Release notes

v0.11.262026-08-10

Large files are much faster to add and draw

Two hidden costs are gone.

Adding a file no longer copies it. EasyAnalysis now points at the file where it already is, so a 500 MB file costs nothing to add and a multi-gigabyte one costs nothing either. (Files you upload through a browser are still copied, because those are temporary.)

Drawing a large raster no longer reads the whole thing. The map now reads only the detail it needs to show — about 13× faster on a 500 MB file, and the difference grows the larger the file is.

v0.11.252026-08-10

Fixed: Add Data ▸ From file did nothing

In the workspace, Add Data ▸ From file… never opened anything. It now works, and behaves exactly like the Add Data button in the panel on the left.

See script is now a button on the analysis itself

Every analysis panel has a See script button at the bottom. Previously this was only in the Help menu, where nobody would think to look for it.

v0.11.242026-08-10

Adding data is one button again

There is a single Add Data button. It opens files straight from your computer, with no copying, so size does not affect how long it takes. The separate upload button is gone — there is nothing to choose between.

See the code behind more analyses

Help ▸ See script for this analysis now shows the R code for screens that were previously left out, including Linear regression.

The script shows the model call as it was made. It notes that it does not include preparation the screen did first, such as dropping rows with missing values, so you know what to check before re-running it.

v0.11.232026-08-10

Fixed: choosing a large file gave no feedback at all

The message that was meant to appear as soon as you pick a file — "Reading 1 file (3.7 GB)…" — never actually appeared, because of a bug in how it was wired up. It now works.

For very large files, use Add data from disk rather than uploading. It opens the file straight from your computer with no copying, so a multi-gigabyte file opens as fast as a small one. Uploading a file of several gigabytes through the browser can still stall.

Layer list is cleaner

The coloured dot before each layer name is gone — the row already tells you the layer type.

v0.11.222026-08-10

How to cite EasyAnalysis, inside the app

Help ▸ How to cite… now gives you APA and BibTeX for the exact version you are running. It is also in the in-app documentation as section 10, so the guide in the app matches the one on the website.

Previously this existed only on the website, and its version number had fallen eleven releases behind — which matters, because that text ends up in published work.

Credit

EasyAnalysis is created and developed by Tim Casanda Gibson, now credited in Settings ▸ Acknowledgements alongside the University of Eastern Finland.

v0.11.212026-08-10

Add data from disk — no more waiting for uploads

There is a new Add data from disk button at the top of the Datasets panel. It opens a normal file browser and reads the file straight from your computer.

Nothing is copied, so file size no longer affects how long adding data takes — a multi-gigabyte raster opens as quickly as a small one. This is the recommended way to add data; the old upload button is still there as Upload instead.

It also means a project remembers your actual file. Previously an uploaded file lived in a temporary location that was gone by the next session.

v0.11.202026-08-10

CSV files load dramatically faster

Reading a CSV is now roughly 60 times faster. A large file that previously froze the app for about fifteen seconds now loads in well under a second.

Column names, types and values are unchanged — this is the same data, read a faster way.

v0.11.192026-08-10

Groundwork for making large files fast

Internal only — nothing changes in the app yet.

We measured where time actually goes when loading data, and added a repeatable performance test so that future speed work can be proven rather than claimed. The measurements confirmed that opening large rasters and drawing them are already fast; the wait is the file being copied through the browser.

The fix being built is to let you point at a file already on your computer instead of uploading it.

v0.11.182026-08-10

GeoPackages with several layers now load all of them

A GeoPackage containing more than one layer was loading only the first, and the rest were discarded with no message anywhere in the app — a file with two layers and five features arrived as one layer and three features, looking like it had worked.

All layers now load, each named after the file and the layer inside it, with a message saying how many arrived. Reopening a project restores the right ones.

See the code behind any result

Statistical analyses now have a See script button. It shows the R code that produced your result — with the columns and settings you actually chose — and you can download it as a .R file and run it yourself.

It is the real code, not a description of it, so a result can be checked, repeated or cited by someone who was not sitting at your machine.

v0.11.172026-08-10

Faster startup

The app was spending about 1.6 seconds of every startup preparing native file dialogs that were removed some time ago. That work is gone.

On slow loading of large files

We measured where the time actually goes. Opening a large raster takes well under a second, and preparing it for the map takes under one more — the app itself is not the slow part. Nearly all of the wait is the file being uploaded through the browser.

The real fix is to let you point at a file already on your computer instead of copying it through the browser, and that is being worked on.

v0.11.162026-08-10

There is no upload size limit any more

Files above about 3 GB were being rejected, and — worse — rejected silently: the file simply never appeared, with no error anywhere in the app. A 4 GB raster vanished exactly this way.

The limit is removed entirely. There is no size cap on what you can add.

Large uploads now say they have started

Adding a very large file takes minutes, and until now nothing said so, which made it look like nothing was happening. You now get an immediate note — "Reading 1 file (3.7 GB)… large files can take a while." — as soon as you choose the file.

v0.11.152026-08-09

Search results say where a tool comes from

When a search turns up a WhiteboxTools tool you have not enabled, the results are now headed WhiteboxTools - not enabled yet, so it is clear whose tools you are being offered before you enable anything.

v0.11.142026-08-09

Tidier menus

R Console now has its own place in the menu bar. It used to appear in two different menus and behave slightly differently depending on which one you used.

The More menu is gone. Everything in it — Documentation and References — is already in Help, so it was one more place to look for no benefit.

Results say which engine produced them

When a WhiteboxTools tool finishes, the confirmation now names it, and the layer itself remembers where it came from.

v0.11.132026-08-09

Search now finds WhiteboxTools tools you have not enabled

Searching for a WhiteboxTools tool used to return nothing, because the search only looked at tools already in your menus. It now searches the whole WhiteboxTools catalogue as well. A tool you have not enabled appears under Not enabled yet with an Activate button — one click enables it and opens it.

If the catalogue has not been indexed yet, the search now tells you and offers to do it, instead of quietly finding nothing.

A few WhiteboxTools tools cannot run here, because they write their result next to the input file rather than returning a layer. Those still appear in search so you can see they exist, marked Not supported, rather than offering a button that would do nothing.

You can now see when a WhiteboxTools tool is running

Any tool that comes from WhiteboxTools shows a badge in its panel naming the engine and the underlying tool, so it is clear which results came from where.

v0.11.122026-08-09

Plugins now opens as a dialog

Choosing Plugins used to replace whatever you were working on with a full screen. It now opens as a pop-up, the same way Packages does, so your map and tool panel stay where they were. Everything is in the one dialog: enable WhiteboxTools, index its tools, search them, and switch individual tools on or off.

v0.11.112026-08-09

Plugins now has its own menu

It was buried under Analysis → More, where nobody would look for it. It now sits in the top menu bar next to Packages, which is where you would expect to find it.

v0.11.102026-08-09

Enabling a tool now works straight away

Turning on a WhiteboxTools tool used to require reloading the page before you could use it. It does not any more — enable it and it appears in the tool list immediately, the same way an installed package is available as soon as it is installed.

v0.11.92026-08-09

New: a Plugins screen

Under More → Plugins. EasyAnalysis can now offer tools from other open-source projects, starting with WhiteboxTools and its 484 algorithms for hydrology, terrain, LiDAR and image processing.

Nothing is switched on for you. You enable WhiteboxTools yourself, and then enable individual tools one at a time — so the app only carries what you actually use. The screen credits the people who wrote it: Prof. John Lindsay, with the R package by Qiusheng Wu and Andrew Brown.

Search finds a tool even before you enable it. Type what you are looking for, and if it is not enabled yet you can switch it on right there in the result.

Building the tool list runs in the background, so you can keep working while it does. Start with Index the common tools — 31 tools covering the usual hydrology, terrain and LiDAR workflows — or index all 484 if you want everything.

One thing to know: after enabling a tool, reload the page to use it.

v0.11.82026-08-09

Groundwork for optional tool packs

Internal plumbing that lets EasyAnalysis offer tools from other open-source projects — starting with WhiteboxTools and its 484 spatial algorithms — without slowing the app down.

Nothing changes for you yet, and nothing is switched on. Tools from an external project will stay off until you choose to enable them, they will be credited to the people who wrote them, and you will be able to enable them one at a time. Searching will find a tool even before it is enabled, so you can turn it on at the moment you need it.

v0.11.72026-08-09

The WhiteboxTools hydrology tools could not run on a new machine

Fill depressions and Flow accumulation use WhiteboxTools, which comes in two parts: a small R package, and the WhiteboxTools program itself — a separate download of about 90 MB. The installer set up the package but never the program, and the app only checked for the package. So both tools looked available and then failed with a confusing file error.

The installer now downloads the program too, once. If it is still missing — on a restricted network, say — the tools now say exactly that, and how to fix it, instead of failing obscurely.

v0.11.62026-08-09

Drag layers to reorder them

Grab the handle at the left of any layer row and drag it up or down. The order you set is the drawing order on the map — put a vector outline above a raster, or one raster above another. Right-click a layer for Move to top and Move to bottom, which are quicker in a long list.

Your order is saved with the project, so it is still there when you reopen it.

The layers panel now reads the way every GIS reads: the top row is the top of the map. It used to be the other way round — the list was effectively upside down, so vectors appeared at the bottom of the panel while drawing on top of everything. Nothing looked wrong while the order was fixed, but it would have meant dragging a layer "up" pushed it down.

