Pattern-Based Quick Assists from Mined Refactorings — Looking for Feedback

[carstenartur]

Hi everyone,

JDT UI already has excellent Quick Fixes / Quick Assists. I'd like to share an experimental, complementary prototype and gauge interest.

The idea: Derive Quick Assist suggestions from patterns mined from real code changes, and represent those patterns in a declarative DSL (.sandbox-hint files).. Example:

java.util.Collections.unmodifiableList(java.util.Arrays.asList($args$)) :: sourceVersionGE(9)
=> java.util.List.of($args$)
;;

Placeholders ($x, $args$) match AST subtrees. Guards (:: sourceVersionGE(9), instanceof(...), isNullable(...)) provide conditional matching — 30+ built-in guards are implemented.

What exists: A working prototype (carstenartur/sandbox) with:

• DSL parser, AST matcher extending JDT's ASTMatcher, guard system
• Quick Assist (Ctrl+1) and Cleanup integration in Eclipse
• 10+ bundled rule sets (performance, collections, Java 11+ migration, JUnit 4→5, probable bugs)
• Git history walker + LLM-based rule inference from diffs
• Mining results from eclipse.jdt.core (85 matches), jdt.ui (31 matches), platform.ui (11 matches)
• Report: https://carstenartur.github.io/sandbox/mining-report/

Why it might matter: Traditional cleanups require 500-1000 lines per pattern. A .sandbox-hint rule needs 5-50 lines. Rules are shareable, composable, and can be project-specific.

Open questions: precision/false positives, developer trust in mined rules, UX differentiation from hand-crafted fixes, quality assurance, maintenance, integration points in jdt.ui.

I'm not promising large-scale follow-up — I'm looking for feedback on whether this direction is worth exploring further.

Thanks!