rss

中文 | English

An RSS / news aggregation plugin for DSH: subscribe to multiple RSS / Atom feeds, get one automatically compiled “Today's Worth Reading” digest every day, and have the result injected into the systemPrompt so the model can cite it directly.

Features

• The digest is injected as a systemPrompt section: the model can cite the day's items directly, so asking "what's worth reading today" returns real content. The switch (announceToAgent) is hot — saving takes effect (shown or withdrawn) without a restart; the section lists at most the first 8 items (in digest.items order, category headings kept) and points to that day's digest file for the rest, so every request does not carry an over-long list.
• Scheduled fetching, aggregation by category: generated daily; the modal lets you search, filter by category, and copy the Markdown.
• Feeds maintained inside the panel: built-in channels + custom sources, one-click add from the curated RSSHub catalog, and OPML import/export.
• A failed source degrades in place: when an individual source is down only that source is flagged and the rest still render; the optional AI summary falls back to the original excerpt and explains why.

Transition layer while refreshing

When the “Today's Worth Reading” modal and the settings card refresh, they show a hint in the same visual language as dsh-docker's "target switch transition": a 2px indeterminate shimmer progress bar at the top of the content area, and a blue pill (⟳ Refreshing feeds · current data generated at 14:29, full explanation in title); the bar sits on the lower edge of the category row and the pill rides it, centred — absolutely positioned, no layout shift. It only appears on slow refreshes (the existing 180ms debounce), so millisecond requests never flash; when data lands the body does an 8px slide-up + fade-in, respecting prefers-reduced-motion.

During refresh the body stays dimmed as before (opacity:.55) and the old content remains readable; stale-while-revalidate, the skeleton debounce, and the post-refresh focus/caret and scroll restoration are all retained. The modal's slow path deliberately does not re-render the whole tree, so the transition layer injects two absolutely-positioned nodes instead — miss that path and the overlay never appears.

Implementation notes (the accent-token trap, why absolute positioning is mandatory, the 6 regression assertions) are in ROADMAP § completed (Chinese).

Capabilities

• Built-in channel library (Ruanyifeng's blog, sspai, Solidot, Hacker News, Juejin, ITHome, 36Kr) — just tick the channels you want to show in the settings; 36Kr's official feed is blocked by anti-scraping, so the built-in address is a third-party RSSHub mirror;
• Custom channels are supported: enter your own RSS / Atom URL; on save it is really fetched for validation, and an address that returns no content is reported and not saved; supports OPML import / export (import subscriptions from any RSS reader or website) and bulk import by pasting a URL list (one address per line, or "name, address"), automatically skipping items already subscribed, and after importing you can save and validate them all with one click;
• Feed catalog (multiple sources): the catalog can come from several sources — the built-in curated awesome-rsshub-routes list (official RSS and RSSHub routes, snapshot + silent refresh every 12 hours) + any number of custom OPML catalogs (add OPML addresses in the “Feed catalog” section of the settings card; name / URL are both configurable); results from all sources are merged together, labeled "from xxx" and filterable by source, with search / category filtering / multi-select followed by bulk add (select all / clear / add selected with one click) and live "subscribed / selected" counters; a single failing catalog source does not affect reading the rest of the catalog;
• Parses RSS 2.0 and Atom with zero dependencies, deduplicates by link / id / title, and generates Markdown in reverse chronological order;
• Optional AI summaries: when enabled, the host LLM is called to generate a one-sentence Chinese summary for each item, shown in the digest / modal and appended to the systemPrompt; summaries are cached per entry for 30 days (up to 500 entries, ai-cache.json), and repeated entries are not requested again; a single failure (timeout / non-stop finish / empty output) falls back to the original excerpt only and does not affect other items, while a missing route skips AI summaries for that whole run and records the reason in the digest;
• Generates the day's digest automatically at 08:00 by default; if the day's digest does not exist at plugin startup it is also generated to fill the gap;
• Injects the day's digest into systemPrompt, so the model can cite it directly when the user asks about “Today's Worth Reading”;
• Provides a Web GUI card “RSS / News Aggregation” for managing built-in channel toggles / custom channels / news categories / aggregation settings, with the day's digest refreshed automatically after saving; custom channels support filtering and counting by name / URL, with an inline warning for duplicate URLs; the top of the card previews today's digest directly (item count / source count / generation time / failure warnings) and offers one-click view list, refresh, and copy Markdown;
• The UI uses "block-scoped update" rendering: typing, catalog search, and adding/removing channels rebuild only the affected blocks, so focus / caret / scroll are not lost; unsaved changes are flagged with a badge;
• Provides a “Today's Worth Reading” shortcut in the sidebar below "New Session" (near Task Board / SSH); clicking it opens a modal for reading the news directly; the modal supports search by title / summary / source, category filtering, a "visible / total" counter, and copy Markdown, plus Esc to close and focus trapping;
• The digest is displayed grouped by "category" (items without a category go under "Uncategorized"), and each item is labeled with its source and date; each source's "View more" links straight to its website;
• Manage built-in channel toggles, custom channels (RSS/Atom URL, category, item limit), and news categories on the UI's "Settings" page; a channel's category is chosen from the "News categories" list, and categories in use are merged into that list automatically on save;
• The generated Markdown is stored at ~/.dsh/rss-digest/YYYY-MM-DD.md, together with latest.json for external readers.

