Use include/transcribe/extensions.h as the default header for generated
bindings. It includes transcribe.h plus every family extension header shipped
by the install, so generators see the full public surface in one translation
unit.
Use include/transcribe.h directly only for bindings that intentionally expose
the stable generic ABI and omit family-specific extension structs and telemetry.
The header split is source organization, not a separate library boundary:
family headers include transcribe.h, do not depend on each other, and are
flattened normally by C preprocessors, bindgen, CFFI API-mode builds, cgo
preambles, and Swift/ObjC module maps.
Three binding-neutral artifacts exist so no binding depends on another binding's generated files:
include/transcribe.abihash— the public-ABI digest (sha256/16 over the normalized FFI surface: structs, enums, macros, layout, prototypes). Emitted bybindings/python/_generate/generate.py(the hash oracle) and drift-gated in CI alongside_generated.py. Every first-class binding pins this value: when the header's ABI changes, the hash moves and the binding's CI goes red until its FFI layer is regenerated or consciously reviewed. Comment-only header edits do not move it.contract.json— stamped into every native artifact directory (_native/in provider wheels; the root of extractedtranscribe-native-<tuple>bundles):version,header_hash,backends,lane. A binding validatesversion(pre-1.0: exact base match) andheader_hash(must equal the hash its FFI layer was generated against) BEFORE dlopen. This is the same contract the Python provider enforces via_contract.py.lib/transcribe-link.json— installed bycmake --install(theTRANSCRIBE_INSTALLrules): the machine-readable link interface for non-CMake consumers building from source (the Rust-syscrate'sbuild.rs). Archive order, system libs, frameworks, flags, and — forGGML_BACKEND_DLinstalls —module_dir, the directory to hand totranscribe_init_backends(). Proven per push by the link-smoke CI lane, which compiles a toy C consumer from nothing but this manifest in both static and shared postures.
Every accessor that returns a const char * (transcribe_full_text,
transcribe_detected_language, and the text field of the segment / word /
token row structs) returns a borrowed pointer into session-owned storage. See
the "Result text-pointer lifetime" block at the top of include/transcribe.h
for the full contract. The binding-relevant rule:
- Offline path: the pointer is valid until the next
transcribe_run/transcribe_stream_begin/transcribe_stream_reset/transcribe_session_freeon the same session. - Streaming path: raw result pointers, including
transcribe_full_textand rowtextfields, may be replaced by everytranscribe_stream_feed/transcribe_stream_finalizecall. Bindings that need UI-stable streaming text should usetranscribe_stream_get_text()and expose owned copies ofcommitted_textandtentative_text.
The simplest safe rule for a binding is to copy the bytes at the FFI
boundary, which the marshal-and-copy idioms each language already uses make
automatic: Python ctypes.c_char_p / .decode(), Go C.GoString, Rust
CStr::to_str().to_owned(), Swift String(cString:). A binding that instead
hands out a zero-copy view must scope it to the current callback/update turn
and document that it dies at the next stream mutation.
When adding a new family extension, update:
include/transcribe/<family>.hwith the typed struct, kind constant, andtranscribe_<family>_<name>_ext_init()function declaration.include/transcribe/extensions.hwith one new include.docs/extension-kinds.mdwith the registered FourCC value and the slot (RUNorSTREAM) the kind is legal on.