Runs

Runs draws one tool's calls to the model, end to end, as a waterfall: where the time went, call by call, and what each call sent and got back.

Sidebar item: Runs (below Live) Route: /runsAdded in: v0.8.7 (October 2026) Requires: calls on the Live page. Point a tool at http://localhost:11435 and let it work. To see what each call sent and got back, switch on Keep the prompt text on the Live page first.

What's the point?

An agent on a local model is slow, and the question is always where. In the model? Loading it? Reading a prompt it already read? In the tools between calls? Live shows every call from every tool as it happens; Runs takes one tool's calls and lays them out in time, so the slow part stands out.

What a run is

A run is one tool's calls that follow each other within two minutes. A longer pause starts a new run. Inside a run, a step starts wherever the tool began a fresh conversation: with the text kept, that is where the next call stops carrying the previous conversation; without it, where a prompt worth keeping found the cache empty while the model stayed loaded.

Reading the waterfall

One row per model call, placed in time. Each call's bar is split into what Ollama reports:

• Waiting: time the call spent before Ollama worked on it, plus anything Ollama did not account for.
• Loading the model: the model was not in memory. Often the biggest single cost of a run.
• Reading the prompt: the part of the prompt that was not in the cache. The cached part costs almost nothing, so it does not show as time; the share is written in the Operation column.
• Writing the answer.

Between two calls, a tools row is the time from the model's answer to the tool's next call, named after the tools the model asked for. An agent row is the same gap when no tool was asked for: the tool's own work. Tool time is measured from outside; nothing inside a tool is traced.

Flags in red say what went wrong on a call: model loaded, cache empty, rejected (the prompt was larger than the loaded context), error.

What each row opens

Click a row and the panel on the right shows it.

• A model call: its numbers and phases, then Input (what was sent: the conversation, the tool schemas, with the point where the prompt stopped matching the previous call) and Output (what came back: the answer, the model's thinking when it shows it, and each tool call with its arguments). Both have a Pretty and a Raw view.
• A tools row: each tool the model asked for, with its arguments and the result the tool sent back in its next call.

Every row has its own address, for example /runs?run=…&row=call-3. Copy link puts it on the clipboard; opening it lands on that row. From a call, Open in Prompt X-ray shows what its prompt was made of, and the X-ray links back.

What it cannot see

• Inside the tools. A database query in the middle of a tool is part of the tools row, not a row of its own.
• Text for calls recorded while Keep the prompt text was off. The timings are still exact.
• The answer of a call recorded before Keep text kept answers, when that call ended a conversation: it never came back in a later call.