Upgrade to v8

Review the sections that apply to your library code, generated applications, or hosted editor. Export existing models and keep a copy of application files before regenerating them.

Python and dependencies

BESSER requires Python 3.11 or newer; CI tests 3.11 and 3.12. tiktoken is a new core dependency, and httpx is bounded to <0.28. The backend pins FastAPI to 0.110.x and Starlette to 0.36.x and adds Anthropic, ruff, and pyflakes. Install the agents extra for direct OpenAI/Anthropic library use:

python -m pip install "besser[agents]"

Generated backend requests

In backend mode, <Class>Create schemas used by POST and PUT no longer accept these server-owned attributes: an id that is not a declared primary key, createdAt / updatedAt, and derived attributes. A declared primary key (is_id=True) stays client-supplied regardless of its name.

Update clients to match the generated Create schema. If an id must come from the client, mark it is_id in the model. Relationship fields now use the referenced primary-key type, which may be a string rather than an integer.

On the non-owning side of a one-to-one association the Create schema no longer contains a link field. Set that link from the side holding the foreign key.

Default values and method bodies

The SQLAlchemy, Pydantic, Python, Django, and backend generators coerce defaults to the attribute type and emit them as literals. Defaults such as "abc" for an integer, executable expressions, empty non-string defaults, or unknown enumeration literals raise InvalidDefaultValueError during generation. Use a literal value of the declared type.

Python and Django generators infer simple method bodies such as getters, setters, boolean predicates, and __str__. Other methods without a modeled implementation now raise NotImplementedError when called. Provide an implementation before depending on those methods at runtime.

Django generation

DjangoGenerator.generate() raises generation errors instead of printing them and returning. Handle the exception in callers that need recovery. Generation uses output_dir without changing the process working directory. It refuses to replace a non-empty directory without manage.py; choose a fresh output directory for a new project.

Model behavior

  • Renaming a class updates association-end names derived from it. Custom role names are preserved; a rename that introduces a collision is skipped.

  • DomainModel.validate() warns about mandatory association cycles and multiple associations between the same classes, and reports duplicate association-end names as errors. Review these findings before generation.

  • AssociationClass now forwards its timestamp and metadata correctly.

  • GUIModel accepts stylesheet; DataBinding accepts aggregation. Unknown aggregation names raise ValueError. Existing constructor calls can omit both options.

Imports and backend defaults

Uploaded B-UML files may contain only statements supported by the code builders. Arbitrary Python in an import is rejected; an if __name__ == "__main__": block is ignored. Export your model with the builders rather than adding executable setup logic to an import file.

DEFAULT_SQL_DIALECT is now sqlite. Requests without a dialect no longer use the invalid standard SQLAlchemy dialect.

Spec-Driven Agent callers

Import LLMGenerator from besser.spec_driven_agent. Code written against the unpublished besser.generators.llm prototype must use the new path.

Shell tools and toolchain validation are off by default in the library and backend. Python callers must explicitly pass allow_shell_tools=True and/or enable_toolchain_validation=True when needed. The web backend uses deployment environment variables; a request cannot enable these permissions.

Runtime execution requires a working Linux bubblewrap sandbox unless a development deployment has explicitly opted out. Unavailable checks are reported as unverified. See Validation and auto-fix.

Hosted editor deployment

Before deploying v8 images:

  1. Add besser-wme-smartgen using the smartgen_worker image.

  2. Configure its four security_opt entries, LLM environment variables, shared workspace, and network as in the production runbook.

  3. Route /besser_api/spec-driven/ to the worker, keeping GitHub push and import endpoints on the backend.

  4. Deploy compatible BESSER, WME, and Modeling Agent versions.

  5. Verify build stamps, sandbox startup, public API, and a small generation run.

Use Production deployment for exact configuration. The deploy workflow checks for the service and image; it does not establish that every sandbox setting or proxy route is correct.