mcdview

Turn a SQL data model into one self-contained HTML page you can open, share and explore: an overview of every table, click one to isolate it with its relations, field-level detail, search. No server, no build step, no dependencies — the page opens straight from file://.

Try it on mcdview.dev (drop a .sql/.dbm, get a shareable link) · Live demo · Example diagrams

Overview → rearrange → click a table to isolate it with its related tables → search. Open the live, interactive version.

The same tool renders a model right here

mcdview model.sql --to-mermaid emits a Mermaid erDiagram that GitHub and GitLab draw natively — no image, no hosting:

erDiagram
    categorie {
        serial id PK
        text nom
        text slug
    }
    produit {
        serial id PK
        text nom
        bigint prix
        integer categorie_id FK
        text sku
    }
    client {
        serial id PK
        text nom
        text email
        text telephone
    }
    commande {
        serial id PK
        integer client_id FK
        date passee_le
        bigint total
        text statut
    }
    ligne {
        serial id PK
        integer commande_id FK
        integer produit_id FK
        integer quantite
        numeric remise
    }
    avis {
        serial id PK
        integer produit_id FK
        integer client_id FK
        integer note
        text commentaire
    }
    categorie ||--o{ produit : ""
    client ||--o{ commande : ""
    commande ||--o{ ligne : ""
    produit ||--o{ ligne : ""
    produit ||--o{ avis : ""
    client ||--o{ avis : ""

What you get

• Explore any model. Overview grouped by schema, click a table to isolate it with its neighbours, field detail (types, NOT NULL, DEFAULT, PK 🔑, clickable FK 🔗, indexes, comments), search by table or column, drag to rearrange (links follow live, positions remembered), force-directed relayout, a minimap and level-of-detail for big models, light/dark toggle, SVG export, #schema.table permalinks, keyboard shortcuts (/ search, r rearrange, Esc overview). One HTML file, works offline.
• Eight input formats. PostgreSQL and ~15 dialects (MySQL/MariaDB, SQLite, SQL Server, Oracle…) via sqlglot; pgModeler .dbm, dbdiagram.io .dbml, Prisma, MySQL Workbench .mwb, Rails db/schema.rb, Mermaid erDiagram, Drizzle schema.ts.
• Diff two versions. --diff old colors what was added, removed and changed (tables, columns, foreign keys, indexes), detects renames, and can write a JSON summary (see below).
• Mermaid export. --to-mermaid for a diagram that renders in any Markdown file.
• Live database. --db postgresql://… / mysql://… reads a running schema (CLI only).
• Zero dependency by default. Python stdlib in, vanilla JS out. Hostile input is HTML-escaped and DoS-budgeted; the upload surface is hardened (XML/zip caps, converter timeouts).
• Tested at scale. 839 real-world schemas, synthetic models to 5000 tables, zero crashes.

Compare two versions

mcdview new.sql --diff old.sql — added tables/columns/FKs in green, removed in red (kept, struck through), changed in amber, with a legend and a "show only what moved" filter:

Install

pipx install mcdview                 # the CLI, isolated
pipx install "mcdview[dialects]"     # + sqlglot for MySQL/SQLite/… input

Or run the script straight from a checkout — it has no dependency of its own: ./mcdview.py model.sql. The .dbm/.dbml/.prisma inputs still need their external converter in the PATH (see Dialects).

Usage

mcdview model.sql                    # (or ./mcdview.py from a checkout)
./mcdview.py model.sql
./mcdview.py model.sql -o explorer.html --titre "My project"
./mcdview.py model.dbm          # pgModeler model (needs pgmodeler-cli)
./mcdview.py model.dbml         # dbdiagram.io model (needs @dbml/cli)
./mcdview.py schema.prisma      # Prisma schema (needs prisma)
./mcdview.py model.mwb          # MySQL Workbench model (native, no tool)
./mcdview.py db/schema.rb       # Rails schema (native, no tool)
./mcdview.py diagram.mmd        # Mermaid erDiagram (native, .mmd/.md)
./mcdview.py schema.ts          # Drizzle ORM schema (native, no tool)
./mcdview.py --db postgresql://user:pw@host/db   # dump a live database (CLI only)
./mcdview.py new.sql --diff old.sql              # compare two versions of a model

Then open the generated HTML file in a browser.

--diff BASELINE compares the main model against an older one (each may be any supported format) and colors the result: tables, columns and foreign keys that were added show green, removed ones red and struck through (kept on the diagram so you can see what went), changed ones amber. A legend appears at the bottom with per-category counts and a "N touched" button that hides the unchanged tables to frame just what moved. Handy for reviewing a migration before running it.

--db reads the schema straight from a running database instead of a file: postgresql://… shells out to pg_dump -s, mysql://… to mysqldump --no-data. It is a command-line-only feature — never wire it behind a public service, since it would let a caller point the process at any host it can reach (SSRF).

A .dbm input needs pgmodeler-cli in the PATH: mcdview delegates the SQL generation to it (pgModeler resolves its relationships at export time), and reuses the table positions drawn in the model.

