feat(job): declarative YAML job files for named relspec workflows

Add `relspec job list` and `relspec job run <name>` driven by YAML job
manifests (relspec.yml / relspec.<name>.yml), so multi-file merge and
conversion workflows can be expressed declaratively instead of as long
shell command lines.

v1 contract (see docs/JOB_FILES.md):
- `command` is a closed allow-list (convert, merge, scripts-list); no
  field accepts a shell string or executable path.
- Deterministic discovery: default file first, then named files sorted
  lexically; all files merged into one namespace; duplicate job names
  across files are a hard error.
- Every path resolves relative to the job file's directory; absolute,
  home-relative and directory-escaping paths are rejected at validation.
- Database credentials referenced by env-var name via `conn_env:`;
  connection strings are never stored and are redacted from logs/plan.
- Full validation (version, unknown fields, command/format, per-command
  input/output shape, path traversal, depends_on targets, dependency
  cycles) runs before anything is read, written or executed; per-job
  pre-flight then checks input existence, script dirs, env vars and the
  output overwrite policy for the whole plan.
- `depends_on` closure runs in deterministic topological order;
  `--no-deps` runs only the named job.
- `--dry-run` (alias `--plan`) prints the resolved plan and exits 0
  without touching inputs, outputs or databases.
- A failing job propagates the underlying non-zero exit status, logs
  FAILED (never OK), and writes no success marker.

pkg/jobs is side-effect free (discovery/parse/validate/plan only);
execution adapters live in cmd/relspec/job.go. Includes unit tests for
discovery, validation, planning and path safety, plus CLI tests for
end-to-end convert/merge, scripts-list across multiple directories,
dry-run, dependency chains, exit-code propagation and log redaction.

Deferred: live `scripts execute` from jobs, split/inspect/diff/templ
commands, job-to-job output wiring, log rotation/retention.

Refs #20

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

README.md

@@ -106,6 +106,49 @@ Modes: `database` (default) · `schema` · `table` · `script`
 
 Template functions: string utils (`toCamelCase`, `toSnakeCase`, `pluralize`, …), type converters (`sqlToGo`, `sqlToTypeScript`, …), filters, loop helpers, safe access.
 
+### `job` — Declarative job files
+
+Run named jobs from a `relspec.yml` manifest instead of repeating long command lines.
+
+```bash
+# List jobs discovered in ./relspec.yml and ./relspec.<name>.yml (deterministic)
+relspec job list
+
+# Validate and print the plan without running anything
+relspec job run build-schema --plan
+
+# Run a job (and its declared dependencies)
+relspec job run build-schema
+```
+
+```yaml
+# relspec.yml
+version: 1
+jobs:
+  build-schema:
+    command: convert            # closed allow-list: convert | merge | scripts-list
+    description: Merge the DBML sources and emit PostgreSQL DDL
+    inputs:
+      - path: schema/core.dbml
+        format: dbml
+      - path: schema/tenant.dbml
+        format: dbml
+    output:
+      format: pgsql
+      path: build/schema.sql
+      overwrite: true
+    options:
+      flatten_schema: false
+    logfile: .relspec/log/build-schema.log
+```
+
+The job system is **not** a shell: `command` is a fixed enum, every path is
+resolved relative to the job file and may not escape it, and remote database
+credentials are referenced by environment-variable name (`conn_env:`) and
+redacted from logs. The whole plan — unknown commands/formats, duplicate job
+names, missing inputs, path traversal, dependency cycles — is validated before
+any job runs. See [docs/JOB_FILES.md](docs/JOB_FILES.md).
+
 ### `edit` — Interactive TUI editor
 
 ```bash

cmd/relspec/job.go

@@ -0,0 +1,567 @@
+package main
+
+import (
+	"fmt"
+	"io"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+	"time"
+
+	"github.com/spf13/cobra"
+
+	"git.warky.dev/wdevs/relspecgo/pkg/jobs"
+	"git.warky.dev/wdevs/relspecgo/pkg/merge"
+	"git.warky.dev/wdevs/relspecgo/pkg/models"
+	"git.warky.dev/wdevs/relspecgo/pkg/readers"
+	"git.warky.dev/wdevs/relspecgo/pkg/readers/sqldir"
+	wpgsql "git.warky.dev/wdevs/relspecgo/pkg/writers/pgsql"
+)
+
+var (
+	jobDir    string
+	jobFiles  []string
+	jobDryRun bool
+	jobNoDeps bool
+)
+
+var jobCmd = &cobra.Command{
+	Use:   "job",
+	Short: "Run declarative RelSpec jobs from job files",
+	Long: `Run named jobs declared in job files instead of repeating command-line arguments.
+
+A job file is a YAML manifest (relspec.yml, or relspec.<name>.yml for extra
+files) describing one or more jobs. Each job names a RelSpec command plus its
+inputs, output and options:
+
+  version: 1
+  jobs:
+    build-schema:
+      command: convert
+      description: Merge the DBML sources and emit PostgreSQL DDL
+      inputs:
+        - path: schema/core.dbml
+          format: dbml
+        - path: schema/tenant.dbml
+          format: dbml
+      output:
+        format: pgsql
+        path: build/schema.sql
+        overwrite: true
+      options:
+        flatten_schema: false
+      logfile: .relspec/log/build-schema.log
+
+Rules and guarantees:
+  - command is a closed allow-list (convert, merge, scripts-list). Arbitrary
+    shell strings are never executed.
+  - Every path is relative to the directory holding the job file and may not
+    escape it. Absolute and home-relative paths are rejected.
+  - Remote database credentials are referenced by environment-variable name
+    via conn_env; connection strings are never stored in the manifest and are
+    redacted from logs and diagnostics.
+  - Discovery and listing are deterministic.
+  - The whole plan is validated - unknown commands/formats, duplicate job
+    names, missing inputs, path traversal, dependency cycles - before any job
+    runs. Nothing is read, written or executed when validation fails.
+  - A failed job propagates the underlying non-zero exit status and writes no
+    success marker.`,
+}
+
+var jobListCmd = &cobra.Command{
+	Use:   "list",
+	Short: "List jobs discovered in job files (deterministic order)",
+	RunE:  runJobList,
+}
+
+var jobRunCmd = &cobra.Command{
+	Use:   "run <job-name>",
+	Short: "Run a named job (and its dependencies) from a job file",
+	Args:  cobra.ExactArgs(1),
+	RunE:  runJobRun,
+}
+
+func init() {
+	for _, c := range []*cobra.Command{jobListCmd, jobRunCmd} {
+		c.Flags().StringVar(&jobDir, "dir", ".", "Directory to discover job files in")
+		c.Flags().StringSliceVar(&jobFiles, "file", nil, "Explicit job file(s) to load (repeatable); disables discovery")
+	}
+	jobRunCmd.Flags().BoolVar(&jobDryRun, "dry-run", false, "Validate and print the execution plan without running anything")
+	jobRunCmd.Flags().BoolVar(&jobDryRun, "plan", false, "Alias for --dry-run")
+	jobRunCmd.Flags().BoolVar(&jobNoDeps, "no-deps", false, "Run only the named job, skipping its declared dependencies")
+
+	jobCmd.AddCommand(jobListCmd)
+	jobCmd.AddCommand(jobRunCmd)
+}
+
+// loadJobSet discovers or loads the requested job files and runs full
+// validation. The returned Set is safe to plan and execute.
+func loadJobSet() (*jobs.Set, error) {
+	paths := jobFiles
+	if len(paths) == 0 {
+		discovered, err := jobs.Discover(jobDir)
+		if err != nil {
+			return nil, err
+		}
+		paths = discovered
+	} else {
+		for i, p := range paths {
+			if _, err := os.Stat(p); err != nil {
+				return nil, fmt.Errorf("job file %q: %w", p, err)
+			}
+			paths[i] = p
+		}
+	}
+	set, err := jobs.Load(paths)
+	if err != nil {
+		return nil, err
+	}
+	if err := set.Validate(); err != nil {
+		return nil, err
+	}
+	return set, nil
+}
+
+func runJobList(cmd *cobra.Command, args []string) error {
+	set, err := loadJobSet()
+	if err != nil {
+		return err
+	}
+	out := cmd.OutOrStdout()
+
+	fmt.Fprintf(os.Stderr, "\n=== RelSpec Jobs ===\n")
+	fmt.Fprintf(os.Stderr, "Job files:\n")
+	for _, f := range set.Files {
+		fmt.Fprintf(os.Stderr, "  - %s\n", f)
+	}
+	fmt.Fprintln(os.Stderr)
+
+	names := set.Names()
+	if len(names) == 0 {
+		fmt.Fprintln(out, "(no jobs defined)")
+		return nil
+	}
+
+	nameW, cmdW, srcW := len("NAME"), len("COMMAND"), len("SOURCE")
+	for _, n := range names {
+		j := set.Jobs[n]
+		nameW = maxInt(nameW, len(n))
+		cmdW = maxInt(cmdW, len(j.Command))
+		srcW = maxInt(srcW, len(j.SourceFile))
+	}
+	fmt.Fprintf(out, "%-*s  %-*s  %-*s  %s\n", nameW, "NAME", cmdW, "COMMAND", srcW, "SOURCE", "DESCRIPTION")
+	for _, n := range names {
+		j := set.Jobs[n]
+		fmt.Fprintf(out, "%-*s  %-*s  %-*s  %s\n", nameW, n, cmdW, j.Command, srcW, j.SourceFile, j.Description)
+	}
+	return nil
+}
+
+func runJobRun(cmd *cobra.Command, args []string) error {
+	set, err := loadJobSet()
+	if err != nil {
+		return err
+	}
+	return executeJobPlan(set, args[0], jobDryRun, jobNoDeps, cmd.OutOrStdout())
+}
+
+// executeJobPlan resolves the plan for name, runs pre-flight checks over
+// EVERY job in the plan, and only then executes. When dryRun is set it prints
+// the plan and returns without touching any input, output or database.
+func executeJobPlan(set *jobs.Set, name string, dryRun, noDeps bool, out io.Writer) error {
+	plan, err := set.Plan(name, !noDeps)
+	if err != nil {
+		return err
+	}
+
+	// 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))
+	for i, j := range plan {
+		rj, perr := preflightJob(j)
+		if perr != nil {
+			return fmt.Errorf("job %q: %w", j.Name, perr)
+		}
+		resolved[i] = rj
+	}
+
+	if dryRun {
+		fmt.Fprintf(out, "RelSpec job plan for %q (dry run - nothing executed):\n\n", name)
+		for i, rj := range resolved {
+			printResolvedJob(out, i+1, len(resolved), rj)
+		}
+		return nil
+	}
+
+	for _, rj := range resolved {
+		if err := executeResolvedJob(rj); err != nil {
+			// Propagate the underlying failure; no success marker is written.
+			return fmt.Errorf("job %q failed: %w", rj.job.Name, err)
+		}
+	}
+	fmt.Fprintf(os.Stderr, "\n=== Job %q complete ===\n", name)
+	return nil
+}
+
+// resolvedJob is a job with every manifest path turned into a checked
+// absolute filesystem path and every conn_env resolved to its value.
+type resolvedJob struct {
+	job           *jobs.Job
+	root          string
+	inputs        []resolvedInput
+	scriptDirs    []string
+	outputPath    string // "" when the output is a database
+	outputConn    string // resolved connection string (secret)
+	outputConnEnv string
+	logPath       string
+	secrets       []string // resolved secret values to redact from logs
+}
+
+type resolvedInput struct {
+	format  string
+	path    string // "" when the input is a database
+	conn    string // resolved connection string (secret)
+	connEnv string
+}
+
+func preflightJob(j *jobs.Job) (*resolvedJob, error) {
+	root := j.Dir()
+	rj := &resolvedJob{job: j, root: root}
+
+	if j.Logfile != "" {
+		p, err := jobs.SafeJoin(root, j.Logfile)
+		if err != nil {
+			return nil, fmt.Errorf("logfile: %w", err)
+		}
+		rj.logPath = p
+	}
+
+	for i, in := range j.Inputs {
+		ri := resolvedInput{format: strings.ToLower(in.Format)}
+		if in.ConnEnv != "" {
+			v, ok := os.LookupEnv(in.ConnEnv)
+			if !ok || v == "" {
+				return nil, fmt.Errorf("input[%d]: environment variable %q (conn_env) is not set", i, in.ConnEnv)
+			}
+			ri.conn = v
+			ri.connEnv = in.ConnEnv
+			rj.secrets = append(rj.secrets, v)
+		} else {
+			p, err := jobs.SafeJoin(root, in.Path)
+			if err != nil {
+				return nil, fmt.Errorf("input[%d]: %w", i, err)
+			}
+			info, err := os.Stat(p)
+			if err != nil {
+				return nil, fmt.Errorf("input[%d]: %s: file not found", i, in.Path)
+			}
+			if info.IsDir() {
+				return nil, fmt.Errorf("input[%d]: %s: is a directory, not a file", i, in.Path)
+			}
+			ri.path = p
+		}
+		rj.inputs = append(rj.inputs, ri)
+	}
+
+	for _, d := range j.ScriptDirs {
+		p, err := jobs.SafeJoin(root, d)
+		if err != nil {
+			return nil, fmt.Errorf("script_dir %q: %w", d, err)
+		}
+		info, err := os.Stat(p)
+		if err != nil {
+			return nil, fmt.Errorf("script_dir %q: not found", d)
+		}
+		if !info.IsDir() {
+			return nil, fmt.Errorf("script_dir %q: not a directory", d)
+		}
+		rj.scriptDirs = append(rj.scriptDirs, p)
+	}
+
+	if j.Output != nil {
+		if j.Output.ConnEnv != "" {
+			v, ok := os.LookupEnv(j.Output.ConnEnv)
+			if !ok || v == "" {
+				return nil, fmt.Errorf("output: environment variable %q (conn_env) is not set", j.Output.ConnEnv)
+			}
+			rj.outputConn = v
+			rj.outputConnEnv = j.Output.ConnEnv
+			rj.secrets = append(rj.secrets, v)
+		} else {
+			p, err := jobs.SafeJoin(root, j.Output.Path)
+			if err != nil {
+				return nil, fmt.Errorf("output: %w", err)
+			}
+			if _, err := os.Stat(p); err == nil && !j.Output.Overwrite {
+				return nil, fmt.Errorf("output %s already exists (set output.overwrite: true to replace it)", j.Output.Path)
+			}
+			rj.outputPath = p
+		}
+	}
+	return rj, nil
+}
+
+func printResolvedJob(out io.Writer, n, total int, rj *resolvedJob) {
+	j := rj.job
+	fmt.Fprintf(out, "[%d/%d] %s\n", n, total, j.Name)
+	fmt.Fprintf(out, "      command: %s\n", j.Command)
+	if j.Description != "" {
+		fmt.Fprintf(out, "      description: %s\n", j.Description)
+	}
+	fmt.Fprintf(out, "      job file: %s\n", j.SourceFile)
+	for _, ri := range rj.inputs {
+		if ri.path != "" {
+			fmt.Fprintf(out, "      input: %s (%s)\n", ri.path, ri.format)
+		} else {
+			fmt.Fprintf(out, "      input: env:%s (%s)\n", ri.connEnv, ri.format)
+		}
+	}
+	for _, d := range rj.scriptDirs {
+		fmt.Fprintf(out, "      script dir: %s\n", d)
+	}
+	if rj.outputPath != "" {
+		fmt.Fprintf(out, "      output: %s (%s)\n", rj.outputPath, j.Output.Format)
+	} else if rj.outputConnEnv != "" {
+		fmt.Fprintf(out, "      output: env:%s (%s)\n", rj.outputConnEnv, j.Output.Format)
+	}
+	if rj.logPath != "" {
+		fmt.Fprintf(out, "      logfile: %s\n", rj.logPath)
+	}
+	fmt.Fprintln(out)
+}
+
+// executeResolvedJob runs a single already-validated job.
+func executeResolvedJob(rj *resolvedJob) (err error) {
+	lg, closeLog, lerr := newJobLogger(rj.logPath, rj.secrets)
+	if lerr != nil {
+		return lerr
+	}
+	defer func() { closeLog(err) }()
+
+	lg.logf("=== job %q (%s) started at %s ===", rj.job.Name, rj.job.Command, time.Now().Format(time.RFC3339))
+
+	switch rj.job.Command {
+	case jobs.CommandConvert:
+		err = runConvertJob(rj, lg)
+	case jobs.CommandMerge:
+		err = runMergeJob(rj, lg)
+	case jobs.CommandScriptsList:
+		err = runScriptsListJob(rj, lg)
+	default:
+		err = fmt.Errorf("unsupported command %q", rj.job.Command)
+	}
+	if err != nil {
+		lg.logf("FAILED: %v", err)
+	} else {
+		lg.logf("OK")
+	}
+	return err
+}
+
+func runConvertJob(rj *resolvedJob, lg *jobLogger) error {
+	db, err := readJobInputs(rj, lg)
+	if err != nil {
+		return err
+	}
+	return writeJobOutput(rj, db, lg)
+}
+
+func runMergeJob(rj *resolvedJob, lg *jobLogger) error {
+	opts := &merge.MergeOptions{
+		SkipDomains:   rj.job.Options.SkipDomains,
+		SkipRelations: rj.job.Options.SkipRelations,
+		SkipEnums:     rj.job.Options.SkipEnums,
+		SkipViews:     rj.job.Options.SkipViews,
+		SkipSequences: rj.job.Options.SkipSequences,
+	}
+	var base *models.Database
+	for i, ri := range rj.inputs {
+		db, err := readOneJobInput(ri)
+		if err != nil {
+			return fmt.Errorf("input[%d]: %w", i, err)
+		}
+		if base == nil {
+			base = db
+			lg.logf("merge target: %s", inputLabel(ri))
+			continue
+		}
+		lg.logf("merging: %s", inputLabel(ri))
+		merge.MergeDatabases(base, db, opts)
+	}
+	base.UpdateDate()
+	return writeJobOutput(rj, base, lg)
+}
+
+func runScriptsListJob(rj *resolvedJob, lg *jobLogger) error {
+	type row struct {
+		priority int
+		sequence uint
+		name     string
+		dir      string
+		lines    int
+	}
+	var rows []row
+	for _, dir := range rj.scriptDirs {
+		reader := sqldir.NewReader(&readers.ReaderOptions{
+			FilePath: dir,
+			Metadata: map[string]any{
+				"schema_name":   valueOr(rj.job.Options.Schema, "public"),
+				"database_name": "database",
+			},
+		})
+		db, err := reader.ReadDatabase()
+		if err != nil {
+			return fmt.Errorf("%s: %w", dir, err)
+		}
+		if len(db.Schemas) == 0 {
+			continue
+		}
+		for _, s := range db.Schemas[0].Scripts {
+			lines := strings.Count(s.SQL, "\n")
+			if len(s.SQL) > 0 && !strings.HasSuffix(s.SQL, "\n") {
+				lines++
+			}
+			rows = append(rows, row{s.Priority, s.Sequence, s.Name, dir, lines})
+		}
+	}
+	sort.Slice(rows, func(i, j int) bool {
+		if rows[i].priority != rows[j].priority {
+			return rows[i].priority < rows[j].priority
+		}
+		if rows[i].sequence != rows[j].sequence {
+			return rows[i].sequence < rows[j].sequence
+		}
+		if rows[i].name != rows[j].name {
+			return rows[i].name < rows[j].name
+		}
+		return rows[i].dir < rows[j].dir
+	})
+	lg.logf("found %d script(s) across %d director(y/ies):", len(rows), len(rj.scriptDirs))
+	lg.logf("%-4s %-9s %-9s %-30s %-6s %s", "No.", "Priority", "Sequence", "Name", "Lines", "Directory")
+	for i, r := range rows {
+		lg.logf("%-4d %-9d %-9d %-30s %-6d %s", i+1, r.priority, r.sequence, r.name, r.lines, r.dir)
+	}
+	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
+	for i, ri := range rj.inputs {
+		db, err := readOneJobInput(ri)
+		if err != nil {
+			return nil, fmt.Errorf("input[%d]: %w", i, err)
+		}
+		lg.logf("read input: %s", inputLabel(ri))
+		if base == nil {
+			base = db
+		} else {
+			merge.MergeDatabases(base, db, &merge.MergeOptions{})
+		}
+	}
+	if base == nil {
+		return nil, fmt.Errorf("no inputs produced a database")
+	}
+	return base, nil
+}
+
+func readOneJobInput(ri resolvedInput) (*models.Database, error) {
+	if ri.conn != "" {
+		return readDatabaseForConvert(ri.format, "", ri.conn)
+	}
+	return readDatabaseForConvert(ri.format, ri.path, "")
+}
+
+func inputLabel(ri resolvedInput) string {
+	if ri.path != "" {
+		return fmt.Sprintf("%s (%s)", ri.path, ri.format)
+	}
+	return fmt.Sprintf("env:%s (%s)", ri.connEnv, ri.format)
+}
+
+// writeJobOutput writes db to the job's output target (file or database).
+func writeJobOutput(rj *resolvedJob, db *models.Database, lg *jobLogger) error {
+	o := rj.job.Options
+	format := strings.ToLower(rj.job.Output.Format)
+
+	if rj.outputConn != "" {
+		if format != "pgsql" {
+			return fmt.Errorf("database output is only supported for pgsql (got %q)", rj.job.Output.Format)
+		}
+		lg.logf("writing output to database env:%s", rj.outputConnEnv)
+		writerOpts := newWriterOptions("", o.Package, o.FlattenSchema, "", "", o.ContinueOnError)
+		writerOpts.Metadata = map[string]interface{}{"connection_string": rj.outputConn}
+		return wpgsql.NewWriter(writerOpts).WriteDatabase(db)
+	}
+
+	if err := os.MkdirAll(filepath.Dir(rj.outputPath), 0o755); err != nil {
+		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, "")
+}
+
+// --- logging + redaction ---------------------------------------------------
+
+type jobLogger struct {
+	file    io.Writer
+	secrets []string
+}
+
+// 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) {
+	lg := &jobLogger{secrets: secrets}
+	if path == "" {
+		return lg, func(error) {}, nil
+	}
+	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+		return nil, nil, fmt.Errorf("failed to create log directory: %w", err)
+	}
+	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)
+	}
+	lg.file = f
+	return lg, func(runErr error) {
+		if runErr != nil {
+			fmt.Fprintf(f, "%s job ended with error\n", time.Now().Format(time.RFC3339))
+		}
+		_ = f.Close()
+	}, nil
+}
+
+func (l *jobLogger) logf(format string, args ...interface{}) {
+	line := l.redact(fmt.Sprintf(format, args...))
+	fmt.Fprintf(os.Stderr, "      %s\n", line)
+	if l.file != nil {
+		fmt.Fprintf(l.file, "%s %s\n", time.Now().Format(time.RFC3339), line)
+	}
+}
+
+func (l *jobLogger) redact(s string) string {
+	for _, sec := range l.secrets {
+		if sec != "" {
+			s = strings.ReplaceAll(s, sec, "***")
+		}
+	}
+	return maskPassword(s)
+}
+
+// --- small helpers -------------------------------------------------------
+
+func maxInt(a, b int) int {
+	if a > b {
+		return a
+	}
+	return b
+}
+
+func valueOr(v, def string) string {
+	if v == "" {
+		return def
+	}
+	return v
+}

cmd/relspec/job_test.go

@@ -0,0 +1,377 @@
+package main
+
+import (
+	"bytes"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/spf13/cobra"
+
+	"git.warky.dev/wdevs/relspecgo/pkg/jobs"
+)
+
+func writeFile(t *testing.T, path, content string) {
+	t.Helper()
+	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(path, []byte(content), 0o644); err != nil {
+		t.Fatal(err)
+	}
+}
+
+// jobFixture creates a job-file project with two DBML sources and returns the
+// project directory.
