GoSQLX v1.16.0 Changelog

Covers everything merged since v1.15.0: the October 2026 performance audit (#566) with large allocation cuts across tokenizer, parser, AST and LSP; the batch-2/batch-3 correctness sweep (round-tripping SQL(), formatter and gosqlx format, render-time injection hardening); new pkg/highlight SQL syntax highlighting across API, CLI, LSP, WASM and playground; a fully dialect-aware LSP; a hardened gosqlx-mcp HTTP server with ChatGPT app support and a Claude Code plugin; editor-accurate positions (a tab is one column); and pygosqlx 1.16.0 with dialect support, lint, security scan, transpile and prebuilt wheels. The release is source-compatible: exported signatures are unchanged apart from additions. Runtime behaviour does change (rendered SQL, token positions, error codes and text, parse acceptance including MySQL ||, CLI refusals, MCP HTTP limits), so read Behavior changes before upgrading.

Behavior changes

Token positions match editors: a tab is one column and Start skips comments: the tokenizer counts a tab as one column (it was four), so a\tb puts b at column 3. The first token after one or more comments, and the EOF token after a trailing comment, now start at the token itself rather than at the comment. Columns stay 1-based byte offsets. Parse errors, LSP diagnostics, semantic tokens, CLI output and the playground move accordingly, and error carets repeat the source line's tabs. Consumers that mapped columns back to text with tab width 4 must treat them as plain byte offsets. Comment.Start/End are unchanged. (#826, fixes #797)
|| is logical OR under MySQL and MariaDB: with the mysql or mariadb dialect (API, linter, CLI --dialect, LSP), a = 1 || b now parses as (a = 1) OR b, a BinaryExpression{Operator: "OR", PipesOr: true}, instead of a = (1 || b). SQL() and the formatter write || back. The linter and the security scanner treat it as OR. Pass parser.WithPipesAsConcat() / gosqlx.WithPipesAsConcat() to keep concatenation (PIPES_AS_CONCAT). Transpiling a || concatenation to MySQL/MariaDB now emits CONCAT(...), and a MySQL || OR becomes OR for other targets. Other dialects are unchanged. (#790, fixes #781)
Composite field access is dialect-gated: (expr).field parses under PostgreSQL, BigQuery, generic and the default dialect. MySQL, MariaDB, SQL Server, SQLite, Oracle, Snowflake, ClickHouse and Redshift return E4001 ("composite field access ... is not supported in "). Array subscripts on parenthesized values are not gated. (#843, fixes #839)
L006 reports a byte column: the column-alignment rule reports the 1-based byte column of a misaligned column's first character (a tab is one column), consistent with token and error positions. Tabs still count as 4 spaces when comparing alignment, so which lines are flagged is unchanged. (#840, fixes #830)
Parenthesized array operands keep their AST grouping: (items)[1] now retains a ParenthesizedExpression in the subscript's Array field. This preserves arithmetic precedence and lets (items)[1].name round-trip without becoming the invalid items[1].name. (#834, fixes #833)
gosqlx-mcp HTTP hardening: rate limiting now keys on the TCP peer address and no longer trusts X-Forwarded-For (which let any client evade limits and grow limiter memory without bound). Behind a proxy, set GOSQLX_MCP_TRUSTED_PROXY_HEADER (e.g. CF-Connecting-IP, as render.yaml does) or GOSQLX_MCP_TRUSTED_PROXY_HOPS. Limiter memory is capped with least-recently-seen eviction; GET (SSE) streams are capped at 4 per client, and a GET refused by that cap costs no rate-limit budget; JSON-RPC batches are charged per tool call, and a malformed non-empty body or empty batch costs 1. Denials are HTTP 429 with Retry-After and the request id (a denied initialize, alone or in a batch, keeps HTTP 200 for mcp-go clients). Oversized bodies (over 64 KB) get 413, bodies not received within 30s get 408 or a closed connection on any route, idle connections close after 120s and headers are capped at 32 KB. Bearer tokens are compared in constant time and the Bearer scheme is case-insensitive. Tool input problems (missing, empty or non-string sql, parse/format failures, wrongly typed options) are now isError tool results naming the argument instead of JSON-RPC -32603 internal errors, and indent_size must be an integer from 0 to 16. With neither trusted-proxy setting, gosqlx-mcp logs a startup warning that all clients behind a proxy share one bucket. The opt-in GOSQLX_MCP_TRUSTED_PROXY_SECRET_HEADER/GOSQLX_MCP_TRUSTED_PROXY_SECRET pair makes the trusted header and hops apply only to requests that carry the shared secret (constant-time compare); other requests are keyed by the peer address (see docs/MCP_GUIDE.md). New pkg/mcp API: NewRateLimitMiddleware(cfg, next), Config.TrustedProxyHeader, Config.TrustedProxyHops, Config.TrustedProxySecretHeader, Config.TrustedProxySecret. (#831, #841, fixes #838)
MCP format_sql uppercases more keywords: uppercase_keywords: true now also uppercases operator and function keywords (AND, OR, COUNT); indent_size affects nested blocks only, as now documented. (#823)
Comments now count toward MaxTokens: tokens plus collected comments may not exceed MaxTokens (1,000,000). Past that the tokenizer returns the existing E1007 token-limit error at the comment or token that crosses the limit. Before, comment-only input was bounded only by MaxInputSize; 10 MB of --\n collected 3.4M comments and peaked at about 960 MB RSS. Pooled tokenizers drop a Comments slice whose capacity exceeds 1024 entries on Reset, and ReleaseAST does the same for AST.Comments. (#597, part of #566)
MaxRecursionDepth (E2007) now covers unary +/-, NOT and derived tables: these three paths recursed outside parseExpression, so their nesting was limited only by input size. They now share its depth counter and RecursionDepthLimitError (limit 100): 99 nested unary operators or NOTs in a select list or WHERE parse and 100 return E2007. (#597, part of #566)
$name placeholders keep their text: $ followed by a name with no closing $ used to produce a bare "$" placeholder and drop the name. It is now one placeholder token whose value is "$name", matching @name and Oracle :name. Token counts, types and spans are unchanged, but formatter output and fingerprints of such queries change. $1, $$...$$ and $tag$...$tag$ are unaffected. (#597, part of #566)
Set-operation chains render in linear time: the formatter used to copy the rendered left side at every level of a left-deep UNION/INTERSECT/ EXCEPT chain. It now writes the chain once. Output is byte-identical; only time and memory change (80,000 UNIONs: 4.1 s and 46 GB allocated -> 6 ms and 101 MB). Format, Fingerprint, Normalize and Transpile benefit too. (#597, part of #566)
Token values reference the input (zero-copy): identifiers, numbers, comments, placeholders, bracket identifiers, dollar-quoted bodies and clean quoted literals/identifiers are now substrings of the input rather than per-token copies, and so are the AST strings built from them. Keeping any of them keeps the whole input string alive, so callers that cache values should strings.Clone them. Tokenize([]byte) copies the input once, so tokens never alias the caller's mutable buffer. ExtractTables/ExtractColumns/ ExtractFunctions/ExtractMetadata (and the Qualified variants) and schema.LoadFromDDL/LoadCatalogFromDDL clone their result strings so they don't pin the input. Tokenizer.Reset and Parser.Reset now release the previous input instead of keeping it reachable from the pool. (#595, fixes #546 and #548; #586, fixes #552)
Oracle :1 / :name are bind variables: under DialectOracle, : followed by digits or an identifier is one TokenTypePlaceholder (value includes the colon). As a consequence, Oracle-mode arr[1:n] (not valid Oracle) now fails with "expected ], got PLACEHOLDER". Other dialects are unchanged. (#595, fixes #520)
INSERT ... WITH ... SELECT is accepted, and gosqlx format keeps leading CTEs: the INSERT error for a missing source now reads "expected VALUES, SELECT, or WITH" instead of "expected VALUES or SELECT" (still E2002), and a CTE written without AS now reports "expected AS, got LPAREN" instead of "expected column name". The CLI formatter used to silently drop a leading WITH on INSERT, UPDATE and DELETE, which changed the statement's meaning; it now emits it. (#584, fixes #537)
Multi-part names and parenthesized join groups: column references with three or more parts now parse in every dialect. The extra qualifiers are folded into Identifier.Table as a dotted string (for example Table: "db.t"), and Identifier.SQL() quotes each dotted part separately. More than five parts is rejected with E2004. A leading parenthesized join group in FROM is flattened into the statement's join chain, so its AST and SQL() match the flat query. Shapes that flattening would get wrong (a comma list after or inside a group), and an alias or ClickHouse FINAL after a group, are rejected with E4001. ExtractColumnsQualified splits a dotted qualifier at its last dot (db.t.col -> {Schema: db, Table: t, Name: col}). (#585, fixes #521)
AST visitors see the tree's own nodes: Children(), Walk and Inspect now return pointers into the parent's slices for value-typed children (table references, joins, ORDER BY items, WHEN clauses, update assignments, frame bounds, MERGE source/target and similar) instead of heap copies. A visitor that mutates them now changes the AST, and the change shows in SQL() and the formatter. These pointers become invalid after ReleaseAST. (#588, fixes #557)
ReleaseAST only pools node types the parser draws from pools: Identifier, LiteralValue, BinaryExpression, FunctionCall, most other expressions and the SELECT/INSERT/UPDATE/DELETE statements are no longer put back by ReleaseAST/ReleaseStatements; they become garbage. The typed Put* functions keep full semantics for hand-built trees. Nodes returned by Get* are now fully zeroed, and typed Put* wrappers release nested subtrees (for example FunctionCall.Over/Filter) instead of dropping them. (#581, fixes #555 and #556)
Schema loaders: the PostgreSQL loader reads pg_constraint and now requires PostgreSQL 11 or later. Read-only roles now see foreign keys, and foreign keys into another schema are reported with the referenced table's bare name. Tables with no foreign keys (and MySQL tables with no indexes) carry a nil slice instead of an empty one, which serializes as JSON null. (#582, fixes #579; #570, fixes #565)
LSP Document.Lines is shared: DocumentManager.Get returns a fresh *Document header, but its Lines now aliases the manager's immutable slice. Treat it as read-only. (#573, fixes #564)
gosqlx format renders through pkg/formatter and refuses output that does not round-trip: the CLI's private renderer is gone; cmd/gosqlx/cmd/sql_formatter.go is now a thin adapter over formatter.FormatStatement. SQLFormatter, NewSQLFormatter and Format keep their signatures, FormatterOptions gains a Dialect field and AlignColumns is now ignored. CAST, ::, EXTRACT, INTERVAL, ANY/ALL, tuples, subscripts, ARRAY, ROLLUP/CUBE/GROUPING SETS, WITHIN GROUP, set-operation subqueries, ON CONFLICT, RETURNING, PIVOT and ALTER TABLE used to be corrupted with exit 0 and now format correctly. Output layout follows pkg/formatter (CLI separators ;\n\n / ; , no trailing semicolon). If the formatted text does not re-parse, or re-parses to a different query, the command exits non-zero with "formatted output does not re-parse" and format -i leaves the file untouched. It also refuses output that would drop meaningful identifier quotes or change a string literal. With --dialect mysql/mariadb the output must keep exactly the input's double-quoted tokens, because MySQL reads "..." as a string: WHERE name = "bob" is refused rather than rewritten as name = bob, and `select` rather than rewritten as "select". ANSI_QUOTES-style input ("users"."id", FROM "users") is formatted as before. (#599, #624)
Extract* and ExtractMetadata return sorted results; L006/L009 are deterministic: ExtractTables, ExtractTablesQualified, ExtractColumns, ExtractColumnsQualified, ExtractFunctions and ExtractMetadata (and the MCP extract_metadata tool, the cbinding JSON and integrations built on them) came out in random map order. Strings now sort in ascending byte order and QualifiedName values by Schema, then Table, then Name. Costs about 3-15% ns/op on the Extract benchmarks; allocs/op are unchanged. L006 now picks the earliest continuation-column indent when counts tie, and the text-fallback path of L009 lists violations in first-seen table order. Counts, messages and lines are unchanged. (#618, #619, #625)
Children() of CreateTable, FunctionDesc, TriggerExecBody and TriggerReferencing returns pointers: CreateTable.Children() returns &c.Name, &c.Columns[i] and &c.Constraints[i] instead of value copies, so type switches on *ObjectName, *ColumnDef, *TableConstraint and *FunctionDesc match and edits made inside Inspect or Walk reach the parent AST. Visit order is unchanged. (#617, #621)
Quoted identifiers that need quotes keep them on render: ast.Identifier gains Quoted, set for ANSI double-quoted names. SQL(), FormatAST, Normalize, Fingerprint and Transpile output now re-quotes such identifiers, and CTE names and column lists, CREATE [MATERIALIZED] VIEW names and column lists and FROM/JOIN table names that are reserved words or contain non-identifier characters. New helpers ast.QuoteName and ast.QuoteQualifiedName. SELECT "select" used to render as SELECT select. A name quoted only for its case ("Total", "MixedCase") renders bare, so it keeps matching the alias or column list it refers to; the AST does not record quoting for those. Fingerprints of "id" and id are unchanged and still equal; only names that need quotes (such as "select") now normalize with them. SQL Server #temp, Snowflake @stage/path and names with Unicode combining marks stay bare. (#611, #632)
Smart quotes inside ASCII-quoted tokens are content: a string or identifier opened with ' or " now ends only at the same ASCII quote. “ ” ‘ ’ « » inside it are kept as written instead of being read as the closing quote. Tokens opened with a smart quote keep their previous rules. (#632)
SQL Server table hints must be bare words: WITH (NOLOCK, UPDLOCK) is unchanged, but WITH (INDEX = ix) and quoted or bracketed hints are now parse errors; they used to be accepted and the hint silently dropped. SQL Server now accepts double-quoted constraint names (CONSTRAINT "x y"), and every dialect accepts CAST(x AS "type"). A non-ASCII spelling such as lıke is stored as LIKE; the formatter keeps WITH in WITH CHECK OPTION; and a hand-built MINUS set operator renders as the equivalent EXCEPT. (#604, #632)
Readable-style formatting emits one trailing ;: with AddSemicolon (readable style, FormatString default) only the outermost statement ends with ;. CTE bodies, IN/EXISTS/scalar subqueries, set-operation operands and the query of INSERT ... SELECT / CREATE VIEW no longer get an inner ; or ;;, and a set-operation chain ends with a single ;. Compact style is unchanged. (#602, #631)
ClickHouse SELECT ... SETTINGS ... INTO OUTFILE is rejected: the SETTINGS clause is now parsed rather than skipped, so the trailing INTO OUTFILE that used to be swallowed returns a syntax error, matching SELECT ... INTO OUTFILE elsewhere. Malformed SETTINGS lists and a bare FORMAT after SETTINGS also return syntax errors; a trailing FORMAT <name> after SETTINGS is still accepted. (#600, #631)
A bare returning alias inside an INSERT source or UPDATE ... FROM is treated as RETURNING: while parsing an INSERT source query or an UPDATE ... FROM list, an unquoted returning is no longer taken as an implicit table or select alias. Quoted returning and aliases inside nested derived tables, IN/EXISTS, CTE bodies and scalar subqueries are unaffected. RETURNING is also matched case-insensitively. (#610, #623)
Qualified names with more than 5 parts return E2004: dotted object names in FROM, DML targets and DDL (TABLE, VIEW, INDEX, USE and MySQL statements) are capped at 5 parts, the limit already applied to column references. Every caller (FROM, JOIN, INSERT/UPDATE/DELETE, CREATE, DROP, TRUNCATE, REFRESH, SHOW/DESCRIBE/REPLACE, USE, ClickHouse MV TO) reports this E2004 rather than rewrapping it as E2002 "expected table name", and Snowflake's permissive CREATE STAGE/SEQUENCE/FILE FORMAT path now rejects such names instead of keeping a truncated RawSQL. Names of 1 to 5 parts render byte-identically. (#613, #627)
Error context windows are bounded on long lines: (*Error).Error(), FormatMultiLineContext and FormatContextWindow clip source lines over 120 bytes to about 60 bytes either side of the error column (UTF-8 safe, marked with ...) and cap caret padding. A 100 KB single-line query used to produce a 205 KB E1002 string and now produces about 400 bytes. Lines of 120 bytes or less render byte-identically. (#615, #626)
Pooled tokenizers release oversized lineStarts: Tokenizer.Reset keeps the backing array only up to 8192 entries (64 KiB) and otherwise replaces it with a 16-entry slice, so a pooled tokenizer no longer holds about 72 MB after an input with millions of lines. Token output, positions and errors are unchanged. (#614, #626)
Postgres schema loader reports index columns and key flags correctly: Index.Columns is in index key order and excludes INCLUDE columns, which move to the new Index.IncludeColumns. Expression keys appear as pg_get_indexdef text instead of being dropped, and indexes on partitioned parents are listed. PK/UNIQUE column flags come from pg_constraint, so each column appears once, IsUnique is set only for a single-column UNIQUE and a SELECT-only role sees the same schema as the owner. ForeignKey gains RefSchema (empty for MySQL and SQLite). (#620, #622)
Named windows round-trip: OVER w now renders as OVER w instead of OVER (w), and WINDOW w AS (...) no longer repeats its name inside the parentheses. The non-standard WINDOW w AS w2 is normalized to WINDOW w AS (w2). (#608, #629)
LSP semantic token legend changed: the non-standard identifier token type is replaced by the standard variable, which now covers table, column and alias names and quoted identifiers (double-quoted identifiers were string before). Function names are function and type names are type. Comments and placeholders ($1, ?, :name, @var) are now emitted, as comment and parameter. AND, OR and NOT stay keyword. Positions are UTF-16 code units. Clients with custom rules keyed on the old identifier type (for example Neovim @lsp.type.identifier highlight links) must switch to variable. (#825)
pygosqlx requires Python 3.10 or later: 3.8 and 3.9 are dropped, and the package metadata lives only in pyproject.toml (setup.py is a shim that tags platform wheels). (#851)

