arcade-agent isn’t a heuristic bolted onto an LLM. It descends fromARCADE — the Architecture Recovery, Change, and Decay Evaluation workbench developed at the University of Southern California — and is built by one of its researchers. The recovery algorithms, the smell catalogue, and the drift baseline each trace to published, peer-reviewed work.
Author’s profile: Google Scholar — 686 citations, h-index 9 (as of 2026-07-21).
ARCADE (ESEC/FSE 2020) recovered architectures across hundreds of versions of open-source systems to measure how they change and decay. arcade-agent keeps that pipeline — parse, recover, detect smells, compute metrics — and changes the consumer: instead of producing datasets for researchers, it serves bounded architectural context to coding agents over MCP, and turns the version-to-version comparison into agit-committed baseline diffed in CI.
acdc — Pattern-based clustering: subgraph dominance, support-library detection.Introduced by Tzerpos & Holt, “ACDC: An Algorithm for Comprehension-Driven Clustering” (WCRE 2000).
wca — Agglomerative hierarchical clustering over entity-dependency similarity (UEM).In the lineage of Maqbool & Babri’s Weighted Combined Algorithm (CSMR 2004).
arc — Concern-based recovery: entities clustered by what they are about, not just what they touch.After Garcia, Krka & Medvidović, “Obtaining Ground-Truth Software Architectures” (ICSE 2013). This implementation assigns concern vectors with an LLM instead of MALLET topic modeling.
limbo — Information-theoretic agglomerative clustering — merges by minimal information loss.Andritsos, Dumitriu & Tzerpos, “Information-Theoretic Software Clustering” (IEEE TSE 2005).
pkg — Package-structure baseline — the null hypothesis every recovery result is judged against.Standard practice in the recovery-evaluation literature.
M. Schmitt Laser, N. Medvidović, D. M. Le, J. Garcia · ESEC/FSE 2020 · 52 citations
The research workbench this project is named for — and reimagined from.
D. M. Le, P. Behnamghader, J. Garcia, D. Link, A. Shahbazian, N. Medvidović · MSR 2015 · Best Paper Award · 149 citations
Architectures change measurably between versions, even when nobody decides they should. This is why the baseline is committed and diffed.
D. M. Le, D. Link, A. Shahbazian, N. Medvidović · ICSA 2018 · 117 citations
How architectural smells accumulate over a system’s lifetime — the study behind detect_smells’ catalogue.
D. M. Le, S. Karthik, M. S. Laser, N. Medvidović · ICSA 2021 · 8 citations
Decay metrics predict where bugs and churn land next — the reason drift is worth catching at PR time.
P. Behnamghader, D. M. Le, J. Garcia, D. Link, A. Shahbazian, N. Medvidović · Empirical Software Engineering, 2017 · 73 citations
Recovery algorithms disagree; using several views beats trusting one — why five algorithms ship, not one.
D. M. Le, N. Medvidović · ICSE 2016 (doctoral track) · 37 citations
Predicting implementation problems from architectural information alone.
D. M. Le, C. Carrillo, R. Capilla, N. Medvidović · WICSA 2016 · 86 citations
Decay as a sustainability problem, not just a code-quality one.
D. M. Le · PhD dissertation, University of Southern California, 2018 · 4 citations
Citation counts from Google Scholar, 2026-07-21. Full publication list on the profile.