GitHub Copilot CLI's Rubber Duck Uses a Rival AI to Catch Your Bugs

GitHub Copilot CLI goes GA with a cross-model Rubber Duck critic and on-device voice input that never leaves your machine

·
·
·
  • Rubber Duck GA: A built-in cross-model critic agent that reviews your plan using a model from a different AI family (Claude reviews GPT sessions and vice versa).
  • 74.7% gap closed: Claude Sonnet + Rubber Duck closes 74.7% of the performance gap between Sonnet and Opus on real-world coding tasks.
  • On-device voice input GA: Hold Space to dictate prompts; audio is transcribed locally using OS speech APIs, never sent to the cloud.
  • Prompt scheduling (experimental): New /every and /after commands let you run prompts on a timer within a session.
  • Redesigned terminal UI (experimental): Tabbed navigation for Issues, PRs, and Gists; accessible color modes including high-contrast and colorblind.
  • Free to existing subscribers: All features included in every Copilot plan; install with npm install -g @github/copilot.

GitHub Copilot CLI just landed four new generally available features, and two of them are worth stopping for. The headline additions are Rubber Duck , a built-in second-opinion agent that critiques your work using a model from a completely different AI family , and on-device voice input that transcribes your speech locally without ever touching a cloud API. Both shipped at Microsoft Build 2026 and are live now on every paid Copilot tier.

The Problem with One Model Reviewing Itself

The goal of Rubber Duck is to catch early planning mistakes before they compound into harder-to-fix downstream errors , a known weakness of single-model AI agents. The core issue is structural: a model reviewing its own output can only catch what its training allows it to see. The same blind spots that caused a bug are the same blind spots that will prevent it from noticing one.

Rubber Duck is a built-in review agent that runs alongside your main Copilot CLI session. After drafting a plan, mid-implementation on something complex, or after writing tests, the primary agent hands its work to the rubber duck for a second opinion. The rubber duck looks for blind spots, design flaws, and overlooked edge cases. It reports back with concrete, actionable feedback , it doesn't make file changes. Your main agent decides what to do with the feedback.

The architectural decision that makes this genuinely interesting: the critic always comes from a different model family than your session model. If you're running Claude, the rubber duck uses a GPT model. If you're on GPT, it uses Claude. Different training, different assumptions, different failure modes.

Terminal showing GitHub Copilot CLI listing open GitHub issues assigned to a user

The Numbers Behind It

In performance testing on real-world coding problems, combining Claude Sonnet with Rubber Duck closed 74.7% of the performance gap between Sonnet and the more capable Opus tier. The effect was strongest on the hardest problems , those spanning three or more source files and requiring 70 or more agent steps , where Sonnet with Rubber Duck scored 3.8% higher than Sonnet alone, and 4.8% higher on the hardest subset.

GitHub published concrete examples of the kinds of bugs it catches:

  • An async scheduler that would start and exit immediately without running any jobs
  • A loop that silently overwrote the same dictionary key on every iteration, discarding three of four search categories with no error thrown
  • A cross-file conflict where new code stopped writing a Redis key that three other files depended on, leaving the UI and cleanup paths silently broken on deployment

These are exactly the class of bugs that are easy to miss in code review and painful to debug in production. You can also invoke it manually at any point with /rubber-duck.

Voice Input That Stays on Your Machine

Voice input is now generally available in GitHub Copilot CLI, and the implementation detail that matters most is where the transcription runs: locally. Audio never leaves your machine. GitHub uses the built-in speech recognition APIs on Windows and macOS , no API call to a cloud transcription service.

Two ways to use it:

  • Hold the space bar on your keyboard and talk to input a prompt.
  • Press Ctrl+X followed by V to start recording, speak your prompt, then press any key to stop recording and insert the transcription.

The first time you enable voice input, the CLI guides you through downloading the runtime and picking a speech-to-text model. English and Spanish are supported. The privacy-first design matters for enterprise teams on sensitive codebases , no audio ever hits a third-party transcription API.

Terminal showing a custom agents menu with numbered options including accessibility, code review, and SEO optimization agents

Scheduled Prompts and a New Terminal UI

Two more features round out the release. Prompt scheduling (currently experimental, enable with /experimental on) adds /every and /after slash commands that let you fire prompts on a timer within your session:

dockerfile
# Run frontend tests every 30 minutes
/every 30m run the frontend tests
# Summarize token usage once, after an hour
/after 1h how many tokens have I used this session
# Open the schedule manager
/every

This turns Copilot CLI from a reactive chat tool into something closer to an ambient background agent. The redesigned terminal UI (also experimental) adds tabbed navigation , press Tab to switch between your session, the repo's Issues, Pull Requests, and personal Gists without leaving the CLI. The redesign also adds new color modes including default, github, dim, high-contrast, and colorblind to match your terminal and your eyes.

What's GA vs. Experimental

FeatureStatusHow to enable
Rubber DuckGenerally AvailableBuilt-in, or /rubber-duck
Voice InputGenerally AvailableHold Space, or Ctrl+X V
Prompt SchedulingExperimental/experimental on
New Terminal UIExperimental/experimental on

Getting It

Copilot CLI is included as a core feature of all GitHub Copilot plans (Free, Pro, Pro+, Max, Business, and Enterprise). Access is included in your existing subscription , each interaction draws on your plan's AI Credits allowance. Install via npm with npm install -g @github/copilot, or use the one-liner:

code
curl -fsSL https://gh.io/copilot-install | bash

Update an existing installation by running copilot update in your terminal. For Business and Enterprise users, your org admin needs to enable Copilot CLI first , it then automatically inherits your organization's existing governance policies and branch protections.

The cross-model critic pattern is the real signal here. It's a practical acknowledgment that single-model agentic loops have a ceiling, and that ceiling is most visible on the hardest, multi-file tasks where subtle bugs compound. Rubber Duck is a direct architectural response to that problem , and the performance data suggests it actually works.

Trending
  • No trending articles

Comments

avatar

Next Reads