dsh-trace-collapse

Every time an Agent runs inside the DeepSeek Harness Web UI, it leaves behind a long trail of intermediate steps: multi-turn reasoning, tool calls, steering directives, workflow executions, system-prompt and skill-catalog context injections. Those records are genuinely useful for debugging, but in day-to-day conversation they push the one thing you actually care about — the final output — off-screen and make it tedious to find.

The fix is deliberately simple: a small, restrained collapse button sits above each completed Agent turn. Hit it and every intermediate node in that turn — tool-calls, non-final assistant reasoning steps, steering lines, workflow runs, context-injection rows, even unknown entries — disappears, leaving only your prompt and the final reply. Re-expand any time you need to audit the steps. A settings toggle ("Collapse agent trace by default after final output", on by default) removes the need to click per turn on a daily basis.

Who should use it? If you interact with Agents in DeepSeek Harness many times a day and your workflow is production conversation rather than step-by-step debugging, but you still want the full trail available when something goes wrong, this plugin is a quiet quality-of-life upgrade. It is implemented entirely client-side; no server changes, no performance cost when idle.