mtop: matoleo yote na matoleo ya zamani

mtop now tells you why a local model is slow, or why its answers went wrong.
Why is it slow?
- When Ollama can't fit a model in VRAM it quietly runs part of it on the CPU. The models pane now turns that row amber and shows cpu 38% next to the TTL. Lemonade models running on CPU show up too.
Why is the answer off?
- A prompt that blew the context window gets a ctx 97% or ctx over, rejected marker in the requests pane, with details in the inspector.
- llama.cpp, vLLM and SGLang rejections are read from the error the server sends back, so that one's exact.
- Ollama trims prompts silently, so anything at 90% of the window or more is flagged as likely truncated.
Reasoning models
- Thinking tokens count toward tok/s on OpenAI-style streams (reasoning_content and reasoning).
- The inspector shows the thinking text separately from the answer, for Ollama, OpenAI-style and /v1/responses traffic.
More runners, more GPUs
- llama-swap, Lemonade and SGLang, with u to unload on llama-swap and Lemonade.
- AMD GPUs through amd-smi, with rocm-smi as the fallback.
- AMD and Intel GPUs on Windows, read from the system performance counters (temp and power show n/a there).
Also
- Two new /metrics series: mtop_model_cpu_percent and mtop_ctx_overflow_total. -notify fires for both warnings.
- The models pane now shows name, VRAM and TTL first, so the markers still fit a 100-column terminal.
- Moved to Bubble Tea v2.
- Captured prompt and answer text (-inspect) stays in memory and is no longer written to the history file.
- Fixes: an old Lemonade on :8000 no longer shows as vLLM, llama-swap no longer appears twice, and the by-model view doesn't reshuffle equal rows.
brew and scoop pick this up shortly; winget follows once Microsoft merges the bump.

Catches up with what the runners changed over the summer.
Fixed
- vLLM 0.12 renamed its kv-cache metric, so the pane sat at 0%. Both names are read now.
- Anything answering on a probed port (a dev server on :8000, say) no longer shows up as a phantom vllm or llama.cpp model.
- Streamed OpenAI-style requests without a usage block now get recorded. llama.cpp's own timings are used when the stream carries them, otherwise tok/s is estimated from the chunk count and marked as such. The bytes your client gets are still untouched.
- nvidia-smi [N/A] fields show as n/a instead of a fake 0, and stay out of the energy total.
- A rocm-smi response mtop can't read surfaces as an error instead of a blank GPU pane.
New
- The proxy records /v1/responses traffic, streamed or not.
- llama.cpp router mode (llama-server started without -m) lists every loaded model.
- LM Studio 0.4's v1 REST API, with a fallback for older installs.
- A CHANGELOG, backfilled to 1.0.
brew and scoop pick this up as usual; winget follows once Microsoft merges the bump.

Watch more than one box, see more per request, and a couple of new numbers.
- Multi-host: give -ollama a comma list and mtop stacks the models and GPUs from each machine, tagged by host. Handy if you run models on a couple of boxes.
- GPU util and memory now draw as sparklines over time, next to the live numbers.
- Request inspector: run with -inspect, press i, and you get the last request's prompt, completion, and a load/prompt/decode timing split. Off by default; the text it captures is stripped of control bytes so a model can't smuggle escape sequences into your terminal.
- Session energy on the TOK/S line: watt-hours used and tokens per watt-hour. It's whole-GPU power, so read it as a rough efficiency number.
- compare -openai <url> runs the comparison against llama.cpp, LM Studio or vLLM, not just ollama.
- -mem-alert and -temp-alert to set the alert thresholds instead of the built-in 93% and 87C.
brew and scoop pick this up as usual; winget follows once Microsoft merges the bump.