Squirrel Release Notes

Notable changes to Squirrel: the archive itself and Nutshell AI summarisation. Customer-facing only - updates that change what you see, do, or rely on. Burrow has its own release notes.

Squirrel ships continuously rather than in numbered releases. Changes are gathered into fortnightly sprints, each dated to the Friday it closed, and grouped by the part of Squirrel they affect. Anything not listed under a heading did not change in that area that fortnight.

Sprint ending 2026-10-23

This sprint closes on Friday 23 October; the entries below run to 21 October.

Archive

• Settings display fix rolling out. Some users reported that a settings page showed values different from the ones they had set, although the settings themselves had been saved correctly. A fix is being deployed so that each page shows the values that are set.
• Site Archive Policies shows each site's address. Site names are not unique across a tenant, and the page showed only the name, so two sites called "Finance" could not be told apart. The address now sits under the name in the table, and a site with its own policy carries a Custom policy badge. The search box finds a site by its name or by part of its address, and the Restore and Archivable Files Calculation windows show the address under the name. See site settings.
• Picking a site in the other site windows is unambiguous too. In the Site Custom Policies and Excluded Libraries and Paths windows each result shows its address path, and the full address once a site is picked. The Excluded Libraries and Paths window now looks the site up by the one you picked rather than by name, so two sites with the same name no longer get their libraries mixed up. See site custom policies.
• Short site names can be found. Site search no longer waits for three characters before it starts, because some sites have names of only two.
• File History can search by user. Switch the search to By user and type a name or email address to see what one person restored: every folder restore they asked for, with its scope, size, status and when it finished, and every single-file restore, newest first. When more than one user matches, the page adds by-user and by-site totals. Narrow by site, folder or filename if you like. Restores run by the system or in bulk record no user, so they do not appear. See File History.
• File History shows the right times. Archive, delete, capture and restore times on the page were four to five hours early on a tenant whose database runs on a non-UTC clock. They are now shown correctly. See File History.

Sprint ending 2026-10-09

Archive

• Cost Allocation prices live and archived storage separately. The report had priced every site's live SharePoint size at the archive rate and never read archived data, so a site with nothing archived was charged and a heavily archived site was under-charged. Each group now has a live rate and an archive rate. The report shows live, archived and total cost per group, plus Saved by Archiving, and the group detail lists the cost by site. See Cost Allocation and how to charge storage costs back to departments.
• Cost Allocation shows ranked bars. The storage distribution was a pie with a slice per group, which was unreadable beyond a handful of departments and reused colours. It is now one bar per group, largest first, split into live and archived, with each group's storage, share and monthly cost, a hover for the detail and a click through to the group.
• Scheduled Reports is gone from Cost Allocation. Nothing sent those emails, so every action on the tab failed. Use Export CSV on the report and on each group instead.
• Storage Comparison matches your bill. The archived figure was a running total of what had been archived each month, which never subtracted restores and deletions, so it read higher than the storage you were billed for. It now shows the archived storage under management at its billed size. See analytics.
• Reports no longer time out on a busy archive. The monthly and yearly counts and volumes, the by-site CSVs and the file-type reports added up the whole archive each time they ran, and over a month or more of data they could run past two minutes and fail. They now read daily summaries prepared overnight, plus today's figures live. Each report shows which days the summaries cover, and warns if the overnight run is behind. The Large Files report lists the top 100 files. See reports.
• Site Archive Policies shows an Archived Size column. Beside Site Size, sortable, and showing a dash until a site's archived size has been worked out. It is recalculated daily. See site settings.

Nutshell AI

• A stub in a moved folder keeps its original modified date and editor. After a folder had been moved in SharePoint, a stub written into it lost the file's original Modified date and Editor when its summary was saved. It now keeps both.
• A renamed site is followed. A site renamed in SharePoint could not be found again by its old address, so stubs on it were never updated. Nutshell now finds the site under its new name.
• A SharePoint slowdown no longer costs a summary. When SharePoint throttled the tenant, errors during a stub update could lead to a finished summary being thrown away. A summary is now kept and the update tried again, and a file SharePoint keeps refusing is set aside after a number of tries instead of occupying the queue.
• Large folders no longer hold files up. A folder of more than 5,000 items could stop Nutshell checking its stubs, leaving every file in it waiting. These folders are now read in pages.
• Checked-out stubs are not listed as Locked. A stub checked out to the system in a library that requires check-out was listed on the Nutshell page as Locked. It is now skipped quietly.

