Roadmap

Where Sift is and where it's going. This page is the living plan — it changes as the tool does.

Shipped in v2

• Robust ranking call — adaptive thinking, and clear errors on model refusal or truncated output instead of a broken digest.
• Cost tracking — every run records token counts and USD cost; sift history shows the running total.
• Tuning controls — per-feed weight, a min_score cutoff, and a mute list, all driven from config.toml.
• Email delivery — the weekly digest can be emailed over SMTP, with the password kept in the keychain.
• This site — an all-in-one explainer, usage guide, roadmap, and a browsable archive of every digest, generated from the pipeline itself.
• Full test coverage across the pipeline.

Near-term

• First real end-to-end run — pending an ANTHROPIC_API_KEY in the keychain; once it runs, the first digest gets committed and appears in the archive.
• GitHub Pages — publish this site (served from /docs) once the repo is ready to go public.
• Go public — after a working-tree and git-history secret scan passes.

Ideas / backlog

• A search box over the digest archive.
• Per-category caps (e.g. at most three "business" stories per week).
• A "read it / skip it" feedback signal that nudges future scoring.
• An RSS feed of the digests themselves, for reading in a feed reader.
• Optional push notification (ntfy/Pushover) linking to the latest digest.

Known limitations

• macOS-first: keychain helpers and the launchd schedule assume macOS. The core pipeline is plain Python and runs anywhere with the key in an env var.
• Single reader: the interest profile and digest are built for one person by design. Multi-reader support is explicitly out of scope.
• Local dedup is greedy title-similarity; the model catches the rest during ranking.