Cline Recursive Chain-of-Thought System (CRCT)

CRCT is a large-scale project context management and task orchestration framework designed for VS Code's Cline or RooCode extensions (current version v8.3.0, by Richard Brandes, Apache 2.0 licensed). Its core mechanism recursively decomposes large tasks into subtasks, persisting project state, dependency relationships, and context information as Markdown files within the VS Code workspace, enabling minimal context loading—only essential data enters the LLM context window.

For dependency tracking, CRCT employs a modular hierarchical architecture: a master tracker (module_relationship_tracker.md, doc_tracker.md) manages global dependencies while module-level mini-trackers ({module_name}_module.md) handle local details. Contextual Keys (KeyInfo) and hierarchical keys enable precise dependency localization, with RLE compression controlling data scale. v8.2+ introduces local LLM dependency resolution via llama-cpp-python for batch-validating unconfirmed dependencies, plus determine-dependency for detailed pairwise relationship analysis.

The Symbol Essence Strings (SES) embedding architecture introduced in v8.0+ combines runtime and AST metadata to generate semantic representations of code symbols, paired with Qwen3 Reranker for AI semantic dependency scoring. The system automatically switches between GGUF-formatted Qwen3-4B (GPU mode) and SentenceTransformer (CPU mode). The dual-token mode simultaneously tracks ses_tokens (embedding) and full_tokens (total file) for optimized context window allocation.

Code analysis covers Python (AST) and JavaScript/TypeScript/HTML/CSS (Tree-sitter), detecting TODOs/FIXMEs, empty functions, unused variables, and integrating Pyright. Dependency visualization generates project overview, module-focus, and multi-key-focus diagrams via Mermaid DSL with SVG rendering (mermaid-cli + layout-elk). Visualization performs well under 1000 edges, degrades at 1500+, and times out at 4000+.

The workflow follows four phases—Set-up/Maintenance, Strategy, Execution, Cleanup/Consolidation—controlled through the .clinerules plugin system. HDTA templates provide implementation plan, module, task, and roadmap summary templates. Caching uses TrackerBatchCollector with atomic writes/rollback, mtime-based invalidation, SHA256 stable hashing, and Pickle persistence.

System Requirements

• Recommended (v8.0+): 8GB+ NVIDIA GPU VRAM, 16GB+ RAM, 8GB+ disk
• Minimum (CPU-only): 4GB RAM, 500MB+ disk

Installation

git clone https://github.com/RPG-fan/Cline-Recursive-Chain-of-Thought-System-CRCT-.git
cd Cline-Recursive-Chain-of-Thought-System-CRCT-
pip install -r requirements.txt
npm install  # For mermaid-cli visualization

Note: The system does not auto-install cuda/mps Python dependencies; GPU users must manually install the appropriate PyTorch version.

Configuration & Startup

1. Open the project in VS Code with Cline or RooCode extension installed
2. Copy contents of cline_docs/prompts/core_prompt(put this in Custom Instructions).md to Cline's Custom Instructions field
3. Type Start. in the Cline input box to initialize

Key CLI Commands

• show-keys: Identify keys needing attention (p placeholder, s suggestion, S suggested)
• show-dependencies: Aggregate dependency details for specified keys (inbound/outbound, paths)
• add-dependency: Resolve placeholder or suggestion dependencies
• resolve-placeholders (v8.2+): python -m cline_utils.dependency_system.dependency_processor resolve-placeholders --tracker <path>
• determine-dependency (v8.2+): python -m cline_utils.dependency_system.dependency_processor determine-dependency --source-key <key> --target-key <key>
• visualize-dependencies: Generate Mermaid dependency diagrams
• analyze-project: Full project analysis, auto-generating trackers and diagrams

Config File: .clinerules.config.json — configurable embedding device (cpu/cuda/mps), file exclusion patterns, etc.

Unconfirmed: No independent website, associated academic paper, or dedicated HuggingFace page found; RooCode compatibility details are sparse; SES embedding claims 10x precision improvement without benchmark data.