AgentSkillsCN

orchestrate

当简单的查询无法满足需求时，可进行多轮上下文探索。当用户需要从多个文件中获取全面的上下文信息，或在实现跨越整个代码库的功能时，又或者仅靠 rlm_context_query 无法捕捉到重要的跨文件关联时，可使用此技能。

SKILL.md
--- frontmatter
name: orchestrate
description: Perform multi-round context exploration when a simple query isn't sufficient. Use this when the user needs comprehensive context from multiple files, when implementing features that span the codebase, or when rlm_context_query alone would miss important cross-file relationships.

When you need deep, comprehensive context across multiple files:

  1. •

    Use mcp__snipara__rlm_orchestrate with:

    • •query: The user's question or implementation task
    • •max_tokens: 8000-16000 depending on scope
    • •search_mode: "hybrid" (best results)
  2. •

    The tool performs 3 rounds automatically:

    • •Sections scan - Identifies all relevant sections across files
    • •Ranked search - Scores and ranks the best matches
    • •Raw load - Loads full content of top documents
  3. •

    For loading a single specific file, prefer mcp__snipara__rlm_load_document instead

  4. •

    For a project-wide structural overview, use mcp__snipara__rlm_load_project

  5. •

    To package context for REPL execution, use mcp__snipara__rlm_repl_context

When to choose orchestrate over context_query:

  • •The topic spans 3+ files
  • •You need full file content, not just excerpts
  • •The user is implementing a complex feature
  • •Previous context_query results were insufficient

When to choose context_query instead:

  • •Simple, focused questions
  • •Single-topic lookups
  • •When token budget should be small (< 6K)

Examples:

  • •"Implement OAuth integration" → rlm_orchestrate("OAuth flow implementation", max_tokens=12000)
  • •"How does the payment system work?" → rlm_orchestrate("payment system architecture", max_tokens=8000)
  • •"Refactor the database layer" → rlm_orchestrate("database layer structure and patterns", max_tokens=16000)