DeepSeek reasonix, DeepSeek native coding agent with high caching and low cost

Original Article ↗ Hacker News Discussion ↗

What Reasonix Is and Scope

  • Terminal-based DeepSeek-focused coding agent, claiming “append-only” loops tuned for DeepSeek’s prefix cache to achieve 90%+ cache hits and large cost reductions.
  • Project is explicitly described as independent and not affiliated with DeepSeek, though some commenters are skeptical about that independence.

Caching Behavior & Design

  • Core idea: keep prompts byte-identical in the prefix; avoid timestamps, reordering, tool list changes, or prompt rewrites that break cache.
  • Several developers explain that many existing frameworks (esp. older LangChain-style) historically rebuilt prompts each turn, reducing cache hits.
  • Others argue most modern agents (Claude Code, Codex, Pi, OpenCode, etc.) already avoid obvious cache killers and achieve high hit rates.
  • Debate over tradeoffs: strict append-only to maximize cache vs. occasional compaction/pruning that may hurt cache but improves model quality and UX.
  • Clarifications that cache usually works on shared prefixes, so changing late parts of context or tools lists can still preserve substantial caching.

Comparisons to Other Agents & Harnesses

  • Some say Reasonix’s benefits largely duplicate what OpenCode, Pi, Codex, Crush, and others already do with DeepSeek’s API.
  • Multiple users report 95–98% cache hit rates and low spend using DeepSeek via other harnesses (OpenCode, Claude Code, Pi).
  • Others complain about cache “stability” issues or many open bugs in existing tools, but this is contested.

Model Quality, Cost, and Workflows

  • DeepSeek V4 Pro/Flash widely praised as “cheap and good,” often compared favorably to Claude Sonnet and as somewhat below Claude Opus or GPT‑5.5 in difficult, long-horizon tasks.
  • Common pattern: use a stronger model (Opus/GPT‑5.5) for planning or complex debugging, then DeepSeek for implementation to cut costs.
  • Some report DeepSeek often ignores constraints or declares tasks finished prematurely in complex scenarios; others find it reliable for everyday coding.

Security, Privacy, and Vendor Trust

  • Concerns raised about using Chinese-hosted models due to legal obligations to cooperate with state intelligence; counterpoints reference US/Five Eyes surveillance and argue all big providers pose risks.
  • Consensus: for truly sensitive code or personal data, self-hosted open-weight models are preferred, and DeepSeek’s open weights are seen as a plus.

UX and Implementation Critiques

  • Website widely criticized as “AI-designed slop”: animated hero text causing layout shifts, poor mobile layout, generic purple/gradient aesthetic.
  • TUI described as flickery and slow when typing; dark-theme bias makes it hard to use in light terminals.
  • Some dislike yet another JS/TS-based CLI and wish for lean Rust/Go single binaries; others point to existing Go agents (e.g., Crush) for that niche.

Unclear / Open Questions

  • Unclear how much Reasonix actually outperforms well-tuned existing harnesses in real-world cache savings.
  • Open question whether advanced behaviors like mid-call compaction or more sophisticated multi-file context management are implemented or planned.