go-toml-edit

Zero-dep TOML editing library for Go with comment preservation

Why

Every Go TOML library either discards comments during parsing or provides read-only access to them. If you parse a config file, change one value, and write it back, the comments are gone. Python has tomlkit; Go had nothing. go-toml-edit fills this gap with a lossless AST that preserves every comment, blank line, and formatting detail through arbitrary edits.

What it preserves is content — comments, values, structure. Whitespace is formatting, not content: this library also reformats TOML, and Format() is strict about how output looks.

Before / After

package main

import (
	"fmt"
	"log"

	tomledit "github.com/smm-h/go-toml-edit"
)

func main() {
	doc, err := tomledit.Parse([]byte(`# Server config
[server]
host = "localhost"  # primary host
port = 8080
`))
	if err != nil {
		log.Fatal(err)
	}

	if err := doc.Set("server.port", 9090); err != nil {
		log.Fatal(err)
	}
	fmt.Print(string(doc.Bytes()))
}

Output -- comments survive the edit:

# Server config
[server]
host = "localhost"  # primary host
port = 9090

Installation

go get github.com/smm-h/go-toml-edit

No runtime dependencies. The package imports the standard library and nothing else; the module's BurntSushi/toml and toml-test requirements are test-only (the benchmark comparison and the compliance corpus). A test asserts this over the non-test sources.

Two surfaces, two questions

A parsed document is readable through two surfaces, and which one to use follows from the question being asked.

• The syntactic surface is the AST: Walk, Resolve and the concrete node types. It answers what the file contains, in the form it was written -- spellings, quoting styles, integer bases, comments, blank lines, and the source span of every construct.
• The logical surface is the read-layer: doc.Root() returns a Record whose Entry values are the document's keys in first-appearance order. It answers what the document means, with the spellings folded away. A dotted key, a [header] table and an inline table are indistinguishable through it.

Reads are spelling-blind; writes are structurally conservative. A value write touches value fragments and nothing else, and a structural construct changes only through a structural operation or an explicit Delete.

Feature Comparison

Feature	go-toml-edit	BurntSushi/toml	pelletier/go-toml/v2
Comment preservation	Yes	No	Read-only (unstable)
Round-trip editing	Yes	No	No
Set/Delete/RenameKey API	Yes	No	No
Unmarshal to struct	Yes (strict-only)	Yes	Yes
Marshal from Go values	No	Yes	Yes
Position spans on AST nodes	Yes	No	Unstable parser only
TOML 1.0 compliance	Full	Full	Full
Formatter	Yes	No	No
Document diffing	Yes	No	No
Deep merge	Yes	No	No

Quick Start

Parse and Read

doc, err := tomledit.Parse([]byte(`[server]
host = "localhost"
port = 8080
`))
host, err := doc.GetString("server.host") // "localhost", nil
port, err := doc.GetInt("server.port")    // 8080, nil

Every getter returns (T, error). A missing path, a bad path, a wrong type and a value the target cannot hold exactly are distinct diagnostics -- see Diagnostics.

Read the logical structure

for entry := range doc.Root().Entries() {
	switch entry.Kind() {
	case tomledit.EntryValue:
		node, _ := entry.Node()
		fmt.Printf("%s = %v\n", entry.Key(), node.(tomledit.Scalar).Value())
	case tomledit.EntryRecord:
		rec, _ := entry.Record()
		fmt.Printf("%s is a table of %d keys\n", entry.Key(), rec.Len())
	case tomledit.EntryRecords:
		recs, _ := entry.Records()
		fmt.Printf("%s is an array of %d tables\n", entry.Key(), len(recs))
	}
}

Edit Values

doc.Set("server.port", 9090)               // update existing key
doc.SetCreate("database.host", "db.local") // create key + intermediate table
doc.Delete("server.debug")                 // remove a key (idempotent)
doc.RenameKey("server.host", "address")    // rename the binding, every spelling

Set is a no-op if and only if the bytes it would write are the bytes already there -- so a tool that writes back what it read touches nothing. When they differ, the value is written in the library's canonical spelling, and a container value (a map, a slice, a []Pair) is replaced wholesale.

Structural editing

doc.NewTable("logging")
doc.NewArrayTable("products")
doc.AppendToArray("server.ports", 8082)
doc.RemoveFromArray("server.ports", 0)
doc.PermuteChildren("server", []int{2, 0, 1}) // order[i] = index moving to i
doc.EnsureDefaults([]tomledit.Default{
	{Path: "server.port", Value: 8080},
	{Path: "logging.level", Value: "info"},
})

Comment Management

doc.SetComment("server.port", "default: 8080")
doc.SetLeadingComments("server", []string{
	"Server configuration",
	"See docs for all options",
})

inline, err := doc.GetComment("server.port")           // "default: 8080"
leading, err := doc.GetLeadingComments("server")       // the two lines above

The getters resolve a path to the same node the setters write to, so what one writes the other reads back, and they refuse the same paths with the same error kinds. Both answer normalized text -- the content without the # and the whitespace around it -- which is exactly what the setters take. A node with no inline comment answers the empty string, one with no leading comments answers nil, and neither is an error.

