Merge pull request 'feat(dbml): @postgres/@sqlite dialect directives (#19)' (#27) from issue-19-dbml-directives into master

Reviewed-on: #27

cmd/relspec/job.go

@@ -1,6 +1,7 @@
 package main
 
 import (
+	"bytes"
 	"fmt"
 	"io"
 	"os"
@@ -11,6 +12,8 @@ import (
 
 	"github.com/spf13/cobra"
 
+	"git.warky.dev/wdevs/relspecgo/pkg/diff"
+	"git.warky.dev/wdevs/relspecgo/pkg/inspector"
 	"git.warky.dev/wdevs/relspecgo/pkg/jobs"
 	"git.warky.dev/wdevs/relspecgo/pkg/merge"
 	"git.warky.dev/wdevs/relspecgo/pkg/models"
@@ -18,6 +21,7 @@ import (
 	"git.warky.dev/wdevs/relspecgo/pkg/readers/sqldir"
 	"git.warky.dev/wdevs/relspecgo/pkg/writers"
 	wpgsql "git.warky.dev/wdevs/relspecgo/pkg/writers/pgsql"
+	"git.warky.dev/wdevs/relspecgo/pkg/writers/sqlexec"
 	wtemplate "git.warky.dev/wdevs/relspecgo/pkg/writers/template"
 )
 
@@ -122,6 +126,9 @@ func loadJobSet() (*jobs.Set, error) {
 	if err := set.Validate(); err != nil {
 		return nil, err
 	}
+	for _, w := range set.Warnings {
+		fmt.Fprintf(os.Stderr, "warning: %s\n", w)
+	}
 	return set, nil
 }
 
@@ -180,12 +187,14 @@ func executeJobPlan(set *jobs.Set, name string, dryRun, noDeps bool, out io.Writ
 	// Pre-flight: resolve and check paths, output policy and env vars for the
 	// whole plan before anything runs. A failure here means no job executes.
 	resolved := make([]*resolvedJob, len(plan))
+	byName := make(map[string]*resolvedJob, len(plan))
 	for i, j := range plan {
-		rj, perr := preflightJob(j)
+		rj, perr := preflightJob(j, byName)
 		if perr != nil {
 			return fmt.Errorf("job %q: %w", j.Name, perr)
 		}
 		resolved[i] = rj
+		byName[j.Name] = rj
 	}
 
 	if dryRun {
@@ -217,7 +226,12 @@ type resolvedJob struct {
 	outputConn    string // resolved connection string (secret)
 	outputConnEnv string
 	logPath       string
+	logPolicy     jobs.LogPolicy
 	templatePath  string
+	reportPath    string // "" for a diff summary written to the log
+	reportFormat  string
+	rulesPath     string // "" means inspector defaults
+	selection     *splitSelection
 	secrets       []string // resolved secret values to redact from logs
 }
 
@@ -226,11 +240,12 @@ type resolvedInput struct {
 	path    string // "" when the input is a database
 	conn    string // resolved connection string (secret)
 	connEnv string
+	fromJob string // producer job name when this input came from from_job
 }
 
-func preflightJob(j *jobs.Job) (*resolvedJob, error) {
+func preflightJob(j *jobs.Job, resolvedByName map[string]*resolvedJob) (*resolvedJob, error) {
 	root := j.Dir()
-	rj := &resolvedJob{job: j, root: root}
+	rj := &resolvedJob{job: j, root: root, logPolicy: j.ResolvedLogPolicy()}
 
 	if j.Logfile != "" {
 		p, err := jobs.SafeJoin(root, j.Logfile)
@@ -253,6 +268,20 @@ func preflightJob(j *jobs.Job) (*resolvedJob, error) {
 
 	for i, in := range j.Inputs {
 		ri := resolvedInput{format: strings.ToLower(in.Format)}
+		if in.FromJob != "" {
+			producer, ok := resolvedByName[in.FromJob]
+			if !ok {
+				return nil, fmt.Errorf("input[%d]: from_job %q is not in this plan (do not use --no-deps with from_job inputs)", i, in.FromJob)
+			}
+			if producer.outputPath == "" {
+				return nil, fmt.Errorf("input[%d]: from_job %q does not write a file output", i, in.FromJob)
+			}
+			ri.path = producer.outputPath
+			ri.format = strings.ToLower(producer.job.Output.Format)
+			ri.fromJob = in.FromJob
+			rj.inputs = append(rj.inputs, ri)
+			continue
+		}
 		if in.ConnEnv != "" {
 			v, ok := os.LookupEnv(in.ConnEnv)
 			if !ok || v == "" {
@@ -313,6 +342,43 @@ func preflightJob(j *jobs.Job) (*resolvedJob, error) {
 			rj.outputPath = p
 		}
 	}
+
+	if j.Rules != "" {
+		p, err := jobs.SafeJoin(root, j.Rules)
+		if err != nil {
+			return nil, fmt.Errorf("rules: %w", err)
+		}
+		info, err := os.Stat(p)
+		if err != nil || info.IsDir() {
+			return nil, fmt.Errorf("rules %q: not found or is a directory", j.Rules)
+		}
+		rj.rulesPath = p
+	}
+
+	if j.Report != nil {
+		rj.reportFormat = strings.ToLower(j.Report.Format)
+		if j.Report.Path != "" {
+			p, err := jobs.SafeJoin(root, j.Report.Path)
+			if err != nil {
+				return nil, fmt.Errorf("report: %w", err)
+			}
+			if _, err := os.Stat(p); err == nil && !j.Report.Overwrite {
+				return nil, fmt.Errorf("report %s already exists (set report.overwrite: true to replace it)", j.Report.Path)
+			}
+			rj.reportPath = p
+		}
+	}
+
+	if j.Select != nil {
+		rj.selection = &splitSelection{
+			Schemas:        j.Select.Schemas,
+			Tables:         j.Select.Tables,
+			ExcludeSchemas: j.Select.ExcludeSchemas,
+			ExcludeTables:  j.Select.ExcludeTables,
+			DatabaseName:   j.Select.DatabaseName,
+		}
+	}
+
 	return rj, nil
 }
 
@@ -325,9 +391,12 @@ func printResolvedJob(out io.Writer, n, total int, rj *resolvedJob) {
 	}
 	fmt.Fprintf(out, "      job file: %s\n", j.SourceFile)
 	for _, ri := range rj.inputs {
-		if ri.path != "" {
+		switch {
+		case ri.fromJob != "":
+			fmt.Fprintf(out, "      input: %s (%s) from job %q\n", ri.path, ri.format, ri.fromJob)
+		case ri.path != "":
 			fmt.Fprintf(out, "      input: %s (%s)\n", ri.path, ri.format)
-		} else {
+		default:
 			fmt.Fprintf(out, "      input: env:%s (%s)\n", ri.connEnv, ri.format)
 		}
 	}
@@ -339,15 +408,31 @@ func printResolvedJob(out io.Writer, n, total int, rj *resolvedJob) {
 	} else if rj.outputConnEnv != "" {
 		fmt.Fprintf(out, "      output: env:%s (%s)\n", rj.outputConnEnv, j.Output.Format)
 	}
+	if j.Report != nil {
+		format := valueOr(rj.reportFormat, "default")
+		if rj.reportPath != "" {
+			fmt.Fprintf(out, "      report: %s (%s)\n", rj.reportPath, format)
+		} else {
+			fmt.Fprintf(out, "      report: (log) (%s)\n", format)
+		}
+	}
+	if rj.rulesPath != "" {
+		fmt.Fprintf(out, "      rules: %s\n", rj.rulesPath)
+	} else if j.Command == jobs.CommandInspect {
+		fmt.Fprintf(out, "      rules: (built-in defaults)\n")
+	}
+	if rj.selection != nil {
+		fmt.Fprintf(out, "      select: %s\n", rj.selection.summary())
+	}
 	if rj.logPath != "" {
-		fmt.Fprintf(out, "      logfile: %s\n", rj.logPath)
+		fmt.Fprintf(out, "      logfile: %s (rotate >= %d bytes, keep %d)\n", rj.logPath, rj.logPolicy.MaxSizeBytes, rj.logPolicy.Keep)
 	}
 	fmt.Fprintln(out)
 }
 
 // executeResolvedJob runs a single already-validated job.
 func executeResolvedJob(rj *resolvedJob) (err error) {
-	lg, closeLog, lerr := newJobLogger(rj.logPath, rj.secrets)
+	lg, closeLog, lerr := newJobLogger(rj.logPath, rj.logPolicy, rj.secrets)
 	if lerr != nil {
 		return lerr
 	}
@@ -364,6 +449,14 @@ func executeResolvedJob(rj *resolvedJob) (err error) {
 		err = runScriptsListJob(rj, lg)
 	case jobs.CommandTempl:
 		err = runTemplJob(rj, lg)
+	case jobs.CommandSplit:
+		err = runSplitJob(rj, lg)
+	case jobs.CommandInspect:
+		err = runInspectJob(rj, lg)
+	case jobs.CommandDiff:
+		err = runDiffJob(rj, lg)
+	case jobs.CommandScriptsExec:
+		err = runScriptsExecJob(rj, lg)
 	default:
 		err = fmt.Errorf("unsupported command %q", rj.job.Command)
 	}
@@ -505,6 +598,168 @@ func runScriptsListJob(rj *resolvedJob, lg *jobLogger) error {
 	return nil
 }
 
+func runSplitJob(rj *resolvedJob, lg *jobLogger) error {
+	db, err := readJobInputs(rj, lg)
+	if err != nil {
+		return err
+	}
+	sel := splitSelection{}
+	if rj.selection != nil {
+		sel = *rj.selection
+	}
+	filtered, err := filterDatabaseSelection(db, sel)
+	if err != nil {
+		return fmt.Errorf("split selection: %w", err)
+	}
+	if sel.DatabaseName != "" {
+		filtered.Name = sel.DatabaseName
+	}
+	tables := 0
+	for _, s := range filtered.Schemas {
+		tables += len(s.Tables)
+	}
+	lg.logf("split: selected %d schema(s), %d table(s)", len(filtered.Schemas), tables)
+	return writeJobOutput(rj, filtered, lg)
+}
+
+func runInspectJob(rj *resolvedJob, lg *jobLogger) error {
+	db, err := readJobInputs(rj, lg)
+	if err != nil {
+		return err
+	}
+	config, err := inspector.LoadConfig(rj.rulesPath) // "" -> built-in defaults
+	if err != nil {
+		return fmt.Errorf("load rules: %w", err)
+	}
+	report, err := inspector.NewInspector(db, config).Inspect()
+	if err != nil {
+		return fmt.Errorf("inspection failed: %w", err)
+	}
+
+	var formatted string
+	switch valueOr(rj.reportFormat, "markdown") {
+	case "json":
+		formatted, err = inspector.NewJSONFormatter().Format(report)
+	default:
+		formatted, err = inspector.NewMarkdownFormatter(io.Discard).Format(report)
+	}
+	if err != nil {
+		return fmt.Errorf("format report: %w", err)
+	}
+	if werr := atomicWrite(rj.reportPath, func(tmp string) error {
+		return os.WriteFile(tmp, []byte(formatted), 0o644)
+	}); werr != nil {
+		return werr
+	}
+	lg.logf("inspect: %d error(s), %d warning(s) -> %s",
+		report.Summary.ErrorCount, report.Summary.WarningCount, rj.reportPath)
+	if report.HasErrors() {
+		return fmt.Errorf("inspection found %d error(s)", report.Summary.ErrorCount)
+	}
+	return nil
+}
+
+func runDiffJob(rj *resolvedJob, lg *jobLogger) error {
+	if len(rj.inputs) != 2 {
+		return fmt.Errorf("diff requires exactly 2 inputs, got %d", len(rj.inputs))
+	}
+	source, err := readOneJobInput(rj.inputs[0])
+	if err != nil {
+		return fmt.Errorf("input[0]: %w", err)
+	}
+	lg.logf("diff source: %s", inputLabel(rj.inputs[0]))
+	target, err := readOneJobInput(rj.inputs[1])
+	if err != nil {
+		return fmt.Errorf("input[1]: %w", err)
+	}
+	lg.logf("diff target: %s", inputLabel(rj.inputs[1]))
+
+	result := diff.CompareDatabases(source, target)
+	s := diff.ComputeSummary(result)
+	lg.logf("diff: schemas %d/%d/%d, tables %d/%d/%d, columns %d/%d/%d (missing/extra/modified)",
+		s.Schemas.Missing, s.Schemas.Extra, s.Schemas.Modified,
+		s.Tables.Missing, s.Tables.Extra, s.Tables.Modified,
+		s.Columns.Missing, s.Columns.Extra, s.Columns.Modified)
+
+	format := diff.FormatSummary
+	switch rj.reportFormat {
+	case "json":
+		format = diff.FormatJSON
+	case "html":
+		format = diff.FormatHTML
+	}
+
+	if rj.reportPath == "" {
+		var buf bytes.Buffer
+		if err := diff.FormatDiff(result, format, &buf); err != nil {
+			return fmt.Errorf("format diff: %w", err)
+		}
+		for _, line := range strings.Split(strings.TrimRight(buf.String(), "\n"), "\n") {
+			lg.logf("%s", line)
+		}
+		return nil
+	}
+	if werr := atomicWrite(rj.reportPath, func(tmp string) error {
+		f, err := os.Create(tmp)
+		if err != nil {
+			return err
+		}
+		defer f.Close()
+		return diff.FormatDiff(result, format, f)
+	}); werr != nil {
+		return werr
+	}
+	lg.logf("diff report written: %s", rj.reportPath)
+	return nil
+}
+
+func runScriptsExecJob(rj *resolvedJob, lg *jobLogger) error {
+	schemaName := valueOr(rj.job.Options.Schema, "public")
+	combined := &models.Schema{Name: schemaName}
+	for _, dir := range rj.scriptDirs {
+		reader := sqldir.NewReader(&readers.ReaderOptions{
+			FilePath: dir,
+			Metadata: map[string]any{
+				"schema_name":   schemaName,
+				"database_name": "database",
+			},
+		})
+		db, err := reader.ReadDatabase()
+		if err != nil {
+			return fmt.Errorf("%s: %w", dir, err)
+		}
+		if len(db.Schemas) == 0 {
+			continue
+		}
+		combined.Scripts = append(combined.Scripts, db.Schemas[0].Scripts...)
+	}
+	if len(combined.Scripts) == 0 {
+		lg.logf("no scripts found; nothing to execute")
+		return nil
+	}
+	lg.logf("executing %d script(s) against database env:%s", len(combined.Scripts), rj.outputConnEnv)
+
+	writer := sqlexec.NewWriter(&writers.WriterOptions{
+		Metadata: map[string]any{
+			"connection_string": rj.outputConn,
+			"ignore_errors":     rj.job.Options.ContinueOnError,
+		},
+	})
+	if err := writer.WriteSchema(combined); err != nil {
+		return fmt.Errorf("script execution failed: %w", err)
+	}
+
+	opts := writer.Options()
+	total, _ := opts.Metadata["execution_total"].(int)
+	success, _ := opts.Metadata["execution_success"].(int)
+	failed, _ := opts.Metadata["execution_failed"].(int)
+	lg.logf("executed %d script(s): %d succeeded, %d failed", total, success, failed)
+	if failed > 0 && !rj.job.Options.ContinueOnError {
+		return fmt.Errorf("%d script(s) failed", failed)
+	}
+	return nil
+}
+
 // readJobInputs reads every input and additively merges them into one model.
 func readJobInputs(rj *resolvedJob, lg *jobLogger) (*models.Database, error) {
 	var base *models.Database
@@ -559,7 +814,37 @@ func writeJobOutput(rj *resolvedJob, db *models.Database, lg *jobLogger) error {
 		return fmt.Errorf("failed to create output directory: %w", err)
 	}
 	lg.logf("writing output: %s (%s)", rj.outputPath, format)
-	return writeDatabase(db, format, rj.outputPath, o.Package, o.Schema, o.FlattenSchema, "", "", o.ContinueOnError, "")
+
+	write := func(target string) error {
+		return writeDatabase(db, format, target, o.Package, o.Schema, o.FlattenSchema, "", "", o.ContinueOnError, "")
+	}
+	// Single-file formats are written to a temp file and renamed into place so
+	// a failure never leaves a partial or truncated output. Directory-emitting
+	// formats (gorm/bun/drizzle/typeorm/prisma) write in place.
+	if jobs.SingleFileOutputFormat(format) {
+		return atomicWrite(rj.outputPath, write)
+	}
+	return write(rj.outputPath)
+}
+
+// atomicWrite calls produce with a temp path in the same directory as
+// finalPath, then renames it over finalPath. The temp file is removed on any
+// error so the destination is only ever replaced by a complete file.
+func atomicWrite(finalPath string, produce func(tmpPath string) error) error {
+	dir := filepath.Dir(finalPath)
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		return fmt.Errorf("failed to create output directory: %w", err)
+	}
+	tmp := filepath.Join(dir, fmt.Sprintf(".%s.relspec-tmp-%d", filepath.Base(finalPath), os.Getpid()))
+	if err := produce(tmp); err != nil {
+		_ = os.Remove(tmp)
+		return err
+	}
+	if err := os.Rename(tmp, finalPath); err != nil {
+		_ = os.Remove(tmp)
+		return fmt.Errorf("failed to finalize %s: %w", finalPath, err)
+	}
+	return nil
 }
 
 // --- logging + redaction ---------------------------------------------------
@@ -572,7 +857,7 @@ type jobLogger struct {
 // newJobLogger returns a logger that mirrors to stderr and, when path is set,
 // to a job logfile. Connection strings and known secret values are redacted
 // from everything it writes.
-func newJobLogger(path string, secrets []string) (*jobLogger, func(err error), error) {
+func newJobLogger(path string, policy jobs.LogPolicy, secrets []string) (*jobLogger, func(err error), error) {
 	lg := &jobLogger{secrets: secrets}
 	if path == "" {
 		return lg, func(error) {}, nil
@@ -580,6 +865,7 @@ func newJobLogger(path string, secrets []string) (*jobLogger, func(err error), e
 	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
 		return nil, nil, fmt.Errorf("failed to create log directory: %w", err)
 	}
+	rotateLogIfNeeded(path, policy)
 	f, err := os.OpenFile(path, os.O_CREATE|os.O_WRONLY|os.O_APPEND, 0o644)
 	if err != nil {
 		return nil, nil, fmt.Errorf("failed to open logfile %q: %w", path, err)
@@ -593,6 +879,28 @@ func newJobLogger(path string, secrets []string) (*jobLogger, func(err error), e
 	}, nil
 }
 
