Rubber Duck in Copilot CLI: AI Code Review

Rubber Duck in GitHub Copilot CLI: Code Review by a Second AI Model

GitHub Copilot CLI has gained an experimental feature called Rubber Duck. The tool connects a second AI model to analyze the plan and code, minimizing errors from the primary model. If the main agent uses Claude Sonnet, the review is handled by a GPT-5.4-level model. This helps detect architectural flaws, logical bugs, and inter-file conflicts before execution.

In tests, the Claude Sonnet + Rubber Duck combo narrowed the efficiency gap with Claude Opus by 74.7% on complex tasks. On multi-file projects with long action chains, accuracy improved by 3.8–4.8%. Rubber Duck activates automatically at key points: after planning, complex implementation, or before tests. Manual activation is available via command.

When and How Rubber Duck Works

The tool doesn't scan every step but focuses on critical development stages:

• Planning: checking architecture and action sequences.
• Implementation: analyzing logic, loops, and potential data overwrites.
• Pre-test validation: detecting hidden bugs, like premature scheduler termination.

Example of an error caught by Rubber Duck: code in a loop silently overwrites data, breaking integrity. The feature is available via /experimental in Copilot CLI, which launched in September as a public preview.

Developers note that this dual-model architecture is especially useful for mid-level and senior specialists working with agents in the terminal. Copilot CLI already supports MCP setup, repository memory, and planning mode from the February update.

Benchmarks and Real-World Scenarios

Testing was done on multi-file tasks with complex dependencies. Results:

| Task | Base model (Claude Sonnet) | With Rubber Duck | Improvement |

|------------------------|----------------------------|------------------------|-------------------------|

| Complex chains | ~25.3% success | 74.7% of gap to Opus | +3.8–4.8% accuracy |

| Multi-file projects | Average efficiency | Significant growth | 74.7% gap narrowing |

Rubber Duck catches typical agentic development issues: state conflicts, implicit race conditions, and architectural mismatches. For senior developers, it's a tool for quick audits without manual reviews.

Integration with Copilot CLI streamlines workflows: terminal commands launch agents, and Rubber Duck adds a verification layer. In the future, such audits could become standard for AI-assisted development.

Key Points

• Rubber Duck uses a second model (e.g., GPT-5.4) to check Claude Sonnet, narrowing the gap with top models by 74.7%.
• Activates at key points: plan, implementation, tests; manual mode via command.
• Detects architectural errors, logical bugs, and inter-file conflicts.
• Available experimentally via /experimental in Copilot CLI.
• Suited for agent tasks, repository memory, and MCP setup.

— Editorial Team