Your existing maps look exactly the same. Only the panel's reading order changed; nothing restacks.

The basemap stays pinned at the bottom and cannot be dragged — it is background tiles rather than one of your layers.

v0.11.52026-08-09

The attribute table now has proper window controls

Reading a wide table in a short strip at the bottom of the map was cramped. The table now has minimise, maximise and close buttons in its header:

• Maximise fills the whole map area, which is the point — a full-size table instead of a letterbox.
• Minimise collapses it to its header.
• Close hides it; Attribute table in the map menu brings it back.
• Drag the header up or down to set any height in between. The header has always shown a resize cursor; now it actually resizes.

Your choice also stays put. Previously, collapsing the table and then selecting a layer, toggling visibility or changing the basemap made it spring back open, because the panel was rebuilt from scratch each time and forgot its state.

v0.11.42026-08-09

The cross-validation warning now looks like a warning

The message that says part of your data was left out of a cross-validated score was correct, and sat in the right place — but it was rendered in the same faint style tables use for ordinary labels. A statement that the number next to it is unreliable read like a footnote.

It now appears as a highlighted chip beside the score, in a colour that adapts to whichever theme you are using. It still sits next to the number rather than somewhere else on the page, because that is the only place it will actually be read.

v0.11.32026-08-09

Error messages now stay on screen until you dismiss them

When something failed, the app explained why — and then deleted the explanation a few seconds later. Almost every error message in the app behaved this way.

Those messages are usually the most useful thing on screen: they name the cause and what to do about it. Losing them meant you were told how to fix the problem and then had it taken away before you could act.

Error messages now remain until you close them. If the same error happens repeatedly it updates in place instead of stacking up. Ordinary confirmations ("Raster added to the project") still fade on their own, since those are meant to be glanced at.

The Data Quality pop-ups are gone

Selecting a dataset raised a stack of warnings about missing values, duplicate rows and similar issues — every time, one pop-up per issue. Because they appeared whenever a dataset became active rather than only when it was first loaded, simply clicking between datasets replayed warnings you had already read.

They no longer appear. The underlying checks still exist and will return somewhere you can open when you want them, rather than as an interruption you cannot decline.

v0.11.22026-08-08

Cross-validation accuracy could be reported too high

If you have recorded a cross-validated accuracy from Logistic Regression or Classification, please re-run it. On some datasets the figure shown was optimistic.

Cross-validation splits your data into parts, fits the model several times, and pools the results. When one of those fits could not be completed, the app used to carry on quietly:

• Logistic Regression left those rows out of the pooled result, so the accuracy was calculated from less data than the label suggested — but still presented as a full cross-validation.
• Classification was affected more seriously. It fits one model per class. When a class could not be modelled at all, it was still allowed to compete for the prediction, and could even win it — so some predictions came from a model that never existed, and those counted towards the accuracy as though they were real.

This matters because fits do not fail at random. They fail on the awkward parts of a dataset — a rare category, an unusual split — which are exactly the cases a model finds hardest. Leaving them out made results look better than they were.

Now:

• A class that could not be modelled takes no part in the prediction.
• A row that no model could handle is left unscored rather than guessed, and is excluded from the accuracy instead of being counted as a mistake.
• Anything left out is stated next to the number it affects, for example: "Incomplete: 1 of 5 folds could not be fitted, so 18 rows (20% of the data) are NOT included."
• A clean run shows no such message, so nothing changes when everything fits.

Mixed-effects models already handled this correctly; their wording is now clearer about how many rows were used.

Time series: failures now explain themselves

Seasonal decomposition said only "Decomposition failed" — it now shows the reason. The stationarity test used to print nothing when it could not run, which made a failed test look like one the screen never offered; it now says it could not be computed.

v0.11.12026-08-08

Fixed — the assistant no longer drops figures without saying so

When the Co-Analyst ran a model for you, any figure it could not compute simply vanished from its answer. Ask for a mixed-effects model and the R² could be missing with no indication why; ask for an ANOVA and the post-hoc comparison could be absent as though it had never been requested.

The assistant answers from what the analysis hands it, so a figure that disappeared silently left it unable to tell you it was missing, let alone why. Those figures now come back with a plain explanation instead — the model is still reported in full, and the answer states which figure could not be computed and the reason.

Affects the linear model, ANOVA post-hoc comparisons, mixed-effects models and random forest.

Changed — the reference page no longer names internal code

The Methods reference described model-quality figures by naming an internal function. It now describes what is computed, which is the part that is useful. Package functions such as MASS::polr() are still named — those you may need to cite.

v0.11.02026-08-08

Added — analyse a map layer, and put the results back on the map

Until now the analysis side and the map side could not reach each other. A shapefile's attributes could not be modelled at all without leaving EasyAnalysis, and nothing a model produced could be shown on the map. Both directions now work.

Attributes to Table takes a vector layer and makes its attribute table an ordinary dataset, so any of the statistical methods can model it.

Predictions to map layer does the return trip: after fitting, it adds the fitted values and residuals to the layer the data came from, as new columns. Colour the layer by one of them in the Layers panel and you are looking at your model on the map.

Available on robust regression, Poisson, negative binomial, GAM and GLMM.

About accuracy. A model normally leaves out rows with missing values, so results do not line up one-to-one with the features. Rather than assume they do, the link between the data and the layer is recorded when the attributes are exported and checked again before anything is written:

• results always land on the features they were computed from, and features the model left out are marked as having no value rather than being given someone else's;
• if the layer has been edited since the attributes were exported, the write is refused with an explanation, because attaching results to shifted features would produce a map that looks right and is wrong;
• renaming a layer is harmless — the link does not depend on the name;
• existing columns are never overwritten.

v0.10.272026-08-06

Added — raster symbology

Single-band rasters were always drawn the same way: the first band, one fixed colour ramp, the full range of values and a fixed transparency. All four are now yours to set, in the same place as the vector symbology — expand the layer in the Layers panel.

• Band — choose which band to display, instead of always the first.
• Palette — five colour ramps, and a Reverse option.
• Stretch — decide which values the colours cover: the full range, 2–98%, 5–95%, or limits you type in. This is the one that usually matters. A handful of extreme or no-data pixels can drag the full range so far that everything real ends up in the middle of the ramp and the image looks flat. Clipping to 2–98% brings it back.
• Classes — keep a continuous ramp, or split the values into 3 to 9 bands.
• Opacity — see the basemap or another layer underneath.

Settings are kept per layer and saved with the project.

v0.10.262026-08-06

Added — delete features from a layer

You can now remove features from a vector layer. Select the rows you want in the attribute table, turn on Edit in the attribute panel, and press Delete.

Editing is off by default and has to be switched on deliberately, so nothing can be removed by a stray click. The button shows that the layer is editable while it is on, and switching to a different layer turns it off again.

Undo edit reverses the last deletion, up to five steps per layer. Deleting every feature at once is refused — remove the layer itself if that is what you want.

Fixed — switching tabs still showed a disconnection after a while

The previous version ignored connection drops for 30 seconds after you left a tab, which was not long enough: coming back after a few minutes still showed the panel. There is no time limit now. A drop that happens while you are on another tab is simply held, and only reported if the connection is genuinely still gone once you are looking at the page again.

v0.10.252026-08-06

Fixed — switching browser tabs no longer looks like a disconnection

Moving to another tab and coming back could show the "connection lost" panel even though nothing was wrong. Browsers slow down or drop the connection for tabs you are not looking at, and the app was treating that as a failure.

It now ignores connection drops while a tab is in the background, and waits a moment before reporting anything — so a brief interruption that recovers on its own passes unnoticed. A genuine disconnection, such as after the computer sleeps, still reports as before.

v0.10.242026-08-06

Added — a Reference page: what EasyAnalysis actually computes

There is a new Reference page on the website describing what happens when you run an analysis: the R function behind each method, the variables it needs, the options it offers, and what the results mean. It covers 14 statistical methods and 50 spatial operations.

The page is generated from the application itself. The method names, variable roles, options and the underlying function calls are read out of the app when the page is built, so they cannot drift away from what the software actually does. Assumptions and caveats are written by hand, because those cannot be derived from code.

If you are publishing a result, this is the page that tells you what to describe in your methods section.

Changed — Documentation is now "Getting started"

The old Documentation page was doing two jobs: helping new users find their way around, and serving as the reference for what the app does. It is now Getting started — install, the workspace, menus, file formats, projects, privacy and troubleshooting — and the method detail lives on the new Reference page. Existing links still work.

v0.10.232026-08-06

Fixed — undo history for deleted datasets was never released

Undo keeps up to five steps of history for each dataset. When a dataset was removed from the project its history was meant to be released with it, but never was, so a long session slowly accumulated saved copies of data that no longer existed. Removing a dataset now frees its history straight away.

Undo itself is unchanged: still five steps, still separate for each dataset, and it tells you how many steps remain each time you use it.

v0.10.222026-08-06

Added — map symbology for vector layers

Layers were always drawn in the same green. You can now style each one, and the settings are saved with the project.

Expand a layer in the Layers panel and choose how it is drawn:

• Single symbol — one fill and outline colour for the whole layer.
• Categorised — a colour per distinct value of a column, such as species or land-cover class.
• Graduated — a numeric column split into 3 to 9 shaded classes, such as volume or height.

Categorised and graduated layers get a legend showing what each colour means. Five colour ramps are available, four of which stay readable for the most common forms of colour blindness.

