GitHub Copilot's Rubber Duck Review Comes to VS Code
GitHub CopilotGitHub Copilot's Rubber Duck agent — an experimental "second opinion" reviewer that debuted in Copilot CLI in April 2026 — is now available inside VS Code's Agent Host sessions as part of the VS Code 1.135 release. Rubber Duck runs a complementary model from a different AI family than the one driving the main agent, checking plans, code, and tests for missed details and edge cases. GitHub says pairing Claude Sonnet 4.6 with a Rubber Duck running GPT-5.4 closed 74.7% of the performance gap to the more capable Claude Opus 4.6 on SWE-Bench Pro benchmarks.
Key Takeaways
- Rubber Duck, GitHub Copilot's cross-model "second opinion" reviewer, has expanded from Copilot CLI (where it launched in April 2026) to VS Code's Agent Host sessions in version 1.135.
- Developers can invoke it manually with the
/rubber-duckslash command, or let Copilot trigger it automatically after a plan, a complex implementation, or before running tests. - The reviewer always runs a different model family than the primary agent — pairing a Claude orchestrator with a GPT-5.4 reviewer, for example.
- GitHub's own benchmarking found the Sonnet + Rubber Duck pairing closed 74.7% of the performance gap to Claude Opus 4.6 on SWE-Bench Pro.
- The biggest accuracy gains showed up on multi-file, multi-step tasks, where a single model is more prone to missing context.
- The feature remains labeled experimental, included with any GitHub Copilot subscription that has access to VS Code Agent Host sessions.
Sources & Mentions
1 external resource covering this update
Rubber Duck Expands From the CLI to VS Code
GitHub Copilot's Rubber Duck feature, an experimental second-opinion reviewer first introduced in GitHub Copilot CLI in April 2026, has arrived in Visual Studio Code as part of the VS Code 1.135 release. Inside a Copilot agent host session, developers can now invoke /rubber-duck to bring in a complementary model that critiques the primary agent's plans, code, and tests before work is considered finished.
How It Works
Rubber Duck is built around cross-model review rather than a single model checking its own work. When a developer selects a Claude model as the primary orchestrator, Rubber Duck runs on GPT-5.4 as the reviewer, and the pairing flips when the orchestrator changes. GitHub Copilot can trigger Rubber Duck automatically at three points in an agent session: after the agent drafts a plan, after a particularly complex implementation, and after writing tests but before running them. Developers can also invoke it manually at any point in the session.
Why It Matters
According to GitHub's own benchmarking, pairing Claude Sonnet 4.6 with a Rubber Duck reviewer running GPT-5.4 closed 74.7% of the performance gap between Sonnet and the significantly more expensive Claude Opus 4.6 model on SWE-Bench Pro. The gains were most pronounced on harder, multi-file tasks requiring dozens of steps, where a single model is more likely to miss context or make compounding mistakes. Bringing Rubber Duck to VS Code means developers no longer have to switch to the CLI to get a second opinion on agent-generated work — it is now available directly inside the editor where most Copilot agent sessions already run.
Availability
Rubber Duck in VS Code is currently marked experimental and ships as part of VS Code 1.135, available to any Copilot subscriber using Agent Host sessions in the editor.