v0.3.0

Full coverage of every B-UML DSL and generator, deeper class-diagram /
state-machine / agent references, a complete editorial-review pass, a
marketing/positioning rework of the README, and field fixes surfaced by real
agent runs.

Added

• Coverage of all remaining B-UML DSLs in besser-user — seven new
  source-verified references:
  • references/object-models.md — object/instance models (objects,
    attribute values, links), incl. the fluent builder and OCL test data.
  • references/feature-models.md — software product lines (features,
    groups, configurations).
  • references/ocl.md — writing and attaching OCL Constraints; bocl
    evaluation; which generators consume OCL.
  • references/deployment.md — deployment models for the
    TerraformGenerator.
  • references/neural-networks.md — NN models for the PyTorch/TF
    generators.
  • references/quantum.md — quantum circuits for the QiskitGenerator.
  • references/project.md — bundling models + metadata.
• Two previously undocumented generators in besser-generators:
  • SupabaseGenerator (Postgres DDL + RLS) in references/persistence.md.
  • JSONObjectGenerator (JSON data from an ObjectModel) in
    references/python-and-data.md.
• Much deeper class-diagram, state-machine, and agent references
  (community contribution): metamodel.md renamed to class-diagram.md and
  expanded (AssociationClass, GeneralizationSet,
  is_abstract/is_optional/is_read_only, navigability, composition
  direction, void methods, end-name uniqueness); state machines gain final
  states, conditions, fallback bodies, config properties, and the
  STATE_MACHINE method-implementation pattern; agents gain LLM integration,
  entities/slot-filling, intent-classifier config, RAG, DBReply,
  ReasoningState, tools/skills/workspaces, and rich platform replies.
• The skills now cover all 19 BESSER generators and every B-UML
  metamodel/DSL; reference-layout tables, trigger descriptions, and the
  README were updated to surface them.
• README "What your agent can build" capability block — surfaces the full
  BESSER surface an agent can drive: every B-UML model type, all 19 generators,
  the deterministic-codegen-then-build-on-top workflow, troubleshooting, and
  extending BESSER itself. Cross-links the standalone
  uml-drawing skill and adds an
  Agent Skills directory link.
• One-call B-UML → SVG rendering in besser-user and the README — POST a
  model to BESSER v7.9.0's headless /besser_api/get-svg endpoint and embed the
  result, no browser required.
• "Generate the baseline, then build on top" guidance in besser-generators:
  regenerate the deterministic baseline freely (cheap and identical each run) and
  put customizations in separate files so a re-run never clobbers them.

Changed

• besser-dev restructured from a 487-line monolith into a lean overview
  • routing table plus five references/ (adding-a-generator,
    adding-a-metamodel, testing, docs-and-build, contributing-workflow),
    matching the progressive-disclosure pattern of the other skills.
• De-duplicated drift-prone content across skills: the web-editor
  registry snippet is now single-sourced in besser-dev; the
  composite-generator tree lives only in debugging.md; the
  besser-troubleshooting generator-failure table is scoped to error-message
  lookups with a clear hand-off to besser-generators.
• besser-generators editorial fixes: merged the redundant
  reference-layout/generator-picker tables; replaced the partial
  output-directory table with an exhaustive-by-rule "Output locations"
  call-out; normalized the Supabase/JSON-object entries to the house table
  style; clarified that BAL/CODE method bodies are inserted verbatim.
• Consistency pass: standardized reference "Gotchas" headings, added
  tables of contents to the longer references, and moved per-file version
  stamps to a single note; relaxed the attribute-naming guidance to match
  what BESSER actually enforces (no spaces/hyphens; snake_case or camelCase).
• README refreshed for v7.9.0: BESSER badge bumped to v7.9.0, the drawing
  section now leads with the one-call headless SVG endpoint, and the codegen
  bullet is reframed around deterministic baselines you extend on top — spend
  tokens only on what is genuinely custom. The intro now links directly to the
  uml-drawing repo instead of an in-page anchor.

Fixed

• Corrected a broken copy-paste example (an always-false chained-comparison
  assertion) in object-models.md, and the chat_history type in
  agents.md. Reconciled CONTRIBUTING.md's description-voice rule with the
  trigger style all four skills actually use.
• Stopped recommending default_value on enum-typed attributes across
  besser-user, besser-troubleshooting, and besser-generators'
  debugging.md. It passes validate(), but the Pydantic/SQLAlchemy generators
  serialize the literal's raw repr() → invalid Python (e.g. SyntaxError: leading zeros…), and the headless SVG endpoint rejects it. Set the initial
  value in the app/auth layer or as a DB column default instead. (Primitive
  defaults like True or 0 round-trip fine.)
• Class diagrams are delivered as B-UML, never Mermaid (besser-user). The
  authoritative artifact is the validate()-checked B-UML model; a rendered
  SVG/PNG is the picture and a quick ASCII sketch is fine as a preview, but
  Mermaid (un-validated, can't be generated from) is no longer produced.

