| name | rlm-utils |
| description | Recursive Language Models utilities for partitioning large datasets and aggregating results from sub-agents in OdooClaw. |
Skill: RLM Utilities (Recursive Language Models)
This skill provides essential tools for implementing the Recursive Language Models (RLM) reasoning strategy within OdooClaw. It allows the agent to handle large datasets by partitioning them into smaller, manageable chunks and aggregating results from sub-agents.
The strategy follows a context-centric decomposition philosophy: keep the root context small, recursively process subsets, and then consolidate.
Capabilities
1. Data Partitioning (rlm_partition)
Splits a large list of Odoo records (from search_read) into multiple JSON files stored in the workspace.
- Goal: Avoid context window overflow and context rot.
- Workflow:
search_read -> rlm_partition -> spawn/subagent for each chunk.
2. Result Aggregation (rlm_aggregate)
Combines results from multiple files or sub-agent outputs into a single report.
- Goal: Consolidate findings from the "Map" phase of recursion.
- Workflow: Collect results from sub-agents ->
rlm_aggregate -> Final User Response.
Usage Guide (RLM Strategy)
When faced with a query requiring deep analysis of many records (e.g., "Analyze the last 100 invoices for patterns"):
- Decompose: Use
odoo-mcp tools to fetch records, then immediately use rlm_partition if the list is large.
- REPL Variable Store: Treats
odooclaw/workspace/tmp/rlm/ as a variable store.
- Recursive Processing: Launch sub-agents to process each file path returned by
rlm_partition.
- Reduce: Summarize the individual results using
rlm_aggregate or a final reasoning step.
RLM patterns used in OdooClaw
- Peek/Grep first: Before full recursion, narrow scope with targeted domains/filters in
odoo-mcp queries.
- Partition + Map: Split large result sets into chunks and process each chunk in isolated sub-agents.
- Reduce: Consolidate chunk-level outputs into one final user answer.
- Context hygiene: Keep long raw data in workspace files, not in the main chat context.
Rules
- Always use absolute paths provided by the tools.
- Clean up temporary files if they are no longer needed (future).
- Prefer smaller chunks (e.g., 5-10 records) for maximum precision in sub-agents.
- Prefer deterministic chunk sizes for reproducibility (e.g., 10, 20, 50).