Fluent Cursor

host, err := doc.Key("database").Key("host").String()
port, err := doc.Key("database").Key("port").Int()

Strict decoding

type Config struct {
	Title  string `toml:"title"`
	Server struct {
		Host string `toml:"host"`
		Port int    `toml:"port"`
	} `toml:"server"`
}
cfg, err := tomledit.Unmarshal[Config](data)

Decoding is strict and strictness is the only mode: an unknown key, an unknown table, a value of a refused kind, a value the target cannot hold exactly, and a missing toml:"...,required" key are all errors. Key matching is exact -- a document key differing only in case matches nothing. Every independent violation is collected and reported together, in document order -- with one placement rule source order cannot give: a record's missing required keys are reported after that record's own entries, and among themselves in lexicographic key order.

The entry points return the value they build -- a *T, pointing at memory they allocated -- so a failed decode leaves nothing to inspect: Unmarshal[T](data), Decode[T](doc), DecodeNode[T](node), and DecodeOver[T](doc, seed) -- the last overlaying a document on a seed the factory builds, and reporting which paths the document supplied.

Descriptor validation

For a schema known only at runtime, describe it as data instead of as a struct:

spec := &tomledit.Spec{Fields: map[string]tomledit.Field{
	"host": {Kind: tomledit.FieldKindString, Required: true},
	"port": {Kind: tomledit.FieldKindInteger},
}}
err := doc.Validate(spec)                 // diagnostics only
values, err := doc.DecodeSpec(spec)       // validated, then native Go data

DecodeSpec is atomic: it returns a map only when the document has no violations at all.

Format

formatted := doc.Format(tomledit.WithIndentWidth(2))

Format re-renders everything in canonical spellings. It preserves the writer's blank-line grouping at document and table-body level -- a run of blank lines becomes one. One gap it opens itself, unconditionally: a table header written flush against what precedes it always gets a blank line above it. What the library preserves is content -- the comments -- while whitespace is formatting, and Format is where the library is strict about output looking good, so the separation is not a caller's option. The insertion never removes a blank line and never doubles one already there.

Walk

Walk is the syntactic traversal: it hands the visitor the concrete nodes, with the spellings and trivia they were written with.

doc.Walk(func(path string, node tomledit.Node) error {
	fmt.Printf("%s = %v\n", path, node.(tomledit.Scalar).Value())
	return nil
}, tomledit.WalkLeaves)

Value() is on Scalar, the sub-interface the value-carrying node kinds implement; Node itself carries Type(), Span(), Raw(), Comment() and LeadingComments().

Position Spans

Every AST node carries its source position from the most recent parse, enabling custom semantic validation with precise line/column diagnostics:

doc.Walk(func(path string, node tomledit.Node) error {
	sp := node.Span() // 1-based line/column, plus a byte offset
	fmt.Printf("%s at %d:%d-%d:%d\n", path,
		sp.Start.Line, sp.Start.Column, sp.End.Line, sp.End.Column)
	return nil
}, tomledit.WalkLeaves)

Span() is available on all nodes: tables and array-tables (covering only the [header]), key-value pairs (key through value), keys, all scalars, arrays, inline tables, and comments. Span.End is exclusive -- the position immediately after the node's last byte.

Span policy: spans reflect the last Parse. Edit operations do not recompute them: nodes created programmatically (via Set, Merge, ...) return the zero span (Span.IsValid() reports false), and edited nodes keep their parse-time spans. Re-parse the output of Bytes() when fresh positions are needed after editing.

Files

doc, err := tomledit.ParseFile("config.toml") // remembers the filename
err = doc.WriteFile("config.toml")            // atomic, round-trip checked

A document loaded with ParseFile names the file in every diagnostic it later produces. WriteFile proves the rendered bytes re-parse to themselves before replacing the destination in one step.

Diagnostics

_, err := doc.GetInt("server.port")

if errors.Is(err, tomledit.ErrTypeMismatch) { /* ... */ }

var diag *tomledit.Error
if errors.As(err, &diag) {
	fmt.Println(diag.Kind, diag.Path, diag.Pos.Line, diag.Expected, diag.Got)
}

Every failure that depends on the document -- parse, decode, edit, access -- is an *Error with a kind, a path and a position, rendered in the compiler convention (config.toml:12:7: server.port: expected integer, got string). A decode reports its violations together as an *Errors aggregate, reachable with errors.As.

Diff

changes := tomledit.Diff(a, b)
for _, c := range changes {
	fmt.Printf("%s: %s\n", c.Kind, c.Path)
}
// removed: age
// added: email
// modified: name

Diff reads values, never spellings: two documents writing one value differently report no difference, so a document always compares equal to itself. Types are not bridged -- 1 and 1.0 are a modification.

Merge

base, _ := tomledit.Parse([]byte(`[server]
host = "localhost"
`))
defaults, _ := tomledit.Parse([]byte(`[server]
host = "0.0.0.0"
port = 8080
`))
base.Merge(defaults)
// host keeps "localhost" (already set); port added from defaults.

Performance

go-toml-edit retains the full AST -- source positions, trivia, comment nodes, and the lexeme every value was written as -- because that is what comment-preserving round-trip editing needs. A decode-only library discards all of it, so it parses faster and allocates less. Reach for a decode-only library when you only need Go values out of a TOML file; reach for this one when the comments have to survive the write-back.

Run go test -bench . to measure on your own machine. benchmarks.txt holds a captured run beside BurntSushi/toml.

API Reference

pkg.go.dev/github.com/smm-h/go-toml-edit

Deferred work

A struct-to-TOML Marshal, the remaining structural-manipulation suite (node moves, positional insertion, table/inline conversion), external node construction with spelling control, and a streaming parser are deferred rather than planned. TOML 1.1 is not supported: this library targets TOML 1.0 only. See todo/ for details.

License

MIT