jx — JSON encoding and decoding

github.com/go-faster/jx is a zero-allocation, RFC 7159 JSON library used as the foundation of ogen. It trades the convenience of encoding/json for direct control over every byte — which means you get to make choices that matter for correctness and performance.

---

Three rules that matter most

1. StrBytes, Num, and Raw reference the internal buffer

These methods return slices that alias the decoder's read buffer. They are valid only until the next decoder call.

// BUG: key is overwritten before use
keys = append(keys, key)

// OK: copy immediately
keys = append(keys, string(key))  // or append([]byte{}, key...)

The same applies to jx.Num returned by d.Num() and jx.Raw returned by d.Raw(). Use the *Append variants to copy into your own buffer:

n, err = d.NumAppend(n[:0])   // safe: appends into your slice
raw, err = d.RawAppend(raw[:0])

2. Capture enables multi-pass decoding — byte buffers only

d.Capture(f) saves decoder state, runs f, then rolls back. Use it to peek at a discriminator field before deciding how to decode:

var kind string
if err := d.Capture(func(d *jx.Decoder) error {
    return d.ObjBytes(func(d *jx.Decoder, key []byte) error {
        if string(key) == "type" {
            v, err := d.StrBytes()
            if err != nil {
                return err
            }
            kind = string(v) // copy — buffer is temporary
            return err
        }
        return d.Skip()
    })
}); err != nil {
    return err
}

// Decoder is reset to before the Capture call.
switch kind {
case "foo":
    return s.Foo.Decode(d)
case "bar":
    return s.Bar.Decode(d)
}

Capture does not work with io.Reader decoders. It only works when the decoder was created with DecodeBytes / DecodeStr / ResetBytes (i.e., the full input is already in memory).

3. jx.Writer is faster but requires manual commas

jx.Encoder tracks a comma-state stack and inserts commas automatically. jx.Writer omits that bookkeeping — ~1.7× faster in benchmarks, but every non-first element must be preceded by an explicit w.Comma() call. Use Writer in generated or hot-path code where the structure is statically known; use Encoder for hand-written code.

---

Decoder

Creating a decoder

d := jx.DecodeBytes(data)        // byte slice
d := jx.DecodeStr(`{"k":"v"}`)   // string literal
d := jx.Decode(r, 512)           // io.Reader, buffer size 512

// Pool reuse — decoder is reset on PutDecoder
d := jx.GetDecoder()
defer jx.PutDecoder(d)
d.ResetBytes(data)

Decoding objects

Prefer ObjBytes over Obj — it avoids allocating a string for every key. The key slice is only valid inside the callback; using string(key) in the switch statement is safe and cheap.

func (s *MyStruct) Decode(d *jx.Decoder) error {
    return d.ObjBytes(func(d *jx.Decoder, key []byte) error {
        switch string(key) {
        case "name":
            v, err := d.Str()
            if err != nil {
                return err
            }
            s.Name = v
        case "count":
            v, err := d.Int()
            if err != nil {
                return err
            }
            s.Count = v
        default:
            return d.Skip() // must skip unknown fields or the decoder stalls
        }
        return nil
    })
}

Decoding arrays

return d.Arr(func(d *jx.Decoder) error {
    v, err := d.Str()
    if err != nil {
        return err
    }
    s.Items = append(s.Items, v)
    return nil
})

Peeking at the next type

d.Next() returns the type of the next value without consuming it.

switch d.Next() {
case jx.String:
    v, err := d.Str()
    ...
case jx.Null:
    if err := d.Null(); err != nil { return err }
    // value is null
case jx.Number:
    v, err := d.Int()
    ...
}

Zero-copy string

StrBytes avoids a heap allocation by returning a slice into the decoder buffer. Use it immediately — it is only valid until the next decoder call. Do not store the result.

raw, err := d.StrBytes()
if err != nil {
    return err
}
// Use raw here — e.g. switch, hex.Decode, or hand to a library that
// accepts []byte (like uuid.ParseBytes). Do NOT store raw in a field.

When you need to retain the string but want allocation-free reuse across calls, use StrAppend with your own buffer:

var buf []byte // reuse across iterations

buf, err = d.StrAppend(buf[:0]) // appends decoded string into buf
name = string(buf)              // copy once, into the final destination

Numbers

jx.Num is a []byte type. Like StrBytes, the value returned by d.Num() aliases the decoder buffer. Use NumAppend to own the bytes.

n, err := d.NumAppend(n[:0]) // append into your own slice
i64, err := n.Int64()
f64, err := n.Float64()

String-encoded numbers

Some APIs encode numbers inside JSON strings ("count": "42"). Decode them by extracting the string bytes and feeding them to a nested decoder — no allocation because StrBytes is zero-copy:

raw, err := d.StrBytes() // e.g. []byte("42")
if err != nil {
    return err
}
n, err := jx.DecodeBytes(raw).Int()

---

Encoder

Commas are inserted automatically. The zero value is valid.

