DeepSeek-Reasonix: DeepSeek-native terminal AI coding agent

My notes

Summary

Reasonix is an MIT-licensed, terminal-first AI coding agent built exclusively for DeepSeek, with its entire loop engineered around DeepSeek’s byte-stable prefix-cache mechanic to keep token costs low across long sessions. It is a direct alternative to Claude Code / Cursor / Aider, trading multi-provider flexibility for deep single-backend cost efficiency.

Key Insight

• Cache stability as an architectural invariant, not a toggle. The loop is designed so the prompt prefix stays byte-stable, maximizing DeepSeek’s prefix-cache hits. This is why it is DeepSeek-only by design - coupling to one backend is the feature.
• Hard cost number: one real user’s single day (01-05-2026) logged 435M input tokens at 99.82% cache hit, costing ~$12 vs ~$61 with no cache. ~5x savings purely from cache discipline.
• Three architecture pillars: (1) cache-first loop, (2) tool-call repair, (3) cost control. These solve problems generic agent frameworks “don’t even see” because they target a different cache mechanic.
• Drop-in Claude compatibility: reads .claude/skills/<name>/SKILL.md and ~/.claude/skills/ alongside its native paths, so Claude-format skills (e.g. OpenSpec workflows) load without an adapter.
• Two modes: code (filesystem + shell tools, SEARCH/REPLACE review, plan mode, project memory) vs chat (tools-off thinking partner with MCP, global memory only).
• Honest non-goals: for hardest-leaderboard reasoning / PhD proofs they recommend Claude Opus; for air-gapped/zero-cost they point to Aider + Ollama. DeepSeek positioned as competitive on coding, cheap per task.
• Requires Node >= 22; needs a paid DeepSeek API key. Desktop is a Tauri GUI prerelease (unsigned installers).