Before: Single-Model Loop

Self-reflection alone keeps repeating the same blind spots from the same training data

Plan Implement Test Fix self-reflection (repeats the same blind spots)

With Rubber Duck: Critique From a Different Family

A complementary family (GPT-5.4) critiques each step of the main (Claude) flow

Plan Implement Write tests Done Duck review #1 Question assumptions Duck review #2 Spot deps & edge cases Duck review #3 Flag coverage gaps A model from a different family critiques the main loop at three checkpoints Claude (blue) = orchestrator / GPT-5.4 (amber) = reviewer
Claude Sonnet 4.6 / Opus 4.6 / Haiku 4.5 (orchestrator)
GPT-5.4 (Rubber Duck reviewer)

Source: GitHub Blog "GitHub Copilot CLI combines model families for a second opinion" (April 6, 2026)