tree-sitter-okf

A tree-sitter grammar for Open Knowledge Format (OKF) v0.2 documents: markdown with YAML frontmatter, organised into knowledge bundles.

Status: 0.2.0. The version follows OKF: 0.2.x targets OKF v0.2 (see docs/releasing.md). It implements the whole grammar specification (milestones M0–M5). See CHANGELOG.md for what is verified and the known limitations.

Why a grammar of its own?

.md files in an OKF bundle are more than markdown:

1. The document shape is OKF's. A frontmatter block plus a body, with rules about where frontmatter is and isn't allowed.
2. The profile is OKF's. GFM tables (for # Schema), footnotes (for per-claim attribution), fenced and indented code, and cross-links. tree-sitter-markdown has no footnotes and gates GFM behind build flags.
3. The semantic join crosses the frontmatter/body boundary. Frontmatter sources[].id is joined to footnote labels in the body. Only a single tree can express that in one query.

So this grammar parses the whole document into one tree:

---
type: Metric
tags: [finance]
sources:
- id: gl
  resource: /tables/gl
---

# Gross Margin

Revenue minus COGS.[^gl]

[^gl]: General ledger.

(source_file
  frontmatter: (frontmatter
    (block_mapping
      (block_mapping_pair key: (plain_scalar) value: (plain_scalar))
      (block_mapping_pair key: (plain_scalar) value: (flow_sequence (plain_scalar)))
      (block_mapping_pair key: (plain_scalar)
        value: (block_sequence
          (block_sequence_item
            (block_mapping
              (block_mapping_pair key: (plain_scalar) value: (plain_scalar))
              (block_mapping_pair key: (plain_scalar) value: (plain_scalar))))))))
  body: (body
    (section
      (atx_heading (atx_h1_marker) heading_content: (inline))
      (paragraph (inline (footnote_reference label: (footnote_label))))
      (footnote_definition label: (footnote_label) (paragraph (inline))))))

• Frontmatter is parsed natively as OKF-YAML, a documented subset of YAML. Anything outside it becomes a yaml_unsupported node, never an ERROR. A missing closing --- is a MISSING "---", so tools can say exactly what is wrong.
• The body uses the rules of tree-sitter-markdown, with block and inline merged into one grammar. GFM is always on, footnotes are added, and node names match upstream. Existing markdown queries and habits carry over. The deltas are listed in vendor/tree-sitter-markdown/MERGE.md.
• Every node type is documented in docs/node-types.md.

Installing

The packages are not published yet. Until they are, install from this repository (for example npm install github:antstanley/tree-sitter-okf, or pip install git+https://github.com/antstanley/tree-sitter-okf).

The generated parser uses tree-sitter ABI 14 (spec D12). That is the highest ABI Neovim 0.10 loads, and every current runtime supports it.

const Parser = require('tree-sitter');
const OKF = require('tree-sitter-okf');

const parser = new Parser();
parser.setLanguage(OKF);
const tree = parser.parse(source);
console.log(tree.rootNode.toString());

import tree_sitter, tree_sitter_okf

parser = tree_sitter.Parser(tree_sitter.Language(tree_sitter_okf.language()))
tree = parser.parse(source.encode())

Queries

The OKF queries use only node patterns and #match?, so they give the same results in every host.

Host helpers

Some OKF facts need something the parse tree does not have: a filename, a clock, the absence of a key, or what a value means. The Node and Python bindings ship reference helpers for them. The helpers classify documents, compute concept ids, turn frontmatter into a plain value, and report trust tier, lifecycle status and staleness. They also join citations to footnotes and produce OKF conformance findings.

const { okf } = require('tree-sitter-okf');
okf.classify('metrics/gross-margin.md', tree); // 'concept'
okf.trustTier(tree);                           // 'human-reviewed'
okf.citations(tree).unresolved;                // footnotes with no source
okf.conformance('metrics/gross-margin.md', tree);

The helpers are specified in docs/host-helpers.md and pinned by shared fixtures, so the two implementations agree.

Editor setup

OKF documents are .md files, and okf is a superset of the markdown that tree-sitter-markdown parses. It is safe to register okf for markdown buffers, ideally only in OKF bundles (spec D11).

Neovim (nvim-treesitter)

-- nvim-treesitter `master` branch API
local parsers = require('nvim-treesitter.parsers').get_parser_configs()
parsers.okf = {
  install_info = {
    url = 'https://github.com/antstanley/tree-sitter-okf',
    files = { 'src/parser.c', 'src/scanner.c' },
    branch = 'main',
  },
}
vim.treesitter.language.register('okf', 'markdown')

Then run :TSInstall okf, and copy queries/*.scm into ~/.config/nvim/queries/okf/. nvim-treesitter does not fetch queries for parsers it does not ship. To use okf only inside bundles, register it from an autocommand that checks for the bundle's root index.md, instead of globally.

Helix

# languages.toml
[[language]]
name = "okf"
scope = "source.okf"
file-types = ["md"]
roots = ["index.md"]
injection-regex = "okf"

[[grammar]]
name = "okf"
source = { git = "https://github.com/antstanley/tree-sitter-okf", rev = "<commit sha>" }

Helix needs a commit SHA for rev, so use the commit of a release tag.

Then run hx --grammar fetch && hx --grammar build, and copy queries/*.scm into ~/.config/helix/runtime/queries/okf/.

Command line

npx tree-sitter parse path/to/concept.md
npx tree-sitter query queries/okf/fields.scm path/to/concept.md

Development

npm install                 # tree-sitter CLI and Node bindings
npm run generate            # tree-sitter generate --abi 14
npm test                    # corpus: tree-sitter test
script/test                 # every suite (below); --bench adds the benchmark

script/test needs a Python environment with tree-sitter, PyYAML, tree-sitter-yaml and this package installed (pip install -e .).

Test fixtures in test/fixtures/bundles/ are the four official OKF example bundles, pinned by LOCK.json and refreshed with script/sync-fixtures. Upstream markdown is refreshed with script/revendor (see MERGE.md).

Releases are cut with changesets. Run npx changeset for every user-facing change. Tagging the merged "version packages" PR publishes to npm, crates.io and PyPI by trusted publishing (see docs/releasing.md).

The experimental [[wiki link]] and #tag dialects are off by default. Build them with OKF_DIALECT_WIKILINK=1 OKF_DIALECT_TAGS=1 npm run generate.

Repository layout

grammar.js               assembles the layers below
grammar/                 block.js, inline.js, common.js (merged markdown), frontmatter.js (OKF-YAML)
src/scanner.c            the external scanner: markdown block + inline, OKF-YAML frontmatter
queries/                 editor queries and the OKF query library
bindings/                C, Go, Node, Python, Rust, Swift (+ host helpers for Node, Python)
docs/                    grammar spec, OKF-YAML subset, node types, host helpers
test/                    corpus, fixture bundles, query and helper fixtures
vendor/tree-sitter-markdown/   pinned upstream sources and MERGE.md
script/                  test, differential, property, bench and maintenance scripts
bench/                   benchmark baseline

License

MIT (see LICENSE). The markdown rules derive from tree-sitter-markdown (MIT). The OKF example bundles used as test fixtures are Apache-2.0. See NOTICE.