Skip to content

ENTRY.md

Latest commit

 

History

History
92 lines (74 loc) · 4.57 KB

ENTRY.md

File metadata and controls

92 lines (74 loc) · 4.57 KB
name wiki-vault
description Knowledge vault MCP server + companion skills for structured wiki compilation. Use this skill whenever you need to search, query, lint, or compile knowledge into the vault. Provides 7 MCP tools for ranked search, graph traversal, and batch validation, plus 8 skills (2 tactics + 6 SOPs) for knowledge compilation workflows.

Wiki Vault

Lightweight knowledge vault with structured graph, BM25 search, and batch validation. The vault stores research knowledge as interconnected markdown pages following the Karpathy llm-wiki pattern: Sources (immutable raw material) → Wiki (LLM-maintained synthesis) → Schema (configuration).

Philosophy

  • Single unified vault. All knowledge lives in one vault regardless of research domain. Cross-domain connections emerge naturally.
  • CC handles CRUD directly. File creation, editing, and deletion are CC's native operations. The MCP server provides only what CC cannot do efficiently: ranked search, graph traversal, batch validation, and index maintenance.
  • Obsidian-compatible. Pages use standard markdown with YAML frontmatter and [[wikilinks]]. The vault is browsable in Obsidian without modification.
  • Graph is first-class. Every page should connect to the knowledge graph via typed edges. Orphans are failures.

Four-Level Hierarchy

ENTRY.md (this file)
  → Tactic (2): multi-step orchestration patterns
    → SOP (6): single-responsibility operations
      → Tool (7): atomic MCP operations

Skill Routing

Signal Skill
编译研究成果、compile findings、ingest material → knowledge-compilation (tactic)
vault 健康检查、lint、orphan cleanup、merge duplicates → vault-maintenance (tactic)
搜索 vault、find pages、check duplicates → wiki-search (SOP)
图谱查询、explore neighborhood、find connections → wiki-graph-query (SOP)
添加关系、create edge、link pages → wiki-add-edge (SOP)
导入源材料、ingest source、capture raw material → wiki-ingest-source (SOP)
创建/更新页面、compile page、synthesize → wiki-compile-page (SOP)
运行 lint、fix issues、validate vault → wiki-lint-fix (SOP)

MCP Tools (7)

Tool Purpose
vault_search BM25 full-text search with type/tag filters
vault_add_edge Create typed edge (validates paths, rejects duplicates)
vault_query_graph BFS traversal from node (direction, depth, edge_type filter)
vault_graph_stats Global or per-node graph statistics
vault_lint Batch validation with optional auto-fix
vault_index Rebuild BM25 index (incremental or full)
vault_info Returns vault metadata (root, directories, types, stats) for session init

Entity Types (9)

Type Directory Purpose
source sources/ Immutable raw material (papers, web content, data)
concept concepts/ Synthesized understanding of a topic
entity entities/ Named entities (people, orgs, tools, datasets)
claim claims/ Testable assertions with confidence scores
relation relations/ Named relationships between concepts
question questions/ Open research questions
evidence evidence/ Supporting/refuting evidence with confidence
failure failures/ Documented failed approaches
topic topics/ High-level topic containers

Edge Types (10)

Edge Type Semantics
component_of A is part of B
instance_of A is an example of B
supported_by A has evidence from B
contradicts A conflicts with B
supersedes A replaces B
derived_from A was built from B
addresses A answers/solves B
raises A creates/implies B
failed_for A didn't work for B
related_to Weak association

Read/Write Rules

  • CC writes files directly. Page creation, editing, deletion — all via CC's native file tools.
  • MCP for queries only. Search, graph traversal, stats, lint — operations that need index/graph state.
  • Sources are immutable. Once created, source pages are never edited.
  • Wiki pages evolve. Updated as understanding deepens, confidence changes, new evidence arrives.
  • Always search before creating. Deduplication is a pre-condition for page creation.

Context-Management Integration

  • Campaign start: context-init (load/create campaign context file)
  • After each strategy completes: context-checkpoint + knowledge-compilation
  • Vault state persists across sessions via the file system — no special serialization needed