+// rotateLogIfNeeded renames path -> path.1 -> path.2 ... up to policy.Keep
+// when path has grown to policy.MaxSizeBytes or more. The oldest file beyond
+// Keep is deleted. A zero/negative MaxSizeBytes disables rotation.
+func rotateLogIfNeeded(path string, policy jobs.LogPolicy) {
+	if policy.MaxSizeBytes <= 0 {
+		return
+	}
+	info, err := os.Stat(path)
+	if err != nil || info.Size() < policy.MaxSizeBytes {
+		return
+	}
+	if policy.Keep < 1 {
+		_ = os.Remove(path)
+		return
+	}
+	_ = os.Remove(fmt.Sprintf("%s.%d", path, policy.Keep))
+	for i := policy.Keep - 1; i >= 1; i-- {
+		_ = os.Rename(fmt.Sprintf("%s.%d", path, i), fmt.Sprintf("%s.%d", path, i+1))
+	}
+	_ = os.Rename(path, path+".1")
+}
+
 func (l *jobLogger) logf(format string, args ...interface{}) {
 	line := l.redact(fmt.Sprintf(format, args...))
 	fmt.Fprintf(os.Stderr, "      %s\n", line)

cmd/relspec/job_test.go

@@ -376,6 +376,286 @@ jobs:
 	}
 }
 
+func TestJobRun_SplitJob(t *testing.T) {
+	dir := jobFixture(t, `version: 1
+jobs:
+  extract:
+    command: split
+    inputs:
+      - path: schema/core.dbml
+        format: dbml
+      - path: schema/tenant.dbml
+        format: dbml
+    select:
+      tables: [users]
+    output:
+      format: json
+      path: build/subset.json
+      overwrite: true
+`)
+	set := mustLoadSet(t, filepath.Join(dir, "relspec.yml"))
+	if err := executeJobPlan(set, "extract", false, false, &bytes.Buffer{}); err != nil {
+		t.Fatalf("execute split job: %v", err)
+	}
+	out, err := os.ReadFile(filepath.Join(dir, "build", "subset.json"))
+	if err != nil {
+		t.Fatalf("read split output: %v", err)
+	}
+	s := string(out)
+	if !strings.Contains(s, "users") {
+		t.Fatalf("split output missing selected table:\n%s", s)
+	}
+	if strings.Contains(s, "posts") {
+		t.Fatalf("split output should have excluded posts:\n%s", s)
+	}
+}
+
+func TestJobRun_InspectJob(t *testing.T) {
+	dir := jobFixture(t, `version: 1
+jobs:
+  lint:
+    command: inspect
+    inputs:
+      - path: schema/core.dbml
+        format: dbml
+    report:
+      format: json
+      path: build/report.json
+      overwrite: true
+    logfile: .relspec/lint.log
+`)
+	set := mustLoadSet(t, filepath.Join(dir, "relspec.yml"))
+	// Default rules only warn, so the job succeeds.
+	if err := executeJobPlan(set, "lint", false, false, &bytes.Buffer{}); err != nil {
+		t.Fatalf("execute inspect job: %v", err)
+	}
+	if _, err := os.ReadFile(filepath.Join(dir, "build", "report.json")); err != nil {
+		t.Fatalf("expected report file: %v", err)
+	}
+	logData, _ := os.ReadFile(filepath.Join(dir, ".relspec", "lint.log"))
+	if !strings.Contains(string(logData), "inspect:") {
+		t.Fatalf("logfile missing inspect summary:\n%s", logData)
+	}
+}
+
+func TestJobRun_InspectJobFailsOnRuleError(t *testing.T) {
+	dir := jobFixture(t, `version: 1
+jobs:
+  lint:
+    command: inspect
+    inputs:
+      - path: schema/core.dbml
+        format: dbml
+    rules: rules.yaml
+    report:
+      format: json
+      path: build/report.json
+      overwrite: true
+    logfile: .relspec/lint.log
+`)
+	// A rule set to "error" level for a violation the fixture triggers.
+	writeFile(t, filepath.Join(dir, "rules.yaml"), `version: "1.0"
+rules:
+  primary_key_naming:
+    enabled: enforce
+    function: primary_key_naming
+    pattern: "^id_"
+    message: "Primary key columns should start with id_"
+`)
+	set := mustLoadSet(t, filepath.Join(dir, "relspec.yml"))
+	err := executeJobPlan(set, "lint", false, false, &bytes.Buffer{})
+	if err == nil || !strings.Contains(err.Error(), "error(s)") {
+		t.Fatalf("expected inspect job to fail on rule error, got %v", err)
+	}
+	logData, _ := os.ReadFile(filepath.Join(dir, ".relspec", "lint.log"))
+	if !strings.Contains(string(logData), "FAILED") {
+		t.Fatalf("failed inspect job should log FAILED:\n%s", logData)
+	}
+}
+
+func TestJobRun_DiffJob(t *testing.T) {
+	dir := jobFixture(t, `version: 1
+jobs:
+  compare:
+    command: diff
+    inputs:
+      - path: schema/core.dbml
+        format: dbml
+      - path: schema/tenant.dbml
+        format: dbml
+    report:
+      format: json
+      path: build/diff.json
+      overwrite: true
+`)
+	set := mustLoadSet(t, filepath.Join(dir, "relspec.yml"))
+	if err := executeJobPlan(set, "compare", false, false, &bytes.Buffer{}); err != nil {
+		t.Fatalf("execute diff job: %v", err)
+	}
+	out, err := os.ReadFile(filepath.Join(dir, "build", "diff.json"))
+	if err != nil {
+		t.Fatalf("read diff report: %v", err)
+	}
+	if len(out) == 0 {
+		t.Fatal("diff report is empty")
+	}
+}
+
+func TestJobRun_FromJobWiring(t *testing.T) {
+	dir := jobFixture(t, `version: 1
+jobs:
+  a:
+    command: convert
+    inputs:
+      - path: schema/core.dbml
+        format: dbml
+    output:
+      format: json
+      path: build/a.json
+      overwrite: true
+  b:
+    command: convert
+    inputs:
+      - from_job: a
+    output:
+      format: yaml
+      path: build/b.yaml
+      overwrite: true
+`)
+	set := mustLoadSet(t, filepath.Join(dir, "relspec.yml"))
+	if err := executeJobPlan(set, "b", false, false, &bytes.Buffer{}); err != nil {
+		t.Fatalf("execute from_job chain: %v", err)
+	}
+	if _, err := os.Stat(filepath.Join(dir, "build", "a.json")); err != nil {
+		t.Fatalf("producer output missing: %v", err)
+	}
+	out, err := os.ReadFile(filepath.Join(dir, "build", "b.yaml"))
+	if err != nil {
+		t.Fatalf("consumer output missing: %v", err)
+	}
+	if !strings.Contains(string(out), "users") {
+		t.Fatalf("consumer did not consume producer output:\n%s", out)
+	}
+}
+
+func TestJobRun_LogRotation(t *testing.T) {
+	dir := jobFixture(t, `version: 1
+jobs:
+  build:
+    command: scripts-list
+    script_dirs: [migrations]
+    log_max_size: "150B"
+    log_keep: 2
+    logfile: .relspec/build.log
+`)
+	writeFile(t, filepath.Join(dir, "migrations", "1_001_a.sql"), "CREATE TABLE a();\n")
+	logPath := filepath.Join(dir, ".relspec", "build.log")
+	writeFile(t, logPath, strings.Repeat("x", 300)+"\n")
+
+	set := mustLoadSet(t, filepath.Join(dir, "relspec.yml"))
+	if err := executeJobPlan(set, "build", false, false, &bytes.Buffer{}); err != nil {
+		t.Fatalf("execute job: %v", err)
+	}
+	rotated, err := os.ReadFile(logPath + ".1")
+	if err != nil {
+		t.Fatalf("expected rotated logfile build.log.1: %v", err)
+	}
+	if !strings.Contains(string(rotated), strings.Repeat("x", 300)) {
+		t.Fatalf("rotated logfile should hold the old content")
+	}
+	fresh, err := os.ReadFile(logPath)
+	if err != nil {
+		t.Fatalf("expected fresh logfile: %v", err)
+	}
+	if strings.Contains(string(fresh), strings.Repeat("x", 300)) {
+		t.Fatalf("fresh logfile should not contain the rotated-out content:\n%s", fresh)
+	}
+	if !strings.Contains(string(fresh), "OK") {
+		t.Fatalf("fresh logfile should hold the new run:\n%s", fresh)
+	}
+}
+
+func TestJobRun_AtomicOutputLeavesOriginalOnFailure(t *testing.T) {
+	dir := jobFixture(t, `version: 1
+jobs:
+  x:
+    command: convert
+    inputs:
+      - path: schema/core.dbml
+        format: dbml
+    output:
+      format: json
+      path: build/out.json
+      overwrite: true
+`)
+	// Seed the destination, then make its parent directory read-only so the
+	// rename step fails. The seeded file must survive intact.
+	seeded := filepath.Join(dir, "build", "out.json")
+	writeFile(t, seeded, `{"seeded":true}`)
+	if err := os.Chmod(filepath.Join(dir, "build"), 0o500); err != nil {
+		t.Skipf("cannot chmod: %v", err)
+	}
+	t.Cleanup(func() { _ = os.Chmod(filepath.Join(dir, "build"), 0o755) })
+
+	set := mustLoadSet(t, filepath.Join(dir, "relspec.yml"))
+	if err := executeJobPlan(set, "x", false, false, &bytes.Buffer{}); err == nil {
+		t.Skip("write unexpectedly succeeded (running as root?)")
+	}
+	if err := os.Chmod(filepath.Join(dir, "build"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	data, err := os.ReadFile(seeded)
+	if err != nil {
+		t.Fatalf("seeded file gone: %v", err)
+	}
+	if !strings.Contains(string(data), "seeded") {
+		t.Fatalf("seeded file was corrupted: %s", data)
+	}
+}
+
+func TestJobRun_ScriptsExecMissingConnEnv(t *testing.T) {
+	dir := jobFixture(t, `version: 1
+jobs:
+  migrate:
+    command: scripts-exec
+    script_dirs: [migrations]
+    output:
+      conn_env: RELSPEC_TEST_EXEC_MISSING
+    logfile: .relspec/migrate.log
+`)
+	writeFile(t, filepath.Join(dir, "migrations", "1_001_a.sql"), "CREATE TABLE a();\n")
+	os.Unsetenv("RELSPEC_TEST_EXEC_MISSING")
+	set := mustLoadSet(t, filepath.Join(dir, "relspec.yml"))
+	err := executeJobPlan(set, "migrate", false, false, &bytes.Buffer{})
+	if err == nil || !strings.Contains(err.Error(), "conn_env") {
+		t.Fatalf("expected missing conn_env error, got %v", err)
+	}
+}
+
+func TestJobRun_ScriptsExecDryRun(t *testing.T) {
+	dir := jobFixture(t, `version: 1
+jobs:
+  migrate:
+    command: scripts-exec
+    script_dirs: [migrations]
+    output:
+      conn_env: RELSPEC_TEST_EXEC_CONN
+`)
+	writeFile(t, filepath.Join(dir, "migrations", "1_001_a.sql"), "CREATE TABLE a();\n")
+	t.Setenv("RELSPEC_TEST_EXEC_CONN", "postgres://u:secretpw@h/db")
+	set := mustLoadSet(t, filepath.Join(dir, "relspec.yml"))
+	var buf bytes.Buffer
+	if err := executeJobPlan(set, "migrate", true, false, &buf); err != nil {
+		t.Fatalf("dry run: %v", err)
+	}
+	if strings.Contains(buf.String(), "secretpw") {
+		t.Fatalf("plan leaked secret:\n%s", buf.String())
+	}
+	if !strings.Contains(buf.String(), "env:RELSPEC_TEST_EXEC_CONN") {
+		t.Fatalf("plan should name the env var:\n%s", buf.String())
+	}
+}
+
 func TestJobRun_TemplDatabaseMode(t *testing.T) {
 	dir := jobFixture(t, `version: 1
 jobs:

cmd/relspec/prisma_options.go

@@ -10,6 +10,7 @@ func newReaderOptions(filePath, connString string) *readers.ReaderOptions {
 		FilePath:         filePath,
 		ConnectionString: connString,
 		Prisma7:          prisma7,
+		StrictDirectives: strictDirectives,
 	}
 }
 
@@ -22,5 +23,6 @@ func newWriterOptions(outputPath, packageName string, flattenSchema bool, nullab
 		NullableArrays:   nullableArrays,
 		Prisma7:          prisma7,
 		ContinueOnError:  continueOnError,
+		StrictDirectives: strictDirectives,
 	}
 }

cmd/relspec/root.go

@@ -14,6 +14,7 @@ var (
 	buildDate        = "unknown"
 	prisma7          bool
 	noVersion        bool
+	strictDirectives bool
 )
 
 func init() {
@@ -72,6 +73,7 @@ func init() {
 	rootCmd.AddCommand(reportCmd)
 	rootCmd.PersistentFlags().BoolVar(&prisma7, "prisma7", false, "Use Prisma 7 generator conventions when reading/writing Prisma schemas")
 	rootCmd.PersistentFlags().BoolVar(&noVersion, "no-version", false, "Suppress the RelSpec version header")
+	rootCmd.PersistentFlags().BoolVar(&strictDirectives, "strict-directives", false, "Fail on unknown or untranslatable DBML dialect directives (@postgres:, @sqlite:, …)")
 }
 
 // printVersionHeader prints the "RelSpec <version> (built: <date>)" banner

cmd/relspec/split.go

@@ -205,8 +205,52 @@ func runSplit(cmd *cobra.Command, args []string) error {
 	return nil
 }
 
-// filterDatabase filters the database based on provided criteria
+// splitSelection is the schema/table selection for a split, independent of the
+// CLI flag globals so the job runner can build one directly.
+type splitSelection struct {
+	Schemas        []string
+	Tables         []string
+	ExcludeSchemas []string
+	ExcludeTables  []string
+	DatabaseName   string
+}
+
+// summary renders a one-line human description of the selection.
+func (s splitSelection) summary() string {
+	var parts []string
+	if len(s.Schemas) > 0 {
+		parts = append(parts, "schemas="+strings.Join(s.Schemas, ","))
+	}
+	if len(s.Tables) > 0 {
+		parts = append(parts, "tables="+strings.Join(s.Tables, ","))
+	}
+	if len(s.ExcludeSchemas) > 0 {
+		parts = append(parts, "exclude_schemas="+strings.Join(s.ExcludeSchemas, ","))
+	}
+	if len(s.ExcludeTables) > 0 {
+		parts = append(parts, "exclude_tables="+strings.Join(s.ExcludeTables, ","))
+	}
+	if s.DatabaseName != "" {
+		parts = append(parts, "database_name="+s.DatabaseName)
+	}
+	if len(parts) == 0 {
+		return "(all schemas/tables)"
+	}
+	return strings.Join(parts, " ")
+}
+
+// filterDatabase filters the database based on the CLI split flags.
 func filterDatabase(db *models.Database) (*models.Database, error) {
+	return filterDatabaseSelection(db, splitSelection{
+		Schemas:        parseCommaSeparated(splitSchemas),
+		Tables:         parseCommaSeparated(splitTables),
+		ExcludeSchemas: parseCommaSeparated(splitExcludeSchema),
+		ExcludeTables:  parseCommaSeparated(splitExcludeTables),
+	})
+}
+
+// filterDatabaseSelection filters db down to the schemas/tables named by sel.
+func filterDatabaseSelection(db *models.Database, sel splitSelection) (*models.Database, error) {
 	filteredDB := &models.Database{
 		Name:            db.Name,
 		Description:     db.Description,
@@ -220,11 +264,11 @@ func filterDatabase(db *models.Database) (*models.Database, error) {
 		Domains:         db.Domains, // Keep domains for now
 	}
 
-	// Parse filter flags
+	// Selection criteria
-	includeSchemas := parseCommaSeparated(splitSchemas)
+	includeSchemas := sel.Schemas
-	includeTables := parseCommaSeparated(splitTables)
+	includeTables := sel.Tables
-	excludeSchemas := parseCommaSeparated(splitExcludeSchema)
+	excludeSchemas := sel.ExcludeSchemas
-	excludeTables := parseCommaSeparated(splitExcludeTables)
+	excludeTables := sel.ExcludeTables
 
 	// Convert table names to lowercase for case-insensitive matching
 	includeTablesLower := make(map[string]bool)

docs/DBML_DIRECTIVES.md

@@ -0,0 +1,115 @@
+# DBML Dialect Directives
+
+DBML has no dialect-neutral way to express database-specific features such as
+PostgreSQL table partitioning or SQLite `WITHOUT ROWID`. RelSpec adds **dialect
+directives** — explicit, parseable lines embedded in a `.dbml` file that are:
+
+- stored losslessly in the intermediate model (under each object's `Metadata`),
+- preserved unchanged through a `DBML → model → DBML` round-trip,
+- translated to SQL **only** by the writer for the matching dialect
+  (`@postgres:` clauses appear in PostgreSQL output, never in SQLite output, and
+  vice-versa).
+
+## Grammar
+
+A directive is a single line, matched on its trimmed content:
+
+```
+@<namespace>[(<target>)]: <args>
+```
+
+| Part | Rules |
+|------|-------|
+| `namespace` | `^[a-z][a-z0-9_]*$` — e.g. `postgres`, `sqlite`. Future dialects allowed. |
+| `(target)` | Optional. A **column name** only, valid only on a directive line inside a table body. Bare or single/double quoted. |
+| `args` | Everything after the first `:`, trimmed. Otherwise preserved **verbatim**. Must be non-empty. |
+
+The **key** of a directive is derived: the lowercased first whitespace-delimited
+token of `args` (`partition by RANGE (created_at)` → `partition`). It drives
+duplicate detection and writer dispatch.
+
+## Location
+
+Where the line appears determines which object it attaches to:
+
+| Position in the file | Attaches to |
+|----------------------|-------------|
+| Before the first `Table {` | database (`db.Metadata`) |
+| Table body, no `(target)` | that table |
+| Table body, `(col)` target | column `col` of that table (error if `col` is unknown) |
+| Inside an `indexes { }` block | the **most recently listed** index entry in that block; `(target)` is not allowed |
+
+```dbml
+@postgres: search_path myapp          -- database
+
+Table myapp.events {
+  id bigint [pk]
+  created_at timestamp [not null]
+  @postgres(id): identity always      -- column "id"
+  @postgres: partition by RANGE (created_at)   -- table
+  @postgres: tablespace fast_data              -- table
+  @sqlite: without rowid                        -- table
+
+  indexes {
+    (created_at) [name: 'idx_events_created']
+    @postgres: with (fillfactor=90)   -- index "idx_events_created"
+    @postgres: tablespace idx_space    -- index "idx_events_created"
+  }
+}
+```
+
+## Duplicate policy
+
+- **Repeatable by default** — every directive with the same `(namespace, key)` at
+  one location is kept, in source order.
+- **Singletons** raise a line-numbered error on a second occurrence at the same
+  location. Current singletons: `postgres` `partition`, `tablespace`, `inherits`,
+  `storage`, `compression`, `identity`; `sqlite` `without`, `strict`, `collate`.
+
+## Strict mode
+
+CLI flag `--strict-directives` (also `ReaderOptions.StrictDirectives` /
+`WriterOptions.StrictDirectives`):
+
+- **Reader**: an unknown namespace or key is a hard error. Without strict mode it
+  is stored and preserved silently, and round-trips unchanged.
+- **PostgreSQL / SQLite writer**: a directive for **that** writer's own dialect
+  whose key it cannot translate is a hard error. Without strict mode, translatable
+  keys are emitted and the rest are skipped. Directives for other dialects are
+  always ignored, never emitted.
+
+## Errors
+
+All are line-numbered (`dbml: line N: …`):
+
+- no colon, or empty `args`
+- namespace empty or not matching `[a-z][a-z0-9_]*`
+- `(target)` naming an unknown column, or used at the top level / in an `indexes` block
+- a directive in the catalog used at a location it is not valid for
+- duplicate singleton at the same location
+- (strict mode) unknown `(namespace, key)`
+
+## Supported directive matrix
+
+### `@postgres`
+
+| Key | Locations | SQL emitted | Notes |
+|-----|-----------|-------------|-------|
+| `partition` | table | `PARTITION BY <args>` appended to `CREATE TABLE` | e.g. `@postgres: partition by RANGE (created_at)` |
+| `inherits` | table | `INHERITS (<args>)` — args verbatim | |
+| `with` | table, index | `WITH (<params>)` | On an index, wins over `WITH` derived from the index comment. `@postgres: with (fillfactor=90)` |
+| `tablespace` | table, index | `TABLESPACE <name>` | Emitted after `WITH`, before `WHERE` on indexes |
+| `storage` | column | `STORAGE <mode>` in the column definition | e.g. `@postgres(blob): storage external` |
+| `compression` | column | `COMPRESSION <method>` | |
+| `identity` | column | `identity always` → `GENERATED ALWAYS AS IDENTITY`; `identity default` / `identity by default` → `GENERATED BY DEFAULT AS IDENTITY` | |
+
+### `@sqlite`
+
+| Key | Locations | SQL emitted | Notes |
+|-----|-----------|-------------|-------|
+| `without` | table | `WITHOUT ROWID` table option | `@sqlite: without rowid` |
+| `strict` | table | `STRICT` table option | `WITHOUT ROWID` is emitted before `STRICT` |
+| `collate` | column | ` COLLATE <name>` in the column definition | e.g. `@sqlite(name): collate NOCASE` |
+
+Unknown namespaces and keys not in these tables are still preserved losslessly
+(and round-trip through the DBML writer) whenever strict mode is off.

