Parallel routing
All three reviews hit simultaneously. Wall time tracks the slowest model.
Developer tools
Parallelize security, logic, and style reviews instantly.
Latency
55%
15s → 6.5s per PR
Cached patterns
10K+
78% re-analysis cut
Parallel reviews
3
security + logic + style
Problem
Sequential review was the bottleneck, not compute.
A 12–15s review pipeline that ran security → logic → style sequentially killed developer flow. Every PR waited for a human-unfriendly delay because reviews were chained.
Solution
Three models. One Promise.all.
Fan out three parallel model calls via BR. Security on Sonnet; logic on GPT-4o; style on auto:floor. Wall time = the slowest of the three (~6.2s).
Semantic cache on patterns
Utility functions, handlers, parsers hit cache on subsequent PRs. Style reviews hit 45% cache.
auto:floor for style
Well-defined rules route to the cheapest qualifying model. 8× cost cut, zero quality loss.
“A 6-second review fits a glance workflow. A 15-second review kills context. Developers noticed the switch immediately.”