+func jobFixture(t *testing.T, manifest string) string {
+	t.Helper()
+	dir := t.TempDir()
+	writeFile(t, filepath.Join(dir, "schema", "core.dbml"), "Table users {\n  id int [pk]\n  name varchar\n}\n")
+	writeFile(t, filepath.Join(dir, "schema", "tenant.dbml"), "Table posts {\n  id int [pk]\n  title varchar\n}\n")
+	writeFile(t, filepath.Join(dir, "relspec.yml"), manifest)
+	return dir
+}
+
+func mustLoadSet(t *testing.T, files ...string) *jobs.Set {
+	t.Helper()
+	set, err := jobs.Load(files)
+	if err != nil {
+		t.Fatalf("load: %v", err)
+	}
+	if err := set.Validate(); err != nil {
+		t.Fatalf("validate: %v", err)
+	}
+	return set
+}
+
+const convertMergeManifest = `version: 1
+jobs:
+  build-schema:
+    command: convert
+    description: Merge DBML sources to PostgreSQL DDL
+    inputs:
+      - path: schema/core.dbml
+        format: dbml
+      - path: schema/tenant.dbml
+        format: dbml
+    output:
+      format: pgsql
+      path: build/schema.sql
+      overwrite: true
+    logfile: .relspec/log/build.log
+`
+
+func TestJobRun_ConvertMultiFileMerge(t *testing.T) {
+	dir := jobFixture(t, convertMergeManifest)
+	set := mustLoadSet(t, filepath.Join(dir, "relspec.yml"))
+
+	if err := executeJobPlan(set, "build-schema", false, false, &bytes.Buffer{}); err != nil {
+		t.Fatalf("executeJobPlan: %v", err)
+	}
+
+	out, err := os.ReadFile(filepath.Join(dir, "build", "schema.sql"))
+	if err != nil {
+		t.Fatalf("expected output file: %v", err)
+	}
+	sql := string(out)
+	if !strings.Contains(sql, "users") || !strings.Contains(sql, "posts") {
+		t.Fatalf("merged output missing tables:\n%s", sql)
+	}
+
+	logData, err := os.ReadFile(filepath.Join(dir, ".relspec", "log", "build.log"))
+	if err != nil {
+		t.Fatalf("expected logfile: %v", err)
+	}
+	if !strings.Contains(string(logData), "OK") {
+		t.Fatalf("logfile missing success marker:\n%s", logData)
+	}
+}
+
+func TestJobRun_DryRunDoesNotExecute(t *testing.T) {
+	dir := jobFixture(t, convertMergeManifest)
+	set := mustLoadSet(t, filepath.Join(dir, "relspec.yml"))
+
+	var buf bytes.Buffer
+	if err := executeJobPlan(set, "build-schema", true, false, &buf); err != nil {
+		t.Fatalf("dry run error: %v", err)
+	}
+	if !strings.Contains(buf.String(), "dry run") {
+		t.Fatalf("expected dry-run banner, got: %s", buf.String())
+	}
+	if _, err := os.Stat(filepath.Join(dir, "build", "schema.sql")); !os.IsNotExist(err) {
+		t.Fatal("dry run must not create the output file")
+	}
+	if _, err := os.Stat(filepath.Join(dir, ".relspec", "log", "build.log")); !os.IsNotExist(err) {
+		t.Fatal("dry run must not create the logfile")
+	}
+}
+
+func TestJobRun_ValidationFailureNoExecution(t *testing.T) {
+	badManifest := `version: 1
+jobs:
+  evil:
+    command: convert
+    inputs:
+      - path: ../../../etc/passwd
+        format: dbml
+    output:
+      format: json
+      path: build/out.json
+    logfile: .relspec/evil.log
+`
+	dir := jobFixture(t, badManifest)
+	if _, err := jobs.Load([]string{filepath.Join(dir, "relspec.yml")}); err != nil {
+		// structural load ok; validation should reject
+		t.Fatalf("unexpected load error: %v", err)
+	}
+	set, _ := jobs.Load([]string{filepath.Join(dir, "relspec.yml")})
+	if err := set.Validate(); err == nil {
+		t.Fatal("expected validation failure for path traversal")
+	}
+	// Nothing should have been produced.
+	if _, err := os.Stat(filepath.Join(dir, "build")); !os.IsNotExist(err) {
+		t.Fatal("validation failure must not create output dir")
+	}
+	if _, err := os.Stat(filepath.Join(dir, ".relspec")); !os.IsNotExist(err) {
+		t.Fatal("validation failure must not create logfile dir")
+	}
+}
+
+func TestJobRun_MissingInputNoExecution(t *testing.T) {
+	manifest := `version: 1
+jobs:
+  x:
+    command: convert
+    inputs:
+      - path: schema/does-not-exist.dbml
+        format: dbml
+    output:
+      format: json
+      path: build/out.json
+    logfile: .relspec/x.log
+`
+	dir := jobFixture(t, manifest)
+	set := mustLoadSet(t, filepath.Join(dir, "relspec.yml"))
+
+	err := executeJobPlan(set, "x", false, false, &bytes.Buffer{})
+	if err == nil || !strings.Contains(err.Error(), "not found") {
+		t.Fatalf("expected missing-input error, got %v", err)
+	}
+	if _, err := os.Stat(filepath.Join(dir, "build")); !os.IsNotExist(err) {
+		t.Fatal("missing input must not create output dir")
+	}
+	if _, err := os.Stat(filepath.Join(dir, ".relspec")); !os.IsNotExist(err) {
+		t.Fatal("missing input must not create logfile")
+	}
+}
+
+func TestJobRun_MissingConnEnvNoExecution(t *testing.T) {
+	manifest := `version: 1
+jobs:
+  remote:
+    command: convert
+    inputs:
+      - format: pgsql
+        conn_env: RELSPEC_TEST_MISSING_CONN
+    output:
+      format: json
+      path: build/out.json
+    logfile: .relspec/remote.log
+`
+	dir := jobFixture(t, manifest)
+	os.Unsetenv("RELSPEC_TEST_MISSING_CONN")
+	set := mustLoadSet(t, filepath.Join(dir, "relspec.yml"))
+
+	err := executeJobPlan(set, "remote", false, false, &bytes.Buffer{})
+	if err == nil || !strings.Contains(err.Error(), "conn_env") {
+		t.Fatalf("expected missing conn_env error, got %v", err)
+	}
+	if _, err := os.Stat(filepath.Join(dir, ".relspec")); !os.IsNotExist(err) {
+		t.Fatal("missing conn_env must not create logfile")
+	}
+}
+
+func TestJobRun_ExitCodePropagation(t *testing.T) {
+	// gorm output without options.package makes the underlying writer fail.
+	manifest := `version: 1
+jobs:
+  fail:
+    command: convert
+    inputs:
+      - path: schema/core.dbml
+        format: dbml
+    output:
+      format: gorm
+      path: build/models
+      overwrite: true
+    logfile: .relspec/fail.log
+`
+	dir := jobFixture(t, manifest)
+	set := mustLoadSet(t, filepath.Join(dir, "relspec.yml"))
+
+	err := executeJobPlan(set, "fail", false, false, &bytes.Buffer{})
+	if err == nil {
+		t.Fatal("expected underlying failure to propagate")
+	}
+	if !strings.Contains(err.Error(), "job \"fail\" failed") {
+		t.Fatalf("error should identify the failing job: %v", err)
+	}
+	// Logfile records the failure and no misleading success marker.
+	logData, _ := os.ReadFile(filepath.Join(dir, ".relspec", "fail.log"))
+	if strings.Contains(string(logData), "\nOK\n") || strings.HasSuffix(strings.TrimSpace(string(logData)), "OK") {
+		t.Fatalf("failed job must not log OK:\n%s", logData)
+	}
+	if !strings.Contains(string(logData), "FAILED") {
+		t.Fatalf("failed job should log FAILED:\n%s", logData)
+	}
+}
+
+func TestJobRun_DependencyChainExecutes(t *testing.T) {
+	manifest := `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
+    depends_on: [a]
+    inputs:
+      - path: schema/tenant.dbml
+        format: dbml
+    output:
+      format: json
+      path: build/b.json
+      overwrite: true
+`
+	dir := jobFixture(t, manifest)
+	set := mustLoadSet(t, filepath.Join(dir, "relspec.yml"))
+
+	if err := executeJobPlan(set, "b", false, false, &bytes.Buffer{}); err != nil {
+		t.Fatalf("executeJobPlan: %v", err)
+	}
+	for _, f := range []string{"a.json", "b.json"} {
+		if _, err := os.Stat(filepath.Join(dir, "build", f)); err != nil {
+			t.Fatalf("expected %s to be produced: %v", f, err)
+		}
+	}
+}
+
+func TestJobRun_ScriptsListMultipleDirs(t *testing.T) {
+	dir := t.TempDir()
+	writeFile(t, filepath.Join(dir, "migrations", "core", "1_001_create_users.sql"), "CREATE TABLE users();\n")
+	writeFile(t, filepath.Join(dir, "migrations", "tenant", "1_002_create_posts.sql"), "CREATE TABLE posts();\n")
+	writeFile(t, filepath.Join(dir, "migrations", "tenant", "2_001_add_index.sql"), "CREATE INDEX x ON posts(id);\n")
+	manifest := `version: 1
+jobs:
+  list-all:
+    command: scripts-list
+    script_dirs:
+      - migrations/core
+      - migrations/tenant
+    logfile: .relspec/scripts.log
+`
+	writeFile(t, filepath.Join(dir, "relspec.yml"), manifest)
+	set := mustLoadSet(t, filepath.Join(dir, "relspec.yml"))
+
+	if err := executeJobPlan(set, "list-all", false, false, &bytes.Buffer{}); err != nil {
+		t.Fatalf("executeJobPlan: %v", err)
+	}
+	logData, err := os.ReadFile(filepath.Join(dir, ".relspec", "scripts.log"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	s := string(logData)
+	iUsers := strings.Index(s, "create_users")
+	iPosts := strings.Index(s, "create_posts")
+	iIndex := strings.Index(s, "add_index")
+	if iUsers < 0 || iPosts < 0 || iIndex < 0 {
+		t.Fatalf("expected all scripts listed:\n%s", s)
+	}
+	if !(iUsers < iPosts && iPosts < iIndex) {
+		t.Fatalf("scripts not in priority/sequence order:\n%s", s)
+	}
+	if !strings.Contains(s, "found 3 script(s) across 2") {
+		t.Fatalf("expected multi-directory summary:\n%s", s)
+	}
+}
+
+func TestJobRun_ConnEnvRedactedInPlan(t *testing.T) {
+	manifest := `version: 1
+jobs:
+  remote:
+    command: convert
+    inputs:
+      - format: pgsql
+        conn_env: RELSPEC_TEST_PLAN_CONN
+    output:
+      format: json
+      path: build/out.json
+`
+	dir := jobFixture(t, manifest)
+	secret := "postgres://user:supersecret@db.example/app"
+	t.Setenv("RELSPEC_TEST_PLAN_CONN", secret)
+	set := mustLoadSet(t, filepath.Join(dir, "relspec.yml"))
+
+	var buf bytes.Buffer
+	if err := executeJobPlan(set, "remote", true, false, &buf); err != nil {
+		t.Fatalf("dry run: %v", err)
+	}
+	if strings.Contains(buf.String(), "supersecret") || strings.Contains(buf.String(), secret) {
+		t.Fatalf("plan leaked secret:\n%s", buf.String())
+	}
+	if !strings.Contains(buf.String(), "env:RELSPEC_TEST_PLAN_CONN") {
+		t.Fatalf("plan should reference the env var name:\n%s", buf.String())
+	}
+}
+
+func TestJobLogger_Redaction(t *testing.T) {
+	lg := &jobLogger{secrets: []string{"topsecret"}}
+	got := lg.redact("connecting with password topsecret and postgres://u:p@h/db")
+	if strings.Contains(got, "topsecret") {
+		t.Fatalf("secret not redacted: %q", got)
+	}
+	if !strings.Contains(got, "***") {
+		t.Fatalf("expected redaction marker: %q", got)
+	}
+}
+
+func TestJobList_DeterministicOutput(t *testing.T) {
+	manifest := `version: 1
+jobs:
+  zebra:
+    command: convert
+    inputs: [{path: schema/core.dbml, format: dbml}]
+    output: {format: json, path: build/z.json}
+  alpha:
+    command: convert
+    inputs: [{path: schema/core.dbml, format: dbml}]
+    output: {format: json, path: build/a.json}
+`
+	dir := jobFixture(t, manifest)
+
+	run := func() string {
+		jobDir = dir
+		jobFiles = nil
+		cmd := &cobra.Command{}
+		var buf bytes.Buffer
+		cmd.SetOut(&buf)
+		if err := runJobList(cmd, nil); err != nil {
+			t.Fatalf("runJobList: %v", err)
+		}
+		return buf.String()
+	}
+	first := run()
+	if strings.Index(first, "alpha") > strings.Index(first, "zebra") {
+		t.Fatalf("jobs not sorted:\n%s", first)
+	}
+	if first != run() {
+		t.Fatal("job list output not deterministic")
+	}
+}

cmd/relspec/root.go

@@ -62,6 +62,7 @@ func init() {
 	rootCmd.AddCommand(diffCmd)
 	rootCmd.AddCommand(inspectCmd)
 	rootCmd.AddCommand(scriptsCmd)
+	rootCmd.AddCommand(jobCmd)
 	rootCmd.AddCommand(assetsCmd)
 	rootCmd.AddCommand(templCmd)
 	rootCmd.AddCommand(editCmd)

docs/JOB_FILES.md

@@ -0,0 +1,223 @@
+# RelSpec Job Files
+
+Job files let you declare named, repeatable RelSpec workflows in YAML and run
+them with `relspec job run <name>` instead of retyping long command lines.
+
+```bash
+relspec job list                     # deterministic list of discovered jobs
+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)
+
+This is the smallest coherent contract that is safe and useful end to end.
+Anything not listed under "Supported" is intentionally deferred.
+
+### Not a shell
+
+`command` is a **closed allow-list**. There is no field anywhere that accepts a
+shell string, an executable path, or arbitrary arguments. Adding a new command
+means adding a vetted adapter in the RelSpec source.
+
+| command        | what it does                                                        |
+|----------------|--------------------------------------------------------------------|
+| `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 |
+| `scripts-list` | deterministically list SQL scripts across one or more directories   |
+
+Deferred (documented, not implemented here): `scripts` execution against a live
+database, `split`, `inspect`, `diff`, `templ`, job-to-job output wiring,
+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
+
+`relspec job` (no `--file`) scans `--dir` (default `.`) for:
+
+1. `relspec.yml` / `relspec.yaml` (the default file), then
+2. `relspec.<name>.yml` / `relspec.<name>.yaml` (extra files),
+
+each group sorted lexically. Order is stable across runs. Use `--file <path>`
+(repeatable) to load explicit files and skip discovery.
+
+All discovered/selected files are merged into one job namespace. A job name
+defined by **more than one file is a hard error** naming both files. YAML maps
+already forbid duplicate keys within a single file.
+
+### Paths
+
+* Every path (`inputs[].path`, `output.path`, `script_dirs[]`, `logfile`) is
+  **relative to the 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.
+
+### Credentials
+
+* Database inputs (`format: pgsql` / `mssql`) and database execution outputs
+  (`format: pgsql` with `conn_env`) reference an **environment variable name**
+  via `conn_env:`. The connection string itself is never stored in the
+  manifest.
+* A `conn_env` value that looks like a connection string (contains `:`, `/`,
+  `@`, `=`, spaces) is rejected.
+* Missing/empty environment variables are reported during pre-flight, before
+  execution.
+* Job logs and `--plan` output show `env:<NAME>`, never the value. Resolved
+  secret values and anything matching a connection-string password are
+  redacted (`***`) from the logfile and diagnostics.
+
+### Validation happens before execution
+
+`relspec job list` and `relspec job run` both fully validate the selected set
+first. Nothing is read, written, connected to, or executed if validation fails.
+Checks include:
+
+* schema `version` (must be `1`), unknown YAML fields rejected
+* duplicate job names across files
+* unknown / missing `command`
+* per-command input/output shape (`convert`/`merge` need inputs + output;
+  `scripts-list` needs `script_dirs` and forbids inputs/output)
+* unknown input/output `format`
+* path traversal / absolute / home-relative paths
+* `depends_on` targets exist
+* dependency cycles (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 `script_dir` exists and is a directory
+* every `conn_env` variable is set
+* `output.path` does not already exist unless `output.overwrite: true`
+
+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
+  topological order (deterministic), then the job. `--no-deps` runs only the
+  named job.
+* `--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".
+
+## Schema reference
+
+```yaml
+version: 1                       # required, must be 1
+jobs:
+  <job-name>:
+    command: convert | merge | scripts-list   # required
+    description: "free text"                   # optional, shown by `job list`
+    depends_on: [other-job, ...]               # optional
+    inputs:                                    # convert (≥1) / merge (≥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)
+      - migrations/core
+      - migrations/tenant
+    output:                                    # convert / merge (required)
+      format: pgsql
+      path: build/schema.sql                   # file output, OR:
+      conn_env: TARGET_DB_URL                  # execute against DB (pgsql only)
+      overwrite: false                         # default false
+    options:
+      flatten_schema: false
+      schema: public
+      package: models                          # for gorm/bun output
+      continue_on_error: false                 # pgsql 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
+```
+
+### Supported input formats
+
+`dbml`, `dctx`, `drawdb`, `graphql`, `json`, `yaml`, `gorm`, `bun`, `drizzle`,
+`prisma`, `typeorm`, `sqlite` (file, via `path`); `pgsql`, `mssql`
+(live, via `conn_env`).
+
+### Supported output formats
+
+`dbml`, `dctx`, `drawdb`, `graphql`, `json`, `yaml`, `gorm`, `bun`, `drizzle`,
+`prisma`, `typeorm`, `pgsql`, `mssql`, `sqlite` (file, via `path`); `pgsql` also
+supports `conn_env` to execute the generated DDL against a live database.
+
+## Examples
+
+### Merge many schema files, emit PostgreSQL DDL
+
+```yaml
+version: 1
+jobs:
+  build-schema:
+    command: convert
+    inputs:
+      - { path: schema/core.dbml, format: dbml }
+      - { path: schema/billing.dbml, format: dbml }
+      - { path: schema/tenant.dbml, format: dbml }
+    output:
+      format: pgsql
+      path: build/schema.sql
+      overwrite: true
+    logfile: .relspec/log/build-schema.log
+```
+
+### Multiple script directories
+
+```yaml
+version: 1
+jobs:
+  migration-order:
+    command: scripts-list
+    script_dirs:
+      - migrations/core
+      - migrations/tenant
+      - migrations/reporting
+    logfile: .relspec/log/migration-order.log
+```
+
+### Job depending on another job
+
+```yaml
+version: 1
+jobs:
+  build-schema:
+    command: convert
+    inputs:
+      - { path: schema/core.dbml, format: dbml }
+      - { path: schema/tenant.dbml, format: dbml }
+    output: { format: json, path: build/schema.json, overwrite: true }
+  build-docs:
+    command: convert
+    depends_on: [build-schema]
+    inputs:
+      - { path: schema/core.dbml, format: dbml }
+    output: { format: yaml, path: build/schema.yaml, overwrite: true }
+```
+
+### Reading from a remote database
+
+```yaml
+version: 1
+jobs:
+  snapshot-prod:
+    command: convert
+    inputs:
+      - format: pgsql
+        conn_env: PROD_DB_URL      # export PROD_DB_URL=postgres://...
+    output:
+      format: dbml
+      path: snapshots/prod.dbml
+      overwrite: true
+```

examples/jobs/.gitignore

@@ -0,0 +1,3 @@
+# Generated by `relspec job run` in this example project.
+/build/
+/.relspec/

examples/jobs/migrations/core/1_001_create_users.sql

@@ -0,0 +1,4 @@
+CREATE TABLE users (
+    id SERIAL PRIMARY KEY,
+    email VARCHAR NOT NULL UNIQUE
+);

examples/jobs/migrations/tenant/1_002_create_posts.sql

@@ -0,0 +1,5 @@
+CREATE TABLE posts (
+    id SERIAL PRIMARY KEY,
+    user_id INT NOT NULL REFERENCES users(id),
+    title VARCHAR NOT NULL
+);

examples/jobs/migrations/tenant/2_001_add_posts_user_idx.sql

@@ -0,0 +1 @@
+CREATE INDEX posts_user_id_idx ON posts(user_id);

examples/jobs/relspec.yml

@@ -0,0 +1,45 @@
+# Example RelSpec job file. See docs/JOB_FILES.md for the full reference.
+#
+#   cd examples/jobs
+#   relspec job list
+#   relspec job run build-schema --plan
+#   relspec job run build-schema
+version: 1
+jobs:
+  build-schema:
+    command: convert
+    description: Merge the DBML sources and emit PostgreSQL DDL
+    inputs:
+      - path: schema/core.dbml
+        format: dbml
+      - path: schema/tenant.dbml
+        format: dbml
+    output:
+      format: pgsql
+      path: build/schema.sql
+      overwrite: true
+    options:
+      flatten_schema: false
+    logfile: .relspec/log/build-schema.log
+
+  build-json:
+    command: convert
+    description: Also emit a JSON schema once build-schema succeeds
+    depends_on: [build-schema]
+    inputs:
+      - path: schema/core.dbml
+        format: dbml
+      - path: schema/tenant.dbml
+        format: dbml
+    output:
+      format: json
+      path: build/schema.json
+      overwrite: true
+
+  migration-order:
+    command: scripts-list
+    description: Show the combined execution order across script directories
+    script_dirs:
+      - migrations/core
+      - migrations/tenant
+    logfile: .relspec/log/migration-order.log

examples/jobs/schema/core.dbml

@@ -0,0 +1,5 @@
+Table users {
+  id int [pk, increment]
+  email varchar [not null, unique]
+  created_at timestamp
+}

examples/jobs/schema/tenant.dbml

@@ -0,0 +1,6 @@
+Table posts {
+  id int [pk, increment]
+  user_id int [not null, ref: > users.id]
+  title varchar [not null]
+  body text
+}

linux/arch/PKGBUILD

@@ -1,6 +1,6 @@
 # Maintainer: Hein (Warky Devs) <hein@warky.dev>
 pkgname=relspec
-pkgver=1.0.73
+pkgver=1.0.74
 pkgrel=1
 pkgdesc="RelSpec is a comprehensive database relations management tool that reads, transforms, and writes database table specifications across multiple formats and ORMs."
 arch=('x86_64' 'aarch64')

linux/centos/relspec.spec

@@ -1,5 +1,5 @@
 Name:           relspec
-Version:        1.0.73
+Version:        1.0.74
 Release:        1%{?dist}
 Summary:        RelSpec is a comprehensive database relations management tool that reads, transforms, and writes database table specifications across multiple formats and ORMs.
 

pkg/diff/diff.go

@@ -4,6 +4,8 @@ import (
 	"fmt"
 	"reflect"
 	"sort"
+	"strconv"
+	"strings"
 
 	"git.warky.dev/wdevs/relspecgo/pkg/models"
 )
@@ -229,11 +231,13 @@ func compareColumns(source, target map[string]*models.Column) *ColumnDiff {
 
 func compareColumnDetails(source, target *models.Column) map[string]any {
 	changes := make(map[string]any)
+	sourceType, sourceLength, sourceDefault := comparableColumn(source)
+	targetType, targetLength, targetDefault := comparableColumn(target)
 
-	if source.Type != target.Type {
+	if sourceType != targetType {
 		changes["type"] = map[string]string{"source": source.Type, "target": target.Type}
 	}
-	if source.Length != target.Length {
+	if sourceLength != targetLength {
 		changes["length"] = map[string]int{"source": source.Length, "target": target.Length}
 	}
 	if source.Precision != target.Precision {
@@ -245,8 +249,8 @@ func compareColumnDetails(source, target *models.Column) map[string]any {
 	if source.NotNull != target.NotNull {
 		changes["not_null"] = map[string]bool{"source": source.NotNull, "target": target.NotNull}
 	}
-	if !reflect.DeepEqual(source.Default, target.Default) {
-		changes["default"] = map[string]any{"source": source.Default, "target": target.Default}
+	if !reflect.DeepEqual(sourceDefault, targetDefault) {
+		changes["default"] = map[string]any{"source": sourceDefault, "target": targetDefault}
 	}
 	if source.AutoIncrement != target.AutoIncrement {
 		changes["auto_increment"] = map[string]bool{"source": source.AutoIncrement, "target": target.AutoIncrement}
@@ -258,6 +262,28 @@ func compareColumnDetails(source, target *models.Column) map[string]any {
 	return changes
 }
 
+// comparableColumn accepts DBML's compact type/default spelling as well as
+// PostgreSQL's normalized fields (for example varchar(255) vs varchar + 255).
+func comparableColumn(column *models.Column) (string, int, any) {
+	typeName := strings.TrimSpace(column.Type)
+	defaultValue := column.Default
+	lower := strings.ToLower(typeName)
+	if i := strings.Index(lower, " default "); i >= 0 {
+		if defaultValue == nil {
+			defaultValue = strings.TrimSpace(typeName[i+len(" default "):])
+		}
+		typeName = strings.TrimSpace(typeName[:i])
+	}
+	length := column.Length
+	if open := strings.LastIndex(typeName, "("); open >= 0 && strings.HasSuffix(typeName, ")") {
+		if parsed, err := strconv.Atoi(strings.TrimSpace(typeName[open+1 : len(typeName)-1])); err == nil && length == 0 {
+			length = parsed
+		}
+		typeName = strings.TrimSpace(typeName[:open])
+	}
+	return strings.ToLower(typeName), length, defaultValue
+}
+
 func compareIndexes(source, target map[string]*models.Index) *IndexDiff {
 	diff := &IndexDiff{
 		Missing:  make([]*models.Index, 0),
@@ -265,12 +291,27 @@ func compareIndexes(source, target map[string]*models.Index) *IndexDiff {
 		Modified: make([]*IndexChange, 0),
 	}
 
-	// Find missing and modified indexes
+	// Match by name first, then by definition. PostgreSQL and DBML can assign
+	// different names to the same index (for example, posts_user_id_title_idx
+	// and uidx_posts_user_id_title), so a name-only comparison reports false
+	// drift after a merge/diff round trip.
+	unmatchedSource := make(map[string]*models.Index, len(source))
+	unmatchedTarget := make(map[string]*models.Index, len(target))
+	for name, index := range source {
+		unmatchedSource[name] = index
+	}
+	for name, index := range target {
+		unmatchedTarget[name] = index
+	}
+
 	for _, name := range sortedKeys(source) {
 		srcIdx := source[name]
-		if tgtIdx, exists := target[name]; !exists {
-			diff.Missing = append(diff.Missing, srcIdx)
-		} else {
+		tgtIdx, exists := target[name]
+		if !exists {
+			continue
+		}
+		delete(unmatchedSource, name)
+		delete(unmatchedTarget, name)
 		if changes := compareIndexDetails(srcIdx, tgtIdx); len(changes) > 0 {
 			diff.Modified = append(diff.Modified, &IndexChange{
 				Name:    name,
@@ -280,19 +321,57 @@ func compareIndexes(source, target map[string]*models.Index) *IndexDiff {
 			})
 		}
 	}
-	}
 
-	// Find extra indexes
-	for _, name := range sortedKeys(target) {
-		tgtIdx := target[name]
-		if _, exists := source[name]; !exists {
-			diff.Extra = append(diff.Extra, tgtIdx)
+	// Pair remaining indexes by their structural identity, independent of the
+	// generated/name field. The sorted iteration makes ambiguous matches
+	// deterministic; duplicate definitions are still represented as separate
+	// indexes by consuming one target at a time.