Added

Claude Code plugin and marketplace: /plugin marketplace add ajitpratap0/GoSQLX then /plugin install gosqlx@gosqlx installs the gosqlx plugin (plugins/gosqlx/): the hosted MCP server, five SQL skills (analysis, formatting, security review, linting, dialects), four slash commands, a read-only sql-reviewer agent and an opt-out PostToolUse hook that validates edited .sql files when the gosqlx CLI is installed. (#823)
ChatGPT app support in gosqlx-mcp (OpenAI Apps SDK and MCP Apps): analyze_sql links an embedded report widget (ui://gosqlx/analysis-v1.html, text/html;profile=mcp-app); all tools now return structuredContent alongside the unchanged text JSON, carry securitySchemes/invocation-status metadata, and the server sends usage instructions. The server reports gosqlx.Version; /health reports the real tool count; the optional GOSQLX_MCP_OPENAI_APPS_CHALLENGE env var serves /.well-known/openai-apps-challenge for app verification; idle MCP sessions have their transport state swept 30–45 minutes after their last request (SessionIdleTTL, sweep every TTL/2) instead of living until restart, and session IDs are generated statelessly and redacted from transport error logs (#835). Server.Shutdown(ctx) stops the transport, and Server.Start sets read-header and idle timeouts. The analyze_sql widget caps long lists, shows failed calls as errors and avoids blocked clipboard calls. Setup: integrations/chatgpt/README.md, details: docs/CHATGPT_APP.md. (#823; #831, #835)
SQL syntax highlighting (pkg/highlight): highlight.Spans(sql, opts...) classifies SQL into sorted, non-overlapping byte-offset spans tagged with one of 13 stable categories (keyword, function, type, identifier, quoted-identifier, string, number, constant, operator, punctuation, comment, placeholder, error). It is dialect-aware via WithDialect, never fails (input after the first lexical error becomes one error span), and detects functions in call position and types in type position. Renderers: ANSI / WriteANSI (256-color, true color or attributes only), HTML / WriteHTML (classes or inline styles) and CSS, with dark, light and mono themes. Spans marshal to JSON. Built on top of it:
- gosqlx.Highlight, gosqlx.HighlightANSI and gosqlx.HighlightHTML.
- gosqlx highlight [file|SQL] (alias hl, reads stdin) with --format ansi|html|json|css, --theme, --dialect, --color auto|always|never (honors TTY, NO_COLOR and COLORTERM) and --standalone for a full HTML page.
- LSP textDocument/semanticTokens/full uses pkg/highlight (see Behavior changes), with per-line splitting of multi-line tokens and coverage of every dialect keyword and all comments.
- WASM gosqlxHighlight(sql[, dialect]) returns spans with UTF-16 offsets; the gosqlx.dev playground uses it for editor highlighting.
See docs/HIGHLIGHTING.md. (#825; #836)
Dialect-aware LSP server: the dialect from initializationOptions.dialect now drives every feature, not only semantic tokens. Diagnostics, formatting and document symbols tokenize and parse with the dialect's rules (MySQL LIMIT 5, 10 is valid under mysql, ILIKE is reported under mysql); completion and hover add the dialect's own keywords. The server also handles workspace/didChangeConfiguration with {"gosqlx": {"dialect": "..."}}: it switches dialect without a restart, re-publishes diagnostics for open documents and, when the client advertises refreshSupport, sends workspace/semanticTokens/refresh. Invalid values fall back to generic, as in initialize. The default generic dialect keeps the previous behaviour for every feature. See docs/LSP_GUIDE.md. (#828)
--dialect flag for the parse, format, analyze, and lint CLI commands: dialect-specific SQL (e.g. MySQL LIMIT offset, count) can now be parsed by these commands, matching the existing validate --dialect. Unknown dialect names are rejected up front. The default remains PostgreSQL. pkg/linter.Linter gains a SetDialect(keywords.SQLDialect) method so the linter can tokenize/parse dialect-specific syntax, and lint --security now parses with the selected dialect. (#526)
? positional placeholders for MySQL, MariaDB and SQLite: under these dialects a bare ? is now a TokenTypePlaceholder (the bind parameter used by their database/sql drivers) instead of the PostgreSQL JSON key-existence operator. PostgreSQL and the other dialects keep ?/?|/?& as JSON operators, and the default dialect is unchanged. (#527, fixes #522)
Oracle bind variables: :1, :10, :name and :new_a parse under DialectOracle and round-trip byte-for-byte through SQL() and the formatter. Previously WHERE id = :1 failed with "unexpected token: COLON". (#595, fixes #520)
Tokenizer.TokenizeString(s string) and Tokenizer.TokenizeStringContext(ctx, s string): tokenize a string in place without the []byte round trip. Both return []models.TokenWithSpan; token values reference s. (#595, fixes #546)
ClickHouse SAMPLE and ORDER BY ... WITH FILL round-trip: SAMPLE now renders, and WITH FILL [FROM x] [TO y] [STEP z] is stored in the new WithFillClause instead of being skipped. (#534)
ClickHouse tuple/array element access with a dot (d.1, tuple.2). (#535)
JOIN ... USING (cols) is kept: stored in the new JoinClause.Using (one *ast.Identifier per column, Condition left nil) and rendered as USING (c1, c2) by SQL(), pkg/formatter and gosqlx format, instead of ON cols. Children(), Walk and ReleaseAST cover the columns, and the schema validator checks each one. (#603, #630)
Parenthesized set operations: a derived table in FROM, JOIN or a comma list can be a set operation, a WITH query or a doubly parenthesized query, and a statement or set-operation operand can be written as (query), e.g. (SELECT 1) UNION (SELECT 2). New TableReference.SetSubquery and SetOperation.LeftParenthesized / RightParenthesized; parentheses that change meaning (A EXCEPT (B UNION C)) are kept on render. ORDER BY/LIMIT the AST cannot represent returns E4001. (#612, #630)
Window frame GROUPS unit, EXCLUDE clause, named-window refinement and INTERVAL qualifiers: GROUPS frames and EXCLUDE CURRENT ROW | GROUP | TIES | NO OTHERS (new WindowFrame.Exclude) parse and render; OVER (w ORDER BY z) refines an existing window via the new WindowSpec.BaseWindow; INTERVAL '1' DAY and INTERVAL '1-2' YEAR TO MONTH parse into the new IntervalExpression.Qualifier, which fixes d + INTERVAL '1' DAY rendering as INTERVAL '1' AS DAY. GROUPS and EXCLUDE are not gated by dialect. (#608, #609, #629)
INSERT forms: RETURNING after INSERT ... SELECT, INSERT ... DEFAULT VALUES (new InsertStatement.DefaultValues), MySQL/MariaDB INSERT ... SET (new InsertStatement.SetSyntax), parenthesized query sources (INSERT INTO t (SELECT ...)) and UPDATE ... FROM with a comma-separated table list. SQL() and FormatAST render DEFAULT VALUES and the SET form; UPDATE ... FROM ... JOIN returns E4001. (#610, #623)
ClickHouse table functions and parametric aggregates: numbers(10), remote(...), file, s3, cluster, generateRandom and the keyword-named url, merge, values and null parse in FROM as a TableReference with TableFunc. quantile(0.5)(DISTINCT x) keeps DISTINCT with the argument group; DISTINCT or ORDER BY in the parameter group and a third paren group return E2004. A [...] subscript or slice directly after a function call (groupArray(x)[1], array_agg(x)[1]) parses as ArraySubscriptExpression / ArraySliceExpression; SQL Server keeps count(*) [cnt] as an alias, and SQLite behaves as before. (#605, #606, #628)
ClickHouse SELECT ... SETTINGS is kept: parsed into the new SelectStatement.Settings ([]ast.QuerySetting, stored on the rightmost SELECT of a set operation) and rendered by SQL() and pkg/formatter. (#600, #631)
Rendering of IGNORE NULLS, named arguments, ON DUPLICATE KEY UPDATE and INSERT ... OUTPUT: SQL() and pkg/formatter (compact and readable) used to drop these; they now render, with name => value via the new (*ast.NamedArgument).SQL(). (#600, #631)
SQL Server table hints and PIVOT rendering: SQL() renders WITH (NOLOCK)-style hints and PIVOT/UNPIVOT with the alias as AS alias (FROM s PIVOT (...) AS p used to render as FROM s p), and pkg/formatter renders hints. (#604, #632)
ALTER TABLE rendering in pkg/formatter: FormatStatement and FormatAST render ALTER TABLE operations (ADD/DROP COLUMN, RENAME, ADD/DROP CONSTRAINT, ALTER COLUMN ...) instead of the bare token ALTER. (#599, #624)
New schema fields: db.Index.IncludeColumns and db.ForeignKey.RefSchema (see Behavior changes). (#620, #622)
ast.FieldAccessExpression: represents composite-field selection with a traversable base expression, field name, quote flag and source position. (#834, fixes #833)
Table-function column alias lists and WITH ORDINALITY: FROM and JOIN table functions accept AS g(a, b), and PostgreSQL also accepts WITH ORDINALITY with or without an alias list (unnest(arr) WITH ORDINALITY AS t(elem, n)). They are exposed as ast.TableReference.ColumnAliases (*ast.ColumnAliasList) and TableReference.WithOrdinality, plus the new dialect.Capabilities.SupportsCompositeFieldAccess. SQL() and the formatter round-trip them. (#843, fixes #839)
WithPipesAsConcat options, ast.BinaryExpression.PipesOr and dialect.Capabilities.PipesAsLogicalOr (see Behavior changes). (#790)
gosqlx.LintWithDialect(sql, dialect): runs the full built-in rule set with a per-dialect cached linter, so AST rules also run on SQL that parses only in that dialect. Unknown dialects return ErrUnsupportedDialect. (#780)
pygosqlx 1.16.0: dialects, lint, security scan and transpile: parse, validate, validate_detailed, format and the extract_* functions take an optional dialect= argument (for example dialect="sqlserver"), and pygosqlx.dialects() lists the supported names. An unknown dialect raises UnsupportedDialectError, which is also a ValueError. New pygosqlx.lint() runs the gosqlx lint rule set (L001-L010, optional dialect= and max_line_length=), pygosqlx.security_scan() returns typed SQL injection findings with severity counts, and pygosqlx.transpile(sql, from_dialect, to_dialect) converts SQL between dialects. libgosqlx gains the gosqlx_*_dialect variants and gosqlx_lint, gosqlx_security_scan, gosqlx_transpile and gosqlx_dialects; each returns a JSON string the caller frees with gosqlx_free. The existing exports are unchanged. transpile() raises the new TranspileError when SQL cannot be parsed or rewritten, and results come back as the LintResult/LintViolation and SecurityScanResult/SecurityFinding dataclasses. (#850)
pygosqlx prebuilt wheels and PyPI release pipeline: wheels bundle the native library for Linux x86_64/aarch64 (manylinux_2_17), macOS x86_64/arm64 (13+) and Windows AMD64, and are tested on Python 3.10-3.14. pygosqlx is now versioned in lockstep with GoSQLX (it was 0.2.0), and each GoSQLX release tag publishes the matching version to PyPI through Trusted Publishing (.github/workflows/pygosqlx-release.yml), starting with pip install pygosqlx==1.16.0. On a platform with no prebuilt wheel, installing from the sdist now fails with an error unless the native library has been built first, instead of installing a package that cannot load; set PYGOSQLX_ALLOW_MISSING_LIB=1 to install anyway and supply GOSQLX_LIB_PATH at runtime. (#851, #852)
rules.CoreRules / rules.DefaultCoreConfig (new package pkg/linter/rules): the core L001-L010 rule set, defined once and shared by gosqlx lint, gosqlx action, the MCP lint_sql tool, the C binding's gosqlx_lint and gosqlx.Lint. Each keeps its existing settings and output; CoreConfig covers the L003, L004 and L005 limits. (#853)

