tigicion/dao-code

What it solves

Dao Code is a terminal-native AI coding assistant designed to provide high-capability software engineering automation while minimizing token costs. It specifically addresses the accessibility barriers of proprietary agents (like Claude Code) for users in certain regions and the high operational costs associated with large-context LLMs.

How it works

Built around the DeepSeek V4 model, the agent operates in a loop: it streams reasoning, requests tool calls (reading files, writing code, executing shell commands), and processes results through an approval gate. To keep costs extremely low, it employs "byte-stable" system prefixes and cache-reusing forks for reflection and memory, maximizing DeepSeek's prefix-cache hit rates. It also features a self-verifying memory system that prunes stale facts against live code at startup and an auto-compaction mechanism to handle context limits.

Who it’s for

Developers who want a powerful, terminal-based AI agent for bug fixing and feature implementation, particularly those seeking a cost-effective alternative to expensive closed-model agents or those who need a tool that is easily accessible without restrictive account requirements.

Highlights

• Extreme Cost Efficiency: Achieves high cache hit rates (up to 95.8% in tests) to significantly reduce input token costs.
• Long-Task Robustness: Includes shadow-git checkpoints for easy restoration, crash recovery, and a "Definition-of-Done" verification system.
• Advanced Memory: Cross-session memory that is deterministically verified against the current codebase upon startup to prevent hallucinations.
• Safety & Control: A layered permission system (allow/ask/deny) and a "circuit breaker" to detect when the agent is stuck in a loop.
• Extensibility: Supports MCP (Model Context Protocol), custom subagents, slash commands, and lifecycle hooks.