<p><strong>TLDR:</strong> Anthropic accidentally shipped a sourcemap file in a Claude Code npm update, exposing hundreds of thousands of lines of TypeScript and triggering a wave of community analysis, rewrites, and DMCA drama. OpenAI reset Codex usage limits for the sixth time in March, with the team admitting they still don’t fully understand why demand keeps spiking. Anthropic’s DMCA response caught legitimate open-source forks in the crossfire, drawing sharp criticism. Meanwhile, leaked references point to unreleased features and internal model codenames that suggest a significant product shift is coming.</p>
<hr>
<h2>Claude Code source leak</h2>
<p>An npm release of Claude Code version 2.1.88 shipped with a 57MB sourcemap file, exposing somewhere between 390,000 and 600,000 lines of TypeScript depending on who’s counting. The leak revealed an 11-step prompt pipeline, anti-distillation poison pills (fake tool definitions designed to mislead competitors training on the output), an undercover mode that strips internal references in external repos, a frustration detector triggered by profanity, and references to unreleased features including a background agent called Chyus/Chyros, a coordinator mode for parallel sub-agents, and internal model codenames like Capiara, Numbat, and Tengu. The community moved fast: Python and Rust rewrites appeared within 48 hours, one repo hit 50,000 stars in under two hours, and the code spread well beyond what DMCA requests could contain.</p>
<h2>Anthropic’s DMCA response</h2>
<p>Anthropic filed takedown requests targeting 97 repositories with copies of the proprietary source, which is defensible. The problem is GitHub extended those takedowns to forks of Anthropic’s own open-source Claude Code SDK, catching legitimate projects that never touched the leaked code. The optics are bad. Anthropic has a public repo called “claude code” that gets forked under the same name, and blanket takedowns that can’t distinguish between proprietary leaks and licensed forks look either careless or hostile. Several developers are calling it DMCA abuse. The community’s read is that this reflects a broader anti-open-ecosystem posture, and the leak itself has made that harder to argue against.</p>
<h2>Codex usage limits reset six times in March</h2>
<p>OpenAI’s Thomas Sottiaux announced another Codex limit reset, the sixth in March alone. His explanation was direct: dashboards show users hitting rate limits at an elevated rate, the team doesn’t fully understand why, and resetting is the cautious call while they investigate. A cluster of fraudulent accounts was also identified and banned, freeing up some compute. Six resets in a month means demand is consistently outrunning whatever ceiling OpenAI sets, and the team is still working out the shape of the problem. Separately, Codex’s core is being open-sourced, which should make the ecosystem dynamics even harder to predict.</p>
<h2>Leaked Anthropic roadmap: Conway and Epitaxy</h2>
<p>Beyond the sourcemap, two separate leaks point to bigger product moves. Conway is reportedly an always-on agent platform with its own UI, browser control, webhook invocation, and a new CNW ZIP extension standard. Epitaxy appears to be a unified interface merging Claude Code’s local execution with web-based task launching under a single <a href="http://claude.ai" target="_blank" rel="noopener nofollow">claude.ai</a> route, with a distinct dark CRT-style design. Internal references suggest the next frontier model (Mythos or Capiara) is already at version 8 internally with a 1 million token context window and a fast mode. The speculation is that Epitaxy and Mythos launch together as a coordinated product shift rather than incremental updates.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Claude Code renderer rewrite:</strong> The team shipped an experimental virtual viewport renderer with mouse support and a persistent bottom-anchored prompt input, a quality-of-life improvement users had been requesting for a while.</li>
<li><strong>Claude Code usage limits:</strong> Multiple users report hitting 49% of weekly usage limits midweek on the same plan that previously let them run seven agents all day without issue. Aggressive auto-compaction is also being reported, possibly an intentional fix or a bug.</li>
<li><strong>Claude Code vs Cursor comparison:</strong> A hands-on Android development test gave Claude Code a clear win over Cursor on code quality, architectural adherence, and feature correctness, while using fewer tokens (343K vs 391K) despite slightly longer thinking time.</li>
<li><strong>Autoresearch benchmarked against hyperparameter optimization:</strong> Claude Code achieved only 0.0222 bits-per-byte improvement on a deep learning task versus 0.0293 for CMA-ES, while costing $27.21 in API fees. Classic algorithms won, and the author argues LLMs shouldn’t be applied where better specialized tools already exist.</li>
<li><strong>Docker Sandboxes open beta:</strong> Docker launched isolated microVM sandboxes for running AI agents in fully autonomous mode, supporting Claude Code, Codex, Copilot CLI, Gemini CLI, and others. Startup takes seconds, no shared state, no host access.</li>
<li><strong>Cloudflare Dynamic Workers open beta:</strong> V8 isolate-based runtime for executing AI-generated code, claiming 100x faster startup than containers. Priced at $0.002 per unique Worker loaded per day, waived during beta. Security analysts have flagged prompt injection and recursive execution risks.</li>
<li><strong>Cursor code review improvements:</strong> Selecting a diff and asking Cursor why it’s there, viewing images and videos of results, and testing straight from the browser are now supported, making in-editor review a more complete workflow than GitHub’s interface for some use cases.</li>
<li><strong>OpenCode /init redesign:</strong> The next release turns /init into a guided <a href="http://AGENTS.md" target="_blank" rel="noopener nofollow">AGENTS.md</a> setup that reads the repo and figures out what agents actually need to know, rather than requiring manual configuration.</li>
<li><strong>AI-generated merged code at 27.4%:</strong> DX’s Q1 2026 preview across 500+ organizations shows a rise from 22%, with daily AI users hitting 30.8%. The relative stability is attributed to teams not yet adapting workflows to newer models, not a capability plateau.</li>
<li><strong>VS Code ships weekly with AI:</strong> The VS Code team reports commit volume more than doubled and they’re closing nearly 3x as many issues year-over-year after shifting to weekly releases using parallel agent sessions, automated triage, and mandatory Copilot review before human review.</li>
<li><strong>MiniMax M2.5:</strong> China’s open-weights model reportedly approaches Claude Opus coding performance at significantly lower cost, with full open weights.</li>
<li><strong>Meta semi-formal reasoning prompting:</strong> Structured logical certificates that require models to explicitly state assumptions before drawing conclusions achieved up to 93% accuracy on patch verification tasks, outperforming standard agentic reasoning across all tested tasks.</li>
</ul>


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

Agentic Digest

Anthropic accidentally shipped a 57MB sourcemap in Claude Code v2.1.88, exposing up to 600,000 lines of TypeScript including an 11-step prompt pipeline, anti-distillation poison pills, an undercover mode, a frustration detector, and references to unreleased features like background agent Chyus/Chyros and internal model codenames. The community responded with rapid rewrites and forks, while Anthropic's DMCA takedowns caught legitimate open-source forks in the crossfire, drawing criticism. OpenAI reset Codex usage limits for the sixth time in March, with the team admitting they don't fully understand the demand spikes. Leaked references also point to two major unreleased Anthropic products: Conway (an always-on agent platform) and Epitaxy (a unified interface merging Claude Code with web-based task launching). Additional highlights include Docker's isolated microVM sandboxes for AI agents, Cloudflare Dynamic Workers open beta, DX data showing AI-generated merged code at 27.4%, and VS Code doubling commit volume after adopting weekly AI-assisted releases.

Claude Code source leak dominates the day, Codex hits usage limits for the sixth time