AI-RSS-Digest

中文说明 · Complete Agent workflow

An Agent-driven RSS reading workflow: review reading value, group events, suppress cross-edition repetition, and produce deep reads and short summaries.

RSS candidates → bounded body recovery → local preflight → Agent editorial review → validated selection
→ Tier1 full text + deep read → Tier2 summary → validate + format → edition ledger

Features

• Type-specific scoring of significance, novelty, evidence, reader relevance and utility. Local code computes scores and enforces caps.
• Same-occurrence representatives, seven-day publication history and evidence-required follow-ups.
• Pending decisions for insufficient evidence or unverified publication dates; incomplete reviews fail closed.
• Versioned candidate/policy/prompt/history snapshots; complete decision receipts for replay and evaluation.
• Review every seven completed editions, human labels and offline calibration. No scheduler or external LLM API is installed.
• Chinese A-tier feeds enabled, including AIHOT official RSS. AIHOT pages are not scraped.

Install and migration

Requires Node.js >=18. Install this repository into your Agent's skills directory under ai-rss-digest, then run:

npm ci --omit=dev --no-audit --no-fund

Display name: AI-RSS-Digest. Package/skill ID: ai-rss-digest. Version 4 replaces ak-92rss-digest. Verify the new installation before retiring the old discovery entry. Repository: ai-freer/AI-RSS-Digest. The development directory is named AI-RSS-Digest; store runtime data outside installed skills.

Run

Follow SKILL.md, the single authoritative workflow. Set RUN to a unique absolute output directory:

node scripts/rss-fetcher.js --hours 24 --outFile "$RUN/fetch-raw.json"
node scripts/editorial-cli.js supplement --fetchFile "$RUN/fetch-raw.json" --bodyBudget 20 --outFile "$RUN/fetch-ready.json"
node scripts/editorial-cli.js prepare --fetchFile "$RUN/fetch-ready.json" --outFile "$RUN/editorial-packet.json" --promptFile "$RUN/editorial-prompt.md"
# Agent reviews ALL reviewable candidate IDs and writes editorial-review.json.
node scripts/editorial-cli.js select --packetFile "$RUN/editorial-packet.json" --reviewFile "$RUN/editorial-review.json" --outFile "$RUN/fetch-selected.json"

Subsequent steps use fetch-selected.json. Raw rankings are discovery aids, not editorial approval. Zero selected articles is valid; do not fill with rejected items.

Default history: ~/.local/share/ai-rss-digest/ledger.json. Separate audiences/projects should pass the same explicit --ledgerFile to prepare, record and retrospective. Record only a validated, formatted local edition; this is not delivery confirmation.

Verification

node --test scripts/editorial.test.js
npm test

These are offline behavior checks, not proof of improved editorial accuracy. The initial threshold of 65 needs real human labels. See design and attribution, technical contracts and templates.

Version 4.1 uses XML DOM parsing, reports body coverage, and keeps local pending/deferred records outside model prompts. Every original candidate remains accountable. Low body coverage or pending-input is a collection/material issue, not evidence that no worthwhile articles exist.

4.4.1: Restore missing Tier1 source from the selected candidate while preserving an explicitly provided source.

4.4.0: Preserve complete structured text, reject overlong Tier2 summaries for rewriting, and make prompt excerpts grapheme-safe. JSON consumers must use fullContentMeta.summaryOnly; visible badges are rendered only in Markdown/HTML. Document source/language semantics and reject explicit placeholder links before deduplication, review and publication.

4.3.2: Add opt-in environment host skips and normalized failure codes; preserve quality warnings and reviewed tier membership during fallback; cover RSS body reads with request timeouts; clarify per-article evidence and daily advice prompts.

4.3.1: Reject security challenge pages and explicitly truncated X posts; preserve RSS and propagate summary-only labels for all sources. Regression fixtures come from the October 25 live run.

4.3.0: Fetch explicit original links once for reviewed AIHOT Tier1 articles; preserve RSS evidence and label failed or unavailable originals as summary-only.

4.2.1: Fix Chinese sentence boundaries and reject summaries still too short after rebuilding; accept both CLI argument styles and expose ledger loading diagnostics without changing the default path.

4.2.0: Incrementally discover Paul Graham essays from the official index: baseline first, new bodies only, persistent retry/cache, and explicit newly discovered longform review without invented publication dates.

4.1.2: Pause the stale Paul Graham feed while retaining its definition and reason; daily fetching and quality monitoring skip disabled feeds.

4.1.1: AIHOT quotas follow original publishers from explicit RSS links, with a configurable unknown-publisher bucket. Scoring and event deduplication still apply.