Annoying and alarming things about OpenCode

Original Article ↗ Hacker News Discussion ↗

Criticism of the popular OpenCode AI coding harness centers on security vulnerabilities, heavy resource use, brittle prompt caching, and opinionated system prompts that can silently alter code style or even strip comments. Commenters argue that these flaws reflect broader problems with current “agentic” LLM tools, which often run untrusted code with weak sandboxing and misleading permission systems, creating serious supply-chain and privacy risks. While many still find OpenCode highly productive—especially due to its free or cheap access to models—others are migrating to alternatives like Pi, Codex, or custom harnesses, often combined with stronger external sandboxes or local-only models.

Overall reception of OpenCode

  • Many agree the post is hyperbolic but see its core criticisms as mostly accurate.
  • Several users report OpenCode made them very productive and remains their favorite harness.
  • Others have already migrated (or now plan to) to alternatives like Pi / OhMyPi, Codex, Kilo, Mimo, Maki, Aider, etc.

Security, permissions, and sandboxing

  • Multiple serious security issues and RCEs are noted; some are reportedly fixed, others unclear.
  • The textual command-filter / “allowlist” is widely viewed as weak or misleading from a security standpoint.
  • Strong consensus: do not trust any coding agent on your real filesystem; run it inside a sandbox/VM (bwrap, sandbox-exec, flatpak, landlock, etc.).
  • Debate over whether sandboxing should be integrated into harnesses vs handled by separate tools.

Prompt cache, compaction, and performance

  • Frequent cache misses due to system prompt mutations (date, AGENTS.md changes, etc.) are a major annoyance and token sink.
  • Compaction/pruning is seen as slow, buggy, and often counterproductive; some disable it via env vars.
  • Others say backend caching (e.g., DeepSeek, vLLM) can mask client inefficiencies.
  • OpenCode developers mention that problematic pruning is disabled by default and that v2 changes aim to avoid cache busting.

System prompts, comments, and LSP

  • Default prompts are criticized as huge, messy, and enforcing dubious policies (e.g., “no comments”), which causes unwanted comment deletion.
  • Some agree with minimal-comments output; others explicitly instruct agents to add many comments and fight the defaults.
  • LSP integration divides opinion: some find it a killer feature for refactors and symbol lookups; others see little benefit and high token cost.

Governance, UX, and project health

  • Large backlog of open issues, aggressive stale bot, and sparse PR acceptance fuel claims that the repo is “open source in name only.”
  • Complaints about bloat, high CPU/RAM usage, and a confusing new UI (tabs, lost workspace support).
  • Maintainers respond that v2 addresses several raised issues but acknowledge GitHub issue noise.

Broader view on LLM agents and tone

  • Many note the same structural problems apply to most agentic CLIs, not just OpenCode.
  • Some see the article as effectively anti-AI-for-SWE; others frame LLMs as “just tools” that require realistic expectations and strong isolation.
  • The aggressive, mocking tone of the article is polarizing; some enjoy the rant, others find it unfair or demoralizing.