In the page:

• wheel: zoom, drag: pan;
• hover a table: lights up its foreign-key links;
• click a table: isolates it with its related tables (the others fade out, the view re-frames) and shows its detail on the right — fields, types, NOT NULL, DEFAULT, PK 🔑, clickable FK 🔗, table and column comments, list of referencing tables; the URL gets a #schema.table permalink that reopens straight on that table;
• drag a table to rearrange the diagram, the links follow live;
• "rearrange" button: a force-directed relayout that spreads the tables to unclutter the links (disabled above 400 tables, where it would be slow);
• schema chips (bottom-left, when there are several schemas): click to frame that schema's zone;
• Escape or "overview": back to the full, re-framed overview;
• search with table name autocompletion.

Options

Without --dbm, mcdview computes an automatic layout: one zone per schema, tables arranged in balanced columns, related tables pulled together.

Dialects

PostgreSQL uses the built-in parser (no dependency). For any other dialect, pip install sqlglot and mcdview reads it — MySQL/MariaDB, SQLite, SQL Server (tsql), Oracle, DuckDB, Snowflake, BigQuery, Redshift, ClickHouse, Trino, Spark, Hive. --dialect auto (the default) uses the PostgreSQL parser first and, when it finds no table, tries several sqlglot dialects and keeps the one that parses the most tables; pass --dialect mysql (etc.) to force one. Model files are read through an upstream converter (an optional dependency, like the SQL dialects): pgModeler .dbm via pgmodeler-cli, dbdiagram.io .dbml via @dbml/cli, and Prisma schema.prisma via the prisma CLI. MySQL Workbench .mwb, Rails db/schema.rb, Mermaid erDiagram (.mmd/.mermaid/.md) and Drizzle ORM schema.ts are read natively (no external tool needed). Other proprietary formats (Navicat, ERwin, Oracle SQL Developer Data Modeler…) have no reliable converter: export their SQL instead (Workbench-style Forward Engineer), which mcdview reads.

--fk-audit is for models where every table carries audit columns (created_by, modified_by...) pointing to a users table: those FKs turn that table into a hub linked to everything and clutter the graph.

What mcdview reads from the DDL

• CREATE TABLE [schema.]table (...): columns, types, NOT NULL, DEFAULT (tables without an explicit schema go to public);
• primary keys, inline (CONSTRAINT ... PRIMARY KEY) or added afterwards (ALTER TABLE ... ADD CONSTRAINT ... PRIMARY KEY, pg_dump -s style);
• ALTER TABLE ... ADD CONSTRAINT ... FOREIGN KEY ... REFERENCES ..., including composite keys and targets without an explicit column list (resolved against the target's primary key);
• partitions (PARTITION OF, ATTACH PARTITION): hidden from the diagram, their constraints are carried back to the parent table;
• COMMENT ON TABLE and COMMENT ON COLUMN.

A pg_dump -s dump works as is. Views, functions and data are ignored.

Examples

Each example ships as a .sql in exemples/; the browsable pages are rebuilt by CI on every push:

The real-world schemas are trimmed to their DDL (no data); each file keeps its source and license in a header comment. To regenerate, e.g.:

./mcdview.py exemples/mediatheque.sql --titre "Médiathèque (démo)" --fk-audit '_idmodificateur_fk$' --lang en
./mcdview.py exemples/pagila.sql --titre "Pagila (DVD rental)" --lang en

Development

Plain Python 3, no dependency. All the rendering lives in templates/explorateur.html (inline CSS/JS); the data is injected as JSON in place of __DONNEES__.

Tested at scale

The harness under tests/ last ran mcdview over 13,408 model files: 9,913 .sql plus native .ts (Drizzle), Mermaid, Rails schema.rb and MySQL Workbench .mwb, harvested from public GitHub repositories, plus synthetic models up to 5000 tables. Results: zero crashes and zero timeouts (2 s per-file budget). Every loadable schema produced a valid page (DOM-verified on a sample in headless Chrome); MySQL and SQLite .sql inputs are detected and get a clear hint. Median page build: 3 ms; worst case (a 5.7 MB model): 1.3 s. Separately, pgModeler .dbm models convert through pgModeler's CLI in a frozen-image CI job (a --fix-model repair pass catches models saved by older versions).

Model data (names, types, comments, defaults) is HTML-escaped and the JSON data island escapes every <, so a hostile schema cannot inject markup or script into the page — this matters when generating pages from files you did not write. tests/test_securite.py checks the escaping and a ReDoS/DoS time budget on hostile inputs (tests/malveillant/).

tests/tester.py runs mcdview end-to-end on every committed example and every pgModeler sample model, checks pinned table/FK counts and times each run; tests/grand_banc.py runs the whole local corpus (--strict fails on any exception, --dbm adds the pgModeler models, --chrome N DOM-validates N sampled pages); tests/rapatrier.sh, tests/moissonner.py and tests/generer_synthetique.py fill the local, uncommitted corpus. tests/test_dialectes.py parses committed non-PostgreSQL fixtures, including tests/dialectes/boutique.mariadb.sql — a frozen mysqldump --no-data dump (backtick quoting, named inline CONSTRAINT FKs, a composite primary key) that pins the exact SQL a --db mysql://… run feeds the parser, so the --db path stays covered without a live server. The pre-commit hook (git config core.hooksPath .githooks) runs the pinned regression, the security suite and the strict corpus campaign.