Sprint ending 2026-09-25

Archive

• Site sizes read in the right unit. A site under a gigabyte was being rounded to "0 GB" on the Site Archive Policies page - on a large tenant that was most of the list. Sizes now show in MB (or "<1 MB") below a gigabyte, GB up to a terabyte and TB above, with the decimals kept, and sorting by size orders on the real value instead of tying every small site at zero. See site settings.
• The site table always reaches its Actions. The Site Archive Policies table keeps a minimum width and scrolls sideways in a narrow window, so the Edit, Archive, Restore and Preview buttons on the right cannot be squeezed off the page.
• A new tenant is not a fault. The Storage analytics page reads "No data collected yet" where a tenant has no billing history, rather than looking like a database error. The series simply starts at your first invoice. See analytics.
• The end-user restore page can wear your colours. Users reach it from a stub link, often without noticing they have left SharePoint. The palette is now set per tenant and applied on the server, so the page renders in your colours immediately rather than flashing the stock ones first, and the browser tab reads "SharePoint Restore". Ask support to set it up. See end-user restore.
• File History reads restores correctly. A restored file shows the restore as its most recent event, and a restore too old to carry a date says so rather than showing a misleading one. See File History.
• The Archived Files total sits with the page title. The count has moved off the pill beside the heading and onto the header subtitle line, which is the shape the Orphaned Files page already used. It still carries the note that the figure is recalculated nightly rather than counted live, now as a tooltip on the number. See Archived Files.
• The Excluded Libraries and Paths window now lists a site's libraries. Picking a site left the library dropdown empty, which made the screen difficult to use for the thing it exists to do - excluding a library before anything in it is archived. It offered only libraries that had already been archived, if it offered anything at all. It now reads the site's actual library inventory, fills in immediately rather than after a pause, suggests folder paths under the site the way Archive and restore by path always has, and works on sub-sites. See excluded libraries and paths.
• The Archive button in SharePoint no longer hangs for a signed-out user. Clicking Archive on a file while not signed in to the portal sent you to sign in and then landed you on a page that sat on "Processing archive request..." indefinitely, with no request actually in flight - indistinguishable from an archive that had failed. The file and site references now survive the sign-in and the request goes through. A link that is genuinely missing those references says so rather than appearing to stall. The same fix covers the folder archive and folder restore pages.
• Choose which buttons get deployed. Two switches on the Button Deployment page decide what the nightly pass puts out, both on by default. Deploy buttons to folders off means buttons reach files only, so a user cannot action a whole folder in one click. Deploy Archive buttons off deploys Restore buttons only - users can recover their own archived content but cannot archive anything on demand, which keeps archiving decisions with your policy and your administrators while leaving self-service recovery intact. Both are greyed out while deployment itself is off, and whatever you chose is remembered.
• Button Deployment reads in the order you set it up. On the Button Deployment screen, the settings that decide which buttons are deployed now sit inside the Enable Button Deployment panel rather than beside it, so the page reads top to bottom: turn deployment on, choose the buttons, then choose where they go. Those settings also grey out while deployment is off, instead of showing as enabled while doing nothing. Whatever you had chosen is remembered and comes back when you turn deployment on again.

