Release v0.3.0 · ai-forever/gpt2giga

[0.3.0] - 2026-08-03

Universal Provider Bridge: startup-owned provider profiles and public model aliases select an exact provider/model/capability/matrix context without request-supplied destinations, credentials, or fallback.
OpenAI Responses ownership: native GigaChat Responses remains the stable compatibility path for hosted tools, attachments, state, v1/v2, and SSE. Route-selected normalized Responses is technical_preview until its accepted corpus reaches native parity.
Provider adapters: added bounded OpenAI-compatible, Anthropic, and Gemini upstream adapters with hermetic fake-provider conformance, streaming, tools, usage, and stable errors; public-protocol cells without integrated routes remain explicitly blocked before provider I/O.
Supervisor contract: --config, redacted --inspect-config, /ready, /bridge/models, and /bridge/capabilities provide a content-free machine contract for an installed sidecar artifact.

Security and lifecycle: immutable alias selection, destination authorization, redirect/private-network guards, graceful bounded shutdown, and no-fallback behavior have dedicated security and corpus coverage.
Bridge performance: profile admission and machine manifests are materialized once, while public projections return isolated copies without reparsing provider configuration.
Documentation: added schema/provider examples, the 16-cell loss matrix, migration, rollback, and a GigaLoom-compatible supervisor sequence.
Model inventory: GigaChat models are discovered dynamically; provider-profile v2 distinguishes model_inventory: dynamic from optional exact aliases and static external-provider profiles.

Full Changelog: v0.2.7...v0.3.0