What problem does it solve? Standard top-k retrieval fails when a task requires reading or aggregating over more material than fits in any model's context window, such as searching an entire repository or performing full literature-review aggregation. This Skill implements the Recursive Language Model (RLM) pattern, where the corpus lives as an external variable in a sandboxed kernel and the model programmatically decomposes and recursively delegates over it. ## Core Features & Use Cases - Host/Kernel Authority Boundary: Separates the process running model-generated Python from the host that owns database access, memory writes, and provider API keys, with tiered sandboxing from subprocess to gVisor. - Async Subagent Spawning: rlm_spawn() returns an admission handle immediately without blocking, and child results arrive later via agent_message.send(), preserving parallelism and crash isolation. - Host-Enforced Depth and Budgets: Recursion depth defaults to 1 and is enforced on the host side so a compromised kernel cannot raise its own budget. - Use Case: A legal-discovery-style task asks whether any file in a large repository performs a specific action. The router detects the corpus exceeds the model's context budget, spawns child agents over corpus partitions, and aggregates their messages into a structured answer. ## Quick Start Ask the agent to determine whether any file in this repository performs a specific action by recursively delegating over the full corpus using the RLM orchestrator.
