core/parsing/function_gemma_wire library

FunctionGemma's wire format, in one place.

Both halves of the protocol live here: the tokens, and the format_argument macro of the model's own chat_template.jinja. Declarations (chat.dart) and tool responses (extensions.dart) both render through it, so the two cannot drift apart the way the encoder and decoder once did.

Constants

functionGemmaEndCall → const String

functionGemmaEndDecl → const String

functionGemmaEndResp → const String

functionGemmaEscape → const String

functionGemmaStartCall → const String

functionGemmaStartDecl → const String

functionGemmaStartResp → const String

Functions

functionGemmaArgument(dynamic value, {bool escapeKeys = true}) → String

The template's format_argument macro: strings are escape-wrapped, booleans and numbers stay bare, lists and maps recurse.

functionGemmaDictsort(Iterable<String> keys) → List<String>

Jinja's dictsort, which defaults to case_sensitive=False. A plain .sort() would put Beta before alpha; the template does the reverse.

functionGemmaDouble(double value) → String

Python's str(float). The template is Jinja, so every number in the prompt was formatted by Python. Both languages print the shortest round-trip digits, but they switch to exponent notation at different magnitudes: Python below 1e-4 and from 1e16, Dart below 1e-6 and from 1e21.

functionGemmaFold(String key) → String

Python's str.lower(), which Jinja's dictsort folds keys with. Dart maps İ (U+0130) to a plain i; Python appends a combining dot, which sorts after it.

functionGemmaResponseBody(Object? response) → String

The body of response:NAME{...}.

functionGemmaScalar(dynamic value) → String

Jinja renders a bare {{ value }} through Python's str(), so booleans capitalise, null becomes None, and floats follow Python's notation.