[richardwooding]

Summary

Adds a seventh dialect, OpenSearch SQL (the OpenSearch SQL plugin, POST _plugins/_sql), plus an index-mapping type provider, following the add-sql-dialect pattern.

Dialect — dialect/opensearch

• MySQL-flavoured syntax: CONCAT(), LOCATE(), MOD(), ? placeholders, DATE_ADD/DATE_SUB, CAST(x AS LONG|DOUBLE|STRING|BOOLEAN|TIMESTAMP), (DAYOFWEEK(x) + 5) % 7 (Monday=0 like the other dialects).
• matches() → (col REGEXP 'p') = 1 — the plugin's REGEXP returns an integer, so a bare REGEXP in WHERE fails.
• object fields are dot paths (p.dims.width); has() → IS NOT NULL.
• nested fields: p.variants.exists(v, v.sku == "A") → EXISTS (SELECT * FROM p.variants AS v WHERE v.sku = 'A') (the plugin's PartiQL form).
• in against a literal list → IN (…); anything else, plus array indexing, size() of a list, split/join, bytes literals, filter/map/exists_one/all and time-zone extraction, is rejected at conversion time with ErrUnsupportedDialectFeature rather than emitting SQL the engine cannot run.
• Index analysis: not supported (OpenSearch indexes every mapped field itself; no IndexAdvisor).

Provider — opensearch

• NewTypeProviderWithClient(ctx, *opensearchapi.Client) (caller owns the client), LoadTableSchema(ctx, index) via GET /<index>/_mapping.
• object properties → nested schemas (ObjectType("products.dims")), nested properties → lists of nested schemas; multi-fields ignored; index names validated against OpenSearch naming rules.
• New direct dependency: github.com/opensearch-project/opensearch-go/v4 (only opensearchapi is imported).

Two additions to dialect.Dialect (pass-through in the six existing dialects, no output changes for them)

• SupportsSubqueryExpressions() bool — gates filter/map/exists_one/transformList, which are written as value-position subqueries. WriteArraySubqueryOpen is void, so this is the only way a dialect can refuse them.
• WriteBooleanFieldPredicate(w, writeField) — a boolean column used directly as a predicate (whole condition, AND/OR/NOT operand, CASE WHEN condition, comprehension predicate). OpenSearch writes col = true: its plugin silently matches every document for a bare col and rejects NOT col.

Verification

• make lint clean; go test -short -race green for every package (the Docker-only pg testcontainer tests were skipped locally).
• All 65 shared test cases have an OpenSearch entry (check_testcase_coverage.py opensearch).
• Live verification: opensearch_integration_test.go starts opensearchproject/opensearch:3.5.0 via testcontainers, loads the mapping through the provider, and executes the generated SQL for 24 expressions (comparisons, bare and negated booleans, in, regex incl. inline (?i), contains, ternary, EXTRACT, timestamp literal + arithmetic, object dot paths, has(), nested exists, parameterized ? binding). It skips when no container runtime is reachable. Findings from that run shaped the dialect: format=json is V1-only (the test uses the default JDBC format), bare booleans are ignored by the engine, REGEXP returns int, and the nested-query handler rejects NOT/!=/IN/REGEXP inside the subquery (documented in the README's OpenSearch notes).

Docs

README (count, badge, comparison table, provider import, index table, placeholder list, an "OpenSearch SQL notes" section), CLAUDE.md, docs/parameterized-queries.md, CHANGELOG [Unreleased], and the add-sql-dialect / add-cel-feature skill references.

🤖 Generated with Claude Code