Encoding a struct

func (s *MyStruct) Encode(e *jx.Encoder) {
    e.ObjStart()
    defer e.ObjEnd()

    e.FieldStart("name")
    e.Str(s.Name)

    e.FieldStart("count")
    e.Int(s.Count)

    if s.Optional != "" {   // omit zero-value optional fields
        e.FieldStart("optional")
        e.Str(s.Optional)
    }
}

Encoding arrays

e.ArrStart()
for _, v := range items {
    e.Str(v)   // comma inserted automatically before each non-first element
}
e.ArrEnd()

Nullable values

e.FieldStart("value")
if ptr == nil {
    e.Null()
} else {
    e.Str(*ptr)
}

Pool reuse

e := jx.GetEncoder()
defer jx.PutEncoder(e)   // resets on return

s.Encode(e)
result := append([]byte{}, e.Bytes()...)  // copy before defer runs

Primitive methods

e.Str(v string)
e.ByteStr(v []byte)   // encodes bytes as a JSON string, no allocation
e.Int(v int) / e.Int64 / e.UInt64 / e.Int32 ...
e.Float64(v float64)
e.Bool(v bool)
e.Null()
e.Raw(v []byte)       // embed pre-encoded JSON verbatim
e.RawStr(v string)    // same, from string
e.Base64(v []byte)
e.Num(v jx.Num)

Encoding pre-formatted values efficiently

When a value has a fixed or bounded byte representation, you can encode it with zero allocations by formatting into a stack-allocated array and writing the result as raw JSON.

Fixed-size value (e.g. UUID — 38 bytes including quotes):

// Pre-encode with quotes included; write as raw bytes.
const quoted = 38 // 36 chars + 2 quotes
var dst [quoted]byte
dst[0] = '"'
dst[quoted-1] = '"'
hexEncode((*[36]byte)(dst[1:37]), id) // your hex-encode function
e.Raw(dst[:])

Variable but bounded value (e.g. timestamp, duration):

// AppendFormat into a stack buffer; ByteStr encodes it as a JSON string.
var buf [64]byte
b := v.AppendFormat(buf[:0], time.RFC3339)
e.ByteStr(b)   // writes "...", no allocation

The key rule: e.Raw(b) embeds b verbatim (caller provides quotes if needed); e.ByteStr(b) wraps b in JSON string quotes and escaping.

---

Writer (faster, manual commas)

jx.Writer exposes Buf []byte directly and skips the comma-state stack, making it the fastest way to produce JSON. You are responsible for calling w.Comma() between every pair of adjacent values.

Encoding a struct with Writer

For statically known fields, embed the comma and colon directly into RawStr literals — this is what code generators do:

func (s *MyStruct) Write(w *jx.Writer) {
    w.ObjStart()
    w.RawStr(`"name":`)       // first field — no leading comma
    w.Str(s.Name)
    w.RawStr(`,"count":`)     // subsequent fields — leading comma in literal
    w.Int(s.Count)
    w.ObjEnd()
}

For dynamic fields (map keys, optional fields), track first-element state explicitly:

w.ObjStart()
first := true
for k, v := range m {
    if !first {
        w.Comma()
    }
    first = false
    w.FieldStart(k)
    w.Str(v)
}
w.ObjEnd()

Encoding arrays with Writer

w.ArrStart()
for i, v := range items {
    if i != 0 {
        w.Comma()
    }
    w.Str(v)
}
w.ArrEnd()

Pool reuse

w := jx.GetWriter()
defer jx.PutWriter(w)

s.Write(w)
result := append([]byte{}, w.Buf...)  // copy before defer runs

---

Encoder vs Writer

	jx.Encoder	jx.Writer
Commas	Automatic	Manual (w.Comma())
Indentation	e.SetIdent(n)	Not supported
Benchmark speed	Fast	~1.7× faster
When to use	Hand-written code	Generated / hot-path code

---

Conventional interfaces

ogen-generated code uses these signatures — implement them on your types to stay interoperable:

func (s *MyStruct) Encode(e *jx.Encoder) { ... }  // encoding
func (s *MyStruct) Decode(d *jx.Decoder) error { ... }  // decoding

---

Common mistakes

Mistake	Fix
Storing StrBytes / Num / Raw result beyond the callback	Copy: string(b) or append([]byte{}, b...)
No d.Skip() in ObjBytes default branch	Decoder stalls; always skip
Capture on an io.Reader-backed decoder	Only works with byte-backed decoders
Adding your own commas with Encoder	Encoder is automatic; extra commas corrupt output
Forgetting commas with Writer	Writer never adds commas; every non-first value needs w.Comma()
Forgetting e.Bytes() before PutEncoder	Buffer is reset on pool return; copy first