Skip to content

Changelog

User-visible changes to Retrace, the Sauvrn extension.

This page logs behavior you can actually see — what gets captured, how Find and Ask work, what the window and the Space manager do, what changed in onboarding. It does not list internal refactors, build-system housekeeping, or test-only updates.

The version number in the extension's Settings page (and in chrome://extensions under "Details") is the source of truth for which build you have. If it is ahead of the newest heading below, you are on a build whose changes are not written up here yet.

Unreleased

Changed

  • The rail and the tab switcher are gone. Retrace is now two surfaces: the Retrace window, which opens over the page you are on (Control+J, or the Retrace icon in your toolbar) with Find, Ask and Threads, and the Space manager in Chrome's side panel (Control+K) with your Spaces and their tools. Nothing is drawn on the page's edge and the page is never pushed aside. Opening a Find, Ask or Threads result still lands you on the passage, in a new tab or the one you already have open. The rail's own settings (its mode and its notch) are gone with it. This supersedes every rail and switcher entry below.

Fixed

  • Removing the model removes the model. Settings → Remove the model now asks the desktop app to delete the staged weights — every file in the pinned weight manifest, its part-files and verification markers, and the download's resume state — and the app then restarts itself so nothing is still holding a copy in memory. It reports which files went, or tells you it could not and leaves them alone. Your captured memory is untouched either way, and the next time Ask needs the model it downloads again. This supersedes "Clear-model reset" under 0.1.10 beta below, where the same button only tore down a browser-side cache and freed nothing on a fresh install.

Security and privacy

  • Usage counts are ON during the alpha, and the switch is gone. This supersedes "Usage counts are opt-in and off by default" under 0.1.10 beta below, which described the shipped behaviour at the time and no longer describes it. Settings → Account now discloses the counts instead of switching them: once a day Retrace sends one row of content-free counts — which tools you clicked, searches and how you narrowed them, pages reopened and whether they landed, Asks and their outcomes, pages remembered, spaces, errors and failed round trips, and speeds as coarse bands. Never page content, titles, URLs, queries, answers, the names you give spaces, or your theme. The row rides your license token, so it is tied to your account rather than anonymous, and rows are kept 90 days. After the alpha the switch returns and the counts go back to off by default.

0.1.10 beta

The first publicly-installable beta. Everything below is "what's actually shipping in this version" unless explicitly marked otherwise.

New

  • Visible version stamp. The extension version is now shown in the extension's Settings page — in the footer, in the Status panel, and in the feedback form's preview — so you can tell us exactly what build you're on when you write to support.
  • In-product feedback. A "Send feedback" form is built into the extension. Submissions land in our support inbox. You can tick on a diagnostic payload — extension version, build, browser, OS, and the hostname of the page you're on — and the form shows you exactly what will be sent before it goes; leave the boxes unticked and only your message is sent. Tokens you paste accidentally are scrubbed before anything leaves your machine.
  • Movable Find / Ask overlay. The in-page overlay window for Find and Ask can be dragged to wherever it doesn't block the page you're reading.
  • Onboarding "reload open tabs" button. During first-run, you can re-trigger capture on the tabs you already had open so your existing session shows up in Find without needing to revisit each tab manually.
  • First-run model download progress. A monotonic progress bar shows the one-time AI-model download during setup. It never regresses and tells you when the download is done so you know Ask is ready.
  • Encryption at rest. The sensitive bodies of captured pages and capture events — text, title, summary — are encrypted on disk using a non-extractable WebCrypto key. The key never leaves your browser and you never have to manage it. The fields that back the local search indexes stay in the clear on your own disk by design: the visit trail (URL, hostname, timestamp) and the search embeddings. One case is exempt: if your browser profile can't open the encrypted store, Retrace keeps your memory in unencrypted browser storage rather than dropping it, and pages saved during that fallback are not encrypted and are never swept back. Settings → Manage memory prints a Limited storage line while that is the store answering.
  • Clear-model reset. Settings gained a dedicated "Clear model" reset tier, which left your captured memory alone. What it actually deleted in this build was the leftover in-browser model cache from installs that predate the in-browser engine's retirement — a fresh install has none, so on one the button freed nothing, and it never reached the model weights the desktop app downloads. Superseded by "Removing the model removes the model" under Unreleased above.
  • Google Workspace capture. Retrace now captures the body of Google Docs, the cell values of Google Sheets, and the text on Google Slides while you read them. Slides remembers which slide you were on.
  • Workspace comment capture. Comments on Docs, Sheets, and Slides are captured opportunistically when they're rendered visible.
  • Conversational Ask. Ask redirects when you change topic mid-thread instead of repeating the previous answer. Robotic filler ("Based on the provided context...") is stripped before the response is shown.
  • Pricing page. app.sauvrn.ai/pricing is live, wired to checkout.