docs/JOB_FILES.md

@@ -9,10 +9,11 @@ relspec job run build-schema --plan  # validate + print plan, execute nothing
 relspec job run build-schema         # run the job (and its dependencies)
 ```
 
-## Design contract (first release)
+## Design contract
 
-This is the smallest coherent contract that is safe and useful end to end.
+This is a deliberately small, safe contract. Every capability is offline-testable
-Anything not listed under "Supported" is intentionally deferred.
+except live database execution (`scripts-exec`), which is validated and planned
+offline and only connects at run time.
 
 ### Not a shell
 
@@ -24,14 +25,15 @@ means adding a vetted adapter in the RelSpec source.
 |----------------|--------------------------------------------------------------------|
 | `convert`      | read one or more input schemas, additively merge them, write one output |
 | `merge`        | like `convert` but requires ≥2 inputs and exposes `skip_*` merge options |
+| `split`        | read one or more schemas, keep the selected schemas/tables, write one output |
 | `scripts-list` | deterministically list SQL scripts across one or more directories   |
+| `scripts-exec` | execute SQL scripts across one or more directories against a live PostgreSQL database |
 | `templ`        | apply a custom Go text template to one or more input schemas          |
+| `inspect`      | validate one or more schemas against rules and write a report       |
+| `diff`         | compare exactly two schemas and write a differences report          |
 
-Deferred (documented, not implemented here): `scripts` execution against a live
+`convert`, `merge` and `split` are **producers**: their file output can be fed
-database, `split`, `inspect`, `diff`, job-to-job output wiring,
+directly into another job with `from_job` (see below).
-log rotation/retention. Live SQL execution already exists as
-`relspec scripts execute`; wiring it into the job runner is a follow-up because
-it needs live database credentials and cannot be covered by offline tests.
 
 ### Discovery and precedence
 
@@ -49,12 +51,16 @@ already forbid duplicate keys within a single file.
 
 ### Paths
 
-* Every path (`inputs[].path`, `output.path`, `script_dirs[]`, `logfile`) is
+* Every path (`inputs[].path`, `output.path`, `report.path`, `rules`,
-  **relative to the directory containing the job file that declared the job**,
+  `script_dirs[]`, `template`, `logfile`) is **relative to the directory
-  not the process working directory.
+  containing the job file that declared the job**, not the process working
+  directory.
 * Absolute paths, `~`-relative paths and any path that resolves outside the job
   file directory (`../`, `a/../../b`, …) are **rejected during validation** —
   before anything runs.
+* At run time each path is additionally resolved through its symlinks: a symlink
+  inside the job-file directory that points outside it is rejected before the
+  path is opened.
 
 ### Credentials
 
@@ -76,58 +82,104 @@ already forbid duplicate keys within a single file.
 first. Nothing is read, written, connected to, or executed if validation fails.
 Checks include:
 
-* schema `version` (must be `1`), unknown YAML fields rejected
+* schema `version` — **forward-permissive**: any version `>= 1` is accepted.
+  An omitted `version` is treated as the current one. A version newer than this
+  build understands loads best-effort (unknown YAML fields are ignored and a
+  warning is printed); at the current version unknown YAML fields are still
+  rejected.
 * duplicate job names across files
 * unknown / missing `command`
-* per-command input/output shape (`convert`/`merge` need inputs + output;
+* per-command input/output shape:
-  `scripts-list` needs `script_dirs` and forbids inputs/output)
+  * `convert` needs ≥1 input + output; `merge` needs ≥2 inputs + output
+  * `split` needs ≥1 input + a file output, plus an optional `select:` block
+  * `scripts-list` needs `script_dirs` and forbids inputs/output
+  * `scripts-exec` needs `script_dirs` and `output.conn_env` (pgsql only)
+  * `inspect` needs ≥1 input + `report:` (format `markdown`|`json`)
+  * `diff` needs **exactly 2** inputs + `report:` (format `summary`|`json`|`html`)
 * unknown input/output `format`
+* `from_job` targets exist, are producers (`convert`/`merge`/`split`) and write a
+  single-file output
 * path traversal / absolute / home-relative paths
-* `depends_on` targets exist
+* `depends_on` and `from_job` targets exist
-* dependency cycles (reported as `a -> b -> c -> a`)
+* dependency cycles over the combined `depends_on` + `from_job` graph
+  (reported as `a -> b -> c -> a`)
 
 Then, immediately before running, per-job pre-flight resolves paths and checks:
 
-* every input file exists and is a file
+* every input file exists and is a file (a `from_job` input is exempt — its
+  producer runs earlier in the same plan)
 * every `script_dir` exists and is a directory
 * every `conn_env` variable is set
-* `output.path` does not already exist unless `output.overwrite: true`
+* `output.path` / `report.path` does not already exist unless the matching
+  `overwrite: true` is set
+* `rules` (inspect), when given, exists and is a file
+* symlinks in every resolved path stay inside the job-file directory
 
 If any pre-flight check fails for **any** job in the plan, **no** job runs.
 
 ### Execution and exit codes
 
-* `relspec job run <name>` runs the job's `depends_on` closure first, in
+* `relspec job run <name>` runs the job's dependency closure first
-  topological order (deterministic), then the job. `--no-deps` runs only the
+  (`depends_on` plus any `from_job` producers), in topological order
-  named job.
+  (deterministic), then the job. `--no-deps` runs only the named job and is
+  incompatible with `from_job` inputs.
 * `--dry-run` (alias `--plan`) prints the resolved plan and exits 0 without
   touching inputs, outputs or databases.
 * A failing job returns the underlying non-zero status (the process exits 1)
   and the error names the job. The logfile records `FAILED: <error>`; a
   successful job records `OK`. No separate success-marker file is written, so a
   failure can never leave a stale "success".
+* `inspect` fails the job when the report contains rule **errors** (enforced
+  rules); warnings do not fail it. `diff` never fails on differences.
+* Single-file outputs and reports are written to a temporary file in the target
+  directory and atomically renamed into place, so an interrupted run never
+  leaves a partial file. Directory-emitting formats (`gorm`, `bun`, `drizzle`,
+  `typeorm`, `prisma`) are written in place.
+
+### Logfile rotation
+
+When a job has a `logfile`, it is size-rotated before each run. Defaults are
+**5 MB** with **3** rotated files kept (`build.log` → `build.log.1` → …). Override
+per job with `log_max_size` / `log_keep`, or for a whole file with a top-level
+`defaults:` block. `log_max_size` accepts `B`/`KB`/`MB`/`GB` suffixes (e.g.
+`"512KB"`, `"5MB"`).
 
 ## Schema reference
 
 ```yaml
-version: 1                       # required, must be 1
+version: 1                       # optional; any value >= 1 is accepted
+defaults:                        # optional, file-wide
+  log_max_size: 5MB              # B / KB / MB / GB
+  log_keep: 3
 jobs:
   <job-name>:
-    command: convert | merge | scripts-list   # required
+    command: convert | merge | split | scripts-list | scripts-exec | templ | inspect | diff
     description: "free text"                   # optional, shown by `job list`
     depends_on: [other-job, ...]               # optional
-    inputs:                                    # convert (≥1) / merge (≥2)
+    inputs:                                    # convert (≥1) / merge (≥2) / split (≥1) / inspect (≥1) / diff (exactly 2)
       - path: relative/file.dbml               # file inputs
         format: dbml
       - format: pgsql                          # live-connection inputs
         conn_env: SOURCE_DB_URL                # env var NAME
-    script_dirs:                               # scripts-list (≥1)
+      - from_job: build-schema                 # consume another job's file output
+    script_dirs:                               # scripts-list / scripts-exec (≥1)
       - migrations/core
       - migrations/tenant
     template: templates/schema.tmpl             # templ (required)
     mode: table                                 # templ: database/schema/script/table
     filename_pattern: "{{.Name}}.go"           # templ multi-output modes
-    output:                                    # convert / merge (required)
+    select:                                    # split (optional; default = keep everything)
+      schemas: [public]
+      tables: [users, orders]
+      exclude_schemas: []
+      exclude_tables: []
+      database_name: SubsetDB                  # optional rename of the output database
+    rules: .relspec-rules.yaml                 # inspect (optional; built-in defaults if omitted)
+    report:                                    # inspect (required) / diff (required)
+      format: json                             # inspect: markdown|json ; diff: summary|json|html
+      path: build/report.json                  # required, except a diff "summary" (goes to the log)
+      overwrite: false
+    output:                                    # convert / merge / split (required); scripts-exec (required, conn_env)
       format: pgsql
       path: build/schema.sql                   # file output, OR:
       conn_env: TARGET_DB_URL                  # execute against DB (pgsql only)
@@ -136,13 +188,15 @@ jobs:
       flatten_schema: false
       schema: public
       package: models                          # for gorm/bun output
-      continue_on_error: false                 # pgsql output
+      continue_on_error: false                 # pgsql / scripts-exec output
       skip_relations: false                    # merge only
       skip_enums: false
       skip_views: false
       skip_domains: false
       skip_sequences: false
-    logfile: .relspec/log/<job-name>.log       # optional; appended to
+    logfile: .relspec/log/<job-name>.log       # optional; appended to, size-rotated
+    log_max_size: 5MB                           # optional per-job override
+    log_keep: 3                                 # optional per-job override
 ```
 
 For `templ`, `inputs` use the same file or `pgsql`/`conn_env` source forms as
@@ -150,6 +204,11 @@ schema conversion. `output` is optional (empty means stdout); when present it
 contains only `path` and `overwrite`, because templates do not select a schema
 writer format.
 
+A `from_job` input takes no `path`, `format` or `conn_env`: it resolves to the
+named job's `output.path` and inherits its format, and implies a dependency on
+that job. The producer must be a `convert`, `merge` or `split` job writing a
+single-file output.
+
 ### Supported input formats
 
 `dbml`, `dctx`, `drawdb`, `graphql`, `json`, `yaml`, `gorm`, `bun`, `drizzle`,
@@ -230,3 +289,77 @@ jobs:
       path: snapshots/prod.dbml
       overwrite: true
 ```
+
+### Chain jobs with `from_job`, then lint the result
+
+```yaml
+version: 1
+jobs:
+  build-json:
+    command: convert
+    inputs:
+      - { path: schema/core.dbml, format: dbml }
+      - { path: schema/tenant.dbml, format: dbml }
+    output: { format: json, path: build/schema.json, overwrite: true }
+  lint-schema:
+    command: inspect
+    inputs:
+      - from_job: build-json      # implies depends_on: [build-json]
+    rules: .relspec-rules.yaml    # optional; built-in rules if omitted
+    report:
+      format: markdown
+      path: build/lint-report.md
+      overwrite: true
+```
+
+`relspec job run lint-schema` runs `build-json` first, then inspects its output.
+The job fails (exit 1) if any enforced rule is violated.
+
+### Split a subset out of a larger schema
+
+```yaml
+version: 1
+jobs:
+  posts-only:
+    command: split
+    inputs:
+      - { path: schema/core.dbml, format: dbml }
+      - { path: schema/tenant.dbml, format: dbml }
+    select:
+      tables: [posts]
+    output: { format: dbml, path: build/posts.dbml, overwrite: true }
+```
+
+### Diff two schemas
+
+```yaml
+version: 1
+jobs:
+  drift:
+    command: diff
+    inputs:                       # exactly two
+      - { path: build/schema.json, format: json }
+      - format: pgsql
+        conn_env: PROD_DB_URL
+    report:
+      format: summary             # summary → logfile; json/html need a path
+```
+
+`diff` reports differences and always exits 0.
+
+### Execute migration scripts against a live database
+
+```yaml
+version: 1
+jobs:
+  apply-migrations:
+    command: scripts-exec
+    script_dirs:
+      - migrations/core
+      - migrations/tenant
+    output:
+      conn_env: TARGET_DB_URL     # pgsql only; no path
+    options:
+      continue_on_error: false
+    logfile: .relspec/log/apply-migrations.log
+```

examples/jobs/relspec.yml

@@ -4,7 +4,14 @@
 #   relspec job list
 #   relspec job run build-schema --plan
 #   relspec job run build-schema
+#   relspec job run lint-schema        # inspect, consuming build-json's output
 version: 1
+
+# File-wide defaults. Individual jobs may override log_max_size / log_keep.
+defaults:
+  log_max_size: 2MB
+  log_keep: 5
+
 jobs:
   build-schema:
     command: convert
@@ -43,3 +50,30 @@ jobs:
       - migrations/core
       - migrations/tenant
     logfile: .relspec/log/migration-order.log
+
+  lint-schema:
+    command: inspect
+    description: Validate build-json's output against the built-in rules
+    # No depends_on needed: the from_job input implies a dependency on build-json.
+    inputs:
+      - from_job: build-json
+    report:
+      format: markdown
+      path: build/lint-report.md
+      overwrite: true
+    logfile: .relspec/log/lint-schema.log
+
+  posts-only:
+    command: split
+    description: Extract just the posts table into its own DBML file
+    inputs:
+      - path: schema/core.dbml
+        format: dbml
+      - path: schema/tenant.dbml
+        format: dbml
+    select:
+      tables: [posts]
+    output:
+      format: dbml
+      path: build/posts.dbml
+      overwrite: true

pkg/jobs/jobs.go

@@ -20,24 +20,52 @@ import (
 	"os"
 	"path/filepath"
 	"sort"
+	"strconv"
 	"strings"
 
 	"gopkg.in/yaml.v3"
 )
 
-// SchemaVersion is the only job-file schema version this build understands.
+// CurrentSchemaVersion is the highest job-file schema version this build was
-const SchemaVersion = 1
+// written for. MinSchemaVersion is the oldest it still accepts. A file that
+// declares a version in between loads normally; a newer version loads
+// best-effort with a warning (see Load); an older-than-minimum version is a
+// hard error.
+const (
+	CurrentSchemaVersion = 1
+	MinSchemaVersion     = 1
+)
+
+// Built-in logfile rotation policy, used when neither the job nor its file's
+// defaults block sets one.
+const (
+	defaultLogMaxSizeBytes int64 = 5 << 20 // 5 MiB
+	defaultLogKeep               = 3
+)
 
 // Command names are a closed allow-list. Arbitrary strings are rejected.
 const (
 	CommandConvert     = "convert"      // read one or more schema files, optionally merge, write one output
 	CommandMerge       = "merge"        // additive merge of two or more schema files into one output
 	CommandScriptsList = "scripts-list" // deterministically list SQL scripts across one or more directories
+	CommandScriptsExec = "scripts-exec" // execute SQL scripts across one or more directories against a live database
 	CommandTempl       = "templ"        // apply a custom Go text template to one or more schemas
+	CommandSplit       = "split"        // extract selected schemas/tables into a separate output
+	CommandInspect     = "inspect"      // validate one or more schemas against rules and write a report
+	CommandDiff        = "diff"         // compare exactly two schemas and write a differences report
 )
 
 // SupportedCommands lists every accepted command, in help order.
-var SupportedCommands = []string{CommandConvert, CommandMerge, CommandScriptsList, CommandTempl}
+var SupportedCommands = []string{
+	CommandConvert, CommandMerge, CommandScriptsList, CommandScriptsExec,
+	CommandTempl, CommandSplit, CommandInspect, CommandDiff,
+}
+
+// producerCommands are commands whose output is a schema file that another job
+// may consume via from_job.
+var producerCommands = map[string]bool{
+	CommandConvert: true, CommandMerge: true, CommandSplit: true,
+}
 
 // readerFormats are the file-based input formats a job may declare (path).
 var readerFormats = map[string]bool{
@@ -61,17 +89,51 @@ var writerFormats = map[string]bool{
 // live database) is supported instead of writing a file.
 var execOutputFormats = map[string]bool{"pgsql": true}
 
+// singleFileFormats are output formats that emit exactly one file (as opposed
+// to a directory of files). Only these are eligible for atomic temp+rename
+// writes and for being consumed by another job via from_job.
+var singleFileFormats = map[string]bool{
+	"json": true, "yaml": true, "dbml": true, "dctx": true, "drawdb": true,
+	"graphql": true, "pgsql": true, "mssql": true, "sqlite": true,
+}
+
+// SingleFileOutputFormat reports whether format writes exactly one file.
+func SingleFileOutputFormat(format string) bool {
+	return singleFileFormats[strings.ToLower(format)]
+}
+
+// diffReportFormats and inspectReportFormats are the report.format values
+// accepted by the diff and inspect commands respectively.
+var (
+	diffReportFormats    = map[string]bool{"summary": true, "json": true, "html": true}
+	inspectReportFormats = map[string]bool{"markdown": true, "json": true}
+)
+
 // File is the on-disk shape of a single job file.
 type File struct {
 	Version  int             `yaml:"version"`
+	Defaults *Defaults       `yaml:"defaults"`
 	Jobs     map[string]*Job `yaml:"jobs"`
 }
 
+// Defaults carries file-wide settings that individual jobs may override.
+type Defaults struct {
+	// LogMaxSize is a human-readable size ("5MB", "512KB", "1GB"). Empty
+	// means "use the built-in default".
+	LogMaxSize string `yaml:"log_max_size"`
+	// LogKeep is how many rotated logfiles to retain. Zero means "use the
+	// built-in default".
+	LogKeep int `yaml:"log_keep"`
+}
+
 // Job is one named job within a job file.
 type Job struct {
 	// Name and SourceFile are populated by Load, not parsed from YAML.
 	Name       string `yaml:"-"`
 	SourceFile string `yaml:"-"`
+	// fileDefaults is the Defaults block of the file that declared this job,
+	// captured by Load. nil when the file had none.
+	fileDefaults *Defaults `yaml:"-"`
 
 	Command         string   `yaml:"command"`
 	Description     string   `yaml:"description"`
@@ -82,8 +144,13 @@ type Job struct {
 	Mode            string   `yaml:"mode"`
 	FilenamePattern string   `yaml:"filename_pattern"`
 	Output          *Output  `yaml:"output"`
+	Rules           string   `yaml:"rules"`
+	Report          *Report  `yaml:"report"`
+	Select          *Select  `yaml:"select"`
 	Options         Options  `yaml:"options"`
 	Logfile         string   `yaml:"logfile"`
+	LogMaxSize      string   `yaml:"log_max_size"`
+	LogKeep         *int     `yaml:"log_keep"`
 }
 
 // Input is one declared input schema.
@@ -94,6 +161,10 @@ type Input struct {
 	// ConnEnv is the NAME of an environment variable holding a connection
 	// string, used with database formats. The value is never stored here.
 	ConnEnv string `yaml:"conn_env"`
+	// FromJob names another job in the set whose file output is used as this
+	// input. It implies a dependency on that job. Path/Format/ConnEnv must be
+	// empty when FromJob is set; the format is inherited from the producer.
+	FromJob string `yaml:"from_job"`
 }
 
 // Output is the declared output target.
@@ -104,6 +175,105 @@ type Output struct {
 	Overwrite bool   `yaml:"overwrite"`
 }
 