Only columns that suit the mode are offered, so you cannot accidentally colour by something that gives every feature its own colour. Class boundaries for graduated layers are chosen so each class holds a similar number of features, which keeps skewed data readable.

Outline width and fill opacity can be adjusted in every mode. Selected features stay outlined in red whatever styling you choose, so a selection is never lost in the colour scheme.

Symbology is reached in two ways: right-click a layer in the Layers panel and choose Symbology…, or use View ▸ Layer ▸ Symbology… for the layer you are working in. Either one selects the layer, switches to the map if you were on the data view, and opens its settings.

Full details are in the documentation under Map symbology.

v0.10.202026-08-06

Fixed — clicking a point on the map did nothing

Identify worked on polygons but not on point layers. Clicking a point marker was being handled differently by the map, and the code was watching for the wrong thing, so the click never registered.

Every feature drawn on the map now carries its own identity, so a click reports exactly which feature was hit rather than working it out from coordinates. Points, lines and polygons all identify the same way, and clicking a feature of a layer you were not already working in switches to that layer.

Changed — the map only moves when you ask it to

The map used to zoom automatically to frame your data the first time a layer was added. It no longer moves on its own: use Zoom to layers, Zoom to active layer, or the Zoom to button beside the selection count.

This also means selecting features no longer disturbs the view you have set up.

v0.10.192026-08-06

Fixed — the identify panel stayed only for a moment

Clicking a feature showed its attributes, but the panel closed again as soon as you clicked anywhere on the map. It now stays open until you close it with the X, so you can read it, compare it with the table, and pan around without losing it. Closing it leaves the feature selected.

Fixed — the identify panel ignored the theme

It was drawn on a white background with dark text whatever theme was in use, which made it hard to read on any of the dark themes. It now uses the theme's own colours, including a clearly visible close button.

v0.10.182026-08-06

Added — click the map to identify a feature

Clicking a feature on the map now selects it: the feature is outlined in red, its row is highlighted in the attribute table, and a panel shows its attributes at the point you clicked. Clicking empty space clears the selection.

It works both ways round now — click a row to find it on the map, or click the map to find it in the table.

Clicking a raster shows the value of every band at that point instead, with empty cells shown as "no data".

Points and lines are matched to whatever is nearest within a few pixels of where you clicked, and that allowance scales with how far you are zoomed in, so it stays usable at every scale.

Added — Zoom to selected

A Zoom to button next to the selection count fits the map to the features you have selected. Selecting a single point zooms to a sensible scale around it rather than as far in as the map can go.

v0.10.172026-08-06

Fixed — selecting a row now actually highlights the feature

The feature added in the previous version did not work: selecting rows in the attribute table changed nothing on the map. The right features were being found; they were simply never drawn. Selecting a row now outlines those features on the map.

Changed — the highlight is red, and drawn as an outline

Selected features are outlined in red. The outline carries the highlight rather than a solid fill, so the selection stays visible over satellite imagery and you can still see the feature underneath instead of it being covered up.

v0.10.162026-08-06

Added — select a feature in the attribute table and see it on the map

Selecting rows in a layer's attribute table now highlights those features on the map in amber. Select several and they all highlight; the header shows how many are selected and offers Clear, which matters because selections build up as you sort and page through a large table.

Selecting a row no longer redraws the map, so it stays quick even with a large raster underneath, and the highlight survives anything that rebuilds the map. Switching to a different layer clears the selection, since row numbers only mean something for the layer they came from.

This is the first half of working with features directly; clicking the map to identify a feature comes next.

Added — how to cite EasyAnalysis

The documentation had a placeholder where the citation should be. There is now a full citation in APA and BibTeX on the documentation page, and a CITATION.cff in the repository, so GitHub's Cite this repository button produces it for you.

If you used a published method through EasyAnalysis, please cite that method as well — its paper is listed on the app's References screen. Citing the tool does not replace citing the method.

v0.10.152026-08-06

Fixed — tables of your own data were limited to 100 rows a page

Whatever the size of your dataset, the largest page a table would show was 100 rows, and there was no way to ask for more — so a table of thousands of rows looked like it held 100.

The attribute table, the data view and the View Data window now let you choose 10 / 25 / 50 / 100 / 500 / 1000 / All rows a page. The data view had a second, separate limit that stopped it at 200 rows regardless; that is gone too, so all three now show the whole dataset.

v0.10.142026-08-06

Fixed — the app no longer becomes unusable after the computer sleeps

Closing a laptop lid, or leaving the machine to sleep, broke the connection between the page and EasyAnalysis. All you got was a grey screen with no way to recover — the app looked permanently broken, and the only way back was to find and run the launcher again.

There is now a panel that explains what happened and offers a Reconnect button. It also checks whether EasyAnalysis is still running and tells you which situation you are in:

• still running (the usual case after sleep) — Reconnect picks up where you left off
• no longer running — it tells you to start EasyAnalysis again from the Desktop shortcut

Your project is saved as you work, so nothing is lost either way.

v0.10.132026-08-06

Fixed — the attribute table stopped at 200 features

A layer's attribute table only ever showed its first 200 features, with nothing to indicate there were more — so a shapefile with thousands of features looked like it had 200. It now shows every feature, and you can select rows in it.

Changed — two ways to run EasyAnalysis, and only two

Installing is done once, from the terminal. On Windows that install then creates an EasyAnalysis shortcut on your Desktop and in the Start Menu, and from then on you just double-click it — no terminal in day-to-day use.

The downloadable Windows setup file offered briefly in the previous version has been withdrawn. Windows security features restrict downloaded files of that type, so it was not a dependable way to install. The one-line command remains the supported route on every platform, and is unchanged.

A properly signed Windows installer is still planned; that is the version that will work without any of these restrictions.

v0.10.122026-08-05

Fixed — the documentation still described the old way in

The walkthrough told you to "keep the terminal window open while you work" and to re-run the install command every time you wanted to start the app. Neither is true any more: there is a Quit button, and on Windows there is a Desktop shortcut. The reference page never mentioned either, and the project's README offered no install instructions at all.

All four now describe the same thing: download and double-click on Windows, or use the one-line command on any platform; restart from the Desktop shortcut; close with Quit. The security prompt Windows shows for downloaded installers is explained rather than left as a surprise.

The terminal remains a fully supported way to install and run EasyAnalysis — it is the route on macOS and Linux, and nothing about it has changed.

v0.10.112026-08-05

Added — EasyAnalysis now has its icon, everywhere

The Desktop shortcut added in the previous version used a generic PowerShell icon. It now uses the EasyAnalysis mark, built at seven sizes so it stays sharp from the taskbar to large-icon view.

Fixed — the website and the app had no icon at all

The favicon had been in the project since July, but it was never placed where the website could serve it and no page ever referred to it. So the site showed a blank icon in browser tabs and bookmarks, and favicon.ico returned "not found" — despite an earlier release note saying this was done. It works now.

The app's own browser tab had the same problem and now shows the icon too.

v0.10.102026-08-05

Added — no more terminal: download it, double-click it, done

Until now the only way in was to open PowerShell and paste a command — and because nothing created a shortcut, that was needed every single time you wanted to open the app, not just to install it. Someone who installed last week had no way back in except finding that command again.

Two changes fix it:

• The installer now creates an EasyAnalysis shortcut on your Desktop and in the Start Menu. After the first install you just double-click it. It starts in seconds, because it skips everything already downloaded.
• There is a downloadable installer on the website for people who would rather not touch a terminal at all. The one-line command still works and is still documented for anyone who prefers it.

The shortcut opens the app minimised rather than hidden, so if something ever goes wrong there is still a window to look at. Close the app with the Quit button.

Windows may warn that the downloaded file came from the internet — choose More info → Run anyway. That warning appears for any installer that has not been signed with a paid certificate; it is not a sign that anything is wrong.

(The shortcut is Windows-only for now, and it still uses a generic icon.)

v0.10.92026-08-05

Fixed — nothing was clickable in v0.10.7 and v0.10.8

The Quit button added in v0.10.7 carried a hidden line-break character in its confirmation message. That one character was invalid where it landed, which stopped the whole block of the app's interface code from loading — so every button in the top bar went dead at once. Projects would not open, Settings would not open, and the app otherwise looked completely normal, which made it hard to tell anything was wrong.

Fixed, and a check now runs over the app's interface code to make sure it is valid before a release. That check was written against this exact bug and confirmed to catch it — the previous tests only checked the message text was present, never that the code around it worked.

If you are on v0.10.7 or v0.10.8, update. Those versions are unusable.

v0.10.82026-08-05

Added — the site can now be found

The website had no sitemap and no robots.txt, so nothing told a search engine or an AI assistant which pages existed. Both now exist, and AI crawlers are welcomed explicitly rather than left to guess — people increasingly find tools by asking an assistant, and there is nothing here worth withholding.

Every page also gained a canonical URL, link-preview tags (so a shared link shows a title and summary instead of a bare address), and structured data describing what EasyAnalysis is, what it costs, and what it runs on.

Fixed — the summary written for AI assistants was unreachable, and wrong

llms.txt is a plain-language description of the tool intended for AI assistants. It was sitting in the wrong folder, so the published site returned "not found" for it, and its contents still described the old browser-based version — claiming the app ran inside the browser with no install, which stopped being true some time ago. It also used the assistant's former name.

It is now published at easyanalysis.dev/llms.txt and describes the current app: a local install, the real setup commands, and what it can actually do.

