barkup
barkup lets an application define a typed tree grammar (pages, sections, components, whatever the domain calls for) and read and write it as an HTML dialect. The codec is config-driven: you declare the node types, their attributes, and their coercions, and barkup gives you parsing, serialization, byte-for-byte id preservation, and round-trip guarantees over that grammar.
The reason it exists is the editing era. When both humans and LLM agents author the same structured document, the format they meet in matters enormously. Our barkup-bench research series measured how models handle typed-tree editing across formats and strategies, and barkup is the library those findings shipped into: stable ids that survive rewrites, focused views so the model sees only the subtree it is editing, and validation of the whole document on every change.
It is the production format behind Replicator's design-tree editor, and the subject of one of the larger pre-registered LLM benchmark series we know of.
bun add @kevinpeckham/barkupnpm →GitHub →Benchmark dashboard →
What It Does
- Config-driven grammar: declare node types, attributes, and coercions once
- Byte-for-byte id preservation through parse and serialize cycles
- Whole-tree rewrite and id-anchored patch editing strategies, both benchmarked
- Focused views: hand the model one subtree without losing document identity
- Round-trip guarantees enforced by test, in the barkdown family tradition