DAO CODE is a terminal-native AI coding assistant for DeepSeek V4, prioritizing cost-efficiency, user experience, and availability. It enables in-terminal code interaction, robust long task handling, and self-verification. Targeting developers, it maximizes the potential of cost-effective models through advanced engineering.

How It Works

Operating with DeepSeek V4's 1M context, DAO CODE's core innovation is sophisticated caching engineering. It achieves high prefix cache hit rates (95.8% on SWE-bench) by maintaining system prefix byte stability and implementing reflection/memory layers that reuse the main cache, drastically reducing costs. Long tasks benefit from automatic context compression, session logging, crash recovery, and shadow Git checkpoints for stability. A constitutional priority system guides actions, prioritizing safety.

Quick Start & Requirements

• Installation: One-click script (curl ... | sh), downloadable binaries, or npm (npm i -g dao-code).
• Prerequisites: DeepSeek API key (configurable via env vars or guided setup), Node.js >= 20 for npm installs.
• Documentation: README, /help command.

Highlighted Details

• Cost-Effectiveness: Achieves significant savings (~30x vs. Claude Opus) via advanced caching, with SWE-bench tasks costing as little as ¥1.07.
• Cache Engineering: Leverages DeepSeek's prefix cache with byte-stable prefixes and fork-based cache reuse, automatically compressing context near the 1M limit.
• Long Task Robustness: Features session logging, crash recovery, shadow Git checkpoints, verification-driven completion (/dod), and deadlock detection.
• Extensibility: Supports custom agents, commands, skills, MCP, hooks, and layered permissions, compatible with Claude Code configurations.

Maintenance & Community

Actively iterated upon (v0.1.20 released). Contribution guidelines are in CONTRIBUTING.md, community standards in CODE_OF_CONDUCT.md. Security issues should be reported privately. No specific community channels are listed.

Licensing & Compatibility

Released under the permissive MIT License, allowing commercial use. Compatible with Claude Code configuration formats for skills, memory, hooks, and MCP.

Limitations & Caveats

Under continuous iteration, suggesting evolving features. Relies on DeepSeek V4 models and API keys. Operational costs are tied to API usage. Advanced features may have a learning curve.