<p><strong>TLDR:</strong> OpenAI shipped a plugin system for Codex with 20+ integrations at launch, plus hooks support — two features that meaningfully close the gap with Claude Code. Meanwhile, Anthropic’s rate limit situation is getting loud: multiple $200/month Max subscribers hit walls within hours and some have already switched. Cursor’s cloud agents crossed one million commits in two weeks. Qwen 3.5 27B is turning heads as a local model worth taking seriously.</p>
<hr>
<h2>Codex plugin launch</h2>
<p>OpenAI rolled out plugins for Codex with over 20 integrations at launch: Slack, Figma, Linear, Sentry, Notion, Box, Gmail. They’re packaged as installable bundles combining MCP configs, reusable workflows, and app integrations, available across the Codex app, CLI, and VS Code extension. Enterprise admins get a governance layer to push or block specific plugins across teams. The one gap: OpenAI hasn’t opened self-serve publishing yet, so it’s still behind GitHub Copilot Extensions and Cursor on third-party marketplace depth. shadcn/ui shipped as an official Codex plugin on day one.</p>
<h2>Claude Code rate limits</h2>
<p>The complaints are consistent enough to be a real signal. Multiple developers on $200/month Max plans are hitting rate limits within two to three hours of a session, mid-project, with no warning. At least a few have cancelled and moved to Codex. One developer who built $4,664 MRR on Claude Code called the 529 errors and outages this week “unusable” for building a company. Anthropic also cut Opus 4.6 rate limits without notice, which is a separate frustration from the outages. The reliability story is getting harder to defend.</p>
<h2>Cursor cloud agents: one million commits</h2>
<p>Cursor’s cloud agents generated over one million commits in two weeks, with minimal human intervention. The agents run code on their own compute. It’s a striking number — not because commits are a great quality metric, but because it shows autonomous agent loops running at a scale that wasn’t plausible even six months ago.</p>
<h2>GitHub Copilot training data opt-out</h2>
<p>Starting April 24, GitHub will use interaction data from Copilot Free, Pro, and Pro+ users to train AI models by default. That includes prompts, generated outputs, code snippets, file context, and repository structures. Users can opt out through Privacy settings. Business and Enterprise customers are unaffected — their contracts already handle this separately. The asymmetry is the real story: enterprise customers get data protection by default, individual users have to find the opt-out themselves.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Codex hooks with conditional filtering:</strong> Claude Code hooks now support an <code>if</code> field using permission rule syntax, so hooks only fire on specific commands rather than every match — useful in long sessions where certain hooks only apply to one command.</li>
<li><strong>Claude Code PowerShell support:</strong> Native PowerShell execution landed, no workarounds needed — relevant for Windows-heavy shops.</li>
<li><strong>Qwen 3.5 27B local performance:</strong> Early reactions are strong. Developers running it on MacBook Air hardware are summarizing tens of thousands of words in seconds. Worth testing if you’re tracking local model quality.</li>
<li><strong>Gemini 3.1 Flash Live:</strong> Google launched a new speech-to-speech model with day-zero support in Pipecat. Reactions from people who worked on it are positive, though Gemini’s coding reputation has taken hits lately from paying customers frustrated with limits and quality regression.</li>
<li><strong>GLM-5.1 released:</strong> Z AI’s post-training update to GLM-5 is available to GLM Coding Plan users now, with open-source release expected in the first half of April. Early testing shows strong agentic task performance, though it regresses on general chat.</li>
<li><strong>SlopCodeBench results:</strong> A University of Wisconsin-Madison benchmark tested 11 models — including Claude Opus 4.6 and GPT-5.4 — on 20 problems requiring agents to extend their own code across 93 checkpoints. Zero models finished end-to-end. The benchmark is specifically designed to expose compounding failures in self-extending code.</li>
<li><strong>GitHub Copilot training data:</strong> The opt-out is in Privacy settings — worth checking if you use Copilot in private repos and haven’t looked at this yet.</li>
<li><strong>Anthropic’s Operon agent:</strong> Reportedly in development for Claude Desktop, targeting biology research workflows. Think Cowork but scoped to scientific sessions with artifact management and Skills support.</li>
<li><strong>jai sandbox tool:</strong> A lightweight Linux sandbox for running AI agents with copy-on-write overlay isolation. Three modes (Casual, Strict, Bare), no Docker or images required. Positions itself as a blast-radius reducer rather than a hardened security boundary.</li>
<li><strong><a href="http://FastRuby.io" target="_blank" rel="noopener nofollow">FastRuby.io</a> Claude Code Skills:</strong> Open-sourced three skills encoding their Rails upgrade methodology, built from 60,000+ developer-hours of real upgrade work. Covers dual-boot setup, load-defaults alignment, and sequential version-by-version upgrades.</li>
</ul>


AI coding assistants, code generation tools, and vibe coding workflows.

Agentic Digest

OpenAI launched a plugin system for Codex with 20+ integrations (Slack, Figma, Linear, Sentry, Notion) plus hooks with conditional filtering, narrowing the gap with Claude Code. Anthropic is facing backlash as $200/month Max subscribers hit rate limits within hours mid-session, with some users switching to Codex. Cursor's cloud agents crossed one million commits in two weeks. GitHub will use Copilot Free/Pro interaction data for AI training by default starting April 24, with an opt-out available in Privacy settings. Qwen 3.5 27B is drawing attention as a capable local model, and a new SlopCodeBench benchmark found zero out of 11 models could complete end-to-end self-extending code tasks.

Codex gets plugins and hooks, Claude Code rate limits push users out