License

MIT. The example schemas keep their own licenses, noted in each file's header.

Changelog

v0.29.1 — Remember the collapsed detail panel (2026-09-01)

• Folding the detail panel is now remembered across loads (localStorage, like the theme). A page reopened with the panel folded starts folded and skips the open-then-close intro, instead of replaying it every time — the annoyance when the same page is reloaded in an iframe or a time-lapse viewer. A first, never-seen visit still plays the intro. Added a regression test (tests/test_panneau.py).

v0.29.0 — Edge-anchored links, self-reference loops (2026-09-01)

• Links attach to the box side facing the other table (left/right when side by side, top/bottom when stacked) instead of running box-centre to box-centre. No more edges starting inside a table or cutting straight through it; the layout reads like a proper ERD.
• A self-referencing foreign key (e.g. reports_to) now draws a small loop on the table's corner, so the hierarchy is visible — the old routing collapsed it to an invisible point.
• The hovered or isolated table's links are painted above the tables, so a highlighted relationship is never hidden behind a box. The SVG export uses the same anchoring. Added a headless-Chrome regression test (tests/test_liens.py).

v0.28.2 — Fix the theme toggle and a stuck audit-FK checkbox (2026-09-01)

• The theme button no longer needs two clicks to reach dark from a light OS. It cycled system → light → dark, but "system" and "light" look identical on a light OS, so the first click seemed to do nothing. A click now always changes the visible theme, skipping any step that would look the same.
• The "audit FKs" checkbox is hidden again when a model has no audit FKs. A #hud label { display: flex } rule overrode the hidden attribute, so the checkbox showed even with nothing to toggle.

v0.28.1 — Fix links pointing into space when saved positions load (2026-08-31)

• A viewer who had dragged tables saw the FK links briefly point into empty space on the next open. Restoring the saved layout re-applied each box's position after the page had settled, so the boxes animated from their auto-placement spot to their saved spot and the first links were drawn against them mid-slide. The restore now jumps the boxes into place with no animation, so the links are correct from the first paint. Added a headless-Chrome regression test (tests/test_positions.py).

v0.28.0 — Auto attribution badge on branded pages (2026-08-31)

• A branded page (--logo or --home-url) now carries a discreet clickable mcdview badge by default, linking to https://mcdview.dev. Branding drops the header link back to mcdview, so the badge keeps the attribution. An explicit --credit still sets its own text and url; a plain, unbranded page stays bare.
• New --no-credit opts out of the badge, even on a branded page.

v0.27.1 — Fix SVG export from a zoomed-out diagram (2026-08-31)

• SVG export now always draws full table boxes. Exporting while the diagram was zoomed out (header-only, level-of-detail view) sized each box to its header alone, so the columns spilled out with no background or border and the FK links anchored to the header center. The export reflows to full detail before measuring, so its output no longer depends on the current zoom level.

v0.27.0 — --to-dico data dictionary on the CLI (2026-08-30)

• --to-dico outputs a Markdown data dictionary (one ## schema.table section with a column / type / key table) instead of the HTML page — to -o if given, else stdout. Same content as the in-page export → Markdown, now scriptable for a docs/CI pipeline like --to-mermaid and --to-preview.

v0.26.2 — Auto-dismiss the hint banner (2026-08-30)

• The transient hint banner (e.g. "No probable missing FK", "No match", a cycle/type/ref count) now clears itself after ~2.6 s, so it no longer sits over the menus that open below it. The join-path prompts stay put until you pick the two tables.

v0.26.1 — Parse inline-opening multi-line CREATE TABLE (2026-08-30)

• The built-in PostgreSQL parser now reads a CREATE TABLE whose ( is followed by columns on the same line and whose body continues on later lines (closing ) elsewhere) — e.g. CREATE TABLE t (id int,\n FOREIGN KEY …\n);. Previously it sat between the two fast passes and was silently dropped in a mixed file (only rescued by sqlglot when the whole file had zero regex-parseable tables). The single-line pass now extends to the statement's closing ); within a bounded window (a C-level search, no copy), so the ReDoS/DoS budget is unchanged (the orphan-openers stress stays ~0.4 s).

v0.26.0 — SVG preview, view deep-links, schema lint (2026-08-30)

• --to-preview emits a static SVG snapshot of the diagram (auto-layout, schema-colored headers, columns, FK links) — a social/OpenGraph card image.
• View deep-links: a generated page reads its query string — ?focus=<table> isolates a table, theme=dark|light forces the theme, audit=show reveals audit FKs, and hubs / cycles open on that analysis — so a precise view can be embedded in a doc.
• --lint checks the schema against rules (missing primary key, unindexed foreign key, disconnected table, non-lower_snake_case names, FK without a target) and prints JSON violations (rule, table, column, severity); --fail-on error|warning|info turns it into a CI gate.

v0.25.1 — Firefox fit fix; grouped export menu (2026-08-30)