+// Report is the output target for the inspect and diff commands.
+type Report struct {
+	// Format is the report format: diff accepts summary|json|html, inspect
+	// accepts markdown|json. Empty means the command's default.
+	Format    string `yaml:"format"`
+	Path      string `yaml:"path"`
+	Overwrite bool   `yaml:"overwrite"`
+}
+
+// Select carries the schema/table selection for the split command.
+type Select struct {
+	Schemas        []string `yaml:"schemas"`
+	Tables         []string `yaml:"tables"`
+	ExcludeSchemas []string `yaml:"exclude_schemas"`
+	ExcludeTables  []string `yaml:"exclude_tables"`
+	DatabaseName   string   `yaml:"database_name"`
+}
+
+// LogPolicy is the resolved logfile rotation policy for a job.
+type LogPolicy struct {
+	MaxSizeBytes int64
+	Keep         int
+}
+
+// ResolvedLogPolicy returns the effective rotation policy: the job's own
+// overrides win, then its file's defaults block, then the built-in default.
+func (j *Job) ResolvedLogPolicy() LogPolicy {
+	p := LogPolicy{MaxSizeBytes: defaultLogMaxSizeBytes, Keep: defaultLogKeep}
+
+	if j.fileDefaults != nil {
+		if n, err := parseHumanSize(j.fileDefaults.LogMaxSize); err == nil && n > 0 {
+			p.MaxSizeBytes = n
+		}
+		if j.fileDefaults.LogKeep > 0 {
+			p.Keep = j.fileDefaults.LogKeep
+		}
+	}
+	if n, err := parseHumanSize(j.LogMaxSize); err == nil && n > 0 {
+		p.MaxSizeBytes = n
+	}
+	if j.LogKeep != nil && *j.LogKeep >= 0 {
+		p.Keep = *j.LogKeep
+	}
+	return p
+}
+
+// effectiveDeps returns the union of explicit depends_on entries and the jobs
+// referenced by from_job inputs, deduplicated in stable order.
+func (j *Job) effectiveDeps() []string {
+	seen := map[string]bool{}
+	var deps []string
+	add := func(name string) {
+		if name == "" || name == j.Name || seen[name] {
+			return
+		}
+		seen[name] = true
+		deps = append(deps, name)
+	}
+	for _, d := range j.DependsOn {
+		add(d)
+	}
+	for _, in := range j.Inputs {
+		add(in.FromJob)
+	}
+	return deps
+}
+
+// parseHumanSize parses a byte size such as "5MB", "512 KB", "1gb" or a bare
+// byte count. An empty string returns (0, nil) so callers can fall back.
+func parseHumanSize(s string) (int64, error) {
+	s = strings.TrimSpace(s)
+	if s == "" {
+		return 0, nil
+	}
+	upper := strings.ToUpper(s)
+	mult := int64(1)
+	// Check multi-character suffixes before the bare "B".
+	for _, u := range []struct {
+		suffix string
+		m      int64
+	}{
+		{"KB", 1 << 10}, {"MB", 1 << 20}, {"GB", 1 << 30}, {"B", 1},
+	} {
+		if strings.HasSuffix(upper, u.suffix) {
+			mult = u.m
+			upper = strings.TrimSpace(strings.TrimSuffix(upper, u.suffix))
+			break
+		}
+	}
+	n, err := strconv.ParseFloat(upper, 64)
+	if err != nil {
+		return 0, fmt.Errorf("invalid size %q", s)
+	}
+	if n < 0 {
+		return 0, fmt.Errorf("negative size %q", s)
+	}
+	return int64(n * float64(mult)), nil
+}
+
 // Options carries the subset of command flags a job file may set.
 type Options struct {
 	FlattenSchema   bool   `yaml:"flatten_schema"`
@@ -127,6 +297,9 @@ type Set struct {
 	Files []string
 	// Jobs is keyed by job name.
 	Jobs map[string]*Job
+	// Warnings holds non-fatal load-time messages (e.g. a newer-than-known
+	// schema version). Callers should surface these to the user.
+	Warnings []string
 }
 
 // Names returns all job names in deterministic (sorted) order.
@@ -199,15 +372,35 @@ func Load(paths []string) (*Set, error) {
 		if err != nil {
 			return nil, fmt.Errorf("failed to read job file %q: %w", path, err)
 		}
+
+		// Peek at the version first so a newer file can be parsed leniently
+		// (unknown fields ignored) instead of failing outright.
+		var probe struct {
+			Version int `yaml:"version"`
+		}
+		if err := yaml.Unmarshal(data, &probe); err != nil {
+			return nil, fmt.Errorf("invalid job file %q: %w", path, err)
+		}
+		version := probe.Version
+		if version == 0 {
+			version = CurrentSchemaVersion
+		}
+		if version < MinSchemaVersion {
+			return nil, fmt.Errorf("job file %q: unsupported version %d (this build accepts %d or newer)", path, version, MinSchemaVersion)
+		}
+		strict := version <= CurrentSchemaVersion
+		if !strict {
+			set.Warnings = append(set.Warnings, fmt.Sprintf(
+				"job file %q declares version %d, newer than this build understands (%d); loading best-effort and ignoring unknown fields",
+				path, version, CurrentSchemaVersion))
+		}
+
 		dec := yaml.NewDecoder(strings.NewReader(string(data)))
-		dec.KnownFields(true)
+		dec.KnownFields(strict)
 		var f File
 		if err := dec.Decode(&f); err != nil {
 			return nil, fmt.Errorf("invalid job file %q: %w", path, err)
 		}
-		if f.Version != SchemaVersion {
-			return nil, fmt.Errorf("job file %q: unsupported version %d (expected %d)", path, f.Version, SchemaVersion)
-		}
 		if len(f.Jobs) == 0 {
 			return nil, fmt.Errorf("job file %q: no jobs defined", path)
 		}
@@ -220,6 +413,7 @@ func Load(paths []string) (*Set, error) {
 			}
 			job.Name = name
 			job.SourceFile = path
+			job.fileDefaults = f.Defaults
 			origin[name] = path
 			set.Jobs[name] = job
 		}
@@ -240,13 +434,30 @@ func (s *Set) Validate() error {
 			errs = append(errs, fmt.Sprintf("job %q: %s", name, msg))
 		}
 	}
-	// Dependency references + cycles.
+	// Dependency references + cycles + from_job wiring.
 	for _, name := range s.Names() {
-		for _, dep := range s.Jobs[name].DependsOn {
+		j := s.Jobs[name]
+		for _, dep := range j.DependsOn {
 			if _, ok := s.Jobs[dep]; !ok {
 				errs = append(errs, fmt.Sprintf("job %q: depends_on unknown job %q", name, dep))
 			}
 		}
+		for i, in := range j.Inputs {
+			if in.FromJob == "" {
+				continue
+			}
+			producer, ok := s.Jobs[in.FromJob]
+			if !ok {
+				errs = append(errs, fmt.Sprintf("job %q: input[%d] from_job references unknown job %q", name, i, in.FromJob))
+				continue
+			}
+			if !producerCommands[producer.Command] || producer.Output == nil ||
+				producer.Output.Path == "" || !SingleFileOutputFormat(producer.Output.Format) {
+				errs = append(errs, fmt.Sprintf(
+					"job %q: input[%d] from_job %q must name a convert/merge/split job that writes a single-file output",
+					name, i, in.FromJob))
+			}
+		}
 	}
 	if cycle := s.findCycle(); cycle != "" {
 		errs = append(errs, fmt.Sprintf("dependency cycle detected: %s", cycle))
@@ -262,7 +473,8 @@ func (j *Job) validate() []string {
 	var e []string
 
 	switch j.Command {
-	case CommandConvert, CommandMerge, CommandScriptsList, CommandTempl:
+	case CommandConvert, CommandMerge, CommandScriptsList, CommandScriptsExec,
+		CommandTempl, CommandSplit, CommandInspect, CommandDiff:
 	case "":
 		e = append(e, "missing command")
 		return e
@@ -282,6 +494,7 @@ func (j *Job) validate() []string {
 	}
 	checkPath("logfile", j.Logfile)
 	checkPath("template", j.Template)
+	checkPath("rules", j.Rules)
 	for _, in := range j.Inputs {
 		checkPath("input path", in.Path)
 	}
@@ -291,6 +504,13 @@ func (j *Job) validate() []string {
 	if j.Output != nil {
 		checkPath("output path", j.Output.Path)
 	}
+	if j.Report != nil {
+		checkPath("report path", j.Report.Path)
+	}
+
+	if _, err := parseHumanSize(j.LogMaxSize); err != nil {
+		e = append(e, fmt.Sprintf("log_max_size: %v", err))
+	}
 
 	switch j.Command {
 	case CommandConvert, CommandMerge:
@@ -350,11 +570,113 @@ func (j *Job) validate() []string {
 		if j.Output != nil && j.Output.Format != "" {
 			e = append(e, "output.format is not valid for command \"templ\"")
 		}
+	case CommandSplit:
+		if len(j.Inputs) < 1 {
+			e = append(e, "command \"split\" requires at least 1 input")
+		}
+		for i, in := range j.Inputs {
+			e = append(e, validateInput(i, in)...)
+		}
+		if len(j.ScriptDirs) > 0 {
+			e = append(e, "script_dirs is not valid for command \"split\"")
+		}
+		if j.Report != nil {
+			e = append(e, "report is not valid for command \"split\" (use output)")
+		}
+		if j.Output == nil {
+			e = append(e, "missing output")
+		} else {
+			if j.Output.ConnEnv != "" {
+				e = append(e, "command \"split\" writes a file; output.conn_env is not supported")
+			}
+			e = append(e, validateOutput(*j.Output)...)
+		}
+	case CommandInspect:
+		if len(j.Inputs) < 1 {
+			e = append(e, "command \"inspect\" requires at least 1 input")
+		}
+		for i, in := range j.Inputs {
+			e = append(e, validateInput(i, in)...)
+		}
+		if len(j.ScriptDirs) > 0 {
+			e = append(e, "script_dirs is not valid for command \"inspect\"")
+		}
+		if j.Output != nil {
+			e = append(e, "output is not valid for command \"inspect\" (use report)")
+		}
+		e = append(e, validateReport(j.Report, "inspect", inspectReportFormats, "markdown")...)
+	case CommandDiff:
+		if len(j.Inputs) != 2 {
+			e = append(e, "command \"diff\" requires exactly 2 inputs (source, target)")
+		}
+		for i, in := range j.Inputs {
+			e = append(e, validateInput(i, in)...)
+		}
+		if len(j.ScriptDirs) > 0 {
+			e = append(e, "script_dirs is not valid for command \"diff\"")
+		}
+		if j.Output != nil {
+			e = append(e, "output is not valid for command \"diff\" (use report)")
+		}
+		e = append(e, validateReport(j.Report, "diff", diffReportFormats, "summary")...)
+	case CommandScriptsExec:
+		if len(j.ScriptDirs) == 0 {
+			e = append(e, "command \"scripts-exec\" requires at least one script_dir")
+		}
+		if len(j.Inputs) > 0 {
+			e = append(e, "inputs is not valid for command \"scripts-exec\"")
+		}
+		if j.Report != nil {
+			e = append(e, "report is not valid for command \"scripts-exec\"")
+		}
+		if j.Output == nil || j.Output.ConnEnv == "" {
+			e = append(e, "command \"scripts-exec\" requires output.conn_env (an environment variable name holding a connection string)")
+		} else {
+			if j.Output.Path != "" {
+				e = append(e, "command \"scripts-exec\" executes against a database; output.path is not supported")
+			}
+			f := strings.ToLower(j.Output.Format)
+			if f != "" && f != "pgsql" {
+				e = append(e, fmt.Sprintf("command \"scripts-exec\" only supports pgsql databases (got %q)", j.Output.Format))
+			}
+			if looksLikeSecret(j.Output.ConnEnv) {
+				e = append(e, "output: conn_env must be an environment variable name, not a connection string")
+			}
+		}
+	}
+	return e
+}
+
+// validateReport checks a Report block for the inspect/diff commands.
+func validateReport(r *Report, cmd string, allowed map[string]bool, defFmt string) []string {
+	if r == nil {
+		return []string{fmt.Sprintf("command %q requires a report block", cmd)}
+	}
+	var e []string
+	f := strings.ToLower(r.Format)
+	if f == "" {
+		f = defFmt
+	}
+	if !allowed[f] {
+		names := make([]string, 0, len(allowed))
+		for k := range allowed {
+			names = append(names, k)
+		}
+		sort.Strings(names)
+		e = append(e, fmt.Sprintf("command %q report.format %q is not supported (use: %s)", cmd, r.Format, strings.Join(names, ", ")))
+	}
+	// A diff summary may be written to the log; everything else needs a path.
+	summaryToLog := cmd == "diff" && f == "summary"
+	if r.Path == "" && !summaryToLog {
+		e = append(e, fmt.Sprintf("command %q requires report.path", cmd))
 	}
 	return e
 }
 
 func validateTemplInput(i int, in Input) []string {
+	if in.FromJob != "" {
+		return fromJobInputShape(i, in)
+	}
 	var e []string
 	if in.Format == "" {
 		return []string{fmt.Sprintf("input[%d]: missing format", i)}
@@ -383,7 +705,27 @@ func validateTemplInput(i int, in Input) []string {
 	return e
 }
 
+// fromJobInputShape checks the structural rules for an input that pulls its
+// schema from another job's output. The referenced job's existence and kind
+// are checked in Set.Validate, which can see the whole set.
+func fromJobInputShape(i int, in Input) []string {
+	var e []string
+	if in.Path != "" {
+		e = append(e, fmt.Sprintf("input[%d]: from_job takes no path", i))
+	}
+	if in.Format != "" {
+		e = append(e, fmt.Sprintf("input[%d]: from_job inherits the producer's format; drop format", i))
+	}
+	if in.ConnEnv != "" {
+		e = append(e, fmt.Sprintf("input[%d]: from_job takes no conn_env", i))
+	}
+	return e
+}
+
 func validateInput(i int, in Input) []string {
+	if in.FromJob != "" {
+		return fromJobInputShape(i, in)
+	}
 	var e []string
 	if in.Format == "" {
 		e = append(e, fmt.Sprintf("input[%d]: missing format", i))
@@ -487,9 +829,42 @@ func SafeJoin(root, rel string) (string, error) {
 	if rp == ".." || strings.HasPrefix(rp, ".."+string(filepath.Separator)) {
 		return "", fmt.Errorf("path %q escapes the job file directory", rel)
 	}
+	// Symlink hardening: resolve symlinks on the root and on the deepest
+	// existing ancestor of the target, and require the target to still live
+	// inside the resolved root. This catches a symlink inside the job-file
+	// directory that points outside it.
+	realRoot, err := filepath.EvalSymlinks(absRoot)
+	if err != nil {
+		return "", fmt.Errorf("cannot resolve job file directory: %w", err)
+	}
+	realAnc, err := filepath.EvalSymlinks(deepestExistingAncestor(joined))
+	if err != nil {
+		return "", fmt.Errorf("cannot resolve path %q: %w", rel, err)
+	}
+	if realAnc != realRoot {
+		if r, err := filepath.Rel(realRoot, realAnc); err != nil ||
+			r == ".." || strings.HasPrefix(r, ".."+string(filepath.Separator)) {
+			return "", fmt.Errorf("path %q resolves outside the job file directory via a symlink", rel)
+		}
+	}
 	return joined, nil
 }
 
+// deepestExistingAncestor returns p itself if it exists, otherwise the nearest
+// existing parent directory (falling back to the filesystem root).
+func deepestExistingAncestor(p string) string {
+	for {
+		if _, err := os.Lstat(p); err == nil {
+			return p
+		}
+		parent := filepath.Dir(p)
+		if parent == p {
+			return p
+		}
+		p = parent
+	}
+}
+
 // Plan returns the jobs to execute for name in dependency order. When
 // includeDeps is false only the named job is returned (its declared
 // dependencies are still validated to exist and be acyclic by Validate).
@@ -514,7 +889,7 @@ func (s *Set) Plan(name string, includeDeps bool) ([]*Job, error) {
 		}
 		inProgress[n] = true
 		j := s.Jobs[n]
-		deps := append([]string(nil), j.DependsOn...)
+		deps := j.effectiveDeps()
 		sort.Strings(deps)
 		for _, d := range deps {
 			if _, ok := s.Jobs[d]; !ok {
@@ -543,7 +918,7 @@ func (s *Set) findCycle() string {
 	dfs = func(n string) []string {
 		color[n] = 1
 		stack = append(stack, n)
-		deps := append([]string(nil), s.Jobs[n].DependsOn...)
+		deps := s.Jobs[n].effectiveDeps()
 		sort.Strings(deps)
 		for _, d := range deps {
 			if _, ok := s.Jobs[d]; !ok {

pkg/jobs/jobs_test.go

@@ -56,13 +56,82 @@ func TestLoadRejectsUnknownFields(t *testing.T) {
 	}
 }
 
-func TestLoadRejectsBadVersion(t *testing.T) {
+func TestLoadWarnsOnNewerVersion(t *testing.T) {
 	dir := t.TempDir()
 	p := filepath.Join(dir, "relspec.yml")
-	write(t, p, "version: 2\njobs:\n  a:\n    command: convert\n")
+	// A newer version loads best-effort with a warning, and unknown fields
-	_, err := Load([]string{p})
+	// from the newer schema are ignored rather than rejected.
-	if err == nil || !strings.Contains(err.Error(), "unsupported version") {
+	write(t, p, "version: 99\njobs:\n  a:\n    command: convert\n"+
-		t.Fatalf("expected unsupported version error, got %v", err)
+		"    inputs:\n      - path: a.dbml\n        format: dbml\n"+
+		"    output:\n      format: json\n      path: out.json\n"+
+		"    future_field: whatever\n")
+	set, err := Load([]string{p})
+	if err != nil {
+		t.Fatalf("newer version should load, got %v", err)
+	}
+	if len(set.Warnings) == 0 {
+		t.Fatal("expected a warning about the newer version")
+	}
+	if err := set.Validate(); err != nil {
+		t.Fatalf("validate: %v", err)
+	}
+}
+
+func TestLoadAcceptsOmittedVersion(t *testing.T) {
+	dir := t.TempDir()
+	p := filepath.Join(dir, "relspec.yml")
+	write(t, p, "jobs:\n  a:\n    command: convert\n"+
+		"    inputs:\n      - path: a.dbml\n        format: dbml\n"+
+		"    output:\n      format: json\n      path: out.json\n")
+	set, err := Load([]string{p})
+	if err != nil {
+		t.Fatalf("omitted version should load, got %v", err)
+	}
+	if len(set.Warnings) != 0 {
+		t.Fatalf("omitted version should not warn, got %v", set.Warnings)
+	}
+}
+
+func TestLoadStillRejectsUnknownFieldsAtCurrentVersion(t *testing.T) {
+	dir := t.TempDir()
+	p := filepath.Join(dir, "relspec.yml")
+	write(t, p, "version: 1\njobs:\n  a:\n    command: convert\n    bogus: true\n")
+	if _, err := Load([]string{p}); err == nil {
+		t.Fatal("expected unknown-field rejection at the current version")
+	}
+}
+
+func TestParseHumanSize(t *testing.T) {
+	cases := []struct {
+		in   string
+		want int64
+		bad  bool
+	}{
+		{"", 0, false},
+		{"512", 512, false},
+		{"512B", 512, false},
+		{"1KB", 1 << 10, false},
+		{"5MB", 5 << 20, false},
+		{"1gb", 1 << 30, false},
+		{" 2 MB ", 2 << 20, false},
+		{"nonsense", 0, true},
+		{"-1MB", 0, true},
+	}
+	for _, c := range cases {
+		got, err := parseHumanSize(c.in)
+		if c.bad {
+			if err == nil {
+				t.Errorf("parseHumanSize(%q): expected error", c.in)
+			}
+			continue
+		}
+		if err != nil {
+			t.Errorf("parseHumanSize(%q): %v", c.in, err)
+			continue
+		}
+		if got != c.want {
+			t.Errorf("parseHumanSize(%q) = %d, want %d", c.in, got, c.want)
+		}
 	}
 }
 
