dsh-ui-tool-graph

After an agent finishes a turn, its trajectory log hides a wealth of efficiency signals — which tool consumed the most wall time, which calls kept failing, how many tokens each round actually burned, which calls repeated themselves verbatim — yet no human can easily parse all of that from a raw event stream. dsh-ui-tool-graph folds the entire session log into a single weighted relationship graph, making tool-call economics visible at a glance with zero token cost, deterministic replay, and no extra model calls.

The core view is a radial weighted graph: nodes are tools, sized by whichever metric you pick (total duration, call count, or errors) with an outer ring proportional to error rate; edges are consecutive dispatch pairs within a turn, weighted by transition count. Above the graph, rule-derived insights surface the dominant bottleneck, error-prone tools, verbatim repeat calls (cache or dedup candidates), consecutive same-tool chains, model-generation waits, and the token-heaviest step, each tagged with a severity dot. A per-tool ranking table lets you drill into every individual call with its duration and argument excerpt, and a session-totals strip shows calls, failures, tool wall time, longest average gap, and output tokens.

Who it is for: any developer or operator driving agents through the dsh Web UI. When you want a clear answer on where this round was slow, repetitive, or optimizable, open the Tool Graph tab for a self-contained efficiency audit. You can even copy the generated analysis prompt to the clipboard or queue it as your next message, stepping straight into the optimization loop.