Rubber Duck Thursdays - Let's build

Here's a comprehensive summary of the YouTube video transcript:

Key Concepts

• Rubber Duck Thursdays: A recurring live stream session focused on coding, development, and exploring new technologies, particularly within the GitHub ecosystem.
• GitHub Copilot: An AI-powered coding assistant that provides code suggestions, completions, and can assist with various development tasks.
• MCP (Model Context Protocol): A protocol that enables AI agents to understand user intent and interact with APIs to perform actions, such as making moves in a game.
• Custom Instructions/Agents: Features within GitHub Copilot that allow users to define specific behaviors, tools, and prompts for AI agents, tailoring their functionality.
• Plan Mode: A specific mode within Copilot that focuses on generating detailed plans and specifications for development tasks, often leading to more structured AI output.
• Handoff: A mechanism where an AI agent can pass a task or a generated plan to another agent or a different mode for implementation.
• Dependency Management: The process of updating and managing software libraries and packages, often involving tools like Dependabot.
• CI (Continuous Integration): An automated process for building and testing code changes, crucial for maintaining code quality.
• OIDC (OpenID Connect): An authentication protocol used to verify the identity of users and systems, often employed in CI/CD pipelines for secure access to cloud services.

Summary of Content

Introduction and Catch-up

The stream begins with the host welcoming viewers back to "Rubber Duck Thursdays" after a break. The host expresses enthusiasm for reconnecting with the community and learning about what they are working on. The agenda for the session is outlined: reviewing recent GitHub changelog updates, revisiting a previously built MCP server for a Tic-Tac-Toe game, and exploring new features like Copilot custom instructions, custom agents, and demonstrating Copilot's assistance with dependency updates. The host also mentions the recent public preview of Gemini 3 Pro as another new development to consider.

GitHub Changelog Highlights

The host walks through recent updates from the GitHub changelog, emphasizing key features:

• Managing Copilot Coding Agent Tasks in VS Code: A new hub within Visual Studio Code allows users to track and manage various Copilot agents (Codex, CLI, Cloud, Chat) without leaving the IDE.
• New GitHub Actions OIDC Token Claims: Enhanced OIDC token claims provide more granular tracing of tokens to specific jobs and compute, improving security and enabling least privilege policies.
• GPT 5.1, 5.1 Codex, and 5.1 Codex Mini in Public Preview: New models are now available for GitHub Copilot, with different versions supporting chat, ask, edit, and agent modes across various editors. Enterprise users may need admin enablement.
• Configure Copilot Coding Agent as a Bypass Actor for Rule Sets: Copilot agents can now be configured to bypass certain rule sets, such as signed commits, offering flexibility.
• Fine-Grained Permissions for Copilot Usage Metrics: An enterprise role (enterprise copilot metrics) allows access to a dashboard for viewing Copilot usage metrics.
• Permissions Block Repository Admins for Installing GitHub Apps: Organization owners can now prevent repository admins from installing GitHub apps on their repositories, enhancing governance.
• Billing Date Standardization: For self-served credit card metered enterprise customers, billing dates are now standardized to the first of the month.
• Improved Enterprise License Consumption Reporting: Enhancements have been made to the CSV export for reporting on license consumption, especially for outside collaborators.
• Migrating Repositories with GitHub-Owned Blob Storage: This feature is now generally available, simplifying repository migration to GitHub Enterprise Cloud by removing the need for users to provide their own blob storage.
• MCP Registry and Allowlist Controls for VS Code Stable: MCP registries and allowlist controls are now available in VS Code stable, enhancing governance for MCP.
• Copilot Coding Agent for Eclipse in Public Preview: Copilot is now available for Eclipse users.
• Plan Mode in Public Preview: Plan mode is now available in JetBrains, Eclipse, and Xcode, allowing for iterative planning and agent implementation.
• GitHub Copilot Next/Edit Suggestions in Public Preview: This feature is available for Xcode and Eclipse.
• Isolated Sub-Agents for JetBrains, Eclipse, and Xcode: This feature, now in public preview, helps manage context bloat by allowing sub-agents to work on specific subtasks.
• Unified Code-to-Cloud Artifact Risk Visibility: Integration between Microsoft Defender for Cloud and GitHub security features is in public preview.
• Custom Agents Available in GitHub Copilot for JetBrains, Eclipse, and Xcode: Custom agents are now supported across these IDEs.
• Auto Model Selection for Copilot: This feature in JetBrains, Xcode, and Eclipse optimizes model availability and offers a 10% discount on model multipliers for premium users.
• Enhanced MCP Support for GitHub Copilot: Improved MCP support with OAuth is available for JetBrains, Eclipse, and Xcode.
• Gemini 3 Pro in Public Preview for GitHub Copilot: This new model is available for use.
• GitHub Copilot CLI Updates: New models, enhanced code search, and better image support have been added.
• CodeQL Updates: Support for Swift 6.2, new Java queries, and improved analysis accuracy for various languages and frameworks.