+	remainingTarget := make(map[string][]*models.Index)
+	for _, name := range sortedKeys(unmatchedTarget) {
+		index := unmatchedTarget[name]
+		key := indexDefinitionKey(index)
+		remainingTarget[key] = append(remainingTarget[key], index)
+	}
+	for _, name := range sortedKeys(unmatchedSource) {
+		srcIdx := unmatchedSource[name]
+		key := indexDefinitionKey(srcIdx)
+		candidates := remainingTarget[key]
+		if len(candidates) == 0 {
+			diff.Missing = append(diff.Missing, srcIdx)
+			continue
+		}
+		tgtIdx := candidates[0]
+		remainingTarget[key] = candidates[1:]
+		if changes := compareIndexDetails(srcIdx, tgtIdx); len(changes) > 0 {
+			diff.Modified = append(diff.Modified, &IndexChange{
+				Name:    srcIdx.Name,
+				Source:  srcIdx,
+				Target:  tgtIdx,
+				Changes: changes,
+			})
 		}
 	}
 
+	for _, key := range sortedKeys(remainingTarget) {
+		for _, index := range remainingTarget[key] {
+			diff.Extra = append(diff.Extra, index)
+		}
+	}
 	return diff
 }
 
+func indexDefinitionKey(index *models.Index) string {
+	return fmt.Sprintf("%t:%s:%s", index.Unique, strings.Join(index.Columns, ","), strings.Join(index.Include, ","))
+}
+
+func comparableIndexType(indexType string) string {
+	indexType = strings.ToLower(strings.TrimSpace(indexType))
+	if indexType == "" {
+		return "btree"
+	}
+	return indexType
+}
+
 func compareIndexDetails(source, target *models.Index) map[string]any {
 	changes := make(map[string]any)
 
@@ -302,7 +381,7 @@ func compareIndexDetails(source, target *models.Index) map[string]any {
 	if source.Unique != target.Unique {
 		changes["unique"] = map[string]bool{"source": source.Unique, "target": target.Unique}
 	}
-	if source.Type != target.Type {
+	if comparableIndexType(source.Type) != comparableIndexType(target.Type) {
 		changes["type"] = map[string]string{"source": source.Type, "target": target.Type}
 	}
 	if source.Where != target.Where {
@@ -312,7 +391,26 @@ func compareIndexDetails(source, target *models.Index) map[string]any {
 	return changes
 }
 
+// Compare constraints.
+// Primary-key constraints are excluded: a PK is already represented by the
+// column's IsPrimaryKey flag, which compareColumns already compares. The
+// PostgreSQL reader additionally materialises each PK as a primary_key
+// constraint and a unique btree index; the DBML reader keeps PKs as column
+// flags only. Comparing the constraint maps directly would therefore report
+// every PK as an "extra" constraint and the generated index as an "extra"
+// index on a freshly-applied schema. Filtering them here keeps the round
+// trip stable without losing real PK information.
 func compareConstraints(source, target map[string]*models.Constraint) *ConstraintDiff {
+	filteredSource := filterPrimaryKeyConstraints(source)
+	filteredTarget := filterPrimaryKeyConstraints(target)
+	sourceByKey := make(map[string]*models.Constraint, len(filteredSource))
+	targetByKey := make(map[string]*models.Constraint, len(filteredTarget))
+	for _, constraint := range filteredSource {
+		sourceByKey[constraintCompareKey(constraint)] = constraint
+	}
+	for _, constraint := range filteredTarget {
+		targetByKey[constraintCompareKey(constraint)] = constraint
+	}
 	diff := &ConstraintDiff{
 		Missing:  make([]*models.Constraint, 0),
 		Extra:    make([]*models.Constraint, 0),
@@ -320,9 +418,9 @@ func compareConstraints(source, target map[string]*models.Constraint) *Constrain
 	}
 
 	// Find missing and modified constraints
-	for _, name := range sortedKeys(source) {
-		srcCon := source[name]
-		if tgtCon, exists := target[name]; !exists {
+	for _, name := range sortedKeys(sourceByKey) {
+		srcCon := sourceByKey[name]
+		if tgtCon, exists := targetByKey[name]; !exists {
 			diff.Missing = append(diff.Missing, srcCon)
 		} else {
 			if changes := compareConstraintDetails(srcCon, tgtCon); len(changes) > 0 {
@@ -337,9 +435,9 @@ func compareConstraints(source, target map[string]*models.Constraint) *Constrain
 	}
 
 	// Find extra constraints
-	for _, name := range sortedKeys(target) {
-		tgtCon := target[name]
-		if _, exists := source[name]; !exists {
+	for _, name := range sortedKeys(targetByKey) {
+		tgtCon := targetByKey[name]
+		if _, exists := sourceByKey[name]; !exists {
 			diff.Extra = append(diff.Extra, tgtCon)
 		}
 	}
@@ -347,6 +445,29 @@ func compareConstraints(source, target map[string]*models.Constraint) *Constrain
 	return diff
 }
 
+// filterPrimaryKeyConstraints drops primary_key constraints from a single
+// map. Primary keys are compared by the column IsPrimaryKey flag in
+// compareColumns, so comparing the primary_key constraints here only
+// produces duplicate "extra" entries (every PK is extra on the DBML side).
+// Other constraint types are preserved untouched.
+func filterPrimaryKeyConstraints(m map[string]*models.Constraint) map[string]*models.Constraint {
+	out := make(map[string]*models.Constraint, len(m))
+	for name, c := range m {
+		if c.Type == models.PrimaryKeyConstraint {
+			continue
+		}
+		out[name] = c
+	}
+	return out
+}
+
+func constraintCompareKey(constraint *models.Constraint) string {
+	if constraint.Type != models.ForeignKeyConstraint {
+		return constraint.SQLName()
+	}
+	return fmt.Sprintf("fk:%s:%s:%s:%s:%s:%s", strings.ToLower(constraint.Schema), strings.ToLower(constraint.Table), strings.Join(constraint.Columns, ","), strings.ToLower(constraint.ReferencedSchema), strings.ToLower(constraint.ReferencedTable), strings.Join(constraint.ReferencedColumns, ","))
+}
+
 func compareConstraintDetails(source, target *models.Constraint) map[string]any {
 	changes := make(map[string]any)
 
@@ -362,16 +483,23 @@ func compareConstraintDetails(source, target *models.Constraint) map[string]any
 	if !reflect.DeepEqual(source.ReferencedColumns, target.ReferencedColumns) {
 		changes["referenced_columns"] = map[string][]string{"source": source.ReferencedColumns, "target": target.ReferencedColumns}
 	}
-	if source.OnDelete != target.OnDelete {
+	if normalizeConstraintAction(source.OnDelete) != normalizeConstraintAction(target.OnDelete) {
 		changes["on_delete"] = map[string]string{"source": source.OnDelete, "target": target.OnDelete}
 	}
-	if source.OnUpdate != target.OnUpdate {
+	if normalizeConstraintAction(source.OnUpdate) != normalizeConstraintAction(target.OnUpdate) {
 		changes["on_update"] = map[string]string{"source": source.OnUpdate, "target": target.OnUpdate}
 	}
 
 	return changes
 }
 
+func normalizeConstraintAction(action string) string {
+	if strings.EqualFold(strings.TrimSpace(action), "NO ACTION") {
+		return ""
+	}
+	return strings.ToUpper(strings.TrimSpace(action))
+}
+
 func compareRelationships(source, target map[string]*models.Relationship) *RelationshipDiff {
 	diff := &RelationshipDiff{
 		Missing:  make([]*models.Relationship, 0),

pkg/diff/diff_test.go

@@ -301,6 +301,22 @@ func TestCompareIndexes(t *testing.T) {
 				return len(d.Modified) == 1 && d.Modified[0].Name == "idx_name"
 			},
 		},
+		{
+			name: "equivalent indexes with different generated names",
+			source: map[string]*models.Index{
+				"uidx_posts_user_id_title": {
+					Name: "uidx_posts_user_id_title", Columns: []string{"user_id", "title"}, Unique: true,
+				},
+			},
+			target: map[string]*models.Index{
+				"posts_user_id_title_idx": {
+					Name: "posts_user_id_title_idx", Columns: []string{"user_id", "title"}, Unique: true, Type: "btree",
+				},
+			},
+			want: func(d *IndexDiff) bool {
+				return len(d.Missing) == 0 && len(d.Extra) == 0 && len(d.Modified) == 0
+			},
+		},
 	}
 
 	for _, tt := range tests {

pkg/jobs/jobs.go

@@ -0,0 +1,517 @@
+// Package jobs implements RelSpec declarative job files.
+//
+// A job file is a small YAML manifest that names one or more jobs and,
+// for each job, the RelSpec command to run plus its inputs, output and
+// options. It lets users run "relspec job run build-schema" instead of
+// repeating long command lines.
+//
+// The job-file system is deliberately NOT a shell: "command" is a closed
+// enum of vetted RelSpec workflows, every path is resolved relative to the
+// directory holding the job file and may not escape it, and remote database
+// credentials are referenced by environment-variable name only - never
+// embedded in the manifest. All discovery, parsing and validation in this
+// package is side-effect free; nothing here reads input schemas, opens
+// database connections or writes output. Execution lives in the CLI layer
+// and only runs after Validate and the caller's pre-flight checks pass.
+package jobs
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+
+	"gopkg.in/yaml.v3"
+)
+
+// SchemaVersion is the only job-file schema version this build understands.
+const SchemaVersion = 1
+
+// 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
+)
+
+// SupportedCommands lists every accepted command, in help order.
+var SupportedCommands = []string{CommandConvert, CommandMerge, CommandScriptsList}
+
+// readerFormats are the file-based input formats a job may declare (path).
+var readerFormats = map[string]bool{
+	"dbml": true, "dctx": true, "drawdb": true, "graphql": true, "json": true,
+	"yaml": true, "gorm": true, "bun": true, "drizzle": true, "prisma": true,
+	"typeorm": true, "sqlite": true,
+}
+
+// inputDBFormats are input formats that can only come from a live connection,
+// referenced by conn_env.
+var inputDBFormats = map[string]bool{"pgsql": true, "mssql": true}
+
+// writerFormats are the output formats a job may declare.
+var writerFormats = map[string]bool{
+	"dbml": true, "dctx": true, "drawdb": true, "graphql": true, "json": true,
+	"yaml": true, "gorm": true, "bun": true, "drizzle": true, "prisma": true,
+	"typeorm": true, "pgsql": true, "mssql": true, "sqlite": true,
+}
+
+// execOutputFormats are output formats for which conn_env (execute against a
+// live database) is supported instead of writing a file.
+var execOutputFormats = map[string]bool{"pgsql": true}
+
+// File is the on-disk shape of a single job file.
+type File struct {
+	Version int             `yaml:"version"`
+	Jobs    map[string]*Job `yaml:"jobs"`
+}
+
+// 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:"-"`
+
+	Command     string   `yaml:"command"`
+	Description string   `yaml:"description"`
+	DependsOn   []string `yaml:"depends_on"`
+	Inputs      []Input  `yaml:"inputs"`
+	ScriptDirs  []string `yaml:"script_dirs"`
+	Output      *Output  `yaml:"output"`
+	Options     Options  `yaml:"options"`
+	Logfile     string   `yaml:"logfile"`
+}
+
+// Input is one declared input schema.
+type Input struct {
+	Path string `yaml:"path"`
+	// Format is the RelSpec reader format (dbml, json, yaml, pgsql, ...).
+	Format string `yaml:"format"`
+	// 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"`
+}
+
+// Output is the declared output target.
+type Output struct {
+	Format    string `yaml:"format"`
+	Path      string `yaml:"path"`
+	ConnEnv   string `yaml:"conn_env"`
+	Overwrite bool   `yaml:"overwrite"`
+}
+
+// Options carries the subset of command flags a job file may set.
+type Options struct {
+	FlattenSchema   bool   `yaml:"flatten_schema"`
+	Schema          string `yaml:"schema"`
+	Package         string `yaml:"package"`
+	ContinueOnError bool   `yaml:"continue_on_error"`
+	SkipRelations   bool   `yaml:"skip_relations"`
+	SkipEnums       bool   `yaml:"skip_enums"`
+	SkipViews       bool   `yaml:"skip_views"`
+	SkipDomains     bool   `yaml:"skip_domains"`
+	SkipSequences   bool   `yaml:"skip_sequences"`
+}
+
+// Dir returns the directory that a job's relative paths resolve against:
+// the directory containing the job file that declared it.
+func (j *Job) Dir() string { return filepath.Dir(j.SourceFile) }
+
+// Set is the merged view of all discovered/selected job files.
+type Set struct {
+	// Files is the sorted list of job files that contributed jobs.
+	Files []string
+	// Jobs is keyed by job name.
+	Jobs map[string]*Job
+}
+
+// Names returns all job names in deterministic (sorted) order.
+func (s *Set) Names() []string {
+	names := make([]string, 0, len(s.Jobs))
+	for n := range s.Jobs {
+		names = append(names, n)
+	}
+	sort.Strings(names)
+	return names
+}
+
+// Discover returns the job files in dir in deterministic order. The default
+// file "relspec.yml"/"relspec.yaml" sorts first, followed by named files
+// "relspec.<name>.yml"/"relspec.<name>.yaml" in lexical order.
+func Discover(dir string) ([]string, error) {
+	if dir == "" {
+		dir = "."
+	}
+	entries, err := os.ReadDir(dir)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read directory %q: %w", dir, err)
+	}
+	var defaults, named []string
+	for _, e := range entries {
+		if e.IsDir() {
+			continue
+		}
+		name := e.Name()
+		if !isJobFileName(name) {
+			continue
+		}
+		full := filepath.Join(dir, name)
+		if name == "relspec.yml" || name == "relspec.yaml" {
+			defaults = append(defaults, full)
+		} else {
+			named = append(named, full)
+		}
+	}
+	sort.Strings(defaults)
+	sort.Strings(named)
+	return append(defaults, named...), nil
+}
+
+func isJobFileName(name string) bool {
+	for _, ext := range []string{".yml", ".yaml"} {
+		if name == "relspec"+ext {
+			return true
+		}
+		if strings.HasPrefix(name, "relspec.") && strings.HasSuffix(name, ext) {
+			return true
+		}
+	}
+	return false
+}
+
+// Load parses every path, rejects unknown fields and unsupported versions,
+// and merges all jobs into one Set. A job name defined by more than one file
+// is a hard error. Load performs structural checks only; call Validate for
+// full semantic validation.
+func Load(paths []string) (*Set, error) {
+	if len(paths) == 0 {
+		return nil, fmt.Errorf("no job files found (looked for relspec.yml / relspec.<name>.yml)")
+	}
+	set := &Set{Jobs: map[string]*Job{}}
+	origin := map[string]string{} // job name -> first file that defined it
+
+	for _, path := range paths {
+		data, err := os.ReadFile(path)
+		if err != nil {
+			return nil, fmt.Errorf("failed to read job file %q: %w", path, err)
+		}
+		dec := yaml.NewDecoder(strings.NewReader(string(data)))
+		dec.KnownFields(true)
+		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)
+		}
+		for name, job := range f.Jobs {
+			if job == nil {
+				return nil, fmt.Errorf("job file %q: job %q is empty", path, name)
+			}
+			if prev, dup := origin[name]; dup {
+				return nil, fmt.Errorf("duplicate job %q defined in both %q and %q", name, prev, path)
+			}
+			job.Name = name
+			job.SourceFile = path
+			origin[name] = path
+			set.Jobs[name] = job
+		}
+		set.Files = append(set.Files, path)
+	}
+	return set, nil
+}
+
+// Validate runs full semantic validation over the whole set and returns a
+// single error describing every problem found. It never touches the
+// filesystem beyond what Load already read; existence of input files and
+// environment variables is checked by the caller immediately before
+// execution.
+func (s *Set) Validate() error {
+	var errs []string
+	for _, name := range s.Names() {
+		for _, msg := range s.Jobs[name].validate() {
+			errs = append(errs, fmt.Sprintf("job %q: %s", name, msg))
+		}
+	}
+	// Dependency references + cycles.
+	for _, name := range s.Names() {
+		for _, dep := range s.Jobs[name].DependsOn {
+			if _, ok := s.Jobs[dep]; !ok {
+				errs = append(errs, fmt.Sprintf("job %q: depends_on unknown job %q", name, dep))
+			}
+		}
+	}
+	if cycle := s.findCycle(); cycle != "" {
+		errs = append(errs, fmt.Sprintf("dependency cycle detected: %s", cycle))
+	}
+	if len(errs) > 0 {
+		sort.Strings(errs)
+		return fmt.Errorf("job file validation failed:\n  - %s", strings.Join(errs, "\n  - "))
+	}
+	return nil
+}
+
+func (j *Job) validate() []string {
+	var e []string
+
+	switch j.Command {
+	case CommandConvert, CommandMerge, CommandScriptsList:
+	case "":
+		e = append(e, "missing command")
+		return e
+	default:
+		e = append(e, fmt.Sprintf("unsupported command %q (supported: %s)", j.Command, strings.Join(SupportedCommands, ", ")))
+		return e
+	}
+
+	// Path safety for every declared path.
+	checkPath := func(label, p string) {
+		if p == "" {
+			return
+		}
+		if err := checkRelPath(p); err != nil {
+			e = append(e, fmt.Sprintf("%s %q: %v", label, p, err))
+		}
+	}
+	checkPath("logfile", j.Logfile)
+	for _, in := range j.Inputs {
+		checkPath("input path", in.Path)
+	}
+	for _, d := range j.ScriptDirs {
+		checkPath("script_dir", d)
+	}
+	if j.Output != nil {
+		checkPath("output path", j.Output.Path)
+	}
+
+	switch j.Command {
+	case CommandConvert, CommandMerge:
+		minInputs := 1
+		if j.Command == CommandMerge {
+			minInputs = 2
+		}
+		if len(j.Inputs) < minInputs {
+			e = append(e, fmt.Sprintf("command %q requires at least %d input(s)", j.Command, minInputs))
+		}
+		for i, in := range j.Inputs {
+			e = append(e, validateInput(i, in)...)
+		}
+		if len(j.ScriptDirs) > 0 {
+			e = append(e, fmt.Sprintf("script_dirs is not valid for command %q", j.Command))
+		}
+		if j.Output == nil {
+			e = append(e, "missing output")
+		} else {
+			e = append(e, validateOutput(*j.Output)...)
+		}
+	case CommandScriptsList:
+		if len(j.ScriptDirs) == 0 {
+			e = append(e, "command \"scripts-list\" requires at least one script_dir")
+		}
+		if len(j.Inputs) > 0 {
+			e = append(e, "inputs is not valid for command \"scripts-list\"")
+		}
+		if j.Output != nil {
+			e = append(e, "output is not valid for command \"scripts-list\"")
+		}
+	}
+	return e
+}
+
+func validateInput(i int, in Input) []string {
+	var e []string
+	if in.Format == "" {
+		e = append(e, fmt.Sprintf("input[%d]: missing format", i))
+		return e
+	}
+	f := strings.ToLower(in.Format)
+	switch {
+	case inputDBFormats[f]:
+		if in.ConnEnv == "" {
+			e = append(e, fmt.Sprintf("input[%d]: format %q requires conn_env (an environment variable name)", i, in.Format))
+		}
+		if in.Path != "" {
+			e = append(e, fmt.Sprintf("input[%d]: format %q takes conn_env, not path", i, in.Format))
+		}
+	case readerFormats[f]:
+		if in.Path == "" {
+			e = append(e, fmt.Sprintf("input[%d]: missing path", i))
+		}
+		if in.ConnEnv != "" {
+			e = append(e, fmt.Sprintf("input[%d]: format %q does not use conn_env", i, in.Format))
+		}
+	default:
+		e = append(e, fmt.Sprintf("input[%d]: unsupported input format %q", i, in.Format))
+	}
+	if looksLikeSecret(in.ConnEnv) {
+		e = append(e, fmt.Sprintf("input[%d]: conn_env must be an environment variable name, not a connection string", i))
+	}
+	return e
+}
+
+func validateOutput(o Output) []string {
+	var e []string
+	if o.Format == "" {
+		e = append(e, "output: missing format")
+		return e
+	}
+	f := strings.ToLower(o.Format)
+	if !writerFormats[f] {
+		e = append(e, fmt.Sprintf("output: unsupported output format %q", o.Format))
+		return e
+	}
+	if o.ConnEnv != "" {
+		if !execOutputFormats[f] {
+			e = append(e, fmt.Sprintf("output: conn_env (live database execution) is not supported for format %q", o.Format))
+		}
+		if o.Path != "" {
+			e = append(e, "output: set either path or conn_env, not both")
+		}
+	} else if o.Path == "" {
+		e = append(e, "output: missing path")
+	}
+	if looksLikeSecret(o.ConnEnv) {
+		e = append(e, "output: conn_env must be an environment variable name, not a connection string")
+	}
+	return e
+}
+
+// looksLikeSecret reports whether s looks like a connection string rather
+// than a bare environment-variable name.
+func looksLikeSecret(s string) bool {
+	if s == "" {
+		return false
+	}
+	return strings.ContainsAny(s, ":/@ =") || strings.Contains(s, "//")
+}
+
+// checkRelPath rejects absolute paths and any path that escapes its root.
+func checkRelPath(p string) error {
+	if p == "" {
+		return fmt.Errorf("empty path")
+	}
+	if filepath.IsAbs(p) {
+		return fmt.Errorf("absolute paths are not allowed; use a path relative to the job file")
+	}
+	if strings.HasPrefix(p, "~") {
+		return fmt.Errorf("home-relative paths are not allowed")
+	}
+	clean := filepath.ToSlash(filepath.Clean(p))
+	if clean == ".." || strings.HasPrefix(clean, "../") {
+		return fmt.Errorf("path escapes the job file directory")
+	}
+	return nil
+}
+
+// SafeJoin resolves rel against root and guarantees the result stays inside
+// root. It is the single choke point for turning a manifest path into a
+// filesystem path.
+func SafeJoin(root, rel string) (string, error) {
+	if err := checkRelPath(rel); err != nil {
+		return "", err
+	}
+	absRoot, err := filepath.Abs(root)
+	if err != nil {
+		return "", err
+	}
+	joined := filepath.Join(absRoot, rel)
+	rp, err := filepath.Rel(absRoot, joined)
+	if err != nil {
+		return "", err
+	}
+	if rp == ".." || strings.HasPrefix(rp, ".."+string(filepath.Separator)) {
+		return "", fmt.Errorf("path %q escapes the job file directory", rel)
+	}
+	return joined, nil
+}
+
+// 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).
+func (s *Set) Plan(name string, includeDeps bool) ([]*Job, error) {
+	root, ok := s.Jobs[name]
+	if !ok {
+		return nil, fmt.Errorf("unknown job %q (known: %s)", name, strings.Join(s.Names(), ", "))
+	}
+	if !includeDeps {
+		return []*Job{root}, nil
+	}
+	var order []*Job
+	visited := map[string]bool{}
+	inProgress := map[string]bool{}
+	var visit func(n string) error
+	visit = func(n string) error {
+		if visited[n] {
+			return nil
+		}
+		if inProgress[n] {
+			return fmt.Errorf("dependency cycle at job %q", n)
+		}
+		inProgress[n] = true
+		j := s.Jobs[n]
+		deps := append([]string(nil), j.DependsOn...)
+		sort.Strings(deps)
+		for _, d := range deps {
+			if _, ok := s.Jobs[d]; !ok {
+				return fmt.Errorf("job %q depends on unknown job %q", n, d)
+			}
+			if err := visit(d); err != nil {
+				return err
+			}
+		}
+		inProgress[n] = false
+		visited[n] = true
+		order = append(order, j)
+		return nil
+	}
+	if err := visit(name); err != nil {
+		return nil, err
+	}
+	return order, nil
+}
+
+// findCycle returns a human-readable cycle path, or "" if the graph is acyclic.
+func (s *Set) findCycle() string {
+	color := map[string]int{} // 0 unvisited, 1 in progress, 2 done
+	var stack []string
+	var dfs func(n string) []string
+	dfs = func(n string) []string {
+		color[n] = 1
+		stack = append(stack, n)
+		deps := append([]string(nil), s.Jobs[n].DependsOn...)
+		sort.Strings(deps)
+		for _, d := range deps {
+			if _, ok := s.Jobs[d]; !ok {
+				continue
+			}
+			switch color[d] {
+			case 0:
+				if c := dfs(d); c != nil {
+					return c
+				}
+			case 1:
+				// Found a back edge; build the cycle slice.