• Fix a Firefox regression where the diagram's foreign-key edges could appear at their provisional positions before the initial fit settled. The edges now stay hidden until a real re-fit runs — re-fit is also retried on window.load and by the safety timeout (which now re-fits, not just reveals), and it is skipped once the viewer has panned or zoomed, so it never resets a view in use.
• The SVG, Markdown and DOT exports are grouped under one ⬇ export dropdown to declutter the toolbar.

v0.25.0 — Graph analysis: join paths, impact, cycles, hub heatmap (2026-08-30)

• Join path (⇢ jointure): pick two tables and mcdview highlights the shortest foreign-key path between them and writes the SELECT … JOIN … ON … skeleton (copy button included) — the "how do I join these?" answer.
• Impact analysis: from an isolated table, "▲ depends on" / "▼ depended on by" highlight everything it reaches upstream (its dependencies) or downstream (what would break if you change it), following FKs transitively.
• Cycles (↻ cycles): highlight tables sitting on a circular foreign-key dependency (Tarjan SCC), including self-references — a schema smell that breaks insert ordering and dumps.
• Hub heatmap (◉ hubs): shade every table by its number of FK relations, so the central hub tables stand out at a glance, with a degree badge.
• Model summary (∑ résumé): tables, FKs, schemas, average columns, the most-referenced table, orphan tables (no FK) and how many sit on a cycle.
• Enriched search: type:<sql type> highlights every table with a column of that type, ref:<table> highlights the tables that reference it, and a plain query now falls back to a fuzzy (subsequence) name match.
• Text exports: a Markdown data dictionary (⬇ MD) and a Graphviz DOT graph (⬇ DOT), alongside the existing SVG export.
• Copy DDL: an isolated table's panel reconstructs its CREATE TABLE (columns, primary key, foreign keys) to the clipboard.
• Probable missing FKs (≈ FK ?): flags <x>_id columns with no declared foreign key when a table named <x> exists — a heuristic hint to review.
• Print stylesheet: printing (or saving to PDF) drops the interactive chrome and lays the diagram and panel out on white.
• All are pure client-side, single-file, zero-dependency, and available in the English UI too.

v0.24.9 — Detect FKs declared inside CREATE TABLE (2026-08-30)

• Foreign keys written inside the CREATE TABLE body are now detected: a column-level inline col int REFERENCES other(id), a table-level FOREIGN KEY (col) REFERENCES other(id) (single or composite), a column-less REFERENCES other that leans on the target's primary key, and self-references. Previously only ALTER TABLE ... ADD ... FOREIGN KEY (pg_dump style) was read, so hand-written schemas showed 0 FKs and every table floated unlinked. The sqlglot path already handled these; this fixes the built-in PostgreSQL parser.
• --diagnose gained a soft disconnected_tables anomaly (several tables and not one FK), which is almost always a parser gap rather than a real model.
• Leading-comma DDL style (,\n col TYPE, common in hand-written schemas) now parses: the comma leading each entry no longer hides the column, its primary key, or a table-level foreign key.

v0.24.8 — Version stamp moved into the detail panel (2026-08-30)

• The version stamp now sits discreetly in the detail panel's bottom-right corner, like a signature, instead of the bottom-left corner where it pushed the schema legend up. It rides with the panel: hidden while the panel is collapsed (the default), shown when the panel is open.

v0.24.7 — Discreet version stamp (2026-08-30)

• Each generated page now shows which mcdview built it, faintly in the bottom-left corner (below the schema legend). Handy when a diagram was saved or shared a while ago: you can tell at a glance which version produced it, since old pages are never regenerated in place.

v0.24.6 — Columns in the isolated view on huge models; no edge flash (2026-08-30)

• On very large models (over 800 tables) the diagram builds header-only nodes to stay fast, so a table showed as a name pill even when you zoomed in or isolated it — its columns only appeared in the side panel. Isolating a table now re-injects the columns into the focused table and its neighbours (a handful of nodes), so the isolated view reads like it does on a small model. Returning to the overview strips them back to header-only, keeping the overview cheap.
• No more edge flash on first paint (Firefox): the FK edges stay hidden until the initial refit routes them against settled boxes, then fade in over ~120 ms. The tables show immediately (no cold-start blank screen), and a safety timeout reveals the edges even if the animation frame is starved (background tab) so they can never stay hidden.

v0.24.5 — Deterministic, flash-free panel intro (2026-08-30)

• The panel intro now plays the same way on every load: the page starts with the panel collapsed (no flash of the panel before it tucks away), shows it open, then slides it shut over ~1.8s. The collapsed/open choice is sticky within the session (the toggle) but no longer remembered across loads — a past click can't make the intro silently skip or the panel snap shut.

v0.24.4 — Firefox initial-fit fix (2026-08-30)

• Firefox computed the first zoom-to-fit before layout settled, so a freshly opened page showed the tables mis-scaled and the FK links mis-routed until the first hover. The fit is now recomputed after fonts are ready and two frames have painted (a no-op in Chrome, the correction Firefox needs).
• The panel intro-close now animates reliably (the slow transition is established a frame before the transform changes, so it no longer snaps shut instantly on some browsers).

v0.24.3 — Intro reveal for the detail panel (2026-08-30)