Changed

tokenizer.NewWithKeywords no longer consults its keyword table: the tokenizer's keyword field was never read, so it was removed. NewWithKeywords keeps its signature and nil check. keywords.Conflicts() now reflects only explicit keywords.New calls, because parsing no longer calls it. (#595, fixes #550)

Fixed

PostgreSQL composite-returning function queries parse correctly: (get_item(payload)).name and (get_item(payload)).* retain their base expression in a new ast.FieldAccessExpression, including quoted field names and chained field/subscript access. SQL rendering, AST traversal, extraction, security scanning and expression release preserve the nested expression. Direct JOIN LATERAL get_item(t.payload) AS f references now parse their arguments under the same dialect gate as FROM table functions. Any keyword, reserved ones included, can follow the dot ((a).order, (f(x)).from), as in PostgreSQL, and renders unquoted. Composite access is dialect-gated (see Behavior changes). (#834, fixes #833; #843, fixes #839)
Arithmetic and || on the right of a comparison or LIKE/ILIKE/REGEXP pattern parse: WHERE a = 1 + 2 failed with "expected statement, got PLUS". Comparisons still do not chain. (#734, fixes #596)
Pooled and zero-value parsers report the default capabilities: GetParser().Capabilities() and a literal &parser.Parser{} returned all-false capabilities instead of matching NewParser(). (#735, fixes #590; #789, fixes #784)
Oracle CONNECT BY accepts general expressions: PRIOR is a prefix operator, so CONNECT BY PRIOR id = parent_id + 1 parses. A column named prior is still an identifier. (#791, fixes #783)
Snowflake VariantPath renders: WHERE src:a = 1 came out as WHERE : = 1 from SQL() and the formatter. Keys that are not bare words are double-quoted. (#786, fixes #782)
Lower- and mixed-case CROSS JOIN, FULL JOIN and GROUPING SETS parse, so formatter output with lower-case keywords round-trips. (#844, fixes #842)
Playground Lint and Analyze: Lint runs all 30 rules (it ran 3, and the tab always reported clean SQL), Analyze reports a parse error instead of scoring unparseable SQL, security findings and optimization suggestions are listed, and the editor marks parse errors. (#780, fixes #737, #738, #739, #758; #796)
gosqlx.Version and gosqlx --version report the release version: v1.15.0 still reported 1.14.0, and so did the MCP server info and /health.
gosqlx_version() double free in the C binding: it returned a cached, process-wide pointer although the binding documents every result as caller-owned, so the second pygosqlx.version() call in a process freed it again and aborted with "pointer being freed was not allocated". It now returns a fresh string on every call; free it with gosqlx_free. pygosqlx.__version__ reports the packaged version, and editable installs work with current setuptools. (#847)
LSP serverInfo.version and release binaries report the real version: the LSP server reported a hard-coded 1.0.0 in serverInfo.version and now reports the library version. The gosqlx binary bundled with the VS Code extension prints the release tag from gosqlx --version; set the CLI version at build time with -ldflags "-X github.com/ajitpratap0/GoSQLX/cmd/gosqlx/cmd.Version=<version>". GoReleaser no longer passes a -X main.version flag that the example binaries ignored. (#846) An unstamped CLI now reports gosqlx.Version, so the CLI and library versions cannot drift, and release.yml fails a tag that does not match gosqlx.Version. (#853)
Tokenizer pool no longer leaks dialect state: a tokenizer whose dialect was changed via SetDialect (e.g. by validate --dialect mysql) is now restored to the default (PostgreSQL) dialect when returned to the pool via PutTokenizer. Previously Reset did not clear the dialect/keywords, so a later GetTokenizer caller could receive a tokenizer configured for the wrong dialect and silently mis-parse. The restore lives in PutTokenizer (the pool boundary), not Reset, because Reset also runs at the start of every Tokenize call where the configured dialect must persist. The keyword table is rebuilt only when a non-default dialect was configured, so the common path is unaffected. (#525)
Parentheses in expressions are preserved: (a = 1 OR b = 2) AND c = 3 used to re-render as a = 1 OR b = 2 AND c = 3, which changed the predicate. The parser now keeps grouping in a new ast.ParenthesizedExpression node. (#532, fixes #519)
IS NOT NULL no longer renders as IS NULL, and ClickHouse parametric aggregates keep their parameter group in SQL() (quantileTDigest(0.95)(value) stayed quantileTDigest(value)). (#529, fixes #530; AST half of #531)
gosqlx format keeps ClickHouse parametric aggregate parameters: the CLI renderer also dropped the (params) group. (#583, fixes #531)
gosqlx format no longer panics on NOT EXISTS (subquery). (#591, fixes #587)
Numeric window-frame offsets are kept: ROWS BETWEEN 1 PRECEDING AND CURRENT ROW rendered as ROWS BETWEEN PRECEDING AND CURRENT ROW in SQL(), pkg/formatter and the CLI. The CLI also stopped adding BETWEEN to a single-bound frame. (#592, fixes #589)
INSERT INTO t (cols) WITH c AS (...) SELECT ... parses in every dialect (it failed with E2002 "expected VALUES or SELECT, got WITH"). A WITH is also no longer dropped from a CTE query whose body is a chain of two or more set operations. (#584, fixes #537)
MySQL three-part identifiers and parenthesized join groups parse: SHOW CREATE VIEW output such as `sakila`.`city`.`city` and FROM ((a JOIN b ON ...) JOIN c ON ...) is accepted. (#585, fixes #521)
Schema loaders: PostgreSQL composite foreign keys are no longer cross-multiplied, same-named constraints on different tables stay separate, partition clone constraints are skipped, and quoted index column names decode exactly. The MySQL foreign-key join also matches on table name. (#582, fixes #579)
Schema loaders no longer deadlock or return empty tables: a *sql.DB with MaxOpenConns=1 used to deadlock and a pooled SQLite :memory: database returned tables with no columns, because cursors were nested. SQLite table names are passed as bound parameters, so names containing quotes or backslashes load, and PostgreSQL PK/UNIQUE flags no longer leak between tables that share a column name. (#570, fixes #565)
AST release-path defects: PutSelectStatement no longer releases a derived-table subquery twice through Joins[i].Left, which could make two later GetSelectStatement calls return the same object. Pooled expression nodes no longer carry stale fields. (#581, fixes #555)
Long comment runs no longer grow the stack: the tokenizer skips consecutive comments in a loop instead of recursing once per comment. (#597)
NOT EXISTS (subquery) rendered as NOT (EXISTS (...) NOT ), which failed to re-parse with E2002. SQL(), FormatAST and FormatString now render NOT EXISTS (...); the AST shape is unchanged. (#601, #632)
Table-function arguments in FROM were dropped on render: generate_series(1,10) AS g came out as generate_series g. SQL() and pkg/formatter now render the call for PostgreSQL, Snowflake, BigQuery and ClickHouse table functions, and transform.ReplaceTable renames table-function references. (#607, #628)
Bogus context in 15 parser errors: nine E2002 and six E2004 sites passed an explanatory sentence as the SQL source, so Error() printed it as line 1 of the user's query. The context is now empty and the sentence is the error's Hint; code, message and position are unchanged. The BETWEEN, LIKE, REGEXP and IN-subquery operand wrappers likewise stop printing a fake 1 | ) snippet and keep the inner error as Cause. (#616, #627)
gosqlx format dropped or mangled constructs (see Behavior changes): JOIN ... USING printed ON LIST, and WINDOW, INTERVAL, UPDATE ... FROM and DEFAULT VALUES were lost. (#599, #603, #608, #609, #610, #624)
Readable-style output that did not re-parse: inner ; and ;; removed (#602, #631); readable-style re-parse failures over the testdata corpus fell from 36 to 6.
Tokenizer error locations point at the offending construct: an unterminated dollar-quoted string ($$/ $tag$ ) and an unterminated SQL Server [identifier] now report their opening delimiter instead of the end of input, and a double-quoted identifier broken by a newline now reports its opening quote on the correct column. These errors, invalid escape sequences (which point at the character after the backslash) and StringLiteralReader errors now use the same rule as token spans: 1-based line, 1-based byte column, a tab counting as one column. Before, they used internal rune counts, so the column drifted after multibyte characters and could be 0, and a dollar quote could report the wrong line. The two errors with no position (keywords cannot be nil, input too large) report column 1 instead of 0. Error codes and messages are unchanged. (#827, follows #826)
Dialect lists now agree everywhere: the --dialect help on parse, format, analyze, lint and validate, the CLI's unknown-dialect error, config-file validation, the MCP validate_sql schema, the playground WASM bridge, the GitHub Action and the VS Code gosqlx.dialect setting all list or accept the same 11 names: generic, postgresql, mysql, mariadb, sqlserver, oracle, sqlite, snowflake, bigquery, redshift, clickhouse. Before, help listed 7, the error omitted generic and clickhouse, pkg/config rejected generic, bigquery, redshift and clickhouse, and the VS Code setting offered 6 (and suggested mysql for mariadb). The Go lists derive from the new keywords.DialectNames(), and a test checks the VS Code extension and the example configs and READMEs against it. Names stay case-sensitive in the CLI. transpile --from/--to keep their own list of transpilable dialects. (#829)

Performance

Numbers are quoted from each PR (benchstat on an Apple M4 unless stated; the #597 figures are single before/after measurements).

