Version 7.10.0
===============

**BPMN 2.0 support**. Adds a full BPMN metamodel, a
vendor-neutral BPMN 2.0 XML generator, and a matching BPMN diagram type in the
web modeling editor — process modeling, alongside class/object/state-machine/
agent modeling, is now a first-class BESSER diagram type. Also adds OCL
constraint validation to the Django generator and fixes a Web App generator
bug affecting locally-built agent containers.

BPMN metamodel
---------------

* New :doc:`../../buml_language/model_types/bpmn` (``besser.BUML.metamodel.bpmn``):
  ``BPMNModel``, ``Process``, flow nodes (``Task`` with ``TaskType`` — user,
  service, script, manual, business rule, send, receive —, ``StartEvent``,
  ``EndEvent``, intermediate catch/throw events with typed event definitions,
  gateways — exclusive, parallel, inclusive, event-based —, ``SubProcess``,
  ``CallActivity``), ``SequenceFlow``, pools/lanes via ``LaneSet``/``Lane``,
  ``DataObject``/``DataStore`` references, and text ``Annotation``.

BPMN generator
--------------

* New :doc:`../../generators/bpmn` (``BPMNGenerator``) emits plain BPMN 2.0
  XML from a ``BPMNModel`` — openable in any BPMN-aware tool (Camunda 7/8,
  Flowable, bpmn-js, …). No engine-specific execution semantics are emitted,
  only the process structure (flow nodes, sequence flows, gateways, events,
  pools/lanes, data objects/stores, sub-processes), and the output round-trips
  through any conformant modeller.
* New ``besser.utilities.buml_code_builder.bpmn_model_builder.bpmn_model_to_code``
  emits a ``BPMNModel`` back to executable B-UML Python code.

Web Modeling Editor
--------------------

* New **BPMN diagram type** in the editor: the full BPMN element palette
  (events, tasks, gateways, sequence flows, pools/lanes, data objects/stores,
  annotations, call activities), property popups, and a default-layout saga
  that auto-arranges newly dropped BPMN flow nodes.
* Backend wiring end-to-end: ``process_bpmn_diagram`` / ``bpmn_buml_to_json``
  converters, a ``BPMNGenerator`` registration (category ``business_process``)
  in the generator registry so ``/generate-output`` produces ``.bpmn`` files,
  BPMN support in the ``/export-buml`` and B-UML-file-import (``/get-json-model``)
  auto-detection paths, and BPMN participation in project-level cross-diagram
  validation.

Django Generator
-----------------

* **OCL constraint validation**: the :doc:`Django generator <../../generators/django/django_admin_panel>`
  now reuses the same OCL parser as the :doc:`Pydantic generator <../../generators/pydantic>`
  and emits a ``clean()`` method on each constrained class, so OCL invariants
  defined on the domain model are enforced when creating or editing entities
  through Django's admin panel or a ``ModelForm``. Unlike Pydantic's
  per-field validators (which run at construction time), Django's ``clean()``
  runs via ``full_clean()`` — automatic through the admin panel/``ModelForm``,
  but not on a bare ``.save()`` unless the caller calls ``full_clean()``
  first.

Modeling Agent
---------------

* **BPMN support** (`BESSER-PEARL/modeling-agent#7 <https://github.com/BESSER-PEARL/modeling-agent/pull/7>`_):
  the AI modeling assistant gains BPMN diagram support — creating,
  editing, and removing BPMN elements through the same conversational
  actions already available for other diagram types, with guardrails
  against LLM hallucination (e.g. refusing to delete elements that don't
  exist). This lives in the separate ``modeling-agent`` repository, which
  is not part of BESSER's release/versioning; merge and deploy it
  alongside this release (``./scripts/deploy.sh agent``) to get
  AI-assisted BPMN editing in the web editor.

Fixes
-----

* **Agent container crash on classical intent recognition (local Docker
  builds)**: ``WebAppGenerator``'s per-agent ``agents/<slug>/Dockerfile``
  never installed PyTorch, so any agent configured with
  ``intentRecognitionTechnology: classical`` crashed on startup
  (``NameError: name 'nn' is not defined`` from ``besser-agentic-framework``'s
  NLP engine) when run via ``docker compose up``. The Render/GitHub deploy
  pipeline already handled this correctly by installing
  ``torch==2.6.0+cpu``/``scikit-learn==1.6.1`` for classical agents, but that
  logic never reached the local Docker Compose/ZIP output. ``WebAppGenerator``
  now resolves each agent's configured intent-recognition technology and
  conditionally installs the same pinned PyTorch/scikit-learn versions in
  ``agent.Dockerfile.j2`` when needed, matching the Render pipeline.

Internal updates
-----------------

* Sandboxed ``exec()`` hardening, Python-keyword name warnings, and
  ``isinstance``-based (rather than class-name-string) event-definition
  legality checks were applied across the new BPMN converters during review.