Nutshell AI

• Detailed now sends the whole document to the AI. Reading a document and giving all of it to the model are two different steps, and they had quietly diverged: a relevance filter sat between them and passed on about 60% of the text, so 40% of a document never reached the AI in Detailed mode even though every page had been read. Detailed now skips that filter entirely. If you chose Detailed for completeness, it is now doing what you picked it for. See summary types.
• Summaries are laid out properly in the stub. The AI writes in markdown, and stubs previously understood only part of it - sub-headings and bullets arrived as literal asterisks and ran together into one paragraph. Headings, lists and bold now render as formatting, and the parts of a long Detailed summary are numbered so it is clear it continues rather than repeats. See stub files.
• Fewer summaries that are not summaries. Answers that come back as a data dump rather than a summary are now rejected and retried, documents whose text layer is garbled are sent to OCR instead of being summarised from nonsense, and Extract handles form-shaped and invoice-shaped documents rather than forcing an outline onto them.
• Spreadsheets are profiled rather than narrated. A table is not prose, and summarising it as though it were produced a vague paragraph about a precise file. Nutshell now reads a workbook and states what is in it - per sheet, under its real sheet name: how many rows and columns, what each column holds, which values dominate it and how often, what range the numbers cover. Those figures are counted from the file rather than written by the AI, so they are right either way, and one AI pass turns them into prose with the counted facts underneath. In Extract, where no AI runs, the profile is the summary. It is also far quicker: a 1,700-row permissions export took dozens of AI passes and produced near-identical paragraphs, because every slice of a table looks like every other slice. It is now a single pass. See supported file types.
• Large workbooks are read instead of refused. Worksheets above a size limit were skipped outright, and on workbooks built in certain ways no cell was read at all - so a stub could read "The spreadsheet is too large to be loaded into the system" about a file that is a perfectly ordinary readable table. The size limit is gone, every way a workbook can store its text is now read, and a very large workbook is read to a budget instead: the stub describes the opening portion of its largest sheets and says so when it does, so a sample is never presented as the whole file. The workbook itself is archived intact and restores whole.
• Content missing from long summaries. Several separate faults could drop whole sections of a long document from its stub, in each case silently. A summary the model wrapped in a code fence was discarded as though it were a script. A deliberately short answer was rejected with no fallback, despite one existing. A reply that merely quoted a phrase like "please provide proof of delivery" was mistaken for the model refusing the task. And PDFs that position every letter separately - which arrives as F a l t a s N o M e s - corrupted everything downstream, at worst losing two thirds of a document. All four are fixed, and a chunk that genuinely cannot be summarised still falls back to a plain extract rather than vanishing.
• Registers and path lists are summarised as what they are. A spreadsheet or export that is a list of records with no sentences in it was being given the instructions for writing prose, and a small model asked for prose about content that contains none tends to answer the instructions instead of the content. That kind of content now gets its own short, direct instruction, which is both faster and more accurate.
• Unreadable scans are caught more reliably. The check that stops a badly scanned document being summarised as though it were clean could be skipped entirely when a scan was so poor that almost no words were recognised - so the very worst scans were the ones that slipped through. A scan with almost nothing legible is now treated as failed rather than as unmeasured, and a short but clean page still passes.
• A document that fails to process is no longer discarded. A defect in a processing script could make large spreadsheets fail instantly and, after retries, drop the file and its record from the queue without a usable explanation - the file was never summarised and nothing said why. Workers now refuse to start in that condition and name what is wrong rather than failing anonymously and discarding work. Archived content was never affected; what was lost was the summary.