Linear tokenizer position tracking: line/column lookup no longer rescans from the start of the line for every token, which made long single-line input quadratic. LargeSQL_100KB 20.6 ms -> 1.96 ms, VeryLargeSQL_1MB 1755 ms -> 21.8 ms, LargeInClause 4.87 ms -> 136 us; allocs/op unchanged. Also adds ASCII fast paths for keyword and identifier scanning and a counted token-slice capacity. (#577, fixes #544, #545, #547, #549)
Zero-copy token values: tokenizer allocs/op drop 88-93% (Tokenizer/ComplexSQL 120 -> 14, MediumSQL_10KB 2823 -> 203, LargeSQL_100KB 28139 -> 1988) and gosqlx parse allocs/op drop 37-66% (Parse_ComplexQuery 193 -> 74). B/op rises 1-6% on the tokenizer, and Tokenize_SimpleQuery on the []byte path is about 3% (11 ns) slower. (#595, fixes #546)
Escape-free quoted literals are substrings: quote-heavy inputs go from 400+ to 2 allocs/op and run 31-56% faster; TokenizeString/ComplexSQL 13 -> 5 allocs/op. The unused global buffer pool is gone. (#595, fixes #548)
Dialect-aware entry points use the pools: ParseBytesWithDialect 21.8 us / 70 allocs -> 4.06 us / 28 allocs (the same as ParseBytes), ParseTree with a dialect 22.0 us -> 1.25 us, SetDialect 35.6 us / 76 allocs -> 0 allocs, transpiler.Transpile 20.2 us -> 1.69 us and transform.ParseSQLWithDialect 35.9 us -> 0.84 us. (#595, fixes #550)
Allocation-free parser preprocessing: keyword classification and compound-token expansion no longer allocate. Parser allocs/op drop 20-50% (SimpleSelect_1000_Columns 2020 -> 1018, SELECT_Complex 198 -> 114), and qualified identifiers and clause parsing lose an allocation each. (#578, fixes #551, #553, #554)
Parser-owned preprocess buffer: ParserSimpleSelect 9 -> 7 allocs/op and 1696 -> 816 B/op, ParserComplexSelect 22 -> 20 allocs/op (-64% B/op). (#586, fixes #552)
Writer-based SQL() serialization: a statement renders into one pooled buffer. BenchmarkSQL/7line 57 -> 1 allocs/op and 15line 122 -> 1, about 3x faster; output is byte-identical. (#575, fixes #558)
Children() stops heap-copying nodes: Inspect 47 -> 26 allocs/op and 3216 -> 928 B/op, Lint/Realistic 728 -> 580 allocs/op, Extract* 20-45% fewer allocations. (#588, fixes #557)
Cheap PutExpression after large lists: the work queue is cleared only to its high-water mark. PutLeafAfterBigList 2995 ns -> 19 ns. (#581, fixes #555)
ExtractMetadata in one AST pass: 114 -> 24 allocs/op, 8264 -> 2208 B/op, about 2.4-2.8x faster. (#572, fixes #561)
SplitStatements slices the source: 855 -> 1 allocs/op, 49.3 -> 6.4 KiB/op and 32-41% faster on a 200-statement script. (#569, fixes #559)
gosqlx.Lint builds its rule set once: Lint/Tiny 169 -> 137 allocs/op. L015 no longer upper-cases the whole input (1 -> 0 allocs/op, 96 KiB -> 0 B on ~100 KB of SQL). (#568, fixes #562)
AST and parser release in batch helpers: ValidateMultiple, ParseMultiple (error paths), Analyze, Transpile and the transform parse helpers return ASTs and parsers to their pools, so pool Get/Put counts balance. (#571, fixes #560)
ScanSQL literal prefilters: regexes whose required literals are absent are skipped. ScanSQL 26.9 us -> 4.7 us (-82%) and malicious-input benchmarks 56-90% faster, with identical findings. (#574, fixes #563)
LSP splits each document once per request: BuildDiagnostics on 10,000 lines 789 MiB / 512 ms -> 1.2 MiB / 6.8 ms, Hover 5296 B / 9 allocs -> 368 B / 5 allocs, SignatureHelp 3535 B / 50 allocs -> 224 B / 4 allocs, and sendMessage writes each frame once (11 -> 7 allocs). (#573, fixes #564)
Schema loaders batch their queries: Load runs 4 statements instead of 3N+1. SQLite at 100 tables: 16.7K -> 9.9K allocs/op, 679 -> 250 KiB/op, 3.59 -> 1.92 ms. (#570, fixes #565)
Linear inline-comment detection: 1 MiB of spaces followed by 20,000 comments tokenizes in 9.8 ms instead of 10.3 s. (#597)
Performance baselines refreshed: performance_baselines.json re-measured on 2026-10-13. Baselines are now about 2.5x the Apple M4 median, because GitHub's ubuntu-latest runners measure about 1.5-3.2x slower than M4 on these queries (up to ~3.2x), so a 1.5x baseline would fail CI's slowest runs. (#566)
Qualified-name parsing is linear: parseQualifiedName collects up to 5 parts in a stack array and joins them once instead of concatenating per part. A 64 KB a.a.a... chain went from 66 ms and 1.17 GB (and parsed) to 8 ms and 10.5 MB (E2004); allocs/op on 3- and 5-part names fell 10% and 25%. (#613, #627)
Long error lines are cheap to render: error context uses a bounded line scan. Long-line Error() went from 49 us / 937 KiB to 1.4 us / 1.3 KiB per op, and the context-format benchmarks are about 20% faster. (#615, #626)
Measured costs of batch-3 correctness fixes: ast.Identifier grows from 48 to 64 bytes for the Quoted flag, raising parser B/op by about 4-6%, and rendering with name quoting makes SQL() about 8% slower with unchanged allocations (#611, #632). TableReference grows from 128 to 136 bytes for parenthesized set operations: BenchmarkParserSimpleSelect +4.8% and +16 B/op (#612, #630). Reallocating lineStarts per call on very large inputs adds about 2.4% B/op above roughly 410 KB (#614, #626). Sorted Extract* results cost 3-15% ns/op on the Extract benchmarks (#618, #625).

Security

Render-time SQL injection closed: quoted names and some literal and keyword values were written into rendered SQL verbatim, so SELECT 1 AS "a; DROP TABLE x; --" FROM t came back from gosqlx.Format, SQL(), Normalize or Transpile as two statements. Every name position (SELECT, table, derived-table, JOIN, PIVOT, UPDATE/DELETE and MERGE aliases, window names, DML targets and column lists, DDL names, constraint and function names, data types) is now quoted when it is not a plain identifier or is reserved. INTERVAL, PRAGMA, table-hint and PIVOT values are escaped, and the formatter's -- UNSUPPORTED: comment covers every line. Keyword fields (join type, set operator, DROP object type, interval unit, constraint, partition, ALTER, MERGE, frame and lock types, FK actions, FETCH, SAMPLE, table options, view WITH option, SHOW kind) are checked against exact allow-lists; a hand-built value outside its list is dropped or rendered so that the output fails to parse, never as an extra statement. Found by the batch-3 adversarial review and validated with hostile-name and AST-mutation fuzzing across all dialects. (#632)
Bounded input-driven work: comment counts, nesting depth for unary operators, NOT and derived tables, comment runs and set-operation rendering are now bounded or linear (see Behavior changes). These were found by the pre-merge security review. (#597)
Tokenizer CPU-DoS on long single lines fixed by linear position tracking. (#577, fixes #544)
Code-scanning and Dependabot alerts: every GitHub Actions uses: is pinned to a commit SHA, workflows default to read-only permissions, install commands use pinned versions or lockfiles, blog slugs are validated (3 XSS alerts), Docker base images are pinned by digest, and OpenTelemetry v1.47.0, filippo.io/edwards25519 v1.1.1 and golang.org/x/sys v0.49.0 are picked up. (#567)
Scanner descends into table-function arguments: url(...), file(...) and similar ClickHouse table functions are detected as out-of-band access (also via a FROM|JOIN <fn>( pattern), and the scanner descends into array subscripts and slices so pg_sleep(5)[1] is not a bypass. The check is dialect-agnostic by design. (#605, #607, #628)
Scanner walks nested set-operation trees: checkUnionInjection handles right-nested and parenthesized UNION/EXCEPT trees, so x UNION (SELECT ... FROM pg_user UNION ...) is no longer missed after parenthesized set operations became parseable. (#612, #630)
ANSI highlighting neutralizes terminal control characters: an ESC or other control character inside a string literal or comment could inject OSC/CSI sequences when highlighting untrusted SQL. C0 controls other than tab/LF/CR are shown as Unicode control pictures (ESC becomes U+241B), DEL as U+2421, and C1 controls and invalid UTF-8 as U+FFFD. HTML output was already escaped. (#836)
gosqlx-mcp no longer logs session IDs: transport error paths reachable by an unauthenticated response-shaped POST logged the client-supplied Mcp-Session-Id. Non-allowlisted error arguments are now redacted and generated IDs scrubbed, and the stateless session-ID manager is pinned so terminated IDs are never retained. (#835)
Dependency patches: certifi 2026.10.19 (CI Sentry script) and markdown-it 14.3.3 (VS Code extension). (#837)

Internal

CI: claude-code-action v1, fork-PR review skip, Scorecard v2.4.4 and a valid glama-sync workflow (#536, #538, #539, #540); .gitattributes enforcing LF line endings for golden fixtures (#594).
Website and playground: the playground WASM is built during deployment, revalidated by browsers and no longer committed (#543, #580).
Docs: the CHANGELOG restored after the #526 merge emptied it (#593).
Website and docs: gosqlx.dev redesign (design tokens, light-mode parity, type scale, home, docs IA and navigation, compare and benchmarks pages), SEO/a11y/performance passes, /changelog with one page per release, llms.txt and raw-markdown alternates, and a reproducible competitor benchmark harness (benchmarks/compare, #719) feeding the site. (#687-#832)
CI: VS Code extension publishes with Entra ID OIDC (#685, #686), website CI and Lighthouse budgets, weekly claims-drift check, Vercel preview opt-out with [vercel skip] (#821, #832), and internal review/plan files untracked (#822). Testcontainer startup in the schema loader tests is bounded by the test deadline, so a slow image pull skips instead of failing the package (#837).
Docs: README rewrite with a compile-checked quickstart, refreshed pkg.go.dev docs for ParseTree (#713), all 30 lint rules documented (#785), the CHANGELOG 1.0.0 upgrade guide and 1.9.1 entry (#824); Server.Start/Server.Handler document that they do not apply RateLimitMiddleware, and docs/CHATGPT_APP.md explains that ChatGPT users share OpenAI's per-IP limit (#841).
CI: a Python Bindings workflow tests pygosqlx and the raw ctypes ABI on Linux and macOS with Python 3.10 and 3.13 (#847); release.yml pins the GoReleaser current and previous tags to the v* series so pygosqlx-v* tags cannot truncate the Go release changelog (#851), and the previous-tag lookup skips the floating Marketplace v1 tag (#852); vscode-publish.yml stamps the tag version into the bundled binary and its smoke test fails if --version does not match, and the README quickstart job pins the latest tagged release explicitly (#846).
Docs: new gosqlx.dev pages for Syntax Highlighting, the ChatGPT App, Upgrading GoSQLX (migration guide) and the highlight package API reference; docs/api/lsp.md documents semantic tokens and runtime dialect changes via workspace/didChangeConfiguration; docs/api/linter.md documents Linter.SetDialect and gosqlx.LintWithDialect; MCP_GUIDE.md states that Server.Start/Server.Handler apply no rate limiting unless wrapped in NewRateLimitMiddleware and that lint_sql runs only L001-L010; the migration guide and VS Code extension changelog gain their missing v1.15.0 sections; models.Location documents byte columns and UTF-16 LSP conversion. (#848) The README describes bigquery and redshift accurately: generic keyword set plus some dialect-gated grammar. (#853)

