tabularasa

Archive, close and refocus Safari tabs. Every open tab lands in a local SQLite inventory. The relevant ones come back in a new window, judged by Jev against your calendar, GitHub activity, and whatever context you hand it. Nothing is lost: every tab you ever closed stays searchable.

macOS only. Node 22.13+ (uses node:sqlite), no npm dependencies. Optional CLIs: gws for Google Calendar, gh for GitHub, claude for free-text requests.

Install

npm install -g @mfittko/tabularasa     # or: npm install -g github:mfittko/tabularasa
tabularasa archive                     # first run: macOS asks whether Node may control Safari; allow it
tabularasa install                     # daily 06:00 LaunchAgent, archive only, plus the Claude Code skill
tabularasa install --close --reopen    # once you trust it: archive, close everything, reopen what's relevant

For development: git clone https://github.com/mfittko/tabularasa && cd tabularasa && npm link.

Jev key: export TYPESAFE_API_KEY=... in ~/.config/typesafe (or the env var).

Use

tabularasa                        # archive + close all tabs, reopen the relevant ones in a new window
tabularasa archive                # archive only, never closes
tabularasa close                  # archive + close, reopen nothing
tabularasa focus                  # dry run: what would be reopened, with probability and reason
tabularasa focus --open --limit 25 --threshold 0.75
tabularasa focus --open --group      # one Safari window per topic cluster (singletons share one window)
tabularasa topic "knowledge graph"        # archive + close everything, open tabs about a topic (90 days) in a new window
tabularasa topic "knowledge graph" --new  # same, but keep the current windows open
tabularasa why "Knowledge Graph"  # explain: candidate? linked? Jev score vs threshold, context lines that share words
tabularasa runs                   # recent archive runs: id, time, tabs, windows, closed?
tabularasa restore                # pick a run, reopen all its tabs, one window per original window
tabularasa restore 25 --pick      # a specific run, choose which tabs come back
tabularasa search playwright      # find archived tabs by title or url
tabularasa forget dependabot      # delete archived tabs matching a term
tabularasa pin "Team dashboard"   # always reopen (own window when grouping); unpin to stop
tabularasa mute youtube           # never reopen, cleanup leaves it alone; unmute to stop; `pins` lists both
tabularasa "drop closed github issues and PRs"   # anything else: Claude Code runs the tab-focus skill

In a Claude Code session the installed tab-focus skill answers "close everything and open only what's relevant today" or "open recent history items on topic X in a new window". There, Claude first gathers Slack messages and recent artifacts into ~/.tab-archive/context.txt, which focus picks up automatically.

How picking works

1. Context: calendar events yesterday through tomorrow (gws, declined and recurring-noise filtered), PRs and issues involving you from the last 7 days (gh), an optional --topic, and ~/.tab-archive/context.txt if present. Open PRs awaiting your review are ingested into the archive so they compete as candidates.
2. Candidates: archive links seen in the last --days (7) days, at most 400.
3. A tab whose URL appears in the context is linked (1.0). Every other candidate is one yes/no question to Jev, "is this tab relevant to the state?", 40 per request. When a page was ingested (tabularasa ingest, or install --ingest for 30 new pages every morning), its description and first 300 characters ride along in the question, which helps with generic titles. Keep those at or above --threshold (0.6), at most --limit (15).
4. Pairwise pass over the picks, two questions per pair: "same underlying content?" drops the lower-scored duplicate (threshold 0.4, measured against control pairs at or below 0.19), and "same piece of work?" orders the tabs as a nearest-neighbour chain so related tabs sit together.
5. --open creates one new Safari window via AppleScript with the picks as tabs. With --group the chain is cut where neighbour similarity drops below 0.5 and each cluster gets its own window; clusters of one tab share a trailing window. tabularasa install --close --reopen --group does this every morning.

About 12 seconds and 3 Jev requests for 130 candidates.

Archive

~/.tab-archive/tabs.db:

• links one row per URL: title, first_seen, last_seen, times_seen (runs it was open in)
• tabs one row per tab per run: window_id, window_index, tab_index, is_current, title, source
• runs one row per run: ran_at, tabs_found, new_links, closed

sqlite3 -column ~/.tab-archive/tabs.db "SELECT last_seen, title, url FROM links ORDER BY last_seen DESC LIMIT 30;"

Guarantees

• Safari is never launched to archive. If it isn't running, the run records zero tabs.
• Windows close only after the archive transaction committed. A failed write leaves everything open.
• Only http(s) URLs are stored. No secrets in files or logs; the Jev key stays in ~/.config/typesafe.
• Only Jev (tab titles and URLs plus the context text), Google and GitHub are contacted.

Uninstall

tabularasa uninstall      # LaunchAgent + skill; keeps ~/.tab-archive
npm unlink -g @mfittko/tabularasa

Test

npm test