Changed #

• LICENSE carries a copyright notice for the work done on this line, next to the upstream one it has always kept, as the BSD-3-Clause terms require. AUTHORS.md already recorded who did what; the licence file now says the same thing.
• The README says where to look for both, rather than describing the licence file as untouched.

No code changed.

First release published to pub.flutter-io.cn. graphql_schema3 is now a hosted dependency rather than a git one, which is what publishing requires and what lets a consumer resolve the whole stack from pub.

Fixed #

• Descriptions and deprecation reasons are escaped before they are written into the generated source. code_builder escapes the quote but leaves the backslash and the dollar sign alone, so a doc comment mentioning a Windows path emitted 'C:\temp', where \t is a tab, and one mentioning $rate emitted a string interpolation that does not compile. Every piece of authored text now goes through safeLiteralString.
• The enum builder assembles code_builder expressions like the rest of the generator instead of pasting its type together as source text and escaping the single quote by hand.
• One rule decides a description. @GraphQLDocumentation(description:) is an explicit override and wins over the doc comment everywhere; the schema preferred the annotation while the generated header preferred the comment, so a class carrying both was described two different ways in the same file.
• The resolved-type cache is keyed on the position as well as the type name. A union first resolved as a return value answered for the same union seen later as a parameter, where it should have been rejected.

Added #

• A strip_class_prefixes builder option. The class name prefix dropped when naming a GraphQL type was written into the generator; it belongs to whoever names the classes. Left unset it keeps stripping Bmc, so nothing changes for an existing build; set it to [] to strip nothing.
• Tests for the naming rule, the option, and the escaping. 39 in all.

Removed #

• collectFields, resolverImportFor, isTypeJsonKey, isTypeJsonValue and jsonValueTypeChecker: none had a caller. The unused packageName parameter went with them, along with three locals a nested block re-declared over the identical ones already in scope.

Changed #

• An input class no longer runs inferType over every field only to discard the answer: an input field takes its type from graphQLTypeForInputField, which is the only one that knows how to tie a recursive input back to itself.

Note on generated output #

Regenerating the 179 models of the consuming project gives byte-identical output on 161 of the 162 generated files. The one difference is the fix: a doc comment holding an authored \n inside a JSON example used to reach the schema as a real line break, and now reaches it as the two characters the comment actually spells.

Removed #

• The angel3_serialize_generator, angel3_serialize and angel3_model dependencies, and with them belatuk_code_buffer and charcode. angel3_model was never imported. Of angel3_serialize_generator the generator used four symbols: dateTimeTypeChecker, already shadowed by a local one and hidden from both imports; serializableTypeChecker, whose five call sites tested for an @serializable annotation no model carries, so every branch was dead; and BuildContext / buildContext, replaced by GraphQLBuildContext. That last one only ever answered two questions, the name of the generated variable and the JSON key of a field, and its resolveFieldName was a no-op here because autoSnakeCaseNames was always false. The point was never the six packages: it was that angel3 dictated which analyzer this package could use, which is what held it seven majors back for months.
• The recase dependency. camelCase and snakeCase were the only conversions used; they now live in lib/src/string_case.dart with the same word-splitting rules, so the generated names are unchanged.
• The build_config and collection dependencies, neither of which was imported.
• lib/src/extensions.dart. It papered over the Element / Element2 split and the current analyzer covers all of it, EnumElement.constants included.
• The angel3_serialize.g.part entry in required_inputs. Consumers no longer run the angel3_serialize builders at all, which drops two phases from their build.
• The naming rule inherited from angel3, which stripped a leading _ or a leading/trailing entity from the class name before deriving the variable name. A leading _ is dropped anyway by the word splitting; a class named FooEntity now yields fooEntityGraphQLType rather than fooGraphQLType.

Changed #

• analyzer 7 to 14, source_gen 3 to 4, build 3 to 4 and build_runner 2.7 to 2.16. Migrated to the current element model: no more Element2 / element3 / fields2, and isSynthetic on a field is now !isOriginDeclaration.
• The SDK floor moves from 3.8 to 3.11, which is what analyzer 14 requires. This package is a dev dependency, so the floor does not reach an app.
• A doc comment on a field or a method now becomes its GraphQL description. It always should have: applyDescription is called for every field. The compatibility layer around Element2 answered null for anything but a class, so those descriptions were silently dropped. Class descriptions are unaffected. Nothing else about the generated schema changes: field names, types, nullability, resolvers and enum values are byte for byte what they were.

Added #

• A test suite: 32 tests over the builder and the case conversions. There was none. The builder tests stub the two annotation libraries by URL, so they need neither graphql_schema3 from git nor a network.
• A CI workflow running format, analyze and test on the stable and beta SDKs.

• Fixed the type resolvers, and carried description and @Deprecated over onto enum values.

• Initial release