+				for i, x := range stack {
+					if x == d {
+						return append(append([]string(nil), stack[i:]...), d)
+					}
+				}
+				return []string{d, d}
+			}
+		}
+		stack = stack[:len(stack)-1]
+		color[n] = 2
+		return nil
+	}
+	for _, n := range s.Names() {
+		if color[n] == 0 {
+			if c := dfs(n); c != nil {
+				return strings.Join(c, " -> ")
+			}
+		}
+	}
+	return ""
+}

pkg/jobs/jobs_test.go

@@ -0,0 +1,251 @@
+package jobs
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+func write(t *testing.T, path, content string) {
+	t.Helper()
+	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(path, []byte(content), 0o644); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func TestDiscoverDeterministicOrder(t *testing.T) {
+	dir := t.TempDir()
+	for _, n := range []string{
+		"relspec.yml", "relspec.zeta.yml", "relspec.alpha.yaml",
+		"relspec.beta.yml", "notes.yml", "relspec.txt",
+	} {
+		write(t, filepath.Join(dir, n), "version: 1\njobs: {}\n")
+	}
+	got, err := Discover(dir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var bases []string
+	for _, p := range got {
+		bases = append(bases, filepath.Base(p))
+	}
+	want := []string{"relspec.yml", "relspec.alpha.yaml", "relspec.beta.yml", "relspec.zeta.yml"}
+	if strings.Join(bases, ",") != strings.Join(want, ",") {
+		t.Fatalf("discover order = %v, want %v", bases, want)
+	}
+
+	// Second call must return the identical order.
+	got2, _ := Discover(dir)
+	for i := range got {
+		if got[i] != got2[i] {
+			t.Fatalf("discover not deterministic: %v vs %v", got, got2)
+		}
+	}
+}
+
+func TestLoadRejectsUnknownFields(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 error for unknown field")
+	}
+}
+
+func TestLoadRejectsBadVersion(t *testing.T) {
+	dir := t.TempDir()
+	p := filepath.Join(dir, "relspec.yml")
+	write(t, p, "version: 2\njobs:\n  a:\n    command: convert\n")
+	_, err := Load([]string{p})
+	if err == nil || !strings.Contains(err.Error(), "unsupported version") {
+		t.Fatalf("expected unsupported version error, got %v", err)
+	}
+}
+
+func TestLoadRejectsDuplicateJobAcrossFiles(t *testing.T) {
+	dir := t.TempDir()
+	a := filepath.Join(dir, "relspec.yml")
+	b := filepath.Join(dir, "relspec.extra.yml")
+	write(t, a, jobFileConvert("build"))
+	write(t, b, jobFileConvert("build"))
+	_, err := Load([]string{a, b})
+	if err == nil || !strings.Contains(err.Error(), "duplicate job") {
+		t.Fatalf("expected duplicate job error, got %v", err)
+	}
+}
+
+func jobFileConvert(name string) string {
+	return "version: 1\njobs:\n  " + name + ":\n    command: convert\n" +
+		"    inputs:\n      - path: a.dbml\n        format: dbml\n" +
+		"    output:\n      format: json\n      path: out.json\n"
+}
+
+func loadOne(t *testing.T, content string) *Set {
+	t.Helper()
+	dir := t.TempDir()
+	p := filepath.Join(dir, "relspec.yml")
+	write(t, p, content)
+	set, err := Load([]string{p})
+	if err != nil {
+		t.Fatalf("load: %v", err)
+	}
+	return set
+}
+
+func TestValidateUnknownCommand(t *testing.T) {
+	set := loadOne(t, "version: 1\njobs:\n  x:\n    command: rm-rf\n")
+	err := set.Validate()
+	if err == nil || !strings.Contains(err.Error(), "unsupported command") {
+		t.Fatalf("want unsupported command, got %v", err)
+	}
+}
+
+func TestValidateShellStringCommandRejected(t *testing.T) {
+	set := loadOne(t, "version: 1\njobs:\n  x:\n    command: \"bash -c 'echo hi'\"\n")
+	if err := set.Validate(); err == nil {
+		t.Fatal("expected arbitrary shell command to be rejected")
+	}
+}
+
+func TestValidateMissingInputs(t *testing.T) {
+	set := loadOne(t, "version: 1\njobs:\n  x:\n    command: convert\n    output:\n      format: json\n      path: o.json\n")
+	if err := set.Validate(); err == nil || !strings.Contains(err.Error(), "at least 1 input") {
+		t.Fatalf("want missing input error, got %v", err)
+	}
+}
+
+func TestValidateUnknownFormat(t *testing.T) {
+	set := loadOne(t, "version: 1\njobs:\n  x:\n    command: convert\n"+
+		"    inputs:\n      - path: a.xyz\n        format: xyz\n"+
+		"    output:\n      format: json\n      path: o.json\n")
+	if err := set.Validate(); err == nil || !strings.Contains(err.Error(), "unsupported input format") {
+		t.Fatalf("want unsupported input format, got %v", err)
+	}
+}
+
+func TestValidatePathTraversalRejected(t *testing.T) {
+	cases := []string{"../secret.dbml", "/etc/passwd", "~/x.dbml", "a/../../b.dbml"}
+	for _, bad := range cases {
+		set := loadOne(t, "version: 1\njobs:\n  x:\n    command: convert\n"+
+			"    inputs:\n      - path: \""+bad+"\"\n        format: dbml\n"+
+			"    output:\n      format: json\n      path: o.json\n")
+		if err := set.Validate(); err == nil {
+			t.Fatalf("path %q: expected rejection", bad)
+		}
+	}
+}
+
+func TestValidateOutputTraversalRejected(t *testing.T) {
+	set := loadOne(t, "version: 1\njobs:\n  x:\n    command: convert\n"+
+		"    inputs:\n      - path: a.dbml\n        format: dbml\n"+
+		"    output:\n      format: json\n      path: ../../evil.json\n")
+	if err := set.Validate(); err == nil {
+		t.Fatal("expected output path traversal rejection")
+	}
+}
+
+func TestValidateConnEnvMustBeName(t *testing.T) {
+	set := loadOne(t, "version: 1\njobs:\n  x:\n    command: convert\n"+
+		"    inputs:\n      - format: pgsql\n        conn_env: \"postgres://u:p@h/db\"\n"+
+		"    output:\n      format: json\n      path: o.json\n")
+	if err := set.Validate(); err == nil || !strings.Contains(err.Error(), "environment variable name") {
+		t.Fatalf("want conn_env name error, got %v", err)
+	}
+}
+
+func TestValidateDependsOnUnknown(t *testing.T) {
+	set := loadOne(t, "version: 1\njobs:\n  x:\n    command: convert\n    depends_on: [nope]\n"+
+		"    inputs:\n      - path: a.dbml\n        format: dbml\n"+
+		"    output:\n      format: json\n      path: o.json\n")
+	if err := set.Validate(); err == nil || !strings.Contains(err.Error(), "unknown job") {
+		t.Fatalf("want unknown dependency error, got %v", err)
+	}
+}
+
+func TestValidateDependencyCycle(t *testing.T) {
+	content := "version: 1\njobs:\n" +
+		jobBlock("a", "b") + jobBlock("b", "c") + jobBlock("c", "a")
+	set := loadOne(t, content)
+	err := set.Validate()
+	if err == nil || !strings.Contains(err.Error(), "cycle") {
+		t.Fatalf("want cycle error, got %v", err)
+	}
+}
+
+func jobBlock(name, dep string) string {
+	return "  " + name + ":\n    command: convert\n    depends_on: [" + dep + "]\n" +
+		"    inputs:\n      - path: a.dbml\n        format: dbml\n" +
+		"    output:\n      format: json\n      path: " + name + ".json\n"
+}
+
+func TestPlanTopologicalOrder(t *testing.T) {
+	content := "version: 1\njobs:\n" +
+		"  base:\n    command: convert\n    inputs:\n      - path: a.dbml\n        format: dbml\n    output:\n      format: json\n      path: base.json\n" +
+		"  mid:\n    command: convert\n    depends_on: [base]\n    inputs:\n      - path: a.dbml\n        format: dbml\n    output:\n      format: json\n      path: mid.json\n" +
+		"  top:\n    command: convert\n    depends_on: [mid]\n    inputs:\n      - path: a.dbml\n        format: dbml\n    output:\n      format: json\n      path: top.json\n"
+	set := loadOne(t, content)
+	if err := set.Validate(); err != nil {
+		t.Fatalf("validate: %v", err)
+	}
+	plan, err := set.Plan("top", true)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var order []string
+	for _, j := range plan {
+		order = append(order, j.Name)
+	}
+	if strings.Join(order, ",") != "base,mid,top" {
+		t.Fatalf("plan order = %v, want [base mid top]", order)
+	}
+
+	solo, err := set.Plan("top", false)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(solo) != 1 || solo[0].Name != "top" {
+		t.Fatalf("no-deps plan = %v, want [top]", solo)
+	}
+}
+
+func TestSafeJoinStaysInsideRoot(t *testing.T) {
+	root := t.TempDir()
+	if _, err := SafeJoin(root, "sub/dir/file.sql"); err != nil {
+		t.Fatalf("expected ok, got %v", err)
+	}
+	if _, err := SafeJoin(root, "../escape"); err == nil {
+		t.Fatal("expected escape rejection")
+	}
+	if _, err := SafeJoin(root, "/abs"); err == nil {
+		t.Fatal("expected absolute rejection")
+	}
+}
+
+func TestShippedExampleIsValid(t *testing.T) {
+	path := filepath.Join("..", "..", "examples", "jobs", "relspec.yml")
+	set, err := Load([]string{path})
+	if err != nil {
+		t.Fatalf("load example: %v", err)
+	}
+	if err := set.Validate(); err != nil {
+		t.Fatalf("example manifest failed validation: %v", err)
+	}
+	if _, err := set.Plan("build-json", true); err != nil {
+		t.Fatalf("plan example: %v", err)
+	}
+}
+
+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") {
+		t.Fatalf("want script_dir required error, got %v", err)
+	}
+	set = loadOne(t, "version: 1\njobs:\n  s:\n    command: scripts-list\n    script_dirs: [migrations, extra]\n")
+	if err := set.Validate(); err != nil {
+		t.Fatalf("expected valid scripts-list job, got %v", err)
+	}
+}

pkg/pgsql/extensions.go

@@ -0,0 +1,459 @@
+package pgsql
+
+import (
+	"sort"
+	"strings"
+)
+
+// Extension describes a PostgreSQL extension RelSpec recognizes, along with the schema
+// artefacts that imply it: the types it provides (declared on TypeSpec.Extension), the
+// index access methods and operator classes it installs, and the functions whose use in a
+// default, check constraint, index predicate, or view body requires it.
+type Extension struct {
+	Name        string
+	Category    string
+	Description string
+
+	// Requires lists extensions that must be created before this one.
+	Requires []string
+
+	// IndexMethods are access methods usable as Index.Type.
+	IndexMethods []string
+
+	// OperatorClasses are operator classes the extension installs.
+	OperatorClasses []string
+
+	// Functions are function names whose use implies the extension.
+	Functions []string
+
+	// FunctionPrefixes match whole families of functions (e.g. "st_" for PostGIS).
+	FunctionPrefixes []string
+}
+
+// postgresExtensions is the set of extensions RelSpec knows how to detect and emit.
+var postgresExtensions = map[string]Extension{
+	"amcheck": {
+		Name: "amcheck", Category: "integrity",
+		Description: "Verifies B-tree and related structure consistency to help detect corruption.",
+		Functions:   []string{"bt_index_check", "bt_index_parent_check", "verify_heapam"},
+	},
+	"btree_gin": {
+		Name: "btree_gin", Category: "indexing",
+		Description: "Adds GIN operator classes for common scalar data types.",
+	},
+	"btree_gist": {
+		Name: "btree_gist", Category: "indexing",
+		Description: "Adds GiST operator classes for common scalar data types and exclusion constraints.",
+	},
+	"citext": {
+		Name: "citext", Category: "text",
+		Description: "Provides case-insensitive text columns and operators.",
+		Functions:   []string{"citext"},
+	},
+	"fuzzystrmatch": {
+		Name: "fuzzystrmatch", Category: "text",
+		Description: "Adds phonetic and fuzzy matching helpers like Soundex and Levenshtein.",
+		Functions: []string{
+			"soundex", "difference", "levenshtein", "levenshtein_less_equal",
+			"metaphone", "dmetaphone", "dmetaphone_alt",
+		},
+	},
+	"hstore": {
+		Name: "hstore", Category: "document",
+		Description:     "Adds a lightweight key/value data type for semi-structured attributes.",
+		OperatorClasses: []string{"gin_hstore_ops", "gist_hstore_ops", "hash_hstore_ops", "btree_hstore_ops"},
+		Functions: []string{
+			"hstore", "akeys", "avals", "skeys", "svals",
+			"hstore_to_json", "hstore_to_jsonb", "hstore_to_array", "hstore_to_matrix",
+		},
+	},
+	"http": {
+		Name: "http", Category: "integration",
+		Description: "Lets SQL functions make outbound HTTP requests.",
+		Functions: []string{
+			"http", "http_get", "http_post", "http_put", "http_patch", "http_delete",
+			"http_head", "urlencode",
+		},
+	},
+	"pg_background": {
+		Name: "pg_background", Category: "jobs",
+		Description: "Runs SQL asynchronously in PostgreSQL background workers.",
+		Functions:   []string{"pg_background_launch", "pg_background_result", "pg_background_detach"},
+	},
+	"pg_cron": {
+		Name: "pg_cron", Category: "scheduling",
+		Description:      "Schedules recurring SQL jobs inside PostgreSQL.",
+		FunctionPrefixes: []string{"cron."},
+	},
+	"pg_jsonschema": {
+		Name: "pg_jsonschema", Category: "validation",
+		Description: "Validates json and jsonb values against JSON Schema.",
+		Functions:   []string{"json_matches_schema", "jsonb_matches_schema", "jsonschema_is_valid"},
+	},
+	"pg_partman": {
+		Name: "pg_partman", Category: "partitioning",
+		Description:      "Automates time-based and serial-based partition management.",
+		FunctionPrefixes: []string{"partman."},
+	},
+	"pg_qualstats": {
+		Name: "pg_qualstats", Category: "observability",
+		Description: "Tracks predicate usage in WHERE and JOIN clauses for tuning and index advice.",
+	},
+	"pg_repack": {
+		Name: "pg_repack", Category: "maintenance",
+		Description: "Rebuilds bloated tables and indexes online with minimal locking.",
+	},
+	"pg_search": {
+		Name: "pg_search", Category: "search",
+		Description: "Provides ParadeDB full-text and relevance search features.",
+		// bm25 is also the access method name used by pg_textsearch; pg_search is the
+		// canonical provider, so a bm25 index resolves to it.
+		IndexMethods:     []string{"bm25"},
+		FunctionPrefixes: []string{"paradedb."},
+	},
+	"pg_stat_statements": {
+		Name: "pg_stat_statements", Category: "observability",
+		Description: "Tracks normalized query execution statistics.",
+	},
+	"pg_textsearch": {
+		Name: "pg_textsearch", Category: "search",
+		Description: "Adds BM25-style text search support.",
+	},
+	"pg_trgm": {
+		Name: "pg_trgm", Category: "text",
+		Description:     "Adds trigram similarity search and fast fuzzy matching indexes.",
+		OperatorClasses: []string{"gin_trgm_ops", "gist_trgm_ops"},
+		Functions: []string{
+			"similarity", "word_similarity", "strict_word_similarity",
+			"show_trgm", "show_limit", "set_limit",
+		},
+	},
+	"pgcrypto": {
+		Name: "pgcrypto", Category: "security",
+		Description: "Adds hashing, encryption, random bytes, and UUID helpers.",
+		// gen_random_uuid is deliberately absent: it is built in since PostgreSQL 13.
+		Functions: []string{
+			"crypt", "gen_salt", "gen_random_bytes", "digest", "hmac",
+			"pgp_sym_encrypt", "pgp_sym_decrypt", "pgp_pub_encrypt", "pgp_pub_decrypt",
+			"armor", "dearmor",
+		},
+	},
+	"pgrouting": {
+		Name: "pgrouting", Category: "geospatial",
+		Description:      "Adds routing and graph algorithms on top of PostGIS data.",
+		Requires:         []string{"postgis"},
+		FunctionPrefixes: []string{"pgr_"},
+	},
+	"pgstattuple": {
+		Name: "pgstattuple", Category: "maintenance",
+		Description: "Reports table and index tuple density and bloat information.",
+		Functions:   []string{"pgstattuple", "pgstatindex", "pgstatginindex", "pg_relpages"},
+	},
+	"plpython3u": {
+		Name: "plpython3u", Category: "procedural",
+		Description: "Lets you write PostgreSQL functions in Python 3.",
+	},
+	"postgis": {
+		Name: "postgis", Category: "geospatial",
+		Description:  "Adds spatial data types, functions, and indexes.",
+		IndexMethods: nil, // uses the built-in gist/spgist/brin access methods
+		OperatorClasses: []string{
+			"gist_geometry_ops_2d", "gist_geometry_ops_nd", "gist_geography_ops",
+			"spgist_geometry_ops_2d", "spgist_geometry_ops_3d", "spgist_geometry_ops_nd",
+			"brin_geometry_inclusion_ops_2d", "brin_geometry_inclusion_ops_3d",
+			"brin_geometry_inclusion_ops_4d", "brin_geography_inclusion_ops_2d",
+			"btree_geometry_ops", "btree_geography_ops",
+		},
+		FunctionPrefixes: []string{"st_"},
+		Functions: []string{
+			"geometrytype", "addgeometrycolumn", "dropgeometrycolumn", "updategeometrysrid",
+			"find_srid", "postgis_version", "postgis_full_version",
+		},
+	},
+	"postgis_raster": {
+		Name: "postgis_raster", Category: "geospatial",
+		Description: "Adds the raster type and raster analysis functions.",
+		Requires:    []string{"postgis"},
+	},
+	"postgis_topology": {
+		Name: "postgis_topology", Category: "geospatial",
+		Description:      "Adds topology-aware spatial models and validation tools.",
+		Requires:         []string{"postgis"},
+		FunctionPrefixes: []string{"topology."},
+	},
+	"postgres_fdw": {
+		Name: "postgres_fdw", Category: "federation",
+		Description: "Connects PostgreSQL tables to other PostgreSQL servers.",
+	},
+	"timescaledb": {
+		Name: "timescaledb", Category: "time-series",
+		Description: "Adds hypertables, compression, retention, and time-series optimizations.",
+		Functions: []string{
+			"create_hypertable", "add_dimension", "time_bucket", "time_bucket_gapfill",
+			"add_retention_policy", "add_compression_policy", "locf", "interpolate",
+		},
+	},
+	"unaccent": {
+		Name: "unaccent", Category: "text",
+		Description: "Removes accents and diacritics for normalized text search.",
+		Functions:   []string{"unaccent"},
+	},
+	"uuid-ossp": {
+		Name: "uuid-ossp", Category: "utility",
+		Description: "Generates UUIDs using several algorithms.",
+		Functions: []string{
+			"uuid_generate_v1", "uuid_generate_v1mc", "uuid_generate_v3",
+			"uuid_generate_v4", "uuid_generate_v5",
+			"uuid_nil", "uuid_ns_dns", "uuid_ns_url", "uuid_ns_oid", "uuid_ns_x500",
+		},
+	},
+	"vector": {
+		Name: "vector", Category: "ai/search",
+		Description:  "Adds vector data types and similarity search for embeddings.",
+		IndexMethods: []string{"hnsw", "ivfflat"},
+		OperatorClasses: []string{
+			"vector_l2_ops", "vector_ip_ops", "vector_cosine_ops", "vector_l1_ops",
+			"halfvec_l2_ops", "halfvec_ip_ops", "halfvec_cosine_ops", "halfvec_l1_ops",
+			"sparsevec_l2_ops", "sparsevec_ip_ops", "sparsevec_cosine_ops", "sparsevec_l1_ops",
+			"bit_hamming_ops", "bit_jaccard_ops",
+		},
+		Functions: []string{"l2_distance", "inner_product", "cosine_distance", "l1_distance", "vector_dims", "vector_norm"},
+	},
+	"vchord": {
+		Name: "vchord", Category: "ai/search",
+		Description:  "Adds VectorChord scalable disk-friendly vector indexes compatible with pgvector data types.",
+		Requires:     []string{"vector"},
+		IndexMethods: []string{"vchordrq", "vchordg"},
+	},
+	"ltree": {
+		Name: "ltree", Category: "document",
+		Description:     "Adds a hierarchical label tree type.",
+		OperatorClasses: []string{"gist_ltree_ops", "gin_ltree_ops", "gist__ltree_ops"},
+		Functions:       []string{"subltree", "subpath", "nlevel", "lca", "ltree2text", "text2ltree"},
+	},
+}
+
+// extensionIndexMethods maps an index access method to the extension providing it.
+var extensionIndexMethods = buildExtensionIndex(func(ext Extension) []string { return ext.IndexMethods })
+
+// extensionOperatorClasses maps an operator class to the extension providing it.
+var extensionOperatorClasses = buildExtensionIndex(func(ext Extension) []string { return ext.OperatorClasses })
+
+// extensionFunctions maps a function name to the extension providing it.
+var extensionFunctions = buildExtensionIndex(func(ext Extension) []string { return ext.Functions })
+
+// extensionFunctionPrefixes maps a function name prefix to the extension providing it.
+var extensionFunctionPrefixes = buildExtensionIndex(func(ext Extension) []string { return ext.FunctionPrefixes })
+
+func buildExtensionIndex(keys func(Extension) []string) map[string]string {
+	index := make(map[string]string)
+	for _, ext := range postgresExtensions {
+		for _, key := range keys(ext) {
+			// Deterministic on collision: the alphabetically first extension wins.
+			if existing, ok := index[key]; ok && existing < ext.Name {
+				continue
+			}
+			index[key] = ext.Name
+		}
+	}
+	return index
+}
+
+// LookupExtension returns the registered extension by name.
+func LookupExtension(name string) (Extension, bool) {
+	ext, ok := postgresExtensions[strings.ToLower(strings.TrimSpace(name))]
+	return ext, ok
+}
+
+// IsKnownExtension reports whether the named extension is registered.
+func IsKnownExtension(name string) bool {
+	_, ok := LookupExtension(name)
+	return ok
+}
+
+// GetExtensions returns every registered extension name, sorted.
+func GetExtensions() []string {
+	names := make([]string, 0, len(postgresExtensions))
+	for name := range postgresExtensions {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+	return names
+}
+
+// IndexMethodExtension returns the extension providing an index access method
+// ("hnsw" -> "vector", "vchordrq" -> "vchord"). Built-in methods return "".
+func IndexMethodExtension(method string) string {
+	return extensionIndexMethods[strings.ToLower(strings.TrimSpace(method))]
+}
+
+// OperatorClassExtension returns the extension providing an operator class
+// ("gin_trgm_ops" -> "pg_trgm"). Built-in operator classes return "".
+func OperatorClassExtension(opClass string) string {
+	return extensionOperatorClasses[strings.ToLower(strings.TrimSpace(opClass))]
+}
+
+// ExtensionsForExpression returns the extensions whose functions appear in a SQL
+// expression such as a column default, check constraint, index predicate, or view body.
+// The result is sorted and deduplicated.
+func ExtensionsForExpression(expression string) []string {
+	if strings.TrimSpace(expression) == "" {
+		return nil
+	}
+
+	lower := strings.ToLower(expression)
+	found := make(map[string]bool)
+
+	for _, call := range sqlFunctionCalls(lower) {
+		if ext, ok := extensionFunctions[call]; ok {
+			found[ext] = true
+			continue
+		}
+		for prefix, ext := range extensionFunctionPrefixes {
+			if strings.HasPrefix(call, prefix) {
+				found[ext] = true
+				break
+			}
+		}
+	}
+
+	if len(found) == 0 {
+		return nil
+	}
+
+	names := make([]string, 0, len(found))
+	for name := range found {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+	return names
+}
+
+// sqlFunctionCalls returns the lowercase names of every function call in an expression.
+// A call is an identifier (optionally schema-qualified) immediately followed by "(".
+func sqlFunctionCalls(lowerExpression string) []string {
+	calls := make([]string, 0, 4)
+	end := 0
+
+	for i := 0; i < len(lowerExpression); i++ {
+		if lowerExpression[i] != '(' {
+			continue
+		}
+
+		end = i
+		// Allow whitespace between the identifier and its opening parenthesis.
+		for end > 0 && isSQLSpace(lowerExpression[end-1]) {
+			end--
+		}
+
+		start := end
+		for start > 0 && isSQLIdentifierByte(lowerExpression[start-1]) {
+			start--
+		}
+		if start == end {
+			continue
+		}
+		// A leading digit means this is not an identifier (e.g. "2(").
+		if lowerExpression[start] >= '0' && lowerExpression[start] <= '9' {
+			continue
+		}
+		calls = append(calls, lowerExpression[start:end])
+	}
+
+	return calls
+}
+
+func isSQLIdentifierByte(b byte) bool {
+	switch {
+	case b >= 'a' && b <= 'z', b >= 'A' && b <= 'Z', b >= '0' && b <= '9':
+		return true
+	case b == '_', b == '.':
+		return true
+	default:
+		return false
+	}
+}
+
+func isSQLSpace(b byte) bool {
+	return b == ' ' || b == '\t' || b == '\n' || b == '\r'
+}
+
+// SortExtensions orders extension names so that dependencies come first (postgis before
+// postgis_topology, vector before vchord), with alphabetical order breaking ties.
+// Duplicates are removed; unknown names are kept and sorted alphabetically.
+func SortExtensions(names []string) []string {
+	unique := make(map[string]bool, len(names))
+	for _, name := range names {
+		name = strings.ToLower(strings.TrimSpace(name))
+		if name != "" {
+			unique[name] = true
+		}
+	}
+	if len(unique) == 0 {
+		return nil
+	}
+
+	pending := make([]string, 0, len(unique))
+	for name := range unique {
+		pending = append(pending, name)
+	}
+	sort.Strings(pending)
+
+	sorted := make([]string, 0, len(pending))
+	emitted := make(map[string]bool, len(pending))
+
+	var emit func(name string, seen map[string]bool)
+	emit = func(name string, seen map[string]bool) {
+		if emitted[name] || seen[name] {
+			return
+		}
+		seen[name] = true
+
+		if ext, ok := LookupExtension(name); ok {
+			for _, dependency := range ext.Requires {
+				// Only order dependencies that are actually being created.
