Terminal Coder That Runs on DeepSeek and a Shoestring

What it does

DAO CODE (command dao) is a terminal-native AI assistant that reads your codebase, writes patches, runs shell commands, and debugs failures without leaving the terminal. It targets DeepSeek V4’s 1M context window and gates destructive actions behind an allow/ask/deny approval system. The project is positioned as a fully open-source alternative to Claude Code, but it achieves utility through tight engineering rather than expensive model capacity.

The interesting bit

The entire architecture is obsessed with DeepSeek’s prefix cache: system prompts, tool tables, and memory are laid out in byte-stable order to push cache hit rates above 95% on real tasks, cutting per-session costs to roughly ¥0.15. Memory and reflection layers run as forked sub-agents that reuse the main cache prefix, so self-correction and cross-session recall add almost no token cost. For long autonomous tasks, it maintains shadow git checkpoints outside your actual .git directory and can crash-recover sessions with dao -c.

Key highlights

• Real SWE-bench-style evaluation: 13/14 tasks solved, with a 95.8% cache hit rate across 3.89M input tokens for a total cost of ¥1.07
• Claims ~30× cost savings versus Claude Opus 4.8 and ~18× versus Sonnet 4.6 when replaying identical token traces
• Self-verifying memory: cross-session facts are deterministically validated against live code on startup, and stale entries are garbage-collected
• Long-task robustness: shadow git checkpoints, automatic context compression near the 1M limit, periodic scope-correction, and crash recovery
• 24 built-in tools with allow/ask/deny permission tiers, MCP support, custom skills, and steering (typing queued while a round streams)

Verdict

Worth a look if you want a fully open-source, terminal-native coding agent that runs locally and optimizes aggressively for token economics on DeepSeek. Skip it if you are locked into Anthropic or OpenAI models and do not care about per-task costs.