gpt2giga v0.2.6a1
Pre-release
Pre-release
[0.2.6a1] - 2026-07-28
Added
- Normalized protocol bridge v1: a versioned semantic-loss matrix and mandatory
admit_protocol_bridge_request()validate OpenAI-, Anthropic-, and Gemini-shaped requests before I/O and reject unsupported or unmodeled semantic loss. - OpenAI-compatible upstream adapter: the internal Harness-owned execution path supports exact profile/model binding, strict model discovery, bounded Chat Completions streaming, function tools, usage, cancellation, and normalized transport/provider errors for reviewed OpenAI-compatible and vLLM profiles.
- Anthropic normalization: Messages, SSE, function tools, stop reasons, usage, errors, and
count_tokensnow have direct request/response projections through the normalized core when normalization mode is enabled. - Gemini bridge contracts: function calls/results, JSON Schema output, typed inline image references, streaming, usage, safety errors, model discovery, and
countTokensare frozen by golden fixtures and the shared normalized model.
Changed
- Repository separation: the source repository, documentation site,
automation, package metadata, and coverage baseline now describe only the
gpt2gigacompatibility gateway. GigaLoom (formerly Unified Harness) moved
tokrakenalt/gigaloom, while
legacy source documentation URLs retain migration notices. - Normalized mode:
GPT2GIGA_NORMALIZATION_MODE=onnow covers the admitted Anthropic Messages v1 subset and Gemini token counting; the inertGPT2GIGA_EXPERIMENTAL_NORMALIZED_LAYERkey was removed. - Runtime dependencies:
httpxis now a declared direct gateway dependency for the OpenAI-compatible upstream transport.
Fixed
- Streaming closure: malformed or incomplete SSE, events after a terminal choice, invalid usage ordering, disconnects, timeouts, and provider failures now end in stable fail-closed error/cancellation contracts without inventing token counts.
- Gateway security: request-scoped Harness model overrides require a protocol-bound HMAC signature; forwarded client IPs are accepted only from trusted proxies, while error text and protocol extensions enter traffic/OTel data only under explicit content capture with redaction and truncation.