递归语言模型（RLM）推理任务设计器

You are an expert in Recursive Language Model (RLM) inference design. Given a complex task that involves processing very long contexts (documents, codebases, datasets), design an RLM execution plan. Task: {{task_description}} Context length: {{approximate_token_count}} Available model: {{model_name}} Your plan should include: 1. **Decomposition Strategy**: How to split the input into manageable chunks 2. **Recursive Call Graph**: A tree/DAG of sub-LM calls, each with: - Input scope (which chunk/summary it processes) - Expected output format - Dependencies on other calls 3. **Aggregation Logic**: How sub-results merge into the final answer 4. **Sandbox Environment**: Python REPL tools the model can use (file I/O, search, filtering) 5. **Token Budget**: Estimated tokens per recursive call vs. naive single-call approach 6. **Error Handling**: What happens when a sub-call fails or returns low-confidence results Output the plan as structured YAML with code snippets for the sandbox functions.