Changelog

All notable changes to FindThatPage are documented here.

1.9.1 — 2026-05-04 — “This site” defaults ON when the tab is pivotable

Follow-up to 1.9.0. Opening the popup/overlay/full-search on an http(s) page now defaults straight into the This site pill — so you see your history for the current site immediately, no click required.

Auto-fallback safety net: if you’re on a site you’ve never indexed before (and therefore “This site” would filter to nothing), the first search response silently drops the scope back to Recent. You never open to an empty list.

Manual clicks on any pill disable the auto-fallback for that session, so your choice always wins.

On pages where there’s no pivotable domain (chrome://, about:, extension pages), behavior is unchanged — pill hidden, Recent is the default.

1.9.0 — 2026-05-04 — “This site” empty-state filter

New third pill in the empty-state sort toggle: This site. Click it with the search box empty to see only pages from the domain of your currently-active tab, sorted by frecency.

• Reading Kubernetes docs and want a quick recap of everything you’ve read on kubernetes.io? Click “This site.”
• Reading a GitHub PR and want to see your prior GitHub pages? Same — one click.
• Start typing → scope drops automatically so full-text search runs across your entire memory. Clear the query → scope returns.

Available in:

• Popup — learns the active tab via browser.tabs.query.
• Overlay (Ctrl+Shift+K on any page) — learns the host page via a URL param passed by the content script.
• Full-tab search — learns the launching tab via a URL param passed by the background’s open-search command.

Pill is hidden when there’s no pivotable domain (chrome://, about:, extension pages, data:/blob: URIs) — offering it there would point at zero results.

No schema change. No persistence: opens in its default off state every popup/overlay/tab, so it’s opt-in per visit.

Tests

• +6 tests for getSourceDomain() covering http/https extraction, www. stripping, all unpivotable protocols, malformed input, and subdomains. 318 tests green total.

1.8.7 — 2026-05-04 — Top sources “Delete pages” actually deletes

• Fix: Clicking Delete pages on a row in the Top sources section used to pre-fill the “Delete all pages from a domain” text input further down the page and scroll to it — forcing a second click on a different button to finally delete. That’s two half-working UIs stitched together. Now the row button runs the delete directly (confirm dialog + toast, same as the typed-domain path), so one click = one delete.

1.8.6 — 2026-05-04 — Fix adjacent-element text gluing (searchability)

Reported: indexing an Amazon phonetool page with DOM like <span>Sup Soln-Fxd</span><h3>Harsh Shah</h3> stored the body text as "Sup Soln-FxdHarsh Shah" — two consecutive elements with no space between their text content. Searching for Harsh then returned no result, because the token in the index was fxdharsh.

Root cause: both extraction paths (Mozilla Readability output + legacy tag stripper) called Node.textContent, which concatenates all descendant text verbatim with no separator. Whatever the DOM looks like, textContent gives you one fused string. Block-level boundaries are invisible to it.

Fix: replaced both paths with extractTextWithSeparators(), a block-aware walker that emits a space around every block-level element (<p>, <h1>–<h6>, <div>, <li>, <br>, <td>, etc., 40+ tags). Inline elements (<span>, <em>, <a>, <strong>) still concat without a gap, so <em>bold</em>face stays boldface and doesn’t hurt phrase-match fidelity.

No schema change. Newly indexed pages extract cleanly from now on. Previously indexed pages keep their glued text — revisit the tab to re-index and pick up the fix, or wait for the next auto-reindex pass (10 min on visible tabs).

Tests

• +7 new tests: the exact Soln-Fxd / Harsh Shah regression, block vs. inline behavior, <br> handling, whitespace collapse, empty-root edge case. 312 tests green total.

1.8.5 — 2026-05-04 — Show matching text from deep in the page

Result cards were showing each page’s summary instead of the actual phrase that matched your query. When the match was buried deep in the page body — which is most of the time for long articles and docs — there was no visual cue about why a result made it in.

Root cause: FTS5 is configured in contentless mode (no body text stored in the FTS index — just the inverted index, to save ~50% of DB size). SQLite’s built-in snippet() can’t quote from a contentless index, so it was returning empty strings for deep-body matches. The card code then fell back to page.summary.

