Context timeline

Single calls tell you about one request. The timeline shows what happens across a multi-turn run - the “context creep” that quietly inflates every later call.

The Context timeline view: a run picker, cards for turns, context growth, history share at the last turn, and run cost, plus a stacked bar chart of prompt tokens per turn where the history band grows turn over turn.

How runs are detected

llmprof groups consecutive calls into a run automatically: when a call’s messages start with a previous call’s full message list, it is treated as the next turn. No code change required. To group calls explicitly instead, send an x-llmprof-session header. (Details in Architecture.)

Reading the chart

Each bar is one turn, stacked by component. The story to look for: the history (assistant) band grows turn over turn while system prompt and tool schemas stay flat - that is context creep, and it is what makes turn 10 cost several times more than turn 1.

The cards

Turns - how many calls in this run.
Context growth - last turn’s prompt tokens vs the first turn’s (e.g. 2.4x).
History at last turn - what share of the final call is history + tool results, the usual culprit.
Run cost - total spend across the run.

Pick a different run from the dropdown at the top right.