What problem does it solve? When switching mj-agent from the Ark cloud provider to a local OpenAI-compatible LLM endpoint on DGX-Spark (vLLM, SGLang, Ollama, TGI, or llama.cpp), misconfigured base URLs, mismatched model IDs, or unreachable hosts cause opaque runtime failures. This Skill verifies the endpoint before the agent runtime depends on it. ## Core Features & Use Cases - Three-step health probe: checks LLM_BASE_URL presence, queries /v1/models (with automatic Ollama /api/tags fallback), and runs a 1-token chat completion smoke test. - Model ID validation: compares the endpoint's loaded model list against LLM_MODEL_ID in .env and reports mismatches with corrective actions. - Actionable troubleshooting: maps failure modes (connection refused, 404, 401, 422, 500) to concrete fixes such as port checks, base URL path corrections, and API key alignment. - Use Case: After switching LLM_PROVIDER to local-openai-compat, run the probe to confirm the DGX vLLM endpoint at 192.168.0.189:8000 is reachable, serves the expected model, and answers a minimal chat request before launching LangGraph Studio. ## Quick Start Ask the agent to run a DGX endpoint healthcheck to verify the local LLM endpoint is reachable and serving the configured model.
