Claude Code CLI was broken

Original Article ↗ Hacker News Discussion ↗

Bug, Cause, and Workarounds

  • CLI crashed because the tool parsed CHANGELOG.md as structured data and broke when version headers started including dates (e.g. 2.1.0 (2026-01-07)), which semver couldn’t compare.
  • Users shared a sed-based hot patch that wraps all version-string comparisons in semver.coerce() inside cli.js.
  • A simpler workaround: replace the cached changelog with a minimal # Changelog file and make it read‑only, preventing bad parsing.
  • The issue is reported as fixed via a GitHub PR, but many note how surprising it is that a changelog formatting tweak could break the entire CLI.

Testing, Quality, and “Vibe Coding”

  • Many criticize the lack of even basic integration tests that would have caught “CLI fails to start.”
  • There’s broader concern that Claude Code is “vibe coded”: changelog treated as a data source, permissions logic behaving inconsistently, and prior unresolved concurrency issues that maintainers called “unlikely to ever be fully fixed.”
  • Some see this as evidence of poor engineering culture at a company heavily promoting AI-written code; others argue the bug is just human error and not inherently an indictment of LLM use.

AI-Written Code, Velocity Claims, and Skepticism

  • A referenced claim that Claude wrote essentially all recent Claude Code changes, with huge PR/LOC counts, kicks off debate about:
    • Whether such velocity can be meaningfully reviewed.
    • LOC as a useless metric, especially with AI churn.
    • Whether “agentic” workflows can legitimately parallelize enough to justify those numbers.
  • Some users say Claude Code is both: highly productive and clearly “vibe coded,” implying vibe coding can still yield valuable tools; others are unconvinced.

Alternatives and Ecosystem

  • Opencode is discussed as an alternative: open source, works with subscriptions and APIs, supports multiple providers, but mixed reports on planning quality vs native tools.
  • Users note large open issue/PR counts in both Opencode and Claude Code, which some see as a red flag.
  • There’s a brief mention of a temporary period where rate limits seemed disabled, suggesting usage tracking also broke.

Permissions, Security, and UX

  • Several comments describe inconsistent or leaky permission behavior (file access, command execution, webfetch), leading some to isolate Claude Code in containers/jails.
  • Users want deterministic, non-AI-mediated permission enforcement and clearer UX when commands are run.