Reasonix is a terminal coding agent built only for DeepSeek models. Its loop is designed around DeepSeek's prefix cache: every layer keeps the prompt prefix byte-stable across a long session so cached tokens stay cacheable and costs stay low. One reported single-day workload (435M input tokens) hit a 99.82% cache rate at about $12, against roughly $61 for the same workload uncached on v4-flash. The architecture is organized around three pillars: a cache-first loop, tool-call repair, and cost control.
The TypeScript 0.x line is in maintenance mode, receiving bug fixes only, and active development moved to a Go rewrite on the main-v2 branch, which ships as a single binary with four entry points: terminal, desktop app, local web UI (reasonix serve), and editors over the Agent Client Protocol (ACP). Plan mode, permissions, a workspace sandbox, and per-turn checkpoints are meant to keep a long autonomous run readable and reversible.
reasonix code is the default mode and the only one with filesystem and shell tools: the assistant proposes SEARCH/REPLACE edits and nothing touches disk until /apply. reasonix chat is a tools-off mode with MCP servers and web search attached, and reasonix run "task" streams a one-shot result to stdout. Configuration lives in a JSON file with per-project overrides. Skills are Markdown playbooks that run inline or as an isolated subagent, and Claude-format skills load alongside them, so tooling such as OpenSpec that emits them works without an adapter. A Tauri desktop client (prerelease) and a QQ channel that relays messages into a running session are also available. Stated non-goals: multi-provider support, IDE integration, and air-gapped or free use (a paid DeepSeek API key is required, and Aider with Ollama or Continue are suggested for those cases).
Features
- Cache-first loop: prompt prefix kept byte-stable so DeepSeek prefix-cache hits persist across long sessions
- Reviewed edits: SEARCH/REPLACE proposals with a cell-diff renderer, applied only on /apply, with automatic checkpoints
- MCP servers: stdio, SSE, and Streamable HTTP servers from config.json or --mcp
- Hooks and permissions: shell commands on PreToolUse (gating), PostToolUse, UserPromptSubmit, and Stop, plus a per-workspace exact-prefix shell allowlist
- Memory: user-private user, feedback, project, and reference entries pinned into the prefix
- Web search: Bing by default, switchable to Baidu AI Search, SearXNG, Metaso, Tavily, Perplexity, Exa, Brave, or Ollama with /search-engine
- Semantic index: reasonix index builds a local code index with Ollama or any OpenAI-compatible embedding endpoint
- Session tooling: persistent per-workspace sessions, transcript replay, diff, events, stats, an /effort knob, and an embedded web dashboard
