[brandur]

🤖 Align SQLite URLs across the Go, Java, JavaScript, Ruby, and Rust CLIs. Both sqlite:path and sqlite://path select the same database, including sqlite::memory: and sqlite://:memory:. Offline migrate-get also accepts the bare sqlite: and sqlite:// dialect hints without opening a database.

Go and JavaScript now accept the short form. Java strips the optional // before calling JDBC so relative paths stay relative and the long memory form works. Ruby translates either prefix for the selected Active Record or Sequel adapter. Rust's SQL export now accepts the same short forms as its migration commands. Ruby's sqlite3: aliases and Java's native jdbc:sqlite: URLs remain supported.

Regression tests cover file paths, alternate forms reopening the same database, memory databases, and offline export. The shared URL reference documents portable forms and driver-specific aliases/options.

Validation:

• GOMAXPROCS=4 GOFLAGS=-p=2 make test lint.
• Java CLI tests, executable integration tests, formatting, and Maven verification on JDK 21.
• JavaScript build, 121 CLI tests, lint, and formatting.
• Full Ruby suite, final CLI regression tests, lint, and Steep type checks.
• Rust CLI migration/export tests with both backends and each backend alone; make lint/rust on Rust 1.99.1.

[bgentry]

🤖 Agent review summary. Blake reviewed each of these comments before they were posted.

Issues that likely warrant fixing before merge:

• Active Record misreads relative paths: a regression that can target the wrong file, with a one-line fix and spec change.
• bench fails on sqlite::memory:: with a one-line fix and regression test.
• An empty path reports success on a throwaway database: with a Go fix and regression test. The other CLIs would need the same rule.

Non-blocking:

• Java upgrade note
• URLs that select different files by language

The merge conflict with master is only in CHANGELOG.md.

Separately, and already on master: PGDATABASE overrides --database-url in the Go CLI, because RunCommand checks pgEnvConfigured() first. With PGDATABASE set, migrate-up --database-url sqlite:f.db ignores the URL and goes to Postgres. That contradicts the CLI's help text, which says --database-url takes precedence, and could run migrations against the wrong database. It's outside this PR, but probably deserves its own fix.

[bgentry]

🤖 Regression for Active Record: it reads the first segment after // as a host, so this breaks relative paths and can target the wrong file. For example, sqlite3://storage/s.db resolves to host storage and database /s.db, and sqlite3://rel.db resolves to no database at all. That means sqlite3:rel.db now fails with "No database file specified", and sqlite3:storage/s.db targets /s.db, which would succeed for a process that can write to /. Both worked before this change. It also breaks the documented sqlite://storage/river.sqlite3.

Dropping the // fixes it. A subdirectory path, a plain filename, and an absolute path all created the right file through the real CLI:

Suggested change

	"sqlite3://#{sqlite_path}"
	"sqlite3:#{sqlite_path}"

The specs pass by accident because the relative path climbs out through ../../../, and dropping the first .. as the "host" still lands on the right absolute file. Creating the temp directory under the working directory exposes this. In ruby/spec/migration_cli_shared_examples.rb:

-      Dir.mktmpdir("river-cli-urls-") do |directory|
+      # Stay under the working directory so the relative path has no `../`
+      # prefix, which makes an adapter that misreads its first segment visible.
+      Dir.mktmpdir("river-cli-urls-", Dir.pwd) do |directory|

With that spec change alone, all four prefixes fail on Active Record and pass on Sequel. With the fix as well, the spec file passes.

[brandur]

🤖 Fixed in 9c435d99: Active Record now receives sqlite3:#{sqlite_path} without the extra //. The regression spec creates its temporary directory under the working directory, as suggested. All four prefixes failed on Active Record before the fix and now pass with both Active Record and Sequel.

[bgentry]

🤖 Accepting the short form here exposes a gap in bench. Its in-memory migration check at river_cli.go:415 only looks at parsedDatabaseURL.Host, but url.Parse("sqlite::memory:") leaves Host empty and puts :memory: in Opaque. So river bench --database-url sqlite::memory: skips the migration and fails with SQL logic error: no such table: river_job (1), while sqlite://:memory: works.

