Loading ToolBench
Handle 10M+ token contexts with chunking, sub-queries, and local Ollama inference
Server has 10 tools with consistent naming (rlm_* prefix) and descriptions present for all. However, parameter descriptions are minimal or missing for several tools. Input schemas are visible and properly typed, but output schemas are not documented. Error handling is basic. Tool composition is reasonable but some tools combine multiple concerns (e.g., rlm_sub_query_batch handles chunking, querying, and parallel execution). No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels declared in metadata.
Check Claude SDK availability and authentication
Check code firewall status and configuration
Check Ollama availability and system requirements
Split a loaded context into manageable chunks for processing
Delete a loaded context and its chunks
Execute Python code with optional firewall security checks
Retrieve a loaded context by ID
List all loaded contexts
Output schemas not documented. Tools return data but LLMs cannot predict field structure, forcing parsing guesses and downstream tool chaining failures.
rlm_sub_query_batch combines chunking, querying, and parallel execution in one tool. Should split into separate tools (chunk → query → parallelize) for agent composability.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite metadata declaring risk levels (READ_ONLY, DESTRUCTIVE, IRREVERSIBLE). Agents cannot infer safety properties.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 52 | 2026-07-28+ | v2 |
Load external context (file, URL, or raw text) into the RLM system for processing
Process multiple chunks in parallel with sub-LLM calls
rlm_execute_code accepts arbitrary Python code with optional firewall checks. No dry-run or confirmation step before execution. Agents can accidentally execute destructive code.
Parameter descriptions are minimal. E.g., 'provider' in rlm_sub_query_batch lacks explanation of when to use 'ollama' vs 'claude-sdk'. 'parallel_count' lacks guidance on valid range or performance implications.