Fix: client-side snippet synthesis. After FTS returns its ranked rows, for any row whose highlight columns are empty, we batch-load pages.text (already stored in the main table), find the first location of your query tokens, and slice ~180 chars of context around it with the matched phrase wrapped in <mark>. Runs only on the visible result set (default 30 rows) — no FTS index bloat, no schema change, no per-page network round-trips.

• Matches are case-insensitive but preserve the page’s original casing in the rendered snippet.
• Multi-term queries highlight every occurrence in the snippet window, not just the first. Overlapping spans (e.g. new + york + new york) collapse into a single merged highlight.
• Ellipses (…) bracket the snippet when it’s mid-sentence on either side.

Tests

• +10 new tests for buildBodySnippet covering empty input, deep matches, ellipsis behavior, multi-term merge, Unicode/casing, and maxLength clamps. 305 tests green.
• No schema change.

1.8.4 — 2026-05-03 — Related: co-visit first, topical second

User feedback made it clear that Related had the wrong mental model. The feature was computing “topically similar pages” (shared keywords), but the expected behavior is “pages opened in the same sitting.” Rewritten to match:

• Primary signal — co-visit, ±10-minute window. Every page whose most-recent visit lands within 10 minutes of the source’s latest visit qualifies outright, no keyword overlap required. This is what “related” meant all along.
• Secondary signal — topical inference. Pages outside the time window still appear if they share ≥2 meaningful keywords with the source (or ≥1 AND the same site). Catches “I read about this last week and again today.”
• Ranking. Co-visited results always come first, tightest time-diff wins. Topical-only results fill the tail.
• Reason chips. “Visited 3 min apart” / “Visited seconds apart” / “2 shared keywords” / “Same site · example.com” — tells you why each suggestion made it in.

Previous ±4h window was catastrophically wrong: it swept in every unrelated tab in the same browser session. Tightening to ±10 min fixes the reported “Amazon approvals” → “Chrome extensions” problem.

1.8.3 — 2026-05-03 — Related now actually means related

Reports on 1.8.2 showed Related still surfacing unrelated pages: clicking Related on an Amazon approvals page pulled up “What’s new in Chrome extensions” because both were visited in the same browser session. Three underlying problems, all fixed:

• Co-visit alone no longer qualifies. The 4-hour visit window is an incidental signal (browsers naturally co-locate unrelated tabs), not a topical one. It’s demoted from “qualifier” to “tiebreaker in the final ranking.”
• Stronger topical bar. A candidate must share ≥3 meaningful keywords with the source page — or share ≥1 meaningful keyword AND come from the same site. Below that, it’s dropped.
• Boilerplate stoplist. Generic keywords (page, home, view, settings, website, loading, 40+ more) used to drive false matches because they appeared in nearly every page’s keyword set. They’re now filtered out of the topical-match path; real discriminating vocabulary (webassembly, kubernetes, typescript) still counts.

Tests

• +7 new tests for the FTS keyword sanitizer covering punctuation, numeric-only fragments, short tokens, and the stoplist. 295 tests green total.
• No schema change.

1.8.2 — 2026-05-03 — Related correctness fix

• Fix: Clicking Related on a page with keywords like node.js, c++, iOS-15, or any term containing ., +, -, : crashed with sqlite3_step() rc= 1 SQLITE_ERROR — those punctuation chars blow up the FTS5 MATCH parser. Keywords are now sanitized to [a-z0-9]+ before being spliced into the query; non-alphanumeric fragments are stripped, the longest remaining stem per keyword survives.
• Fix: Related panel was surfacing unrelated pages. The old query fired as soon as any keyword prefix matched (e.g. docker* matching “documentation”) and always labelled the result “Shared keywords” regardless. Now a candidate must share ≥2 distinct keywords — or clear a real signal bar (same site / co-visited ±4h) — to appear. Reasons now show the actual hit count: “3 shared keywords”.

