sql2er

Use the bundled CLI from the repo root:

AGENT=skills/sql2er/scripts/sql2er-agent.mjs

State lives in one JSON file passed with --state <path>; reuse the same state path for every command in a diagram session.

Core workflow

1. Generate the diagram:

node $AGENT generate --input schema.sql --state er.json

Use --hide-attrs only when the user wants an overview/skeleton, the schema is dense, or attributes are not important. Attribute visibility is fixed at generate; export never adds or removes attributes. If fields, PKs, or column-level semantics matter, generate with attributes.

Use --comment when table/column/ref comments carry readable business labels. Set labels before tuning layout because regenerating resets positions.

If generated names are not semantic enough, relabel nodes before layout tuning:

node $AGENT labels mode comment --state er.json
node $AGENT labels set rel-orders-users-user_id-7 "placed by" --state er.json
node $AGENT labels batch --file labels.json --state er.json

2. Inspect diagnostics:

node $AGENT describe --state er.json

Use DIAGNOSTICS for structural decisions: crossings, overlaps, disconnected components, attribute overlaps, and attribute-line crossings. Use the ASCII map only as a rough orientation aid. Copy exact entity ids from the ENTITIES block for manual position edits.

3. Improve layout as needed:

node $AGENT layout optimal --state er.json
node $AGENT attrs compact --state er.json
node $AGENT swap entity-users-0 entity-orders-1 --state er.json
node $AGENT move entity-users-0 600 400 --state er.json
node $AGENT nudge entity-users-0 20 -10 --state er.json

Automatic overlap avoidance is on by default and moves attributes first, then relationship diamonds, never entity rectangles. Use avoid off only when the user explicitly wants to preserve exact positions.

Prefer layout optimal after major edits. Use attrs compact for the smallest attribute footprint and attrs moderate for a more even ring. Use swap, move, and nudge only when diagnostics or visual review show a local issue.

4. Export and review:

node $AGENT export svg --out er.svg --state er.json
node $AGENT export png --out er.png --state er.json
node $AGENT export drawio --out er.drawio --state er.json
node $AGENT export html --out er.html --state er.json --lang zh

Export HTML only when the user wants a portable manual fine-tuning editor; choose --lang zh|en for the page language. Always export SVG or PNG and inspect the image before finishing. Ensure labels are legible, the layout is balanced, and nodes/attribute ellipses do not overlap or sit on top of lines. If overlap or crossings are provably unavoidable, minimize them and make each remaining crossing/overlap clear.

When HTML is exported, tell the user the best manual fine-tuning path: open the HTML, enable continuous force layout briefly, turn it off, enable automatic avoidance once, turn it off, double-click nodes to edit the label text inside them, then hand-tune the remaining local details.

Cleanup rules

• crossings / overlaps: run layout optimal; for a small local crossing, try swap.
• attrOverlaps / attrCrossings: compare attrs compact and attrs moderate, then keep the clearer result.
• Persistent attribute issues: run describe --details to identify the exact attribute/node/line pairs before local edits.
• Unwanted automatic movement: run avoid off; turn it back on with avoid on before final review when possible.
• COMPONENTS > 1: keep related clusters in one diagram, or use export <fmt> --split for separate files.
• Awkward aspect ratio: try rotate 90.
• Busy attribute-heavy diagram: regenerate with --hide-attrs only if a skeleton-only answer satisfies the user.

References

• Read references/commands.md for full command syntax, flags, describe output schema, export behavior, and recipes.
• Read references/data-model.md when node ids, edge types, cardinality, placeholder entities, self-references, or semantic labels matter.