RefDiff-n

RefDiff-n is a refactoring detection tool applicable to diverse programming languages, built on top of RefDiff. It mitigates RefDiff's language dependence so that refactorings can be detected uniformly across languages. Plugins are currently provided for seven languages: Java, C, JavaScript, PHP, Python, Go, and Ruby.

Key design points:

• Parsing: a single parsing infrastructure (Tree-sitter) replaces per-language parsers.
• Definition extraction: the only language-specific part is a set of declarative Tree-sitter queries that convert syntax trees into RefDiff's language-independent Code Structure Tree (CST).
• Call graph construction: lightweight string matching between method/function names and tokens, avoiding per-language semantic analysis.
• LLM-assisted plugin implementation: tests and queries for a new language are generated with an LLM (Gemini) against a Universal Ctags pseudo-oracle; see GEMINI.md and plugin-generator/.

Repository layout

Directory	Contents
refdiff-core	Detection core (from RefDiff)
refdiff-universal	Language-common plugin (Tree-sitter based) with per-language queries and tests
plugin-generator	Assets for LLM-assisted plugin generation (procedure documented in GEMINI.md)
evaluation	Accuracy evaluation harness (precision/recall on Java; evaluation/oracle contains the RefDiff evaluation datasets)
investigation	Empirical study harness (refactoring-type distribution and Extract LOC across seven languages)
refdiff-java, refdiff-c, refdiff-go	Original RefDiff plugins (kept for comparison)
doc	Additional documentation

Requirements

• git
• JDK 8 or later
• python3 (with venv)

Reproducing the accuracy evaluation (RQ1)

Precision:

./gradlew :evaluation:installDist
cd evaluation
bash ./script/detect.sh                      # detection over the sampled Java projects
cd calculator
python3 -m venv venv && venv/bin/pip install -r requirements.txt
venv/bin/python -m src.manual.random         # sample instances per type (seed=42)
# annotate the sampled CSVs with TP/FP in the "manual check result" column, then:
venv/bin/python -m src.manual.count

Recall:

cd evaluation
./build/install/evaluation/bin/evaluation java
cd calculator
venv/bin/python -m src.main -d ../detection-result/java/java-<timestamp>.csv

Reproducing the empirical study (RQ2/RQ3)

./gradlew :investigation:installDist
cd investigation
bash ./script/investigate.sh                 # detection over the 69 subject projects
# The number of commits analyzed per project is configured by COMMIT_DEPTH in
# investigation/src/main/java/investigation/Executor.java.
cd calculator
python3 -m venv venv && venv/bin/pip install -r requirements.txt
venv/bin/python -m src.rq2.dist              # RQ2: distribution per language
venv/bin/python -m src.rq2.dist_repo         # RQ2: distribution per repository
# RQ3 requires no additional detection run: the RQ2 results already contain LOC.
venv/bin/python -m src.rq3.loc               # RQ3: LOC of extraction sources/targets
venv/bin/python -m src.rq3.loc_repo          # RQ3: per repository

License

MIT, inherited from RefDiff. The original RefDiff README is preserved in README-refdiff.md.