Behavior changes

Token positions match editors: a tab is one column and Start skips comments: the tokenizer counts a tab as one column (it was four), so a\tb puts b at column 3. The first token after one or more comments, and the EOF token after a trailing comment, now start at the token itself rather than at the comment. Columns stay 1-based byte offsets. Parse errors, LSP diagnostics, semantic tokens, CLI output and the playground move accordingly, and error carets repeat the source line's tabs. Consumers that mapped columns back to text with tab width 4 must treat them as plain byte offsets. Comment.Start/End are unchanged. (#826, fixes #797)

|| is logical OR under MySQL and MariaDB: with the mysql or mariadb dialect (API, linter, CLI --dialect, LSP), a = 1 || b now parses as (a = 1) OR b, a BinaryExpression{Operator: "OR", PipesOr: true}, instead of a = (1 || b). SQL() and the formatter write || back. The linter and the security scanner treat it as OR. Pass parser.WithPipesAsConcat() / gosqlx.WithPipesAsConcat() to keep concatenation (PIPES_AS_CONCAT). Transpiling a || concatenation to MySQL/MariaDB now emits CONCAT(...), and a MySQL || OR becomes OR for other targets. Other dialects are unchanged. (#790, fixes #781)

Composite field access is dialect-gated: (expr).field parses under PostgreSQL, BigQuery, generic and the default dialect. MySQL, MariaDB, SQL Server, SQLite, Oracle, Snowflake, ClickHouse and Redshift return E4001 ("composite field access ... is not supported in "). Array subscripts on parenthesized values are not gated. (#843, fixes #839)

L006 reports a byte column: the column-alignment rule reports the 1-based byte column of a misaligned column's first character (a tab is one column), consistent with token and error positions. Tabs still count as 4 spaces when comparing alignment, so which lines are flagged is unchanged. (#840, fixes #830)

Parenthesized array operands keep their AST grouping: (items)[1] now retains a ParenthesizedExpression in the subscript's Array field. This preserves arithmetic precedence and lets (items)[1].name round-trip without becoming the invalid items[1].name. (#834, fixes #833)

gosqlx-mcp HTTP hardening: rate limiting now keys on the TCP peer address and no longer trusts X-Forwarded-For (which let any client evade limits and grow limiter memory without bound). Behind a proxy, set GOSQLX_MCP_TRUSTED_PROXY_HEADER (e.g. CF-Connecting-IP, as render.yaml does) or GOSQLX_MCP_TRUSTED_PROXY_HOPS. Limiter memory is capped with least-recently-seen eviction; GET (SSE) streams are capped at 4 per client, and a GET refused by that cap costs no rate-limit budget; JSON-RPC batches are charged per tool call, and a malformed non-empty body or empty batch costs 1. Denials are HTTP 429 with Retry-After and the request id (a denied initialize, alone or in a batch, keeps HTTP 200 for mcp-go clients). Oversized bodies (over 64 KB) get 413, bodies not received within 30s get 408 or a closed connection on any route, idle connections close after 120s and headers are capped at 32 KB. Bearer tokens are compared in constant time and the Bearer scheme is case-insensitive. Tool input problems (missing, empty or non-string sql, parse/format failures, wrongly typed options) are now isError tool results naming the argument instead of JSON-RPC -32603 internal errors, and indent_size must be an integer from 0 to 16. With neither trusted-proxy setting, gosqlx-mcp logs a startup warning that all clients behind a proxy share one bucket. The opt-in GOSQLX_MCP_TRUSTED_PROXY_SECRET_HEADER/GOSQLX_MCP_TRUSTED_PROXY_SECRET pair makes the trusted header and hops apply only to requests that carry the shared secret (constant-time compare); other requests are keyed by the peer address (see docs/MCP_GUIDE.md). New pkg/mcp API: NewRateLimitMiddleware(cfg, next), Config.TrustedProxyHeader, Config.TrustedProxyHops, Config.TrustedProxySecretHeader, Config.TrustedProxySecret. (#831, #841, fixes #838)

MCP format_sql uppercases more keywords: uppercase_keywords: true now also uppercases operator and function keywords (AND, OR, COUNT); indent_size affects nested blocks only, as now documented. (#823)

Comments now count toward MaxTokens: tokens plus collected comments may not exceed MaxTokens (1,000,000). Past that the tokenizer returns the existing E1007 token-limit error at the comment or token that crosses the limit. Before, comment-only input was bounded only by MaxInputSize; 10 MB of --\n collected 3.4M comments and peaked at about 960 MB RSS. Pooled tokenizers drop a Comments slice whose capacity exceeds 1024 entries on Reset, and ReleaseAST does the same for AST.Comments. (#597, part of #566)

MaxRecursionDepth (E2007) now covers unary +/-, NOT and derived tables: these three paths recursed outside parseExpression, so their nesting was limited only by input size. They now share its depth counter and RecursionDepthLimitError (limit 100): 99 nested unary operators or NOTs in a select list or WHERE parse and 100 return E2007. (#597, part of #566)

$name placeholders keep their text: $ followed by a name with no closing $ used to produce a bare "$" placeholder and drop the name. It is now one placeholder token whose value is "$name", matching @name and Oracle :name. Token counts, types and spans are unchanged, but formatter output and fingerprints of such queries change. $1, $$...$$ and $tag$...$tag$ are unaffected. (#597, part of #566)

Set-operation chains render in linear time: the formatter used to copy the rendered left side at every level of a left-deep UNION/INTERSECT/ EXCEPT chain. It now writes the chain once. Output is byte-identical; only time and memory change (80,000 UNIONs: 4.1 s and 46 GB allocated -> 6 ms and 101 MB). Format, Fingerprint, Normalize and Transpile benefit too. (#597, part of #566)

Token values reference the input (zero-copy): identifiers, numbers, comments, placeholders, bracket identifiers, dollar-quoted bodies and clean quoted literals/identifiers are now substrings of the input rather than per-token copies, and so are the AST strings built from them. Keeping any of them keeps the whole input string alive, so callers that cache values should strings.Clone them. Tokenize([]byte) copies the input once, so tokens never alias the caller's mutable buffer. ExtractTables/ExtractColumns/ ExtractFunctions/ExtractMetadata (and the Qualified variants) and schema.LoadFromDDL/LoadCatalogFromDDL clone their result strings so they don't pin the input. Tokenizer.Reset and Parser.Reset now release the previous input instead of keeping it reachable from the pool. (#595, fixes #546 and #548; #586, fixes #552)

Oracle :1 / :name are bind variables: under DialectOracle, : followed by digits or an identifier is one TokenTypePlaceholder (value includes the colon). As a consequence, Oracle-mode arr[1:n] (not valid Oracle) now fails with "expected ], got PLACEHOLDER". Other dialects are unchanged. (#595, fixes #520)

INSERT ... WITH ... SELECT is accepted, and gosqlx format keeps leading CTEs: the INSERT error for a missing source now reads "expected VALUES, SELECT, or WITH" instead of "expected VALUES or SELECT" (still E2002), and a CTE written without AS now reports "expected AS, got LPAREN" instead of "expected column name". The CLI formatter used to silently drop a leading WITH on INSERT, UPDATE and DELETE, which changed the statement's meaning; it now emits it. (#584, fixes #537)

Multi-part names and parenthesized join groups: column references with three or more parts now parse in every dialect. The extra qualifiers are folded into Identifier.Table as a dotted string (for example Table: "db.t"), and Identifier.SQL() quotes each dotted part separately. More than five parts is rejected with E2004. A leading parenthesized join group in FROM is flattened into the statement's join chain, so its AST and SQL() match the flat query. Shapes that flattening would get wrong (a comma list after or inside a group), and an alias or ClickHouse FINAL after a group, are rejected with E4001. ExtractColumnsQualified splits a dotted qualifier at its last dot (db.t.col -> {Schema: db, Table: t, Name: col}). (#585, fixes #521)

AST visitors see the tree's own nodes: Children(), Walk and Inspect now return pointers into the parent's slices for value-typed children (table references, joins, ORDER BY items, WHEN clauses, update assignments, frame bounds, MERGE source/target and similar) instead of heap copies. A visitor that mutates them now changes the AST, and the change shows in SQL() and the formatter. These pointers become invalid after ReleaseAST. (#588, fixes #557)

ReleaseAST only pools node types the parser draws from pools: Identifier, LiteralValue, BinaryExpression, FunctionCall, most other expressions and the SELECT/INSERT/UPDATE/DELETE statements are no longer put back by ReleaseAST/ReleaseStatements; they become garbage. The typed Put* functions keep full semantics for hand-built trees. Nodes returned by Get* are now fully zeroed, and typed Put* wrappers release nested subtrees (for example FunctionCall.Over/Filter) instead of dropping them. (#581, fixes #555 and #556)

Schema loaders: the PostgreSQL loader reads pg_constraint and now requires PostgreSQL 11 or later. Read-only roles now see foreign keys, and foreign keys into another schema are reported with the referenced table's bare name. Tables with no foreign keys (and MySQL tables with no indexes) carry a nil slice instead of an empty one, which serializes as JSON null. (#582, fixes #579; #570, fixes #565)

LSP Document.Lines is shared: DocumentManager.Get returns a fresh *Document header, but its Lines now aliases the manager's immutable slice. Treat it as read-only. (#573, fixes #564)

gosqlx format renders through pkg/formatter and refuses output that does not round-trip: the CLI's private renderer is gone; cmd/gosqlx/cmd/sql_formatter.go is now a thin adapter over formatter.FormatStatement. SQLFormatter, NewSQLFormatter and Format keep their signatures, FormatterOptions gains a Dialect field and AlignColumns is now ignored. CAST, ::, EXTRACT, INTERVAL, ANY/ALL, tuples, subscripts, ARRAY, ROLLUP/CUBE/GROUPING SETS, WITHIN GROUP, set-operation subqueries, ON CONFLICT, RETURNING, PIVOT and ALTER TABLE used to be corrupted with exit 0 and now format correctly. Output layout follows pkg/formatter (CLI separators ;\n\n / ; , no trailing semicolon). If the formatted text does not re-parse, or re-parses to a different query, the command exits non-zero with "formatted output does not re-parse" and format -i leaves the file untouched. It also refuses output that would drop meaningful identifier quotes or change a string literal. With --dialect mysql/mariadb the output must keep exactly the input's double-quoted tokens, because MySQL reads "..." as a string: WHERE name = "bob" is refused rather than rewritten as name = bob, and `select` rather than rewritten as "select". ANSI_QUOTES-style input ("users"."id", FROM "users") is formatted as before. (#599, #624)

Extract* and ExtractMetadata return sorted results; L006/L009 are deterministic: ExtractTables, ExtractTablesQualified, ExtractColumns, ExtractColumnsQualified, ExtractFunctions and ExtractMetadata (and the MCP extract_metadata tool, the cbinding JSON and integrations built on them) came out in random map order. Strings now sort in ascending byte order and QualifiedName values by Schema, then Table, then Name. Costs about 3-15% ns/op on the Extract benchmarks; allocs/op are unchanged. L006 now picks the earliest continuation-column indent when counts tie, and the text-fallback path of L009 lists violations in first-seen table order. Counts, messages and lines are unchanged. (#618, #619, #625)

Children() of CreateTable, FunctionDesc, TriggerExecBody and TriggerReferencing returns pointers: CreateTable.Children() returns &c.Name, &c.Columns[i] and &c.Constraints[i] instead of value copies, so type switches on *ObjectName, *ColumnDef, *TableConstraint and *FunctionDesc match and edits made inside Inspect or Walk reach the parent AST. Visit order is unchanged. (#617, #621)

Quoted identifiers that need quotes keep them on render: ast.Identifier gains Quoted, set for ANSI double-quoted names. SQL(), FormatAST, Normalize, Fingerprint and Transpile output now re-quotes such identifiers, and CTE names and column lists, CREATE [MATERIALIZED] VIEW names and column lists and FROM/JOIN table names that are reserved words or contain non-identifier characters. New helpers ast.QuoteName and ast.QuoteQualifiedName. SELECT "select" used to render as SELECT select. A name quoted only for its case ("Total", "MixedCase") renders bare, so it keeps matching the alias or column list it refers to; the AST does not record quoting for those. Fingerprints of "id" and id are unchanged and still equal; only names that need quotes (such as "select") now normalize with them. SQL Server #temp, Snowflake @stage/path and names with Unicode combining marks stay bare. (#611, #632)

Smart quotes inside ASCII-quoted tokens are content: a string or identifier opened with ' or " now ends only at the same ASCII quote. “ ” ‘ ’ « » inside it are kept as written instead of being read as the closing quote. Tokens opened with a smart quote keep their previous rules. (#632)

