v0.14.0

Highlights

A small minor release on top of v0.13.1. The headline is a clearer, more discoverable way to control where the PydanticAI adapter places its durable checkpoints — the new checkpoint_strategy="calls" | "turn" option — alongside a new durable chatbot example that shows off long-horizon kitaru.wait(), and a bump to the latest upstream ZenML.

Added

• PydanticAI checkpoint_strategy="calls" | "turn" — the preferred public spelling for where the adapter draws its checkpoint boundaries. "calls" stays the default and maps to the existing per-model/tool/MCP checkpoint behavior; "turn" maps to one checkpoint per agent run. Existing granular_checkpoints=True | False code keeps working unchanged.
• Durable chatbot example (examples/chatbot/) — models an entire conversation as a single PydanticAI agent with one human-in-the-loop tool. It uses kitaru.wait() to release compute between turns, so a session can sleep for minutes or days and resume exactly where it left off when the user replies.

Changed

• Standardized adapter docs and examples around the shared checkpoint_strategy concept, while keeping framework-specific boundary names such as PydanticAI "turn", OpenAI Agents "runner_call", LangGraph "graph_call", and Claude Agent SDK "invocation".
• Bumped the minimum ZenML dependency, server image, and Helm subchart versions to 0.94.6 so Kitaru tracks the latest upstream ZenML release.

Full Changelog: v0.13.1...v0.14.0