Version 8.0.0

This release adds the experimental Spec-Driven Agent: generate an application from a reviewed model and written requirements, then validate and repair the generated files within a budget. It also improves generated GUIs, backend data contracts, model imports, and assistant reconnects.

Important

Existing users should read Upgrade to v8 before upgrading. Hosted deployments need a separate generation worker and reverse-proxy routes before rolling out the v8 backend.

What’s new

  • Application generation with AI. The Spec-Driven Agent customises a deterministic BESSER scaffold when one fits, or authors a project for another stack. The editor shows progress, cost, findings, and downloadable output. Runs support cancellation, recovery, and modification of an earlier result. Start with Build an application with AI.

  • Generated screens preserve more of the design. Shared GUI stylesheets, class-bound forms, create/update/delete buttons, and chart aggregations reach the React and full web app output.

  • More consistent generated backends. Create schemas omit server-owned fields, links use the declared primary-key types, and entity creation is atomic. Generated code handles mandatory foreign-key cycles and associations targeting association classes.

  • Provider options. Configured free and sponsored tiers, user-supplied keys, GPT-6 and Claude 5.x models, and authorised local OpenAI-compatible endpoints are supported. See Providers and models.

  • Version provenance. Generated archives include BESSER_GENERATION.md. Code and model exports record the BESSER version where supported, and editor JSON exports include version metadata.

Reliability and security

  • Uploaded B-UML is loaded through an AST allowlist. Default values and user-provided strings are escaped when emitted into generated source.

  • Authorised shell tools and checks that execute generated code run in a bubblewrap sandbox. Unavailable checks are reported as unverified. Completion is not a guarantee that every check ran; review the findings.

  • Downloads, workspace copies, and GitHub pushes skip links and exclude runtime artefacts. Run credentials are redacted from delivered files and streamed events.

  • Assistant replies carry turn identifiers so reconnect replay can be deduplicated. Partially applied model edits are surfaced to the user.

  • Generated declaration order is stable. Fixes also cover object-diagram parsing, GUI re-import, quantum exports, and Django self-associations.

Upgrade and deployment

The coordinated release comprises BESSER, the Web Modeling Editor, and the Modeling Agent. Deploy compatible versions together. Production uses separate backend and smartgen-worker Docker targets; the worker needs its documented sandbox settings, shared workspace, and proxy routes.

Follow Production deployment for the service definitions and checks. After deployment, verify that the worker sandbox starts, the public API answers, and a small editor generation run completes.