+				if unique[dependency] {
+					emit(dependency, seen)
+				}
+			}
+		}
+
+		emitted[name] = true
+		sorted = append(sorted, name)
+	}
+
+	for _, name := range pending {
+		emit(name, make(map[string]bool))
+	}
+	return sorted
+}
+
+// ExtensionDependencies returns the extensions a given extension requires, sorted.
+func ExtensionDependencies(name string) []string {
+	ext, ok := LookupExtension(name)
+	if !ok || len(ext.Requires) == 0 {
+		return nil
+	}
+	requires := append([]string(nil), ext.Requires...)
+	sort.Strings(requires)
+	return requires
+}
+
+// QuoteExtensionName quotes an extension name when it is not a bare SQL identifier,
+// e.g. uuid-ossp -> "uuid-ossp".
+func QuoteExtensionName(name string) string {
+	name = strings.TrimSpace(name)
+	if name == "" {
+		return ""
+	}
+	for i := 0; i < len(name); i++ {
+		b := name[i]
+		switch {
+		case b >= 'a' && b <= 'z', b == '_':
+		case b >= '0' && b <= '9' && i > 0:
+		default:
+			return `"` + strings.ReplaceAll(name, `"`, `""`) + `"`
+		}
+	}
+	return name
+}

pkg/pgsql/extensions_test.go

@@ -0,0 +1,165 @@
+package pgsql
+
+import (
+	"reflect"
+	"strings"
+	"testing"
+)
+
+func TestExtensionRegistryConsistency(t *testing.T) {
+	for name, ext := range postgresExtensions {
+		if name != ext.Name {
+			t.Errorf("extension registered as %q has Name %q", name, ext.Name)
+		}
+		if name != strings.ToLower(name) {
+			t.Errorf("extension %q must be registered lowercase", name)
+		}
+		if ext.Description == "" || ext.Category == "" {
+			t.Errorf("extension %q is missing a category or description", name)
+		}
+		for _, dependency := range ext.Requires {
+			if !IsKnownExtension(dependency) {
+				t.Errorf("extension %q requires unregistered extension %q", name, dependency)
+			}
+		}
+	}
+}
+
+// Every extension named by a type in the type registry must itself be registered,
+// otherwise a column type would ask for a CREATE EXTENSION nothing knows how to order.
+func TestTypeExtensionsAreRegistered(t *testing.T) {
+	for typeName, spec := range postgresBaseTypes {
+		if spec.Extension == "" {
+			continue
+		}
+		if !IsKnownExtension(spec.Extension) {
+			t.Errorf("type %q declares unregistered extension %q", typeName, spec.Extension)
+		}
+	}
+}
+
+func TestIndexMethodExtension(t *testing.T) {
+	tests := map[string]string{
+		"hnsw":     "vector",
+		"ivfflat":  "vector",
+		"HNSW":     "vector",
+		"vchordrq": "vchord",
+		"vchordg":  "vchord",
+		"bm25":     "pg_search",
+		"btree":    "",
+		"gin":      "",
+		"":         "",
+	}
+
+	for method, want := range tests {
+		if got := IndexMethodExtension(method); got != want {
+			t.Errorf("IndexMethodExtension(%q) = %q, want %q", method, got, want)
+		}
+	}
+}
+
+func TestOperatorClassExtension(t *testing.T) {
+	tests := map[string]string{
+		"gin_trgm_ops":         "pg_trgm",
+		"gist_trgm_ops":        "pg_trgm",
+		"vector_cosine_ops":    "vector",
+		"halfvec_l2_ops":       "vector",
+		"gist_ltree_ops":       "ltree",
+		"gist_geometry_ops_2d": "postgis",
+		"jsonb_path_ops":       "",
+		"array_ops":            "",
+		"":                     "",
+	}
+
+	for opClass, want := range tests {
+		if got := OperatorClassExtension(opClass); got != want {
+			t.Errorf("OperatorClassExtension(%q) = %q, want %q", opClass, got, want)
+		}
+	}
+}
+
+func TestExtensionsForExpression(t *testing.T) {
+	tests := []struct {
+		name       string
+		expression string
+		want       []string
+	}{
+		{"empty", "", nil},
+		{"no functions", "status = 'active'", nil},
+		{"builtin only", "now()", nil},
+		{"uuid-ossp default", "uuid_generate_v4()", []string{"uuid-ossp"}},
+		{"gen_random_uuid is builtin", "gen_random_uuid()", nil},
+		{"pgcrypto", "crypt(password, gen_salt('bf'))", []string{"pgcrypto"}},
+		{"postgis prefix", "ST_Area(geom) > 0", []string{"postgis"}},
+		{"paradedb prefix", "paradedb.snippet(body)", []string{"pg_search"}},
+		{"jsonschema", "json_matches_schema('{}', payload)", []string{"pg_jsonschema"}},
+		{"whitespace before paren", "unaccent ('crème')", []string{"unaccent"}},
+		{"multiple sorted", "ST_X(geom) = levenshtein(a, b)::float", []string{"fuzzystrmatch", "postgis"}},
+		{"column named like function", "similarity_score > 0.5", nil},
+		{"numeric prefix ignored", "2(3)", nil},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := ExtensionsForExpression(tt.expression); !reflect.DeepEqual(got, tt.want) {
+				t.Errorf("ExtensionsForExpression(%q) = %v, want %v", tt.expression, got, tt.want)
+			}
+		})
+	}
+}
+
+func TestSortExtensions(t *testing.T) {
+	tests := []struct {
+		name  string
+		input []string
+		want  []string
+	}{
+		{"empty", nil, nil},
+		{"alphabetical", []string{"pg_trgm", "citext"}, []string{"citext", "pg_trgm"}},
+		{"deduplicated", []string{"vector", "vector", " VECTOR "}, []string{"vector"}},
+		{"dependency first", []string{"vchord", "vector"}, []string{"vector", "vchord"}},
+		{
+			"postgis dependants",
+			[]string{"postgis_topology", "pgrouting", "postgis"},
+			[]string{"postgis", "pgrouting", "postgis_topology"},
+		},
+		{"dependency not requested", []string{"vchord"}, []string{"vchord"}},
+		{"unknown names kept", []string{"zzz_custom", "citext"}, []string{"citext", "zzz_custom"}},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := SortExtensions(tt.input); !reflect.DeepEqual(got, tt.want) {
+				t.Errorf("SortExtensions(%v) = %v, want %v", tt.input, got, tt.want)
+			}
+		})
+	}
+}
+
+func TestQuoteExtensionName(t *testing.T) {
+	tests := map[string]string{
+		"vector":    "vector",
+		"pg_trgm":   "pg_trgm",
+		"uuid-ossp": `"uuid-ossp"`,
+		"PostGIS":   `"PostGIS"`,
+		"":          "",
+	}
+
+	for name, want := range tests {
+		if got := QuoteExtensionName(name); got != want {
+			t.Errorf("QuoteExtensionName(%q) = %q, want %q", name, got, want)
+		}
+	}
+}
+
+func TestExtensionDependencies(t *testing.T) {
+	if got := ExtensionDependencies("vchord"); !reflect.DeepEqual(got, []string{"vector"}) {
+		t.Errorf("ExtensionDependencies(vchord) = %v, want [vector]", got)
+	}
+	if got := ExtensionDependencies("citext"); got != nil {
+		t.Errorf("ExtensionDependencies(citext) = %v, want nil", got)
+	}
+	if got := ExtensionDependencies("not_an_extension"); got != nil {
+		t.Errorf("ExtensionDependencies(not_an_extension) = %v, want nil", got)
+	}
+}

pkg/pgsql/index_options.go

@@ -0,0 +1,248 @@
+package pgsql
+
+import (
+	"strconv"
+	"strings"
+)
+
+// Index access-method storage parameters, the WITH (...) clause of CREATE INDEX. RelSpec
+// carries them through the model in Index.Comment, so the parsing here is deliberately
+// strict: only well-formed "key = value" pairs survive, and comment prose is discarded.
+//
+// Value forms accepted:
+//   - bare tokens: lists=100, m=16, deduplicate_items=true
+//   - quoted strings: key_field='id' (pg_search bm25)
+//   - dollar-quoted blocks: options=$$ [build.internal] lists=[4096] $$ (vchord)
+
+// ExtractWithClause returns the contents of the first WITH (...) clause in s, without the
+// surrounding parentheses. Parentheses inside quoted and dollar-quoted values are ignored,
+// so a vchord TOML block survives intact. Returns "" when there is no WITH clause.
+func ExtractWithClause(s string) string {
+	lower := strings.ToLower(s)
+
+	for offset := 0; ; {
+		idx := strings.Index(lower[offset:], "with")
+		if idx < 0 {
+			return ""
+		}
+		start := offset + idx
+		offset = start + 4
+
+		// "with" must stand as its own word.
+		if start > 0 && isSQLIdentifierByte(s[start-1]) {
+			continue
+		}
+
+		pos := offset
+		for pos < len(s) && isSQLSpace(s[pos]) {
+			pos++
+		}
+		if pos >= len(s) || s[pos] != '(' {
+			continue
+		}
+
+		if end, ok := matchClosingParen(s, pos); ok {
+			return s[pos+1 : end]
+		}
+		return ""
+	}
+}
+
+// matchClosingParen returns the index of the ')' matching the '(' at open, skipping over
+// quoted and dollar-quoted spans.
+func matchClosingParen(s string, open int) (int, bool) {
+	depth := 0
+	for i := open; i < len(s); i++ {
+		switch s[i] {
+		case '\'':
+			end, ok := skipQuoted(s, i)
+			if !ok {
+				return 0, false
+			}
+			i = end
+		case '$':
+			if end, ok := skipDollarQuoted(s, i); ok {
+				i = end
+			}
+		case '(':
+			depth++
+		case ')':
+			depth--
+			if depth == 0 {
+				return i, true
+			}
+		}
+	}
+	return 0, false
+}
+
+// skipQuoted returns the index of the closing quote of the single-quoted string starting
+// at start, treating ” as an escaped quote.
+func skipQuoted(s string, start int) (int, bool) {
+	for i := start + 1; i < len(s); i++ {
+		if s[i] != '\'' {
+			continue
+		}
+		if i+1 < len(s) && s[i+1] == '\'' {
+			i++
+			continue
+		}
+		return i, true
+	}
+	return 0, false
+}
+
+// skipDollarQuoted returns the index of the last byte of the dollar-quoted block starting
+// at start ($tag$ … $tag$). Reports false when start does not open one.
+func skipDollarQuoted(s string, start int) (int, bool) {
+	tagEnd := strings.IndexByte(s[start+1:], '$')
+	if tagEnd < 0 {
+		return 0, false
+	}
+	tag := s[start : start+1+tagEnd+1]
+	for i := start + 1; i < len(tag); i++ {
+		if !isSQLIdentifierByte(tag[i]) && tag[i] != '$' {
+			return 0, false
+		}
+	}
+
+	closing := strings.Index(s[start+len(tag):], tag)
+	if closing < 0 {
+		return 0, false
+	}
+	return start + len(tag) + closing + len(tag) - 1, true
+}
+
+// SplitStorageParameters splits a WITH clause body on top-level commas, leaving quoted and
+// dollar-quoted values untouched.
+func SplitStorageParameters(clause string) []string {
+	parts := make([]string, 0, 4)
+	depth := 0
+	start := 0
+
+	for i := 0; i < len(clause); i++ {
+		switch clause[i] {
+		case '\'':
+			if end, ok := skipQuoted(clause, i); ok {
+				i = end
+			}
+		case '$':
+			if end, ok := skipDollarQuoted(clause, i); ok {
+				i = end
+			}
+		case '(', '[':
+			depth++
+		case ')', ']':
+			depth--
+		case ',':
+			if depth == 0 {
+				parts = append(parts, clause[start:i])
+				start = i + 1
+			}
+		}
+	}
+	parts = append(parts, clause[start:])
+
+	trimmed := make([]string, 0, len(parts))
+	for _, part := range parts {
+		if part = strings.TrimSpace(part); part != "" {
+			trimmed = append(trimmed, part)
+		}
+	}
+	return trimmed
+}
+
+// ParseStorageParameter splits one "key = value" storage parameter. It reports false for
+// anything that is not a well-formed parameter, which is how comment prose is filtered out.
+func ParseStorageParameter(part string) (key, value string, ok bool) {
+	key, value, found := strings.Cut(part, "=")
+	if !found {
+		return "", "", false
+	}
+
+	key = strings.ToLower(strings.TrimSpace(key))
+	value = strings.TrimSpace(value)
+	if key == "" || value == "" || !isBareIdentifier(key) {
+		return "", "", false
+	}
+	if !isStorageParameterValue(value) {
+		return "", "", false
+	}
+	return key, value, true
+}
+
+// FormatStorageParameters renders a WITH clause body as a canonical "key = value" list,
+// dropping anything malformed. Returns "" when nothing survives.
+func FormatStorageParameters(clause string) string {
+	params := make([]string, 0, 4)
+	for _, part := range SplitStorageParameters(clause) {
+		key, value, ok := ParseStorageParameter(part)
+		if !ok {
+			continue
+		}
+		params = append(params, key+" = "+value)
+	}
+	return strings.Join(params, ", ")
+}
+
+// NormalizeStorageParameterValue unquotes a value that PostgreSQL rendered as a string but
+// that is really a number, so that pg_indexes output (lists='100') and hand-written models
+// (lists=100) normalize identically. Non-numeric quoted values keep their quotes because
+// some access methods require a string (pg_search's key_field='id').
+func NormalizeStorageParameterValue(value string) string {
+	value = strings.TrimSpace(value)
+	if len(value) < 2 || value[0] != '\'' || value[len(value)-1] != '\'' {
+		return value
+	}
+
+	inner := strings.ReplaceAll(value[1:len(value)-1], "''", "'")
+	if _, err := strconv.ParseFloat(inner, 64); err == nil {
+		return inner
+	}
+	if strings.EqualFold(inner, "true") || strings.EqualFold(inner, "false") {
+		return strings.ToLower(inner)
+	}
+	return value
+}
+
+func isBareIdentifier(s string) bool {
+	if s == "" {
+		return false
+	}
+	for i := 0; i < len(s); i++ {
+		b := s[i]
+		switch {
+		case b >= 'a' && b <= 'z', b >= 'A' && b <= 'Z', b == '_':
+		case b >= '0' && b <= '9' && i > 0:
+		default:
+			return false
+		}
+	}
+	return true
+}
+
+// isStorageParameterValue reports whether value is a bare token, a complete quoted string,
+// or a complete dollar-quoted block.
+func isStorageParameterValue(value string) bool {
+	switch {
+	case value == "":
+		return false
+	case value[0] == '\'':
+		end, ok := skipQuoted(value, 0)
+		return ok && end == len(value)-1
+	case value[0] == '$':
+		end, ok := skipDollarQuoted(value, 0)
+		return ok && end == len(value)-1
+	}
+
+	for i := 0; i < len(value); i++ {
+		b := value[i]
+		switch {
+		case b >= 'a' && b <= 'z', b >= 'A' && b <= 'Z', b >= '0' && b <= '9':
+		case b == '_', b == '.', b == '-', b == '+':
+		default:
+			return false
+		}
+	}
+	return true
+}

pkg/pgsql/index_options_test.go

@@ -0,0 +1,111 @@
+package pgsql
+
+import (
+	"reflect"
+	"testing"
+)
+
+func TestExtractWithClause(t *testing.T) {
+	tests := []struct {
+		name  string
+		input string
+		want  string
+	}{
+		{"empty", "", ""},
+		{"no clause", "opclass=vector_cosine_ops", ""},
+		{"simple", "WITH (lists=100)", "lists=100"},
+		{"lowercase", "with (m = 16, ef_construction = 64)", "m = 16, ef_construction = 64"},
+		{
+			"index definition",
+			"CREATE INDEX i ON t USING ivfflat (embedding vector_cosine_ops) WITH (lists='100')",
+			"lists='100'",
+		},
+		{"paren inside quotes", "with (key_field='id(x)')", "key_field='id(x)'"},
+		{"dollar quoted", "with (options = $$f(x)$$)", "options = $$f(x)$$"},
+		{"word boundary", "swith (lists=100)", ""},
+		{"not followed by paren", "with lists=100", ""},
+		{"unterminated", "with (lists=100", ""},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := ExtractWithClause(tt.input); got != tt.want {
+				t.Errorf("ExtractWithClause(%q) = %q, want %q", tt.input, got, tt.want)
+			}
+		})
+	}
+}
+
+func TestSplitStorageParameters(t *testing.T) {
+	tests := []struct {
+		name  string
+		input string
+		want  []string
+	}{
+		{"empty", "", []string{}},
+		{"single", "lists=100", []string{"lists=100"}},
+		{"multiple", "m = 16, ef_construction = 64", []string{"m = 16", "ef_construction = 64"}},
+		{"comma in quotes", "key_field='a,b', m=16", []string{"key_field='a,b'", "m=16"}},
+		{"comma in dollar quotes", "options=$$a,b$$, m=16", []string{"options=$$a,b$$", "m=16"}},
+		{"comma in brackets", "options=[1,2], m=16", []string{"options=[1,2]", "m=16"}},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := SplitStorageParameters(tt.input); !reflect.DeepEqual(got, tt.want) {
+				t.Errorf("SplitStorageParameters(%q) = %v, want %v", tt.input, got, tt.want)
+			}
+		})
+	}
+}
+
+func TestParseStorageParameter(t *testing.T) {
+	tests := []struct {
+		name      string
+		input     string
+		wantKey   string
+		wantValue string
+		wantOK    bool
+	}{
+		{"bare", "lists=100", "lists", "100", true},
+		{"spaced and uppercased key", " Lists = 100 ", "lists", "100", true},
+		{"quoted", "key_field='id'", "key_field", "'id'", true},
+		{"dollar quoted", "options=$$a$$", "options", "$$a$$", true},
+		{"boolean", "deduplicate_items=true", "deduplicate_items", "true", true},
+		{"float", "fillfactor=90.5", "fillfactor", "90.5", true},
+		{"no equals", "please drop everything", "", "", false},
+		{"empty value", "lists=", "", "", false},
+		{"quoted key rejected", "'lists'=100", "", "", false},
+		{"injection rejected", "lists=100); drop table t", "", "", false},
+		{"unterminated quote rejected", "key_field='id", "", "", false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			key, value, ok := ParseStorageParameter(tt.input)
+			if key != tt.wantKey || value != tt.wantValue || ok != tt.wantOK {
+				t.Errorf("ParseStorageParameter(%q) = (%q, %q, %v), want (%q, %q, %v)",
+					tt.input, key, value, ok, tt.wantKey, tt.wantValue, tt.wantOK)
+			}
+		})
+	}
+}
+
+func TestNormalizeStorageParameterValue(t *testing.T) {
+	tests := map[string]string{
+		"'100'":   "100",
+		"'90.5'":  "90.5",
+		"'true'":  "true",
+		"'id'":    "'id'",
+		"100":     "100",
+		"$$a,b$$": "$$a,b$$",
+		"'":       "'",
+		"''":      "''",
+	}
+
+	for input, want := range tests {
+		if got := NormalizeStorageParameterValue(input); got != want {
+			t.Errorf("NormalizeStorageParameterValue(%q) = %q, want %q", input, got, want)
+		}
+	}
+}

pkg/pgsql/types_registry.go

@@ -2,6 +2,7 @@ package pgsql
 
 import (
 	"sort"
+	"strconv"
 	"strings"
 )
 
@@ -9,6 +10,14 @@ import (
 type TypeSpec struct {
 	SupportsLength    bool
 	SupportsPrecision bool
+
+	// SupportsTypeModifier marks types whose "(...)" modifier is opaque and must be
+	// preserved verbatim (e.g. vector(1536), geometry(Point,4326)) instead of being
+	// decomposed into Length/Precision/Scale.
+	SupportsTypeModifier bool
+
+	// Extension is the PostgreSQL extension providing the type; empty for built-ins.
+	Extension string
 }
 
 var postgresBaseTypes = map[string]TypeSpec{
@@ -104,14 +113,28 @@ var postgresBaseTypes = map[string]TypeSpec{
 	"void":   {},
 
 	// Common extensions
-	"citext":    {},
-	"hstore":    {},
-	"ltree":     {},
-	"lquery":    {},
-	"ltxtquery": {},
-	"vector":    {}, // pgvector: keep explicit modifier form (vector(dim))
-	"halfvec":   {}, // pgvector: keep explicit modifier form (halfvec(dim))
-	"sparsevec": {}, // pgvector: keep explicit modifier form (sparsevec(dim))
+	"citext":    {Extension: "citext"},
+	"hstore":    {Extension: "hstore"},
+	"ltree":     {Extension: "ltree"},
+	"lquery":    {Extension: "ltree"},
+	"ltxtquery": {Extension: "ltree"},
+
+	// pgvector: modifier form is opaque (vector(dim), sparsevec(dim))
+	"vector":    {SupportsTypeModifier: true, Extension: "vector"},
+	"halfvec":   {SupportsTypeModifier: true, Extension: "vector"},
+	"sparsevec": {SupportsTypeModifier: true, Extension: "vector"},
+
+	// PostGIS: geometry/geography carry an opaque modifier (geometry(PointZ,4326))
+	"geometry":      {SupportsTypeModifier: true, Extension: "postgis"},
+	"geography":     {SupportsTypeModifier: true, Extension: "postgis"},
+	"box2d":         {Extension: "postgis"},
+	"box3d":         {Extension: "postgis"},
+	"geometry_dump": {Extension: "postgis"},
+	"geomval":       {Extension: "postgis"},
+	"spheroid":      {Extension: "postgis"},
+	"valid_detail":  {Extension: "postgis"},
+	"raster":        {SupportsTypeModifier: true, Extension: "postgis_raster"},
+	"topogeometry":  {Extension: "postgis_topology"},
 }
 
 var postgresTypeAliases = map[string]string{
@@ -346,3 +369,72 @@ func stripArraySuffixes(t string) string {
 func normalizeTypeToken(t string) string {
 	return strings.Join(strings.Fields(strings.TrimSpace(t)), " ")
 }
+
+// SupportsTypeModifier reports if this SQL type carries an opaque "(...)" modifier
+// that must be preserved verbatim (e.g. vector(1536), geometry(Point,4326)).
+func SupportsTypeModifier(sqlType string) bool {
+	base := CanonicalizeBaseType(ExtractBaseTypeLower(sqlType))
+	spec, ok := postgresBaseTypes[base]
+	return ok && spec.SupportsTypeModifier
+}
+
+// TypeExtension returns the PostgreSQL extension providing the given type
+// ("postgis", "vector", "citext", …). Built-in types return "".
+func TypeExtension(sqlType string) string {
+	base := CanonicalizeBaseType(ExtractBaseTypeLower(sqlType))
+	return postgresBaseTypes[base].Extension
+}
+
+// IsSpatialType reports whether the type comes from PostGIS (geometry, geography,
+// raster, topogeometry, …).
+func IsSpatialType(sqlType string) bool {
+	return strings.HasPrefix(TypeExtension(sqlType), "postgis")
+}
+
+// IsVectorType reports whether the type comes from pgvector (vector, halfvec, sparsevec).
+func IsVectorType(sqlType string) bool {
+	return TypeExtension(sqlType) == "vector"
+}
+
+// TypeModifier returns the raw "(...)" modifier of a SQL type without the parentheses,
+// or "" when the type has none. Array suffixes are ignored.
+// Example: geometry(PointZ,4326)[] -> "PointZ,4326".
+func TypeModifier(sqlType string) string {
+	t := stripArraySuffixes(normalizeTypeToken(sqlType))
+	start := strings.Index(t, "(")
+	end := strings.LastIndex(t, ")")
+	if start < 0 || end < start {
+		return ""
+	}
+	return strings.TrimSpace(t[start+1 : end])
+}
+
+// SpatialSRID returns the SRID declared in a PostGIS type modifier, or 0 when absent.
+// Example: geometry(Point,4326) -> 4326.
+func SpatialSRID(sqlType string) int {
+	if !IsSpatialType(sqlType) {
+		return 0
+	}
+	parts := strings.Split(TypeModifier(sqlType), ",")
+	if len(parts) < 2 {
+		return 0
+	}
+	srid, err := strconv.Atoi(strings.TrimSpace(parts[len(parts)-1]))
+	if err != nil {
+		return 0
+	}
+	return srid
+}
+
+// SpatialGeometryType returns the geometry subtype declared in a PostGIS type modifier
+// ("Point", "MultiPolygonZ", …), or "" when absent.