@@ -239,6 +308,135 @@ func TestShippedExampleIsValid(t *testing.T) {
 	}
 }
 
+func TestFromJobWiring(t *testing.T) {
+	content := "version: 1\njobs:\n" +
+		"  producer:\n    command: convert\n" +
+		"    inputs:\n      - path: a.dbml\n        format: dbml\n" +
+		"    output:\n      format: json\n      path: build/schema.json\n" +
+		"  consumer:\n    command: convert\n" +
+		"    inputs:\n      - from_job: producer\n" +
+		"    output:\n      format: yaml\n      path: build/schema.yaml\n"
+	set := loadOne(t, content)
+	if err := set.Validate(); err != nil {
+		t.Fatalf("validate: %v", err)
+	}
+	plan, err := set.Plan("consumer", true)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(plan) != 2 || plan[0].Name != "producer" || plan[1].Name != "consumer" {
+		t.Fatalf("plan = %v, want [producer consumer]", plan)
+	}
+}
+
+func TestFromJobRejectsNonProducer(t *testing.T) {
+	content := "version: 1\njobs:\n" +
+		"  lister:\n    command: scripts-list\n    script_dirs: [migrations]\n" +
+		"  consumer:\n    command: convert\n" +
+		"    inputs:\n      - from_job: lister\n" +
+		"    output:\n      format: yaml\n      path: out.yaml\n"
+	set := loadOne(t, content)
+	if err := set.Validate(); err == nil || !strings.Contains(err.Error(), "from_job") {
+		t.Fatalf("want from_job producer error, got %v", err)
+	}
+}
+
+func TestFromJobRejectsUnknownJob(t *testing.T) {
+	content := "version: 1\njobs:\n" +
+		"  consumer:\n    command: convert\n" +
+		"    inputs:\n      - from_job: ghost\n" +
+		"    output:\n      format: yaml\n      path: out.yaml\n"
+	set := loadOne(t, content)
+	if err := set.Validate(); err == nil || !strings.Contains(err.Error(), "unknown job") {
+		t.Fatalf("want unknown job error, got %v", err)
+	}
+}
+
+func TestFromJobCycleDetected(t *testing.T) {
+	content := "version: 1\njobs:\n" +
+		"  a:\n    command: convert\n" +
+		"    inputs:\n      - from_job: b\n" +
+		"    output:\n      format: json\n      path: a.json\n" +
+		"  b:\n    command: convert\n" +
+		"    inputs:\n      - from_job: a\n" +
+		"    output:\n      format: json\n      path: b.json\n"
+	set := loadOne(t, content)
+	if err := set.Validate(); err == nil || !strings.Contains(err.Error(), "cycle") {
+		t.Fatalf("want cycle error, got %v", err)
+	}
+}
+
+func TestSplitJobValidation(t *testing.T) {
+	set := loadOne(t, "version: 1\njobs:\n  s:\n    command: split\n"+
+		"    inputs:\n      - path: a.dbml\n        format: dbml\n")
+	if err := set.Validate(); err == nil || !strings.Contains(err.Error(), "missing output") {
+		t.Fatalf("want missing output, got %v", err)
+	}
+	set = loadOne(t, "version: 1\njobs:\n  s:\n    command: split\n"+
+		"    inputs:\n      - path: a.dbml\n        format: dbml\n"+
+		"    select:\n      tables: [users]\n"+
+		"    output:\n      format: json\n      path: out.json\n")
+	if err := set.Validate(); err != nil {
+		t.Fatalf("expected valid split job, got %v", err)
+	}
+}
+
+func TestInspectJobValidation(t *testing.T) {
+	set := loadOne(t, "version: 1\njobs:\n  i:\n    command: inspect\n"+
+		"    inputs:\n      - path: a.dbml\n        format: dbml\n")
+	if err := set.Validate(); err == nil || !strings.Contains(err.Error(), "report") {
+		t.Fatalf("want report required, got %v", err)
+	}
+	set = loadOne(t, "version: 1\njobs:\n  i:\n    command: inspect\n"+
+		"    inputs:\n      - path: a.dbml\n        format: dbml\n"+
+		"    report:\n      format: json\n      path: build/report.json\n")
+	if err := set.Validate(); err != nil {
+		t.Fatalf("expected valid inspect job, got %v", err)
+	}
+}
+
+func TestDiffJobValidation(t *testing.T) {
+	set := loadOne(t, "version: 1\njobs:\n  d:\n    command: diff\n"+
+		"    inputs:\n      - path: a.dbml\n        format: dbml\n"+
+		"    report:\n      format: summary\n")
+	if err := set.Validate(); err == nil || !strings.Contains(err.Error(), "exactly 2 inputs") {
+		t.Fatalf("want exactly 2 inputs, got %v", err)
+	}
+	set = loadOne(t, "version: 1\njobs:\n  d:\n    command: diff\n"+
+		"    inputs:\n      - path: a.dbml\n        format: dbml\n"+
+		"      - path: b.dbml\n        format: dbml\n"+
+		"    report:\n      format: summary\n")
+	if err := set.Validate(); err != nil {
+		t.Fatalf("expected valid diff job, got %v", err)
+	}
+}
+
+func TestScriptsExecValidation(t *testing.T) {
+	set := loadOne(t, "version: 1\njobs:\n  x:\n    command: scripts-exec\n"+
+		"    script_dirs: [migrations]\n")
+	if err := set.Validate(); err == nil || !strings.Contains(err.Error(), "conn_env") {
+		t.Fatalf("want output.conn_env required, got %v", err)
+	}
+	set = loadOne(t, "version: 1\njobs:\n  x:\n    command: scripts-exec\n"+
+		"    script_dirs: [migrations]\n"+
+		"    output:\n      conn_env: TARGET_DB_URL\n")
+	if err := set.Validate(); err != nil {
+		t.Fatalf("expected valid scripts-exec job, got %v", err)
+	}
+}
+
+func TestSafeJoinRejectsSymlinkEscape(t *testing.T) {
+	root := t.TempDir()
+	outside := t.TempDir()
+	link := filepath.Join(root, "link")
+	if err := os.Symlink(outside, link); err != nil {
+		t.Skipf("symlink not supported: %v", err)
+	}
+	if _, err := SafeJoin(root, "link/x.sql"); err == nil {
+		t.Fatal("expected rejection of a path escaping via a symlink")
+	}
+}
+
 func TestScriptsListValidation(t *testing.T) {
 	set := loadOne(t, "version: 1\njobs:\n  s:\n    command: scripts-list\n")
 	if err := set.Validate(); err == nil || !strings.Contains(err.Error(), "script_dir") {

pkg/models/directives.go

@@ -0,0 +1,237 @@
+package models
+
+import (
+	"fmt"
+	"sort"
+	"strings"
+)
+
+// Directive is a dialect-specific instruction embedded in a source schema
+// (currently DBML) that is preserved losslessly in the intermediate model and
+// consumed only by the writer for its namespace. Directives are stored in the
+// Metadata map of the object they apply to, under DirectivesMetadataKey.
+//
+// Example DBML: `@postgres: partition by RANGE (created_at)` parses to
+// Directive{Namespace: "postgres", Key: "partition", Args: "partition by RANGE (created_at)"}.
+type Directive struct {
+	// Namespace is the dialect the directive targets, e.g. "postgres" or "sqlite".
+	Namespace string `json:"namespace" yaml:"namespace"`
+	// Key is the lowercased first token of Args, used for duplicate detection
+	// and writer dispatch.
+	Key string `json:"key,omitempty" yaml:"key,omitempty"`
+	// Args is the verbatim argument text following the "@namespace:" prefix.
+	Args string `json:"args" yaml:"args"`
+	// Line is the 1-based source line the directive was read from, when known.
+	Line int `json:"line,omitempty" yaml:"line,omitempty"`
+}
+
+// DirectivesMetadataKey is the Metadata map key under which the ordered list of
+// dialect directives for an object is stored.
+const DirectivesMetadataKey = "directives"
+
+// DirectiveKey derives the Key for a directive from its argument text: the
+// lowercased first whitespace-delimited token.
+func DirectiveKey(args string) string {
+	fields := strings.Fields(args)
+	if len(fields) == 0 {
+		return ""
+	}
+	return strings.ToLower(fields[0])
+}
+
+// AddDirective appends d to the directive list stored in meta. The caller is
+// responsible for ensuring meta is non-nil (all Init* constructors allocate it).
+// If d.Key is empty it is derived from d.Args.
+func AddDirective(meta map[string]any, d Directive) {
+	if meta == nil {
+		return
+	}
+	if d.Key == "" {
+		d.Key = DirectiveKey(d.Args)
+	}
+	existing := GetDirectives(meta)
+	existing = append(existing, d)
+	meta[DirectivesMetadataKey] = existing
+}
+
+// GetDirectives returns the directives stored in meta, sorted deterministically
+// by (Namespace, Line, Args). It tolerates both a freshly built []Directive and
+// the []any of map[string]any produced by a JSON/YAML round-trip.
+func GetDirectives(meta map[string]any) []Directive {
+	if meta == nil {
+		return nil
+	}
+	raw, ok := meta[DirectivesMetadataKey]
+	if !ok || raw == nil {
+		return nil
+	}
+
+	var out []Directive
+	switch v := raw.(type) {
+	case []Directive:
+		out = append(out, v...)
+	case []any:
+		for _, item := range v {
+			if d, ok := directiveFromAny(item); ok {
+				out = append(out, d)
+			}
+		}
+	}
+
+	sort.SliceStable(out, func(i, j int) bool {
+		if out[i].Namespace != out[j].Namespace {
+			return out[i].Namespace < out[j].Namespace
+		}
+		if out[i].Line != out[j].Line {
+			return out[i].Line < out[j].Line
+		}
+		return out[i].Args < out[j].Args
+	})
+	return out
+}
+
+// directiveFromAny decodes a single directive from the loosely typed forms that
+// survive a JSON or YAML round-trip (map[string]any / map[any]any).
+func directiveFromAny(item any) (Directive, bool) {
+	switch m := item.(type) {
+	case Directive:
+		return m, true
+	case map[string]any:
+		return directiveFromStringMap(m), true
+	case map[any]any:
+		sm := make(map[string]any, len(m))
+		for k, val := range m {
+			if ks, ok := k.(string); ok {
+				sm[ks] = val
+			}
+		}
+		return directiveFromStringMap(sm), true
+	}
+	return Directive{}, false
+}
+
+func directiveFromStringMap(m map[string]any) Directive {
+	d := Directive{}
+	if s, ok := m["namespace"].(string); ok {
+		d.Namespace = s
+	}
+	if s, ok := m["key"].(string); ok {
+		d.Key = s
+	}
+	if s, ok := m["args"].(string); ok {
+		d.Args = s
+	}
+	switch n := m["line"].(type) {
+	case int:
+		d.Line = n
+	case int64:
+		d.Line = int(n)
+	case float64:
+		d.Line = int(n)
+	}
+	if d.Key == "" {
+		d.Key = DirectiveKey(d.Args)
+	}
+	return d
+}
+
+// DirectivesForNamespace returns the directives in meta that target ns, in the
+// deterministic order of GetDirectives.
+func DirectivesForNamespace(meta map[string]any, ns string) []Directive {
+	all := GetDirectives(meta)
+	if len(all) == 0 {
+		return nil
+	}
+	out := make([]Directive, 0, len(all))
+	for _, d := range all {
+		if d.Namespace == ns {
+			out = append(out, d)
+		}
+	}
+	return out
+}
+
+// HasDirective reports whether meta contains a directive with the given
+// namespace and key.
+func HasDirective(meta map[string]any, ns, key string) bool {
+	for _, d := range GetDirectives(meta) {
+		if d.Namespace == ns && d.Key == key {
+			return true
+		}
+	}
+	return false
+}
+
+// DirectiveSpec describes a documented directive in the catalog.
+type DirectiveSpec struct {
+	// Singleton means only one directive with this namespace/key may appear at
+	// a single location; a second one is a parse error.
+	Singleton bool
+	// Locations lists the location kinds the directive is valid at
+	// ("database", "table", "column", "index").
+	Locations []string
+}
+
+// Location kinds a directive may attach to.
+const (
+	DirectiveLocationDatabase = "database"
+	DirectiveLocationTable    = "table"
+	DirectiveLocationColumn   = "column"
+	DirectiveLocationIndex    = "index"
+)
+
+// DirectiveCatalog is the set of documented directives per namespace. It is used
+// for strict-mode validation in readers and writers; unknown namespaces/keys are
+// still preserved losslessly when strict mode is off.
+var DirectiveCatalog = map[string]map[string]DirectiveSpec{
+	"postgres": {
+		"partition":   {Singleton: true, Locations: []string{DirectiveLocationTable}},
+		"tablespace":  {Singleton: true, Locations: []string{DirectiveLocationTable, DirectiveLocationIndex}},
+		"inherits":    {Singleton: true, Locations: []string{DirectiveLocationTable}},
+		"with":        {Singleton: false, Locations: []string{DirectiveLocationTable, DirectiveLocationIndex}},
+		"storage":     {Singleton: true, Locations: []string{DirectiveLocationColumn}},
+		"compression": {Singleton: true, Locations: []string{DirectiveLocationColumn}},
+		"identity":    {Singleton: true, Locations: []string{DirectiveLocationColumn}},
+	},
+	"sqlite": {
+		"without": {Singleton: true, Locations: []string{DirectiveLocationTable}},
+		"strict":  {Singleton: true, Locations: []string{DirectiveLocationTable}},
+		"collate": {Singleton: true, Locations: []string{DirectiveLocationColumn}},
+	},
+}
+
+// LookupDirectiveSpec returns the catalog spec for a namespace/key and whether
+// it is documented.
+func LookupDirectiveSpec(ns, key string) (DirectiveSpec, bool) {
+	keys, ok := DirectiveCatalog[ns]
+	if !ok {
+		return DirectiveSpec{}, false
+	}
+	spec, ok := keys[key]
+	return spec, ok
+}
+
+// DirectiveLocationAllowed reports whether a documented directive may appear at
+// the given location. Unknown directives (not in the catalog) are allowed
+// everywhere so they can be preserved.
+func DirectiveLocationAllowed(ns, key, location string) bool {
+	spec, ok := LookupDirectiveSpec(ns, key)
+	if !ok {
+		return true
+	}
+	for _, l := range spec.Locations {
+		if l == location {
+			return true
+		}
+	}
+	return false
+}
+
+// FormatDirectiveLine renders a directive back to its DBML source form, e.g.
+// "@postgres: partition by RANGE (created_at)" or "@postgres(id): identity always".
+func FormatDirectiveLine(d Directive, target string) string {
+	if target != "" {
+		return fmt.Sprintf("@%s(%s): %s", d.Namespace, target, d.Args)
+	}
+	return fmt.Sprintf("@%s: %s", d.Namespace, d.Args)
+}

pkg/models/directives_test.go

@@ -0,0 +1,133 @@
+package models
+
+import (
+	"encoding/json"
+	"testing"
+)
+
+func TestDirectiveKey(t *testing.T) {
+	cases := map[string]string{
+		"partition by RANGE (created_at)": "partition",
+		"WITHOUT ROWID":                   "without",
+		"  strict  ":                      "strict",
+		"":                                "",
+	}
+	for args, want := range cases {
+		if got := DirectiveKey(args); got != want {
+			t.Errorf("DirectiveKey(%q) = %q, want %q", args, got, want)
+		}
+	}
+}
+
+func TestAddDirectiveDerivesKey(t *testing.T) {
+	meta := map[string]any{}
+	AddDirective(meta, Directive{Namespace: "postgres", Args: "partition by RANGE (x)", Line: 2})
+	AddDirective(meta, Directive{Namespace: "postgres", Key: "tablespace", Args: "tablespace fast", Line: 3})
+
+	got := GetDirectives(meta)
+	if len(got) != 2 {
+		t.Fatalf("got %d directives, want 2", len(got))
+	}
+	if got[0].Key != "partition" {
+		t.Errorf("derived key = %q, want %q", got[0].Key, "partition")
+	}
+	if got[1].Key != "tablespace" {
+		t.Errorf("explicit key = %q, want %q", got[1].Key, "tablespace")
+	}
+}
+
+func TestAddDirectiveNilMeta(t *testing.T) {
+	// Must not panic.
+	AddDirective(nil, Directive{Namespace: "postgres", Args: "strict"})
+}
+
+func TestGetDirectivesOrdering(t *testing.T) {
+	meta := map[string]any{}
+	AddDirective(meta, Directive{Namespace: "sqlite", Args: "strict", Line: 9})
+	AddDirective(meta, Directive{Namespace: "postgres", Args: "with (b)", Line: 5})
+	AddDirective(meta, Directive{Namespace: "postgres", Args: "with (a)", Line: 5})
+	AddDirective(meta, Directive{Namespace: "postgres", Args: "partition by x", Line: 2})
+
+	got := GetDirectives(meta)
+	wantArgs := []string{"partition by x", "with (a)", "with (b)", "strict"}
+	if len(got) != len(wantArgs) {
+		t.Fatalf("got %d directives, want %d", len(got), len(wantArgs))
+	}
+	for i, w := range wantArgs {
+		if got[i].Args != w {
+			t.Errorf("directive[%d].Args = %q, want %q", i, got[i].Args, w)
+		}
+	}
+}
+
+func TestGetDirectivesTolerantDecodeAfterJSON(t *testing.T) {
+	meta := map[string]any{}
+	AddDirective(meta, Directive{Namespace: "postgres", Args: "partition by RANGE (created_at)", Line: 4})
+	AddDirective(meta, Directive{Namespace: "sqlite", Args: "without rowid", Line: 6})
+
+	blob, err := json.Marshal(meta)
+	if err != nil {
+		t.Fatalf("marshal: %v", err)
+	}
+	var round map[string]any
+	if err := json.Unmarshal(blob, &round); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+
+	got := GetDirectives(round)
+	if len(got) != 2 {
+		t.Fatalf("got %d directives after JSON round-trip, want 2", len(got))
+	}
+	if got[0].Namespace != "postgres" || got[0].Key != "partition" || got[0].Line != 4 {
+		t.Errorf("post-JSON directive[0] = %+v", got[0])
+	}
+	if got[0].Args != "partition by RANGE (created_at)" {
+		t.Errorf("post-JSON args not verbatim: %q", got[0].Args)
+	}
+	if got[1].Namespace != "sqlite" || got[1].Key != "without" {
+		t.Errorf("post-JSON directive[1] = %+v", got[1])
+	}
+}
+
+func TestDirectivesForNamespaceAndHasDirective(t *testing.T) {
+	meta := map[string]any{}
+	AddDirective(meta, Directive{Namespace: "postgres", Args: "partition by x", Line: 1})
+	AddDirective(meta, Directive{Namespace: "sqlite", Args: "strict", Line: 2})
+
+	pg := DirectivesForNamespace(meta, "postgres")
+	if len(pg) != 1 || pg[0].Key != "partition" {
+		t.Errorf("DirectivesForNamespace(postgres) = %+v", pg)
+	}
+	if !HasDirective(meta, "sqlite", "strict") {
+		t.Error("HasDirective(sqlite, strict) = false, want true")
+	}
+	if HasDirective(meta, "postgres", "tablespace") {
+		t.Error("HasDirective(postgres, tablespace) = true, want false")
+	}
+}
+
+func TestDirectiveLocationAllowed(t *testing.T) {
+	if !DirectiveLocationAllowed("postgres", "partition", DirectiveLocationTable) {
+		t.Error("partition should be allowed at table level")
+	}
+	if DirectiveLocationAllowed("postgres", "partition", DirectiveLocationColumn) {
+		t.Error("partition should not be allowed at column level")
+	}
+	// Unknown directives are allowed everywhere so they can be preserved.
+	if !DirectiveLocationAllowed("postgres", "bogus", DirectiveLocationDatabase) {
+		t.Error("unknown key should be allowed everywhere")
+	}
+	if !DirectiveLocationAllowed("madeup", "x", DirectiveLocationTable) {
+		t.Error("unknown namespace should be allowed everywhere")
+	}
+}
+
+func TestFormatDirectiveLine(t *testing.T) {
+	d := Directive{Namespace: "postgres", Key: "identity", Args: "identity always"}
+	if got := FormatDirectiveLine(d, ""); got != "@postgres: identity always" {
+		t.Errorf("FormatDirectiveLine no target = %q", got)
+	}
+	if got := FormatDirectiveLine(d, "id"); got != "@postgres(id): identity always" {
+		t.Errorf("FormatDirectiveLine with target = %q", got)
+	}
+}

