dirac  by dirac-run

Summary Dirac is an open-source AI coding agent focused on efficiency and context curation, targeting engineers and power users. It drastically cuts API costs (50-80%) and enhances code quality through advanced techniques like Hash Anchored Edits and AST manipulation, making complex refactoring tasks more tractable and cost-effective.

How It Works Dirac utilizes Hash Anchored Edits for precise modification targeting via stable line hashes, avoiding traditional line-numbering issues. Its AST-Native Precision enables accurate structural code manipulation across languages (TypeScript, Python, C++). Multi-File Batching processes multiple files in a single LLM call, significantly reducing latency and API costs. Optimized context curation ensures efficient delivery of relevant information to the LLM.

Quick Start & Requirements Installation: VS Code Extension (Marketplace) or npm CLI (npm install -g dirac-cli). Requires models with native tool calling. Authentication via dirac auth or environment variables (Anthropic, OpenAI, Gemini, etc.). AWS Bedrock integration supported via AWS credentials. Common commands: dirac "prompt", dirac -p "prompt" (Plan Mode), dirac -y "prompt" (Yolo Mode), git diff | dirac "prompt".

Highlighted Details

• Terminal-Bench 2.0 Leaderboard: Ranked first (65.2%) with gemini-3-flash-preview, outperforming Google's baseline and Junie CLI.
• Cost Efficiency: Claims 64.8% average cost reduction.
• Hash-Anchored Edits: Precise targeting via stable line hashes.
• AST-Native Precision: Accurate structural code modifications.
• Multi-File Batching: Single LLM roundtrip for multiple files.

Maintenance & Community A fork of the Cline project, developed by Max Trivedi at Dirac Delta Labs. No specific community channels or roadmap links were detailed in the provided README snippet.

Licensing & Compatibility Licensed under the permissive Apache License 2.0, generally suitable for commercial use.

Limitations & Caveats Exclusively supports models with native tool calling. A minor bug affecting previous cost reporting was noted and a fix submitted. No specific hardware or OS limitations beyond standard development environments were specified.