MCP Server and Tic-Tac-Toe Game Demonstration

The host revisits a previously built MCP server designed for playing Tic-Tac-Toe against an AI agent. The MCP server facilitates understanding user intent and using tool calls to execute game moves. The demonstration shows the AI making moves and the user responding, highlighting the natural language interaction. Some minor regressions are observed, indicating areas for improvement. The host switches between different models (Codex, Sonnet 45) to test their performance.

Copilot Assisting with Dependency Updates and CI Failures

A significant portion of the stream is dedicated to demonstrating how GitHub Copilot can help resolve issues arising from dependency version updates. The host presents a scenario where Dependabot creates a pull request, but the CI pipeline fails due to type check errors after the updates.

The host prompts Copilot with specific instructions:

1. Identify failing CI on a PR.
2. Switch to that PR's branch.
3. Run npm install to align with new dependencies.
4. Implement a fix for the CI issues.

Copilot attempts to diagnose the problem, create to-dos, run npm install, and then attempts to fix the type mismatch. Initially, Copilot's attempts are not entirely successful, leading to a discussion about prompt clarity and AI non-determinism. The host manually intervenes and then tries again with a more specific prompt, eventually leading to a successful resolution where the CI checks pass. The host reflects on the importance of trusting the AI and the iterative nature of debugging with AI assistance.

Exploring Custom Instructions and Custom Agents

The host then transitions to discussing custom instructions and custom agents.

• Custom Instructions: These allow for more targeted AI behavior, either generally or based on file paths (glob approach).
• Custom Agents: Formerly known as "chat modes," custom agents are a concept that translates across IDEs and Copilot coding agents. They allow users to specify tools, prompts, default LLMs, and handoff points.

The host demonstrates creating a custom agent for a 3D Tic-Tac-Toe game. The process involves:

1. Defining the Plan: The host provides a prompt to create a plan for a 3D Tic-Tac-Toe game, including toggling between 2D/3D and 3D navigation.
2. Iterative Refinement: The host uses Sonnet 45 to identify ambiguities in the plan and refine it, asking clarifying questions about user experience, camera angles, and visual effects.
3. Handoff to Implementation: The refined plan can then be handed off to an agent for implementation. The host chooses to have the plan written into a markdown document for team handoff.
4. Agent Implementation (CLI): The host uses the Copilot CLI agent to implement the 3D Tic-Tac-Toe game based on the generated markdown plan. This involves installing dependencies and executing the implementation.
5. Observed Regressions and Improvements: The implementation results in some regressions, but also introduces new features like confetti effects and 3D rotation. The host notes the importance of continuous experimentation and reviewing AI-generated code.

The host emphasizes that custom agents and instructions are powerful tools for tailoring AI behavior and that experimentation is key to leveraging them effectively.

Conclusion and Future Plans

The host concludes by summarizing the session's key takeaways, including the successful demonstration of Copilot assisting with dependency issues and the exploration of custom agents and plan mode. The host reiterates the commitment to regular "Rubber Duck Thursdays" and encourages viewers to return. Future sessions will delve deeper into custom agents, custom instructions, and techniques for using them across different IDEs. The importance of community interaction and continuous learning through experimentation is highlighted.