Rules - Program Graphs for Code Analysis

Rules in Program Graphs for Code Analysis

IDE rules exported with workflows to your editor

Title	Slug
Always add backward edges	`always-add-backward-edges`
Code must be fully compilable	`code-must-be-fully-compilable`	Read the 'Code must be fully compilable' rule from playbook 'Program Graphs for Code Analysis'. Slug: code-must-be-fully-compilable. always_apply: true. All projects must compile successfully before graph construction. Precise type information — including types from external libraries — is required to extract data flow edges and type hierarchy embeddings. Discard any project that cannot be fully compiled.
Randomly sample supertypes during training	`randomly-sample-supertypes-during-training`	Read the 'Randomly sample supertypes during training' rule from playbook 'Program Graphs for Code Analysis'. Slug: randomly-sample-supertypes-during-training. always_apply: true. During training, randomly select a non-empty subset of τ(v) (the full supertype set) for each variable. This acts as dropout over the type lattice and ensures all known types in the lattice are trained. Use the full τ(v) at inference time.
Split dataset by file, not by example	`split-dataset-by-file-not-by-example`
Use exactly 8 GGNN propagation steps	`use-exactly-8-ggnn-propagation-steps`