Rubber Duck Thursdays

Key Concepts

• GitHub Copilot Memory: A persistent storage feature that allows Copilot to remember repository-specific instructions, preferences, and context across different agents and sessions.
• Copilot CLI (Command Line Interface): An AI-powered agent that operates within the terminal, supporting interactive, headless, and server modes.
• Multi-Agent Orchestration: The ability for a main "orchestrator" agent to dispatch tasks to specialized sub-agents (e.g., Code Review, Rubber Duck) to improve output quality and manage context windows.
• GitHub Copilot SDK: A toolkit allowing developers to use the Copilot CLI as a backend for custom AI-powered applications.
• Cross-Model Family Validation: A design pattern where a primary agent (e.g., Anthropic) is paired with a secondary agent from a different model family (e.g., OpenAI) to ensure higher accuracy and reduce hallucinations.

---

1. Recent GitHub Change Log Updates

• Code Coverage on Pull Requests: Now in public preview. It provides an aggregated percentage of code coverage directly within PRs, allowing maintainers to evaluate test completeness before merging.
• GitHub Classroom Retirement: New sign-ups are disabled. Existing classrooms will be supported until August 28, after which the platform will be decommissioned in favor of partner solutions like Codelio and Classroom 50.
• Code Quality Repository Enablement API: New endpoints allow programmatic configuration of code quality settings for repositories. Supported languages include C#, Go, Java, Kotlin, JavaScript, TypeScript, Python, and Ruby.
• Granular Copilot Model Control: Enterprise owners can now define which Copilot models (e.g., OpenAI vs. Anthropic) are available to specific organizations within their enterprise via a refreshed UI.

---

2. GitHub Copilot Memory & Controls

• Functionality: Copilot Memory stores facts about a repository (instructions, preferences) that persist across local CLI sessions and cloud-based agents.
• Management: Users can now delete or update memory logs via the GitHub.com repository settings or through the slash memory command in the CLI.
• Lifecycle: To prevent stale information, memory logs are refreshed/dropped approximately every 28 days.

---

3. Copilot CLI: Methodologies and Modes

• Interactive Mode: The standard back-and-forth chat interface.
• Headless Mode: Designed for automation (e.g., CI/CD pipelines). It uses flags like --no-ask-user and strict URL restrictions to execute tasks without human intervention.
• Server Mode: Exposes a JSON RPC port, allowing external applications to use the Copilot CLI as an AI backend via the Copilot SDK.
• Fleet Command (/fleet): An orchestration feature where a main agent breaks down a complex prompt into sub-tasks, dispatches them to specialized agents, and consolidates the results. This keeps the main context window clean.

---

4. Best Practices and Troubleshooting

• Customization: Users are encouraged to visit the "Awesome Copilot" repository to find community-contributed custom instructions (e.g., security/OWASP standards) to improve agent performance.
• Model Selection: The host suggests mapping specific tasks to specific models (e.g., GPT-3.5 for coding, Sonnet 4.6 for UI design) rather than relying solely on "Auto Mode."
• Licensing Guardrails: The "Suggestions matching public code" setting can be enabled to prevent Copilot from suggesting code that matches existing public repositories, or disabled to allow the agent to provide links and license information for generated code.

---

5. Notable Quotes

• "The beauty with the Copilot CLI... is that you get access to different models from different families... you’re able to have all these models working together and you can leverage their different strengths and substitute their different weaknesses."
• "If you run your main agent using a GPT model, then the rubber duck agent is going to spin up using a cloud model... you basically get the best of both worlds."

---

6. Synthesis and Conclusion

The session highlights a shift toward multi-agent, terminal-native workflows. By leveraging the Copilot CLI's ability to orchestrate sub-agents and maintain persistent memory, developers can automate complex tasks—such as code reviews, security checks, and even conference session planning—directly from their terminal. The key takeaway is that Copilot is not just a "code completion" tool but a programmable AI backend that, when properly configured with custom instructions and model-specific task mapping, significantly enhances developer productivity and code quality.