(No link-preview image yet — one still needs to be made.)

v0.10.72026-08-05

Added — a Quit button

There was no way to close EasyAnalysis from inside the app. Stopping it meant closing the R console window the launcher opened — which is also why the app could not be started from a tidy desktop shortcut: hiding that window would have left no way to stop it at all.

Quit now sits at the right of the top bar, on every screen. It confirms first, stops the R session, and replaces the page with a plain "EasyAnalysis has closed — you can close this tab" message rather than the grey disconnected screen, which looked like a crash.

Fixed — an invisible R process could be left running

Closing the browser cleared the app's data but never shut down the background R session used for heavy, cancellable jobs. That session kept running with nothing attached to it, invisible, until the machine was restarted. It is now shut down both when the browser closes and when you press Quit.

v0.10.62026-08-05

Added — turn the basemap off from the Layers panel

Data layers already had a switch each; the basemap was the one thing on the map with no row and no switch. Turning it off meant hunting for "None" at the bottom of a 14-entry menu.

There is now a Basemap row at the bottom of the Layers panel — where it belongs, since the basemap draws beneath every data layer. Same toggle switch as a layer, but no remove button and no expander: it is tiles, not a project layer. Its label shows which basemap is active, so the row doubles as a readout. It appears even in an empty project, because there is still a map to turn off.

The toggle is deliberately the same state as the menu's existing "None" entry rather than a second flag, so the two can never disagree — and switching back on restores the basemap you were using, not the default.

Added — pick your theme before you start

The theme picker used to live only in the workspace View menu, which does not exist until a project is open. Theme is now the first section in Settings, and the Settings gear is on the topbar from the moment the app loads — including on the Projects screen. Six swatches, each previewing its own background and accent colour.

Both entry points call the same function, so there is one mechanism, not two. Changes apply instantly and are remembered on this computer.

(A black & white theme is still on the list — this change moves where the picker lives, it does not add a palette.)

v0.10.52026-08-05

Fixed — CRS search really does query GDAL/PROJ now

CRS search felt hardcoded, and some coordinate systems could not be found at all. Both symptoms were real, and they had three separate causes. The search function queried PROJ's proj.db correctly — but nothing ever called it with a query. Every picker was built from a static 500-entry list, which selectize then filtered in the browser, so what you typed never reached the database.

• The list stopped at EPSG:32632. It was ordered by numeric code and cut at 500, so British National Grid (27700), UTM 35N (32635), Belgian Lambert 72 (31370), Czech Krovak (5514) and Dutch RD New (28992) simply were not in it.
• Search was one LIKE '%whole query%', so multi-word searches always failed — the real name is "WGS 84 / UTM zone 35N", which %utm 35n% cannot match. Matching is now tokenised: every word you type must appear somewhere in the entry, so a code, a name, or a mix all work.
• RSQLite was in neither dependency list, and it is the only way to read proj.db. On a fresh install every picker silently degraded to 8 hardcoded codes — which is what made the picker feel hardcoded in the first place. It is now an extras dependency.
• The catalogue is attached server-side, which is what makes 6,886 entries practical at all: embedding them measured 509 KB per picker and the app builds five of them.
• mod_raster.R stopped claiming it was "Querying 7,000+ official GDAL/PROJ EPSG…" while offering 500.

Only vertical and engineering systems are withheld, because neither can serve as a horizontal target CRS. Anything not in the catalogue can still be typed in by hand.

Changed

• The workspace's tool-render signal now reports which tool rendered, not just that one did, so a panel can re-arm its own controls without every other panel re-arming too.

v0.10.42026-08-05

Added

• Release notes on the website, kept up to date automatically. Every version's entry is now published at easyanalysis.dev/release-notes, and regenerated automatically whenever a release is added — so the page can never drift behind the app. Each version has its own link, so a support answer can point at one release.
• "What's new in this version" next to the version number in the app's About panel. The notes existed nowhere a user could reach them before.

Fixed

• The About panel read "an Co-Analyst" — a slip from renaming the assistant in the previous version.

v0.10.32026-08-05

Fixed

• More fixed-light surfaces found by an app-wide scan, the same fault as the References page rather than new ones: the confusion-matrix cells on Classification, the F1 bar, the macro-average row in every precision/recall table, and GAM's Yes/No non-linearity column were all fixed pastels that stayed pale on a dark theme. They are translucent tints of the theme's own colours now, so they keep their meaning and follow the theme.
• The Recommender's priority colours, a Download-spatial-data caption, the active dataset row's text, and the Co-Analyst's typing dot and disabled send button were likewise pinned to fixed colours.

Changed

• The Recommender's "Ask AI" button — a robot emoji plus a fourth name for the same panel — is now an icon and "Ask Co-Analyst". (The visual language rules emojis out.)

v0.10.22026-08-05

Fixed

• The References page was unreadable in every theme except light. Each reference card was pinned to a white background with fixed grey text, so on any dark set the card stayed white while the text that had no explicit colour inherited the app's light body colour — which is why some lines read as black and others vanished. The whole page now uses theme colours, including the Implemented / In progress / Cataloged badges.

Changed

• One name for the assistant: "Co-Analyst". The Analysis menu offered it as "AI Assistant" while its own panel header and the top-bar button both said Co-Analyst — three names in the code for one feature. Everything now says Co-Analyst.

v0.10.12026-08-05

Fixed

• Discriminant Analysis validated a different model than the one you configured. Its cross-validation refitted each fold with the package defaults instead of the settings you chose — Kernel DA ran at the default RBF width and cost rather than yours, Maximum Margin ignored its cost, and Locally Linear DA always used 5 neighbours whatever the slider said. Measured on test data: a default-parameter fold agreed with the configured model on only 38% of rows, so the reported accuracy described a substantially different classifier.
• Locally Linear DA's validation silently scored a different method entirely. When that fit falls back to its PCA-decorrelated form (which happens whenever the local covariance is singular), the validation no longer recognised it and quietly cross-validated plain LDA instead. It now validates the method actually used, and says so honestly when it cannot.

v0.10.02026-08-05

Changed

• ANOVA is now a registry entry — the last of the nine screens migrated. Verified against aov() and TukeyHSD() directly: identical ANOVA table, identical Tukey comparisons, identical eta-squared, Cohen's f and leave-one-out cross-validation. No bug was found in this screen.
• All nine screens are migrated. XGBoost, SVM, Decision Tree, Neural Network, PCA/FA/MDS, GAM, Random Forest, Survival and ANOVA now share one variable picker, one result layout and one place to add the next analysis. Adding one is a list entry rather than a new screen.

Fixed across the migration (v0.9.1 – v0.10.0)

Nine screens were ported; ten faults were found in seven of them, every one pre-existing and none reported by a user — because each failed silently or produced plausible-looking numbers:

• Three screens could not produce a result at all. XGBoost errored the moment you pressed Train (a package API had moved underneath it); GAM never fitted a model, ever; Survival's Cox proportional-hazards model never fitted.
• Five screens validated a different model than the one on display — SVM, Decision Tree, Neural Network, GAM and Random Forest all refitted their folds with settings you had not chosen, so the accuracy figures described a model you were not looking at. In SVM's case the cross-validation had never produced any result at all.
• PCA and ANOVA were clean.

v0.9.82026-08-05

Fixed

• The Cox proportional-hazards model never fitted. Adding covariates produced nothing at all, with no error shown. The formula it built used internal column names starting with underscores, which R cannot parse as variable names — so the formula failed before the model was ever attempted, and the failure was silently discarded. Cox models now fit, with the proportional-hazards check alongside them.
• Kaplan-Meier and the log-rank test were unaffected and always worked; only the Cox half was broken.

Changed

• Survival analysis is now a registry entry (migration 8 of 9), verified against the survival package directly: identical survival curves, risk sets, log-rank chi-square and Cox coefficients, under both tie-handling methods.
• An event indicator that is not 0/1 is now refused with a message naming the offending values, instead of producing a meaningless model.
• The Cox and log-rank views explain what to choose when you have not selected covariates or a grouping variable, rather than showing an empty panel.

v0.9.72026-08-05

Fixed

• Random Forest's 10-fold CV described a different forest than the one you trained. It never passed your tree count through, so the CV curve always came from 500-tree forests however you set the slider — and it used the classification sqrt(p) rule for choosing variables per split even on a regression model, whose displayed fit used p/3. Both now match the model on screen.

Changed

• Random Forest is now a registry entry (migration 7 of 9), verified to produce identical out-of-bag predictions and identical variable importance to the module it replaces, for both regression and classification, with the same mtry rules.
• Its partial-dependence plot still sits behind its own button (it is slow on large forests) and now reports clearly when the chosen variable was not one of the model's predictors.

v0.9.62026-08-05

Fixed

• The GAM screen could never fit a model at all. Every "Fit GAM" ended in an error notification. It built its smooth terms as mgcv::s(...), and mgcv does not recognise a namespaced smooth — it identifies one by the term label starting with s(, so the namespaced form was treated as an ordinary variable and the fit failed with invalid type (list) for variable 'mgcv::s(...)'. The screen now fits.
• Its cross-validation also dropped two of your settings — the fold models were built without the smooth type you chose (always the thin-plate default) and with smoothness selection hardcoded to REML. So a cubic-regression GAM selected by GCV.Cp would have been validated as a thin-plate REML fit. Both now match the model on screen.

