rss-news-digest

name	rss-news-digest
description	Use when the user wants RSS-based research/news digests, feed management, or feed health checks.
metadata	{"short-description":"RSS digests and feed management"}

RSS News Digest

Windows Runtime Commands

On native Windows, use the managed Windows runner and the native runtime command target. Set $runtime to the installed runtime root. Multi-agent installs usually use %LOCALAPPDATA%\ai-agents-skills\runtime. Then run:

$runtime = if ($env:AAS_RUNTIME_ROOT) { $env:AAS_RUNTIME_ROOT } else { "$env:LOCALAPPDATA\ai-agents-skills\runtime" }
& "$runtime\run_skill.ps1" "skills/rss-news-digest/run_and_summarize.ps1" <args>
& "$runtime\run_skill.ps1" "skills/rss-news-digest/rss_news_digest.py" <args>

POSIX examples below use run_skill.sh and .sh command targets; use the Windows command target above on native Windows.

Base path

• $AAS_RUNTIME_WORKSPACE/skills/rss-news-digest/

Use the managed runtime runner rather than invoking the RSS script directly.

Shared runner:

• bash "${AAS_RUNTIME_ROOT:-$HOME/.local/share/ai-agents-skills/runtime}/run_skill.sh"

Use cases

• get the research RSS digest
• get jobs/events/general/video digests
• list/search/add/edit/disable feeds
• run feed doctor/health checks

Core execution

bash "${AAS_RUNTIME_ROOT:-$HOME/.local/share/ai-agents-skills/runtime}/run_skill.sh" skills/rss-news-digest/run_rss_news_digest.sh <COMMAND AND ARGS>

Common actions

• run --tag research
• run --all-tags
• summarize-sidecars (raw, sidecar-derived top-item view; not final synthesis)
• run --tag jobs --max-items 20 --per-feed-limit 5
• list-feeds
• add-feed "<URL>" --tag research --priority 5
• edit-feed "<URL>" --tag research --priority 5
• disable-feed "<URL>" / enable-feed "<URL>"
• remove-feed "<URL>"
• backup-feeds --reason "REASON"
• list-backups
• restore-feeds-backup <managed-backup-name>
• export-feeds-tsv --output /tmp/feeds.tsv
• import-feeds-tsv /tmp/feeds.tsv
• doctor
• search-feeds "<query>"

Verified example shapes:

bash "${AAS_RUNTIME_ROOT:-$HOME/.local/share/ai-agents-skills/runtime}/run_skill.sh" skills/rss-news-digest/run_rss_news_digest.sh run --tag research --max-items 25 --per-feed-limit 5

bash "${AAS_RUNTIME_ROOT:-$HOME/.local/share/ai-agents-skills/runtime}/run_skill.sh" skills/rss-news-digest/run_rss_news_digest.sh add-feed "https://example.com/rss.xml" --tag research --priority 5

After execution

If a digest is produced, read the Markdown path reported by the command output and summarize the top items for the user. Each rss-<tag>.md display artifact has an adjacent rss-<tag>.json digest-items.v1 sidecar for machine handoff. Summary-sidecar loading binds each fixed filename to the matching top-level source (rss-<tag> or rss-all) and requires string item titles/links before any summary/history write. It also requires a consistent successful producer run_status; an all-failed or otherwise unsuccessful sidecar cannot be laundered into a successful empty summary. Feed responses are fetched without redirects under a bounded byte limit, and a declared Content-Length must exactly match the complete body actually read. Real fetches run in isolated workers that the parent terminates and reaps under a whole-response wall-clock deadline, independent of the socket idle timeout, so repeated small chunks cannot extend a fetch indefinitely. Process creation itself may remain platform-dependent.

The Markdown, JSON sidecar, and memory stubs are raw external-source artifacts with artifact_role: raw_external_digest and style_applied: false. Do not parse Markdown as a machine protocol or treat these artifacts as the final user-facing synthesis.

