Segno

Typed values, in any order

Pick a kind of rule, then type. The same notation means different things in different places: each place has a schema that says which fields it takes, and every error names what was expected.

Aliases, shortcuts, and one spelling per document

[x], {done} and {yes} all mean the same state. That is kind per span and messy per document, so Segno reports every span written differently from the document's most common spelling — with a fix. One state mark per line:

Try to make it slow

These are the inputs that took other parser libraries from one second to over sixteen, at 2k, 8k and 32k characters. Each step is 4 times longer, so steady parsing takes about 4 times as long; a parser that blows up takes 16 times as long or more.

Build your own parser experimental

Underneath the notation is a grammar engine. Segno only builds a parser it can prove reads in one pass, and says no to a grammar that would need to guess. A grammar is a few rules built from seq, alt, many, opt, ref and character sets; node marks the pieces you want back. Edit the grammar or the text: it compiles and parses as you type. Pick a refused grammar to see the compiler say no.

First user: Lattice

Lattice renders slide decks from Markdown, and its inline code spans carry directives. It had grown 27 hand-written grammars for them, with 21 sigils between them; it is moving every one onto this notation. Some of them, beside what each becomes — try any row in the editor above.

How it compares

Segno beside the six parser libraries its first user evaluated, and beside the hand-written parsers it replaces — every candidate in its own process, on identical inputs, on one machine. Time per input on real spans, then the worst case on hostile input.

Measured 2026-09-29, Node 22, on a shared cloud machine: absolute times move with the machine, the ratios carry. Bracket lists: the 64 in Lattice's decks, parsed into parts. Inline spans: all 4,645, 99% of them ordinary code. "Did not finish": a smaller input already took over half a second, so the next was stopped. Lezer's grammar set does not cover bracket lists. Segno's flat best case on nested brackets is its 31-level cap refusing the input, not a parse. Reproduce: npm run parser:bakeoff:versus.