@deepseek-ai/dsh-client-ui-agent-map

Token usage

Each finalized Assistant usage sample is validated at the public snapshot boundary and totaled as uncached input + output + cache reads + cache writes. reasoningTokens is displayed but is not added again because it is already included in outputTokens. Selecting a model response opens a two-by-two Token card above that node with inputTokens, outputTokens, cacheReadTokens, and reasoningTokens, highlights the same call's dashboard segment, and dims the other segments; selecting a dashboard segment selects that same canvas node. cacheWriteTokens remains part of totals but is omitted from the fixed four-cell quick card. The taller dashboard uses enlarged type, and its expanded call list matches the dashboard width. Its denominator is the sum of the loaded model calls, including Tool-call-only Assistant records, so every segment has one visible Model-lane node. It reports the loaded conversation window rather than a hidden all-history total. Calls without a recorded usage sample are absent instead of being estimated.

The presentation uses the host theme's semantic colors rather than fixed light or dark values: node accents, status text, focus outlines, edge glow, glass surfaces, and the canvas grid therefore retain their hierarchy in both appearance modes. A slow aurora, scanning beam, flowing Tool edges, running-node beacon, selected-node circuit, and spatial detail-card entrance add HUD motion without changing graph geometry or React state. prefers-reduced-motion disables every continuous or entrance animation. Status remains written on every node, so color is never its only indicator.