1.8.1 — 2026-05-03 — Popup Related fix

• Fix: The “Related” button introduced in 1.8.0 never rendered in the popup (the most-used surface). Overlay and full search had it wired; the popup’s ResultCard was missing the onShowRelated prop, which conditionally hides the button entirely. Now wired in all three entrypoints.

1.8.0 — 2026-05-03 — Related pages, bulk forget, auto-pin, polish

New: related pages

• “Related” button on every result card. One tap opens a modal with up to 5 pages that share signal with this one: same topic (keyword overlap), same site, or visited around the same time (±4 hours). Each suggestion explains why it showed up — “Same site · github.com”, “Visited around the same time”, “Shared keywords”.
• Good for “I remember reading about X when I was researching Y” — the kind of connection keyword search alone can’t surface.

Bulk operations

• “Forget all N matches” on the full-tab search page. Deletes every page matching the current query + chips + domain filter. Explicit confirm dialog with the exact count; pinned pages are preserved server-side regardless; capped at 500 deletions per click. Pairs with the existing delete-by-domain for cleanup workflows.
• Top sources section in Options. Shows the top 20 domains by indexed page count with a quick-delete shortcut.

Ergonomics

• Keyboard shortcut for “Index this page now”. New command (index-this-page) — users bind it via chrome://extensions/shortcuts. Shows a small badge on the extension icon after: ✓ green for a new row, ↺ blue for update, ✗ red on failure. Clears after 2s.
• Whole-card click-to-open. Previously only the title/Open button opened pages. Now the whole card is clickable; Cmd/Ctrl/Shift/ middle-click opens in a background tab.
• Copy URL button in the card header (⧉ icon, flips to ✓ after copy).
• Multi-term text fragment highlighting. Chrome’s #:~:text= URLs now include up to 3 highlighted phrases from the snippet instead of just the first — a multi-word query visibly highlights each match on the destination page.
• Favicon broken-URL cache. When one result card’s favicon fails to load, other cards showing the same URL skip the <img> entirely and render the letter fallback — no reflow flash.

New setting

• Auto-pin frequently-visited pages (opt-in, default off). Set a visit-count threshold in Options → Search UX; pages that cross it get auto-pinned so they float to the top of empty-state. Manual unpin still wins until the threshold is crossed again.

Tests

• +3 new tests for multi-fragment text highlight. 288 tests green total. No schema change — everything uses existing columns or computed at query time.

1.7.0 — 2026-05-03 — Readability extraction, language filtering, faceted search

Extraction quality

• Mozilla Readability.js swap. Same engine Firefox Reader View uses. On media/news/blog/docs sites it cuts nav/ads/sidebar/comments/related-posts far more accurately than the old tag stripper. Falls back to the legacy stripper when Readability declines a page or throws — safe by default.
• Content-script bundle grew by ~36KB to accommodate Readability; only loads inside content scripts, background/popup are unchanged.

Language

• Per-page language detection. Every indexed page gets a BCP-47 code — we prefer <html lang="…"> when set, fall back to a tiny heuristic scan over common function words in en/es/fr/de/pt/it. Stored on the page row; import/export round-trips preserve it.
• New lang:CODE query grammar: lang:fr react narrows to French- tagged pages containing “react”. Combines with site:, in:, after:, before:, today, etc.

Faceted search (full-tab page)

• Facet rail below the query hint with three chip groups:
  • Top 8 domains with page counts (click to filter)
  • Time buckets: Last 24h / 7d / 30d / Older (click appends the right time token)
  • Language chips (only shown when the corpus spans 2+ languages — otherwise the group is hidden)
• Counts reflect the current active filters but exclude the facet being counted, so the domain chips still show alternatives when a domain filter is active.
• Runs in parallel with the main search — no added latency.

Tests

• +20 new tests (detectLanguage, langFilter, lang integration tests, Readability extractor smoke test).
• 285 tests green total.

Schema

• Additive: pages.lang TEXT NOT NULL DEFAULT 'en' column added via addColumnIfMissing. Existing rows default to 'en' and get the real detected value on next revisit.

1.6.0 — 2026-05-03 — Time filters, pagination, title highlights, domain collapse

Power-user query grammar

• Time-range filters: after:2024-01-01, before:yesterday, today, yesterday, this-week, this-month. Combine freely with other tokens: after:2024-10-01 react hooks narrows ranked results to the date window. Multiple after: narrows to the tightest floor.

