DeepSeek Ships a Web-Based Agent Runner, and pnpm 12 Hits RC in Rust
DeepSeek Harness signals that Chinese AI labs are now shipping agent tooling built on web-first, plugin-based architectures rather than copying terminal-centric workflows. The pnpm 12 RC means teams can adopt a faster Rust-based package manager without touching their lockfiles or CI scripts.
DeepSeek released Harness, an open-source agent framework that competes directly with Claude Code but swaps the CLI for a web UI. Every feature ships as a plugin on top of Cordis, an existing Node plugin system whose original author DeepSeek has now hired. The project pulled in 30,000 GitHub stars on launch day.
pnpm 12 is now in RC, rewritten in Rust while preserving full backward compatibility with pnpm 11 commands, arguments, config, and lockfile format. Only four documented differences exist between versions. Separately, a 12-year-old weak PRNG in crypto-js versions before 4.0.0 has been exploited to steal wallet keys.
React Native 0.87 converts its public JS API to strict-mode TypeScript with types generated from source, and Deno 2.9.5 adds an experimental QuickJS backend for compiled binaries. On the tooling side, vlt 1.0 launches as an npm replacement with 60-plus CSS-like selectors for querying dependency graphs, half of them security-focused.
DeepSeek hiring Cordis's original author to build Harness's plugin layer is a concrete signal that agent frameworks are being treated as serious platform investments, not experiments.
pnpm's Rust rewrite with full backward compatibility sets a high bar for migration strategy: the team prioritized not breaking existing workflows over chasing new features in the RC.
A 12-year-old PRNG bug in crypto-js that only now led to wallet thefts illustrates how cryptographic weaknesses in widely depended-upon libraries can lie dormant until the incentive to exploit them becomes large enough.
vlt's CSS-selector approach to dependency queries is a genuinely novel interface for package management, turning what is normally a CLI flag memorization problem into a composable query language.