rlm

name: rlm description: "Hub skill for fleet-rlm: when to use dspy.RLM vs ReAct/CodeAct and which workflow skill to load next."

fleet-rlm RLM Hub

Canonical DSPy module: https://dspy.ai/api/modules/RLM/

Module variants overview: https://dspy.ai/diving-deeper/built-in-module-variants/

When to use dspy.RLM

Use RLM when context is too large, unevenly relevant, or benefits from programmatic exploration — not a single full-context LM call.

fleet-rlm routes to dspy.RLM when:

• execution_mode is rlm / rlm_only
• Auto mode detects URL document analysis (url_document_rlm)
• Auto mode detects large context (large_context_rlm, threshold via FLEET_RLM_LARGE_CONTEXT_THRESHOLD)

Other turns are classified by the typed RouteTurnSignature router: direct (ChainOfThought), tools (dspy.ReAct loop), or rlm (sandboxed Python).

REPL contract (Daytona interpreter)

• Large fields are REPL variables; explore with print() and Python.
• Built-ins: llm_query, llm_query_batched, SUBMIT(...).
• Fleet extensions: sub_rlm, sub_rlm_batched, volume load_skill(name).
• Durable storage: /home/daytona/memory/ (skills/system/*.md seeded from scaffold).

Route to a workflow skill

1. Sandbox / REPL / SUBMIT / volume lifecycle → sandbox-execution
2. Recursive child sandboxes / budgets → delegation
3. Signatures / modules / execution modes → dspy-programs
4. Large documents / variable mode → long-context
5. GEPA / offline optimization → optimization
6. Failures / contract drift → diagnostics
7. Volume layout / CRUD → volume-bootstrap
8. Playwright / rendered pages → browser-interaction