dsh-auto-collapse

When using DeepSeek Harness Web, complex model tasks generate a lot of intermediate process—tool-call cards, Think reasoning blocks, context injections—often drowning the final response in a long workflow. dsh-auto-collapse exists to fix that: it collapses these processes into concise one-line summaries, returning the conversation to just the model's actual words.

The plugin's core is a layered collapsing mechanism. After each turn completes, the working process automatically shrinks to a single line like "Processed Xs", leaving only the model's final text; click to expand the full workflow. When expanded, tool-call groups and thinking blocks are further collapsed into chips, and consecutive thinking can be merged into a tertiary summary line. All visuals align with DSH's native design, transitions are smooth, and it fully supports streaming—state survives React re-renders and out-of-order history mounting.

It's ideal for users who frequently inspect agent tool calls and reasoning chains but still want a clean interface. Whether debugging prompts, reviewing workflows, or just chatting, you can replace walls of process information with a single line and keep the focus on the model's real answer. The status label is configurable, and uninstalling restores everything, so there's no lock-in.