@plurnk/plurnk-mimetypes-text-gherkin

text/x-gherkin (Gherkin / .feature BDD) mimetype handler for the plurnk ecosystem. Lean-library tier, backed by @cucumber/gherkin — the cucumber org's reference parser.

why @cucumber/gherkin

Gherkin has no usable tree-sitter grammar, so Tiers 1–3 (clean WASM / dirty WASM / antlr4ng) don't apply. The honest backend is the format's own reference parser: @cucumber/gherkin is pure JS/ESM, parses synchronously (no WASM init), and natively handles every Gherkin construct including localization (80 dialects via # language: xx). The dep precedent is text-pdf's pdfjs — a focused, canonical library beats a hand-rolled scanner once the parse surface (docstrings, data tables, outlines, dialects) gets non-trivial.

what it emits

extractRaw → MimeSymbol[] with 1-indexed line / endLine / column. The AST carries start locations only; endLine is derived as the deepest descendant line (docstrings reach their closing delimiter, tables reach their last row).

| construct | kind | name | container | |---|---|---|---| | Feature | module | feature name | — (top-level) | | Rule | module | rule name | feature name | | Scenario / Scenario Outline | function | scenario name | feature or feature.rule | | Background | function | the keyword (Background) | feature or rule path | | step | field | "{keyword} {text}" (e.g. Given a paid invoice) | dotted scenario path | | Examples: block | field | Examples: {name} (or Examples) | the outline's path |

container is the dot-joined path of enclosing named symbols (e.g. a step under a rule scenario carries Feature.Rule.Scenario). Tags (@smoke) are dropped from symbols — they're render noise; leave them to query.

Localization works for free: keyword extraction reads the parsed keyword text, so a # language: fr file surfaces Fonctionnalité / Scénario / Soit … keywords without special handling.

Malformed .feature input → extractRaw returns [] (mirrors the AntlrExtractor catch-to-empty policy; agents see malformed input rather than an error).

refs-free

This handler emits no references. Gherkin steps bind to step-definitions living in other codebases (Java/Ruby/JS) via regex or cucumber-expressions — nothing in a .feature corpus is name-joinable in-corpus, so there is no honest code-graph edge to emit. References are a code-graph concept; this format has none to offer.

license

MIT.