Changelog

Every release of Nucleo, newest first. Released artifacts are on Maven Central under the group ai.redouble; a published version is never changed in place.

0.1.1 (2026-10-02)

• Every published artifact carries a CycloneDX software bill of materials, attached as <artifact>-<version>-cyclonedx.json and signed with it.
• NOTICE carries the copyright and licence only. The statement on Redouble AI's patents is in the README's License section, where it can be explained.
• A model can no longer author an artifact: an artifact-typed field takes a reference, and the caller receives the registry's object (GHSA-m8pq-5898-5q8j).
• WebFetchTool declares its address policy itself and applies it to every redirect (GHSA-5h79-cgv5-c74w).
• An answer still missing a required field after its corrections fails the call (#13).
• A cancelled job that then fails settles as cancelled and releases its caller (#14).
• The demo's extract, pricing and decision runs read the corpus the demo ships, located by the engine; no request names a folder to read or a file to write, and the pricing report is the response alone. The page shows the folder in place of an input.

0.1 (2026-10-01)

The first public release.

• The runtime: jobs on virtual threads with capacity admission, typed tools, thinkers and doers, scopes and guardrails, byte-identical artifacts, prompt and skill registries, the model catalog with its discovery, and the observability bus.
• Provider modules: Anthropic, OpenAI and OpenAI-compatible endpoints, Azure AI Foundry, AWS Bedrock through Converse and through the Anthropic SDK, and System One decision models.
• Hosts: nucleo-spring-boot-starter for Spring Boot 4 and nucleo-quarkus for Quarkus, the latter building to a native image.
• nucleo-examples, a demo running the same work on both hosts, and the documentation site at docs.redouble.ai.

Known gaps, each tracked as an issue:

• #1: the MAXIMUM compaction level runs the same collapse as AGGRESSIVE, so a conversation whose non-compactable messages alone exceed the context limit is refused rather than made to fit.
• #2: ThinkerInput.toLLMString renders a thinker's input as ad-hoc text rather than strict JSON.

What is unstable at 0.1: the public API has not been through a second release, so a 0.2 may change signatures where a first consumer shows a better shape. Each such change will be named here.