SQL Server table hints must be bare words: WITH (NOLOCK, UPDLOCK) is unchanged, but WITH (INDEX = ix) and quoted or bracketed hints are now parse errors; they used to be accepted and the hint silently dropped. SQL Server now accepts double-quoted constraint names (CONSTRAINT "x y"), and every dialect accepts CAST(x AS "type"). A non-ASCII spelling such as lıke is stored as LIKE; the formatter keeps WITH in WITH CHECK OPTION; and a hand-built MINUS set operator renders as the equivalent EXCEPT. (#604, #632)

Readable-style formatting emits one trailing ;: with AddSemicolon (readable style, FormatString default) only the outermost statement ends with ;. CTE bodies, IN/EXISTS/scalar subqueries, set-operation operands and the query of INSERT ... SELECT / CREATE VIEW no longer get an inner ; or ;;, and a set-operation chain ends with a single ;. Compact style is unchanged. (#602, #631)

ClickHouse SELECT ... SETTINGS ... INTO OUTFILE is rejected: the SETTINGS clause is now parsed rather than skipped, so the trailing INTO OUTFILE that used to be swallowed returns a syntax error, matching SELECT ... INTO OUTFILE elsewhere. Malformed SETTINGS lists and a bare FORMAT after SETTINGS also return syntax errors; a trailing FORMAT <name> after SETTINGS is still accepted. (#600, #631)

A bare returning alias inside an INSERT source or UPDATE ... FROM is treated as RETURNING: while parsing an INSERT source query or an UPDATE ... FROM list, an unquoted returning is no longer taken as an implicit table or select alias. Quoted returning and aliases inside nested derived tables, IN/EXISTS, CTE bodies and scalar subqueries are unaffected. RETURNING is also matched case-insensitively. (#610, #623)

Qualified names with more than 5 parts return E2004: dotted object names in FROM, DML targets and DDL (TABLE, VIEW, INDEX, USE and MySQL statements) are capped at 5 parts, the limit already applied to column references. Every caller (FROM, JOIN, INSERT/UPDATE/DELETE, CREATE, DROP, TRUNCATE, REFRESH, SHOW/DESCRIBE/REPLACE, USE, ClickHouse MV TO) reports this E2004 rather than rewrapping it as E2002 "expected table name", and Snowflake's permissive CREATE STAGE/SEQUENCE/FILE FORMAT path now rejects such names instead of keeping a truncated RawSQL. Names of 1 to 5 parts render byte-identically. (#613, #627)

Error context windows are bounded on long lines: (*Error).Error(), FormatMultiLineContext and FormatContextWindow clip source lines over 120 bytes to about 60 bytes either side of the error column (UTF-8 safe, marked with ...) and cap caret padding. A 100 KB single-line query used to produce a 205 KB E1002 string and now produces about 400 bytes. Lines of 120 bytes or less render byte-identically. (#615, #626)

Pooled tokenizers release oversized lineStarts: Tokenizer.Reset keeps the backing array only up to 8192 entries (64 KiB) and otherwise replaces it with a 16-entry slice, so a pooled tokenizer no longer holds about 72 MB after an input with millions of lines. Token output, positions and errors are unchanged. (#614, #626)

Postgres schema loader reports index columns and key flags correctly: Index.Columns is in index key order and excludes INCLUDE columns, which move to the new Index.IncludeColumns. Expression keys appear as pg_get_indexdef text instead of being dropped, and indexes on partitioned parents are listed. PK/UNIQUE column flags come from pg_constraint, so each column appears once, IsUnique is set only for a single-column UNIQUE and a SELECT-only role sees the same schema as the owner. ForeignKey gains RefSchema (empty for MySQL and SQLite). (#620, #622)

Named windows round-trip: OVER w now renders as OVER w instead of OVER (w), and WINDOW w AS (...) no longer repeats its name inside the parentheses. The non-standard WINDOW w AS w2 is normalized to WINDOW w AS (w2). (#608, #629)

LSP semantic token legend changed: the non-standard identifier token type is replaced by the standard variable, which now covers table, column and alias names and quoted identifiers (double-quoted identifiers were string before). Function names are function and type names are type. Comments and placeholders ($1, ?, :name, @var) are now emitted, as comment and parameter. AND, OR and NOT stay keyword. Positions are UTF-16 code units. Clients with custom rules keyed on the old identifier type (for example Neovim @lsp.type.identifier highlight links) must switch to variable. (#825)

pygosqlx requires Python 3.10 or later: 3.8 and 3.9 are dropped, and the package metadata lives only in pyproject.toml (setup.py is a shim that tags platform wheels). (#851)

Added

Claude Code plugin and marketplace: /plugin marketplace add ajitpratap0/GoSQLX then /plugin install gosqlx@gosqlx installs the gosqlx plugin (plugins/gosqlx/): the hosted MCP server, five SQL skills (analysis, formatting, security review, linting, dialects), four slash commands, a read-only sql-reviewer agent and an opt-out PostToolUse hook that validates edited .sql files when the gosqlx CLI is installed. (#823)

ChatGPT app support in gosqlx-mcp (OpenAI Apps SDK and MCP Apps): analyze_sql links an embedded report widget (ui://gosqlx/analysis-v1.html, text/html;profile=mcp-app); all tools now return structuredContent alongside the unchanged text JSON, carry securitySchemes/invocation-status metadata, and the server sends usage instructions. The server reports gosqlx.Version; /health reports the real tool count; the optional GOSQLX_MCP_OPENAI_APPS_CHALLENGE env var serves /.well-known/openai-apps-challenge for app verification; idle MCP sessions have their transport state swept 30–45 minutes after their last request (SessionIdleTTL, sweep every TTL/2) instead of living until restart, and session IDs are generated statelessly and redacted from transport error logs (#835). Server.Shutdown(ctx) stops the transport, and Server.Start sets read-header and idle timeouts. The analyze_sql widget caps long lists, shows failed calls as errors and avoids blocked clipboard calls. Setup: integrations/chatgpt/README.md, details: docs/CHATGPT_APP.md. (#823; #831, #835)

SQL syntax highlighting (pkg/highlight): highlight.Spans(sql, opts...) classifies SQL into sorted, non-overlapping byte-offset spans tagged with one of 13 stable categories (keyword, function, type, identifier, quoted-identifier, string, number, constant, operator, punctuation, comment, placeholder, error). It is dialect-aware via WithDialect, never fails (input after the first lexical error becomes one error span), and detects functions in call position and types in type position. Renderers: ANSI / WriteANSI (256-color, true color or attributes only), HTML / WriteHTML (classes or inline styles) and CSS, with dark, light and mono themes. Spans marshal to JSON. Built on top of it:

gosqlx.Highlight, gosqlx.HighlightANSI and gosqlx.HighlightHTML.
gosqlx highlight [file|SQL] (alias hl, reads stdin) with --format ansi|html|json|css, --theme, --dialect, --color auto|always|never (honors TTY, NO_COLOR and COLORTERM) and --standalone for a full HTML page.
LSP textDocument/semanticTokens/full uses pkg/highlight (see Behavior changes), with per-line splitting of multi-line tokens and coverage of every dialect keyword and all comments.
WASM gosqlxHighlight(sql[, dialect]) returns spans with UTF-16 offsets; the gosqlx.dev playground uses it for editor highlighting.

See docs/HIGHLIGHTING.md. (#825; #836)

Dialect-aware LSP server: the dialect from initializationOptions.dialect now drives every feature, not only semantic tokens. Diagnostics, formatting and document symbols tokenize and parse with the dialect's rules (MySQL LIMIT 5, 10 is valid under mysql, ILIKE is reported under mysql); completion and hover add the dialect's own keywords. The server also handles workspace/didChangeConfiguration with {"gosqlx": {"dialect": "..."}}: it switches dialect without a restart, re-publishes diagnostics for open documents and, when the client advertises refreshSupport, sends workspace/semanticTokens/refresh. Invalid values fall back to generic, as in initialize. The default generic dialect keeps the previous behaviour for every feature. See docs/LSP_GUIDE.md. (#828)

--dialect flag for the parse, format, analyze, and lint CLI commands: dialect-specific SQL (e.g. MySQL LIMIT offset, count) can now be parsed by these commands, matching the existing validate --dialect. Unknown dialect names are rejected up front. The default remains PostgreSQL. pkg/linter.Linter gains a SetDialect(keywords.SQLDialect) method so the linter can tokenize/parse dialect-specific syntax, and lint --security now parses with the selected dialect. (#526)

? positional placeholders for MySQL, MariaDB and SQLite: under these dialects a bare ? is now a TokenTypePlaceholder (the bind parameter used by their database/sql drivers) instead of the PostgreSQL JSON key-existence operator. PostgreSQL and the other dialects keep ?/?|/?& as JSON operators, and the default dialect is unchanged. (#527, fixes #522)

Oracle bind variables: :1, :10, :name and :new_a parse under DialectOracle and round-trip byte-for-byte through SQL() and the formatter. Previously WHERE id = :1 failed with "unexpected token: COLON". (#595, fixes #520)

Tokenizer.TokenizeString(s string) and Tokenizer.TokenizeStringContext(ctx, s string): tokenize a string in place without the []byte round trip. Both return []models.TokenWithSpan; token values reference s. (#595, fixes #546)

ClickHouse SAMPLE and ORDER BY ... WITH FILL round-trip: SAMPLE now renders, and WITH FILL [FROM x] [TO y] [STEP z] is stored in the new WithFillClause instead of being skipped. (#534)

ClickHouse tuple/array element access with a dot (d.1, tuple.2). (#535)

JOIN ... USING (cols) is kept: stored in the new JoinClause.Using (one *ast.Identifier per column, Condition left nil) and rendered as USING (c1, c2) by SQL(), pkg/formatter and gosqlx format, instead of ON cols. Children(), Walk and ReleaseAST cover the columns, and the schema validator checks each one. (#603, #630)

Parenthesized set operations: a derived table in FROM, JOIN or a comma list can be a set operation, a WITH query or a doubly parenthesized query, and a statement or set-operation operand can be written as (query), e.g. (SELECT 1) UNION (SELECT 2). New TableReference.SetSubquery and SetOperation.LeftParenthesized / RightParenthesized; parentheses that change meaning (A EXCEPT (B UNION C)) are kept on render. ORDER BY/LIMIT the AST cannot represent returns E4001. (#612, #630)

Window frame GROUPS unit, EXCLUDE clause, named-window refinement and INTERVAL qualifiers: GROUPS frames and EXCLUDE CURRENT ROW | GROUP | TIES | NO OTHERS (new WindowFrame.Exclude) parse and render; OVER (w ORDER BY z) refines an existing window via the new WindowSpec.BaseWindow; INTERVAL '1' DAY and INTERVAL '1-2' YEAR TO MONTH parse into the new IntervalExpression.Qualifier, which fixes d + INTERVAL '1' DAY rendering as INTERVAL '1' AS DAY. GROUPS and EXCLUDE are not gated by dialect. (#608, #609, #629)

INSERT forms: RETURNING after INSERT ... SELECT, INSERT ... DEFAULT VALUES (new InsertStatement.DefaultValues), MySQL/MariaDB INSERT ... SET (new InsertStatement.SetSyntax), parenthesized query sources (INSERT INTO t (SELECT ...)) and UPDATE ... FROM with a comma-separated table list. SQL() and FormatAST render DEFAULT VALUES and the SET form; UPDATE ... FROM ... JOIN returns E4001. (#610, #623)

ClickHouse table functions and parametric aggregates: numbers(10), remote(...), file, s3, cluster, generateRandom and the keyword-named url, merge, values and null parse in FROM as a TableReference with TableFunc. quantile(0.5)(DISTINCT x) keeps DISTINCT with the argument group; DISTINCT or ORDER BY in the parameter group and a third paren group return E2004. A [...] subscript or slice directly after a function call (groupArray(x)[1], array_agg(x)[1]) parses as ArraySubscriptExpression / ArraySliceExpression; SQL Server keeps count(*) [cnt] as an alias, and SQLite behaves as before. (#605, #606, #628)

ClickHouse SELECT ... SETTINGS is kept: parsed into the new SelectStatement.Settings ([]ast.QuerySetting, stored on the rightmost SELECT of a set operation) and rendered by SQL() and pkg/formatter. (#600, #631)

Rendering of IGNORE NULLS, named arguments, ON DUPLICATE KEY UPDATE and INSERT ... OUTPUT: SQL() and pkg/formatter (compact and readable) used to drop these; they now render, with name => value via the new (*ast.NamedArgument).SQL(). (#600, #631)

SQL Server table hints and PIVOT rendering: SQL() renders WITH (NOLOCK)-style hints and PIVOT/UNPIVOT with the alias as AS alias (FROM s PIVOT (...) AS p used to render as FROM s p), and pkg/formatter renders hints. (#604, #632)