+func SpatialGeometryType(sqlType string) string {
+	if !IsSpatialType(sqlType) {
+		return ""
+	}
+	modifier := TypeModifier(sqlType)
+	if modifier == "" {
+		return ""
+	}
+	return strings.TrimSpace(strings.Split(modifier, ",")[0])
+}

pkg/pgsql/types_registry_test.go

@@ -145,3 +145,104 @@ func TestEquivalentSQLTypeVariants(t *testing.T) {
 		})
 	}
 }
+
+func TestExtensionTypes(t *testing.T) {
+	tests := []struct {
+		name          string
+		sqlType       string
+		wantKnown     bool
+		wantExtension string
+		wantSpatial   bool
+		wantVector    bool
+		wantModifier  bool
+	}{
+		{"geometry with modifier", "geometry(Point,4326)", true, "postgis", true, false, true},
+		{"geography", "geography", true, "postgis", true, false, true},
+		{"geometry array", "geometry[]", true, "postgis", true, false, true},
+		{"box2d", "box2d", true, "postgis", true, false, false},
+		{"raster", "raster", true, "postgis_raster", true, false, true},
+		{"topogeometry", "topogeometry", true, "postgis_topology", true, false, false},
+		{"vector", "vector(1536)", true, "vector", false, true, true},
+		{"halfvec", "halfvec(768)", true, "vector", false, true, true},
+		{"sparsevec", "sparsevec(1000)", true, "vector", false, true, true},
+		{"citext", "citext", true, "citext", false, false, false},
+		{"builtin text", "text", true, "", false, false, false},
+		{"builtin point is not postgis", "point", true, "", false, false, false},
+		{"unknown type", "mytype", false, "", false, false, false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := IsKnownPostgresType(tt.sqlType); got != tt.wantKnown {
+				t.Errorf("IsKnownPostgresType(%q) = %v, want %v", tt.sqlType, got, tt.wantKnown)
+			}
+			if got := TypeExtension(tt.sqlType); got != tt.wantExtension {
+				t.Errorf("TypeExtension(%q) = %q, want %q", tt.sqlType, got, tt.wantExtension)
+			}
+			if got := IsSpatialType(tt.sqlType); got != tt.wantSpatial {
+				t.Errorf("IsSpatialType(%q) = %v, want %v", tt.sqlType, got, tt.wantSpatial)
+			}
+			if got := IsVectorType(tt.sqlType); got != tt.wantVector {
+				t.Errorf("IsVectorType(%q) = %v, want %v", tt.sqlType, got, tt.wantVector)
+			}
+			if got := SupportsTypeModifier(tt.sqlType); got != tt.wantModifier {
+				t.Errorf("SupportsTypeModifier(%q) = %v, want %v", tt.sqlType, got, tt.wantModifier)
+			}
+		})
+	}
+}
+
+func TestExtensionTypesDoNotSupportLengthOrPrecision(t *testing.T) {
+	for _, sqlType := range []string{"geometry(Point,4326)", "geography", "vector(1536)", "halfvec(768)"} {
+		if SupportsLength(sqlType) {
+			t.Errorf("SupportsLength(%q) = true, want false", sqlType)
+		}
+		if SupportsPrecision(sqlType) {
+			t.Errorf("SupportsPrecision(%q) = true, want false", sqlType)
+		}
+	}
+}
+
+func TestSpatialTypeModifier(t *testing.T) {
+	tests := []struct {
+		sqlType      string
+		wantModifier string
+		wantGeomType string
+		wantSRID     int
+	}{
+		{"geometry(Point,4326)", "Point,4326", "Point", 4326},
+		{"geometry(MultiPolygonZ, 3857)", "MultiPolygonZ, 3857", "MultiPolygonZ", 3857},
+		{"geography(Point)", "Point", "Point", 0},
+		{"geometry", "", "", 0},
+		{"geometry(Point,4326)[]", "Point,4326", "Point", 4326},
+		{"vector(1536)", "1536", "", 0},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.sqlType, func(t *testing.T) {
+			if got := TypeModifier(tt.sqlType); got != tt.wantModifier {
+				t.Errorf("TypeModifier() = %q, want %q", got, tt.wantModifier)
+			}
+			if got := SpatialGeometryType(tt.sqlType); got != tt.wantGeomType {
+				t.Errorf("SpatialGeometryType() = %q, want %q", got, tt.wantGeomType)
+			}
+			if got := SpatialSRID(tt.sqlType); got != tt.wantSRID {
+				t.Errorf("SpatialSRID() = %d, want %d", got, tt.wantSRID)
+			}
+		})
+	}
+}
+
+func TestNormalizeEquivalentSQLTypePreservesExtensionModifiers(t *testing.T) {
+	tests := map[string]string{
+		"geometry(Point,4326)": "geometry(Point,4326)",
+		"vector(1536)":         "vector(1536)",
+		"geography(Point)[]":   "geography(Point)[]",
+	}
+
+	for input, want := range tests {
+		if got := NormalizeEquivalentSQLType(input); got != want {
+			t.Errorf("NormalizeEquivalentSQLType(%q) = %q, want %q", input, got, want)
+		}
+	}
+}

pkg/readers/dbml/reader.go

@@ -571,6 +571,28 @@ func (r *Reader) parseDBML(content string) (*models.Database, error) {
 		}
 	}
 
+	// PostgreSQL readers derive relationships from foreign keys. Do the same
+	// for DBML refs so diffing equivalent schemas compares the same model.
+	for _, schema := range schemaMap {
+		for _, table := range schema.Tables {
+			for _, constraint := range table.Constraints {
+				if constraint.Type != models.ForeignKeyConstraint {
+					continue
+				}
+				name := fmt.Sprintf("%s_to_%s", table.Name, constraint.ReferencedTable)
+				relationship := models.InitRelationship(name, models.OneToMany)
+				relationship.FromTable = table.Name
+				relationship.FromSchema = table.Schema
+				relationship.FromColumns = append([]string(nil), constraint.Columns...)
+				relationship.ToTable = constraint.ReferencedTable
+				relationship.ToSchema = constraint.ReferencedSchema
+				relationship.ToColumns = append([]string(nil), constraint.ReferencedColumns...)
+				relationship.ForeignKey = constraint.Name
+				table.Relationships[name] = relationship
+			}
+		}
+	}
+
 	// Add schemas to database
 	for _, schema := range schemaMap {
 		db.Schemas = append(db.Schemas, schema)

pkg/readers/dbml/reader_test.go

@@ -863,6 +863,13 @@ func TestParseColumn_PostgresTypes(t *testing.T) {
 			wantName: "embedding",
 			wantType: "vector(1536)",
 		},
+		{
+			name:        "postgis geometry with type modifier",
+			line:        "location geometry(Point,4326) [not null]",
+			wantName:    "location",
+			wantType:    "geometry(Point,4326)",
+			wantNotNull: true,
+		},
 		{
 			name:     "multi word timestamp type",
 			line:     "published_at timestamp with time zone",

pkg/readers/pgsql/README.md

@@ -128,6 +128,27 @@ sessions so they are identifiable in `pg_stat_activity`. If you provide
 - Sequence properties
 - Associated tables
 
+## Extension Types (PostGIS, pgvector)
+
+- Extension column types keep their catalog-formatted form: `geometry(Point,4326)`,
+  `geography(Point)`, `vector(1536)`, `halfvec(768)`, `citext`, arrays included.
+- Built-in types are canonicalized and their dimensions moved to
+  `Column.Length` / `Precision` / `Scale`; extension modifiers stay in `Column.Type`.
+- Index access methods are read from the definition as-is: `gist`, `spgist`, `brin`, `hnsw`,
+  `ivfflat`, `vchordrq`, `vchordg`, `bm25`.
+- Operator class and `WITH (...)` parameters have no model field, so they are stored in
+  `Index.Comment` in the form the PostgreSQL writer reads back:
+
+  ```
+  opclass=vector_cosine_ops; with (m=16, ef_construction=64)
+  ```
+
+  Ordering modifiers (`DESC`, `NULLS LAST`, `COLLATE`) are not treated as operator classes.
+  Numeric parameter values are unquoted (`lists='100'` -> `lists=100`); string values keep
+  their quotes (`key_field='id'`), and dollar-quoted values are preserved whole.
+- Installed extensions are read from `pg_extension` into `schema.Metadata["extensions"]`
+  (only extensions RelSpec recognizes), so a read/write round-trip re-creates them.
+
 ## Notes
 
 - Requires PostgreSQL connection permissions

pkg/readers/pgsql/queries.go

@@ -5,6 +5,7 @@ import (
 	"strings"
 
 	"git.warky.dev/wdevs/relspecgo/pkg/models"
+	"git.warky.dev/wdevs/relspecgo/pkg/pgsql"
 )
 
 // querySchemas retrieves all non-system schemas from the database
@@ -46,6 +47,41 @@ func (r *Reader) querySchemas() ([]*models.Schema, error) {
 	return schemas, rows.Err()
 }
 
+// queryExtensions retrieves the extensions installed into a schema. Only extensions RelSpec
+// recognizes are kept, so a round-trip never emits a CREATE EXTENSION the writer cannot
+// order; plpgsql is not registered and is therefore skipped along with other built-ins.
+func (r *Reader) queryExtensions(schemaName string) ([]string, error) {
+	query := `
+		SELECT e.extname
+		FROM pg_extension e
+		JOIN pg_namespace n ON n.oid = e.extnamespace
+		WHERE n.nspname = $1
+		ORDER BY e.extname
+	`
+
+	rows, err := r.conn.Query(r.ctx, query, schemaName)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+
+	extensions := make([]string, 0)
+	for rows.Next() {
+		var name string
+		if err := rows.Scan(&name); err != nil {
+			return nil, err
+		}
+		if pgsql.IsKnownExtension(name) {
+			extensions = append(extensions, name)
+		}
+	}
+	if err := rows.Err(); err != nil {
+		return nil, err
+	}
+
+	return pgsql.SortExtensions(extensions), nil
+}
+
 // queryTables retrieves all tables for a given schema
 func (r *Reader) queryTables(schemaName string) ([]*models.Table, error) {
 	query := `
@@ -502,8 +538,13 @@ func (r *Reader) queryCheckConstraints(schemaName string) (map[string][]*models.
 		FROM information_schema.table_constraints tc
 		JOIN information_schema.check_constraints cc
 			ON tc.constraint_name = cc.constraint_name
+			AND cc.constraint_schema = tc.table_schema
+		JOIN pg_catalog.pg_constraint pc
+			ON pc.conname = tc.constraint_name
+			AND pc.connamespace = (SELECT oid FROM pg_namespace WHERE nspname = tc.table_schema)
 		WHERE tc.constraint_type = 'CHECK'
 		  AND tc.table_schema = $1
+		AND pc.contype = 'c'
 	`
 
 	rows, err := r.conn.Query(r.ctx, query, schemaName)
@@ -543,7 +584,12 @@ func (r *Reader) queryIndexes(schemaName string) (map[string][]*models.Index, er
 			indexname,
 			indexdef
 		FROM pg_indexes
+		JOIN pg_catalog.pg_class idx ON idx.relname = indexname
+		JOIN pg_catalog.pg_index i ON i.indexrelid = idx.oid
+		JOIN pg_catalog.pg_namespace idx_ns ON idx_ns.oid = idx.relnamespace
 		WHERE schemaname = $1
+		  AND idx_ns.nspname = schemaname
+		  AND NOT i.indisprimary
 		ORDER BY schemaname, tablename, indexname
 	`
 
@@ -597,6 +643,7 @@ func (r *Reader) parseIndexDefinition(indexName, tableName, schema, indexDef str
 	}
 
 	// Extract columns - pattern: (column1, column2, ...)
+	opClass := ""
 	columnsRegex := regexp.MustCompile(`\(([^)]+)\)`)
 	if matches := columnsRegex.FindStringSubmatch(indexDef); len(matches) > 1 {
 		columnsStr := matches[1]
@@ -604,8 +651,17 @@ func (r *Reader) parseIndexDefinition(indexName, tableName, schema, indexDef str
 		columnParts := strings.Split(columnsStr, ",")
 		for _, col := range columnParts {
 			col = strings.TrimSpace(col)
+			fields := strings.Fields(col)
+			if len(fields) == 0 {
+				continue
+			}
+			// Remember an explicit operator class (e.g. "embedding vector_cosine_ops")
+			// so the writer can reproduce it; ordering modifiers are not operator classes.
+			if opClass == "" && len(fields) > 1 {
+				opClass = extractIndexOperatorClass(fields[1:])
+			}
 			// Remove any ordering (ASC/DESC) or other modifiers
-			col = strings.Fields(col)[0]
+			col = fields[0]
 			// Remove parentheses if it's an expression
 			if !strings.Contains(col, "(") {
 				index.Columns = append(index.Columns, col)
@@ -613,6 +669,15 @@ func (r *Reader) parseIndexDefinition(indexName, tableName, schema, indexDef str
 		}
 	}
 
+	// Extract access method storage parameters, e.g. WITH (lists='100')
+	storageParams := normalizeIndexStorageParams(pgsql.ExtractWithClause(indexDef))
+
+	// Operator class and storage parameters have no dedicated model fields; carry them in
+	// the comment hint the PostgreSQL writer reads back.
+	if hint := buildIndexHint(opClass, storageParams); hint != "" && index.Comment == "" {
+		index.Comment = hint
+	}
+
 	// Extract WHERE clause for partial indexes
 	whereRegex := regexp.MustCompile(`WHERE\s+(.+)$`)
 	if matches := whereRegex.FindStringSubmatch(indexDef); len(matches) > 1 {
@@ -622,6 +687,52 @@ func (r *Reader) parseIndexDefinition(indexName, tableName, schema, indexDef str
 	return index, nil
 }
 
+// indexOrderingKeywords are column modifiers that are not operator classes.
+var indexOrderingKeywords = map[string]bool{
+	"asc": true, "desc": true, "nulls": true, "first": true, "last": true, "collate": true,
+}
+
+// extractIndexOperatorClass picks the operator class out of a column's trailing modifiers.
+// Returns "" when the modifiers are only ordering keywords.
+func extractIndexOperatorClass(modifiers []string) string {
+	for _, modifier := range modifiers {
+		lower := strings.ToLower(strings.TrimSpace(modifier))
+		if lower == "" || indexOrderingKeywords[lower] {
+			continue
+		}
+		return lower
+	}
+	return ""
+}
+
+// normalizeIndexStorageParams rewrites "m='16', ef_construction='64'" as "m=16,
+// ef_construction=64". Non-numeric values keep their quotes because some access methods
+// require a string literal (pg_search's key_field='id').
+func normalizeIndexStorageParams(params string) string {
+	normalized := make([]string, 0, 4)
+	for _, part := range pgsql.SplitStorageParameters(params) {
+		key, value, ok := pgsql.ParseStorageParameter(part)
+		if !ok {
+			continue
+		}
+		normalized = append(normalized, key+"="+pgsql.NormalizeStorageParameterValue(value))
+	}
+	return strings.Join(normalized, ", ")
+}
+
+// buildIndexHint renders the operator class and storage parameters in the form the
+// PostgreSQL writer parses back out of an index comment.
+func buildIndexHint(opClass, storageParams string) string {
+	parts := make([]string, 0, 2)
+	if opClass != "" {
+		parts = append(parts, "opclass="+opClass)
+	}
+	if storageParams != "" {
+		parts = append(parts, "with ("+storageParams+")")
+	}
+	return strings.Join(parts, "; ")
+}
+
 // normalizePostgresDefault converts a raw PostgreSQL column_default expression into the
 // unquoted string value that the model convention expects. PostgreSQL stores string
 // literal defaults as 'value' or 'value'::type (e.g. '{}'::text[]), while every other

pkg/readers/pgsql/reader.go

@@ -88,6 +88,18 @@ func (r *Reader) ReadDatabase() (*models.Database, error) {
 		}
 		schema.Sequences = sequences
 
+		// Query extensions installed into this schema
+		extensions, err := r.queryExtensions(schema.Name)
+		if err != nil {
+			return nil, fmt.Errorf("failed to query extensions for schema %s: %w", schema.Name, err)
+		}
+		if len(extensions) > 0 {
+			if schema.Metadata == nil {
+				schema.Metadata = make(map[string]any)
+			}
+			schema.Metadata["extensions"] = extensions
+		}
+
 		// Query columns for tables and views
 		columnsMap, err := r.queryColumns(schema.Name)
 		if err != nil {
@@ -278,11 +290,6 @@ func (r *Reader) mapDataType(pgType, udtName, formattedType string, hasNextval b
 		}
 	}
 
-	// information_schema reports arrays generically as "ARRAY" with udt_name like "_text".
-	if strings.EqualFold(pgType, "ARRAY") && strings.HasPrefix(udtName, "_") && len(udtName) > 1 {
-		return udtName[1:] + "[]"
-	}
-
 	// Use the database-formatted type when available. For known built-in types, strip
 	// embedded dimensions (they are stored in column.Length/Precision/Scale separately).
 	// For unknown/custom types, keep the full formatted string (e.g. vector(1536)).
@@ -303,6 +310,13 @@ func (r *Reader) mapDataType(pgType, udtName, formattedType string, hasNextval b
 		return formattedType
 	}
 
+	// information_schema reports arrays generically as "ARRAY" with udt_name like "_text".
+	// Only reached when the catalog-formatted type is unavailable, which is the one case
+	// where the element modifier (e.g. geometry(Point,4326)[]) cannot be recovered.
+	if strings.EqualFold(pgType, "ARRAY") && strings.HasPrefix(udtName, "_") && len(udtName) > 1 {
+		return udtName[1:] + "[]"
+	}
+
 	// Fall back to normalizing the information_schema type name directly.
 	canonical := pgsql.NormalizePGType(normalizedPGType)
 	if pgsql.IsKnownPGBaseType(canonical) {
@@ -327,8 +341,10 @@ func (r *Reader) deriveRelationship(table *models.Table, fk *models.Constraint)
 	relationship := models.InitRelationship(relationshipName, models.OneToMany)
 	relationship.FromTable = table.Name
 	relationship.FromSchema = table.Schema
+	relationship.FromColumns = append([]string(nil), fk.Columns...)
 	relationship.ToTable = fk.ReferencedTable
 	relationship.ToSchema = fk.ReferencedSchema
+	relationship.ToColumns = append([]string(nil), fk.ReferencedColumns...)
 	relationship.ForeignKey = fk.Name
 
 	// Store constraint actions in properties

pkg/readers/pgsql/reader_test.go

@@ -2,6 +2,7 @@ package pgsql
 
 import (
 	"os"
+	"reflect"
 	"testing"
 
 	"git.warky.dev/wdevs/relspecgo/pkg/models"
@@ -359,6 +360,14 @@ func TestDeriveRelationship(t *testing.T) {
 		t.Errorf("Expected ToTable 'users', got '%s'", rel.ToTable)
 	}
 
+	if !reflect.DeepEqual(rel.FromColumns, []string{"user_id"}) {
+		t.Errorf("Expected FromColumns [user_id], got %v", rel.FromColumns)
+	}
+
+	if !reflect.DeepEqual(rel.ToColumns, []string{"id"}) {
+		t.Errorf("Expected ToColumns [id], got %v", rel.ToColumns)
+	}
+
 	if rel.ForeignKey != "fk_orders_user_id" {
 		t.Errorf("Expected ForeignKey 'fk_orders_user_id', got '%s'", rel.ForeignKey)
 	}
@@ -392,3 +401,101 @@ func BenchmarkReader_ReadDatabase(b *testing.B) {
 		}
 	}
 }
+
+func TestParseIndexDefinition_ExtensionIndexes(t *testing.T) {
+	reader := &Reader{}
+
+	tests := []struct {
+		name        string
+		indexDef    string
+		wantType    string
+		wantColumns []string
+		wantComment string
+	}{
+		{
+			name:        "hnsw vector index with storage parameters",
+			indexDef:    "CREATE INDEX idx_docs_embedding ON public.docs USING hnsw (embedding vector_cosine_ops) WITH (m='16', ef_construction='64')",
+			wantType:    "hnsw",
+			wantColumns: []string{"embedding"},
+			wantComment: "opclass=vector_cosine_ops; with (m=16, ef_construction=64)",
+		},
+		{
+			name:        "ivfflat vector index",
+			indexDef:    "CREATE INDEX idx_docs_embedding ON public.docs USING ivfflat (embedding vector_l2_ops) WITH (lists='100')",
+			wantType:    "ivfflat",
+			wantColumns: []string{"embedding"},
+			wantComment: "opclass=vector_l2_ops; with (lists=100)",
+		},
+		{
+			name:        "gist geometry index with default operator class",
+			indexDef:    "CREATE INDEX idx_places_geom ON public.places USING gist (geom)",
+			wantType:    "gist",
+			wantColumns: []string{"geom"},
+			wantComment: "",
+		},
+		{
+			name:        "gist geometry index with explicit operator class",
+			indexDef:    "CREATE INDEX idx_places_geom ON public.places USING gist (geom gist_geometry_ops_nd)",
+			wantType:    "gist",
+			wantColumns: []string{"geom"},
+			wantComment: "opclass=gist_geometry_ops_nd",
+		},
+		{
+			name:        "btree ordering modifiers are not operator classes",
+			indexDef:    "CREATE INDEX idx_users_created ON public.users USING btree (created_at DESC NULLS LAST)",
+			wantType:    "btree",
+			wantColumns: []string{"created_at"},
+			wantComment: "",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			index, err := reader.parseIndexDefinition("idx", "tbl", "public", tt.indexDef)
+			if err != nil {
+				t.Fatalf("parseIndexDefinition() error = %v", err)
+			}
+
+			if index.Type != tt.wantType {
+				t.Errorf("Type = %q, want %q", index.Type, tt.wantType)
+			}
+			if len(index.Columns) != len(tt.wantColumns) {
+				t.Fatalf("Columns = %v, want %v", index.Columns, tt.wantColumns)
+			}
+			for i, col := range tt.wantColumns {
+				if index.Columns[i] != col {
+					t.Errorf("Columns[%d] = %q, want %q", i, index.Columns[i], col)
+				}
+			}
+			if index.Comment != tt.wantComment {
+				t.Errorf("Comment = %q, want %q", index.Comment, tt.wantComment)
+			}
+		})
+	}
+}
+
+func TestMapDataType_ExtensionTypesPreserveModifiers(t *testing.T) {
+	reader := &Reader{}
+
+	tests := []struct {
+		name          string
+		pgType        string
+		udtName       string
+		formattedType string
+		want          string
+	}{
+		{"postgis geometry", "USER-DEFINED", "geometry", "geometry(Point,4326)", "geometry(Point,4326)"},
+		{"postgis geography", "USER-DEFINED", "geography", "geography(Point,4326)", "geography(Point,4326)"},
+		{"postgis geometry without modifier", "USER-DEFINED", "geometry", "geometry", "geometry"},
+		{"pgvector halfvec", "USER-DEFINED", "halfvec", "halfvec(768)", "halfvec(768)"},
+		{"postgis geometry array", "ARRAY", "_geometry", "geometry(Point,4326)[]", "geometry(Point,4326)[]"},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := reader.mapDataType(tt.pgType, tt.udtName, tt.formattedType, false); got != tt.want {
+				t.Errorf("mapDataType() = %q, want %q", got, tt.want)
+			}
+		})
+	}
+}

pkg/writers/pgsql/README.md

@@ -171,6 +171,7 @@ When `include_audit` is enabled, adds:
 - Function-based indexes
 - Concurrent index creation (`CREATE INDEX CONCURRENTLY`) via `Index.Concurrent`
 - Check constraints with expressions
+- Extension types and indexes: PostGIS, pgvector, citext, hstore, ltree (see below)
 
 ## Data Types
 
@@ -186,6 +187,108 @@ Supports all PostgreSQL data types:
 - Network: INET, CIDR, MACADDR
 - Special: ARRAY, HSTORE
 
+## Extension Types (PostGIS, pgvector)
+
+Extension column types are preserved verbatim, including their type modifier:
+
+| Type | Example column type | Extension |
+|------|---------------------|-----------|
+| PostGIS | `geometry(Point,4326)`, `geography(Point)`, `box2d`, `raster` | `postgis`, `postgis_raster`, `postgis_topology` |
+| pgvector | `vector(1536)`, `halfvec(768)`, `sparsevec(1000)` | `vector` |
+| Other | `citext`, `hstore`, `ltree` | `citext`, `hstore`, `ltree` |
+
+`CREATE EXTENSION IF NOT EXISTS <ext>;` is emitted automatically for every extension the
+schema needs. See [Extensions](#extensions).
+
+### Extension Indexes
+
+`Index.Type` selects the access method: `gist`, `spgist`, `brin` (PostGIS), `hnsw`, `ivfflat`
+(pgvector), `vchordrq`, `vchordg` (VectorChord), `bm25` (pg_search).
+
+Operator class and access-method parameters ride in `Index.Comment`:
+
+```
+opclass=vector_l2_ops; with (lists=100)
+```
+
+- `opclass=<name>` — used only when compatible with the column type; otherwise ignored.
+  Bare operator class names in the comment (e.g. `gin_trgm_ops`) are also recognized.
+- `with (k=v, …)` — rendered as `WITH (k = v, …)`. Only well-formed `key = value` pairs are
+  kept, so comment prose never reaches the DDL. Values may be bare (`lists=100`), quoted
+  (`key_field='id'`), or dollar-quoted (`options=$$[build.internal]$$`).
+
+Defaults when no operator class is requested:
+
+| Access method | Column type | Emitted operator class |
+|---------------|-------------|------------------------|
+| `hnsw`, `ivfflat`, `vchordrq`, `vchordg` | `vector` / `halfvec` / `sparsevec` / `bit` | `vector_cosine_ops` / `halfvec_cosine_ops` / `sparsevec_cosine_ops` / `bit_hamming_ops` |
+| `gist`, `spgist`, `brin` | `geometry`, `geography` | none (PostGIS default operator class) |
+| `gin` | text / `jsonb` / array | `gin_trgm_ops` / `jsonb_ops` / `array_ops` |
+
+pgvector defines no default operator class, so a vector index always names one.
+
+```sql
+CREATE INDEX IF NOT EXISTS idx_documents_embedding
+  ON public.documents USING ivfflat (embedding vector_cosine_ops) WITH (lists = 100);
+CREATE INDEX IF NOT EXISTS idx_documents_location
+  ON public.documents USING gist (location);
+```
+
+Migrations only recreate an index when both sides specify a hint and they differ, so a model
+without hints does not churn against a live database.
