What went wrong, if anything, with the calls your Ollama client sent through LLMxRay (port 11435), over several turns: prompts rejected for exceeding the loaded context, a prompt silently cut to fit it, the prompt cache lost on every turn or on one turn, a model running partly on CPU, tool schemas crowding the prompt. Call it after a few turns, and again after you changed something. Use last_turn_metrics instead when you want one turn in numbers rather than a verdict; use local_models instead when the question is which model or context length to pick. Returns text: one line per client seen (turns, tokens, tool calls, prefill time), then "What stands out (N):" followed by one finding per line, each tagged [high], [medium] or [low] and ending with the fix. When nothing stands out it says so. When no turns were seen it explains how to route calls through port 11435. Parameters: client narrows to one tool when several share the machine (substring of the label LLMxRay shows); since_minutes limits the window, so pass it after a change to judge only the turns since. Both optional. Needs LLMxRay running (npx llmxray) on the same machine; if it is not, the answer says so. Reads only; changes nothing. Cost: well under a second, from records LLMxRay already keeps; no model is called.

The most recent turn seen through LLMxRay, as numbers, no judgement: prompt tokens and how many came from cache, prefill time and speed, generated tokens and speed, why generation stopped, what the prompt was made of (system, tools, history, new message, in characters), the context requested, whether the model fit in video memory, and where the prompt stopped matching the previous turn. Use it to budget the next call or to confirm a change helped. Use diagnose_my_turns instead for findings across several turns with fixes; use local_models for what is installed and loaded. Returns text, one fact per line, starting with "turn at client: ... model: ...". When no turns were seen it explains how to route calls through port 11435. Parameter: client picks the latest turn of one tool when several share the machine (substring of the label); omit it for the latest turn from anyone. Needs LLMxRay running (npx llmxray) on the same machine; if it is not, the answer says so. Reads only; changes nothing. Cost: well under a second, from records LLMxRay already keeps; no model is called.

Which models Ollama has pulled on this machine (size on disk, parameters, quantisation, family), which are loaded right now and with what context length, how much of each loaded model sits in video memory, and whether each pulled model would fit in video memory. Use it before choosing a model or a context length. It says nothing about past calls: for those use diagnose_my_turns (findings) or last_turn_metrics (numbers). Returns text: a first line with the GPU, video memory and RAM when known, then "N models pulled:" and one line per model, sorted by name, ending with LOADED now ... or fits in VRAM ... or larger than VRAM: will run partly on CPU. Takes no arguments. Needs Ollama reachable; the machine line needs LLMxRay running too and is simply omitted when it is not. Reads only; changes nothing. Cost: well under a second, two listing calls to Ollama; no model is loaded or called. The answer is a snapshot: a model shown as loaded unloads after its keep_alive, so call again right before deciding.