ALTER TABLE rendering in pkg/formatter: FormatStatement and FormatAST render ALTER TABLE operations (ADD/DROP COLUMN, RENAME, ADD/DROP CONSTRAINT, ALTER COLUMN ...) instead of the bare token ALTER. (#599, #624)

New schema fields: db.Index.IncludeColumns and db.ForeignKey.RefSchema (see Behavior changes). (#620, #622)

ast.FieldAccessExpression: represents composite-field selection with a traversable base expression, field name, quote flag and source position. (#834, fixes #833)

Table-function column alias lists and WITH ORDINALITY: FROM and JOIN table functions accept AS g(a, b), and PostgreSQL also accepts WITH ORDINALITY with or without an alias list (unnest(arr) WITH ORDINALITY AS t(elem, n)). They are exposed as ast.TableReference.ColumnAliases (*ast.ColumnAliasList) and TableReference.WithOrdinality, plus the new dialect.Capabilities.SupportsCompositeFieldAccess. SQL() and the formatter round-trip them. (#843, fixes #839)

WithPipesAsConcat options, ast.BinaryExpression.PipesOr and dialect.Capabilities.PipesAsLogicalOr (see Behavior changes). (#790)

gosqlx.LintWithDialect(sql, dialect): runs the full built-in rule set with a per-dialect cached linter, so AST rules also run on SQL that parses only in that dialect. Unknown dialects return ErrUnsupportedDialect. (#780)

pygosqlx 1.16.0: dialects, lint, security scan and transpile: parse, validate, validate_detailed, format and the extract_* functions take an optional dialect= argument (for example dialect="sqlserver"), and pygosqlx.dialects() lists the supported names. An unknown dialect raises UnsupportedDialectError, which is also a ValueError. New pygosqlx.lint() runs the gosqlx lint rule set (L001-L010, optional dialect= and max_line_length=), pygosqlx.security_scan() returns typed SQL injection findings with severity counts, and pygosqlx.transpile(sql, from_dialect, to_dialect) converts SQL between dialects. libgosqlx gains the gosqlx_*_dialect variants and gosqlx_lint, gosqlx_security_scan, gosqlx_transpile and gosqlx_dialects; each returns a JSON string the caller frees with gosqlx_free. The existing exports are unchanged. transpile() raises the new TranspileError when SQL cannot be parsed or rewritten, and results come back as the LintResult/LintViolation and SecurityScanResult/SecurityFinding dataclasses. (#850)

pygosqlx prebuilt wheels and PyPI release pipeline: wheels bundle the native library for Linux x86_64/aarch64 (manylinux_2_17), macOS x86_64/arm64 (13+) and Windows AMD64, and are tested on Python 3.10-3.14. pygosqlx is now versioned in lockstep with GoSQLX (it was 0.2.0), and each GoSQLX release tag publishes the matching version to PyPI through Trusted Publishing (.github/workflows/pygosqlx-release.yml), starting with pip install pygosqlx==1.16.0. On a platform with no prebuilt wheel, installing from the sdist now fails with an error unless the native library has been built first, instead of installing a package that cannot load; set PYGOSQLX_ALLOW_MISSING_LIB=1 to install anyway and supply GOSQLX_LIB_PATH at runtime. (#851, #852)

rules.CoreRules / rules.DefaultCoreConfig (new package pkg/linter/rules): the core L001-L010 rule set, defined once and shared by gosqlx lint, gosqlx action, the MCP lint_sql tool, the C binding's gosqlx_lint and gosqlx.Lint. Each keeps its existing settings and output; CoreConfig covers the L003, L004 and L005 limits. (#853)

Changed

tokenizer.NewWithKeywords no longer consults its keyword table: the tokenizer's keyword field was never read, so it was removed. NewWithKeywords keeps its signature and nil check. keywords.Conflicts() now reflects only explicit keywords.New calls, because parsing no longer calls it. (#595, fixes #550)

Fixed

PostgreSQL composite-returning function queries parse correctly: (get_item(payload)).name and (get_item(payload)).* retain their base expression in a new ast.FieldAccessExpression, including quoted field names and chained field/subscript access. SQL rendering, AST traversal, extraction, security scanning and expression release preserve the nested expression. Direct JOIN LATERAL get_item(t.payload) AS f references now parse their arguments under the same dialect gate as FROM table functions. Any keyword, reserved ones included, can follow the dot ((a).order, (f(x)).from), as in PostgreSQL, and renders unquoted. Composite access is dialect-gated (see Behavior changes). (#834, fixes #833; #843, fixes #839)

Arithmetic and || on the right of a comparison or LIKE/ILIKE/REGEXP pattern parse: WHERE a = 1 + 2 failed with "expected statement, got PLUS". Comparisons still do not chain. (#734, fixes #596)

Pooled and zero-value parsers report the default capabilities: GetParser().Capabilities() and a literal &parser.Parser{} returned all-false capabilities instead of matching NewParser(). (#735, fixes #590; #789, fixes #784)

Oracle CONNECT BY accepts general expressions: PRIOR is a prefix operator, so CONNECT BY PRIOR id = parent_id + 1 parses. A column named prior is still an identifier. (#791, fixes #783)

Snowflake VariantPath renders: WHERE src:a = 1 came out as WHERE : = 1 from SQL() and the formatter. Keys that are not bare words are double-quoted. (#786, fixes #782)

Lower- and mixed-case CROSS JOIN, FULL JOIN and GROUPING SETS parse, so formatter output with lower-case keywords round-trips. (#844, fixes #842)

Playground Lint and Analyze: Lint runs all 30 rules (it ran 3, and the tab always reported clean SQL), Analyze reports a parse error instead of scoring unparseable SQL, security findings and optimization suggestions are listed, and the editor marks parse errors. (#780, fixes #737, #738, #739, #758; #796)

gosqlx.Version and gosqlx --version report the release version: v1.15.0 still reported 1.14.0, and so did the MCP server info and /health.

gosqlx_version() double free in the C binding: it returned a cached, process-wide pointer although the binding documents every result as caller-owned, so the second pygosqlx.version() call in a process freed it again and aborted with "pointer being freed was not allocated". It now returns a fresh string on every call; free it with gosqlx_free. pygosqlx.__version__ reports the packaged version, and editable installs work with current setuptools. (#847)

LSP serverInfo.version and release binaries report the real version: the LSP server reported a hard-coded 1.0.0 in serverInfo.version and now reports the library version. The gosqlx binary bundled with the VS Code extension prints the release tag from gosqlx --version; set the CLI version at build time with -ldflags "-X github.com/ajitpratap0/GoSQLX/cmd/gosqlx/cmd.Version=<version>". GoReleaser no longer passes a -X main.version flag that the example binaries ignored. (#846) An unstamped CLI now reports gosqlx.Version, so the CLI and library versions cannot drift, and release.yml fails a tag that does not match gosqlx.Version. (#853)

Tokenizer pool no longer leaks dialect state: a tokenizer whose dialect was changed via SetDialect (e.g. by validate --dialect mysql) is now restored to the default (PostgreSQL) dialect when returned to the pool via PutTokenizer. Previously Reset did not clear the dialect/keywords, so a later GetTokenizer caller could receive a tokenizer configured for the wrong dialect and silently mis-parse. The restore lives in PutTokenizer (the pool boundary), not Reset, because Reset also runs at the start of every Tokenize call where the configured dialect must persist. The keyword table is rebuilt only when a non-default dialect was configured, so the common path is unaffected. (#525)

Parentheses in expressions are preserved: (a = 1 OR b = 2) AND c = 3 used to re-render as a = 1 OR b = 2 AND c = 3, which changed the predicate. The parser now keeps grouping in a new ast.ParenthesizedExpression node. (#532, fixes #519)

IS NOT NULL no longer renders as IS NULL, and ClickHouse parametric aggregates keep their parameter group in SQL() (quantileTDigest(0.95)(value) stayed quantileTDigest(value)). (#529, fixes #530; AST half of #531)

gosqlx format keeps ClickHouse parametric aggregate parameters: the CLI renderer also dropped the (params) group. (#583, fixes #531)

gosqlx format no longer panics on NOT EXISTS (subquery). (#591, fixes #587)

Numeric window-frame offsets are kept: ROWS BETWEEN 1 PRECEDING AND CURRENT ROW rendered as ROWS BETWEEN PRECEDING AND CURRENT ROW in SQL(), pkg/formatter and the CLI. The CLI also stopped adding BETWEEN to a single-bound frame. (#592, fixes #589)

INSERT INTO t (cols) WITH c AS (...) SELECT ... parses in every dialect (it failed with E2002 "expected VALUES or SELECT, got WITH"). A WITH is also no longer dropped from a CTE query whose body is a chain of two or more set operations. (#584, fixes #537)

MySQL three-part identifiers and parenthesized join groups parse: SHOW CREATE VIEW output such as `sakila`.`city`.`city` and FROM ((a JOIN b ON ...) JOIN c ON ...) is accepted. (#585, fixes #521)

Schema loaders: PostgreSQL composite foreign keys are no longer cross-multiplied, same-named constraints on different tables stay separate, partition clone constraints are skipped, and quoted index column names decode exactly. The MySQL foreign-key join also matches on table name. (#582, fixes #579)

Schema loaders no longer deadlock or return empty tables: a *sql.DB with MaxOpenConns=1 used to deadlock and a pooled SQLite :memory: database returned tables with no columns, because cursors were nested. SQLite table names are passed as bound parameters, so names containing quotes or backslashes load, and PostgreSQL PK/UNIQUE flags no longer leak between tables that share a column name. (#570, fixes #565)

AST release-path defects: PutSelectStatement no longer releases a derived-table subquery twice through Joins[i].Left, which could make two later GetSelectStatement calls return the same object. Pooled expression nodes no longer carry stale fields. (#581, fixes #555)

Long comment runs no longer grow the stack: the tokenizer skips consecutive comments in a loop instead of recursing once per comment. (#597)

NOT EXISTS (subquery) rendered as NOT (EXISTS (...) NOT ), which failed to re-parse with E2002. SQL(), FormatAST and FormatString now render NOT EXISTS (...); the AST shape is unchanged. (#601, #632)

Table-function arguments in FROM were dropped on render: generate_series(1,10) AS g came out as generate_series g. SQL() and pkg/formatter now render the call for PostgreSQL, Snowflake, BigQuery and ClickHouse table functions, and transform.ReplaceTable renames table-function references. (#607, #628)

Bogus context in 15 parser errors: nine E2002 and six E2004 sites passed an explanatory sentence as the SQL source, so Error() printed it as line 1 of the user's query. The context is now empty and the sentence is the error's Hint; code, message and position are unchanged. The BETWEEN, LIKE, REGEXP and IN-subquery operand wrappers likewise stop printing a fake 1 | ) snippet and keep the inner error as Cause. (#616, #627)

gosqlx format dropped or mangled constructs (see Behavior changes): JOIN ... USING printed ON LIST, and WINDOW, INTERVAL, UPDATE ... FROM and DEFAULT VALUES were lost. (#599, #603, #608, #609, #610, #624)

Readable-style output that did not re-parse: inner ; and ;; removed (#602, #631); readable-style re-parse failures over the testdata corpus fell from 36 to 6.

Tokenizer error locations point at the offending construct: an unterminated dollar-quoted string ($$/ $tag$ ) and an unterminated SQL Server [identifier] now report their opening delimiter instead of the end of input, and a double-quoted identifier broken by a newline now reports its opening quote on the correct column. These errors, invalid escape sequences (which point at the character after the backslash) and StringLiteralReader errors now use the same rule as token spans: 1-based line, 1-based byte column, a tab counting as one column. Before, they used internal rune counts, so the column drifted after multibyte characters and could be 0, and a dollar quote could report the wrong line. The two errors with no position (keywords cannot be nil, input too large) report column 1 instead of 0. Error codes and messages are unchanged. (#827, follows #826)

Dialect lists now agree everywhere: the --dialect help on parse, format, analyze, lint and validate, the CLI's unknown-dialect error, config-file validation, the MCP validate_sql schema, the playground WASM bridge, the GitHub Action and the VS Code gosqlx.dialect setting all list or accept the same 11 names: generic, postgresql, mysql, mariadb, sqlserver, oracle, sqlite, snowflake, bigquery, redshift, clickhouse. Before, help listed 7, the error omitted generic and clickhouse, pkg/config rejected generic, bigquery, redshift and clickhouse, and the VS Code setting offered 6 (and suggested mysql for mariadb). The Go lists derive from the new keywords.DialectNames(), and a test checks the VS Code extension and the example configs and READMEs against it. Names stay case-sensitive in the CLI. transpile --from/--to keep their own list of transpilable dialects. (#829)

Performance

Numbers are quoted from each PR (benchstat on an Apple M4 unless stated; the #597 figures are single before/after measurements).

