Changelog

Typescript Intellisense was becoming quite slow when defining agents, plugins, etc. p90 latency is back to <150 ms.

The plugin.context() API has been improved with initial values

The plugin.methods() API has been replaced by plugin.api() making it more flexible and runtime agnostic

New defineStore() definition builder

Exclude memories builder methods as they are called

Add an higher-level and typesafe RPC API to Transport classes

New client-side config schema and types

First functional version of client

Add support for ZodError serialization

Simplify the plugin context API to enforce immutability without neither relying on complex JS Proxy patterns, nor on the consumers to properly clone values

Plugin.methods() was still be typed with the raw context

Make non-blocking memories running on history change, instead of resources requests

Fix plugin.pick() broken typesafety

Cleanup bundle and package.json metadata

Make peer dependencies optional

Wire plugin lifecycle hooks in the plugin runner

Add set() method to plugin context and onChange() listeners

Solve context.onChange() not properly comparing object values

Refactor and simplify: equal, serialize, and sha256 libraries for consistency

Rename memory.getOutput() to memory.output()

TTS estimated transcript's tokenizer was broken because of regex hoisting

Add a standardized serialization/deserialization library

Refactor plugin dependencies to unify them with the future items dependency and simplify the DX

Make plugins' context definition bound to a Zod schema

Bump dependencies + migrate to Biome v2

Fully functional memories plugin

Wire plugins interceptors to the event loop of the external dependency

New life/define export

Huge refactoring/refinement of plugins' dependencies and methods + draft memories plugin

Fix plugin methods in agent definition not assigning plugin methods after called

Fix tools calls formatting in Mistral LLM provider class

A lot of simplification on the orchestration/generation classes of the core plugin

Fix new Int16Array type error

Mistral LLM provider class wasn't emitting end token

Support Mistral.ai LLM provider

Add all models and available languages to Cartesia TTS config schema

Generation interruptions weren't working in the case the generation had ended, but it's produced content was still being played. This commit fixes that.

Disable output stream throttling in text-only mode