Serious multi-agent orchestration backend with LangGraph, OpenRouter, local RAG tool, and MCP integration
This server has 6 tools with basic descriptions but significant structural gaps. Tool names follow verb_noun convention (rag_search, mcp_call, build_project_context, search_tasks, search_knowledge, evidence_merge), which is good. However, descriptions are minimal (10-60 chars), parameter descriptions are absent or trivial, and input schemas lack proper type definitions and constraints. No output schemas are documented. Error handling is not visible in the provided code. The server acts as an orchestrator calling external MCP tools, but the tool definitions themselves are underdeveloped for LLM reasoning.
Build project context from external MCP tool
Merge and curate evidence from multiple tool execution records
Use MCP tool for external capability execution
Retrieve relevant local knowledge for grounded response generation.
Search knowledge from external MCP tool
Search tasks from external MCP tool
Parameter descriptions missing or trivial. 'query' param in rag_search has description 'Search query for local knowledge retrieval' (50 chars), but no guidance on format, length, or examples. 'tool_name' in mcp_call lacks any description of valid values or how to discover them.
No output schemas documented. Tools return results but LLMs cannot plan downstream calls without knowing what fields to expect. E.g., does rag_search return {results: [{id, text, score}]} or {documents: [...]}? Does search_tasks return pagination info?
Input schemas lack proper type constraints. 'limit' in search_tasks is typed as integer but has no min/max bounds. LLMs could pass limit=999999 or limit=-1. No enum constraints on 'tool_name' in mcp_call, LLMs must guess valid tool names.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | <=2025-11-25 | v2 |
Tool descriptions are too brief (30-60 chars). Baseline for A+ tools is 194 chars. Current descriptions lack WHEN to use the tool, WHAT it returns, and dependencies. E.g., 'Use MCP tool for external capability execution' does not explain when mcp_call is preferred over direct tool invocation.
No error handling guidance visible. If rag_search returns no results, or mcp_call fails because tool_name is invalid, what should the LLM do? No recovery hints, retryability classification, or actionable error messages.