Features

Everything Sift does, grouped by what it's for. Each line names the command or config knob that drives it.

Curation

Multi-feed fetch — any number of RSS/Atom feeds. sift add <url> validates a feed before adding it; sift list shows them all.
Seen-once memory — every URL ever shown is remembered in SQLite, so a story never appears in two digests.
Freshness window — items older than eight days are dropped automatically.
Free local dedup — near-identical headlines from different feeds collapse into one cluster before anything is sent to the model.
One-call ranking — a single Claude call merges remaining duplicates, categorizes (models & research / tooling / infra / policy / business), scores importance 1–10, and writes a two-sentence neutral summary per story.
Primary-source flag — any story whose central claim isn't from a primary source (official blog, paper, announcement) is marked needs verification.

Interest profile — a paragraph of plain language describing your signal vs. noise; it drives every score.
Per-feed weight — weight = 2.0 floats a trusted source's stories up (a score multiplier, plus a trust hint to the model).
Minimum score — min_score drops weak stories before the digest's item cap.
Muted topics — mute = [...] tells the model to down-rank or exclude subjects you never want to see.
Item cap — max_items_per_digest keeps only the top-scored N.
Model choice — model is any current Anthropic id (Opus for quality, Sonnet for cheaper runs).

Self-contained HTML digest — dark-mode-friendly, opens anywhere, no external assets, written to docs/digests/.
JSON digest — the same data as machine-readable JSON for anything downstream.
Email — [email] enabled = true sends each digest over SMTP; re-send any past week with sift email <week>. The password is read from the keychain, never stored or logged.
Live site archive — every digest is browsable from the site's archive index.

Exact per-run cost — token counts and USD are computed from the model's pricing and stored per run.
Run history — sift history lists every run with tokens, cost, and a running total.
Full logging — logs/sift.log records counts and errors for every step.
Free dry-run — sift run --dry-run runs the whole pipeline up to (but not including) the API call and prints exactly what would be sent — $0.

All-in-one — explainer, this feature list, the how-it-works flow + architecture, a usage guide, the roadmap, and the digest archive.
One visual language — a single shared stylesheet; the digests match it but stay self-contained.
Generated from the pipeline — sift run rebuilds the site automatically; sift site rebuilds it on demand.
GitHub Pages–ready — served from main / /docs.

Scheduled runs — a launchd plist runs the pipeline weekly (Sunday 6am) and reads secrets from the keychain at run time.
Resilient — a dead feed, a failed email, or a site-rebuild error is logged and skipped; the digest is never lost.
Schema migration — the SQLite history migrates itself forward in place, so old databases keep working.

Tested — unit + integration tests across the pipeline, with an 80% coverage gate enforced in CI.
Reviewed — the v2 changes went through a multi-agent adversarial review.
Secrets stay out of git — keychain-only secrets, verified by a working-tree and full-history scan before going public.