• On first load the detail panel shows open (so its purpose is visible), then slides shut over ~2s to the collapsed default — clicking a table or the toggle cancels the intro and keeps the panel open. Skipped when a preference is remembered or the URL opens straight on a table.

v0.24.2 — Collapsible detail panel (2026-08-30)

• The detail panel can be collapsed to give the diagram the full width. It starts collapsed and a table click auto-reveals it; once you use the toggle, your choice sticks (collapsed stays collapsed on later clicks, open stays open) and is remembered for next time.

v0.24.1 — Soft-faded diagram edges (2026-08-30)

• The viewport edges fade softly, so a foreign-key link to a table currently off-screen (after panning) fades out at the border instead of stopping hard in mid-air (it used to read as "a link leading to nothing").

v0.24.0 — Robustness, --diagnose, and a big corpus sweep (2026-08-30)

• Swept mcdview over ~15,000 real-world files across every format; fixed every crash found: a phantom REFERENCES column from a wrapped FK, and several sqlglot AST edge cases (exotic defaults, constraint/comment/index nodes) — plus a safety net so no sqlglot AST shape can crash the tool.
• The built-in PostgreSQL parser reads single-line CREATE TABLE t (a int);.
• Faster --dialect auto on large non-PostgreSQL files (dialect picked on a prefix, ~4–6× less parse time).
• Rails FK columns resolve against the real columns and handle irregular plurals (people→person); Drizzle pgTableCreator factories are supported.
• --diagnose prints a JSON verdict (status, dialect, counts, anomalies) and never fails — for a hosting service to flag problematic uploads.

v0.23.3 — Single-line DDL, --version (2026-08-30)

• The built-in PostgreSQL parser now reads a single-line / compact CREATE TABLE t (a int, b numeric(10,2)); (no newline after the open paren), not only the pg_dump (\n layout. Before, such DDL only parsed when sqlglot was installed. Body splitting is depth- and string-aware, so a comma or paren inside a type or a 'literal' never splits a column; still linear, still within the DoS budget.
• --version prints the installed version.
• --db is now covered end-to-end in CI (against a throwaway PostgreSQL service), not only manually.

v0.23.2 — Index diff on the sqlglot path (2026-08-30)

• Indexes and unique constraints are now extracted on the sqlglot path too (CREATE [UNIQUE] INDEX, ALTER … ADD CONSTRAINT … UNIQUE, inline UNIQUE). Previously only the built-in PostgreSQL parser read them, so a single-line CREATE TABLE … (which routes to sqlglot) reported no index changes in a diff. MySQL UNIQUE KEY is picked up as a bonus.

v0.23.1 — Installable from PyPI (2026-08-30)

• Published to PyPI: pipx install mcdview (extra [dialects] pulls in sqlglot). The HTML template and logo ride along as bundled data; running the script straight from a checkout keeps working unchanged. Releases publish automatically on a version tag via PyPI Trusted Publishing (no stored token).
• Fixed a backslash inside an f-string expression that broke the Mermaid export on Python 3.9–3.11 (only 3.12+ tolerated it).

v0.23.0 — Big-model UX, SVG export, richer diff (2026-08-29)

• Big models: a level-of-detail mode (table names only when zoomed out), a clickable minimap, and header-only table nodes above 800 tables (≈20× fewer DOM nodes) make a 1000+ table schema usable.
• SVG export: a toolbar button downloads the diagram as a portable SVG for docs and slides.
• Diff: indexes and unique constraints are now diffed too (shown in the panel, counted in the JSON summary); renames are detected by column overlap, not just an exact match.
• Multiple input files are merged into one model (mcdview a.sql b.sql), cross-file foreign keys resolved.
• Quality of life: dragged/rearranged positions are remembered (localStorage), a light/dark toggle, keyboard shortcuts (/, r, Esc), and --watch to regenerate the page on every save.
• Indexes and unique constraints (CREATE [UNIQUE] INDEX, ADD CONSTRAINT … UNIQUE) are parsed and shown in the detail panel.

v0.22.2 — Security and correctness audit (2026-08-29)

• XSS fix: a schema fill-color from a .dbm was injected into the page unescaped — a hostile model could run script in a viewer's browser. The color is now escaped like all other model data (and ech() now escapes single quotes too, closing attribute-breakout).
• ReDoS fix: the Mermaid and Rails block extraction backtracked catastrophically on crafted input (a ~20 KB .mmd took ~18 s); both now use a bounded string search, and the DoS-budget test covers every parser, not just PostgreSQL.
• Phantom-PK fix: a DEFAULT 'PRIMARY KEY', a CHECK or an inline comment mentioning the words no longer marks a column as a primary key (regression from the inline-PK support).
• Hardening: url_sure now strips tab/newline/control chars before the scheme check (java\tscript: is neutralized); the XML DTD guard scans the whole document; malformed .mwb/.dbm inputs exit cleanly instead of dumping a traceback.
• Perf: the primary-key regexes are precompiled and guarded (~130 ms off a 5000-table parse); the column search suggests columns only on models under 300 tables (avoids a ×10 datalist blow-up).
• Internals: a nouvelle_fk factory joins nouvelle_table/nouvelle_colonne; new tests/test_parser.py; dead code removed.

