GEP-30

The complete java.beans/java.desktop surface of the core jar (Groovy 6 master, 2026-07-29), grouped by how it is reached:

Everything else that touches java.beans (groovy-swing bindings, groovy-console, groovydoc tooling) lives in modules whose whole purpose is desktop UI; they are out of scope and correct as-is.

Two facts make this inventory actionable. First, it is complete and member-verified: during the GROOVY-12199 shadow-migration review, every class and every method/field reference from Groovy code into the beans namespace was resolved against openbeans 1.0.2 — fourteen of the fifteen types resolve fully; only java.beans.Transient is missing. Second, it is small: the spine portion is five files, and the beans API they actually consume (property/event descriptors, decapitalize, cache flush, a marker annotation) is a fraction of java.beans.

Android. Android has never shipped java.beans (beyond a vestigial PropertyChange* subset). The grooid jar therefore bundles openbeans relocated to groovyjarjaropenbeans. and rewrites every java.beans reference — machinery that GROOVY-12199 just migrated from ant-jarjar to shadow, where a string-relocation subtlety (Java9.JAVA8_PACKAGES()) already required a targeted skipStringConstants carve-out. Shading *works, but its failure mode is silent: the Transient gap sat unnoticed for six years because nothing verifies the relocated world is closed. Retiring the spine dependency removes the need for the relocation (and openbeans) for every app that does not itself use the beans-shaped leaves.

Minimized server runtimes. jlink-built images routinely omit java.desktop — it is one of the heaviest platform modules, and headless services have no use for it. Groovy on such an image fails at metaclass initialization. There is no workaround short of adding the module back; Groovy is the reason a headless service ships AWT.

GraalVM native-image. Introspector-based introspection needs reflection metadata for every introspected type, and reachability of java.beans pulls desktop classes into the closed-world analysis. A reflection-only introspector on a documented internal seam is easier to supply metadata for, easier to initialize at build time, and shrinks images. (This GEP does not promise full native-image support — dynamic Groovy has larger obstacles there — but it removes one structural one, and @CompileStatic-heavy applications benefit immediately.)

The java.beans usage predates the module system; when it was written, "the JDK" was one indivisible thing. Post-JDK-9, java.desktop is severable and increasingly severed. Meanwhile the historical reasons Groovy leaned on Introspector have weakened: the MOP’s own property/method model (MetaClass, MetaProperty, CachedClass) has long been the primary mechanism, with java.beans supplying supplementary event-set detection and descriptor compatibility. The remaining coupling is habit plus a thin slice of genuine function — exactly the shape an internal abstraction handles well.

Every current usage falls into one of three tiers, with different obligations:

Introduce a minimal internal abstraction — working name org.apache.groovy.beans.IntrospectionService (final naming during implementation; hanging it off VMPlugin, which already abstracts platform variance, is the natural alternative) — with exactly the operations the spine consumes:

Two implementations:

Selection is a one-time reflective probe at startup (Class.forName("java.beans.Introspector") guarded, never a static link), cached, with the chosen implementation exposed for diagnostics.

The GROOVY-12199 lesson is that this class of dependency regresses silently. Two standing guards:

No public API signatures change. On a full JDK, the default implementation preserves current behavior including custom BeanInfo interop; the golden-equivalence corpus is the gate. The only intentionally observable changes are on platforms where the current behavior is failure: clear diagnostics instead of NoClassDefFoundError, and working compilation where it was impossible.

Spine classes keep their public members (per the project’s compatibility policy for classgen/ast packages). The BeanInfo companion class for records remains generated by default on desktop JDKs; a record compiled on a beans-less platform simply lacks the companion — code reflectively looking it up must already handle its absence (pre-Groovy-4-record classes never had one).

joint-compilation stub generation, Groovydoc, and IDEs are unaffected (no syntax or AST-shape changes). Documentation gains one page: which Groovy features require java.desktop, what the failure looks like without it, and how the introspection fallback differs (custom BeanInfo only). The Android/grooid README notes the phase-4 simplification when it lands.