A run where every attempted feed fails exits nonzero; partial feed failure is reported as degraded. Feed response bytes and total items are bounded across the run; the aggregate response-byte reservation is shared across parallel workers and is acquired before another response is opened. Early status errors release unread reservations, while bounded partial bytes from truncated worker responses are charged to the shared run budget. Before fetching, the producer validates the digest directory, every required digest/sidecar leaf, and the state leaf. It serializes all digest, sidecar, and state payloads before the first digest write and publishes bridge-visible sidecars last. A handled publication failure restores the exact prior sidecar generation and state bytes or absence; rollback failure is explicit. These retry guarantees cover handled command failures, not abrupt termination, power loss, or same-UID races, because there is no cross-file crash journal or transaction lock. Existing seen/health state is read only from a bounded regular non-link-like JSON file; unsafe, oversized, or malformed state—including duplicate JSON object members and non-standard NaN/Infinity constants—is preserved and blocks run/doctor publication, while only true absence starts empty. An ingest-ledger hit is not an unconditional skip: its corresponding owned regular memory stub is checked, a missing stub is rebuilt, and a foreign or link-like occupant fails retryably. An existing stub is owned only when its complete bounded frontmatter, exact digest identity and schema, and fixed body shape match this producer; a marker- only, truncated, or self-asserted foreign file is preserved and fails retryably. Summary limits of 0, 1, or 2 characters are honored exactly without adding an ellipsis beyond the requested bound. Feed add/edit operations validate the complete prospective config before creating a backup or replacing the current file. Bootstrap, legacy migration, and backup restore distinguish a genuinely absent config from any existing directory entry; unsafe live or broken feed/profile symlinks are refused and preserved. Feed TSV imports require exact, unpadded, unique managed headers and a url column; malformed replacement input cannot create a backup or alter the current config. Exact or tracking-normalized duplicate feed URLs are rejected as one ambiguous canonical identity before backup or replacement. Missing optional fields use documented inference, but nonempty explicit values must use a recognized boolean, a managed tag, an integer priority from 0 through 10, and bounded kind/notes text. Add and edit operations use the same range validation; edits never clamp invalid priorities into a different accepted value. Add/edit/remove/enable/disable validate their lookup URL as a bounded absolute HTTP(S) URL before tracking normalization, so an oversized tracking-only suffix cannot alias and mutate a stored feed. Profile JSON is a bounded, strictly shaped map of canonical names to bounded text terms. A present malformed, oversized, invalid-UTF-8, or otherwise invalid profile file fails closed. Terms are case-folded and deduplicated once at admission under a 50,000-character aggregate ceiling; per-item scoring reuses that prepared tuple rather than repeating Unicode/HTML/regex normalization. When names no configured profile, both and stop before feed fetching, state saving, or digest output instead of silently running without the requested filter.

Truncated or malformed HTTP response streams are isolated as errors for their individual feeds. Other feeds may still produce a degraded successful run; the broken response cannot abort the entire multi-feed command. XML containing DTD or entity declarations, more than 500 entries, 20,000 elements or attributes, or 128 levels of nesting is rejected by a streaming structural precheck before feed parsing, and feedparser must recognize a feed version even when it recovers entry-like fragments. Parsed entries are consumed through a bounded worker window rather than copied into an unbounded run-wide list. Remote title, summary, and description fields are type-checked and bounded before profile-term scoring, so an oversized or non-text field cannot amplify matching work or abort the run. Timezone-less RFC feed dates are interpreted as UTC, and published timestamps are emitted in UTC, so freshness and ordering do not depend on the host zone. Calendar values outside the platform's representable UTC range are treated as unknown rather than aborting the feed. Timestamps farther than 24 hours in the future are also normalized to unknown, receive no freshness bonus, emit no published timestamp, and cannot win sorting through a hostile future date. Special-source identities that exceed the shared state/stub key ceiling fall back to a fixed digest of the complete entry identity; they are never truncated into unstable state or colliding memory-stub IDs. arXiv, YouTube, and StackExchange IDs are extracted only from their exact recognized hosts and path shapes under ASCII-explicit path grammars; Unicode case-fold lookalikes and hostile lookalike URLs or titles use the structured fallback instead. Stack Exchange question identity includes the canonical site host because numeric question IDs are site-local; oversized host-plus-ID identities hash the complete candidate before entering state, stubs, or ingest history. Generic fallback identities hash a structured, versioned tuple including source kind, ID, canonical link, and title, so delimiter-bearing fields cannot serialize to the same key. URL canonicalization removes configured tracking parameters with conservative raw-query filtering. Retained segments keep their original structure and order, and strict-decoding failures preserve the untouched query; encoded delimiters, blank/bare fields, invalid octets, and reversed repeated values therefore cannot collapse during item deduplication or feed merge/edit lookup. Host-specific tracking keys such as YouTube si/feature are removed only on recognized YouTube hosts and remain semantic on generic feed URLs.