v0.22.1 — Inline primary keys, revamped README (2026-08-29)

• The PostgreSQL parser now recognises a column-level primary key (id serial PRIMARY KEY), not only a separate PRIMARY KEY (...) line or an ALTER TABLE. The column gets its 🔑 and the constraint no longer leaks into the displayed type.
• README rebuilt around an animated demo, a live-rendered Mermaid diagram and a diff screenshot.

v0.22.0 — Mermaid export (2026-08-29)

• --to-mermaid renders the model as a Mermaid erDiagram. Pasted into a Markdown file, GitHub and GitLab render it natively — a static diagram in the README, no hosting needed (the interactive page stays the way to explore a big model). See docs/diagrams.md for the four examples.

v0.21.0 — Diff polish, rename detection, column search (2026-08-29)

• The diff detail panel now shows a retyped column as old → new (the old type struck through).
• Small single-schema models lay out as a roughly square block instead of a tall vertical band.
• Rename detection: a table dropped and one added with the same set of column names is shown as a rename ("renamed from X", amber) rather than a remove+add pair, and a foreign key untouched apart from the rename stays unchanged. The JSON summary carries renamed_from.
• Search now matches column names too (table.column or a bare column name): it isolates the table and highlights the row.

v0.20.2 — JSON diff summary (2026-08-29)

• --summary FILE (with --diff) writes a machine-readable JSON summary of the changes: per-category counts (tables/columns/FKs added, removed, changed) and the change list, with the previous type kept for retyped columns ("was": "numeric(10,2)"). Lets a caller show change badges or a version timeline without parsing the HTML.

v0.20.1 — Diff legend counts and "touched only" filter (2026-08-29)

• The diff legend now shows how many tables were added, changed and removed, and a "N touched" button hides the unchanged tables to frame just the ones that moved. "Overview" / Escape brings everything back.

v0.20.0 — Schema diff (2026-08-29)

• --diff BASELINE compares the model against an older version (any supported format on either side) and colors the diagram: tables, columns and foreign keys that were added (green), removed (red, struck through, kept on the diagram), or changed (amber). A legend shows the key. Useful to review a migration visually before applying it.
• Internals: the seven parsers now share a single charger() dispatch and the nouvelle_table / nouvelle_colonne factories, so every format flows through the same code — which is what makes the diff work across all of them.

v0.19.3 — Harden the untrusted-upload surface (2026-08-29)

• Model XML (.mwb, .dbm) is refused if it declares a DTD or entities, the vector for entity-expansion ("billion laughs") and XXE attacks the stdlib XML parser does not block on its own.
• The .mwb archive is read through a size cap, so a tiny file cannot decompress into gigabytes (zip bomb).
• Every external converter (pgmodeler-cli, dbml2sql, prisma, pg_dump/mysqldump) runs under a timeout and is killed if it hangs.
• tests/test_securite.py covers all four (generated on the fly, like the ReDoS bombs). This matters for a service that renders files it did not write.

v0.19.2 — Fuller Mermaid erDiagram parsing (2026-08-29)

• A crow's-foot cardinality (o{, }o) is no longer mistaken for an entity block, which previously invented a phantom one-letter table.
• Attribute comments (int id PK "SERIAL") become column comments, shown in the detail panel; %% comment lines and a direction directive inside an entity block are ignored instead of read as columns; comma-separated attribute keys (PK, FK) are accepted.
• Validated on 113 real erDiagram files (raw .mmd and inside Markdown fences): 918 tables, 1002 FKs, zero bogus columns.

v0.19.1 — Links follow tables during "rearrange" (2026-08-29)

• "Rearrange" now redraws the FK links on every frame while the tables glide to their new positions, instead of leaving them behind and snapping them into place once the CSS transition ends.

v0.19.0 — Read a live database via --db (2026-08-28)

• --db postgresql://… / --db mysql://…: dump a running database's schema (pg_dump -s for PostgreSQL, mysqldump --no-data for MySQL/MariaDB) and render it, no intermediate file needed. The positional model argument becomes optional; the page title defaults to the database name.
• MySQL passwords go through the MYSQL_PWD environment variable, never on the process command line.
• FK constraint names are now recovered on the MySQL/sqlglot path (they hang off the wrapping CONSTRAINT node), so --fk-audit works on MySQL and MariaDB models too.
• Command-line only. This feature must not be exposed on a public service: it would let a caller make the process connect to any host it can reach (SSRF). The hosted site keeps taking uploaded files only.

v0.18.0 — Drizzle ORM schema.ts input (2026-08-28)

• Read Drizzle ORM schema.ts natively: each pgTable/mysqlTable/ sqliteTable becomes a table, its type("col") fields columns (with .primaryKey()/.notNull()), and .references(() => t.col) the foreign keys; commented-out definitions are ignored

v0.17.0 — Mermaid erDiagram input (2026-08-28)

• Read Mermaid erDiagram natively, raw .mmd/.mermaid or inside a ```mermaid fence in a .md: entities become tables, attributes columns (PK marker → key), each relationship an FK from the crow's-foot ("many") side to the "one" side

v0.16.0 — Rails schema.rb input (2026-08-28)