Scale

• “Load 30 more” on the full-tab search page. Previous 30-result cap was a ceiling, not a convenience — now it’s just the first page. Cached per-page so paging back and forth is instant.

Speed

• Pre-warm SQLite on SW startup. The first search after Chrome wakes a suspended service worker used to eat 300-600ms compiling a prepared statement; now it’s primed in the bootstrap path. First search feels instant.

UX polish (Tier C bundle)

• Highlighted matches in result titles, not just body snippets. The FTS highlight markers were already computed for titles — now rendered.
• Progressive Esc. First Esc clears the query + chips; a second Esc closes the overlay / popup. Previously a typo + Esc slammed it shut and threw your text away.
• Per-token “Did you mean?” chips. When there are 2+ corrections, you get one clickable chip per suggestion — click replaces only that token (Google/Kagi behaviour) — plus a small “Apply all” if you want every suggestion at once. Single-correction flow unchanged.

Result list ergonomics

• Group-by-domain collapse. When a single domain contributes ≥5 hits to a query (not empty-state, not already domain-filtered), the top 2 results show, then a dashed “N more from domain.com →” row appears. Click to expand. Keeps broad queries like “react” scannable when most hits are from one site.

Tests

• +46 new tests (collapseDomains, timeFilter, applyCorrection).
• 264 total tests green.

1.5.0 — 2026-05-03 — Faster search, cleaner titles, recent-queries, better empty states

Five independent wins, all shipping together. No schema change.

Search speed

• Background LRU query cache (20 entries, 30 s TTL). Repeat searches — re-opening the popup on the same query, clearing-and-retyping, bouncing between two queries — now return in under a millisecond instead of re-running FTS (previously 5-60 ms depending on corpus size). Cache is invalidated in lockstep with every DB write via a generation counter; no stale results after save/delete/import/clear.

Indexing quality

• Title cleanup on index. Strips boilerplate site-name suffixes like “ — React”, “ | GitHub”, “ · Stack Overflow”, “ - Medium” when they look like a site brand (contain the domain root or a known tail word: Home, Blog, Docs, Wiki, News, Store). Keeps the real title; FTS5’s title-column BM25 weight stops being biased by the same repeated brand on every page from that domain.
• Meta description fallback for summary. When the auto-generated first-3-sentences summary is under 40 chars (common on JS-heavy sites that only have pre-render boilerplate at document_idle), fall back to meta[name="description"] → og:description → twitter:description. Better summaries on most SPAs.

UX

• Recent searches (popup). When the search box is empty and you have indexed history, the popup shows clickable chips for your last 8 productive searches (≥3 chars, non-empty result set). Case- insensitive dedupe keeps repeats from filling the list. Persisted locally — never synced across devices, there’s a Clear button.
• Dead-query guard. Queries that parse to nothing (all stopwords, all too-short tokens like “a” or “!!”) used to fall through silently to the empty-state recents, making the UI look like it ignored the input. Now shows “Type at least 3 letters to search.” Empty input still shows recents — that’s the intended empty-state, not the guard.

Tests

• +36 new tests (cleanTitle, queryCache, isQueryDead, recentQueries).
• 231 total tests green.

1.4.1 — 2026-05-01 — Popup UX fixes

• “Recent” now means recent. Default empty-state sort was silently frecency (last-visit time weighted by visit count), so a frequently- visited older page would outrank one you just opened. Now it’s pure last-visit order. “Most visited” sort still works as before.
• “Index this page” now tells you what happened. Clicking it on a URL already in the index used to say “Indexed.” with no count change, looking broken. It’s now “Updated existing entry.” (correct — dedupe by normalized URL) versus “Added to memory.” for a new row.
• Sort toggle added to the popup. Recent / Most-visited tabs appear when no query is entered. Changing it writes to settings so the preference is consistent across the popup, overlay, and full search.

1.4.0 — 2026-05-01 — High-coverage indexing for chat sites

Long AI chats (ChatGPT, Claude, Gemini) now get 3× the indexing coverage of regular pages, with near-real-time re-indexing so streaming replies become searchable within a minute.

