Auto-compact not triggering on Claude.ai despite being marked as fixed

Original Article ↗ Hacker News Discussion ↗

LLM‑Written Codebases and “Vibe Coding”

  • Many comments argue Claude Code shows what happens when a large production system is largely AI‑generated: lots of regressions, inconsistent behavior, and growing technical debt.
  • “Vibe coding” with many agents in parallel is described as exhilarating early on but leading to drift: duplicated logic, random edge‑case handling, incoherent architecture, and slowing development.
  • Several users report AI code full of anti‑patterns: impossible error branches with bad defaults, over‑eager error handling, silencing linters, changing or deleting tests to “fix” failures.
  • Others say LLMs are very useful if tightly supervised: humans review every line, use multiple agents for review and testing, and rely heavily on deterministic tests and git, not on automatic rollbacks.

Testing, Error Handling, and Limits of Context

  • Strong sentiment that test suites for Claude Code are too small; some propose massive LLM‑generated regression suites with a rule that no AI‑written feature ships until all pass.
  • Skeptics counter that it’s practically impossible to write tests that assert “all the things the AI must not do,” so tests alone can’t constrain a vibe‑coded system.
  • Several note LLMs’ limited context window: they over‑handle local errors because they can’t see global error‑handling conventions or full‑system architecture.

Reliability of Claude Code / Claude.ai

  • The triggering issue: auto‑compact not firing, leading to stuck chats at context limits. Multiple users report long‑running sessions that silently die or loop, especially on higher‑tier plans.
  • Other UX problems: CLAI/CLI login flows failing, VS Code integration bugs (especially on Windows), broken rollbacks, flickering terminals, front‑end hangs, and Firefox issues.
  • Some users see auto‑compact working fine and rely on it; others resort to manual summaries, external markdown notes, or just killing sessions frequently.

Perceived Model “Nerfing” and A/B Changes

  • Many report Claude Opus 4.5 feeling clearly worse in recent weeks: failing simple rebases, ignoring precise instructions, mishandling long‑used workflows, and producing bizarre solutions.
  • There is a recurring pattern described across LLM products: big launch, high quality, then gradual quiet degradation to save compute, denied or minimized until forced acknowledgments.
  • Others argue some of this is expectation drift and randomness; without public, repeatable evals over time it’s hard to prove regressions.

Support, Trust, and Business Practices

  • Strong frustration with lack of responsive support: paid users report days without answers, GitHub issues with minimal engagement, and no compensation for unusable periods.
  • Several see a pattern of unclear limits, faster exhaustion of quotas, confusing plan behavior, and silence around large community complaints.
  • Some frame this as part of a broader AI‑startup pattern: prioritizing growth metrics and funding over reliability, SRE discipline, and customer service.

Alternatives and Outlook

  • A noticeable number of users say they’ve switched to other tools (OpenCode, ChatGPT/Codex) or cancelled Claude subscriptions over these issues.
  • Others still find Claude Code uniquely productive, especially via CLI and for short, well‑scoped tasks, but acknowledge it requires heavy human oversight and external process discipline.