• Read Rails db/schema.rb natively (the create_table / add_foreign_key DSL is regular): the implicit id primary key, column types, and foreign keys (default column resolved by singularizing the target table)

v0.15.0 — MySQL Workbench .mwb input (2026-08-28)

• Read MySQL Workbench .mwb models natively — the file is a zip whose document.mwb.xml (GRT object tree) is parsed with the standard library, no external tool: schemas, tables, column types, primary and foreign keys

v0.14.0 — Prisma schema input (2026-08-28)

• Read Prisma schema.prisma models, converted to SQL by prisma migrate diff; the schema's provider (postgres/mysql/sqlite/sqlserver) is detected automatically, MongoDB schemas are reported as unsupported
• The PK/FK parsers now accept double-quoted constraint names (CONSTRAINT "x_pkey" PRIMARY KEY ...), as emitted by Prisma and modern tools

v0.13.0 — DBML input (2026-08-28)

• Read dbdiagram.io .dbml models (converted to SQL by @dbml/cli), the same upstream-converter pattern as .dbm
• The FK parser now also reads unnamed constraints (ADD FOREIGN KEY ... without CONSTRAINT name), which dbml2sql and some dumps emit

v0.12.0 — Attribution badge, safer link URLs (2026-08-28)

• --credit TEXT / --credit-url URL: a slanted rubber-stamp attribution badge (mcdview logo + text) in the bottom-right corner, off by default; text escaped, optional link
• --home-url and --credit-url now reject dangerous URL schemes (javascript:, data:…), keeping only http(s)/mailto/relative links

v0.11.0 — SQL Server routing and ALTER ADD COLUMN (2026-08-28)

• SQL Server DDL ([bracket] identifiers) is routed to the sqlglot tsql parser instead of being mangled by the PostgreSQL parser
• ALTER TABLE ... ADD [COLUMN] name type columns are now read, not just the ones inside the CREATE TABLE body
• Both found by running the content-invariant test over a larger, more dialect-diverse harvested corpus

v0.10.1 — Case-insensitive key matching (2026-08-28)

• PK/FK column names are matched to columns case-insensitively (unquoted SQL identifiers are case-insensitive), so a PLAYERID primary key on a playerID column shows its 🔑 instead of being dropped as a phantom — found by the content-invariant test on a wider harvested corpus

v0.10.0 — Header logo link and custom logo (2026-08-28)

• --home-url URL: the header logo becomes a link (e.g. back to the hosting site); URL escaped, nothing changes when the option is absent
• --logo FILE: replace the header logo with your own image (svg/png/jpg…), embedded as an <img> data URI so a third-party SVG cannot run scripts

v0.9.5 — Parser fixes from content-invariant testing (2026-08-28)

• Strip quoting from column names in PK/FK lists, so a `id` or "timestamp" primary key no longer becomes a phantom (unmatched) key
• Read quoted table/schema names (CREATE TABLE "accounts"), so double-quoted DDL (e.g. Drizzle output) is parsed instead of dropped entirely
• Route obviously non-PostgreSQL files (backticks) to sqlglot rather than letting the regex parser mangle them
• New tests/test_invariants.py: checks the parsed model is internally consistent (no empty types, no phantom PK columns, FK endpoints exist)

v0.9.4 — Auto-dialect tries several parsers (2026-08-28)

• When PostgreSQL yields no table, --dialect auto now tries several sqlglot dialects and keeps the one parsing the most tables, instead of guessing a single one — recovers files that were mis-detected (a 28-table model was read as the wrong dialect and dropped entirely)

v0.9.3 — Square overview for huge schemas (2026-08-28)

• The automatic layout now sizes each schema's column height for a roughly square zone, so a large single-schema model no longer stretches into an unreadable horizontal band (a 1066-table model went from a 41:1 strip to 1.4:1); small models are unchanged

v0.9.2 — Dialect polish (2026-08-28)

• The detected dialect is shown in the toolbar counter for non-PostgreSQL models (e.g. 12 tables · 21 FK · mysql)
• sqlglot column types are lowercased to match the PostgreSQL parser (INT(11) → int(11)), keeping string literals intact

v0.9.1 — Isolated view no longer overlaps (2026-08-28)

• When isolating a table, the star of related tables now sizes its radius to the tables involved and runs a de-overlap pass (centre kept fixed), so a large neighbour (e.g. a 40-column table) no longer covers the isolated one

v0.9.0 — Multi-dialect input via sqlglot (2026-08-28)

• Read MySQL/MariaDB, SQLite and ~15 other dialects through the optional sqlglot backend (PostgreSQL stays dependency-free); --dialect flag, auto sniffs and falls back. On the harvested corpus, files that produced no page dropped from 257 to 18
• Dialect fixtures and test (tests/test_dialectes.py), sqlglot wired into CI and the pre-commit hook

v0.8.5 — Hosted at mcdview.dev (2026-08-28)

• Link to the hosted service at mcdview.dev: upload a .sql or .dbm and get a shareable link, no install needed
• README badges (hosted, demo, CI, license), repository homepage and topics for discoverability

v0.8.4 — Rearrange no longer overlaps big tables (2026-08-28)