Improved

  • Rail stability. The rail mounts idempotently — its width stays stable, and the Find/Ask overlay persists across in-page navigation.
  • Find result quality. Find now dedupes results, scopes the tool filter to the active Space, and matches against open-tab text. Snippets window onto the matched span (including semantic matches), so you see why a result is a hit.
  • Find honesty. When Find falls back to keyword-only because the semantic model is still loading, an inline indicator tells you so. No silent quality changes.
  • Ask transparency. A phase line is now shown during inference — "Reading your activity…" while Ask gathers evidence, then "Thinking…" while it composes — so a slow answer is never a silent answer. Citations on grounded answers come from the verifier, not the model — what you see is what the model actually grounded against.
  • Find / Ask hotkey. The Find hotkey now toggles open/close and switches across modes — no more dead-end overlays.
  • Switcher actuation. Opening a result via the switcher actuates the real tab and previews on each step. Duplicate tabs can be closed inline. The back-arrow returns to the results list.
  • Rail visuals. Lighter rail background, defined right edge, drop shadow onto the canvas; tool icons render on first paint, with an initial-tile fallback if an icon fails to load. Space-logo hover swaps to a panel button without container-shape changes and stays keyboard-accessible.
  • New-tab loader. The "open in new tab" spinner self-clears, so it can never hang permanently. Tool launches always open a fresh tab instead of clobbering the source.
  • Onboarding consistency. The tool catalog is sourced from a single master registry, with an empty-state when nothing matches your search.

Fixed

  • Cold-start hang on Google Docs. The loader no longer gets stuck waiting on a missed signal; the switcher card label is also correct now.
  • Ask never silent-cards. The cold engine now hydrates on Ask, so questions can't fail silently when the engine has gone idle.
  • License revoke under load. Revoking a license is wrapped in a transaction with compare-and-swap so concurrent revokes can't race.
  • Onboarding download bar. No longer regresses on retry.
  • Re-anchor accuracy. Reopening a captured Google Slides source lands you on the right slide; Docs/Sheets re-anchors fall back to a reader-peek when the live DOM has changed. The Slides filmstrip is excluded from live-DOM extraction so it can't poison the anchor.
  • Grounded Ask citations. Citations resolve by containment, not prefix, so source cards match the passage Ask actually used.
  • Redo-onboarding on CSP hosts. Falls back to a dedicated tab when the host page's Content Security Policy blocks the overlay.
  • Settings error messaging. The redo-onboarding error surface now shows the real error string instead of [object Object].

Security and privacy

  • Mechanical injection screen. Attacker-controlled text from captured pages can no longer hijack Ask output — the prompt-injection check is a mechanical filter, not a model judgment, and it never renders the attacker's text.
  • Redactor coverage. The on-device redactor now catches modern secret-token formats (long-lived API keys, JWT-like strings).

  • Usage counts are opt-in and off by default. Settings → Account has a "Share usage counts" switch. Leave it off and the extension sends no usage data at all — not a single request. Turn it on and it sends content-free counters and timings (never page content, titles, URLs, queries, or answers) to api.sauvrn.ai, with your license token attached, so the row is tied to your account. Turning it back off discards anything not yet sent and records the opt-out with us. The counters the extension keeps for itself stay on your machine either way.

    Superseded 2026-09-06 — do not read the paragraph above as current. It described 0.1.10 beta and no longer describes the shipped build. In the alpha builds telemetry is on: there is no switch, and Settings → Account discloses the counts instead of asking for them. What leaves is one row per day to api.sauvrn.ai, carrying your license token, so it is tied to your account rather than anonymous. The daily row carries the calendar day (a date, never a time), whether you used Retrace at all that day, how many times you clicked each supported tool (tool names from our own fixed list — anything else is counted as "other", never named), how many searches you ran and whether you narrowed them by time or by tool (a yes or no, never the range or the tool), how many searches ended with you opening a page, how many remembered pages you reopened and how many landed where they were meant to, how many Asks you asked and how many were answered or declined, how many pages were captured, how many spaces you used and created, those actions added into one total, how many errors happened by kind and how many Find or Ask round trips failed, how long searches and Asks took as coarse bands computed on your device, and roughly how many pages you have remembered in total, as a range. Separately: once per install, which setup steps you reached and the first time you clicked a tool, searched or landed on a reopened page, each with a coarse elapsed bucket; a crash marker naming which part of Retrace broke and what it was doing, never the error text; a bare once-a-day signal that carries no fields at all; and, once, a one-boolean note of the setting itself with no events attached. Never page content, titles, URLs, queries, answers, the names you give spaces, or your theme. Rows are kept 90 days and then deleted. After the alpha the switch returns and the counts go back to off by default. The Unreleased entry at the top of this page is the current one.

Earlier builds

Earlier builds were internal-only. Public history starts at 0.1.10 beta above. If you're upgrading from an internal build, the version stamp in Settings is the source of truth for which behavior you have.

How to read this changelog

  • New is behavior or surfaces that did not exist before.
  • Improved is behavior that existed but works better now.
  • Fixed is a defect that produced a wrong result or a stuck state.
  • Security and privacy is anything that changes the threat model or the data that crosses the network. Always worth reading.

If something behaves differently than you remembered and you can't find it here, that's a bug in the changelog — please tell us via the in-extension feedback form.