GitHub Copilot CLI Offers a 'Second Opinion' Using a Different Model Family

GitHub Copilot CLI has introduced Rubber Duck in experimental mode, adding a structure where a model from a different AI family than the main model reviews the work once more. The core idea is to address the limitations of self-review by the same model after planning, complex implementation, and test writing.

The key is an independent second perspective. When using the Claude family as the orchestrator, Rubber Duck runs on GPT-5.4, and conversely, other model combinations are also continuously being tested. The judgment is that models with different training biases and blind spots, rather than simple self-reflection, can better catch gaps in plans, edge cases, and cross-file conflicts.

Evaluation results support this. The Claude Sonnet 4.6 + GPT-5.4 Rubber Duck combination closed 74.7% of the performance gap with Claude Opus 4.6 alone, and particularly on difficult tasks touching 3 or more files and going 70+ steps, it showed a 3.8% improvement over Sonnet, and in the hardest problem set across three experiments, a 4.8% improvement.

Rubber Duck automatically intervenes at the following points:

• After writing a plan: prevents an incorrectly set structure early on from spreading throughout the subsequent work.
• After complex implementation: catches missed dependencies and edge cases in multi-file changes.
• After writing tests, before execution: filters out coverage gaps or incorrect assertions in advance.

Real-world examples presented include a design problem where a scheduler terminates right after execution, a bug where a loop overwrites the same dict key causing 3 Solr facet categories to disappear, and a cross-file conflict where an email verification flow breaks because a Redis key is no longer used. Users can enable the feature with /experimental, and can also manually request critique and correction at any time if needed.