pkg/models/models.go

@@ -32,6 +32,7 @@ type Database struct {
 	DatabaseType    DatabaseType   `json:"database_type,omitempty" yaml:"database_type,omitempty" xml:"database_type,omitempty"`
 	DatabaseVersion string         `json:"database_version,omitempty" yaml:"database_version,omitempty" xml:"database_version,omitempty"`
 	SourceFormat    string         `json:"source_format,omitempty" yaml:"source_format,omitempty" xml:"source_format,omitempty"` // Source Format of the database.
+	Metadata        map[string]any `json:"metadata,omitempty" yaml:"metadata,omitempty" xml:"-"`
 	UpdatedAt       string         `json:"updatedat,omitempty" yaml:"updatedat,omitempty" xml:"updatedat,omitempty"`
 	GUID            string         `json:"guid" yaml:"guid" xml:"guid"`
 }
@@ -240,6 +241,7 @@ type Column struct {
 	IsPrimaryKey  bool           `json:"is_primary_key" yaml:"is_primary_key" xml:"is_primary_key"`
 	Comment       string         `json:"comment,omitempty" yaml:"comment,omitempty" xml:"comment,omitempty"`
 	Collation     string         `json:"collation,omitempty" yaml:"collation,omitempty" xml:"collation,omitempty"`
+	Metadata      map[string]any `json:"metadata,omitempty" yaml:"metadata,omitempty" xml:"-"`
 	Sequence      uint           `json:"sequence,omitempty" yaml:"sequence,omitempty" xml:"sequence,omitempty"`
 	GUID          string         `json:"guid" yaml:"guid" xml:"guid"`
 }
@@ -263,6 +265,7 @@ type Index struct {
 	Concurrent  bool           `json:"concurrent,omitempty" yaml:"concurrent,omitempty" xml:"concurrent,omitempty"`
 	Include     []string       `json:"include,omitempty" yaml:"include,omitempty" xml:"include,omitempty"` // INCLUDE columns
 	Comment     string         `json:"comment,omitempty" yaml:"comment,omitempty" xml:"comment,omitempty"`
+	Metadata    map[string]any `json:"metadata,omitempty" yaml:"metadata,omitempty" xml:"-"`
 	Sequence    uint           `json:"sequence,omitempty" yaml:"sequence,omitempty" xml:"sequence,omitempty"`
 	GUID        string         `json:"guid" yaml:"guid" xml:"guid"`
 }
@@ -396,6 +399,7 @@ func InitDatabase(name string) *Database {
 		Name:     name,
 		Schemas:  make([]*Schema, 0),
 		Domains:  make([]*Domain, 0),
+		Metadata: make(map[string]any),
 		GUID:     uuid.New().String(),
 	}
 }
@@ -434,6 +438,7 @@ func InitColumn(name, table, schema string) *Column {
 		Name:     name,
 		Table:    table,
 		Schema:   schema,
+		Metadata: make(map[string]any),
 		GUID:     uuid.New().String(),
 	}
 }
@@ -446,6 +451,7 @@ func InitIndex(name, table, schema string) *Index {
 		Schema:   schema,
 		Columns:  make([]string, 0),
 		Include:  make([]string, 0),
+		Metadata: make(map[string]any),
 		GUID:     uuid.New().String(),
 	}
 }

pkg/readers/dbml/README.md

@@ -93,6 +93,50 @@ Ref: posts.user_id > users.id [delete: cascade]
 - Indexes and composite indexes
 - Table notes and column notes
 - Enums
+- Dialect directives (`@postgres:` / `@sqlite:` — see below)
+
+## Dialect directives
+
+Lines of the form `@<namespace>[(<column>)]: <args>` embed database-specific
+features that plain DBML cannot express (partitioning, `WITHOUT ROWID`,
+tablespaces, index storage parameters, …). They are stored losslessly on the
+relevant object's `Metadata` and round-trip unchanged through the DBML writer;
+the PostgreSQL and SQLite writers translate the ones they understand to SQL.
+
+```dbml
+@postgres: search_path myapp
+
+Table myapp.events {
+  id bigint [pk]
+  created_at timestamp [not null]
+  @postgres(id): identity always
+  @postgres: partition by RANGE (created_at)
+  @sqlite: without rowid
+
+  indexes {
+    (created_at) [name: 'idx_events_created']
+    @postgres: with (fillfactor=90)
+  }
+}
+```
+
+| Position | Attaches to |
+|----------|-------------|
+| Before the first `Table {` | database |
+| Table body, no `(target)` | that table |
+| Table body, `(col)` target | column `col` (error if unknown) |
+| Inside `indexes { }` | the most recently listed index entry |
+
+`args` is preserved verbatim; the **key** (lowercased first token) drives
+duplicate detection. Repeated directives are kept in order; catalog "singleton"
+keys error on a second occurrence at the same location. All errors are
+line-numbered.
+
+`ReaderOptions.StrictDirectives` (CLI `--strict-directives`) turns an unknown
+namespace or key into an error instead of preserving it silently.
+
+See [`docs/DBML_DIRECTIVES.md`](../../../docs/DBML_DIRECTIVES.md) for the full
+grammar and the supported-directive matrix.
 
 ## Notes
 

pkg/readers/dbml/directives.go

@@ -0,0 +1,137 @@
+package dbml
+
+import (
+	"fmt"
+	"regexp"
+	"strings"
+
+	"git.warky.dev/wdevs/relspecgo/pkg/models"
+)
+
+// directiveLineRegex matches a dialect directive line:
+//
+//	@postgres: partition by RANGE (created_at)
+//	@postgres(id): identity always
+//
+// Group 1 is the namespace, group 2 the optional (column) target, group 3 the
+// raw argument text (validated separately so error messages can be specific).
+var directiveLineRegex = regexp.MustCompile(`^@([^():]*)(?:\(([^()]*)\))?\s*:(.*)$`)
+
+// namespaceRegex is the grammar for a directive namespace.
+var namespaceRegex = regexp.MustCompile(`^[a-z][a-z0-9_]*$`)
+
+// parsedDirective is a directive line that has been parsed but not yet attached
+// to a model object.
+type parsedDirective struct {
+	namespace string
+	target    string // column name; "" when absent
+	args      string
+	line      int
+}
+
+// parseDirectiveLine parses a single "@namespace[(target)]: args" line.
+func parseDirectiveLine(line string, lineNo int) (parsedDirective, error) {
+	m := directiveLineRegex.FindStringSubmatch(line)
+	if m == nil {
+		return parsedDirective{}, fmt.Errorf(
+			"dbml: line %d: malformed directive %q (expected \"@namespace: args\")", lineNo, line)
+	}
+
+	ns := strings.TrimSpace(m[1])
+	target := strings.TrimSpace(m[2])
+	args := strings.TrimSpace(m[3])
+
+	if !namespaceRegex.MatchString(ns) {
+		return parsedDirective{}, fmt.Errorf(
+			"dbml: line %d: invalid directive namespace %q (must match [a-z][a-z0-9_]*)", lineNo, ns)
+	}
+	if args == "" {
+		return parsedDirective{}, fmt.Errorf("dbml: line %d: directive @%s has no arguments", lineNo, ns)
+	}
+	if target != "" {
+		target = stripQuotes(target)
+	}
+
+	return parsedDirective{namespace: ns, target: target, args: args, line: lineNo}, nil
+}
+
+// attachDirective resolves the target model object from the current parser state
+// and stores the directive in its Metadata, enforcing location, duplicate and
+// strict-mode rules.
+func (r *Reader) attachDirective(
+	pd parsedDirective,
+	db *models.Database,
+	table *models.Table,
+	inTable, inIndexes bool,
+	lastIndex *models.Index,
+) error {
+	strict := r.options != nil && r.options.StrictDirectives
+	key := models.DirectiveKey(pd.args)
+
+	var meta map[string]any
+	var location string
+
+	switch {
+	case inIndexes:
+		if pd.target != "" {
+			return fmt.Errorf("dbml: line %d: directive target (%s) is not allowed inside an indexes block", pd.line, pd.target)
+		}
+		if lastIndex == nil {
+			return fmt.Errorf("dbml: line %d: directive @%s must follow an index definition", pd.line, pd.namespace)
+		}
+		if lastIndex.Metadata == nil {
+			lastIndex.Metadata = make(map[string]any)
+		}
+		meta = lastIndex.Metadata
+		location = models.DirectiveLocationIndex
+
+	case inTable && table != nil:
+		if pd.target != "" {
+			col, ok := table.Columns[pd.target]
+			if !ok {
+				return fmt.Errorf("dbml: line %d: directive target column %q not found in table %q", pd.line, pd.target, table.Name)
+			}
+			if col.Metadata == nil {
+				col.Metadata = make(map[string]any)
+			}
+			meta = col.Metadata
+			location = models.DirectiveLocationColumn
+		} else {
+			if table.Metadata == nil {
+				table.Metadata = make(map[string]any)
+			}
+			meta = table.Metadata
+			location = models.DirectiveLocationTable
+		}
+
+	default:
+		if pd.target != "" {
+			return fmt.Errorf("dbml: line %d: directive target (%s) is only valid inside a table", pd.line, pd.target)
+		}
+		if db.Metadata == nil {
+			db.Metadata = make(map[string]any)
+		}
+		meta = db.Metadata
+		location = models.DirectiveLocationDatabase
+	}
+
+	spec, documented := models.LookupDirectiveSpec(pd.namespace, key)
+
+	if strict && !documented {
+		return fmt.Errorf("dbml: line %d: unknown directive @%s: %s (strict mode)", pd.line, pd.namespace, key)
+	}
+	if documented && !models.DirectiveLocationAllowed(pd.namespace, key, location) {
+		return fmt.Errorf("dbml: line %d: directive @%s: %s is not valid at %s level", pd.line, pd.namespace, key, location)
+	}
+	if documented && spec.Singleton && models.HasDirective(meta, pd.namespace, key) {
+		return fmt.Errorf("dbml: line %d: duplicate @%s directive %q at %s level", pd.line, pd.namespace, key, location)
+	}
+
+	models.AddDirective(meta, models.Directive{
+		Namespace: pd.namespace,
+		Key:       key,
+		Args:      pd.args,
+		Line:      pd.line,
+	})
+	return nil
+}

pkg/readers/dbml/directives_test.go

@@ -0,0 +1,181 @@
+package dbml
+
+import (
+	"strings"
+	"testing"
+
+	"git.warky.dev/wdevs/relspecgo/pkg/models"
+	"git.warky.dev/wdevs/relspecgo/pkg/readers"
+)
+
+func parse(t *testing.T, strict bool, src string) (*models.Database, error) {
+	t.Helper()
+	r := NewReader(&readers.ReaderOptions{StrictDirectives: strict})
+	return r.parseDBML(src)
+}
+
+func firstTable(t *testing.T, db *models.Database) *models.Table {
+	t.Helper()
+	if len(db.Schemas) == 0 || len(db.Schemas[0].Tables) == 0 {
+		t.Fatal("no table parsed")
+	}
+	return db.Schemas[0].Tables[0]
+}
+
+func TestDirectives_AttachAtEachLocation(t *testing.T) {
+	src := `@postgres: search_path myapp
+
+Table myapp.events {
+  id bigint [pk]
+  created_at timestamp [not null]
+  @postgres(id): identity always
+  @postgres: partition by RANGE (created_at)
+
+  indexes {
+    (created_at) [name: 'idx_events_created']
+    @postgres: with (fillfactor=90)
+  }
+}
+`
+	db, err := parse(t, false, src)
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+
+	if !models.HasDirective(db.Metadata, "postgres", "search_path") {
+		t.Errorf("database-level directive missing: %+v", db.Metadata)
+	}
+
+	tbl := firstTable(t, db)
+	if !models.HasDirective(tbl.Metadata, "postgres", "partition") {
+		t.Errorf("table-level directive missing: %+v", tbl.Metadata)
+	}
+
+	col := tbl.Columns["id"]
+	if col == nil || !models.HasDirective(col.Metadata, "postgres", "identity") {
+		t.Errorf("column-level directive missing")
+	}
+	// Verbatim args preserved.
+	if d := models.DirectivesForNamespace(col.Metadata, "postgres"); len(d) != 1 || d[0].Args != "identity always" {
+		t.Errorf("column directive args = %+v", d)
+	}
+
+	var idx *models.Index
+	for _, i := range tbl.Indexes {
+		idx = i
+	}
+	if idx == nil || !models.HasDirective(idx.Metadata, "postgres", "with") {
+		t.Errorf("index-level directive missing: %+v", idx)
+	}
+}
+
+func TestDirectives_RepeatablePreservedAndOrdered(t *testing.T) {
+	src := `Table s.t {
+  id int [pk]
+  @postgres: with (fillfactor=90)
+  @postgres: with (autovacuum_enabled=off)
+}
+`
+	db, err := parse(t, false, src)
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	tbl := firstTable(t, db)
+	got := models.DirectivesForNamespace(tbl.Metadata, "postgres")
+	if len(got) != 2 {
+		t.Fatalf("got %d directives, want 2", len(got))
+	}
+	if got[0].Args != "with (fillfactor=90)" || got[1].Args != "with (autovacuum_enabled=off)" {
+		t.Errorf("repeatable directives out of order: %+v", got)
+	}
+}
+
+func TestDirectives_SingletonDuplicateErrors(t *testing.T) {
+	src := `Table s.t {
+  id int [pk]
+  @postgres: partition by RANGE (a)
+  @postgres: partition by LIST (b)
+}
+`
+	_, err := parse(t, false, src)
+	if err == nil || !strings.Contains(err.Error(), "duplicate") {
+		t.Fatalf("want duplicate error, got %v", err)
+	}
+	if !strings.Contains(err.Error(), "line 4") {
+		t.Errorf("error not line-numbered: %v", err)
+	}
+}
+
+func TestDirectives_MalformedErrors(t *testing.T) {
+	cases := map[string]string{
+		"no colon":       "@postgres partition by x",
+		"empty args":     "@postgres:",
+		"bad namespace":  "@Postgres: partition by x",
+		"numeric prefix": "@1x: foo",
+	}
+	for name, line := range cases {
+		t.Run(name, func(t *testing.T) {
+			src := "Table s.t {\n  id int [pk]\n  " + line + "\n}\n"
+			_, err := parse(t, false, src)
+			if err == nil {
+				t.Fatalf("want error for %q", line)
+			}
+			if !strings.Contains(err.Error(), "line 3") {
+				t.Errorf("error not line-numbered: %v", err)
+			}
+		})
+	}
+}
+
+func TestDirectives_UnknownPreservedNonStrict(t *testing.T) {
+	src := `Table s.t {
+  id int [pk]
+  @postgres: frobnicate all the things
+  @clickhouse: engine MergeTree
+}
+`
+	db, err := parse(t, false, src)
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	tbl := firstTable(t, db)
+	if !models.HasDirective(tbl.Metadata, "postgres", "frobnicate") {
+		t.Error("unknown postgres key not preserved")
+	}
+	if !models.HasDirective(tbl.Metadata, "clickhouse", "engine") {
+		t.Error("unknown namespace not preserved")
+	}
+}
+
+func TestDirectives_StrictErrors(t *testing.T) {
+	src := `Table s.t {
+  id int [pk]
+  @postgres: frobnicate x
+}
+`
+	_, err := parse(t, true, src)
+	if err == nil || !strings.Contains(err.Error(), "strict mode") {
+		t.Fatalf("want strict-mode error, got %v", err)
+	}
+}
+
+func TestDirectives_UnknownColumnTargetErrors(t *testing.T) {
+	src := `Table s.t {
+  id int [pk]
+  @postgres(missing): identity always
+}
+`
+	_, err := parse(t, false, src)
+	if err == nil || !strings.Contains(err.Error(), "not found") {
+		t.Fatalf("want unknown-column error, got %v", err)
+	}
+}
+
+func TestDirectives_WrongLocationErrors(t *testing.T) {
+	// partition is table-only.
+	src := "@postgres: partition by RANGE (x)\n\nTable s.t {\n  id int [pk]\n}\n"
+	_, err := parse(t, false, src)
+	if err == nil || !strings.Contains(err.Error(), "not valid at database level") {
+		t.Fatalf("want location error, got %v", err)
+	}
+}

pkg/readers/dbml/reader.go

@@ -435,11 +435,14 @@ func (r *Reader) parseDBML(content string) (*models.Database, error) {
 	var inIndexes bool
 	var inTable bool
 	var columnSeq uint
+	var lastIndex *models.Index // most recent index in the current Indexes block
+	lineNo := 0
 
 	tableRegex := regexp.MustCompile(`^Table\s+(.+?)\s*{`)
 	refRegex := regexp.MustCompile(`^Ref:\s+(.+)`)
 
 	for scanner.Scan() {
+		lineNo++
 		line := strings.TrimSpace(scanner.Text())
 
 		// Skip empty lines and comments
@@ -447,6 +450,20 @@ func (r *Reader) parseDBML(content string) (*models.Database, error) {
 			continue
 		}
 
+		// Parse a dialect directive (@postgres:, @sqlite:, …). Handled before
+		// table/column/index parsing so directive lines are never mistaken for
+		// columns.
+		if strings.HasPrefix(line, "@") {
+			pd, err := parseDirectiveLine(line, lineNo)
+			if err != nil {
+				return nil, err
+			}
+			if err := r.attachDirective(pd, db, currentTable, inTable, inIndexes, lastIndex); err != nil {
+				return nil, err
+			}
+			continue
+		}
+
 		// Parse Table definition
 		if matches := tableRegex.FindStringSubmatch(line); matches != nil {
 			tableName := matches[1]
@@ -474,8 +491,10 @@ func (r *Reader) parseDBML(content string) (*models.Database, error) {
 			continue
 		}
 
-		// End of table definition
+		// End of table definition. Guarded by !inIndexes so the closing brace
-		if inTable && line == "}" {
+		// of an `indexes { }` block is not mistaken for the end of the table
+		// (which would drop any table-level content that follows it).
+		if inTable && !inIndexes && line == "}" {
 			if currentTable != nil && currentSchema != "" {
 				schemaMap[currentSchema].Tables = append(schemaMap[currentSchema].Tables, currentTable)
 				currentTable = nil
@@ -488,12 +507,14 @@ func (r *Reader) parseDBML(content string) (*models.Database, error) {
 		// Parse indexes section
 		if inTable && (strings.HasPrefix(line, "Indexes {") || strings.HasPrefix(line, "indexes {")) {
 			inIndexes = true
+			lastIndex = nil
 			continue
 		}
 
 		// End of indexes section
 		if inIndexes && line == "}" {
 			inIndexes = false
+			lastIndex = nil
 			continue
 		}
 
@@ -513,6 +534,7 @@ func (r *Reader) parseDBML(content string) (*models.Database, error) {
 			index := r.parseIndex(line, currentTable.Name, currentSchema)
 			if index != nil {
 				currentTable.Indexes[index.Name] = index
+				lastIndex = index
 			}
 			continue
 		}

pkg/readers/reader.go

@@ -28,6 +28,10 @@ type ReaderOptions struct {
 	// Prisma7 enables Prisma 7-specific handling for Prisma schemas.
 	Prisma7 bool
 