+
+## Extensions
+
+`CREATE EXTENSION IF NOT EXISTS <ext>;` is emitted per schema, deduplicated and ordered so
+dependencies come first (`postgis` before `postgis_topology`/`postgis_raster`/`pgrouting`,
+`vector` before `vchord`). Names needing quoting are quoted: `CREATE EXTENSION IF NOT EXISTS "uuid-ossp";`
+
+### Detection
+
+| Source | Example | Extension |
+|--------|---------|-----------|
+| Column type | `vector(1536)`, `geometry(Point,4326)`, `citext`, `ltree` | `vector`, `postgis`, `citext`, `ltree` |
+| Index access method | `hnsw`, `ivfflat` / `vchordrq`, `vchordg` / `bm25` | `vector` / `vchord` / `pg_search` |
+| Operator class | `gin_trgm_ops`, `gist_ltree_ops` | `pg_trgm`, `ltree` |
+| GIN/GiST on a scalar type | `USING gin (views)` | `btree_gin` / `btree_gist` |
+| Function in a default, CHECK, index `WHERE`, or view body | `uuid_generate_v4()`, `crypt()`, `ST_Area()`, `unaccent()`, `json_matches_schema()` | `uuid-ossp`, `pgcrypto`, `postgis`, `unaccent`, `pg_jsonschema` |
+
+`gen_random_uuid()` is built in since PostgreSQL 13 and does not pull in `pgcrypto`.
+
+### Declaring extensions explicitly
+
+Extensions that leave no trace in the schema go in `schema.Metadata["extensions"]`, as a list
+or a comma-separated string. Dependencies are pulled in automatically; unknown names are kept
+as given. The PostgreSQL reader populates this from `pg_extension` for the schemas it reads.
+
+```yaml
+metadata:
+  extensions: [pg_cron, timescaledb, pg_stat_statements]
+```
+
+### Recognized extensions
+
+| Category | Extensions |
+|----------|------------|
+| ai/search | `vector`, `vchord` |
+| document | `hstore`, `ltree` |
+| federation | `postgres_fdw` |
+| geospatial | `postgis`, `postgis_raster`, `postgis_topology`, `pgrouting` |
+| indexing | `btree_gin`, `btree_gist` |
+| integration | `http` |
+| integrity | `amcheck` |
+| jobs / scheduling | `pg_background`, `pg_cron` |
+| maintenance | `pg_repack`, `pgstattuple` |
+| observability | `pg_qualstats`, `pg_stat_statements` |
+| partitioning | `pg_partman` |
+| procedural | `plpython3u` |
+| search | `pg_search`, `pg_textsearch` |
+| security | `pgcrypto` |
+| text | `citext`, `fuzzystrmatch`, `pg_trgm`, `unaccent` |
+| time-series | `timescaledb` |
+| utility | `uuid-ossp` |
+| validation | `pg_jsonschema` |
+
 ## Notes
 
 - Generated SQL is formatted and readable

pkg/writers/pgsql/extensions_test.go

@@ -0,0 +1,260 @@
+package pgsql
+
+import (
+	"reflect"
+	"strings"
+	"testing"
+
+	"git.warky.dev/wdevs/relspecgo/pkg/models"
+	"git.warky.dev/wdevs/relspecgo/pkg/writers"
+)
+
+// buildExtensionSchema returns a single-table schema the extension detection tests mutate.
+func buildExtensionSchema(t *testing.T) (*models.Schema, *models.Table) {
+	t.Helper()
+
+	schema := models.InitSchema("public")
+	table := models.InitTable("documents", "public")
+	schema.Tables = append(schema.Tables, table)
+	return schema, table
+}
+
+func addColumn(table *models.Table, name, sqlType string) *models.Column {
+	col := models.InitColumn(name, table.Name, table.Schema)
+	col.Type = sqlType
+	table.Columns[name] = col
+	return col
+}
+
+func TestRequiredExtensions_Detection(t *testing.T) {
+	tests := []struct {
+		name  string
+		build func(schema *models.Schema, table *models.Table)
+		want  []string
+	}{
+		{
+			name:  "no extensions",
+			build: func(_ *models.Schema, table *models.Table) { addColumn(table, "id", "integer") },
+			want:  nil,
+		},
+		{
+			name: "column type",
+			build: func(_ *models.Schema, table *models.Table) {
+				addColumn(table, "embedding", "vector(1536)")
+				addColumn(table, "name", "citext")
+			},
+			want: []string{"citext", "vector"},
+		},
+		{
+			name: "column default function",
+			build: func(_ *models.Schema, table *models.Table) {
+				addColumn(table, "id", "uuid").Default = "uuid_generate_v4()"
+			},
+			want: []string{"uuid-ossp"},
+		},
+		{
+			name: "check constraint expression",
+			build: func(_ *models.Schema, table *models.Table) {
+				addColumn(table, "geom", "geometry")
+				table.Constraints["chk_geom"] = &models.Constraint{
+					Name:       "chk_geom",
+					Type:       models.CheckConstraint,
+					Expression: "ST_IsValid(geom)",
+				}
+			},
+			want: []string{"postgis"},
+		},
+		{
+			name: "partial index predicate",
+			build: func(_ *models.Schema, table *models.Table) {
+				addColumn(table, "title", "text")
+				table.Indexes["idx_title"] = &models.Index{
+					Name:    "idx_title",
+					Type:    "btree",
+					Columns: []string{"title"},
+					Where:   "similarity(title, 'x') > 0.3",
+				}
+			},
+			want: []string{"pg_trgm"},
+		},
+		{
+			name: "view definition",
+			build: func(schema *models.Schema, table *models.Table) {
+				addColumn(table, "title", "text")
+				schema.Views = append(schema.Views, &models.View{
+					Name:       "v_documents",
+					Schema:     "public",
+					Definition: "SELECT unaccent(title) FROM documents",
+				})
+			},
+			want: []string{"unaccent"},
+		},
+		{
+			name: "index access method",
+			build: func(_ *models.Schema, table *models.Table) {
+				addColumn(table, "body", "text")
+				table.Indexes["idx_body"] = &models.Index{
+					Name:    "idx_body",
+					Type:    "bm25",
+					Columns: []string{"body"},
+					Comment: "with (key_field='id')",
+				}
+			},
+			want: []string{"pg_search"},
+		},
+		{
+			name: "vchord depends on vector",
+			build: func(_ *models.Schema, table *models.Table) {
+				addColumn(table, "embedding", "vector(3)")
+				table.Indexes["idx_embedding"] = &models.Index{
+					Name:    "idx_embedding",
+					Type:    "vchordrq",
+					Columns: []string{"embedding"},
+				}
+			},
+			want: []string{"vector", "vchord"},
+		},
+		{
+			name: "gin on scalar needs btree_gin",
+			build: func(_ *models.Schema, table *models.Table) {
+				addColumn(table, "views", "integer")
+				table.Indexes["idx_views"] = &models.Index{
+					Name:    "idx_views",
+					Type:    "gin",
+					Columns: []string{"views"},
+				}
+			},
+			want: []string{"btree_gin"},
+		},
+		{
+			name: "gist on scalar needs btree_gist",
+			build: func(_ *models.Schema, table *models.Table) {
+				addColumn(table, "views", "integer")
+				table.Indexes["idx_views"] = &models.Index{
+					Name:    "idx_views",
+					Type:    "gist",
+					Columns: []string{"views"},
+				}
+			},
+			want: []string{"btree_gist"},
+		},
+		{
+			name: "gist on geometry uses postgis operator classes",
+			build: func(_ *models.Schema, table *models.Table) {
+				addColumn(table, "location", "geometry(Point,4326)")
+				table.Indexes["idx_location"] = &models.Index{
+					Name:    "idx_location",
+					Type:    "gist",
+					Columns: []string{"location"},
+				}
+			},
+			want: []string{"postgis"},
+		},
+		{
+			name: "gin on jsonb needs no companion",
+			build: func(_ *models.Schema, table *models.Table) {
+				addColumn(table, "payload", "jsonb")
+				table.Indexes["idx_payload"] = &models.Index{
+					Name:    "idx_payload",
+					Type:    "gin",
+					Columns: []string{"payload"},
+				}
+			},
+			want: nil,
+		},
+		{
+			name: "gin on text uses pg_trgm",
+			build: func(_ *models.Schema, table *models.Table) {
+				addColumn(table, "title", "text")
+				table.Indexes["idx_title"] = &models.Index{
+					Name:    "idx_title",
+					Type:    "gin",
+					Columns: []string{"title"},
+				}
+			},
+			want: []string{"pg_trgm"},
+		},
+		{
+			name: "gin on array needs no companion",
+			build: func(_ *models.Schema, table *models.Table) {
+				addColumn(table, "tags", "text[]")
+				table.Indexes["idx_tags"] = &models.Index{
+					Name:    "idx_tags",
+					Type:    "gin",
+					Columns: []string{"tags"},
+				}
+			},
+			want: nil,
+		},
+		{
+			name: "declared in metadata as string",
+			build: func(schema *models.Schema, _ *models.Table) {
+				schema.Metadata = map[string]any{"extensions": "pg_cron, timescaledb"}
+			},
+			want: []string{"pg_cron", "timescaledb"},
+		},
+		{
+			name: "declared in metadata as list",
+			build: func(schema *models.Schema, _ *models.Table) {
+				schema.Metadata = map[string]any{"extensions": []any{"postgis_topology", "pg_stat_statements"}}
+			},
+			// postgis is pulled in as a dependency of postgis_topology and emitted first.
+			want: []string{"pg_stat_statements", "postgis", "postgis_topology"},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			schema, table := buildExtensionSchema(t)
+			tt.build(schema, table)
+
+			if got := requiredExtensions(schema); !reflect.DeepEqual(got, tt.want) {
+				t.Errorf("requiredExtensions() = %v, want %v", got, tt.want)
+			}
+		})
+	}
+}
+
+func TestRequiredExtensions_NilSchema(t *testing.T) {
+	if got := requiredExtensions(nil); got != nil {
+		t.Errorf("requiredExtensions(nil) = %v, want nil", got)
+	}
+}
+
+func TestWriteDatabase_QuotesExtensionNames(t *testing.T) {
+	db := models.InitDatabase("testdb")
+	schema, table := buildExtensionSchema(t)
+	addColumn(table, "id", "uuid").Default = "uuid_generate_v4()"
+	db.Schemas = append(db.Schemas, schema)
+
+	output := writeDatabaseOutput(t, db)
+	if !strings.Contains(output, `CREATE EXTENSION IF NOT EXISTS "uuid-ossp";`) {
+		t.Fatalf("expected quoted extension name, got:\n%s", output)
+	}
+}
+
+func TestGenerateSchemaStatements_ExtensionDependencyOrder(t *testing.T) {
+	schema, table := buildExtensionSchema(t)
+	addColumn(table, "embedding", "vector(3)")
+	table.Indexes["idx_embedding"] = &models.Index{
+		Name:    "idx_embedding",
+		Type:    "vchordrq",
+		Columns: []string{"embedding"},
+	}
+
+	writer := NewWriter(&writers.WriterOptions{})
+	statements, err := writer.GenerateSchemaStatements(schema)
+	if err != nil {
+		t.Fatalf("GenerateSchemaStatements failed: %v", err)
+	}
+
+	joined := strings.Join(statements, "\n")
+	vector := strings.Index(joined, "CREATE EXTENSION IF NOT EXISTS vector")
+	vchord := strings.Index(joined, "CREATE EXTENSION IF NOT EXISTS vchord")
+	if vector < 0 || vchord < 0 {
+		t.Fatalf("expected vector and vchord extensions, got:\n%s", joined)
+	}
+	if vector > vchord {
+		t.Fatalf("expected vector to be created before vchord, got:\n%s", joined)
+	}
+}

pkg/writers/pgsql/migration_writer.go

@@ -164,14 +164,14 @@ func (w *MigrationWriter) WriteMigration(model *models.Database, current *models
 func (w *MigrationWriter) generateSchemaScripts(model *models.Schema, current *models.Schema) ([]MigrationScript, error) {
 	scripts := make([]MigrationScript, 0)
 
-	if schemaRequiresPGTrgm(model) {
+	for _, extension := range requiredExtensions(model) {
 		scripts = append(scripts, MigrationScript{
-			ObjectName: "extension.pg_trgm",
+			ObjectName: "extension." + extension,
 			ObjectType: "create extension",
 			Schema:     model.Name,
 			Priority:   80,
 			Sequence:   len(scripts),
-			Body:       "CREATE EXTENSION IF NOT EXISTS pg_trgm;",
+			Body:       fmt.Sprintf("CREATE EXTENSION IF NOT EXISTS %s;", pgsql.QuoteExtensionName(extension)),
 		})
 	}
 
@@ -653,6 +653,7 @@ func (w *MigrationWriter) generateIndexScripts(model *models.Schema, current *mo
 					Columns:           strings.Join(columnExprs, ", "),
 					Unique:            modelIndex.Unique,
 					Concurrent:        modelIndex.Concurrent,
+					StorageParameters: indexStorageParameters(modelIndex.Comment),
 				})
 				if err != nil {
 					return nil, err
@@ -674,20 +675,31 @@ func (w *MigrationWriter) generateIndexScripts(model *models.Schema, current *mo
 	return scripts, nil
 }
 
+// buildIndexColumnExpressions renders the column list of an index, appending the operator
+// class each column needs for the access method (GIN opclasses, pgvector distance ops,
+// explicitly requested PostGIS opclasses). Columns that cannot be resolved on the table are
+// emitted verbatim.
 func buildIndexColumnExpressions(table *models.Table, index *models.Index, indexType string) []string {
+	return buildIndexColumnExpressionsFiltered(table, index, indexType, false)
+}
+
+// buildIndexColumnExpressionsFiltered is buildIndexColumnExpressions with the option to drop
+// columns that do not exist on the table instead of emitting them verbatim.
+func buildIndexColumnExpressionsFiltered(table *models.Table, index *models.Index, indexType string, skipUnresolved bool) []string {
 	columnExprs := make([]string, 0, len(index.Columns))
 	for _, colName := range index.Columns {
-		colExpr := colName
-		if table != nil {
-			if col, ok := resolveIndexColumn(table, colName); ok && col != nil {
-				colExpr = col.SQLName()
-				if strings.EqualFold(indexType, "gin") {
-					opClass := ginOperatorClassForColumn(col, index.Comment)
-					if opClass != "" {
-						colExpr = fmt.Sprintf("%s %s", col.SQLName(), opClass)
-					}
+		col, ok := resolveIndexColumn(table, colName)
+		if !ok || col == nil {
+			if skipUnresolved {
+				continue
 			}
+			columnExprs = append(columnExprs, colName)
+			continue
 		}
+
+		colExpr := col.SQLName()
+		if opClass := indexOperatorClassForColumn(col, indexType, index.Comment); opClass != "" {
+			colExpr = fmt.Sprintf("%s %s", colExpr, opClass)
 		}
 		columnExprs = append(columnExprs, colExpr)
 	}
@@ -1046,5 +1058,19 @@ func indexesEqual(idx1, idx2 *models.Index) bool {
 			return false
 		}
 	}
+	// Operator class and storage parameters ride along in the index comment. They only
+	// signal a difference when both sides specify one, so an index whose model side omits
+	// the hint is not recreated on every migration.
+	if !indexHintsEqual(extractOperatorClass(idx1.Comment), extractOperatorClass(idx2.Comment)) {
+		return false
+	}
+	return indexHintsEqual(indexStorageParameters(idx1.Comment), indexStorageParameters(idx2.Comment))
+}
+
+// indexHintsEqual compares two optional index hints, treating an unspecified hint as a match.
+func indexHintsEqual(hint1, hint2 string) bool {
+	if hint1 == "" || hint2 == "" {
 		return true
 	}
+	return strings.EqualFold(hint1, hint2)
+}

pkg/writers/pgsql/migration_writer_test.go

@@ -852,3 +852,93 @@ func TestWriteMigration_NilCurrentTreatsDatabaseAsEmpty(t *testing.T) {
 		t.Fatalf("expected CREATE TABLE in migration output, got:\n%s", output)
 	}
 }
+
+func TestWriteMigration_VectorAndPostGISIndexes(t *testing.T) {
+	current := models.InitDatabase("testdb")
+	current.Schemas = append(current.Schemas, models.InitSchema("public"))
+
+	model := models.InitDatabase("testdb")
+	modelSchema := models.InitSchema("public")
+
+	table := models.InitTable("documents", "public")
+
+	embedding := models.InitColumn("embedding", "documents", "public")
+	embedding.Type = "vector(1536)"
+	table.Columns["embedding"] = embedding
+
+	location := models.InitColumn("location", "documents", "public")
+	location.Type = "geometry(Point,4326)"
+	table.Columns["location"] = location
+
+	table.Indexes["idx_documents_embedding"] = &models.Index{
+		Name:    "idx_documents_embedding",
+		Type:    "ivfflat",
+		Columns: []string{"embedding"},
+		Comment: "opclass=vector_cosine_ops; with (lists=100)",
+	}
+	table.Indexes["idx_documents_location"] = &models.Index{
+		Name:    "idx_documents_location",
+		Type:    "gist",
+		Columns: []string{"location"},
+	}
+
+	modelSchema.Tables = append(modelSchema.Tables, table)
+	model.Schemas = append(model.Schemas, modelSchema)
+
+	var buf bytes.Buffer
+	writer, err := NewMigrationWriter(&writers.WriterOptions{})
+	if err != nil {
+		t.Fatalf("Failed to create writer: %v", err)
+	}
+	writer.writer = &buf
+
+	if err := writer.WriteMigration(model, current); err != nil {
+		t.Fatalf("WriteMigration failed: %v", err)
+	}
+
+	output := buf.String()
+	for _, want := range []string{
+		"CREATE EXTENSION IF NOT EXISTS postgis;",
+		"CREATE EXTENSION IF NOT EXISTS vector;",
+		"vector(1536)",
+		"geometry(Point,4326)",
+		"USING ivfflat (embedding vector_cosine_ops) WITH (lists = 100)",
+		"USING gist (location)",
+	} {
+		if !strings.Contains(output, want) {
+			t.Fatalf("expected migration to contain %q, got:\n%s", want, output)
+		}
+	}
+}
+
+func TestIndexesEqual_OperatorClassAndStorageParameters(t *testing.T) {
+	newIndex := func(comment string) *models.Index {
+		return &models.Index{
+			Name:    "idx_documents_embedding",
+			Type:    "hnsw",
+			Columns: []string{"embedding"},
+			Comment: comment,
+		}
+	}
+
+	tests := []struct {
+		name      string
+		comment1  string
+		comment2  string
+		wantEqual bool
+	}{
+		{"identical hints", "opclass=vector_l2_ops", "opclass=vector_l2_ops", true},
+		{"different operator class", "opclass=vector_l2_ops", "opclass=vector_cosine_ops", false},
+		{"different storage parameters", "with (m=16)", "with (m=32)", false},
+		{"unspecified hint on one side", "", "opclass=vector_l2_ops; with (m=16)", true},
+		{"unrelated comments", "primary lookup index", "primary lookup index", true},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := indexesEqual(newIndex(tt.comment1), newIndex(tt.comment2)); got != tt.wantEqual {
+				t.Errorf("indexesEqual() = %v, want %v", got, tt.wantEqual)
+			}
+		})
+	}
+}

pkg/writers/pgsql/templates.go

@@ -140,6 +140,9 @@ type CreateIndexData struct {
 	Columns    string
 	Unique     bool
 	Concurrent bool
+	// StorageParameters holds access-method parameters rendered as WITH (...),
+	// e.g. "lists = 100" for ivfflat or "m = 16, ef_construction = 64" for hnsw.
+	StorageParameters string
 }
 
 // CreateForeignKeyData contains data for create foreign key template

pkg/writers/pgsql/templates/create_index.tmpl

@@ -1,2 +1,2 @@
 CREATE {{if .Unique}}UNIQUE {{end}}INDEX {{if .Concurrent}}CONCURRENTLY {{end}}IF NOT EXISTS {{quote_ident .IndexName}}
-  ON {{qual_table .SchemaName .TableName}} USING {{.IndexType}} ({{.Columns}});
+  ON {{qual_table .SchemaName .TableName}} USING {{.IndexType}} ({{.Columns}}){{if .StorageParameters}} WITH ({{.StorageParameters}}){{end}};

pkg/writers/pgsql/writer.go

@@ -6,8 +6,10 @@ import (
 	"fmt"
 	"io"
 	"os"
+	"regexp"
 	"sort"
 	"strings"
+	"sync"
 	"time"
 
 	"git.warky.dev/wdevs/relspecgo/pkg/models"
@@ -147,8 +149,8 @@ func (w *Writer) GenerateSchemaStatements(schema *models.Schema) ([]string, erro
 		statements = append(statements, fmt.Sprintf("CREATE SCHEMA IF NOT EXISTS %s", schema.SQLName()))
 	}
 
-	if schemaRequiresPGTrgm(schema) {
-		statements = append(statements, `CREATE EXTENSION IF NOT EXISTS pg_trgm`)
+	for _, extension := range requiredExtensions(schema) {
+		statements = append(statements, fmt.Sprintf("CREATE EXTENSION IF NOT EXISTS %s", pgsql.QuoteExtensionName(extension)))
 	}
 
 	// Phase 2: Create sequences
@@ -271,18 +273,12 @@ func (w *Writer) GenerateSchemaStatements(schema *models.Schema) ([]string, erro
 				indexType = "btree"
 			}
 
-			// Build column expressions with operator class support for GIN indexes
-			columnExprs := make([]string, 0, len(index.Columns))
-			for _, colName := range index.Columns {
-				colExpr := colName
-				if col, ok := resolveIndexColumn(table, colName); ok {
-					if strings.EqualFold(indexType, "gin") {
-						if opClass := ginOperatorClassForColumn(col, index.Comment); opClass != "" {
-							colExpr = fmt.Sprintf("%s %s", colName, opClass)
-						}
-					}
-				}
-				columnExprs = append(columnExprs, colExpr)
+			// Build column expressions with operator class support (GIN, pgvector, PostGIS)
+			columnExprs := buildIndexColumnExpressions(table, index, indexType)
+
+			withClause := ""
+			if params := indexStorageParameters(index.Comment); params != "" {
+				withClause = fmt.Sprintf(" WITH (%s)", params)
 			}
 
 			whereClause := ""
@@ -290,8 +286,8 @@ func (w *Writer) GenerateSchemaStatements(schema *models.Schema) ([]string, erro
 				whereClause = fmt.Sprintf(" WHERE %s", index.Where)
 			}
 
-			stmt := fmt.Sprintf("CREATE %sINDEX IF NOT EXISTS %s ON %s USING %s (%s)%s",
-				uniqueStr, quoteIdentifier(index.Name), w.qualTable(schema.SQLName(), table.SQLName()), indexType, strings.Join(columnExprs, ", "), whereClause)
+			stmt := fmt.Sprintf("CREATE %sINDEX IF NOT EXISTS %s ON %s USING %s (%s)%s%s",
+				uniqueStr, quoteIdentifier(index.Name), w.qualTable(schema.SQLName(), table.SQLName()), indexType, strings.Join(columnExprs, ", "), withClause, whereClause)
 			statements = append(statements, stmt)
 		}
 	}
@@ -819,11 +815,14 @@ func (w *Writer) writeCreateSchema(schema *models.Schema) error {
 }
 
 func (w *Writer) writeRequiredExtensions(schema *models.Schema) error {
-	if !schemaRequiresPGTrgm(schema) {
+	extensions := requiredExtensions(schema)
+	if len(extensions) == 0 {
 		return nil
 	}
 
-	fmt.Fprintln(w.writer, "CREATE EXTENSION IF NOT EXISTS pg_trgm;")
+	for _, extension := range extensions {
+		fmt.Fprintf(w.writer, "CREATE EXTENSION IF NOT EXISTS %s;\n", pgsql.QuoteExtensionName(extension))
+	}
 	fmt.Fprintln(w.writer)
 	return nil
 }
@@ -1063,21 +1062,13 @@ func (w *Writer) writeIndexes(schema *models.Schema) error {
 				indexName = fmt.Sprintf("%s_%s_%s", indexType, table.SQLName(), strings.ToLower(columnSuffix))
 			}
 
-			// Build column list with operator class support for GIN indexes
-			columnExprs := make([]string, 0, len(index.Columns))
-			for _, colName := range index.Columns {
-				if col, ok := resolveIndexColumn(table, colName); ok {
-					colExpr := col.SQLName()
-					if strings.EqualFold(index.Type, "gin") {
-						opClass := ginOperatorClassForColumn(col, index.Comment)
-						if opClass != "" {
-							colExpr = fmt.Sprintf("%s %s", col.SQLName(), opClass)
-						}
-					}
-					columnExprs = append(columnExprs, colExpr)
-				}
+			indexType := index.Type
+			if indexType == "" {
+				indexType = "btree"
 			}
 
+			// Build column list with operator class support (GIN, pgvector, PostGIS)
+			columnExprs := buildIndexColumnExpressionsFiltered(table, index, indexType, true)
 			if len(columnExprs) == 0 {
 				continue
 			}
@@ -1087,9 +1078,9 @@ func (w *Writer) writeIndexes(schema *models.Schema) error {
 				unique = "UNIQUE "
 			}
 
-			indexType := index.Type
-			if indexType == "" {
-				indexType = "btree"
+			withClause := ""
+			if params := indexStorageParameters(index.Comment); params != "" {
+				withClause = fmt.Sprintf(" WITH (%s)", params)
 			}
 
 			whereClause := ""
@@ -1104,8 +1095,8 @@ 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;\n\n",
-				w.qualTable(schema.SQLName(), table.SQLName()), indexType, strings.Join(columnExprs, ", "), whereClause)
+			fmt.Fprintf(w.writer, "  ON %s USING %s (%s)%s%s;\n\n",
+				w.qualTable(schema.SQLName(), table.SQLName()), indexType, strings.Join(columnExprs, ", "), withClause, whereClause)
 		}
 	}
 
@@ -1483,7 +1474,69 @@ func isTextTypeWithoutLength(colType string) bool {
 	return strings.EqualFold(colType, "text")
 }
 
-func ginOperatorClassForColumn(col *models.Column, comment string) string {
+// vectorOperatorClasses maps pgvector operator classes to the column base type they
+// apply to. pgvector defines no default operator class, so an hnsw/ivfflat index must
+// always name one explicitly.