What changed

• 150,000-char index cap (was 50,000) for high-coverage domains — covers ~25,000 words of conversation per chat, up from ~8,000.
• 45-second re-index cadence with a 200-char delta threshold (was 10 minutes / 400 chars) — streaming responses surface quickly.
• Default list: chatgpt.com, chat.openai.com, claude.ai, gemini.google.com. Add your own in Options → Long content / chat sites.
• Subdomain-aware matching shared with the excluded-domain logic — sub.chatgpt.com picks up the setting too.

Scope guard

• Privacy rules are unchanged; password-gated / banking / sensitive-path pages are still blocked regardless of high-coverage status.
• Keywords extraction still caps its input at 50k chars to avoid spending serious time on giant conversations.

Tests

• 16 new tests (11 for matchesDomainList — subdomain correctness, suffix spoofing, case sensitivity — and 5 for the extractor maxChars override).
• 195 tests green total.

1.3.0 — 2026-05-01 — Chunked body search, fuzzy fallback, redesigned settings

Search depth

• 4× more content searchable per page. Body text is now indexed as overlapping ~3000-char chunks up to 20,000 chars total, replacing the flat 5,000-char cap. “I read it somewhere” on long Wikipedia / GitHub / blog posts now resolves.
• Schema v3 migration rebuilds FTS from stored text on first run — existing history gains chunked coverage without re-browsing.
• Best-chunk-per-page aggregation keeps ranking meaningful; same term appearing in multiple chunks no longer duplicates the page in results.

Fuzzy “Did you mean?”

• Zero-hit queries now offer a one-click correction (“kuberntes” → “kubernetes”) in the popup, overlay, and full-tab search page.
• Prefix-extension fallback: partial queries like “observabili” that FTS5 porter can’t prefix-match now get suggested the full word.
• Toggleable in Settings → Search experience.

Options page redesign

• Hero header with big-number stats (pages / domains / visits / pinned).
• Sticky section nav. Card-style sections with coloured accent bars.
• Modern switch-style toggles replace plain checkboxes.
• Segmented Recent / Most-visited control.
• Pause shortcuts extended: 15 min / 1 hour / 8 hour.

New user preferences (all sync-mirrored)

• Show/hide “Did you mean?” suggestions
• Default-open results in background tabs
• Toggle word count on result cards
• Persist Recent vs Most-visited sort globally (overridable per session)

New bulk action

• Delete all pages from a domain — retroactive cleanup separate from the exclude-list (exclude prevents future indexing; this clears existing history).

Import & reliability

• Fixes “Import failed — not a FindThatPage export?” for large export files: the prior code bundled the full page list into one 64 MiB-capped message. Now chunked into 200-page batches with per-chunk progress and real error surfacing.
• Tightened isImportablePage guards against malformed exports crashing downstream UI string ops.
• Scheme-check on open: javascript: / data: / file: URLs smuggled via import or storage corruption now resolve to about:blank before reaching browser.tabs.create.

Testing

• 179 tests green (was 139). New coverage for the chunker, chunked search ranking, fuzzy prefix matching, import validation, URL scheme gating, and concurrent import races.
• CI now gates releases on green typecheck + tests.

Performance

• 50k-page bench: worst query 59ms p95 (was ~49ms pre-chunk). Trade of ~1.5× latency for 4× content coverage. See docs/BENCHMARKS.md.

1.2.0 — 2026-04-30 — Bigger empty state, sort toggle, Firefox install polish

• Overlay empty-state: raised the recents cap from 20 → 50 so you can actually scroll recent history without a query.
• New sort toggle in the overlay: Recent (frecency — last visit boosted by visit count) vs. Most visited (raw visit count, tie-break by last visit). Applies to the entire page table, not just the visible 50.
• Clearer count label: Most recent 50 · 35 total / Most visited 50 · 35 total / N matches · 35 total replaces the old 20 of 35 pages, which read like an active filter.
• Sort preference persists per tab via sessionStorage.
• Firefox build fix: offscreen permission is now emitted only for Chromium-class builds. Firefox uses the direct-Worker fallback; it rejected the unknown permission outright before.
• Local install: unpacked builds now mirror to built-files/chrome-mv3/ AND built-files/firefox/. New npm run sync:builds script refreshes both from .output/ between formal releases.