• The rearrange layout now sizes each edge to the two tables it links (a constant target was pulling wide tables into each other), reins in weakly linked tables with a gravity term and a bounded repulsion range, and ends with a de-overlap pass — a 69-table model went from 60 overlapping pairs to zero, in a compact frame

v0.8.3 — Own the pgmodeler-cli image, reliable .dbm CI (2026-08-28)

• docker/pgmodeler-cli/ + image.yml build and publish ghcr.io/gheop/pgmodeler-cli (Fedora 44 + pgModeler 1.2.2, plus a -node variant for container: jobs); the tag is the pgModeler version baked in
• The CI .dbm job now runs inside that pinned image (fixed version → stable counts) and is blocking instead of best-effort

v0.8.2 — Tolerate pgmodeler-cli fix-model segfault (2026-08-28)

• --fix-model on pgmodeler-cli 1.2.2 writes the repaired .dbm in full, then segfaults while freeing the model (memory-layout dependent, systematic in containers). mcdview now judges the repair by the output file, not the exit code, so old .dbm models convert inside a container

v0.8.1 — CI test pipeline (2026-08-28)

• CI now runs on every push and pull request: Python error linting (ruff --select F,E9, py_compile), the pinned regression, the security suite, the strict corpus campaign, HTML structure validation and a headless-Chrome DOM check on the generated example pages; deploy to Pages only runs on main after the tests pass
• tests/valider_html.py validates a page's structure and data island
• A best-effort .dbm job installs pgmodeler-cli on the runner (never blocks a PR)

v0.8.0 — Rearrange button (2026-08-28)

• "Rearrange" button: a dependency-free force-directed relayout (Fruchterman-Reingold) that spreads tables to reduce link crossings — on a 69-table model with its audit FKs hidden, crossings dropped from 662 to 217. Disabled above 400 tables (the O(n²) pass would freeze the tab)

v0.7.0 — Hover, permalinks, hardening (2026-08-28)

• Hover a table to light up its links; #schema.table permalinks that reopen on the right table; DEFAULT values in the detail panel; table comments as tooltips; a table/FK counter in the toolbar
• Security: all injected model data is HTML-escaped and the JSON island escapes <, closing an XSS vector; the CREATE TABLE body is now bounded by string search instead of a lazy regex, killing a ReDoS (a 830 KiB malformed file went from 13 s to 0.3 s)
• Security test suite and a strict corpus campaign wired into the pre-commit hook

v0.6.1 — .dbm repair fallback (2026-08-28)

• When pgmodeler-cli refuses a .dbm (saved by an older pgModeler), mcdview now retries through --fix-model before giving up — on the harvested corpus this takes the refusal rate from 136/296 down to 2/296
• Campaign results published in the README ("Tested at scale")

v0.6.0 — Schema legend, large-scale campaign (2026-08-28)

• Schema legend: one colored chip per schema (name + table count) in the bottom-left corner, click to frame that schema's zone; hidden when the model has a single schema
• A clear hint when the input is MySQL or SQLite DDL instead of PostgreSQL
• Performance: pre-sorted BFS starts (layout of a 5000-table model drops from 894 ms to 56 ms), precompiled column regexes, tables built in one DocumentFragment and links redrawn in a single DOM write
• Large-scale harness (tests/moissonner.py, tests/grand_banc.py): hundreds of real-world schemas harvested from GitHub plus synthetic models up to 5000 tables, per-phase timings and DOM validation

v0.5.0 — Draggable tables, test bench (2026-08-28)

• Tables can be dragged around the diagram, links redraw live; a table moved in the overview keeps its new home position
• Regression/benchmark runner (tests/tester.py) over the examples, the pgModeler sample models and an optional local corpus of big real-world schemas (GitLab: 1066 tables parsed in ~0.7 s), plus a pre-commit hook

v0.4.0 — pgModeler .dbm input, CI-built pages (2026-08-28)

• A .dbm pgModeler model can be passed directly as input: the SQL is produced by pgmodeler-cli and the table positions drawn in the model are reused
• The example pages are rebuilt by GitHub Actions on every push instead of being committed
• Code comments, docstrings and CLI output translated to English

v0.3.0 — Wider DDL support, real-world examples (2026-08-28)

• The parser now accepts unqualified tables (defaulting to the public schema), an opening parenthesis on its own line, primary keys declared via ALTER TABLE (pg_dump -s style), composite foreign keys, FK targets without a column list, and partitioned tables (partitions are folded into their parent)
• 3 real-world example models with browsable pages: Pagila, Northwind, Chinook

v0.2.0 — English UI (2026-08-28)

• --lang {fr,en} option: language of the generated page's interface (search box, help text, detail panel labels)
• The live demo now uses the English UI

v0.1.1 — Public release (2026-08-28)

• Fictional library demo (exemples/mediatheque.sql) with a live version on GitHub Pages
• README in English, screenshot
• Published on GitHub

v0.1.0 — First standalone version (2026-08-27)

• Automatic table layout (one zone per schema, balanced columns, related tables pulled together) — the pgModeler .dbm becomes optional
• --titre, --fk-audit, --dbm, -o options
• SVG logo, "audit FK" checkbox hidden when no FK matches