Elyra Framework 0.3.1

A feature-rich point release: Elyra gains a built-in AI SDK plus the last of the Electron-parity desktop features.

Added

AI SDK (ai feature)

A new Laravel-inspired elyra-ai crate, re-exported as elyra::ai. Keys stay in the Rust backend.

• Agents — anonymous (ai.chat().instructions(…).prompt(…)) and named (impl Agent + ai.prompt(&agent, …)).
• Tools — Tool trait with an automatic tool-use loop across both providers.
• Sub-agents — an Agent used as a tool (sub_agent / AgentTool); delegates run in isolation.
• Structured output — prompt_as::<T>() via serde + schemars.
• Streaming — stream() yields StreamChunks (SSE), ideal for piping to the event bus.
• Images — ai.image(prompt).landscape().generate() (OpenAI gpt-image-1).
• Embeddings — ai.embeddings([...]).generate().
• Vector store (RAG) — in-memory VectorStore + cosine_similarity, portable across the sqlx Any driver.
• AiProvider binds an env-configured Ai client into the container.

Default provider Anthropic (claude-sonnet-5); images gpt-image-1; embeddings text-embedding-3-small. See docs/ai.md.

Single-instance & deep-linking

• App::single_instance() — later launches focus the running window and forward their command line on elyra:second-instance, then exit (portable loopback rendezvous + handshake).
• App::deep_link("myapp") — launch URL via deepLink.initial(), later URLs on elyra:deep-link (macOS open-URL event; Windows/Linux scheme registration). See docs/single-instance.md.

Notes

Everything is compiled, clippy-clean, and tested (21 suites; SQLite locally, MySQL/Postgres in CI, runtime typechecked with TypeScript 7). AI live calls are exercised only with ELYRA_AI_LIVE=1 and keys — they were not run in CI or during development (no keys present), so the live provider path is documented as unverified. RAG ranks in-process by design (the Any driver has no pgvector).

Full changelog: https://github.com/kwhorne/elyra-framework/blob/main/CHANGELOG.md · MIT licensed.