DiffLens

Deterministic Validation Layer for LLM-Generated Code Reviews

DiffLens enforces structural correctness between LLM-generated feedback and actual git diff topologies before comments hit GitHub Pull Requests.

Design Principle

LLMs propose review logic, deterministic systems authorize structural validity. DiffLens treats LLM outputs as untrusted suggestions.

Why It Exists (The Structural Mismatch)

LLM-based code review systems frequently produce unstable line references (referencing deleted lines, out-of-hunk coordinates, or shifted line numbers).

This is not a prompt engineering issue. It is a structural limitation: LLMs operate on token sequences, while diff correctness depends on stateful, line-indexed reasoning over structured edit regions defined by hunk headers (@@ -l,s +l,s @@).

System Flow

Git Diff ──> Structured Diff Parser (AST) ──> [Immutable Reference Map]
                                                           │
LLM Reviewer ──> [Untrusted Suggestion Payload] ───────┼─> (Deterministic Validation Engine)
                                                           │
GitHub API <── [Filtered / Realigned PR Output] <──────┘