Screenshots

“RSS / News Aggregation” card (built-in channel toggles + custom channels + news categories):

Where the settings surface lives depends on the DSH version, but it is always the same form: from 0.2.0-rc.1 it is registered as plugins.bundle.config — inline on the plugin detail page, directly below the description, with no extra ">" step; older hosts (the 0.1.6 line) fall back to the row's ">" sub-page under the Plugins sidebar.

Sidebar “Today's Worth Reading” modal (grouped by category, sources with "View more" linking straight to their websites, manual refresh at the bottom):

Asking the model about the day's news cites the day's digest directly (no need to open the modal):

Configuration example

Pass Config in DSH's plugin configuration:

{
  sources: [
    { name: 'Ruanyifeng blog', url: 'https://www.ruanyifeng.com/blog/atom.xml', category: 'Tech' },
    { name: 'sspai', url: 'https://sspai.com/feed', category: 'Productivity' },
  ],
  maxItemsPerSource: 5,
  maxTotalItems: 30,
  dailyTime: '08:00',
  digestDir: '~/.dsh/rss-digest',
}

When sources is omitted, all built-in channels are enabled by default; when sources is passed, all built-in channels are disabled and only that list is used.

Common configuration options:

AI summaries (optional)

AI summaries are off by default, and are configured only through the ai field of the editable store ~/.dsh/rss.json (editable in the "AI summaries" section of the settings card; no new Config field is added):

{
  "ai": {
    "enabled": true,
    "provider": "deepseek",
    "model": "deepseek-chat",
    "maxItems": 20,
    "concurrency": 3,
    "timeoutMs": 20000
  }
}

If only one of provider / model is filled in, the whole ai config is ignored and a warning is returned in the save response. Summary results are written to ai-cache.json in the digest directory ({ version: 1, entries: { <sha1(link|id|title)>: { text, model, at } } }); a hit that has not expired (30 days) is reused as-is, and beyond 500 entries the oldest are evicted by write time.

Two easy pitfalls (both measured on real runs):

• maxTokens is hard-coded in the plugin (currently 4096), and reasoning models need room for their reasoning budget: reasoning tokens and the final answer share this cap, so too small a value makes the model burn the budget on reasoning before it writes any prose, the finish reason becomes max-tokens, and that item's summary is marked failed. Measured (commandcode/deepseek/deepseek-v4.1-flash, 20 items): 200 → 7 succeed, 1024 → 12, 4096 → all 20 succeed (48 s for the whole run, ~7 s each, still inside timeoutMs). If a heavier reasoning model hits the cap again, the failure reason names the current maxTokens value.
• Failure reasons are surfaced: on partial failure, aiSummary.failures gives the distribution of reasons (up to 3 kinds, with counts), and the Markdown “fetch failed” section states them too (e.g. 原因：终止原因 max-tokens（…）×8). Previously there was only a bare “N failed”, which made it impossible to tell a timeout from rate limiting from a provider error — or to judge whether the plugin was at fault at all.

Development

pnpm --filter @hyzyn/dsh-rss build
pnpm --filter @hyzyn/dsh-rss typecheck

Installing into DSH

dsh plugin --profile web add link:$(pwd)

Or from the repository root:

dsh plugin --profile web add link:$(pwd)/packages/rss

Files