How is the Bun rewrite in Rust going?

Original Article ↗ Hacker News Discussion ↗

Bun’s high‑profile rewrite from Zig to Rust using Anthropic’s Claude model is prompting questions about what “done” really means for an AI‑assisted port of a large runtime. Commenters note that while the Rust version already powers Anthropic’s Claude Code editor and passes extensive tests, there’s still no stable public release, ongoing heavy CI activity, and a large amount of unsafe Rust, all of which suggest significant post‑rewrite hardening work. The thread broadens into a debate over whether such LLM‑driven rewrites are cost‑effective and trustworthy, or mainly a marketing stunt that obscures the long, human‑led effort required to make the result idiomatic, secure, and maintainable.

Status of the Rust rewrite

  • Thread agrees the Zig→Rust translation is functionally done: the Zig code is no longer used and the Rust version runs in production for Claude Code and as a canary (bun upgrade --canary).
  • No formal 1.4 release yet. Some see the lack of a tagged release as evidence the rewrite isn’t “complete” or stable; others argue a long canary period is normal after a major rewrite.
  • A maintainer says the public release is blocked on hitting a promised number of newly passing Node.js tests; release is “most likely” imminent.

Real‑world usage and reliability

  • Claude Code has been running on the Rust runtime for over a month; most users apparently did not notice.
  • Some reports of Bun‑related crashes and memory leaks in Claude Code; others say it’s been stable enough that the rewrite “hasn’t crashed and burned.”
  • Debate over how representative Claude Code is: some argue it uses a large surface of Node/Bun APIs; others say it’s just one app and not proof of broad ecosystem stability.

Safety, “idiomatic Rust”, and tech debt

  • Many comments worry the port is unidiomatic Rust with heavy unsafe and C‑like boundaries, so initial safety isn’t much better than Zig.
  • A rough count of unsafe in the repo appears flat over time, disappointing those who expected progressive reduction via refactoring.
  • Others counter that some unsafe is inherently required (FFI to JavaScriptCore, OS APIs) and that Rust plus tools (linting, miri, tests) make future safety work easier.

Cost, tokens, and CI

  • Initial marketing figure was ~$165k in tokens over ~11 days; commenters estimate total token + CI spend approaching ~$800k as work continued.
  • Some say that’s trivial for a frontier AI company; others note it’s large and potentially misleading as a model for “normal” organizations.
  • CI costs (many platforms/architectures and heavy sharding) are acknowledged as significant and long‑standing, not just rewrite‑driven.

LLMs for large rewrites: hype vs reality

  • Enthusiasts see this as a landmark: a highly complex runtime ported quickly while keeping a large test suite passing; they report similar success using LLMs for ports.
  • Skeptics emphasize long‑term maintenance: LLMs are good at fast mechanical translation, less so at architecture, edge cases, and de‑slopping; the “last 10%” of bugs and bloat is expensive.
  • Concerns raised about shipping large amounts of non‑human‑reviewed LLM code to millions of users, and about Anthropic’s incentives and transparency; others reply that strong tests, Rust’s type system, and dogfooding significantly mitigate risk.