convez.ai
MCP ServerAutomaçãoAvançado
fwornle

MCP Server Semantic Analysis

fwornle/mcp-server-semantic-analysis

Análise semântica de histórico de código com 14 agentes especializados

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP que analisa histórico de código (.specstory files, git history) e gera documentação resumida para sistemas de gestão do conhecimento. Orquestra 14 agentes especializados (8 com LLM, 2 infraestrutura) que extraem, classificam e persistem insights sobre a evolução arquitetural do código.

Instalação

$npx convez-ai fwornle/mcp-server-semantic-analysis

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 1 mês
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Gerar documentação automática de decisões arquiteturais a partir do histórico git
  • Classificar e organizar conhecimento de código em grafo ontológico
  • Deduplicar e validar conteúdo com embeddings semânticos e busca web

Veredito

Ferramenta robusta para times que precisam capturar conhecimento implícito do código. Arquitetura sofisticada com failover de providers LLM e integração forte com km-core, mas exige experiência em orquestração de agentes e gestão de workflows — não é plug-and-play.

README

A tradução para o português deste README ainda está na fila.

mcp-server-semantic-analysis

MCP server driving the wave-analysis workflow, ingest, and ontology classification for the coding knowledge graph — 14 agents (8 LLM-enhanced, 2 infrastructure, orchestration + persistence) writing through @fwornle/km-core.

Configurations Owned

  • Ontology: — (owned by coding — this server READS .data/ontologies/coding-ontology.json for classification, but does not own it)
  • LLM providers: lib/llm/ (6-tier provider chain: Groq → Gemini → Custom → Anthropic → OpenAI → Ollama) + config/agents/*.json (per-agent prompts) + config/workflows/*.json (wave-analysis processOverrides routing — see scripts/configure-wave-analysis-routing.sh)
  • Ingest adapters: src/agents/*.ts — each agent ingests via PersistenceAgent, which writes through @fwornle/km-core's REST router (POST /api/v1/entities)
  • Domain dedup: src/agents/dedup-agent.ts — OpenAI text-embedding-3-small + cosine similarity overlay on top of km-core's baseline LayeredDeduplicator (with Jaccard text-similarity fallback)

Architecture

mcp-server-semantic-analysis architecture

This server exposes 12 MCP tools over stdio (local) or HTTP/SSE on port 3848 (Docker mode). The CoordinatorAgent orchestrates the 14-agent pipeline via workflow definitions in config/workflows/*.json; WaveController drives the multi-wave analysis flow (extract → analyze → classify → persist → dedup → predict → merge). The 8 LLM-enhanced agents share a SemanticAnalyzer service with automatic provider failover; the 2 infrastructure agents handle embedding-based dedup and stale-entity detection. All persistence converges on PersistenceAgent, which writes through km-core's REST contract — mcp-server-semantic-analysis owns no graph storage itself.

14-Agent Catalog (names + 1-line roles):

  • LLM-Enhanced (8):
    1. GitHistoryAgent — git commits since checkpoint + architectural decisions
    2. VibeHistoryAgent — LLM session summaries + pattern analysis
    3. SemanticAnalysisAgent — deep code-vs-conversation correlation
    4. WebSearchAgent — 40% keyword + 60% semantic relevance scoring
    5. InsightGenerationAgent — LLM insight + PlantUML diagram generation
    6. ObservationGenerationAgent — structured observation extraction + domain classification
    7. QualityAssuranceAgent — semantic validation (fragments, generic content)
    8. PersistenceAgent — ontology classifier + canonical write path via km-core
  • Infrastructure (2): 9. DeduplicationAgent — embedding cosine similarity, Jaccard fallback 10. ContentValidationAgent — stale-entity detection + refresh recommendations
  • Orchestration: 11. CoordinatorAgent — workflow orchestration, {{step}} templating 12. WaveController — wave-analysis driver, sub-step events for dashboard progress

For per-agent enhancement detail, the 6-tier LLM provider chain, ontology classification internals, MCP tool reference, use cases, and project structure, see docs/AGENTS.md.

Where to Edit

To add…Edit…Verify
A new agentNew file src/agents/<NewAgent>.ts + register in src/agents/index.ts + add step in src/coordinator.ts workflownpm run build && npm test -- agents
Change LLM routing for a processconfig/workflows/<workflow>.json processOverrides blockRestart container + check /api/ukb/llm-mode on the dashboard
A new wave stepsrc/wave-controller.ts (step registration) + step implementationnpm test -- wave-controller
A new dedup rulesrc/agents/dedup-agent.ts (calculateNameSimilarity / fuzzy thresholds)npm test -- dedup

Related Systems

  • KM-Core — shared store + REST contracts this server writes through (/api/v1/ wire-shape lock)
  • coding — host runtime + observation source (obs-api at localhost:12436)
  • operational-knowledge-management — sister system consuming the same km-core core for RaaS / KPI-FW / business ontologies ("OKM" for short, external BMW GHE repo)

Tests / Verify

cd integrations/mcp-server-semantic-analysis
npm run build
npm test

Submodule build pipeline: This server is a git submodule of the coding parent repo. After any TypeScript source change, BOTH npm run build (inside this submodule) AND a Docker rebuild of coding-services are required for the change to take effect at runtime. See the Rebuilding After Code Changes section in the parent repo's CLAUDE.md for the full two-step recipe. Documentation-only changes do not require a rebuild.

Ver o original no GitHub

Bom para

Relacionados