omnix 0.1.0-dev.3

An open, headless, on-device intelligence engine for Flutter applications.

0.1.0-dev.3

• Add provider-neutral Knowledge documents, chunks, access labels, sources, citations, embedding-space identities, and local or transport-ready semantic query contracts.
• Add a Knowledge coordinator for validated indexing, durable flushes, deterministic access-filtered retrieval, semantic-space compatibility, and token-budgeted RAG context assembly.
• Add a strict versioned codec for transporting semantic queries and attributed retrieval matches across future node boundaries.
• Add a Flutter Gemma Knowledge backend that preserves source attribution and access labels, uses provider-correct document/query embedding tasks, flushes durable writes, and remains independent of the host-selected vector store.
• Add transport-neutral node authentication evidence, verified principals, replaceable authenticators, and Knowledge authorization policies suitable for server-mediated or direct peer-to-peer deployments.
• Add a node Knowledge service that fails closed, rejects expired principals, intersects access grants, and clamps retrieval limits before semantic search.
• Let authenticated node Knowledge requests carry either natural-language text or a precomputed embedding under the same policy boundary, keeping future query modalities extensible without coupling authentication to transport.
• Add a versioned, discriminated node-query encoding for transporting text and embedding requests without provider types.

0.1.0-dev.2

• Add provider-neutral context budgets, token-estimator contracts, and a recent history policy that preserves mandatory system information.
• Allow conversations and agents to replay context-policy-selected history when a session opens.
• Add durable conversation snapshot and repository contracts plus a strict, versioned JSON codec with multimodal message support.
• Add a durable conversation coordinator that restores bounded active context, serializes context restoration with inference, and atomically stores complete successful turns without deleting trimmed history.
• Serialize native conversation and agent-session creation, history replay, and standalone history replacement through the shared inference scheduler.
• Allow hosts to supply an application-owned inference scheduler and retain its lifecycle ownership.
• Allow augmented model input to retain a separate durable user-visible prompt.

0.1.0-dev.1

• Establish the Omnix Flutter and Rust package foundation.
• Add a headless engine lifecycle contract and native bridge handshake.
• Add a framework-neutral, non-preemptive inference scheduler with priority, FIFO, streaming, lifecycle, and immutable state snapshot guarantees.
• Add portable model-manifest contracts and strict validation for the Omnies registry and pinned public LiteRT-LM downloads.
• Add the first thin inference boundary: a plugin-independent conversation API and a Flutter Gemma adapter for streamed text, thinking, and tool calls.
• Add an application runtime that owns engine initialization and deterministic conversation cleanup.
• Add neutral model-source, installation, progress, cancellation, storage, and integrity contracts with a LiteRT-LM-backed Flutter Gemma implementation.
• Add a one-call FlutterGemmaOmnix.createRuntime() composition and a runnable model-installation and streaming-conversation example.
• Add provider-neutral skill and tool contracts, a runtime-owned shared capability registry, observable enablement state, host permission policy, and structured tool execution outcomes.
• Validate tool arguments against a documented JSON Schema subset before permission evaluation and execution.
• Add a Flutter Gemma Agent skill-catalog adapter that makes Omnix enablement authoritative while preserving provider execution metadata.
• Add provider-neutral agent sessions, orchestration events, structured tool outputs, runtime lifecycle ownership, and a Flutter Gemma Agent backend.
• Add a typed, temporary native-chat bridge to support incremental migration of existing history and voice integrations.
• Document the provider-neutral audio and speech boundary without making a dedicated speech-model package mandatory.
• Define capability-oriented module boundaries for Core, Workflow, Knowledge, Nexus, skills, and tools without prematurely splitting packages.
• Add a durable Workflow runtime with atomic task/event transitions, executor ports, retries, cooperative cancellation, restart recovery, and shared inference-scheduler arbitration.
• Make OmnixRuntime own one scheduler for conversation and agent turns, expose it to workflows, and drain active inference safely during shutdown.
• Add a strict, versioned JSON-compatible Workflow codec for portable durable storage adapters.
• Add architecture, contribution, security, and CI foundations.