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

Original Article ↗ Hacker News Discussion ↗

A new terminal-based coding agent called DeepSeek Reasonix aims to exploit DeepSeek V4’s “byte-stable” prefix cache to deliver very high cache hit rates, sharply reducing token costs for long coding sessions. Commenters compare it to existing harnesses like Claude Code, OpenCode, Pi, and Crush, debating whether its cache-first, append-only design is genuinely novel or just repackaging practices many mature agents already use. The conversation broadens into trade-offs between cheap open(-weight) Chinese models vs Western SOTA systems, privacy and espionage concerns, and frustration with AI-generated UIs and the proliferation of yet another bespoke agent tool.

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.