v0.15.0

Highlights

Kitaru v0.15.0 brings live execution streaming to every adapter and the SDK, so you can watch a run unfold in real time instead of waiting for it to finish, and adds an experimental Gemini Interactions adapter. Both sit on top of Kitaru's durable-execution core: the streamed events are best-effort and observational, while the checkpoint output stays the durable, replayable source of truth.

Added

Live execution streaming

• KitaruClient().executions.events(...) consumes server-backed live execution events with kind/checkpoint/correlation filters, SSE cursor reconnects, and clear feature-unavailable errors for local database mode or disabled server streaming.
• Checkpoints can emit their own live events from inside a running body: kitaru.progress(...) and kitaru.events.publish(...) send best-effort progress/custom updates, and Kitaru automatically publishes checkpoint started/completed/failed lifecycle events.
• Streaming across all four adapters, each still preserving its durable checkpoint output as the replay anchor:
  • OpenAI Agents — KitaruRunner.run_stream(...) / run_stream_sync(...) (openai_agents.stream.* events)
  • LangGraph — KitaruGraphRunner.stream(...) / astream(...) (langgraph.stream.* events)
  • Claude Agent SDK — KitaruClaudeRunner.run_stream(...) / run_stream_sync(...) (claude_agent_sdk.stream.* events)
  • PydanticAI — KitaruAgent runs emit pydantic_ai.stream.started/.event/.completed/.failed
  • Gemini Interactions — run_stream(...) / run_stream_sync(...) with conservative live-event privacy defaults
• Runnable, provider-key-gated streaming examples under examples/integrations/... for the adapters.

Gemini Interactions adapter (experimental)

• New kitaru.adapters.gemini adapter with an Interactions-first design: one stable Gemini response maps to one Kitaru checkpoint, non-stable background statuses fail instead of being cached as success, raw provider payload capture is opt-in, and the public API includes an Antigravity managed-agent preset plus explicit cache_identity disambiguation while keeping Google-owned sandbox/tool internals outside Kitaru's replay promise.

Changed

• Adapter documentation is now a first-class Adapters section in the docs, with redirects from the old guide URLs to the new canonical pages.

Fixed

• Fixed PydanticAI adapter compatibility with newer Pydantic AI releases that forward retries= from run_sync() into run(), while preserving Kitaru's legacy output_retries= keyword.

Full Changelog: v0.14.0...v0.15.0