Notes

• All new content was written against BESSER v7.8.3 source with verified API
  citations; several upstream documentation bugs were deliberately not
  reproduced (e.g. the wrong besser.BUML.notations.od.* import path and the
  ObjectModel(instances=…) kwarg). Known-broken paths are flagged in-place
  (e.g. the objectPlantUML parser is non-functional in v7.8.3).
• The headless B-UML → SVG flow and the v7.9.0 badge track BESSER v7.9.0
  (PR #554, ELK auto-layout); all other API citations remain verified against
  v7.8.3 source.

v0.2.1

Fixed

• Corrected 24 API/accuracy errors across the skills, found by a
  systematic audit against BESSER v7.8.3 source:
  • besser-generators: the custom-endpoint example now uses
    Method(code=..., implementation_type=MethodImplementationType.BAL)
    (the previous MethodImplementation class/import did not exist); the
    SQLAlchemy customization example uses sqlalchemy.orm.Session + the
    generated engine (no Session is generated); BAFGenerator output is
    config.yaml, not config.ini.
  • TerraformGenerator(deployment_model=...); the PyTorch/TF generators
    import from besser.generators.nn.{pytorch,tf}.* with generation_type
    as a constructor argument; FlutterGenerator requires main_page and
    is registered in the web editor; many-to-many join-table name is the
    association name only; jinja2.TemplateNotFound; BackendGenerator
    filters invalid HTTP methods silently.
  • Python requirement corrected to 3.11+ (setup.cfg python_requires);
    bocl==1.0.1; Docker base python:3.11-slim; state-machine transitions
    live on State, not StateMachine.
  • besser-dev: SUPPORTED_GENERATORS (not GENERATORS); documented that
    BESSER now ships CI (pytest + ruff).

Added

• Two evals (documentation delivery, SQLGenerator dialect) — the suite is
  now 10 (evals/evals.json), making those behaviors permanent regression
  checks.
• README skills table and trigger table now surface "drawing UML for
  documentation".
• CI: .github/workflows/release.yml automatically creates a GitHub Release
  from the matching CHANGELOG section whenever a v* tag is pushed.

v0.2.0

BESSER Skills v0.2.0.

Highlights

• Modeling for documentation, not only code — besser-user now treats a B-UML model as having two first-class outcomes: generating code and documenting a system as a correct, embeddable UML class diagram. Triggers broadened so the skill activates on "draw/document a correct class diagram" requests even when BESSER isn't named.
• Model delivery — deliver a model as a runnable .py file by default, or embed the same B-UML in Markdown for docs; plus how to import a model into the web editor (Import → B-UML). New references/delivering-models.md.
• README "Drawing correct UML diagrams" section.

Fixed

• SQLGenerator API in references/persistence.md — sql_dialect is a constructor argument and generate() takes none (only SQLAlchemyGenerator takes dbms in generate()). The reference previously showed a call that raises TypeError. This shipped in v0.1.0 and was caught by the v0.2.0 benchmark re-run.

Benchmarks

• benchmarks/iteration-2/: v0.2.0 correctness re-run, graded with-skill vs no-skill baseline against BESSER v7.8.3 — 38/38 (100%) for both with the fixed skill. Performance figures carried over from v0.1.0 (skill-text-only changes).

Full details in CHANGELOG.md.

v0.1.0 — Initial release

Initial public release of BESSER Skills — Agent Skills for BESSER, the low-code model-driven engineering platform.

Included skills

• besser-user — end-user guide: domain modeling, code generation, web editor
• besser-generators — per-generator operations, safe customization, template overrides
• besser-troubleshooting — diagnosis guide for install, import, runtime, and deployment issues
• besser-dev — contributor guide: adding generators, tests, docs, PR workflow

besser-user and besser-generators use the progressive-disclosure pattern: a slim SKILL.md plus references/*.md loaded on demand. Scaffolding scripts are bundled for both besser-user (scaffold_model.py) and besser-dev (scaffold_generator.py).

Benchmark — with skills vs no-skill baseline (8 evals)

Metric	With skills	Without skills (baseline)	Skill advantage
Pass rate	100%	100%	tie
Mean response time	36.5s	138.4s	74% faster
Mean token usage	25,181	48,007	48% fewer

Token usage is also dramatically more consistent with the skill (stddev ~1.4k vs ~19.8k baseline). Full data and per-eval breakdown in benchmarks/iteration-1/benchmark.md.

Install

Works with Claude Code, Cursor, Cline, Windsurf, GitHub Copilot, and any agent that follows the Agent Skills specification.

```bash
npx skills add BESSER-PEARL/BESSER-Skills --all
```

See README.md for full installation and usage instructions.