chore!: Automate data model generation from upcoming CycloneDX 2.0 modularized specification

[jkowalleck]

Description

Currently, the data models in this library are largely written and maintained manually. While this approach has worked so far, it is time-consuming and requires significant effort for both implementation and review. This effort could be better invested in feature development and bug fixing. (Also the issue is that the models code also implement the business logic, like property constraints, etc.)

With the upcoming CycloneDX 2.0 specification, a modularized and machine-readable format will be introduced. This presents an opportunity to rethink how data models are created and maintained in this project.

Reference (work in progress):

• PR https://github.com/CycloneDX/cyclonedx-python-lib/issues
• modularized schema https://github.com/CycloneDX/specification/tree/2.0-dev/schema/2.0/model

---

Problem

• Data models are mostly handwritten and mix in the business logic
• High maintenance overhead
• Repetitive work for contributors
• Slows down development velocity due to review effort

---

Proposal

Leverage the machine-readable specification planned for CycloneDX 2.0 to introduce static code generation for data models (not business logic).

This would involve:

• Parsing the official CycloneDX specification (once available in machine-readable form)
• Generating Python data models automatically
• Integrating generation into the build or release process
• Minimizing manual intervention for future spec updates

There have already been proof-of-concept implementations demonstrating that automated generation of data models from the specification is feasible. These approaches should be revisited, consolidated, and applied as part of this effort.

Pipeline:

flowchart TB
    A["Input: CycloneDX JSON Schema files"]
    B["Preprocessing (if needed)"]
    C["Code Generation"]
    D["Post-processing: formatting, code style, adjustments"]

    subgraph E["Result: Generated Python code"]
        direction LR
        E1["Data Models"]
        E2["Factories"]
        E3["Serialization-specific Normalizers"]
    end

    A --> B --> C --> D --> E

Loading

Philosophy: decouple logic fro mdata

• Models are pure dumb typed data classes.
• Factories implement business logic for instancing data classes with the propeties' constrinats defined in the schema
• normalizers turn data classes into objects that can be handed over to serializers
• serializers take prepared objects and render XML/JSON/PB

---

Possible Tools / Libraries

The following tools could be evaluated as part of this effort:

• datamodel-code-generator — MIT
  https://pypi.org/project/datamodel-code-generator/

• pydantic — MIT
  https://pypi.org/project/pydantic/

• dataclasses-json — MIT
  https://pypi.org/project/dataclasses-json/

• dacite — MIT
  https://pypi.org/project/dacite/

• marshmallow — MIT
  https://pypi.org/project/marshmallow/

• marshmallow-jsonschema — MIT
  https://pypi.org/project/marshmallow-jsonschema/

• jsonschema (validation, not models) — MIT
  https://pypi.org/project/jsonschema/

• quicktype — Apache 2.0
  https://pypi.org/project/quicktype/

• genson (schema generator, reverse direction) — MIT
  https://pypi.org/project/genson/

---

Community Input

Community discussions have already suggested evaluating tools such as:

• datamodel-code-generator
• json-schema-to-pydantic (https://pypi.org/project/json-schema-to-pydantic/)
• jambo (https://pypi.org/project/jambo/)

and

• de/serialization with cattrs (https://pypi.org/project/cattrs/)
  • example: feat!: de/serialize with cattrs #934

These should be considered as primary candidates during evaluation.

see discussions:

• https://cyclonedx.slack.com/archives/CVA0QJEVA/p1769793218405249

---

Expected Benefits

• Significant reduction in maintenance effort
• Improved consistency across models
• Faster adoption of new specification versions
• More time available for feature development and bug fixing

---

Considerations / Open Questions

• What format will the machine-readable spec be published in (e.g., JSON Schema, OpenAPI, etc.)?
  • JSON Schema it is
• Should generated code be committed or generated at build time?
  • decision: generated before build time, and commited to the repo
• How to handle custom logic or extensions on top of generated models?
• Backward compatibility with CycloneDX 1.x
  • easy path: breaking change in the library, and only support 2.0 from then on

Additional Context

This proposal aligns with the direction of CycloneDX 2.0, which aims to make the specification more modular and tooling-friendly. Taking advantage of this early could significantly improve long-term maintainability of this library.

---

Note: This issue is intended as a meta-ticket to collect related subtasks and track overall implementation efforts.

---

Checklist

• have the models created in a deterministic way from schema
• generate the docs
• have examples updated
• have tests
  • test with all schema test cases
  • tests all enum completeness
• write a transition guide in the docs
  • section "Upgrading to vXXX"
  • we might not be able to provide a proper guide at all, since basically everything is rewrite.
  • we might point to the github discussions - section FAQ - to pride at least some help

[jkugler]

Thank you for starting this. I would lean heavily toward Pydantic (which datamodel-code-generator leverages). I have used it heavily, and done code generation with datamodel-code-generator with other standards (namely SARIF).

datamodel-code-generator has a LOT of flexibility (including custom templates) and Pydantic allows us to add in custom validators and other data checking.

So, my vote is for Pydantic and datamodel-code-generator. :)

[jkugler]

The complication we will have with something like CycloneDX is versioning the spec, something I did not have to solve in my SARIF work. I don't know if JSON schema has "variants" and deprecations built in to the spec (if so, then that problem might be solved), but we may need to come up with a way of handling different versions and variants of certain objects, especially when it comes to rendering.

[jkowalleck]

need to look into the details of the respective implementation options.

if nothing fits our needs (probbaly not the case), we might fall back to simple data classes with proper utilization of the "dunder" method generation flags and fields() for default values...