@reponova/lang-sql

SQL language support for RepoNova. Regex-based, statement-oriented parser — no tree-sitter grammar required.

Multi-dialect tolerant by design: PostgreSQL, MySQL, SQLite, T-SQL (SQL Server), and BigQuery / Snowflake all share one extractor. The plugin focuses on the schema-level DDL surface RepoNova's knowledge graph cares about (tables, views, functions, procedures, triggers, indexes, types, sequences, schemas) plus the cross-references between them. Query-body analysis is intentionally out of scope.

Install

reponova lang add @reponova/lang-sql

What it extracts

The extractor pre-processes the source (strips comments, neutralises string literals and PostgreSQL dollar-quoted bodies, honours MySQL DELIMITER ;; / // / $$ directives), splits the source into top-level statements, then dispatches each CREATE / ALTER statement to a per-construct regex matcher.

Symbols (fileNode.kind is always "module"; tagged ["sql"])

| SQL construct | SymbolNode.kind | First decorator | Notes | | --- | --- | --- | --- | | CREATE TABLE (incl. TEMP, UNLOGGED, FOREIGN, VIRTUAL) | class | table | Modifier captured as a secondary decorator (temporary, unlogged, …). Schema-qualified names produce parent: "<schema>". | | CREATE VIEW | class | view | | | CREATE MATERIALIZED VIEW | class | materialized_view | | | CREATE FUNCTION | function | function | PostgreSQL PL/pgSQL bodies (between $$ … $$) are neutralised — symbols inside are not extracted. | | CREATE PROCEDURE / PROC | method | procedure | Supports CREATE OR ALTER PROC (T-SQL) and CREATE DEFINER=… PROCEDURE (MySQL). | | CREATE TRIGGER | function | trigger | Target table captured as a references edge. | | CREATE [UNIQUE] INDEX (incl. CLUSTERED, NONCLUSTERED, BITMAP, FULLTEXT, SPATIAL) | variable | index | unique carried as a secondary decorator. Target table captured as a references edge. | | CREATE TYPE … AS ENUM | enum | enum | | | CREATE TYPE (composite / range / object) | type | type | | | CREATE DOMAIN | type | domain | | | CREATE SEQUENCE | constant | sequence | | | CREATE SCHEMA | module | schema | CREATE SCHEMA AUTHORIZATION user (no explicit name) is silently skipped. |

Cross-references

| Source | Target | Edge kind | | --- | --- | --- | | FOREIGN KEY (col) REFERENCES other(col) (in CREATE TABLE body) | other table | extends | | Inline column-level col TYPE REFERENCES other(col) | other table | extends | | ALTER TABLE foo ADD CONSTRAINT … FOREIGN KEY … REFERENCES bar (the pg_dump / mysqldump style) | other table | extends | | FROM other_table / JOIN other_table (in view / function / procedure body) | other table | references | | CALL / EXEC / EXECUTE other_proc (in function / procedure body) | other proc | calls | | CREATE INDEX … ON other_table | other table | references | | CREATE TRIGGER … ON other_table | other table | references |

Extensions

.sql, .ddl, .dml, .psql, .pgsql, .tsql

Configuration

In reponova.yml:

plugins:
  sql:
    enabled: true       # default: true
    # patterns: []      # override global patterns for SQL files
    # exclude: []       # override global exclude for SQL files

| Property | Type | Default | Description | | --- | --- | --- | --- | | enabled | boolean | true | Enable / disable SQL file detection and extraction | | patterns | string[] | [] | Glob patterns to override global file matching for this plugin | | exclude | string[] | [] | Glob patterns to override global exclusions for this plugin |

Dialect handling

• Identifier quoting: "double", `back`, [bracket], bare. Three-part qualified names (db.schema.name) collapse to schema.name (the database segment is dropped — RepoNova works inside a single connection).
• Comments: -- line, # line (MySQL, only when anchored at the start of a line), /* block */.
• Strings: single quotes with '' escape and PostgreSQL E'…' style escapes. String bodies are neutralised so tokens like FROM/JOIN inside data strings don't produce false-positive edges.
• PostgreSQL dollar-quoted strings: $$ … $$ and $tag$ … $tag$. Bodies are neutralised — function-body symbols are NOT extracted today.
• MySQL DELIMITER: DELIMITER ;;, DELIMITER //, DELIMITER $$ are tracked across the file. Statements end at the active delimiter, NOT at every ;. DELIMITER $$ in a MySQL file disables PG dollar-quote handling for $$ (the two notations are mutually exclusive in practice).
• T-SQL BEGIN … END: nested BEGIN/END blocks are tracked so internal ; characters inside procedure bodies don't terminate the outer CREATE PROCEDURE. END IF, END LOOP, END WHILE, END CASE, END REPEAT close their respective control-flow constructs without affecting the BEGIN depth counter.
• CREATE modifier clauses: OR REPLACE, OR ALTER, DEFINER = user@host (with any quoting), SQL SECURITY DEFINER|INVOKER, ALGORITHM = UNDEFINED|MERGE|TEMPTABLE are all accepted between CREATE and the object keyword (in any order).

Resolution semantics

• No imports: SQL has no module / import semantics. resolveImportPath always returns []. Schema-qualified references are graph edges, not imports.
• Out of scope (intentionally):
  • Query-body symbol extraction (CTE names, subquery aliases) — these are scoped to a single statement and not stable identifiers.
  • CREATE AGGREGATE, CREATE OPERATOR, CREATE RULE, CREATE EVENT, CREATE EXTENSION — niche PostgreSQL/MySQL features outside the schema-level surface the graph layer consumes.
  • PL/pgSQL function-body call graphs — body content is neutralised by the dollar-quote stripper.

Why regex instead of tree-sitter-sql?

The roadmap originally proposed Archetype A with DerekStride/tree-sitter-sql. Investigation showed that no pre-built tree-sitter-sql.wasm is published anywhere in the ecosystem (not by upstream, not by @vscode/tree-sitter-wasm, not by @cursorless/tree-sitter-wasms). Adopting it would have required adding emscripten + tree-sitter-cli to the monorepo's CI matrix solely for one plugin — and the grammar also ships with a known scanner bug requiring a local patch. Meanwhile the schema-level surface RepoNova actually consumes (CREATE TABLE / VIEW / FUNCTION / PROCEDURE / INDEX / TYPE / SEQUENCE + FK & query refs) is well-bounded and regex-tractable across dialects, and a regex extractor gives multi-dialect tolerance for free instead of locking us to one grammar. See ROADMAP §6.1 for details.

License

MIT — see LICENSE.