Changed

• GAM is now a registry entry (migration 6 of 9), verified against mgcv::gam directly: identical coefficients, fitted values and R-squared, including a non-default basis and selection method.
• Its "Predictions to data pool" button is preserved, and GAM now says what went wrong when a fit fails (too large a basis, too many predictors) instead of showing a raw message.

v0.9.52026-08-04

Changed

• PCA / Factor analysis / MDS is now a registry entry (migration 5 of 9), verified against prcomp, cmdscale and factanal directly: identical loadings, scores, standard deviations, MDS coordinates and factor uniquenesses.

• Factor analysis now explains itself when it cannot run. Asking for more factors than the variables support used to surface R's bare complaint; it now says what failed and what to try instead (fewer factors, or Principal axis).

• The method-specific options (rotation, distance metric, component axes) appear only for the method they belong to, and "Colour points by" only for PCA.

• The colour-by column is subset by the same complete-case filter as the data, so it stays aligned with the points when a variable has missing values. (Verified rather than changed — the old screen did this correctly too.)

v0.9.42026-08-04

Fixed

• The Neural Network screen's validation ignored your Max iterations setting. Both of its cross-validation loops trained each fold for a hardcoded 200 iterations, so a network you trained for 1000 was being scored against one trained for 200. Third screen in a row with a validation that measured a different model than the one on display. The folds now use the iteration count you set.

Changed

• Neural Network is now a registry entry (migration 4 of 9), verified to produce identical predictions and an identical final objective value to the module it replaces, for both regression and classification, including a non-default architecture (hidden units, decay, iterations, restarts and scaling all changed together).
• Its validation is computed once when you press Run, and a regression network now refuses a categorical response with a clear message instead of failing inside nnet.

v0.9.32026-08-04

Fixed

• The Decision Tree screen's validation scored the wrong tree. Its per-fold refits were built with rpart's default settings instead of the max depth, cp, min-split and min-bucket you had set — so the validation numbers described a tree you were not looking at. Measured on the test data: the default tree had 8 leaves where the configured one had 2. The folds now use the same controls as the tree on screen.

Changed

• Decision Tree is now a registry entry (migration 3 of 9), verified to produce identical predictions and an identical CP table to the module it replaces, for both regression and classification trees, including non-default depth/cp/min-split settings.
• Its validation is computed once when you press Run, rather than being recomputed every time the results redrew — the same change SVM got.
• A regression tree now refuses a categorical response with a clear message instead of failing inside rpart, and the pruning control is labelled to distinguish it from the separate hold-out validation.

v0.9.22026-08-04

Fixed

• SVM's cross-validation never worked. It always showed "Awaiting SVM CV results…" and no result ever arrived. The refit inside each fold passed the fitted model's kernel back to svm() — but an e1071 model stores the kernel as an integer code (radial = 2), and feeding that back errors with "wrong kernel specification!". The error was swallowed by a tryCatch that returned NULL, so every fold failed silently and the screen waited forever. Now the kernel name is passed, and validation produces a result — and it honours the cost, gamma and scaling you actually set, which the old fold refits ignored.

Changed

• SVM is now a registry entry (migration 2 of 9), verified to produce identical predictions and identical support vectors to the module it replaces, for regression, classification and a non-default kernel.
• SVM validation is computed once, when you press Run. The old screen recomputed the whole k-fold loop inside its render outputs, so every time the metrics table redrew it refitted k models. It now runs once under the progress bar, where a slow job belongs.
• SVM's two cross-validation controls are labelled so you can tell them apart — one is e1071's own built-in check, the other a separate hold-out validation. Both were present before with no indication which was which.

v0.9.12026-08-04

Fixed

• XGBoost was broken and would error the moment you pressed Train. Not a regression — the xgboost package changed its API in version 3.x and the screen was never updated. Two separate breakages: xgboost(data =, params =, verbose =) no longer exists (params was removed and data renamed to x), and xgb.cv()'s best_iteration moved out of the top level into early_stop, so the screen read NULL and passed it straight to the trainer. Both fixed while porting the screen, and the best-iteration lookup now checks both locations and falls back to the minimum of the test metric, so the next API move degrades instead of erroring.
• XGBoost also refuses a binary task on a response that does not have exactly 2 classes, instead of silently encoding a continuous column into hundreds of "classes".

Changed

• XGBoost is now a registry entry rather than its own module — the first of the existing screens migrated onto statistics.R. Same method, same hyperparameters, same views; verified to produce identical predictions to the module's own computation. The old mod_xgboost.R is retired: still present, no longer registered or bound, so the screen appears once rather than twice.

Added

• The registry can render plots, which it could not before. A spec declares drawing functions in plots and the runner binds one renderPlot per entry — a plot needs a device, so unlike a table it cannot simply be returned as UI. Also added boolean options and conditional options (a setting that only appears when it applies).

v0.9.02026-08-04

Added

• Five new analyses, and a registry so the next ones are cheap: Ordinal regression, Robust regression, Poisson regression (counts), Negative binomial, and GLMM (generalised mixed effects). Find them under Regression, or search for them.
• GLMM fills a real gap. The existing Mixed effects screen is nlme::lme, which fits Gaussian responses only — it has no family argument at all, so a yes/no or count outcome with random effects could not be fitted anywhere in the app. The new screen does binomial and Poisson with random intercepts, random slopes, and crossed or nested grouping variables, and reports singular fits and convergence trouble in plain language with the usual remedies rather than as a raw error.
• statistics.R + mod_stat.R — one spec per method, one generic runner, the same move algorithms.R made for spatial operations. Adding an analysis is now a list entry: no new module, no new variable pickers, no new view plumbing. Methods that genuinely do not fit (Tests, Discriminant analysis, Descriptive, Clustering) stay as they are.
• Every registry method shares one variable picker, generated from the roles the method declares — so predictor selection is finally identical across them, and the Co-Analyst sees a registry method exactly as it sees a hand-written screen.
• lme4 added to the optional packages.

Fixed

• Internal notes claimed the shared model-quality function was unused. It is in fact used by the mixed-effects, random forest and linear regression screens, and now by the method registry too, which is what keeps those figures comparable between screens.

v0.8.42026-08-04

Fixed

• Model results were unreadable in light mode. The Model Summary, Performance Metrics and Cross-Validation blocks rendered light grey on white on every model screen. Bootstrap colours <pre> from --bs-emphasis-color-rgb — an R,G,B triplet, not the --bs-emphasis-color the app was overriding — and bslib compiles that triplet once from the default dark palette, so it stayed light text on every theme. It was never only <pre>: the same variable colours code, .well, .navbar and .link-body-emphasis, which is why the problem showed up across many screens. Each palette now emits the triplet, derived from its own ink with grDevices::col2rgb() so it cannot drift.
• Fixed light panels stayed light in dark mode. Twelve inline background-color: #f8f9fa / #fff8e1 / #e9ecef blocks across six modules kept a cream or pale background while the app's light text ran across them — the Quick Builder and Convergence Options panels on Mixed effects being the reported case. Replaced with reusable classes (.ea-subpanel, .ea-subpanel-warn, .formula-box, .ea-row-warn, .ea-row-flat). The warn variants are translucent tints of the semantic colour, so they take their lightness from whatever is behind them and work on every set.
• No fixed-light panel hex remains in any module. The two strip.background values inside mod_da.R's ggplots are deliberately kept: plots render on their own light canvas regardless of theme, so a themed colour there would be wrong.

v0.8.32026-08-04

Changed

• Column types are spelled out. The dataset summary showed tibble/pillar abbreviations (dbl, int, fct, chr, lgl) that mean nothing to anyone who does not write R — and this app exists so people do not have to. They now read number, whole number, text, category, true/false, date, with the R class kept as the cell's tooltip. The coloured badge behind them is gone too: it carried six hardcoded hex values that followed no theme, and the colour never said anything the word did not.
• The Co-Analyst no longer offers suggestion chips. They were the last place the app volunteered a next step, which contradicted its own system prompt — that already forbids the model from proposing one. The Recommend screen is kept and is where suggestions belong.

Added

• Undo now goes back 5 steps, not 1. snap() was already the single choke point every data operation passes through, so the change is one bounded stack. Each undo reports how many steps remain, so the last press reads as "no further undo steps" rather than a dead button. Capped deliberately — each entry is a full copy of the data frame.
• A "Docs" button in the app. The documentation pages have been live on the website for days, but nothing inside the app pointed at them, so users who never visited the site never found them. It sits in the top bar on both the projects screen and the analysis area, and opens in a new tab so a running project is never navigated away from.
• The guided tour covers 9 steps, up from 6 — added the tool search, Undo/Reset, and the new Docs link, clearing the "at least 8" requirement.

