菜单
Process documents and codebases exceeding a single context window using canonical dspy.RLM variable mode in the Daytona REPL.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Delegate recursive work to child RLM sandboxes with budget management. Use when decomposing tasks into sub-queries, fanning out batched work, managing LLM call budgets, or building parent-child RLM hierarchies.
Diagnose fleet-rlm runtime failures, API contract drift, sandbox errors, and observability issues. Use when something is broken — Daytona connection failures, websocket mismatches, escalation not triggering, budget exhaustion, or missing traces.
Optimize fleet-rlm DSPy programs and RLM skill bundles using GEPA with MLflow tracking. Use when running optimization loops, designing feedback metrics, building training datasets, or comparing runs.
Hub skill for fleet-rlm: when to use dspy.RLM vs ReAct/CodeAct and which workflow skill to load next.
Design DSPy signatures and compose runtime modules for fleet-rlm tasks. Use when creating input/output field definitions, choosing between built-in signatures, selecting execution modes, or wiring custom modules.
Fetch and inspect JavaScript-heavy pages with Playwright in a Daytona browser-capable snapshot.
| name | long-context |
| description | Process documents and codebases exceeding a single context window using canonical dspy.RLM variable mode in the Daytona REPL. |
Official references:
dspy.RLM stores large inputs as REPL variables (document_text, context_paths, history, …). The model sees only metadata (name, type, length, preview) and explores with Python:
print(document_text[:2000]) or print(len(document_text)) to peek.re, or open(path) on context_paths to locate relevant sections.llm_query(snippet) or llm_query_batched([...]) on focused excerpts — never the full document.SUBMIT(answer=...).execution_mode=auto routes to large_context_rlm when estimated context ≥ FLEET_RLM_LARGE_CONTEXT_THRESHOLD (default 32_000 chars).context_paths REPL variables with context_manifest metadata.sub_rlm(text) delegates to an isolated child sandbox for heavy map-reduce (see delegation skill).When semantic boundaries matter before delegation:
scripts/semantic_chunk.py — split by structure (markdown, logs, Python, JSON).scripts/rank_chunks.py — rank chunks against the query.Chunking complements dspy.RLM; it does not replace REPL inspection.
llm_query on an entire large variable; slice first.max_llm_calls and max_output_chars; print summaries, not raw dumps.load_skill("long-context") when mounted at /home/daytona/memory/.When the user asks for a verbatim quote or speaker attribution:
SUBMIT — not a numbered list of quotes.document_text with Python search, then slice the typographic quote span verbatim.context_paths in the sandbox.