Alibaba Open Code Review trims diffs, cuts tokens

Alibaba's approach to AI code review proves that deterministic preprocessing is essential for making LLM-driven PR analysis financially viable at scale. However, excluding test files by default is a contentious trade-off that prioritizes token budgets over catching fragile or untested edge cases.

• –Deterministic filtering over raw diff dumping: Passing raw git diffs directly to an LLM explodes context windows and costs; OCR's rule-based pre-slicing cuts review volume by over 70% before dispatching tasks to an LLM.
• –The test file blind spot: Skipping test files avoids spending tokens on boilerplate assertions, but it blinds automated reviews to missing assertions, flawed mocks, or dropped coverage.
• –Hybrid architecture is winning: Rather than relying solely on autonomous coding agents, tools combining AST and static analysis filters with focused LLM reasoning offer the highest signal-to-noise ratio in CI/CD.
• –Pipeline configurability is essential: Teams adopting OCR in production workflows must fine-tune inclusion patterns to balance token conservation with rigorous quality gates.