Fixed

• Undo could corrupt a dataset. The undo snapshot was a single slot shared across every dataset, so switching from A to B and pressing Undo restored A's data into B. The stack is now per dataset, and stacks for deleted datasets are pruned.

v0.8.22026-08-04

Fixed

• The app would not start on a machine without plotly — the failure every fresh install hits. plotly was in neither the core nor the extras list in launcher/deps.R, so the installer reported deps: OK and the app then died at boot with there is no package called 'plotly'. Three causes, all fixed: plotly added to extras; the workspace's output$chart_i <- plotly::renderPlotly(...) binding guarded the way its UI already was (a pkg:: in a binding resolves when the module server is built, so it threw before the UI's requireNamespace() guard could degrade to the static plot); and the observeEvent(workspace_ctx$tool_open()) in server.R moved after the assignment it references, so a failure there no longer surfaces as a misleading "object not found" on every flush. Audited every other eager output$x <- pkg::render* binding — all use core packages, so plotly was the only one exposed.
• Nothing showed that the app was working. Only 12 of 42 modules use withProgress, so on the other 30 — including most model screens — a slow fit looked like a frozen app. Worse, ui.R deliberately disables Shiny's own dimming (--shiny-fade-opacity: 1) on the promise of a "Running pill" that had never been built, leaving less feedback than stock Shiny. Added the global #ea-busy pill: pure CSS keyed off Shiny's shiny-busy class, so it needs no per-module wiring and works even while R is blocked (a server-rendered spinner cannot). Shown only by real request state, never a timer, and only after 400 ms so quick actions do not flash it. Also extended the existing skeleton shimmer to the model canvas.
• Native <select> popups and scrollbars ignored the theme. Page CSS colours the closed control but not the browser-drawn popup list. Every colour set now declares scheme = "light"|"dark" in theme.R and ea_theme_css() emits a real color-scheme declaration; :root declares it too, since a first-time visitor has no data-ea-theme attribute yet while the default palette is dark. Added option colour rules as the Chromium-specific complement.

Docs

• Checked the state of the guided tour and the links to the documentation, and recorded what was actually there rather than what was assumed.

A gap in these notes, stated rather than papered over: work done between v0.8.1 and this release — including the custom domain and the rewritten website — was never given a version or an entry here. These notes are published from a single file, so writing the entry is now part of finishing the work rather than something done afterwards.

v0.8.12026-07-29

Fixed

• Model screens opened with empty variable selectors. Linear regression, ANOVA and Random forest could not be run at all — by hand or by the Co-Analyst — because the workspace renders a module's panel lazily and updateSelectInput fired before the element existed. active_dataset() now depends on ds_refresh, bumped when a tool opens.
• Raster invisible on the map although the view zoomed to it: layers were added by leafletProxy after the map element had been re-created. Tiles, view and layers are built in one pass now. Reprojection also ran at full resolution before downsampling (18.1 s, and a warp failure under memory pressure that tryCatch swallowed) — it is downsampled first, memoised, and the fit comes from the extent alone.
• Zoom to layer was wired to the same input as Zoom to layers, so it always framed everything; it now targets one layer and reports honestly when there is nothing to zoom to.
• Adding a file yanked the map view. The automatic fit now applies only to the first spatial layer; everything after overlays.
• Dark surfaces on the light colour sets (data tables, accordions, the R console, buttons) — Bootstrap's component variables restated from the tokens.
• The R console cleared your script on every Run. It keeps it now.

Added

• R console write-back. Objects a script produces become project layers by class (data.frame, SpatRaster, sf, LAS), so a clip appears in the Layers panel and on the map. Only genuine OUTPUTS: consumed intermediates and aliases of loaded layers are skipped. Modes auto (default) and ask via options(ea.console_sync).
• Plot appearance app-wide — title, axis labels and colour on every screen, stored per screen. ggplot via a renderPlot wrapper; base-R plots per call site, with multi-panel diagnostic grids taking only an overall title.
• .ea-pop, a reusable hover panel behind one icon; plot appearance is its first user.
• Co-Analyst run_in_app — opens the real screen, fills it and presses Run. Every name is verified against the loaded data; near-misses and case differences are refused, never substituted. Linear regression only so far.
• True-colour RGB for multi-band rasters with a per-layer band mapping (detected from the file, never guessed), point clouds drawn and a density slider reaching the file's full point count, a 3D view button, a dockable/floatable R console with line numbers, and a guided tour that now runs inside the workspace.

v0.8.02026-07-27

The unified workspace. The ~35 separate analysis screens are now one workspace: a GeoLibre-style menu bar, a layers panel, a canvas that follows your data, a tool panel and a results dock. Plus a public site and a macOS/Linux installer.

Workspace

• One frame, two views — Map view and Data view, plus a resizable Split. A project with spatial data opens on the map; a table-only project opens on data.
• GeoLibre menu bar in the top bar — Project · Edit · View · Add Data · Processing · Controls · Packages · Settings · Help, with hover fly-outs and a tool search box. Every analysis is launched from Processing; the old per-screen menus are retired.
• Layers panel — every dataset and layer in one list, each with a visibility toggle and an expandable legend / styling.
• Tools open in the side panel, results in the centre. A tool can be floated (draggable, resizable) or minimised to the dock; past results park there as chips.
• R console is a bottom dock that slides up on demand.
• Data & Exploration operations are individual menu entries — pick "Row filtering" and only that operation's controls appear.
• Map: real layer rendering (rasters + vectors), 14 basemaps, and the view is preserved when you switch basemap. The attribute table shows the active vector layer's own attributes.
• Charts: a plot builder with a Static ⇄ Interactive toggle (plotly), and a draggable split between plot and table.

Look & feel

• Six colour sets (Forest, Light, Midnight, Ocean, Plum, Paper) — instant, remembered.
• App icon is now a lightbulb; the app is emoji-free (icons only).
• Fixed the startup dim at the root — Shiny fades every output while it recomputes; now suppressed and replaced with a proper boot screen, shimmer loaders and subtle result reveals (CSS only — no animation library).
• Contrast swept: 0 low-contrast text in all six themes.

Assistant

• Recommend is merged into the Co-Analyst — one surface: recommend a method, then ask it to run the analysis.

Packages

• Packages menu (was "Plugins") with a real CRAN search that tolerates typos, honest per-package progress, and immediate loading — no restart needed.

Fixes

• Single click on a project card opens it (pre-rendered clicks were being lost).
• Map zooms to your raster (the fit is applied when the map is built, not via a proxy call that could be discarded).
• Uploading a file no longer jumps you to another screen.
• LAS/LAZ: reopening a project no longer runs out of memory.
• Memory: the results store is bounded, the CRAN index is released, and session state is cleared on disconnect.

Distribution

• install.sh — one-line install for macOS and Linux (mirrors install.ps1).
• landing/ — public site: landing page with a 3D LiDAR hero, a how-to-use walkthrough and full documentation. Self-contained; no CDN.

Security

• Documented in ARCHITECTURE.md §10: no secrets in the repo, no user data tracked, and the only outbound calls are the Co-Analyst (opt-in), CRAN, basemap tiles and satellite search. Removed unused PowerShell dialog helpers.

v0.7.152026-07-22

• DA plots now show something useful with >1 predictor. The 1-D decision plots (Scatter+Regions/Class Lanes/Class Density) require a single axis; with 2+ predictors they showed "requires exactly 1 numeric predictor". Now they fall back to the 2-D PCA/LD projection (points by class + predicted-class ellipses) so a plot renders whatever's selected. (TODO next: true filled decision-boundary regions in the 2-D PCA plane — needs a grid-predict that also handles the PCA-decorrelated LLDA model.)
• Download hardening. The ea-download blob save now runs against the top document when possible, to escape iframe download restrictions (CSV/TSV export).

v0.7.142026-07-22

• Fix: CSV/TSV export wasn't downloading in the browser build. The standard Shiny downloadHandler doesn't reliably trigger a file download under webR. Replaced it with a message-based download — R builds the file text and sends it to the browser, where a small JS handler saves it as a Blob. Multi-format (CSV / TSV) + optional filename preserved. (Same webR limitation likely affects the hover plot-PNG download — a follow-up.)

v0.7.132026-07-22

• LLDA now works with multiple correlated predictors. klaR::loclda inverts a local covariance; with >1 collinear predictor (e.g. two diameter columns) that's singular (dgesv: system is exactly singular), so the fit died before any plot. Now, on that failure the numeric predictors are decorrelated via PCA (orthogonal → never collinear) and loclda is refit on the components — labelled "Locally Linear DA (PCA-decorrelated)". Multi-predictor LLDA plots (the PCA-projection scatter) now render because the fit succeeds. (A 3D rotating scatter is the next increment.)

v0.7.122026-07-21

• Fix: the Rows overview tile showed a blank icon — icon("rows") isn't a real FontAwesome name; swapped for a valid one.

v0.7.112026-07-21

• Fix: a deleted dataset reappeared in the left data rail. The list now explicitly drops any pool entry whose value is NULL — in this webR build pool[[key]] <- NULL can leave the name behind, so the deleted dataset kept showing. Now it stays gone.
• Fix: CSV download on the Dataset Overview could error. The handler now guards against no active dataset / empty data and sanitizes the filename.
• Clickable overview tiles (drill-down). Each tile on the Dataset Overview now opens a detail modal: Total NA → per-column NA counts + a table of the actual rows with missing values; Columns → type/missing/unique per column; Numeric/Categorical → those columns; Complete rows → the incomplete rows; Rows → the full table.
• Multi-format, named export. The export toolbar now offers CSV / TSV and an optional file name (defaults to the dataset name). TSV opens directly in Excel; native .xlsx export is a follow-up (the writexl wasm bundle needs sorting — shinylive skipped it).

v0.7.102026-07-21

• Added the EasyAnalysis favicon (the EA app icon). Center-cropped to a square PNG, copied into the site root as favicon.png + favicon.ico, and linked from the page <head> (also stops the /favicon.ico 404 in the console).
• First-boot UX (the "doesn't work on new browsers" report). Confirmed by watching a fresh, uncached browser boot the deployed site end-to-end: v0.7.9 boots fine — it just takes ~5–8 min the first time (downloading the R environment). The problem was the splash's safety timeout removing the splash after only 8 min, so a slower first boot got cut off and looked like "loads forever then ends". Raised the timeout to 20 min, and the splash note now says "3–8 minutes (longer on a slow connection) … keep this tab open." The splash still fades immediately once the app is actually ready (ea-app-ready).

