Massive Context MCP

# Project Instructions ## Large Context Handling with RLM This project uses the RLM (Recursive Language Model) MCP server for handling contexts that exceed normal token limits. ### When to Use RLM Tools Autonomously Claude should **automatically** reach for RLM tools when encountering: 1. **Large files** - Any file >100KB or >2000 lines 2. **Multi-file analysis** - Tasks requiring synthesis across multiple documents 3. **Context overflow** - When combined context would exceed ~180K tokens 4. **Deep codebases** - Analyzing entire repositories or large modules 5. **Iterative refinement** - Tasks requiring multiple passes over the same content ### RLM Pattern The standard flow is: **load → chunk → sub-query → aggregate** ``` 1. Load large context into RLM memory (keeps it OUT of main window) 2. Inspect/chunk the context into processable pieces 3. Sub-query each chunk with focused questions (parallel when possible) 4. Store results and aggregate into final answer ``` ### Recognizing RLM-Appropriate Tasks **Use RLM when user says:** - "Analyze this entire codebase/repository" - "Summarize all these files" - "Find patterns across these documents" - "Review all test files for X" - "Extract all instances of Y from this large file" **Use RLM when you observe:** - File read returns >2000 lines - User provides multiple file paths to analyze together - Task requires multiple analytical passes - Context budget is getting tight ### Available RLM Tools **Core workflow:** - `rlm_load_context` - Load large content into RLM memory (returns metadata only) - `rlm_inspect_context` - Preview structure without loading full content - `rlm_chunk_context` - Split by lines/chars/paragraphs - `rlm_get_chunk` - Retrieve specific chunk for processing - `rlm_list_contexts` - See all loaded contexts **Advanced operations:** - `rlm_filter_context` - Regex-based filtering to create focused subsets - `rlm_sub_query` - Send focused question to chunk via sub-LLM - `rlm_sub_query_batch` - Process multiple chunks in parallel (respects concurrency limits) - `rlm_store_result` - Save sub-query results for aggregation - `rlm_get_results` - Retrieve stored results for final synthesis ### Example Usage Pattern ``` User: "Analyze all Python files in src/ for security issues" Claude's approach: 1. Read file list, identify total size 2. Load combined content with rlm_load_context 3. Chunk by files or logical sections 4. Use rlm_sub_query_batch to check each chunk for security patterns 5. Store results with rlm_store_result 6. Aggregate with rlm_get_results 7. Present unified findings ``` ### Key Principle **Preserve main context window** - Don't load massive files into the main conversation. Use RLM to process them externally and return only the insights.