Key Signals

• All three major AI coding platforms now ship plugin ecosystems built on nearly identical architecture. OpenAI launched 20+ plugins for Codex bundling skills, MCP servers, and app integrations into one-click installs for services like Figma, Linear, Slack, and Vercel. Anthropic's Claude Code and Google's Gemini CLI already offered comparable systems, meaning the plugin war is now a feature-parity race rather than a differentiator — and cross-ecosystem portability is explicitly supported. This convergence signals that plugins/extensions will become a standard interface layer across all AI coding agents. [1][2]

• GitHub will begin using Copilot interaction data to train AI models starting April 24, sharing data with Microsoft. The policy applies to all Copilot Free, Pro, and Pro+ users on an opt-out basis, covering inputs, outputs, code context, file names, repository structure, and feature interactions. Business and Enterprise users are exempt. Developer communities reacted negatively to the opt-out model and the individual/corporate asymmetry, though GitHub defends the move as industry-standard practice that will improve models for its 26 million+ Copilot users. [3]

• Copilot CLI shipped two releases in a single day, introducing the /rewind timeline picker and MCP server LLM inference. Version 1.0.13-0 enables MCP servers to request LLM sampling with user approval and drops gemini-3-pro-preview model support, while v1.0.13-1 adds /rewind and double-Esc for rolling back to any point in conversation history. The V8 compile cache now accelerates CLI startup on repeated invocations, and MCP registry lookups gained automatic retries. [4][5]

• Claude Code v2.1.86 delivered 15+ bug fixes alongside meaningful token efficiency gains. The release reduces token overhead when mentioning files with @ by eliminating JSON-escaping of raw content, improves prompt cache hit rates for Bedrock/Vertex/Foundry users, and caps skill descriptions at 250 characters. Critical fixes address resume failures from older sessions, config corruption on Windows, OOM crashes during /feedback, and marketplace plugin permission failures on macOS/Linux. [6]

• A quantitative benchmark across 13 programming languages found Ruby, Python, and JavaScript are 1.4–2.6× faster and cheaper for Claude Code than statically typed alternatives. Running 600 trials of a mini-Git implementation, the top three dynamic languages completed in 73–81 seconds at $0.36–0.39, with notably low variance. Adding type checkers imposed significant overhead: Python/mypy was 1.6–1.7× slower than plain Python, and Ruby/Steep was 2.0–3.2× slower than plain Ruby. Only 3 of 600 runs failed, all in statically typed languages. [7]

• At QCon London 2026, agentic coding was reported to have overtaken tab-completion as the dominant form of AI-assisted development interaction. Teams using tools like Cursor in agentic mode are exhausting product backlogs before lunchtime — a phenomenon never seen before. AWS achieved 10x throughput through agentic coding only after completely re-engineering its testing and deployment pipeline, demonstrating that organizational change must accompany the speed gains. [8]

AI Coding News

• OpenAI's Responses API now provides a complete agent runtime with shell tool, execution loop, containerized workspace, and reusable skills. The shell tool moves beyond Python-only code interpretation to support Go, Java, Node.js, and standard Unix utilities. A built-in agent execution loop lets models propose actions iteratively in controlled environments. Network access is routed through a centralized policy layer with credential placeholders invisible to the model, and context compaction handles long-running tasks that would otherwise exceed token limits. [9]

• OpenBSD rejected an LLM-generated ext4 filesystem implementation over unresolvable copyright concerns, setting a precedent for AI code contributions to open-source projects. The code was created using ChatGPT and Claude Code without reading Linux source files, but OpenBSD maintainer Theo de Raadt stated that current copyright law cannot assign rights to AI-generated output, making redistribution permissions impossible. The contributor acknowledged "forking OpenBSD" might be easier than convincing the project to accept the code. The case raises fundamental questions about who maintains AI-generated code when the submitter does not understand it. [10]

Feature Update

• OpenAI Codex launched a plugin system with 20+ first-party integrations, available across the Codex app, CLI, and VS Code extension. Plugins bundle Markdown-based skills, MCP servers, and app connectors into single-click installs. Highlighted examples include a "build web app" plugin packaging Stripe, Supabase, and Vercel together. A plugin creator tool lets users scaffold new plugins from natural-language descriptions. Self-serve publishing is coming soon. The feature matches existing plugin/extension systems in Claude Code and Gemini CLI, and OpenAI explicitly supports importing plugins from other ecosystems. [1][2]

• Copilot CLI v1.0.13-0 added MCP server LLM inference sampling and removed gemini-3-pro-preview model support. MCP servers can now request LLM inference with user approval through a new review prompt. Fixes address MCP allowlist policy enforcement, BYOM reasoning effort settings, classic PAT error messaging, grep tool memory handling for large files, OAuth in ACP mode, and mouse-support paste corruption. [4]

• Copilot CLI v1.0.13-1 introduced the /rewind timeline picker for conversation history rollback. Double-pressing Esc now opens a timeline picker that can roll back to any point in conversation history, not just the previous snapshot. MCP registry lookups are more reliable with automatic retries and request timeouts. CLI startup is faster due to V8 compile cache reducing parse and compile time on repeated invocations. [5]

• Claude Code v2.1.86 shipped fixes for resume failures, Windows config corruption, and marketplace plugin permissions. The release added X-Claude-Code-Session-Id headers for API proxy aggregation and .jj/.sl VCS directory exclusions for Jujutsu and Sapling. Token efficiency improved through compact line-number format in the Read tool, deduplicated re-reads, and removal of JSON-escaping for @-mentioned file content. The /skills menu now sorts alphabetically and caps descriptions at 250 characters. VS Code extension fixes address "Not responding" false positives during long operations and model defaulting after OAuth token refresh. [6]

• Codex 0.118.0-alpha.3 continued rapid iteration following the v0.117.0 stable release that introduced first-class plugin support. Three alpha builds shipped on March 27, building on the previous day's stable release which made plugins a first-class workflow, introduced path-based sub-agent addresses, and enabled the app-server TUI by default. The alpha releases contain incremental fixes ahead of the next stable version. [11]

• Gemini CLI v0.36.0-preview.5 shipped as the latest preview in a series adding subagent tool filtering, native sandboxing, and Git worktree support. The preview build follows v0.36.0-preview.0 which introduced multi-registry architecture for subagent tool isolation, strict macOS sandboxing via Seatbelt allowlist, native Windows sandboxing, Git worktree support for parallel sessions, and an experimental memory manager agent. [12]