v0.7.92026-07-20

• Critical fix: the app was stuck on the splash and never booted (regression from v0.7.7). mod_docs.R's .doc_b() helper took a single argument but was called with two in the Documentation overview, throwing "unused argument" while building the UI — so the app never rendered and the splash never received its ready signal. .doc_b() now accepts .... (Parse-checks pass on this kind of bug; the full source(global/ui/server) build check is now the gate before every push.)

v0.7.82026-07-19

• Splash: removed the sliding light-sweep (and leftover progress-bar CSS) — no sliding bar element on the splash at all now; the globe/satellite/radar/bars motion stays.
• Co-Pilot: Enter now reliably sends. The old handler clicked the send button, which could fire before Shiny had synced the typed text (sending stale/empty). Enter now passes the input's current value straight to the server (input$enter), race-free.
• Added llms.txt (llmstxt.org) served at /llms.txt — a detailed description of EasyAnalysis for LLMs and search: what it is, features, data formats, analyses, privacy, tech, and author. Copied into the site on every build by webapp_export.R.

v0.7.72026-07-19

• Honest "Running…" progress indicator. A global pill shows the elapsed time ("Running… 0:08", counting up) whenever the app is computing for more than half a second — no fabricated "time remaining" (we can't know an analysis's duration, so we don't pretend to). Driven by Shiny's busy/idle events; works on every screen.
• Co-Pilot status now reflects the real step. Instead of a hardcoded "Thinking…", the progress message updates to what the agent is actually doing — "Reading your data…", "Running a random forest…", "Computing correlations…", etc. — via a status callback threaded through the agent's tool loop.
• New in-app Documentation screen (mod_docs.R): a full user guide — interface, loading data, exploring, running analyses, the AI Co-Pilot, R Console, privacy, and a Tips/FAQ — reachable from the top menu, with a jump-to-section contents list.

v0.7.62026-07-19

• Redesigned loading splash. Replaced the plain screen with an animated hero — a rotating Earth-observation globe (raster shimmer, meridian wireframe), an orbiting satellite, radar pings and rising data bars — over a frosted-glass card with drifting aurora depth and a panning data-lattice. SVG line-icons (no emoji) for the four highlights: AI Co-Pilot, Recommend, Any data, Fully private. Respects prefers-reduced-motion.
• Also includes the v0.7.5 fixes below (delete-name fix, About/Acknowledgements split).

v0.7.52026-07-19

• Fix: deleted dataset's name lingered in the left data panel. Clicking × removed the data from the pool but the label stayed, because reactiveValuesToList() reliably re-fires on key ADDs (uploads appear) but not on key REMOVALs in this webR build. Added a ds_refresh trigger that the delete handler bumps and datasets_list depends on, so the list re-renders and the removed name disappears.
• About / Acknowledgements split. About now reads "a platform for conducting analyses without writing complex code…". Moved "University of Eastern Finland" out of About into a new Acknowledgements section crediting UEF for code contributions and for data used in analyses and testing — EasyAnalysis is an independent build, not a university product.

v0.7.42026-07-19

• Reverted the v0.7.3 boot-load experiment (it broke booting — see below) and confirmed the finding in webapp/global.R: compiled ML/stats packages can't link in this spatial-heavy browser build, late OR at boot. Fix is a separate lean build.
• Retained: .ensure_pkg() retry helper (v0.7.2), Decision Tree pickerInput selector, and the "3–5 minutes" splash estimate.

v0.7.32026-07-19

• Boot-load the compiled analysis packages that fail on late link (experiment). v0.7.2's retry fixed R-level packages (klaR::loclda works), but compiled ones loaded lazily still fail to register native routines: rpart C_rpart not found, xgboost XGDMatrixCreateFromMat_R not found, tidyr bad export type '_ZTINSt3…' (the kmeans cluster-map uses tidyr via factoextra). The proven fix is linking at BOOT (as the lidR stack does). Now webapp/global.R links base64enc first (so boot is safe), then tidyr, rpart, xgboost — a small measured set; will grow once confirmed boot survives the module-link budget in the browser.
• Decision Tree predictor selector now uses the pickerInput chip/search widget (with select-all), matching Random Forest instead of a plain dropdown.

v0.7.22026-07-18

• Fix: bundled optional packages still showed "install package" on first use. The v0.7.1 packages (klaR, kernlab, xgboost, rpart, glmnet, mgcv, survival, car, lavaan, tseries, trend, Hmisc, BayesFactor) are bundled and work, but in the browser (webR) a compiled package's first lazy-load can transiently fail ("file could not be read") even though a retry succeeds — proven in-app: klaR::loclda() fits fine, yet the screen's one-shot requireNamespace() reported it missing. Replaced every guard with a retrying .ensure_pkg() helper (helpers.R) that attempts the load up to 8× before giving up, so the first-try flake no longer surfaces as a false "install package" message.
• Splash note now reads "3–5 minutes" for the first-visit load estimate.

v0.7.12026-07-18

• Bundle the optional analysis packages into the browser build. klaR, kernlab, xgboost, glmnet, mgcv, survival, rpart, car, lavaan, tseries, trend, Hmisc, and BayesFactor are now included, so the Discriminant Analysis (LLDA/RLDA/KDA/MMC), GAM, Survival, Decision Tree, SEM, Bayesian, Time Series, XGBoost, and Ridge/Lasso screens can load them. Added via webapp/_deps.R (shinylive's scanner doesn't follow requireNamespace() guards, so these were simply never being bundled).
• Correction to the v0.7.0 "hard budget" note. The "install package" messages were not (primarily) a link-budget problem — the packages were never included in the build at all. Each has a WebAssembly build on repo.r-wasm.org; they just weren't scanned. (heplots/ggord/mda genuinely have no wasm build and still fall back to string-indirection.)

v0.7.02026-07-18

• Branded loading screen. The ~1-3 min first boot now shows an EasyAnalysis splash: what the app is, an elapsed timer, a "first visit is slow, then instant" note, rotating tips, a privacy line, and an auto-generated "What's new" from this changelog. Fades out when the app is actually ready (ea-app-ready signal). Injected by webapp_export.R from splash_template.html.
• Random Forest "unexpected input" error fixed (see below) — RF itself works (randomForest attaches at boot).
• Known limitation (browser build): some optional compiled ML/stats packages can't load — xgboost, glmnet, kernlab, mgcv, survival, rpart, car, klaR. Their screens show "install package". Root cause: this webR build has a hard budget for how many compiled .so can dynamically link before a libc++ symbol becomes unavailable; the LiDAR stack already consumes most of it, and preloading these pushed past the budget and broke the whole app's boot. Making them work in the browser needs a slimmer-boot strategy (future) — for now they run in the server build. (Not a regression; they were never working in the browser.)
• Fix: Random Forest "unexpected input" training error — column names weren't backtick-quoted in the formula, so names with digits/dots/spaces (VMI codes) broke parsing. Fixed in RF and the same bug in Decision Tree, SVM, Neural Net, and GAM.
• Co-Pilot: no unsolicited advice, stricter grounding, accurate capabilities. The agent no longer volunteers "best next step" suggestions unless you ask for a recommendation; the prompt hardens the no-hallucination rule (only tool outputs / context / image); and it's now given the app's real method list, so it stops claiming built-in methods (e.g. LLDA) are missing — it points you to the screen instead.
• Fix: deleted-dataset name lingering in the file-upload widget — cleared on delete.

v0.6.32026-07-18

• Browser tab title → "EasyAnalysis" (shinylive shipped "Shiny App"). Patched in webapp_export.R.
• Clip LAS to a shapefile. New "Clip LAS to this shapefile" button in the LiDAR tools — clips the point cloud to the polygon selected in Plot Shapefile (from an uploaded vector), matching CRS then lidR::clip_roi. The manual 4-coordinate clip is kept as an alternative.
• Auto-zoom to spatial data: already implemented (raster .zoom_to on load, LAS fitBounds on its location map) — the reason it appeared not to work was the raster display failing, which v0.6.2 fixes. No new code needed; noting it here.

v0.6.22026-07-18

• LiDAR: access the full point cloud. Read cap raised 500k → 5M (full plot-level clouds now load; still a memory safety net for huge files), and the "Max display points" slider now reaches the full loaded cloud (its max is set to the actual point count on upload). Default display stays 60k for a responsive 3D viewer.
• Fix "Install the 'rstac' package". rstac/exactextractr were bundled in v0.6.1 but their requireNamespace() guard runs on a button click = LATE, and exactextractr is compiled → same late-link ABI bug. Now preloaded at boot, so the guards pass. (If you still saw the message, you were on the pre-v0.6.1 deploy.)
• Raster/TIFF display fix. leafem::addGeoRaster (client-side JS renderer) is fragile in the wasm build; added a fallback to leaflet::addRasterImage (server-side PNG overlay via the now-working cairo device). Display errors, previously swallowed silently ("feels like it doesn't load"), are now shown.