One-line fix in cmd/river/rivercli/river_cli.go:

-		if parsedDatabaseURL.Host == ":memory:" {
+		if parsedDatabaseURL.Host == ":memory:" || parsedDatabaseURL.Opaque == ":memory:" {

Suggested regression test, as the first subtest in TestBaseCommandSetSQLiteURLs in river_cli_test.go. The sqlite::memory: case fails on this branch and passes with the fix:

t.Run("BenchMemory", func(t *testing.T) {
	t.Parallel()

	for _, databaseURL := range []string{"sqlite::memory:", "sqlite://:memory:"} {
		t.Run(databaseURL, func(t *testing.T) {
			t.Parallel()

			runFunc(t, databaseURL, "bench", "--num-total-jobs", "10")
		})
	}
})

[brandur]

🤖 Fixed in 9c435d99 by checking Opaque as well as Host. Added the benchmark regression test for both memory URL forms; the short form reproduced the missing-table error before the fix, and both now pass. Full Go make test and make lint also passed.

[bgentry]

🤖 Bare forms are documented only for migrate-get, but migration commands accept them too, and SQLite treats an empty path as a temporary database. So migrate-up --database-url sqlite: reports every migration as applied and exits 0 on Go, Sequel, Java, and Rust, with nothing saved. An unset variable in sqlite:$DB_PATH would hit this and look like a successful deploy step. Active Record raises a raw exception instead, and JavaScript rejects it with a usage error.

For Go, the bare sqlite: is new in this PR, while sqlite:// already behaved this way. Could every CLI reject an empty path except for migrate-get? A tested Go version, in RunCommand in command.go, though there may be a cleaner way to mark commands that don't need a database:

 			case "sqlite":
+				if _, isMigrateGet := any(command).(*migrateGet); urlWithoutProtocol == "" && !isMigrateGet {
+					return false, fmt.Errorf("SQLite database URL (`%s`) needs a path or `:memory:`", *databaseURL)
+				}
+
 				dbPool, err := openSQLitePool(protocol, urlWithoutProtocol)

Suggested regression test, as a subtest between DialectHint and File in TestBaseCommandSetSQLiteURLs. It fails on this branch and passes with the change above:

t.Run("EmptyPathRejected", func(t *testing.T) {
	t.Parallel()

	for _, databaseURL := range []string{"sqlite:", "sqlite://"} {
		t.Run(databaseURL, func(t *testing.T) {
			t.Parallel()

			cli := NewCLI(&Config{Name: "River"})
			cli.SetOut(&bytes.Buffer{})
			cmd := cli.BaseCommandSet()
			cmd.SetArgs([]string{"migrate-up", "--database-url", databaseURL})
			require.ErrorContains(t, cmd.Execute(), "needs a path")
		})
	}
})

[brandur]

🤖 Fixed in 9c435d99. Go, Java, Ruby, and Rust now reject empty SQLite paths before connecting; JavaScript already rejected them and now has matching regression coverage. Tests cover each supported migration command, Go's benchmark, Ruby's sqlite3: aliases, and Java's bare jdbc:sqlite: form. Bare prefixes remain valid for migrate-get, with export tests passing across all five languages. The shared URL reference now documents this rule.

[bgentry]

🤖 Non-blocking, and probably low impact given how few Java users there are yet: sqlite://var/lib/x.db used to open the absolute /var/lib/x.db and now opens ./var/lib/x.db. Worth a brief mention here if you think anyone is relying on it.

[brandur]

🤖 Added an upgrade note to the Java changelog in 9c435d99: sqlite://var/lib/x.db now resolves to ./var/lib/x.db, and callers intending the absolute path should use sqlite:///var/lib/x.db.

[bgentry]

🤖 Non-blocking docs suggestion. "Not portable" undersells this a little, because a URL can silently select a different file by language rather than just lose an option. For example:

• sqlite:q.db?mode=rwc opens q.db on Go and Ruby, but creates a file literally named q.db?mode=rwc on Java and JavaScript.
• sqlite:a%20b.db opens a b.db on Ruby and Rust, but a%20b.db on Go, Java, and JavaScript.
• sqlite::memory:?cache=shared stays in memory on Go, but creates an on-disk file on Java and JavaScript.

Maybe add something like: "Query parameters and percent-encoding are handled differently by each language, and can select a different file. For portable URLs, avoid ? and % in paths, and don't add parameters to :memory:."

[brandur]

🤖 Updated the shared URL reference in 9c435d99 to explain that query parameters and percent encoding can select different files, including turning an in-memory database into an on-disk file. It now recommends avoiding ? and % in portable paths and omitting parameters from :memory:.

[brandur]

Unfortunately the Rust advisories job is failing and there's kind of nothing we can really do about it — there'a SQLite security advisory. libsqlite3-sys fixes it in 0.39, but SQLx requires <= 0.9.0 so there's nothing actionable we can do on our end at the moment.

[bgentry]

@brandur yep no worries, I split apart our required checks so that the Rust advisory one is not marked as required since we don't want it blocking the merge of totally unrelated changes. Thanks for looking into that one!