Overwatch: Learning Patterns in Code Edit Sequences
summary
Overwatch learns recurring edit sequences from IDE session traces by building fine-grained edit graphs, summarizing them into an edit quotient graph, anti‑unifying examples into templates with unary and binary predicates, and agglomeratively clustering patterns. Trained on 135,000 C# versions (12 developers, ~250 hours) it predicts next edits with 78% precision, ~36% recall and an F3 ≈ 70. It surfaces 51 deployable patterns (17 match existing Visual Studio refactorings, 34 are new) and ships as Visual Studio tooling to auto-suggest edits using temporal context.
AI-written summary · about
chapters (8)
- 0:00Start
- 1:18Spatial Context based Suggestions
- 5:29Understand User Editing behaviour before Change...
- 8:25What do traces of IDE sessions look like?
- 15:44Generalizing Edit Sequences
- 16:57Agglomerative Hierarchical Clustering
- 21:17Performance Comparison
- 23:35PROSE (aka.ms/prose)
related talks
23mFast Focus: Sneak Peek at How Visual Studio's Code Editor Works Under the Hood
10mSporq: An Interactive Environment for Exploring Code Using Query-by-Example
01h 32mOptimizing Visual Studio Code for PowerShell, 2023 Edition
09mWhat's new in Visual Studio IntelliCode
47mKnowledge in code is gold — how to extract it and power your team and AI
39mPower of AI and Code Quality and Performance
42mThe Future of Visual Studio
06mreCode: A Lightweight Find-and-Replace Interaction in the IDE for Transforming Code by Example