Replica: prototypes built on your own components
October 28, 2026
GitHub Design released Replica under MIT. The tool builds interactive prototypes using a team’s components, patterns, and documentation. An optional self-hosted Replica Hub supports shared viewing: the interesting idea is that prototyping starts with the product’s existing system.
Source: github.blogCopilot Fusion: orchestration instead of manual selection
October 28, 2026
Fusion became generally available. Users select Standard, High, or Max effort; the system selects and, when necessary, combines models. Organizations retain control of available models through policies; this does not promise identical results on every task.
Source: github.blogLocal models join Fusion
October 28, 2026 · experimental feature
The CLI gained an experimental choice between local and cloud models. Managed Windows ML installation is available on supported Windows devices; manually configured models also work in the CLI on macOS and Linux. A key boundary: routing decisions still happen in the cloud, so local inference does not mean an offline session.
Source: github.blogImages directly in the Copilot workspace
October 28, 2026 · public preview
Copilot can generate and edit images from text prompts, show the result in chat, and save it to a project. The feature is available in VS Code, the app, and CLI; image generation is accounted for separately from the conversational model in AI credits. Organizations can disable it through model policy.
Source: github.blogActions gets a workflow dependency lockfile
October 28, 2026 · generally available
The .github/workflows/actions.lock file pins exact action commits, including transitive dependencies, while keeping readable tag references in workflows. The gh actions-lock extension creates the file; Dependabot updates the corresponding entry with the dependency. Pinning makes runs more reproducible but does not itself prove code is safe.
Source: github.blogAI Scan checks the whole repository
October 28, 2026 · public preview
The new mode analyzes the latest commit on the default branch, not only pull request changes. Findings enter the shared code scanning queue alongside CodeQL; CodeQL duplicates are skipped. A rerun is needed after a fix: merging a patch does not itself close an AI Scan alert.
Source: github.blogDependabot shows the path to a vulnerable function
October 28, 2026 · public preview
Reachability analysis for JavaScript and TypeScript searches for a call chain from your code to a vulnerable dependency function. During preview the feature is free for Enterprise Cloud and requires an npm, Yarn, or pnpm lockfile. Results help prioritize work but do not automatically close alerts.
Source: github.blogAgentic Workflows becomes generally available
October 28, 2026
The GitHub Next project runs agent tasks inside Actions: triaging issues, investigating CI failures, updating documentation, and writing tests. By default the agent receives read permissions; proposed changes undergo separate review. A workflow description becomes another way to express work for which a conventional script is insufficient.
Source: github.blogarch-diagram-skill: architecture maps from source
Found October 29, 2026 · release date not stated
The Claude Code and Codex skill turns a repository or article into an animated architecture diagram. It uses spec.json as input and supports HTML, PNG, MP4, GIF, and Mermaid. A useful way to explain a project’s structure, not proof that a generated diagram accurately describes any code.
Source: github.combig-arrow-on-the-screen: show where to look
Found October 29, 2026 · release date not stated
This macOS utility displays an arrow or card over an interface: it can point at an element, provide a value to copy, or wait for a person’s answer. It does not click buttons or enter text. An interesting way to hand the next step to a person in the application’s context.
Source: github.comdimensional-drawing: anatomy of a UI component
Found October 29, 2026 · release date not stated
The project presents UI components as dimensioned drawings: an overview, layers, corner details, and variable tables. Input can be a file, Storybook, URL, or screenshot. An animated version of the finished sheet is also available; dimensions and explanations should still be checked against the source component.
Source: github.comFloorplan: read an application as a map
Found October 29, 2026 · release date not stated
Floorplan creates an interactive codebase map with plain-language explanations. Users can trace a path from a screen to related system parts, ask questions, and update the map after changes. It works with Claude Code or Codex and requires Node.js 18+. It is a map for exploration, not a substitute for reading critical code.
Source: github.comOpenIDEA: familiar habits in another editor
Found October 29, 2026 · release date not stated
The open Rust/Tauri IDE reproduces familiar keyboard shortcuts and an IntelliJ IDEA-style layout. The repository lists an editor, debugger, Git, terminal, database clients, and HTTP. The project is not affiliated with JetBrains; the author’s speed comparisons should not be assumed to apply to your machine without measurement.
Source: github.comSwarmLLM: local models and shared inference
Found October 29, 2026 · release date not stated
The open application runs chat models on your computer and offers to connect machines over the internet for models too large for one. OpenAI- and Anthropic-compatible APIs are claimed. This selection presents a direction to explore; performance and network operation were not measured here.
Source: github.comTailscale examines SSH privilege boundaries
October 28, 2026
The company published an analysis of issues found and fixes in version 1.104.4. The described SSH vulnerabilities require access to an already authorized device and connections permitted by tailnet policy; they do not allow unauthorized login from an arbitrary machine. The source links separate advisories with conditions and fixes.
Source: tailscale.com