What problem does it solve? When a local multi-agent LLM pipeline misbehaves on constrained CPU-only hardware, operators need a safe way to find the root cause without accidentally changing protected production settings. This Skill runs a structured, read-only diagnostic pass over the SwarmX/Ollama stack and returns an incident-style report. ## Core Features & Use Cases - Ten-point health check: Verifies API health, Ollama reachability, /api/ps model residency, available RAM, pressure tier, active 7B model, OLLAMA_NUM_PARALLEL, model registry consistency, recent stage failures, and OpenClaw provider configuration. - Invariant protection: Enforces hard stops so diagnostics never mutate protected values such as RAM_CRITICAL_MB=800, MAX_CONCURRENT_JOBS=1, OLLAMA_NUM_PARALLEL=1, the SINGLE-7B lock, or production promotion gates. - Incident-style reporting: Returns observed values, likely cause, a safe next action, and a rollback/containment path. - Use Case: A video generation stage fails on a 16 GB CPU-only host. Run the diagnostic to confirm whether memory pressure evicted the resident model or whether the OpenClaw provider is misconfigured, then follow the recommended safe remediation. ## Quick Start Ask the assistant to run the swarmx-doctor diagnostic and report why the current pipeline stage is failing.
