Rubber Duck Thursdays | Rubber-Duck Agent

Key Concepts

• GitHub Copilot CLI: A command-line interface tool for interacting with GitHub Copilot.
• Bring Your Own Key (BYOK): A new feature allowing users to connect Copilot CLI to external model providers (Azure OpenAI, Anthropic, etc.) or local models.
• Rubber Duck Agent: An experimental AI agent that provides a "second opinion" by using a model from a different AI family to review the primary agent's work.
• Dependabot Remediation: The ability to assign dependency vulnerability alerts to AI agents to generate automated, tested pull requests.
• Integrated Browser: A VS Code feature used to troubleshoot web applications directly within the IDE.
• Model Providers: Services like Azure OpenAI or local runtimes like Ollama that can now power Copilot CLI sessions.

---

1. GitHub Copilot Usage Metrics

GitHub has introduced granular usage metrics for Copilot Code Review. Enterprise and organization administrators can now track:

• Who is actively using Copilot for code reviews.
• Who is requesting reviews from the AI.
• Who is applying suggestions provided by the AI. These metrics are available via the API, providing clear visibility into the adoption and engagement levels of AI-assisted code reviews within an organization.

2. Dependabot and AI Remediation

A significant improvement allows Dependabot alerts to be assigned directly to AI coding agents (e.g., Copilot, Claude, or CodeX).

• The Problem: Simple version bumps often fail to address breaking API changes, deprecated methods, or incompatible type signatures.
• The Solution: The assigned agent analyzes the vulnerability, drafts a pull request with the fix, and attempts to resolve any test failures introduced by the update.
• Workflow: Users navigate to the Dependabot alert detail page and assign the task to an agent. Multiple agents can be assigned simultaneously to compare different implementation strategies.

3. Copilot CLI: Bring Your Own Key (BYOK) & Local Models

The Copilot CLI now supports custom model providers, moving beyond GitHub-hosted models.

• Supported Providers: Azure OpenAI, Anthropic, and any OpenAI-compatible endpoint.
• Local Models: Users can run models locally (e.g., via Ollama) and connect them to the CLI.
• Requirements: Models must support tool calling and streaming. A context window of at least 128k tokens is recommended for optimal performance.
• Configuration: Users set environment variables for PROVIDER_TYPE, BASE_URL, API_KEY, and MODEL_NAME.
• Authentication: GitHub authentication becomes optional when using custom providers.

4. The Rubber Duck Agent (Experimental)

This experimental feature introduces a "second opinion" mechanism to improve AI output quality.

• Methodology: When a primary model (e.g., Claude Sonnet 3.5) generates a plan or code, the Rubber Duck agent—running a model from a different AI family (e.g., GPT-4o)—reviews the work.
• Why it matters: Research shows that self-reflection within the same model family is often limited by the model's own training biases. By using a different family, the system can catch errors that the primary model might miss.
• Activation Points:
  1. After drafting a plan: To ensure the strategy is sound.
  2. After complex implementation: To provide a fresh set of eyes on the code.
  3. After writing tests: To catch gaps or "lazy" tests designed only to pass.
  4. Reactive: If the primary agent gets stuck in a loop, the Rubber Duck agent can be invoked to break the "log jam."

5. Practical Demonstrations

• Azure OpenAI Integration: The host demonstrated deploying a model on Microsoft Foundry and configuring the Copilot CLI to use it, successfully monitoring token usage via both the CLI and the Foundry dashboard.
• Local Model Execution: The host connected the CLI to a local Gemma 4 model via Ollama, highlighting the ability to work offline or reduce costs.
• Research Workflow: The host utilized the /research command to gather information on LangChain, followed by a manual request for a Rubber Duck review to critique the generated research report.

Synthesis and Conclusion

The latest updates to GitHub Copilot focus on flexibility and reliability. By allowing users to bring their own models (BYOK) and introducing the Rubber Duck agent, GitHub is shifting from a "one-size-fits-all" AI approach to a more robust, multi-model, and agentic workflow. The ability to automate complex Dependabot remediations and use cross-model critiques significantly reduces the manual burden on developers, ensuring that AI-generated code is not only faster to produce but also more thoroughly vetted and resilient.