RSS state and ingest history have independent byte and record bounds. Existing seen and feed maps that exceed their admission counts fail closed rather than being silently compacted on read. At the exact boundary, repeated seen keys are stable-deduplicated newest-first before retention so duplicates cannot evict distinct history. Digest ingest ownership lives at $AAS_RUNTIME_WORKSPACE/data/research/rss/ingested.json; the former shared data/library/ingested.json is read only as an optional one-time migration source, and only its source: digest records are admitted. An invalid optional legacy ledger is skipped with a warning, while a corrupt dedicated RSS ledger fails closed; this includes live or broken symlink directory entries, which are never replaced as if the ledger were absent. Dedicated JSON rejects duplicate object members before any stub or ledger mutation. Dedicated records use exactly source, id, and processed_at, with unique canonical control-clean IDs and valid UTC producer timestamps. Malformed or duplicate digest records in optional legacy input are skipped rather than weakening the dedicated format. The shared library ledger is never rewritten by this skill. The fixed memory and memory/papers directory entries are admitted with lstat before any stub or ledger work. A link-like papers root—including a Windows reparse point or junction—is a retryable stub failure and cannot redirect an owned Markdown write. Memory-stub filenames retain a readable bounded slug but append the full SHA-256 digest of the complete canonical item key. Distinct valid source IDs therefore cannot be forced into one owned stub through a short hash prefix. Bounded feed-health retention keeps the newest entries, and every feed touched by the current run is moved to that retained tail even when its state key predates the stale history around it.

run_and_summarize.sh and run_and_summarize.ps1 invoke summarize-sidecars --no-history; they never parse the Markdown digest and automatic runs update only last-summary.md. That artifact is explicitly marked artifact_role: raw_external_digest and style_applied: false. An operator may invoke summarize-sidecars without --no-history to create a timestamped raw history file, but that manual history is operator-managed and has no automatic retention policy. Canonical sidecar admission is based on directory entries, not target existence: live or broken symlinks and non-regular optional/aggregate sidecar occupants are errors rather than silently absent inputs. The summary command also admits the digest-directory entry itself before reading or writing, so a live or broken directory symlink cannot redirect last-summary.md, a custom output, or timestamped history. A custom output must be a direct child of that admitted directory and cannot alias a canonical RSS sidecar/Markdown pair, feed/profile config, state, or ingest ledger; collision or parent traversal is rejected before any output/history mutation.

The feed-backup root is likewise admitted as a real non-symlink directory for create, list, resolve, restore, and retention rotation. An unsafe root fails closed before any matching external backup can be written or unlinked. The complete directory index is bounded at 10,000 total entries and fails closed on overflow rather than sampling an arbitrary prefix. Every managed-name entry must be a bounded readable UTF-8 regular file with a plausible UTC filename timestamp no more than five minutes ahead; live/broken symlinks, oversized/unreadable content, and far-future names poison the complete index. Recovery snapshots may contain readable malformed config bytes. Only the 50 newest managed feed snapshots are retained, while the just-created snapshot is mandatory even when within-skew future names sort ahead.

All owned RSS path admissions use the same link-like predicate: POSIX symlinks and Windows reparse points/junctions are rejected before summary, stub, state, or retention work, so a platform-specific link cannot redirect those effects.

Writing Style Gate

For any user-facing RSS digest summary, load writing-style-settings.md before writing. If the digest item or synthesis is mathematical, TCS, graph-theoretic, Lean-related, or LaTeX manuscript prose, also load math-manuscript-style.md. The later agent-authored stored summary should record style_profile_ref, policy_hash, active_overlays, active_requirement_ids, and style_applied; do not accept a bare style_applied: true assertion as sufficient evidence.