1.1.0 — 2026-04-30 — Branding + new icon set

• Replaced the default puzzle-piece icon with a custom magnifying-glass-over-stacked-cards mark rendered at 16 / 32 / 48 / 96 / 128 and a 1024 master source.
• Chrome Web Store assets: 440×280 promo tile and 1400×560 marquee, both matching the in-extension palette (#1d4ed8 → #0f172a).
• Store listing URLs updated to link to the public releases repo.
• No functional changes; search, chips, overlay, and indexing behave identically to v1.0.1.

1.0.1 — 2026-04-30 — CI pipeline verification

• Release workflow hardened: removed a redundant gh auth login call that fought with GH_TOKEN, dropped the noisy upload-artifact step that was hitting storage quota.
• No user-facing changes from v1.0.0. Reinstalling is a no-op.

1.0.0 — 2026-04-30 — First public release

Search

• SQLite FTS5 full-text search across titles, keywords, summaries, domains, and body text
• BM25 ranking with column weighting and frecency boost (visit count × recency)
• Highlighted snippets with Chrome text-fragment deep links (#:~:text=...)
• Chip-based AND filters via Tab key
• Smart prefixes: site:, domain:, in:title, in:body, in:summary, in:keywords
• Stopword skip (the, and, of, …) to reduce noise
• Narrow SELECT (no body column in result rows) for faster round-trips at scale
• Backspace-on-empty-input removes last chip
• Shake-on-reject for chips that parse down to nothing

UI

• Spotlight-style overlay on any page (⌘⇧K / Ctrl+Shift+K)
• Full-screen search tab (opens in dedicated tab)
• Popup with inline search, index-this-page button, and settings link
• Pinned / Recent section dividers in result list
• Per-result actions: Open, Pin, Forget, Exclude domain, More from domain
• 30-second Undo Forget with a restore toast
• Top-domain chips on the full search page
• Dark mode via prefers-color-scheme (CSS custom properties + color-scheme: light dark)
• React error boundaries on every entrypoint
• Options page opens in its own tab (not the chrome://extensions modal)
• Settings page has a filter input that hides non-matching sections

Indexing

• Runs at document_idle with MutationObserver + pushState/popstate hooks for SPAs
• Incremental re-index every 10 minutes on long-lived visible tabs when body text changes by ≥400 chars
• Coalesced batched writes (300ms window)
• “Index this page now” button in the popup, using on-demand content-script injection
• Automatic retention pruning (180 days or 20,000 pages by default, configurable)
• FTS body trimmed to 5,000 most-informative characters (full text still stored for export)

Privacy

• Zero network calls from the extension itself
• Password-field detection blocks indexing on any page with <input type="password">
• Sensitive-domain blocklist (banking, payment processors, auth)
• RFC 1918 + IPv6 ULA + link-local + .internal/.corp/.home.arpa/.local hostname skip
• Incognito never indexed
• Per-domain exclusion list (user-managed)
• Export / import as JSON
• One-click clear-all-memory
• chrome.storage.sync mirrors only preferences (never pages) for cross-device settings

Extension plumbing

• Manifest V3 with offscreen document for SQLite + SAH
• Firefox MV3 support via direct-Worker fallback when chrome.offscreen is unavailable
• On-demand content-script injection via chrome.scripting.executeScript so shortcuts and “Index this page” work on tabs that pre-date install
• Idempotent schema migrations (ALTER TABLE ADD COLUMN IF NOT EXISTS pattern via PRAGMA introspection)
• One-shot FTS-rebuild gate via meta.schema_version to recover from a prior FTS5 contentless-delete bug

Known limitations

• First index of a very long page can block content script for ~100ms on low-end devices
• FTS rebuild on large databases (>5k pages) takes 1-2 seconds on extension reload after schema bumps
• Firefox build is untested against AMO submission rules; works on about:debugging