Statistics reference
Open the right inspector with ⌥⌘I. The Stats section at the top describes the current chat's inference, not saved memories or your Mac's RAM. Model, Engine, and MCP details sit underneath. The MCP section lists connected servers and their model-callable tool counts. Per-call results, errors and timings appear in Activity (⌃`), keeping the inspector focused on current connections.
Context and speed indicators stay hidden until the first message is sent, then fade in. Reduce Motion disables the fade. The speed dial sits first, followed by its graph and context pressure, without a nested card.
Stats describe the currently open conversation, including saved responses from earlier app launches. Live marks an active chat turn; the caption distinguishes generation, engine waiting and tool work. The speed label identifies the latest measured response in that chat; it is not an average for the app session. Extra measurement explanations are available on hover.
Throughput
The native tachometer refreshes four times per second from a rolling one-second estimate. Its needle falls toward zero during pauses. The labelled scale grows when necessary and stays fixed for the rest of that response. The dial retains a numeric readout and an accessible speed value; ~ identifies estimates. Reduce Motion disables needle animation.
During generation, the line and shaded area show estimated tokens per second in one-second intervals, using streamed text, reasoning and generated tool names/arguments. Initial engine waiting is labelled separately and does not fill the graph with artificial zero-speed samples. Dips during actual generation can include server/network stalls. This is a client estimate, not a hardware profiler.
The trace keeps at most 60 samples and starts fresh for each response. Sampling stops when the inspector closes or GOAT becomes inactive. Reopening during a response starts a fresh timing baseline. A finished response shows the engine's decode rate when supplied, otherwise a rate derived from output tokens and decode duration. First-token latency, output tokens, and total request duration appear below it.
While waiting for the engine or working through a tool step, the dial retains the last measured response and labels it accordingly. A long first-token wait can coexist with a healthy final decode rate. The live interval estimate can differ from the final whole-response rate. That is expected.
Context ring
The ring shows the same context pressure as the composer: the current request's token usage against its available budget. During generation the budget can reserve space for the answer. The text identifies an input limit when applicable, and a trimming notice appears if GOAT removed older prompt material to fit.
~ means estimated usage or capacity. ? means the engine or saved chat does not supply enough information yet. An unknown value is not zero. Run another response to populate missing saved-chat usage. This is a request budget indicator, not a promise that all earlier messages remain in context.
Recent responses
Bars compare up to 12 recent responses with usable generation statistics, newest on the right. Click or drag over a bar to inspect its rate. The legend distinguishes engine-reported from derived rates. Different models, prompts, and output settings can change speed, so this is a history view, not a controlled benchmark. Saved chats reuse whatever generation statistics were persisted; missing measurements are not invented.
All charts run natively. They add no network calls or telemetry files. Live traces are temporary; existing saved response statistics follow normal chat persistence.