v0.6.12026-07-18

• Fix: .laz / .tif uploads failed with bad export type for '_ZTINSt3__216__owns_one_stateIcEE'. Root cause: v0.5.0 made lidR/stars lazy-loaded; a heavy C++ .so that links LATE (on first pkg:: use) fails in this webR build — the same failure as the cairo/LAPACK bug. Reverted: lidR, sf, terra, stars are attached at boot again. Heavy C++ packages cannot be lazy-loaded here; lazy-loading is off the table as a fast-start approach for the spatial stack. (The v0.5.0 change was "verified" in the Node harness, which does not reproduce the browser's linking limits — only the browser does.)
• Fix: "Install the 'rstac' package to use satellite search". rstac and exactextractr (both behind requireNamespace() guards, so invisible to the shinylive scanner) are now force-bundled via webapp/_deps.R. Download Spatial Data and zonal stats work in the browser build.
• About panel: added a short description of what EasyAnalysis is (runs in your browser, your data never leaves your machine).
• Upload diagnostics: spatial-file load errors are now persistent and shown in full, plus an empty/unreadable-file guard — to pin down the .laz/.tif upload issue. (Verified in webR that terra/sf/lidR can read TIF/SHP/LAS; the remaining variable is the browser upload→filesystem step, so we need the exact on-screen error to finish the fix.)

v0.6.02026-07-17

• R Console (mod_rconsole.R) — new "R Console" screen. Run arbitrary R directly on your data: the active dataset is df, every loaded dataset is available by name, assignments persist between runs, and a single evaluation captures both printed output and any plot (base or ggplot — no double-eval, so RNG/side-effects behave correctly). Click-to-insert examples, Ctrl+Enter to run.
  • Safe in the browser build: each user runs their own sandboxed webR session in their own browser on their own data. (Would be arbitrary code execution on a shared server; the shipped product is the browser build.)
  • Verified via testServer: text output, persistent state, base+ggplot capture, error handling, the lm/LAPACK path, and clear all work.

v0.5.02026-07-17

• Lazy-load the heavy spatial stack. lidR, stars (and rlas via lidR) are no longer attached in global.R; they load on first use. Every call to them is pkg::fn() qualified (audited: zero unqualified call sites across all modules — the 13 apparent hits were all CSS/strings/labels), and :: loads a namespace on demand, so this needed no code changes. Sessions that never open a LiDAR/Surface screen never fetch them.
  • Bonus: it reinforces the v0.2.2 cairo/LAPACK fix — those libs failed when linking after the spatial stack, which now isn't loaded at boot at all.
  • Ceiling: sf/terra still load at boot because leaflet (needed at UI build for 8 leafletOutput() calls) imports sf. Making those lazy too requires deferring the map screens' UI behind uiOutput() placeholders.
  • Constraint: keep heavy-package calls pkg:: qualified. An unqualified rast() would now fail at runtime.
• Build is now self-healing (webapp_export.R), after a careless cp mod_*.R webapp/ broke two builds:
  • Prunes files global.R doesn't source. mod_gee.R (rgee/reticulate, no wasm build) had poisoned the bundle → "fatal-missing: reticulate".
  • Applies the no-wasm indirection (ggord/heplots → .opt_pkg/.opt_fun) itself, instead of relying on webapp/mod_da.R not being overwritten.
  • Both hard-fail with an explanation if they can't be satisfied.

v0.4.02026-07-17

• Renamed to EasyAnalysis in the UI: top-bar brand and About monogram (SA → EA). Status bar/About already used APP_VERSION.
• Co-Pilot: fixed the 400 error. GPT-5 family models (incl. the configured gpt-5.4-nano) reject the temperature parameter — "Unsupported value: 'temperature' does not support 0.1 with this model". temperature is now only sent for models that accept it. (The model id itself was fine — verified against OpenAI's current model list.)
• Co-Pilot send UX: animated typing indicator while the agent works (a turn with tool calls + vision can take 10–30s and the panel previously looked dead), send button disabled during a turn, double-sends ignored, auto-scroll to the newest message.
• Co-Pilot capabilities: two new agent tools —
  • column_stats(dataset, column) — full numeric summary (quartiles, sd, skewness, IQR outliers) or level frequencies, incl. a skew hint suggesting a transform.
  • correlate(dataset, columns, method) — Pearson/Spearman with the strongest pairs ranked first, for "what relates to what" questions. Both verified end-to-end, including error paths.

v0.3.02026-07-17

• Run button on every model screen. All 18 model screens now fit only on an explicit click (ANOVA, Logistic, Clustering and Discriminant Analysis were still refitting reactively on every input change). This matters most in the browser build, where each fit runs on the user's own single-threaded machine — a stray dropdown click could previously freeze the tab on a random forest. Recorded as UX rule #8 in DESIGN.md. Cheap outputs (descriptive stats, distribution plots) stay live.
• Removed 20 redundant plot-download buttons (+ their 20 handlers) across 14 modules. The global hover overlay already injects a PNG button on every plot — verified working in the wasm build before removing anything. All 43 CSV / raster / GeoJSON downloads are untouched (the overlay can't handle those).

v0.2.22026-07-17

Fix (the real one): blank plots AND "LAPACK routines cannot be loaded".

• Root cause: R links some of its own shared libraries lazily, on first use — cairo.so on the first plot, libRlapack.so on the first linear-algebra call. In this app that first use happens after ~40 packages incl. the heavy C++ side-modules (terra/sf/lidR/rgl) are loaded, and in the browser build that late link fails. Two symptoms, one bug: blank plots + dead model summaries.
• Fix: global.R now touches both at startup — a throwaway PNG device and a 2×2 svd()/solve() — before library(lidR)/sf/terra/rgl. Load order is the entire point; keep those probes above the spatial libraries.
• Verified in the browser: plot renders (real PNG), model summaries work, and the cairo/bad export type/LAPACK console errors are gone.

Correction to v0.2.1: that entry blamed the terra R.js patch. That was wrong — the patch is fine (Node runs terra + plots with it happily). The control experiment that "proved" it changed two variables at once (pristine R.js and a tiny package set), so it isolated nothing. The surgical patch in v0.2.1 is still worth keeping (it is narrower and safer), but it was never the cause. Node cannot reproduce this bug at all — only a real browser can.

v0.2.12026-07-17

Fix: all plots rendered blank in the browser build.

• Root cause: the R.js terra/PROJ patch introduced in v0.1.0 was far too broad. It rewrote the generic wasm stub factory and threw on any unresolved symbol, which broke side-module loading — cairo.so failed (bad export type for '_ZTINSt3__216__owns_one_stateIcEE'), R's PNG device never initialised, and every renderPlot silently produced nothing. It also surfaced as "LAPACK routines cannot be loaded" on screens doing linear algebra.
• Proven by a control app: pristine R.js + identical pins renders plots fine.
• Fix: patch only resolveSymbol(), supplying a no-op only for genuinely unresolved internal_proj_* symbols. No throw, nothing else touched. (R.js is now +83 bytes vs pristine, was +1969.)
• webapp_export.R now writes patched files with writeBin — writeLines was silently converting every LF to CRLF on Windows.
• New serve_local.R: serves webapp_site/ with COOP/COEP headers via httpuv::staticPath, so local testing is cross-origin isolated (webR uses the fast SharedArrayBuffer channel) and Range/HEAD still work for webR's lazy file loading. Use this instead of httpuv::runStaticServer for local testing.

v0.2.02026-07-16

• References screen — new "References" item in the menubar. Lists the published methods implemented in the app (currently Kalliovirta & Tokola 2005), each with full citation, the named method(s) derived, where it's used, and a DOI link. Single-sourced in references.R (APP_REFERENCES), kept in sync with papers/METHODS.md.
• About panel refreshed: name → EasyAnalysis, live APP_VERSION, tagline "A universal scientific analysis platform" (was stale "v0.9.0" / forestry-only).
• Source now lives in the deploy repo — a curated src/ copy of the app source was added to the EasyAnalysis repo (not just the compiled site), so it is a full project repo.

v0.1.02026-07-16

First versioned baseline. The app now exists in two builds from one codebase, with an AI agent and the first paper-derived methodology implemented.

Platform

• Browser build via Shinylive/WebAssembly — the whole app runs in each user's browser, no R server. Built by webapp_export.R (pins shinylive assets 0.10.10 / R 4.5 channel, patches the terra PROJ loader, bumps the SW cache, integrity-checks the bundle).
• Deployment: static site + Vercel serverless AI proxy (api/chat.js); the OpenAI key lives only in Vercel env, never in the client. Cross-origin isolation headers set in vercel.json.

AI Co-Pilot → agent

• The Co-Pilot now RUNS analyses via OpenAI tool-calling (agent_tools.R): list_datasets, describe_dataset, run_analysis (descriptive, lm, anova, ttest, lme, logistic, rf, clustering, pca), using the app's own fitting functions.
• Dual transport (httr on server, sync-XHR via webR in browser); user-supplied key overrides the shared proxy key.

Analysis features

• Dependent-variable transforms in Linear Regression (log / log1p / sqrt / 1/Y), in addition to predictor transforms. Options always visible (nothing hidden).
• Bias-corrected back-transformation to the original Y scale — Kalliovirta & Tokola 2005 — with original-scale metrics. Verified numerically.

Docs

• The project's internal documentation was rewritten around the "universal scientific analysis platform" goal, and a catalogue of published methods was started — each paper mapped to the method it provides and whether that method is implemented yet.

Cataloged papers

• Kalliovirta & Tokola 2005 — stem diameter & tree age models (bias-corrected transformed-Y regression). core methodology implemented.