• Extract keeps far more of what matters. Extract was trimming every section to a fixed 3,000 characters and keeping whatever came first, so the tail of a section - and the dates, amounts, part numbers and warranty terms in it - was simply dropped. The budget now scales with the document, and it is spent on the lines a reader needs rather than the opening ones: a line carrying a date, a sum of money, a percentage, a reference code or obligation language (shall, must, warranty, penalty, deadline) is kept ahead of generic prose. Measured over eight varied documents, the share of distinct facts that survive into the extract went from 28% to 72%. On a 940,000-character equipment manual it went from 15% to 71%, in about 3% of the text.
• Extract reads properly, not just completely. Four smaller fixes in the same area. Passages are no longer cut mid-sentence, and an extract is never returned empty. Outlines drop noise bullets, sections that turned out to be nothing but scanning artefacts, and headings that are really a fragment of a sentence. Lines that a PDF wrapped across two visual lines are put back together before being bulleted, so one sentence is no longer served as two. And the closing dates-and-amounts line is harvested only from lines that actually read cleanly, rather than presenting numbers scraped out of OCR noise as fact.
• Extract no longer explains itself. The banner that opened an extract - naming how the passages were chosen, the source character count, and that no AI was involved - has been removed. It described the pipeline to a reader who did not ask and could not act on it. An extract is now the document title, the key facts line where there is one, and the passages.
• Lite reads as much as Extract, and polishes all of it. Lite is meant to be Extract plus one AI pass, but it had been selecting a smaller extract to begin with, and then asking the model to restate up to 20,000 characters in a single short reply - an eight-to-one squeeze, which is where the facts it had just worked to keep were being lost. Lite now starts from the identical extract, and a long document is polished in as many passes as it takes, split on the headings the document already has. On that same equipment manual the summary went from retaining none of its 94 facts to retaining 20. Most documents are unaffected: 86% still take exactly one AI call.
• Detailed stops repeating itself. Detailed assembles its summary from several passes over a long document, and each pass was introducing the document again - one stub opened by saying what the document was nine separate times - and restating points an earlier part had already made. Only the first part now introduces the document, a continuation that reintroduces it has that framing removed while its facts are kept, and a sentence an earlier part already said is dropped before the parts are joined. The detail itself is untouched, which is the point of the type.
• Summaries keep the structure the AI wrote them in. A cleaning step had been flattening bullet lists, numbered steps and bold phrases into a single run of prose, which reads badly on manuals and specifications, and stripping lines it judged to be model commentary - including one that named a certification standard and a calibration seal. Both are gone. A specification that comes back as a list now reaches the reader as a list, and measured on a representative reply the facts that survive went from 9 of 11 to 11 of 11.
• Scanned documents are read in their own language. Every scan was being read with the English model, whatever language it was in. A Russian contract came back as "Tpunoxenne 2.9 / Koutpaxt Ne 74-E-38-1" and a Japanese patent as "{Oevt4 Sid, ARMS OAM" - not bad scans, the wrong model. Nutshell now detects the script a scan is written in and reads it with the matching model, and the full set of language packs ships with the platform. Measured on the OCR engine's own confidence score, out of 100: the Russian contract went from 26 to 96, the Japanese patent from 25 to 94, and English was unaffected. Documents summarised before this keep their old text until they are re-summarised.
• A scan too poor to read now says so. Where the OCR engine reports low confidence across a document, the stub states plainly that the document is a scan whose text could not be read reliably enough to summarise, and lists the lines that did read clearly - the letterhead, dates and reference numbers that let you recognise the file. Previously that document was summarised anyway: in Extract you got a page of noise, and in the AI types you got fluent prose invented on top of noise, which is worse because it reads as though it is true.
• Nothing is dropped when the model refuses a chunk. A small model occasionally answers about the task rather than doing it. That reply is correctly rejected, but the chunk it belonged to was then quietly left out of the summary - one seven-chunk document lost chunks four and six, nearly a third of it, with nothing saying so. A rejected chunk now falls back to a plain extract of its own text, so the content reaches the reader either way.