Linear tokenizer position tracking: line/column lookup no longer rescans from the start of the line for every token, which made long single-line input quadratic. LargeSQL_100KB 20.6 ms -> 1.96 ms, VeryLargeSQL_1MB 1755 ms -> 21.8 ms, LargeInClause 4.87 ms -> 136 us; allocs/op unchanged. Also adds ASCII fast paths for keyword and identifier scanning and a counted token-slice capacity. (#577, fixes #544, #545, #547, #549)

Zero-copy token values: tokenizer allocs/op drop 88-93% (Tokenizer/ComplexSQL 120 -> 14, MediumSQL_10KB 2823 -> 203, LargeSQL_100KB 28139 -> 1988) and gosqlx parse allocs/op drop 37-66% (Parse_ComplexQuery 193 -> 74). B/op rises 1-6% on the tokenizer, and Tokenize_SimpleQuery on the []byte path is about 3% (11 ns) slower. (#595, fixes #546)

Escape-free quoted literals are substrings: quote-heavy inputs go from 400+ to 2 allocs/op and run 31-56% faster; TokenizeString/ComplexSQL 13 -> 5 allocs/op. The unused global buffer pool is gone. (#595, fixes #548)

Dialect-aware entry points use the pools: ParseBytesWithDialect 21.8 us / 70 allocs -> 4.06 us / 28 allocs (the same as ParseBytes), ParseTree with a dialect 22.0 us -> 1.25 us, SetDialect 35.6 us / 76 allocs -> 0 allocs, transpiler.Transpile 20.2 us -> 1.69 us and transform.ParseSQLWithDialect 35.9 us -> 0.84 us. (#595, fixes #550)

Allocation-free parser preprocessing: keyword classification and compound-token expansion no longer allocate. Parser allocs/op drop 20-50% (SimpleSelect_1000_Columns 2020 -> 1018, SELECT_Complex 198 -> 114), and qualified identifiers and clause parsing lose an allocation each. (#578, fixes #551, #553, #554)

Parser-owned preprocess buffer: ParserSimpleSelect 9 -> 7 allocs/op and 1696 -> 816 B/op, ParserComplexSelect 22 -> 20 allocs/op (-64% B/op). (#586, fixes #552)

Writer-based SQL() serialization: a statement renders into one pooled buffer. BenchmarkSQL/7line 57 -> 1 allocs/op and 15line 122 -> 1, about 3x faster; output is byte-identical. (#575, fixes #558)

Children() stops heap-copying nodes: Inspect 47 -> 26 allocs/op and 3216 -> 928 B/op, Lint/Realistic 728 -> 580 allocs/op, Extract* 20-45% fewer allocations. (#588, fixes #557)

Cheap PutExpression after large lists: the work queue is cleared only to its high-water mark. PutLeafAfterBigList 2995 ns -> 19 ns. (#581, fixes #555)

ExtractMetadata in one AST pass: 114 -> 24 allocs/op, 8264 -> 2208 B/op, about 2.4-2.8x faster. (#572, fixes #561)

SplitStatements slices the source: 855 -> 1 allocs/op, 49.3 -> 6.4 KiB/op and 32-41% faster on a 200-statement script. (#569, fixes #559)

gosqlx.Lint builds its rule set once: Lint/Tiny 169 -> 137 allocs/op. L015 no longer upper-cases the whole input (1 -> 0 allocs/op, 96 KiB -> 0 B on ~100 KB of SQL). (#568, fixes #562)

AST and parser release in batch helpers: ValidateMultiple, ParseMultiple (error paths), Analyze, Transpile and the transform parse helpers return ASTs and parsers to their pools, so pool Get/Put counts balance. (#571, fixes #560)

ScanSQL literal prefilters: regexes whose required literals are absent are skipped. ScanSQL 26.9 us -> 4.7 us (-82%) and malicious-input benchmarks 56-90% faster, with identical findings. (#574, fixes #563)

LSP splits each document once per request: BuildDiagnostics on 10,000 lines 789 MiB / 512 ms -> 1.2 MiB / 6.8 ms, Hover 5296 B / 9 allocs -> 368 B / 5 allocs, SignatureHelp 3535 B / 50 allocs -> 224 B / 4 allocs, and sendMessage writes each frame once (11 -> 7 allocs). (#573, fixes #564)

Schema loaders batch their queries: Load runs 4 statements instead of 3N+1. SQLite at 100 tables: 16.7K -> 9.9K allocs/op, 679 -> 250 KiB/op, 3.59 -> 1.92 ms. (#570, fixes #565)

Linear inline-comment detection: 1 MiB of spaces followed by 20,000 comments tokenizes in 9.8 ms instead of 10.3 s. (#597)

Performance baselines refreshed: performance_baselines.json re-measured on 2026-10-13. Baselines are now about 2.5x the Apple M4 median, because GitHub's ubuntu-latest runners measure about 1.5-3.2x slower than M4 on these queries (up to ~3.2x), so a 1.5x baseline would fail CI's slowest runs. (#566)

Qualified-name parsing is linear: parseQualifiedName collects up to 5 parts in a stack array and joins them once instead of concatenating per part. A 64 KB a.a.a... chain went from 66 ms and 1.17 GB (and parsed) to 8 ms and 10.5 MB (E2004); allocs/op on 3- and 5-part names fell 10% and 25%. (#613, #627)

Long error lines are cheap to render: error context uses a bounded line scan. Long-line Error() went from 49 us / 937 KiB to 1.4 us / 1.3 KiB per op, and the context-format benchmarks are about 20% faster. (#615, #626)

Measured costs of batch-3 correctness fixes: ast.Identifier grows from 48 to 64 bytes for the Quoted flag, raising parser B/op by about 4-6%, and rendering with name quoting makes SQL() about 8% slower with unchanged allocations (#611, #632). TableReference grows from 128 to 136 bytes for parenthesized set operations: BenchmarkParserSimpleSelect +4.8% and +16 B/op (#612, #630). Reallocating lineStarts per call on very large inputs adds about 2.4% B/op above roughly 410 KB (#614, #626). Sorted Extract* results cost 3-15% ns/op on the Extract benchmarks (#618, #625).

Security

Render-time SQL injection closed: quoted names and some literal and keyword values were written into rendered SQL verbatim, so SELECT 1 AS "a; DROP TABLE x; --" FROM t came back from gosqlx.Format, SQL(), Normalize or Transpile as two statements. Every name position (SELECT, table, derived-table, JOIN, PIVOT, UPDATE/DELETE and MERGE aliases, window names, DML targets and column lists, DDL names, constraint and function names, data types) is now quoted when it is not a plain identifier or is reserved. INTERVAL, PRAGMA, table-hint and PIVOT values are escaped, and the formatter's -- UNSUPPORTED: comment covers every line. Keyword fields (join type, set operator, DROP object type, interval unit, constraint, partition, ALTER, MERGE, frame and lock types, FK actions, FETCH, SAMPLE, table options, view WITH option, SHOW kind) are checked against exact allow-lists; a hand-built value outside its list is dropped or rendered so that the output fails to parse, never as an extra statement. Found by the batch-3 adversarial review and validated with hostile-name and AST-mutation fuzzing across all dialects. (#632)

Bounded input-driven work: comment counts, nesting depth for unary operators, NOT and derived tables, comment runs and set-operation rendering are now bounded or linear (see Behavior changes). These were found by the pre-merge security review. (#597)

Tokenizer CPU-DoS on long single lines fixed by linear position tracking. (#577, fixes #544)

Code-scanning and Dependabot alerts: every GitHub Actions uses: is pinned to a commit SHA, workflows default to read-only permissions, install commands use pinned versions or lockfiles, blog slugs are validated (3 XSS alerts), Docker base images are pinned by digest, and OpenTelemetry v1.47.0, filippo.io/edwards25519 v1.1.1 and golang.org/x/sys v0.49.0 are picked up. (#567)

Scanner descends into table-function arguments: url(...), file(...) and similar ClickHouse table functions are detected as out-of-band access (also via a FROM|JOIN <fn>( pattern), and the scanner descends into array subscripts and slices so pg_sleep(5)[1] is not a bypass. The check is dialect-agnostic by design. (#605, #607, #628)

Scanner walks nested set-operation trees: checkUnionInjection handles right-nested and parenthesized UNION/EXCEPT trees, so x UNION (SELECT ... FROM pg_user UNION ...) is no longer missed after parenthesized set operations became parseable. (#612, #630)

ANSI highlighting neutralizes terminal control characters: an ESC or other control character inside a string literal or comment could inject OSC/CSI sequences when highlighting untrusted SQL. C0 controls other than tab/LF/CR are shown as Unicode control pictures (ESC becomes U+241B), DEL as U+2421, and C1 controls and invalid UTF-8 as U+FFFD. HTML output was already escaped. (#836)

gosqlx-mcp no longer logs session IDs: transport error paths reachable by an unauthenticated response-shaped POST logged the client-supplied Mcp-Session-Id. Non-allowlisted error arguments are now redacted and generated IDs scrubbed, and the stateless session-ID manager is pinned so terminated IDs are never retained. (#835)

Dependency patches: certifi 2026.10.19 (CI Sentry script) and markdown-it 14.3.3 (VS Code extension). (#837)

Internal

CI: claude-code-action v1, fork-PR review skip, Scorecard v2.4.4 and a valid glama-sync workflow (#536, #538, #539, #540); .gitattributes enforcing LF line endings for golden fixtures (#594).

Website and playground: the playground WASM is built during deployment, revalidated by browsers and no longer committed (#543, #580).

Docs: the CHANGELOG restored after the #526 merge emptied it (#593).

Website and docs: gosqlx.dev redesign (design tokens, light-mode parity, type scale, home, docs IA and navigation, compare and benchmarks pages), SEO/a11y/performance passes, /changelog with one page per release, llms.txt and raw-markdown alternates, and a reproducible competitor benchmark harness (benchmarks/compare, #719) feeding the site. (#687-#832)

CI: VS Code extension publishes with Entra ID OIDC (#685, #686), website CI and Lighthouse budgets, weekly claims-drift check, Vercel preview opt-out with [vercel skip] (#821, #832), and internal review/plan files untracked (#822). Testcontainer startup in the schema loader tests is bounded by the test deadline, so a slow image pull skips instead of failing the package (#837).

Docs: README rewrite with a compile-checked quickstart, refreshed pkg.go.dev docs for ParseTree (#713), all 30 lint rules documented (#785), the CHANGELOG 1.0.0 upgrade guide and 1.9.1 entry (#824); Server.Start/Server.Handler document that they do not apply RateLimitMiddleware, and docs/CHATGPT_APP.md explains that ChatGPT users share OpenAI's per-IP limit (#841).

CI: a Python Bindings workflow tests pygosqlx and the raw ctypes ABI on Linux and macOS with Python 3.10 and 3.13 (#847); release.yml pins the GoReleaser current and previous tags to the v* series so pygosqlx-v* tags cannot truncate the Go release changelog (#851), and the previous-tag lookup skips the floating Marketplace v1 tag (#852); vscode-publish.yml stamps the tag version into the bundled binary and its smoke test fails if --version does not match, and the README quickstart job pins the latest tagged release explicitly (#846).

Docs: new gosqlx.dev pages for Syntax Highlighting, the ChatGPT App, Upgrading GoSQLX (migration guide) and the highlight package API reference; docs/api/lsp.md documents semantic tokens and runtime dialect changes via workspace/didChangeConfiguration; docs/api/linter.md documents Linter.SetDialect and gosqlx.LintWithDialect; MCP_GUIDE.md states that Server.Start/Server.Handler apply no rate limiting unless wrapped in NewRateLimitMiddleware and that lint_sql runs only L001-L010; the migration guide and VS Code extension changelog gain their missing v1.15.0 sections; models.Location documents byte columns and UTF-16 LSP conversion. (#848) The README describes bigquery and redshift accurately: generic keyword set plus some dialect-gated grammar. (#853)

v1.16.0 — Performance Audit, SQL Highlighting, Dialect-Aware LSP, Claude Code Plugin & ChatGPT App, Correctness Sweep

Behavior changes

Added

Changed

Fixed

Performance

Security

Internal

v1.16.0 — Performance Audit, SQL Highlighting, Dialect-Aware LSP, Claude Code Plugin & ChatGPT App, Correctness Sweep

Behavior changes

Added

Changed

Fixed

Performance

Security

Internal

Behavior changes#

Added#

Changed#

Fixed#

Performance#

Security#

Internal#

Behavior changes#

Added#

Changed#

Fixed#

Performance#

Security#

Internal#

Behavior changes

Added

Changed

Fixed

Performance

Security

Internal

Behavior changes

Added

Changed

Fixed

Performance

Security

Internal