+var vectorOperatorClasses = map[string]string{
+	"vector_l2_ops":        "vector",
+	"vector_ip_ops":        "vector",
+	"vector_cosine_ops":    "vector",
+	"vector_l1_ops":        "vector",
+	"halfvec_l2_ops":       "halfvec",
+	"halfvec_ip_ops":       "halfvec",
+	"halfvec_cosine_ops":   "halfvec",
+	"halfvec_l1_ops":       "halfvec",
+	"sparsevec_l2_ops":     "sparsevec",
+	"sparsevec_ip_ops":     "sparsevec",
+	"sparsevec_cosine_ops": "sparsevec",
+	"sparsevec_l1_ops":     "sparsevec",
+	"bit_hamming_ops":      "bit",
+	"bit_jaccard_ops":      "bit",
+}
+
+// defaultVectorOperatorClasses is the operator class used for an hnsw/ivfflat index when
+// the index comment does not request one. Cosine distance is the common default for
+// embedding columns; override it with an "opclass" hint in the index comment.
+var defaultVectorOperatorClasses = map[string]string{
+	"vector":    "vector_cosine_ops",
+	"halfvec":   "halfvec_cosine_ops",
+	"sparsevec": "sparsevec_cosine_ops",
+	"bit":       "bit_hamming_ops",
+}
+
+// spatialOperatorClasses are the PostGIS operator classes recognized in index comments.
+// PostGIS installs default operator classes for gist/spgist/brin, so these are only
+// emitted when explicitly requested (e.g. the 3D/nD variants).
+var spatialOperatorClasses = map[string]bool{
+	"gist_geometry_ops_2d":            true,
+	"gist_geometry_ops_nd":            true,
+	"gist_geography_ops":              true,
+	"spgist_geometry_ops_2d":          true,
+	"spgist_geometry_ops_3d":          true,
+	"spgist_geometry_ops_nd":          true,
+	"brin_geometry_inclusion_ops_2d":  true,
+	"brin_geometry_inclusion_ops_3d":  true,
+	"brin_geometry_inclusion_ops_4d":  true,
+	"brin_geography_inclusion_ops_2d": true,
+	"btree_geometry_ops":              true,
+	"btree_geography_ops":             true,
+}
+
+// isVectorIndexMethod reports whether the access method indexes pgvector types, which
+// covers both pgvector itself (hnsw, ivfflat) and VectorChord (vchordrq, vchordg).
+func isVectorIndexMethod(method string) bool {
+	switch strings.ToLower(strings.TrimSpace(method)) {
+	case "hnsw", "ivfflat", "vchordrq", "vchordg":
+		return true
+	default:
+		return false
+	}
+}
+
+// indexOperatorClassForColumn returns the operator class to emit for a column in an index
+// of the given access method, honouring an explicit request from the index comment when it
+// is compatible with the column type.
+func indexOperatorClassForColumn(col *models.Column, indexType, comment string) string {
 	if col == nil {
 		return ""
 	}
@@ -1492,15 +1545,20 @@ func ginOperatorClassForColumn(col *models.Column, comment string) string {
 	baseType := pgsql.CanonicalizeBaseType(pgsql.ExtractBaseTypeLower(sqlType))
 	isArray := pgsql.IsArrayType(sqlType)
 	requested := extractOperatorClass(comment)
+	method := strings.ToLower(strings.TrimSpace(indexType))
+	if method == "" {
+		method = "btree"
+	}
 
-	if requested != "" && ginOperatorClassCompatible(baseType, isArray, requested) {
+	if requested != "" && operatorClassCompatible(method, baseType, isArray, requested) {
 		return requested
 	}
 
+	switch {
+	case method == "gin":
 		if isArray {
 			return "array_ops"
 		}
-
 		switch {
 		case isTextGinBaseType(baseType):
 			return "gin_trgm_ops"
@@ -1509,9 +1567,31 @@ func ginOperatorClassForColumn(col *models.Column, comment string) string {
 		default:
 			return requested
 		}
+	case isVectorIndexMethod(method):
+		if isArray {
+			return ""
+		}
+		return defaultVectorOperatorClasses[baseType]
+	default:
+		// gist/spgist/brin/btree have default operator classes (PostGIS included),
+		// so nothing is emitted unless the comment requested a compatible class.
+		return ""
+	}
+}
+
+// ginOperatorClassForColumn is the GIN-specific form of indexOperatorClassForColumn.
+func ginOperatorClassForColumn(col *models.Column, comment string) string {
+	return indexOperatorClassForColumn(col, "gin", comment)
+}
+
+func operatorClassCompatible(method, baseType string, isArray bool, opClass string) bool {
+	if vectorType, ok := vectorOperatorClasses[opClass]; ok {
+		return !isArray && baseType == vectorType && isVectorIndexMethod(method)
+	}
+	if spatialOperatorClasses[opClass] {
+		return !isArray && pgsql.IsSpatialType(baseType)
 	}
 
-func ginOperatorClassCompatible(baseType string, isArray bool, opClass string) bool {
 	switch opClass {
 	case "gin_trgm_ops", "gin_bigm_ops":
 		return !isArray && isTextGinBaseType(baseType)
@@ -1524,6 +1604,10 @@ func ginOperatorClassCompatible(baseType string, isArray bool, opClass string) b
 	}
 }
 
+func ginOperatorClassCompatible(baseType string, isArray bool, opClass string) bool {
+	return operatorClassCompatible("gin", baseType, isArray, opClass)
+}
+
 func isTextGinBaseType(baseType string) bool {
 	switch baseType {
 	case "text", "varchar", "character varying", "char", "character", "string", "citext", "bpchar":
@@ -1533,27 +1617,186 @@ func isTextGinBaseType(baseType string) bool {
 	}
 }
 
-func schemaRequiresPGTrgm(schema *models.Schema) bool {
+// requiredExtensions returns the PostgreSQL extensions a schema depends on, ordered so
+// that dependencies are created first (postgis before postgis_topology, vector before
+// vchord). Extensions are detected from column types, index access methods, resolved
+// operator classes, and function calls in defaults, check constraints, partial index
+// predicates and view definitions. Extensions that leave no trace in the model (pg_cron,
+// timescaledb, postgres_fdw, …) can be declared in schema.Metadata["extensions"].
+func requiredExtensions(schema *models.Schema) []string {
 	if schema == nil {
-		return false
+		return nil
 	}
+
+	required := make(map[string]bool)
+	add := func(names ...string) {
+		for _, name := range names {
+			if name != "" {
+				required[name] = true
+			}
+		}
+	}
+
+	add(declaredExtensions(schema)...)
+
+	for _, view := range schema.Views {
+		if view == nil {
+			continue
+		}
+		add(pgsql.ExtensionsForExpression(view.Definition)...)
+	}
+
 	for _, table := range schema.Tables {
 		if table == nil {
 			continue
 		}
-		for _, index := range table.Indexes {
-			if index == nil || !strings.EqualFold(index.Type, "gin") {
+
+		for _, col := range table.Columns {
+			if col == nil {
 				continue
 			}
+			add(pgsql.TypeExtension(effectiveColumnSQLType(col)))
+			if def, ok := col.Default.(string); ok {
+				add(pgsql.ExtensionsForExpression(def)...)
+			}
+		}
+
+		for _, constraint := range table.Constraints {
+			if constraint == nil {
+				continue
+			}
+			add(pgsql.ExtensionsForExpression(constraint.Expression)...)
+		}
+
+		for _, index := range table.Indexes {
+			if index == nil {
+				continue
+			}
+			add(pgsql.IndexMethodExtension(index.Type))
+			add(pgsql.ExtensionsForExpression(index.Where)...)
+
 			for _, colName := range index.Columns {
 				col, ok := resolveIndexColumn(table, colName)
 				if !ok || col == nil {
 					continue
 				}
-				if ginOperatorClassForColumn(col, index.Comment) == "gin_trgm_ops" {
+				opClass := indexOperatorClassForColumn(col, index.Type, index.Comment)
+				add(pgsql.OperatorClassExtension(opClass))
+				add(btreeCompanionExtension(index.Type, col, opClass))
+			}
+		}
+	}
+
+	extensions := make([]string, 0, len(required))
+	for ext := range required {
+		extensions = append(extensions, ext)
+	}
+
+	// Pull in dependencies, so a declared postgis_topology also creates postgis.
+	for i := 0; i < len(extensions); i++ {
+		for _, dependency := range pgsql.ExtensionDependencies(extensions[i]) {
+			if !required[dependency] {
+				required[dependency] = true
+				extensions = append(extensions, dependency)
+			}
+		}
+	}
+
+	return pgsql.SortExtensions(extensions)
+}
+
+// declaredExtensions reads schema.Metadata["extensions"], which accepts either a list or a
+// comma-separated string. Unknown names are kept: the metadata is an explicit instruction.
+func declaredExtensions(schema *models.Schema) []string {
+	value, ok := schema.Metadata["extensions"]
+	if !ok {
+		return nil
+	}
+
+	var names []string
+	switch declared := value.(type) {
+	case string:
+		names = strings.Split(declared, ",")
+	case []string:
+		names = declared
+	case []any:
+		for _, item := range declared {
+			if name, ok := item.(string); ok {
+				names = append(names, name)
+			}
+		}
+	default:
+		return nil
+	}
+
+	cleaned := make([]string, 0, len(names))
+	for _, name := range names {
+		if name = strings.TrimSpace(name); name != "" {
+			cleaned = append(cleaned, name)
+		}
+	}
+	return cleaned
+}
+
+// btreeCompanionExtension returns btree_gin or btree_gist when a GIN/GiST index covers a
+// scalar type that neither access method has a built-in operator class for. Without the
+// companion extension PostgreSQL rejects the CREATE INDEX outright.
+func btreeCompanionExtension(indexType string, col *models.Column, opClass string) string {
+	if opClass != "" {
+		return ""
+	}
+
+	method := strings.ToLower(strings.TrimSpace(indexType))
+	if method != "gin" && method != "gist" {
+		return ""
+	}
+
+	sqlType := effectiveColumnSQLType(col)
+	if pgsql.IsArrayType(sqlType) {
+		return ""
+	}
+
+	baseType := pgsql.CanonicalizeBaseType(pgsql.ExtractBaseTypeLower(sqlType))
+	if pgsql.TypeExtension(baseType) != "" {
+		// Extension types (geometry, vector, citext, …) ship their own operator classes.
+		return ""
+	}
+
+	if method == "gin" {
+		if nativeGinBaseType(baseType) {
+			return ""
+		}
+		return "btree_gin"
+	}
+	if nativeGistBaseType(baseType) {
+		return ""
+	}
+	return "btree_gist"
+}
+
+// nativeGinBaseType reports whether core PostgreSQL provides a GIN operator class.
+func nativeGinBaseType(baseType string) bool {
+	switch baseType {
+	case "jsonb", "json", "tsvector", "tsquery":
+		return true
+	default:
+		return false
+	}
+}
+
+// nativeGistBaseType reports whether core PostgreSQL provides a GiST operator class.
+func nativeGistBaseType(baseType string) bool {
+	switch baseType {
+	case "tsvector", "tsquery", "point", "box", "circle", "polygon", "line", "lseg", "path", "inet", "cidr":
 		return true
 	}
+	return strings.HasSuffix(baseType, "range") || strings.HasSuffix(baseType, "multirange")
 }
+
+func schemaRequiresPGTrgm(schema *models.Schema) bool {
+	for _, ext := range requiredExtensions(schema) {
+		if ext == "pg_trgm" {
+			return true
 		}
 	}
 	return false
@@ -1642,14 +1885,21 @@ func formatStringList(items []string) string {
 
 // extractOperatorClass extracts operator class from index comment/note
 // Looks for common operator classes like gin_trgm_ops, gist_trgm_ops, etc.
+// explicitOperatorClassPattern matches an "opclass=<name>" hint, the form the PostgreSQL
+// reader uses to carry an index's operator class through the model.
+var explicitOperatorClassPattern = regexp.MustCompile(`(?i)\bopclass\s*=\s*([a-z_][a-z0-9_]*)\b`)
+
 func extractOperatorClass(comment string) string {
 	if comment == "" {
 		return ""
 	}
+
 	lowerComment := strings.ToLower(comment)
-	// Common GIN/GiST operator classes
-	opClasses := []string{"gin_trgm_ops", "gist_trgm_ops", "gin_bigm_ops", "jsonb_ops", "jsonb_path_ops", "array_ops"}
-	for _, op := range opClasses {
+	if matches := explicitOperatorClassPattern.FindStringSubmatch(lowerComment); len(matches) > 1 {
+		return matches[1]
+	}
+
+	for _, op := range knownOperatorClasses() {
 		if strings.Contains(lowerComment, op) {
 			return op
 		}
@@ -1657,6 +1907,35 @@ func extractOperatorClass(comment string) string {
 	return ""
 }
 
+// knownOperatorClasses lists every operator class recognized in an index comment,
+// longest name first so that e.g. gist_geometry_ops_nd wins over a shorter prefix.
+var knownOperatorClasses = sync.OnceValue(func() []string {
+	names := []string{"gin_trgm_ops", "gist_trgm_ops", "gin_bigm_ops", "jsonb_ops", "jsonb_path_ops", "array_ops"}
+	for name := range vectorOperatorClasses {
+		names = append(names, name)
+	}
+	for name := range spatialOperatorClasses {
+		names = append(names, name)
+	}
+	sort.Slice(names, func(i, j int) bool {
+		if len(names[i]) != len(names[j]) {
+			return len(names[i]) > len(names[j])
+		}
+		return names[i] < names[j]
+	})
+	return names
+})
+
+// indexStorageParameters extracts access-method storage parameters from an index comment.
+// Only well-formed "key = value" pairs are kept, so comment prose cannot leak into DDL.
+// Example: "opclass=vector_cosine_ops with (m=16, ef_construction=64)" -> "m = 16, ef_construction = 64".
+func indexStorageParameters(comment string) string {
+	if comment == "" {
+		return ""
+	}
+	return pgsql.FormatStorageParameters(pgsql.ExtractWithClause(comment))
+}
+
 // escapeQuote escapes single quotes in strings for SQL
 func escapeQuote(s string) string {
 	return strings.ReplaceAll(s, "'", "''")

pkg/writers/pgsql/writer_test.go

@@ -1310,3 +1310,203 @@ func TestWriteSchema_UsesStorageTypeForSerialAlterStatements(t *testing.T) {
 		t.Fatalf("expected serial alter to include USING cast, got:\n%s", output)
 	}
 }
+
+// buildVectorSpatialSchema returns a database with a pgvector column and a PostGIS column.
+func buildVectorSpatialSchema(indexType, indexComment string) *models.Database {
+	db := models.InitDatabase("testdb")
+	schema := models.InitSchema("public")
+
+	table := models.InitTable("documents", "public")
+
+	embedding := models.InitColumn("embedding", "documents", "public")
+	embedding.Type = "vector(1536)"
+	table.Columns["embedding"] = embedding
+
+	location := models.InitColumn("location", "documents", "public")
+	location.Type = "geometry(Point,4326)"
+	table.Columns["location"] = location
+
+	if indexType != "" {
+		index := &models.Index{
+			Name:    "idx_documents_embedding",
+			Type:    indexType,
+			Columns: []string{"embedding"},
+			Comment: indexComment,
+		}
+		table.Indexes[index.Name] = index
+	}
+
+	schema.Tables = append(schema.Tables, table)
+	db.Schemas = append(db.Schemas, schema)
+	return db
+}
+
+func writeDatabaseOutput(t *testing.T, db *models.Database) string {
+	t.Helper()
+
+	var buf bytes.Buffer
+	writer := NewWriter(&writers.WriterOptions{})
+	writer.writer = &buf
+
+	if err := writer.WriteDatabase(db); err != nil {
+		t.Fatalf("WriteDatabase failed: %v", err)
+	}
+	return buf.String()
+}
+
+func TestWriteDatabase_VectorAndPostGISColumnsCreateExtensions(t *testing.T) {
+	output := writeDatabaseOutput(t, buildVectorSpatialSchema("", ""))
+
+	for _, want := range []string{
+		"CREATE EXTENSION IF NOT EXISTS postgis;",
+		"CREATE EXTENSION IF NOT EXISTS vector;",
+		"vector(1536)",
+		"geometry(Point,4326)",
+	} {
+		if !strings.Contains(output, want) {
+			t.Fatalf("expected output to contain %q, got:\n%s", want, output)
+		}
+	}
+
+	// postgis must be created before postgis-dependent extensions and stay deterministic
+	if strings.Index(output, "EXISTS postgis;") > strings.Index(output, "EXISTS vector;") {
+		t.Fatalf("expected extensions to be emitted in sorted order, got:\n%s", output)
+	}
+}
+
+func TestWriteDatabase_HNSWIndexUsesDefaultVectorOperatorClass(t *testing.T) {
+	output := writeDatabaseOutput(t, buildVectorSpatialSchema("hnsw", ""))
+
+	if !strings.Contains(output, "USING hnsw (embedding vector_cosine_ops)") {
+		t.Fatalf("expected hnsw index with default vector operator class, got:\n%s", output)
+	}
+	if !strings.Contains(output, "CREATE EXTENSION IF NOT EXISTS vector;") {
+		t.Fatalf("expected pgvector extension, got:\n%s", output)
+	}
+}
+
+func TestWriteDatabase_VectorIndexHonoursRequestedOperatorClassAndStorageParameters(t *testing.T) {
+	output := writeDatabaseOutput(t, buildVectorSpatialSchema("ivfflat", "opclass=vector_l2_ops; with (lists=100)"))
+
+	if !strings.Contains(output, "USING ivfflat (embedding vector_l2_ops) WITH (lists = 100)") {
+		t.Fatalf("expected ivfflat index with requested opclass and storage parameters, got:\n%s", output)
+	}
+}
+
+func TestWriteDatabase_VectorIndexIgnoresIncompatibleOperatorClass(t *testing.T) {
+	output := writeDatabaseOutput(t, buildVectorSpatialSchema("hnsw", "opclass=halfvec_l2_ops"))
+
+	if !strings.Contains(output, "USING hnsw (embedding vector_cosine_ops)") {
+		t.Fatalf("expected halfvec operator class to be rejected for a vector column, got:\n%s", output)
+	}
+}
+
+func TestWriteDatabase_VectorIndexIgnoresCommentProseInStorageParameters(t *testing.T) {
+	output := writeDatabaseOutput(t, buildVectorSpatialSchema("hnsw", "tuned with (m=16, ef_construction=64, drop table foo)"))
+
+	if !strings.Contains(output, "WITH (m = 16, ef_construction = 64)") {
+		t.Fatalf("expected only well-formed storage parameters, got:\n%s", output)
+	}
+	if strings.Contains(output, "drop table") {
+		t.Fatalf("expected prose to be dropped from storage parameters, got:\n%s", output)
+	}
+}
+
+func TestWriteDatabase_GistIndexOnGeometryUsesDefaultOperatorClass(t *testing.T) {
+	db := models.InitDatabase("testdb")
+	schema := models.InitSchema("public")
+
+	table := models.InitTable("places", "public")
+	geom := models.InitColumn("geom", "places", "public")
+	geom.Type = "geometry(Point,4326)"
+	table.Columns["geom"] = geom
+
+	table.Indexes["idx_places_geom"] = &models.Index{
+		Name:    "idx_places_geom",
+		Type:    "gist",
+		Columns: []string{"geom"},
+	}
+
+	schema.Tables = append(schema.Tables, table)
+	db.Schemas = append(db.Schemas, schema)
+
+	output := writeDatabaseOutput(t, db)
+
+	if !strings.Contains(output, "USING gist (geom)") {
+		t.Fatalf("expected gist index to rely on the PostGIS default operator class, got:\n%s", output)
+	}
+}
+
+func TestWriteDatabase_GistIndexHonoursRequestedSpatialOperatorClass(t *testing.T) {
+	db := models.InitDatabase("testdb")
+	schema := models.InitSchema("public")
+
+	table := models.InitTable("places", "public")
+	geom := models.InitColumn("geom", "places", "public")
+	geom.Type = "geometry(PointZ,4326)"
+	table.Columns["geom"] = geom
+
+	table.Indexes["idx_places_geom_nd"] = &models.Index{
+		Name:    "idx_places_geom_nd",
+		Type:    "gist",
+		Columns: []string{"geom"},
+		Comment: "opclass=gist_geometry_ops_nd",
+	}
+
+	schema.Tables = append(schema.Tables, table)
+	db.Schemas = append(db.Schemas, schema)
+
+	output := writeDatabaseOutput(t, db)
+
+	if !strings.Contains(output, "USING gist (geom gist_geometry_ops_nd)") {
+		t.Fatalf("expected requested spatial operator class, got:\n%s", output)
+	}
+}
+
+func TestGenerateDatabaseStatements_VectorIndexIncludesOperatorClassAndParameters(t *testing.T) {
+	db := buildVectorSpatialSchema("hnsw", "opclass=vector_ip_ops; with (m=16)")
+
+	writer := NewWriter(&writers.WriterOptions{})
+	statements, err := writer.GenerateDatabaseStatements(db)
+	if err != nil {
+		t.Fatalf("GenerateDatabaseStatements failed: %v", err)
+	}
+
+	joined := strings.Join(statements, "\n")
+	for _, want := range []string{
+		"CREATE EXTENSION IF NOT EXISTS vector",
+		"CREATE EXTENSION IF NOT EXISTS postgis",
+		"USING hnsw (embedding vector_ip_ops) WITH (m = 16)",
+	} {
+		if !strings.Contains(joined, want) {
+			t.Fatalf("expected statements to contain %q, got:\n%s", want, joined)
+		}
+	}
+}
+
+func TestIndexStorageParameters(t *testing.T) {
+	tests := []struct {
+		name    string
+		comment string
+		want    string
+	}{
+		{"empty", "", ""},
+		{"no with clause", "opclass=vector_cosine_ops", ""},
+		{"single parameter", "with (lists=100)", "lists = 100"},
+		{"multiple parameters", "WITH (m = 16, ef_construction = 64)", "m = 16, ef_construction = 64"},
+		{"quoted value kept", "with (fillfactor='90')", "fillfactor = '90'"},
+		{"bm25 key field", "with (key_field='id')", "key_field = 'id'"},
+		{"dollar quoted value", "with (options = $$[build.internal]\nlists = [4096]$$)", "options = $$[build.internal]\nlists = [4096]$$"},
+		{"dollar quoted value with parens", "with (options = $$f(x)$$, m = 16)", "options = $$f(x)$$, m = 16"},
+		{"prose dropped", "with (lists=100, please drop everything)", "lists = 100"},
+		{"unterminated quote dropped", "with (key_field='id)", ""},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := indexStorageParameters(tt.comment); got != tt.want {
+				t.Errorf("indexStorageParameters(%q) = %q, want %q", tt.comment, got, tt.want)
+			}
+		})
+	}
+}

tests/postgres/issue21/schema.sql

@@ -0,0 +1,30 @@
+CREATE TABLE public.users (
+    id integer NOT NULL,
+    username varchar(255) NOT NULL,
+    email varchar(255) NOT NULL,
+    created_at timestamptz NOT NULL,
+    profile_id integer
+);
+
+CREATE TABLE public.profiles (
+    id integer NOT NULL,
+    bio text,
+    created_at timestamptz NOT NULL
+);
+
+CREATE TABLE public.posts (
+    id integer NOT NULL,
+    user_id integer NOT NULL,
+    title varchar(255) NOT NULL,
+    body text,
+    published_at timestamptz,
+    view_count integer DEFAULT 0
+);
+
+ALTER TABLE public.users ADD PRIMARY KEY (id);
+ALTER TABLE public.profiles ADD PRIMARY KEY (id);
+ALTER TABLE public.posts ADD PRIMARY KEY (id);
+
+CREATE UNIQUE INDEX posts_user_id_title_idx ON public.posts (user_id, title);
+
+ALTER TABLE public.users ADD CONSTRAINT users_profile_id_fkey FOREIGN KEY (profile_id) REFERENCES public.profiles (id);

tests/postgres/issue21/spec.dbml

@@ -0,0 +1,28 @@
+Table users {
+    id integer [pk, not null]
+    username varchar(255) [not null]
+    email varchar(255) [not null]
+    created_at timestamptz [not null]
+    profile_id integer
+}
+
+Table profiles {
+    id integer [pk, not null]
+    bio text
+    created_at timestamptz [not null]
+}
+
+Table posts {
+    id integer [pk, not null]
+    user_id integer [not null]
+    title varchar(255) [not null]
+    body text
+    published_at timestamptz
+    view_count integer default 0
+
+    Indexes {
+        (user_id, title) [unique]
+    }
+}
+
+Ref: users.profile_id > profiles.id