Dspy Rlm

DSPy RLM (Recursive Language Model)

Load this optional Skill only when the request is specifically about explaining or implementing dspy.RLM; ordinary RLM Turns do not need it.

1. load_skill(...) returns a dictionary; read its Skill body from load_result["skill_markdown"]. Read references/rlm-contract.md with read_skill_resource when you need constructor limits, built-ins, or Fleet mapping. That call also returns a dictionary: extract contract = resource_result["content"] instead of slicing the result.
2. Treat dspy.RLM as a Recursive Language Model: a sandboxed REPL code agent. Never call it a Retrieval Language Model or redefine it as RAG / dspy.Retrieve / ReAct.
3. Prefer Python standard-library computation, parsing, search, and aggregation for deterministic work. Use llm_query for one bounded semantic judgment and llm_query_batched for multiple independent semantic judgments with self-contained prompts.
4. Ground any semantic prompts in the extracted contract string. Do not invent DSPy RLM APIs from training priors.
5. For long writeups when Session Workspace is available, follow report-builder / workspace-files: write the full report durably, then issue exactly one typed SUBMIT with every active Signature output.
6. For nontrivial deterministic work, verify in a later iteration using an independent invariant, known reference, higher-precision stability, or a genuinely independent formulation before submitting.

Authority: supported DSPy 3.3.x and https://dspy.ai/api/modules/RLM/ — not Daytona docs as DSPy authority. The current lock resolves 3.3.1.