+	// StrictDirectives makes DBML dialect directives (@postgres:, @sqlite:, …)
+	// fail on an unknown namespace or key instead of preserving them silently.
+	StrictDirectives bool
+
 	// Additional options can be added here as needed
 	Metadata map[string]interface{}
 }

pkg/writers/dbml/README.md

@@ -137,6 +137,41 @@ indexes {
 }
 ```
 
+### Dialect directives
+
+Dialect directives stored on a model object's `Metadata` (namespace `postgres`,
+`sqlite`, …) are re-emitted verbatim, one line per directive, at the location
+they belong to:
+
+```dbml
+@postgres: search_path myapp
+
+Table myapp.events {
+  id bigint [pk]
+  created_at timestamp [not null]
+  @postgres(id): identity always
+  @postgres: partition by RANGE (created_at)
+  @sqlite: without rowid
+
+  indexes {
+    (created_at) [name: 'idx_events_created']
+    @postgres: with (fillfactor=90)
+  }
+}
+```
+
+| Emitted at | From |
+|------------|------|
+| Before the first table | `Database.Metadata` |
+| After a column line, as `@ns(col): …` | `Column.Metadata` |
+| After an index line, inside `indexes { }` | `Index.Metadata` |
+| After the `indexes` block, before `Note:` | `Table.Metadata` |
+
+Output is deterministic (ordered by namespace, then source line, then args), so a
+`DBML → model → DBML` round-trip is idempotent. See
+[`docs/DBML_DIRECTIVES.md`](../../../docs/DBML_DIRECTIVES.md) for the grammar and
+the list of directives the PostgreSQL and SQLite writers translate to SQL.
+
 ## Type Mapping
 
 | SQL Type | DBML Type |

pkg/writers/dbml/directives.go

@@ -0,0 +1,23 @@
+package dbml
+
+import (
+	"git.warky.dev/wdevs/relspecgo/pkg/models"
+)
+
+// directiveLines renders every dialect directive stored in meta back to its DBML
+// source form, one line per directive, each prefixed with indent. When target is
+// non-empty it is emitted as the "(column)" target, e.g.
+// "  @postgres(id): identity always". Order is deterministic (see
+// models.GetDirectives).
+func directiveLines(meta map[string]any, indent, target string) []string {
+	directives := models.GetDirectives(meta)
+	if len(directives) == 0 {
+		return nil
+	}
+
+	lines := make([]string, 0, len(directives))
+	for _, d := range directives {
+		lines = append(lines, indent+models.FormatDirectiveLine(d, target))
+	}
+	return lines
+}

pkg/writers/dbml/directives_test.go

@@ -0,0 +1,97 @@
+package dbml
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+
+	dbmlreader "git.warky.dev/wdevs/relspecgo/pkg/readers/dbml"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"git.warky.dev/wdevs/relspecgo/pkg/models"
+	"git.warky.dev/wdevs/relspecgo/pkg/readers"
+	"git.warky.dev/wdevs/relspecgo/pkg/writers"
+)
+
+const directiveSrc = `@postgres: search_path myapp
+
+Table myapp.events {
+  id bigint [pk]
+  created_at timestamp [not null]
+  @postgres(id): identity always
+  @postgres: partition by RANGE (created_at)
+  @postgres: tablespace fast_data
+  @sqlite: without rowid
+
+  indexes {
+    (created_at) [name: 'idx_events_created']
+    @postgres: with (fillfactor=90)
+  }
+}
+`
+
+func writeDBML(t *testing.T, db *models.Database) string {
+	t.Helper()
+	out := filepath.Join(t.TempDir(), "out.dbml")
+	require.NoError(t, NewWriter(&writers.WriterOptions{OutputPath: out}).WriteDatabase(db))
+	b, err := os.ReadFile(out)
+	require.NoError(t, err)
+	return string(b)
+}
+
+func readDBML(t *testing.T, src string) *models.Database {
+	t.Helper()
+	f := filepath.Join(t.TempDir(), "in.dbml")
+	require.NoError(t, os.WriteFile(f, []byte(src), 0o644))
+	db, err := dbmlreader.NewReader(&readers.ReaderOptions{FilePath: f}).ReadDatabase()
+	require.NoError(t, err)
+	return db
+}
+
+func collectDirectives(db *models.Database) map[string][]string {
+	got := map[string][]string{}
+	add := func(loc string, meta map[string]any) {
+		for _, d := range models.GetDirectives(meta) {
+			got[loc] = append(got[loc], models.FormatDirectiveLine(d, ""))
+		}
+	}
+	add("database", db.Metadata)
+	for _, s := range db.Schemas {
+		for _, tbl := range s.Tables {
+			add("table:"+tbl.Name, tbl.Metadata)
+			for _, c := range tbl.Columns {
+				add("column:"+c.Name, c.Metadata)
+			}
+			for _, i := range tbl.Indexes {
+				add("index:"+i.Name, i.Metadata)
+			}
+		}
+	}
+	return got
+}
+
+func TestDirectives_RoundTrip(t *testing.T) {
+	db1 := readDBML(t, directiveSrc)
+	out1 := writeDBML(t, db1)
+	db2 := readDBML(t, out1)
+	out2 := writeDBML(t, db2)
+
+	assert.Equal(t, out1, out2, "DBML directive output should be idempotent")
+	assert.Equal(t, collectDirectives(db1), collectDirectives(db2), "directives preserved through round-trip")
+
+	// Spot-check each location survived.
+	d := collectDirectives(db2)
+	assert.Contains(t, d["database"], "@postgres: search_path myapp")
+	assert.Contains(t, d["table:events"], "@postgres: partition by RANGE (created_at)")
+	assert.Contains(t, d["table:events"], "@sqlite: without rowid")
+	assert.Contains(t, d["column:id"], "@postgres: identity always")
+	assert.Contains(t, d["index:idx_events_created"], "@postgres: with (fillfactor=90)")
+}
+
+func TestDirectives_WriterEmitsColumnTarget(t *testing.T) {
+	db := readDBML(t, directiveSrc)
+	out := writeDBML(t, db)
+	assert.Contains(t, out, "@postgres(id): identity always")
+}

pkg/writers/dbml/writer.go

@@ -72,6 +72,14 @@ func (w *Writer) databaseToDBML(d *models.Database) string {
 		sb.WriteString("\n")
 	}
 
+	if dirLines := directiveLines(d.Metadata, "", ""); len(dirLines) > 0 {
+		for _, line := range dirLines {
+			sb.WriteString(line)
+			sb.WriteString("\n")
+		}
+		sb.WriteString("\n")
+	}
+
 	for _, schema := range d.Schemas {
 		sb.WriteString(w.schemaToDBML(schema))
 	}
@@ -146,6 +154,11 @@ func (w *Writer) tableToDBML(t *models.Table) string {
 			fmt.Fprintf(&sb, " // %s", column.Comment)
 		}
 		sb.WriteString("\n")
+
+		for _, line := range directiveLines(column.Metadata, "  ", column.Name) {
+			sb.WriteString(line)
+			sb.WriteString("\n")
+		}
 	}
 
 	if len(t.Indexes) > 0 {
@@ -167,10 +180,20 @@ func (w *Writer) tableToDBML(t *models.Table) string {
 				fmt.Fprintf(&sb, " [%s]", strings.Join(indexAttrs, ", "))
 			}
 			sb.WriteString("\n")
+
+			for _, line := range directiveLines(index.Metadata, "    ", "") {
+				sb.WriteString(line)
+				sb.WriteString("\n")
+			}
 		}
 		sb.WriteString("  }\n")
 	}
 
+	for _, line := range directiveLines(t.Metadata, "  ", "") {
+		sb.WriteString(line)
+		sb.WriteString("\n")
+	}
+
 	note := strings.TrimSpace(t.Description + " " + t.Comment)
 	if note != "" {
 		fmt.Fprintf(&sb, "\n  Note: '%s'\n", note)

pkg/writers/pgsql/README.md

@@ -172,6 +172,26 @@ When `include_audit` is enabled, adds:
 - Concurrent index creation (`CREATE INDEX CONCURRENTLY`) via `Index.Concurrent`
 - Check constraints with expressions
 - Extension types and indexes: PostGIS, pgvector, citext, hstore, ltree (see below)
+- DBML dialect directives (`@postgres:` — see below)
+
+### DBML dialect directives
+
+`@postgres:` directives carried on a model object's `Metadata` (typically from a
+DBML source file) are translated to SQL:
+
+| Directive | Location | Emitted |
+|-----------|----------|---------|
+| `@postgres: partition by …` | table | `PARTITION BY …` on `CREATE TABLE` |
+| `@postgres: inherits …` | table | `INHERITS (…)` |
+| `@postgres: with (…)` | table, index | `WITH (…)` (on an index, overrides the comment-derived `WITH`) |
+| `@postgres: tablespace …` | table, index | `TABLESPACE …` |
+| `@postgres(col): storage …` | column | `STORAGE …` |
+| `@postgres(col): compression …` | column | `COMPRESSION …` |
+| `@postgres(col): identity always` / `identity by default` | column | `GENERATED ALWAYS/BY DEFAULT AS IDENTITY` |
+
+Directives for other dialects (`@sqlite:` …) are ignored. With
+`WriterOptions.StrictDirectives` (CLI `--strict-directives`) an untranslatable
+`@postgres:` key is an error. Full reference: [`docs/DBML_DIRECTIVES.md`](../../../docs/DBML_DIRECTIVES.md).
 
 ## Data Types
 

pkg/writers/pgsql/directives.go

@@ -0,0 +1,184 @@
+package pgsql
+
+import (
+	"fmt"
+	"strings"
+
+	"git.warky.dev/wdevs/relspecgo/pkg/models"
+)
+
+// directiveNamespace is the dialect namespace this writer consumes. Directives
+// for other namespaces (e.g. "sqlite") are ignored and never emitted as SQL.
+const directiveNamespace = "postgres"
+
+// pgHandledDirectives maps a directive location to the set of postgres keys this
+// writer knows how to translate. In strict mode an unknown key for this
+// namespace at a supported location is a hard error.
+var pgHandledDirectives = map[string]map[string]bool{
+	models.DirectiveLocationTable:  {"partition": true, "inherits": true, "with": true, "tablespace": true},
+	models.DirectiveLocationColumn: {"storage": true, "compression": true, "identity": true},
+	models.DirectiveLocationIndex:  {"with": true, "tablespace": true},
+}
+
+// checkDirectives validates postgres directives across a schema when strict mode
+// is enabled. It returns an error for any postgres directive whose key this
+// writer cannot translate. With strict mode off it is a no-op.
+func (w *Writer) checkDirectives(schema *models.Schema) error {
+	if w.options == nil || !w.options.StrictDirectives {
+		return nil
+	}
+	for _, table := range schema.Tables {
+		if err := checkObjectDirectives(table.Metadata, models.DirectiveLocationTable, table.Name); err != nil {
+			return err
+		}
+		for _, col := range table.Columns {
+			if err := checkObjectDirectives(col.Metadata, models.DirectiveLocationColumn, table.Name+"."+col.Name); err != nil {
+				return err
+			}
+		}
+		for _, idx := range table.Indexes {
+			if err := checkObjectDirectives(idx.Metadata, models.DirectiveLocationIndex, idx.Name); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+func checkObjectDirectives(meta map[string]any, location, owner string) error {
+	for _, d := range models.DirectivesForNamespace(meta, directiveNamespace) {
+		if !pgHandledDirectives[location][d.Key] {
+			return fmt.Errorf("pgsql: %s: unsupported @postgres directive %q at %s level (strict mode)", owner, d.Key, location)
+		}
+	}
+	return nil
+}
+
+// upperLeadingClause upcases a known leading keyword phrase in a directive
+// argument so the emitted SQL reads conventionally. Identifiers that follow are
+// left untouched.
+func upperLeadingClause(args, lowerPrefix, upperPrefix string) string {
+	args = strings.TrimSpace(args)
+	if strings.HasPrefix(strings.ToLower(args), lowerPrefix) {
+		return upperPrefix + args[len(lowerPrefix):]
+	}
+	return args
+}
+
+// pgTableDirectiveSuffix returns the clause appended after the closing ")" of a
+// CREATE TABLE statement, e.g. " PARTITION BY RANGE (created_at) TABLESPACE fast".
+func pgTableDirectiveSuffix(table *models.Table) string {
+	directives := models.DirectivesForNamespace(table.Metadata, directiveNamespace)
+	if len(directives) == 0 {
+		return ""
+	}
+
+	byKey := firstByKey(directives)
+
+	var parts []string
+	if d, ok := byKey["partition"]; ok {
+		parts = append(parts, upperLeadingClause(d.Args, "partition by", "PARTITION BY"))
+	}
+	if d, ok := byKey["inherits"]; ok {
+		parts = append(parts, upperLeadingClause(d.Args, "inherits", "INHERITS"))
+	}
+	if d, ok := byKey["with"]; ok {
+		parts = append(parts, upperLeadingClause(d.Args, "with", "WITH"))
+	}
+	if d, ok := byKey["tablespace"]; ok {
+		parts = append(parts, upperLeadingClause(d.Args, "tablespace", "TABLESPACE"))
+	}
+	if len(parts) == 0 {
+		return ""
+	}
+	return " " + strings.Join(parts, " ")
+}
+
+// pgColumnDirectiveSuffix returns the clause appended to a column definition,
+// e.g. " STORAGE PLAIN" or " GENERATED ALWAYS AS IDENTITY".
+func pgColumnDirectiveSuffix(col *models.Column) string {
+	directives := models.DirectivesForNamespace(col.Metadata, directiveNamespace)
+	if len(directives) == 0 {
+		return ""
+	}
+	byKey := firstByKey(directives)
+
+	var parts []string
+	if d, ok := byKey["storage"]; ok {
+		parts = append(parts, upperLeadingClause(d.Args, "storage", "STORAGE"))
+	}
+	if d, ok := byKey["compression"]; ok {
+		parts = append(parts, upperLeadingClause(d.Args, "compression", "COMPRESSION"))
+	}
+	if d, ok := byKey["identity"]; ok {
+		parts = append(parts, identityClause(d.Args))
+	}
+	if len(parts) == 0 {
+		return ""
+	}
+	return " " + strings.Join(parts, " ")
+}
+
+// identityClause maps the two documented identity forms to standard SQL,
+// falling back to a verbatim (upcased-keyword) rendering.
+func identityClause(args string) string {
+	switch strings.ToLower(strings.Join(strings.Fields(args), " ")) {
+	case "identity always":
+		return "GENERATED ALWAYS AS IDENTITY"
+	case "identity default", "identity by default":
+		return "GENERATED BY DEFAULT AS IDENTITY"
+	default:
+		return upperLeadingClause(args, "identity", "IDENTITY")
+	}
+}
+
+// pgIndexDirectiveWith returns the parenthesised storage-parameter list from an
+// @postgres: with (...) index directive, e.g. "fillfactor=90", or "".
+func pgIndexDirectiveWith(index *models.Index) string {
+	for _, d := range models.DirectivesForNamespace(index.Metadata, directiveNamespace) {
+		if d.Key != "with" {
+			continue
+		}
+		inner := d.Args
+		if i := strings.Index(inner, "("); i >= 0 {
+			if j := strings.LastIndex(inner, ")"); j > i {
+				return strings.TrimSpace(inner[i+1 : j])
+			}
+		}
+		return strings.TrimSpace(strings.TrimPrefix(strings.ToLower(inner), "with"))
+	}
+	return ""
+}
+
+// pgIndexWithParams returns the storage-parameter list to use for an index,
+// preferring an @postgres: with (...) directive over the given fallback (e.g.
+// one derived from the index comment).
+func pgIndexWithParams(index *models.Index, fallback string) string {
+	if p := pgIndexDirectiveWith(index); p != "" {
+		return p
+	}
+	return fallback
+}
+
+// pgIndexDirectiveTablespace returns the tablespace name from an
+// @postgres: tablespace <name> index directive, or "".
+func pgIndexDirectiveTablespace(index *models.Index) string {
+	for _, d := range models.DirectivesForNamespace(index.Metadata, directiveNamespace) {
+		if d.Key == "tablespace" {
+			return strings.TrimSpace(strings.TrimPrefix(strings.ToLower(d.Args), "tablespace"))
+		}
+	}
+	return ""
+}
+
+// firstByKey indexes directives by key, keeping the first occurrence (the
+// documented postgres keys used here are all singletons).
+func firstByKey(directives []models.Directive) map[string]models.Directive {
+	byKey := make(map[string]models.Directive, len(directives))
+	for _, d := range directives {
+		if _, exists := byKey[d.Key]; !exists {
+			byKey[d.Key] = d
+		}
+	}
+	return byKey
+}

pkg/writers/pgsql/directives_test.go

@@ -0,0 +1,107 @@
+package pgsql
+
+import (
+	"bytes"
+	"strings"
+	"testing"
+
+	"git.warky.dev/wdevs/relspecgo/pkg/models"
+	"git.warky.dev/wdevs/relspecgo/pkg/writers"
+)
+
+func directiveTestDB(t *testing.T) *models.Database {
+	t.Helper()
+	db := models.InitDatabase("testdb")
+	schema := models.InitSchema("public")
+	table := models.InitTable("events", "public")
+
+	id := models.InitColumn("id", "events", "public")
+	id.Type = "bigint"
+	id.IsPrimaryKey = true
+	id.NotNull = true
+	models.AddDirective(id.Metadata, models.Directive{Namespace: "postgres", Args: "identity always"})
+	models.AddDirective(id.Metadata, models.Directive{Namespace: "sqlite", Args: "collate NOCASE"})
+	table.Columns["id"] = id
+
+	created := models.InitColumn("created_at", "events", "public")
+	created.Type = "timestamp"
+	created.NotNull = true
+	table.Columns["created_at"] = created
+
+	models.AddDirective(table.Metadata, models.Directive{Namespace: "postgres", Args: "partition by RANGE (created_at)"})
+	models.AddDirective(table.Metadata, models.Directive{Namespace: "postgres", Args: "tablespace fast_data"})
+	models.AddDirective(table.Metadata, models.Directive{Namespace: "sqlite", Args: "without rowid"})
+
+	idx := models.InitIndex("idx_events_created", "events", "public")
+	idx.Columns = []string{"created_at"}
+	models.AddDirective(idx.Metadata, models.Directive{Namespace: "postgres", Args: "with (fillfactor=90)"})
+	models.AddDirective(idx.Metadata, models.Directive{Namespace: "postgres", Args: "tablespace idx_space"})
+	table.Indexes["idx_events_created"] = idx
+
+	schema.Tables = append(schema.Tables, table)
+	db.Schemas = append(db.Schemas, schema)
+	return db
+}
+
+func TestPgDirectives_WriteDatabasePath(t *testing.T) {
+	var buf bytes.Buffer
+	w := NewWriter(&writers.WriterOptions{})
+	w.writer = &buf
+	if err := w.WriteDatabase(directiveTestDB(t)); err != nil {
+		t.Fatalf("WriteDatabase: %v", err)
+	}
+	out := buf.String()
+
+	for _, want := range []string{
