dsh-ui-digest

A turn-digest overview tab for the DeepSeek Harness (dsh) Web UI.

Long agent replies — hundreds of lines of text, tool cards, diffs, reasoning — bury what you actually asked and what the AI did each round. This plugin adds a Overview (全览) tab to the conversation view ring that folds the whole session into one round per user message:

• the message you sent,
• a one-sentence summary of the AI reply (first sentence of the settled reply text, truncated at 160 chars),
• tool-action chips grouped by category (commands run, files edited, files read, web searches, subagent delegations, …) with file paths extracted from edit/write arguments.

Click a row to expand the full user message and the full reply text. The summary is heuristic, deterministic and purely client-side — no LLM call, no RPC, zero cost — and updates live while a turn streams.

Features

• One line per round — user message, one-sentence AI summary, action chips; expandable rows.
• Whole-conversation overview — the digest covers the loaded history window and pages older messages with a "Load older messages" button.
• Live — a running turn shows as waiting for reply and settles as chunks land.
• Zero cost — heuristic summary, pure function of the session snapshot; no LLM, no persistence, no event family.
• Bilingual — Simplified Chinese and English copy through the dsh locale system.
• Read-only projection — no service, no Context merge, no conversation-node definitions; follows the official ui-trajectory pure-consumer pattern.

Installation

As a dsh bundle (recommended)

The package ships a pre-built client bundle in lib/ and declares dsh.bundle + dsh.client, so it installs with the plugin CLI. Grab the tarball from the latest release and add it (no npm account needed):

dsh plugin add https://github.com/rocklau/dsh-ui-digest/releases/download/v0.1.0/dsh-ui-digest-0.1.0.tgz

This injects the ui-digest row into your profile's composition; the browser plugin is served from lib/client.js. A local checkout works the same way: dsh plugin add ./dsh-ui-digest.

From the deepseek-harness source tree

1. Copy the package: cp -R <this-repo> <dsh-root>/packages/client/ui-digest/
2. Register it in the web composition (3 edits, see the deepseek-harness PR pattern):
   • tsconfig.client.json → add { "path": "./packages/client/ui-digest" } to references
   • packages/bundle/web-app/package.json → add "@deepseek-ai/dsh-ui-digest": "workspace:^"… or keep the package name @deepseek-ai/dsh-client-ui-digest and add it as a workspace dep
   • packages/bundle/web-app/cordis.patch.yml → add a roster row - id: ui-digest / name: <package name>
3. pnpm install && pnpm run build && pnpm dsh web

Requires Node 22.19+ / 24 and pnpm 11 (matching the harness toolchain). The lib/ artifacts in this repo are built with the harness's own tsdown pipeline; rebuild with pnpm run build:lib:client inside the harness checkout.

How it works

The plugin is a pure-consumer browser plugin (mirroring the official ui-trajectory shape):

• registers one tab in the conversation.view slot ring (order 20),
• folds snapshot.chat.nodes via the framework standard kit's useSession into rounds keyed by user message,
• rounds start at append-surface user/steering nodes; assistant evidence is settled text from assistant-step nodes, settled tool-call roots, or the turn-tail closing,
• the summary is firstSentence(replyText, 160) with structured fallbacks (ran tool calls only, no text reply, reply in progress).

Tool categorization:

Category	dsh tools
shell · 运行命令	bash, pwsh, terminal
read · 读取文件	read, read_image
edit · 修改文件	edit, write, str_replace_editor
search · 搜索文件	glob, grep
web · 检索网页	web_search, web_fetch
subagent · 委派子代理	subagent*
todo / skill / goal / plan	todo*, skill*, goal*, plan*

Unknown tools fall through to their raw name. File paths are extracted from path/filePath/oldPath/newPath/uri/root/file keys of edit/write arguments (deduplicated, up to 8 per category).

Known limitations

• Window-scoped — the digest covers the loaded history window; page older messages with "Load older messages". Compacted sessions show only the post-compaction surface plus whatever the window retains.
• Heuristic summaries only — first-sentence extraction is language-neutral but naive: an answer opening with a code block summarizes as its first prose fragment. LLM-generated summaries would require a host-side turn/end listener and an event family, deliberately out of scope.
• No cross-view deep links — rows are a read-only projection; they do not jump back into the Chat view at the matching turn.

Development

# inside a deepseek-harness source checkout with this package in packages/client/
pnpm vitest run packages/client/ui-digest/tests/digest.client.spec.ts   # unit tests (19)
pnpm run build:lib:client                                                # rebuild lib/client.js
pnpm dsh web                                                             # run the Web UI

License

MIT. Built on top of deepseek-harness (MIT).