0.11.4; no runtime or user-facing change — the core install
stays dependency-free.

Changed

• Bumped development dependencies (version-updates group): sqlglot
  (30.13.0 → 30.14.0) and ruff (0.15.22 → 0.16.0); relaxed the uv-build
  build requirement to >=0.11.19,<0.13.0. No runtime or user-facing change;
  the core install stays dependency-free.
• Pinned the ruff rule set ([tool.ruff.lint] select = ["E", "F"]) so the
  ruff 0.16.0 default-rule expansion (59 → 413 rules) stays behavior-neutral;
  adopting the new rules is tracked separately.

Full changelog: v0.11.4...v0.11.5

v0.11.4

Fixed

• Structural link-table detection now covers composite-PK association tables with more than two foreign keys (ternary and higher-order joins), not just two-FK tables. A table whose primary key is made up entirely of foreign keys — with no payload columns — is flagged as a link table (<< (L) link >>) regardless of its name or arity, and higher-order ones are drawn as a star of edges from each parent to the link entity.
• The rule is now shared by every parser backend (Python AST, Core Table(...) synthesis, and the SQL DDL frontend), so it can no longer drift between them (#118).

Full changelog: https://github.com/devsuit-berlin/erdify/blob/v0.11.4/CHANGELOG.md

v0.11.3

Dev-dependency-only patch release — no runtime or user-facing change. The core install stays dependency-free.

Changed

• Bumped development dependencies (version-updates group): ast-serialize, librt, mkdocs-material, mypy, ruff, sqlglot, and django (5.2.15 → 5.2.16).

Full changelog: https://github.com/devsuit-berlin/erdify/blob/main/CHANGELOG.md
Compare: v0.11.2...v0.11.3