AgentSkillsCN

RLM (Recursive Language Model) Skill

RLM（递归语言模型）技能

SKILL.md

RLM (Recursive Language Model) Skill

When to Use RLM

Activate RLM processing when ANY of the following apply:

  1. •Context exceeds threshold: Total input > 100K tokens (~400K chars)
  2. •Multi-document reasoning: Task requires synthesizing 5+ documents
  3. •Aggregation tasks: Questions about "all", "every", "count of", "list all pairs"
  4. •Information-dense queries: Answer depends on most/all of the input (not needle-in-haystack)
  5. •Explicit request: User asks to "analyze this codebase", "process these logs"

Subcall Model Recommendations (Advisory)

The MCP server does not enforce model selection. These are recommendations:

Root ModelSubcall ModelBulk/Map ModelUse Case
Opus 4.5Sonnet 4.5Haiku 4.5Complex synthesis, deep analysis
Sonnet 4.5Haiku 4.5Haiku 4.5Standard workflows
Haiku 4.5Haiku 4.5Haiku 4.5Cost-sensitive, simple aggregation

Critical: Haiku is for bulk passes only. Never invoke Haiku per-line — batch chunks.

Workflow Pattern

  1. •Initialize: rlm.session.create with appropriate config
  2. •Load: rlm.docs.load documents
  3. •Probe: rlm.docs.peek at structure (first lines, format detection)
  4. •Search: rlm.search.query to find relevant sections (lazy-builds BM25)
  5. •Chunk: rlm.chunk.create with appropriate strategy
  6. •Process: rlm.span.get + client subcalls on spans
  7. •Store: rlm.artifact.store results with span provenance
  8. •Synthesize: Aggregate artifacts into final answer
  9. •Close: rlm.session.close

Chunking Strategy Selection

Content TypeStrategyParamsRationale
Source codedelimiter"\ndef |\nclass "Preserve semantic units
Logslinesline_count: 100, overlap: 10Temporal locality
Markdowndelimiter"\n## "Section boundaries
JSON/JSONLlinesline_count: 1Record-level processing
Plain textfixedchunk_size: 50000, overlap: 500Balanced chunks

Cost Guardrails

  • •Max tool calls per session: 500 (default), warn at 400
  • •Max chars per response: 50K (server-enforced)
  • •Max chars per peek: 10K (server-enforced)
  • •Batch aggressively: Prefer 10 docs per subcall over 1 doc per subcall
  • •Cache reuse: Check artifacts before re-querying same span
  • •Use span provenance: Every artifact should trace back to its source span

Anti-patterns to Avoid

❌ One subcall per line (Qwen3-Coder's failure mode — thousands of calls) ❌ Loading entire context into single subcall ❌ Ignoring cached artifacts from prior analysis ❌ Returning raw subcall outputs without synthesis ❌ Forgetting to close session ❌ Ignoring truncated: true in responses