Sprint ending 2026-09-11

Archive

• File History - one search for what happened to a file or folder. A new page that answers “where did it go?” in one place instead of three. Give it a site and a folder, a file path or a filename prefix, and it merges the archive, recycle-bin capture, the processing-site inventory and the restore history into a single per-file answer: who deleted it, whether Squirrel captured it, whether it is archived, orphaned or already restored, and where the copy is now. It leads with a plain-English summary - including saying outright when Squirrel did not delete the files, which is the first thing people assume. Each row links to the page that can recover it, carrying the filename so the search is already run when you arrive. See File History.
• Recycle Bin search finds things it used to miss. Two fixes that customers would have experienced as “the file isn’t there” when it was. Searching with a file’s extension never matched, because a captured copy is stored under a stamped name - matching is now on the base name, so either form works. And only the currently configured processing site was searched, so anything captured before that setting changed was unfindable; search now covers every processing site Squirrel has used. Recovery was never affected either way - it resolves a copy’s location from the record, not the current setting - so nothing was ever actually lost. See Recycle Bin capture.
• Log entries stay on one line. The Log Viewer now scrolls sideways rather than folding a long entry across several rows, which keeps the timestamps aligned and a run of entries scannable. A Wrap long lines switch turns folding back on when you would rather read one long entry in full. The scrolling is contained to the log panel, so the rest of the page stays put.
• The archive overview reworked. Storage Distribution is now two lines on one scale by invoiced month - what remains in SharePoint against the archived running total - which shows one falling as the other rises. The bottom row carries five cards rather than three, and each metric tile shows its own change against the previous period instead of a sparkline, with the fuller trend a click away. See the dashboard.

Nutshell AI

• Two new summary types: Lite and Extract. Nutshell now offers five types rather than three. Both new ones read the whole document, pick out its key passages automatically, keep the document’s own headings where it has them, and finish with a line of the dates, amounts and reference codes found in it. The difference is the last step: Lite hands those passages to the AI for one short pass, so the result reads as written prose at well under half of Brief’s processing per document. Extract skips the AI entirely and shows the passages as they appear, labelled as an extract - for maximum throughput, or where policy says no AI at all. The temperature slider greys out while Extract is selected, because that type never calls the AI. See summary types.
• Moving a stub no longer costs you its summary. A summary is written into the stub shortly after archiving, so a user can move or re-file it in between. Nutshell now follows the stub and corrects the archive path. Previously it tried to drag the file back to the recorded path, which failed outright once the old parent folder was gone - the file was retried a few times and then given up on, with no summary ever written. See stub files.
• Non-English text in PDFs is no longer garbled. Cyrillic, Turkish, accented Latin and CJK characters extracted from PDFs were arriving as mojibake and going into summaries in every mode. Fixed at the extraction step, so affected documents summarise correctly from now on. Documents summarised before this keep the old text until they are re-summarised.

Sprint ending 2026-08-28

Archive

• A refreshed admin portal. Both the archive dashboard and the security dashboard moved to a new visual language - new typography, softer surfaces, and consistent dialogs and tables throughout. Nothing moved that you need to relearn; the layout and the navigation are unchanged.
• Jump to any page with Ctrl+K. Press Ctrl+K (Cmd+K on a Mac) anywhere in the portal for a Go to page box: type a few letters, arrow through grouped results, Enter to go. It is the quickest route to the analytics pages, which sit a level below the main navigation. A button in the top bar does the same for people who would rather click.
• Being refused a page now explains itself. Access is by group membership, so you can sign in successfully and still be refused a particular page. Instead of a silent redirect you get a screen naming the page, a button that copies the support address to your clipboard, and - usefully - the attempt is logged, so support can see it without asking you to reproduce it. See logging in.
• Site actions say what they will do before you click. Hovering the Archive, Restore or Preview action on a site in Site Archive Settings now states its scope. Worth reading: a site-wide Archive takes every file regardless of your age thresholds, while Preview counts what the policy would take and changes nothing.

Nutshell AI

• Brief mode is actually brief. Brief had been given the same writing instructions as Standard, so although it read less of the document its summaries came out at the same length and level of detail. It now produces a genuine executive summary that leads with the single most important point. If you evaluated Brief before and found it indistinguishable from Standard, it is worth another look.
• Boilerplate is dropped rather than summarised. Disclaimers, forward-looking-statement notices and confidentiality, copyright or trademark notices are omitted from summaries entirely rather than paraphrased. A summary that spends its closing sentence noting that a document “includes a disclaimer” has wasted it.
• Large backlogs drain faster. The surge capacity that clears a build-up was capped low enough that it stopped scaling on backlogs of a few hundred files. The ceiling is substantially higher, so a catch-up run after onboarding a large site clears at a higher sustained rate without anyone intervening.

Need help? support@smikar.com.