+		") PARTITION BY RANGE (created_at) TABLESPACE fast_data",
+		"GENERATED ALWAYS AS IDENTITY",
+		"WITH (fillfactor=90) TABLESPACE idx_space",
+	} {
+		if !strings.Contains(out, want) {
+			t.Errorf("missing %q in:\n%s", want, out)
+		}
+	}
+	// sqlite directives must never reach PG output.
+	if strings.Contains(out, "WITHOUT ROWID") || strings.Contains(strings.ToUpper(out), "COLLATE NOCASE") {
+		t.Errorf("sqlite directive leaked into PG output:\n%s", out)
+	}
+}
+
+func TestPgDirectives_WriteSchemaPath(t *testing.T) {
+	var buf bytes.Buffer
+	w := NewWriter(&writers.WriterOptions{})
+	w.writer = &buf
+	if err := w.WriteSchema(directiveTestDB(t).Schemas[0]); err != nil {
+		t.Fatalf("WriteSchema: %v", err)
+	}
+	out := buf.String()
+	if !strings.Contains(out, ") PARTITION BY RANGE (created_at) TABLESPACE fast_data;") {
+		t.Errorf("table suffix missing from WriteSchema path:\n%s", out)
+	}
+	if !strings.Contains(out, "WITH (fillfactor=90) TABLESPACE idx_space") {
+		t.Errorf("index clauses missing from WriteSchema path:\n%s", out)
+	}
+}
+
+func TestPgDirectives_StrictUnknownKeyErrors(t *testing.T) {
+	db := directiveTestDB(t)
+	models.AddDirective(db.Schemas[0].Tables[0].Metadata, models.Directive{Namespace: "postgres", Args: "frobnicate x"})
+
+	var buf bytes.Buffer
+	w := NewWriter(&writers.WriterOptions{StrictDirectives: true})
+	w.writer = &buf
+	err := w.WriteSchema(db.Schemas[0])
+	if err == nil || !strings.Contains(err.Error(), "frobnicate") {
+		t.Fatalf("want strict error for unknown postgres key, got %v", err)
+	}
+}
+
+func TestPgDirectives_StrictIgnoresOtherNamespaces(t *testing.T) {
+	// sqlite directives are present but must not trip PG strict mode.
+	var buf bytes.Buffer
+	w := NewWriter(&writers.WriterOptions{StrictDirectives: true})
+	w.writer = &buf
+	if err := w.WriteSchema(directiveTestDB(t).Schemas[0]); err != nil {
+		t.Fatalf("strict mode should ignore non-postgres directives, got %v", err)
+	}
+}

pkg/writers/pgsql/writer.go

@@ -143,6 +143,10 @@ func (w *Writer) GenerateDatabaseStatements(db *models.Database) ([]string, erro
 func (w *Writer) GenerateSchemaStatements(schema *models.Schema) ([]string, error) {
 	statements := []string{}
 
+	if err := w.checkDirectives(schema); err != nil {
+		return nil, err
+	}
+
 	// Phase 1: Create schema (skip entirely when flattening)
 	if schema.Name != "public" && !w.options.FlattenSchema {
 		statements = append(statements, fmt.Sprintf("-- Schema: %s", schema.Name))
@@ -277,17 +281,22 @@ func (w *Writer) GenerateSchemaStatements(schema *models.Schema) ([]string, erro
 			columnExprs := buildIndexColumnExpressions(table, index, indexType)
 
 			withClause := ""
-			if params := indexStorageParameters(index.Comment); params != "" {
+			if params := pgIndexWithParams(index, indexStorageParameters(index.Comment)); params != "" {
 				withClause = fmt.Sprintf(" WITH (%s)", params)
 			}
 
+			tablespaceClause := ""
+			if ts := pgIndexDirectiveTablespace(index); ts != "" {
+				tablespaceClause = fmt.Sprintf(" TABLESPACE %s", ts)
+			}
+
 			whereClause := ""
 			if index.Where != "" {
 				whereClause = fmt.Sprintf(" WHERE %s", index.Where)
 			}
 
-			stmt := fmt.Sprintf("CREATE %sINDEX IF NOT EXISTS %s ON %s USING %s (%s)%s%s",
+			stmt := fmt.Sprintf("CREATE %sINDEX IF NOT EXISTS %s ON %s USING %s (%s)%s%s%s",
-				uniqueStr, quoteIdentifier(index.Name), w.qualTable(schema.SQLName(), table.SQLName()), indexType, strings.Join(columnExprs, ", "), withClause, whereClause)
+				uniqueStr, quoteIdentifier(index.Name), w.qualTable(schema.SQLName(), table.SQLName()), indexType, strings.Join(columnExprs, ", "), withClause, tablespaceClause, whereClause)
 			statements = append(statements, stmt)
 		}
 	}
@@ -581,8 +590,9 @@ func (w *Writer) generateCreateTableStatement(schema *models.Schema, table *mode
 		columnDefs = append(columnDefs, "  "+def)
 	}
 
-	stmt := fmt.Sprintf("CREATE TABLE IF NOT EXISTS %s (\n%s\n)",
+	stmt := fmt.Sprintf("CREATE TABLE IF NOT EXISTS %s (\n%s\n)%s",
-		w.qualTable(schema.SQLName(), table.SQLName()), strings.Join(columnDefs, ",\n"))
+		w.qualTable(schema.SQLName(), table.SQLName()), strings.Join(columnDefs, ",\n"),
+		pgTableDirectiveSuffix(table))
 	statements = append(statements, stmt)
 
 	return statements, nil
@@ -611,7 +621,7 @@ func (w *Writer) generateColumnDefinition(col *models.Column) string {
 		}
 	}
 
-	return strings.Join(parts, " ")
+	return strings.Join(parts, " ") + pgColumnDirectiveSuffix(col)
 }
 
 func effectiveColumnSQLType(col *models.Column) string {
@@ -678,6 +688,10 @@ func (w *Writer) WriteSchema(schema *models.Schema) error {
 		w.writer = os.Stdout
 	}
 
+	if err := w.checkDirectives(schema); err != nil {
+		return err
+	}
+
 	// Phase 1: Create schema (priority 1)
 	if err := w.writeCreateSchema(schema); err != nil {
 		return err
@@ -884,7 +898,7 @@ func (w *Writer) writeCreateTables(schema *models.Schema) error {
 		}
 
 		fmt.Fprintf(w.writer, "%s\n", strings.Join(columnDefs, ",\n"))
-		fmt.Fprintf(w.writer, ");\n\n")
+		fmt.Fprintf(w.writer, ")%s;\n\n", pgTableDirectiveSuffix(table))
 	}
 
 	return nil
@@ -1079,10 +1093,15 @@ func (w *Writer) writeIndexes(schema *models.Schema) error {
 			}
 
 			withClause := ""
-			if params := indexStorageParameters(index.Comment); params != "" {
+			if params := pgIndexWithParams(index, indexStorageParameters(index.Comment)); params != "" {
 				withClause = fmt.Sprintf(" WITH (%s)", params)
 			}
 
+			tablespaceClause := ""
+			if ts := pgIndexDirectiveTablespace(index); ts != "" {
+				tablespaceClause = fmt.Sprintf(" TABLESPACE %s", ts)
+			}
+
 			whereClause := ""
 			if index.Where != "" {
 				whereClause = fmt.Sprintf(" WHERE %s", index.Where)
@@ -1095,8 +1114,8 @@ func (w *Writer) writeIndexes(schema *models.Schema) error {
 
 			fmt.Fprintf(w.writer, "CREATE %sINDEX %sIF NOT EXISTS %s\n",
 				unique, concurrently, indexName)
-			fmt.Fprintf(w.writer, "  ON %s USING %s (%s)%s%s;\n\n",
+			fmt.Fprintf(w.writer, "  ON %s USING %s (%s)%s%s%s;\n\n",
-				w.qualTable(schema.SQLName(), table.SQLName()), indexType, strings.Join(columnExprs, ", "), withClause, whereClause)
+				w.qualTable(schema.SQLName(), table.SQLName()), indexType, strings.Join(columnExprs, ", "), withClause, tablespaceClause, whereClause)
 		}
 	}
 

pkg/writers/sqlite/README.md

@@ -118,6 +118,21 @@ CREATE TABLE "posts" (
 - **Check Constraints**: Generated as comments (should be added to CREATE TABLE manually)
 - **Indexes**: Generated without PostgreSQL-specific features (no GIN, GiST, operator classes)
 
+## DBML dialect directives
+
+`@sqlite:` directives carried on a model object's `Metadata` (typically from a
+DBML source file) are translated to SQL:
+
+| Directive | Location | Emitted |
+|-----------|----------|---------|
+| `@sqlite: without rowid` | table | `WITHOUT ROWID` table option |
+| `@sqlite: strict` | table | `STRICT` table option (after `WITHOUT ROWID`) |
+| `@sqlite(col): collate …` | column | ` COLLATE …` in the column definition |
+
+Directives for other dialects (`@postgres:` …) are ignored. With
+`WriterOptions.StrictDirectives` (CLI `--strict-directives`) an untranslatable
+`@sqlite:` key is an error. Full reference: [`docs/DBML_DIRECTIVES.md`](../../../docs/DBML_DIRECTIVES.md).
+
 ## Output Structure
 
 Generated SQL follows this order:

pkg/writers/sqlite/directives.go

@@ -0,0 +1,84 @@
+package sqlite
+
+import (
+	"fmt"
+	"strings"
+
+	"git.warky.dev/wdevs/relspecgo/pkg/models"
+)
+
+// directiveNamespace is the dialect namespace this writer consumes. Directives
+// for other namespaces (e.g. "postgres") are ignored and never emitted as SQL.
+const directiveNamespace = "sqlite"
+
+// sqliteHandledDirectives maps a directive location to the set of sqlite keys
+// this writer knows how to translate. In strict mode an unknown key for this
+// namespace at a supported location is a hard error.
+var sqliteHandledDirectives = map[string]map[string]bool{
+	models.DirectiveLocationTable:  {"without": true, "strict": true},
+	models.DirectiveLocationColumn: {"collate": true},
+}
+
+// checkDirectives validates sqlite directives across a schema when strict mode
+// is enabled. With strict mode off it is a no-op.
+func (w *Writer) checkDirectives(schema *models.Schema) error {
+	if w.options == nil || !w.options.StrictDirectives {
+		return nil
+	}
+	for _, table := range schema.Tables {
+		if err := checkObjectDirectives(table.Metadata, models.DirectiveLocationTable, table.Name); err != nil {
+			return err
+		}
+		for _, col := range table.Columns {
+			if err := checkObjectDirectives(col.Metadata, models.DirectiveLocationColumn, table.Name+"."+col.Name); err != nil {
+				return err
+			}
+		}
+		for _, idx := range table.Indexes {
+			if err := checkObjectDirectives(idx.Metadata, models.DirectiveLocationIndex, idx.Name); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+func checkObjectDirectives(meta map[string]any, location, owner string) error {
+	for _, d := range models.DirectivesForNamespace(meta, directiveNamespace) {
+		if !sqliteHandledDirectives[location][d.Key] {
+			return fmt.Errorf("sqlite: %s: unsupported @sqlite directive %q at %s level (strict mode)", owner, d.Key, location)
+		}
+	}
+	return nil
+}
+
+// sqliteTableOptions returns the trailing table-option clause for a CREATE TABLE
+// statement, e.g. "WITHOUT ROWID, STRICT". WITHOUT ROWID is emitted before
+// STRICT, matching SQLite's own grammar ordering.
+func sqliteTableOptions(table *models.Table) string {
+	var opts []string
+	if models.HasDirective(table.Metadata, directiveNamespace, "without") {
+		opts = append(opts, "WITHOUT ROWID")
+	}
+	if models.HasDirective(table.Metadata, directiveNamespace, "strict") {
+		opts = append(opts, "STRICT")
+	}
+	return strings.Join(opts, ", ")
+}
+
+// sqliteColumnCollate returns a " COLLATE <name>" clause for a column carrying an
+// @sqlite(col): collate <name> directive, or "".
+func sqliteColumnCollate(col *models.Column) string {
+	for _, d := range models.DirectivesForNamespace(col.Metadata, directiveNamespace) {
+		if d.Key != "collate" {
+			continue
+		}
+		name := strings.TrimSpace(strings.TrimPrefix(strings.TrimSpace(d.Args), "collate"))
+		name = strings.TrimSpace(name)
+		if name == "" {
+			return ""
+		}
+		return " COLLATE " + name
+	}
+	return ""
+}

pkg/writers/sqlite/directives_test.go

@@ -0,0 +1,82 @@
+package sqlite
+
+import (
+	"bytes"
+	"strings"
+	"testing"
+
+	"git.warky.dev/wdevs/relspecgo/pkg/models"
+	"git.warky.dev/wdevs/relspecgo/pkg/writers"
+)
+
+func sqliteDirectiveDB(t *testing.T) *models.Database {
+	t.Helper()
+	db := models.InitDatabase("testdb")
+	schema := models.InitSchema("public")
+	table := models.InitTable("events", "public")
+
+	id := models.InitColumn("id", "events", "public")
+	id.Type = "bigint"
+	id.IsPrimaryKey = true
+	id.NotNull = true
+	table.Columns["id"] = id
+
+	name := models.InitColumn("name", "events", "public")
+	name.Type = "varchar(200)"
+	name.NotNull = true
+	models.AddDirective(name.Metadata, models.Directive{Namespace: "sqlite", Args: "collate NOCASE"})
+	// A postgres directive on the same column must be ignored by the sqlite writer.
+	models.AddDirective(name.Metadata, models.Directive{Namespace: "postgres", Args: "storage plain"})
+	table.Columns["name"] = name
+
+	models.AddDirective(table.Metadata, models.Directive{Namespace: "sqlite", Args: "without rowid"})
+	models.AddDirective(table.Metadata, models.Directive{Namespace: "sqlite", Args: "strict"})
+	models.AddDirective(table.Metadata, models.Directive{Namespace: "postgres", Args: "partition by RANGE (id)"})
+
+	schema.Tables = append(schema.Tables, table)
+	db.Schemas = append(db.Schemas, schema)
+	return db
+}
+
+func TestSqliteDirectives_TableOptionsAndCollate(t *testing.T) {
+	var buf bytes.Buffer
+	w := NewWriter(&writers.WriterOptions{})
+	w.writer = &buf
+	if err := w.WriteDatabase(sqliteDirectiveDB(t)); err != nil {
+		t.Fatalf("WriteDatabase: %v", err)
+	}
+	out := buf.String()
+
+	if !strings.Contains(out, ") WITHOUT ROWID, STRICT;") {
+		t.Errorf("missing table options clause:\n%s", out)
+	}
+	if !strings.Contains(out, `"name" TEXT COLLATE NOCASE NOT NULL`) {
+		t.Errorf("missing column COLLATE clause:\n%s", out)
+	}
+	// postgres directives must never reach sqlite output.
+	if strings.Contains(strings.ToUpper(out), "PARTITION BY") || strings.Contains(strings.ToUpper(out), "STORAGE PLAIN") {
+		t.Errorf("postgres directive leaked into sqlite output:\n%s", out)
+	}
+}
+
+func TestSqliteDirectives_StrictUnknownKeyErrors(t *testing.T) {
+	db := sqliteDirectiveDB(t)
+	models.AddDirective(db.Schemas[0].Tables[0].Metadata, models.Directive{Namespace: "sqlite", Args: "frobnicate x"})
+
+	var buf bytes.Buffer
+	w := NewWriter(&writers.WriterOptions{StrictDirectives: true})
+	w.writer = &buf
+	err := w.WriteDatabase(db)
+	if err == nil || !strings.Contains(err.Error(), "frobnicate") {
+		t.Fatalf("want strict error for unknown sqlite key, got %v", err)
+	}
+}
+
+func TestSqliteDirectives_StrictIgnoresPostgres(t *testing.T) {
+	var buf bytes.Buffer
+	w := NewWriter(&writers.WriterOptions{StrictDirectives: true})
+	w.writer = &buf
+	if err := w.WriteDatabase(sqliteDirectiveDB(t)); err != nil {
+		t.Fatalf("strict mode should ignore postgres directives, got %v", err)
+	}
+}

pkg/writers/sqlite/template_functions.go

@@ -25,6 +25,7 @@ func GetTemplateFuncs(opts *writers.WriterOptions) template.FuncMap {
 		"join":           strings.Join,
 		"lower":          strings.ToLower,
 		"upper":          strings.ToUpper,
+		"column_collate": sqliteColumnCollate,
 	}
 }
 

pkg/writers/sqlite/templates.go

@@ -45,6 +45,7 @@ type TableTemplateData struct {
 	Columns      []*models.Column
 	PrimaryKey   *models.Constraint
 	ForeignKeys  []ForeignKeyTemplateData
+	TableOptions string
 }
 
 // ForeignKeyTemplateData contains data for an inline FOREIGN KEY clause
@@ -193,6 +194,7 @@ func BuildTableTemplateData(schema string, table *models.Table) TableTemplateDat
 		Columns:      columns,
 		PrimaryKey:   pk,
 		ForeignKeys:  fks,
+		TableOptions: sqliteTableOptions(table),
 	}
 }
 

pkg/writers/sqlite/templates/create_table.tmpl

@@ -1,7 +1,7 @@
 CREATE TABLE {{quote_ident (qualified_table_name .Schema .Name)}} (
 {{- $hasAutoIncrement := false}}
 {{- range $i, $col := .Columns}}{{if $i}},{{end}}
-    {{quote_ident $col.Name}} {{map_type $col.Type}}{{if is_autoincrement $col}}{{$hasAutoIncrement = true}} PRIMARY KEY AUTOINCREMENT{{else}}{{if $col.NotNull}} NOT NULL{{end}}{{if ne (format_default $col) ""}} DEFAULT {{format_default $col}}{{end}}{{end}}
+    {{quote_ident $col.Name}} {{map_type $col.Type}}{{column_collate $col}}{{if is_autoincrement $col}}{{$hasAutoIncrement = true}} PRIMARY KEY AUTOINCREMENT{{else}}{{if $col.NotNull}} NOT NULL{{end}}{{if ne (format_default $col) ""}} DEFAULT {{format_default $col}}{{end}}{{end}}
 {{- end}}
 {{- if and .PrimaryKey (not $hasAutoIncrement)}}{{if gt (len .Columns) 0}},{{end}}
     PRIMARY KEY ({{range $i, $colName := .PrimaryKey.Columns}}{{if $i}}, {{end}}{{quote_ident $colName}}{{end}})
@@ -9,4 +9,4 @@ CREATE TABLE {{quote_ident (qualified_table_name .Schema .Name)}} (
 {{- range .ForeignKeys}},
     FOREIGN KEY ({{range $i, $col := .Columns}}{{if $i}}, {{end}}{{quote_ident $col}}{{end}}) REFERENCES {{quote_ident (qualified_table_name .ForeignSchema .ForeignTable)}} ({{range $i, $col := .ForeignColumns}}{{if $i}}, {{end}}{{quote_ident $col}}{{end}}){{if .OnDelete}} ON DELETE {{.OnDelete}}{{end}}{{if .OnUpdate}} ON UPDATE {{.OnUpdate}}{{end}}
 {{- end}}
-);
+){{if .TableOptions}} {{.TableOptions}}{{end}};

pkg/writers/sqlite/writer.go

@@ -186,6 +186,10 @@ func tableSchemaName(schema string) string {
 func (w *Writer) WriteSchema(schema *models.Schema) error {
 	tableSchema := tableSchemaName(schema.Name)
 
+	if err := w.checkDirectives(schema); err != nil {
+		return err
+	}
+
 	// SQLite doesn't have schemas, so we just write a comment (skip for the
 	// default schema, since its tables aren't actually being prefixed)
 	if tableSchema != "" {

pkg/writers/writer.go

@@ -86,6 +86,10 @@ type WriterOptions struct {
 	// Prisma7 enables Prisma 7-specific output for Prisma writers.
 	Prisma7 bool
 
+	// StrictDirectives makes dialect directive translation fail on an
+	// unsupported key for the writer's own namespace instead of skipping it.
+	StrictDirectives bool
+
 	// ContinueOnError instructs SQL writers to prepend `\set ON_ERROR_STOP off`
 	// to their output so that psql continues past errors instead of stopping.
 	ContinueOnError bool