dsh-ui-digest

Anyone who has run a few long agent sessions in DeepSeek Harness knows the pain: the AI spews hundreds of lines of reasoning, tool-call cards, diffs, and code blocks, and when you scroll back through the conversation you have lost track of what you actually asked and what the model did in each round. dsh-ui-digest solves this with a single light touch — it adds an Overview tab to the conversation view ring that folds the entire session into one scannable line per user message, so you can reconstruct the arc of the whole conversation in about thirty seconds.

Each line shows three things: the original message you sent, a one-sentence summary of the AI reply (first sentence of the settled text, truncated at 140 characters), and a row of tool-action chips. The chips are grouped by category — commands run, files edited, files read, web searches, subagent delegations — with file paths extracted directly from edit and write arguments, deduplicated to a maximum of eight per category. Click any row to expand the full question and full reply without hunting through a wall of text. The summary is a purely client-side heuristic: no LLM call, no RPC, no persistence. While a turn is streaming the row shows a waiting-for-reply state and settles as chunks land. The plugin follows the official ui-trajectory pure-consumer pattern — no service registration, no Context merge, no conversation-node definitions — so it leaves zero footprint on the host. Copy is available in both Simplified Chinese and English through the dsh locale system.

It is purpose-built for a specific moment: you are doing multi-turn development, subagent orchestration, or iterative refinement in the dsh Web UI, and you need to explain the arc of a session to a colleague or simply reorient yourself before the next prompt. It does not replace the Chat view and does not deep-link across views; it is a read-only projection of the session, zero-cost and zero-dependency, ready to use the moment you install it.