swarmxq-model-orchestrator

Call resolveCanonicalTag() on the operator's canonical tag, then call ModelOrchestrator.evictIncompatible(tag) before ollamaPull(). This guarantees only one 7B-class model is inference-active at any time, satisfying the SINGLE-7B LOCK.

The SINGLE-7B LOCK permits only one 7B-class model to be inference-active simultaneously because CPU-only hardware has exactly one inference thread. On 16 GB hosts a second model like Pilot may stay RAM-resident but idle, never actively inferencing.

getRamSnapshot() reads MemAvailable from /proc/meminfo, and readPressure() classifies it as normal, high, or critical. High pressure triggers a backoff delay of 1000-30000 ms; at or below 800 MB (RAM_CRITICAL_MB) the system fails immediately with PRESSURE_CRITICAL.

No. Even though OLLAMA_MAX_LOADED_MODELS=2 allows Pilot (~3 GB) and one 7B model (~5 GB) to be RAM-resident together, only one may actively inference. Two concurrent 7B inferences violate the SINGLE-7B contract regardless of available RAM.

Legacy tags like phi4-fast, deepseek-reasoner, and qwen-worker must pass through resolveCanonicalTag(), which maps them to canonical tags and throws on unknown values. Hard-coded aliases bypass validation and desynchronize the TypeScript and Python operator maps.

shouldAutoEnableLowRamMode() returns true below the full-pipeline threshold, switching stages to the lite model instruct-phi4-lite-q4km-prod, setting maxLoadedModels to 1, disabling Pilot keep-alive, and reducing frames and resolution. Degradation must be logged at warn level.