tangledgroup Dirac v0.2.89
Overview
Dirac is an open-source AI coding agent that achieves 64.8% cost reduction compared to competing agents (Cline, Kilo, Opencode, Roo) while maintaining 100% accuracy on complex real-world refactoring tasks across repositories like HuggingFace Transformers, VS Code, and Django. It topped the Terminal-Bench-2 leaderboard with a 65.2% score using gemini-3-flash-preview.
Its core innovation: instead of search-and-replace edits that force the LLM to repeat unchanged code (O(S+R) output tokens), Dirac uses hash-anchored edits that reduce output to only the replacement text (O(R)). Combined with AST-native tools via tree-sitter, multi-file batching, and tight context curation, this makes large-scale refactoring tractable in single tasks.
Dirac is a fork of Cline and does not use MCP — it implements its own tool protocol directly.
When to Use
- Deploying Dirac as a coding agent (CLI or VS Code extension)
- Understanding the hash-anchored edit engine architecture
- Integrating with Dirac's tool system or building custom tools
- Configuring subagents, worktrees, or checkpoints for safe parallel editing
- Debugging anchor reconciliation or AST-based refactoring workflows
Core Concepts
Hash-Anchored Edits: Each line in a file is assigned a stable word anchor (e.g., Moderator§def process():). When the LLM proposes an edit, it references anchors instead of repeating full search text. The AnchorStateManager uses FNV-1a content hashing internally and Myers Diff to reconcile anchors after edits — unchanged lines keep their exact anchor, only new/changed lines get fresh ones. This eliminates cascading invalidation.
AST-Native Precision: Dirac embeds tree-sitter parsers for TypeScript, Python, C++, Go, Rust, Java, Ruby, PHP, and more. Tools like get_file_skeleton, get_function, find_symbol_references, replace_symbol, and rename_symbol operate on the AST rather than raw text, enabling structural refactoring with 100% syntactic accuracy.
Multi-File Batching: All edit_file blocks in a single LLM turn are grouped by file path, validated in memory, and applied atomically. Diagnostics run in parallel across all modified files after save. This reduces roundtrip latency and API costs significantly.
High-Bandwidth Context: Dirac only reads what the LLM needs. File skeletons (AST definition lines with anchors) replace full file reads. The ContextLoader extracts file paths, directories, and symbol names from user prompts to auto-enrich context. Conversation condensation (/smol) truncates history when needed.
Advanced Topics
Architecture Overview: Codebase modules, core flow (extension → webview → controller → task), CLI vs VS Code extension → Architecture Overview
Hash-Anchored Edits Deep Dive: AnchorStateManager internals, FNV-1a hashing, Myers Diff reconciliation, single-token word anchors, § delimiter, validator, reconciler lifecycle → Hash-Anchored Edits
AST-Native Tools: tree-sitter integration, ASTAnchorBridge, file skeleton extraction, get_function, find_symbol_references, replace_symbol, rename_symbol, symbol index service → AST-Native Tools
Tool System: All 28 tools (DiracDefaultTool enum), handler patterns, approval workflow, auto-approval, read-only vs write tools → Tool System
Context Curation: ContextLoader, context condensation, symbol enrichment, high-bandwidth strategy → Context Curation
Subagents, Worktrees & Checkpoints: Parallel subagent spawning, git worktree isolation, checkpoint-based diff tracking → Subagents, Worktrees & Checkpoints
CLI Reference: Commands, flags, slash commands, authentication, environment variables, piping → CLI Reference
VS Code Extension: Sidebar, settings, Dirac Rules, browser integration, theme support → VS Code Extension