Rewrote looksLikeMap from 6 layered heuristics down to 3 simple rules:
1. All numeric keys → map (certain)
2. Any key composed of words → struct (certain)
3. Everything else → map (safe default)
The asymmetry is deliberate: a map misidentified as a struct explodes
into hundreds of fields, while a struct misidentified as a map is
compact and easy to fix.
Word detection splits keys on _ and camelCase, checks segments for
pronounceability (vowel/consonant rhythm). Short words (id, at, on)
only count when paired with a strong 3+ char word.
Also adds raw path type naming with Item suffix for collections
(FriendsItem instead of Friend), real API test fixtures from PokeAPI
and SWAPI, and UUID-based test data replacing synthetic abc123 keys.
Add tools/jsontypes library and tools/jsontypes/cmd/jsonpaths CLI.
Given a JSON sample (file, URL, or stdin), walks the structure,
detects maps vs structs, infers optional fields from multiple
instances, and produces typed definitions.
Output formats (--format):
- json-paths: flat type path notation (default)
- go: struct definitions with json tags and union support
- typescript: interfaces with optional/nullable fields
- jsdoc: @typedef annotations
- zod: validation schemas with type inference
- python: TypedDict classes
- sql: CREATE TABLE with FK relationships
- json-schema: draft 2020-12
- json-typedef: RFC 8927
Features:
- Interactive prompts for ambiguous structure (map vs struct, same
vs different types), with --anonymous mode for non-interactive use
- Answer replay: saves prompt answers to .answers files for iterative
refinement
- URL fetching with local caching and sensitive param stripping
- Curl-like auth: -H, --bearer, --user, --cookie, --cookie-jar
- Discriminated union support with sealed interfaces, unique-field
probing, and CHANGE ME comments for type/kind discriminators
- Extensive round-trip compilation tests for generated Go code
