nodebench-mcp 2.18.1 → 2.18.2

package/README.md

@@ -205,11 +205,15 @@ Research → Risk → Implement → Test (3 layers) → Eval → Gate → Learn
 
 ### Recommended Workflow: Start with Default
 
-The `default` preset includes:
+The `default` preset includes 50 tools in 3 groups:
 
-1. **Discovery tools** — 6 tools: `findTools`, `getMethodology`, `check_mcp_setup`, `discover_tools`, `get_tool_quick_ref`, `get_workflow_chain`
-2. **Core methodology** — 38 tools: verification, eval, quality_gate, learning, flywheel, recon, security, boilerplate
-3. **Self-escalate** — Add `--preset full` when you need vision, web, files, or parallel agents
+1. **Discovery tools (6)** — *"What tool should I use?"* — `findTools`, `getMethodology`, `check_mcp_setup`, `discover_tools`, `get_tool_quick_ref`, `get_workflow_chain`. These help agents find the right tool via keyword search, 14-strategy hybrid search, workflow chains, and methodology guides.
+
+2. **Dynamic loading tools (6)** — *"Add/remove tools from my session"* — `load_toolset`, `unload_toolset`, `list_available_toolsets`, `call_loaded_tool`, `smart_select_tools`, `get_ab_test_report`. These let agents manage their own context budget by loading toolsets on demand and unloading them when done.
+
+3. **Core methodology (38)** — *"Do the work"* — verification (8), eval (6), quality_gate (4), learning (4), flywheel (4), recon (7), security (3), boilerplate (2). These are the AI Flywheel tools that enforce structured research, risk assessment, 3-layer testing, quality gates, and persistent knowledge.
+
+**Self-escalate**: Add `--preset full` when you need vision, web, files, or parallel agents.
 
 This approach minimizes token overhead while ensuring agents have access to the complete methodology when needed.
 

package/dist/index.js

@@ -1308,7 +1308,7 @@ Use NodeBench tools when you need to:
 - Audit SEO, analyze Figma flows, detect Android flicker
 - Call LLMs (GPT, Claude, Gemini) for analysis and extraction
 Start with discover_tools("<your task>") to find the right tool.`;
-const server = new Server({ name: "nodebench-mcp-methodology", version: "2.18.1" }, {
+const server = new Server({ name: "nodebench-mcp-methodology", version: "2.18.2" }, {
     capabilities: { tools: { listChanged: true }, prompts: {} },
     instructions: SERVER_INSTRUCTIONS,
 });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "nodebench-mcp",
-  "version": "2.18.1",
+  "version": "2.18.2",
   "description": "Make AI agents catch the bugs they normally ship. 175 MCP tools across 34 domains: AI Flywheel methodology (6-phase verification + eval-driven development), structured research, 3-layer testing, quality gates, persistent knowledge, LLM calling, security analysis, visual regression, academic writing, local file parsing, flicker detection, Figma flow analysis, contract compliance scoring, TOON encoding (~40% token savings), pattern mining, git workflow compliance, SEO audit, voice bridge, agent mailbox, progressive discovery with 8-mode hybrid search + neural embeddings (local HuggingFace / Google / OpenAI fallback), model-tier complexity routing, agent contract (front-door + anti-rationalization), session memory, 6 GAIA media solvers, adaptive architect (structural code analysis + concept verification + implementation planning), email (SMTP/IMAP send+read+draft+setup wizard), RSS/Atom feed monitoring with research digest pipeline scaffolding, Agent-as-a-Graph bipartite embedding search (arxiv:2511.18194) with type-conditioned wRRF, MCP annotations (category/phase/complexity), critter intentionality checkpoint. Default: 39 tools (complete AI Flywheel, ~76% fewer). Full: 175 tools.",
   "type": "module",
   "bin": {