What I've been building

A couple of things that got in your way on the phone now behave.

• A confirmation message no longer sits over the bottom navigation bar, so you can still reach the buttons for moving around the app.
• Fixed sliding panels where tapping a field lower down scrolled the panel’s title and buttons off the top of the screen.

The package now runs its own tests, so a change to a shared component is checked before it can reach the apps that use it.

The work went into how the interface building blocks are wired together, so a shared piece can be updated in one place, with nothing that changes what you see or do.

The screen for judging two model answers and the results screen were both simplified to fewer controls.

• You can now mark what went wrong on each answer with tags inside that answer’s box, so a fault stays tied to it.
• The whole answer box is now the vote, so choosing a winner is one tap and both parts of the screen use the same wording.
• The results screen now leads with the ranking, and the weight sliders fold behind a summary that still shows their current values.
• The download button now says how many scenarios it covers, so the filter is clear where you choose the file.
• The CSV and full-page downloads now sit on the results screen, labelled by what they do, so they’re where you read the results.
• Fixed a false worker-unreachable message before any run, button-styled links showing as plain text, mismatched export wording, and judge links opening the wrong run’s answers.

A database stall that could freeze a task snooze for half a minute now gives up within about five seconds and shows an error.

An overdue task that repeats monthly can now be snoozed to the next day without opening the calendar.

• You can now move an overdue monthly task to tomorrow with the quick snooze button, where before only the calendar picker would accept it.
• Fixed the tutorial’s demo habit sometimes missing its first day, and a sample task landing on the wrong day for some time zones.

The homepage cards for two projects now open with a moment you have been in, instead of describing how the app works.

• The Anecdote card now leads with reaching a performance review and having lost track of the year’s wins, so its point lands faster.
• The Remind me later card now opens on calendar alerts you dismiss and never act on, which is the gap it fills.
• Fixed the build log page, which had stopped showing new entries since late August even though they were being written.

A couple of things that were displaying wrong got corrected.

• New file-upload and stat-row controls, plus a two-line option for dropdowns, were added to the shared component set for a tool being built later.
• Fixed theme labels on story cards showing as loose grey words not readable pills, and a long name stretching a dropdown past its panel.

The tool for comparing model answers lost the clutter that sat between you and the work.

• The run screen now opens on the active run, showing its name, progress and money spent, with one button to resume it.
• Starting a new run is now a folded panel at the foot of the page, with a proper upload button and a preview of the settings.
• The three screens show Run, Judge and Results as links at every width, so the narrow-screen menu and its extra taps are gone.
• An expired sign-in now shows one banner with a reload button on every screen, so you keep the navigation and do not lose your place.
• Fault labels on the judge screen now come from the run itself, so a run no longer offers faults that do not apply to it.
• Fixed votes silently dropped after a session expired, and a run status that could stick on the words “Checking the run”.

The story screens were adjusted so the text that matters is easier to pick out at a glance.

• The heading above a story now reads as a proper heading, where before it looked no heavier than the paragraph beneath it.
• The theme tags on your story list now sit back quietly, so a story’s title is the first thing you notice.

The date picker used in the planner and in Anecdote moved to a newer version of the library behind it, with no change to how it looks or works for you.

A couple of problems that made the tool unusable at the wrong moment are now sorted.

• You can now open the tool again, where every page had been refusing you with a message that wrongly said your session had expired.
• Fixed an uploaded run setup that named a judging method the tool cannot run, which used to cancel your active run and only fail later.

Snoozing a task to a day you choose now works on a phone, where before most of the calendar ignored your taps.

• You can now tap any day in the snooze calendar to move a task there, where before most days on screen did nothing.
• The calendar now greys out days it cannot use, past days and days the task already runs on, so you can see them before tapping.
• You can now snooze a repeating task to its own weekday in a later week, which the app used to refuse without a word.
• Setting up your profile now flags a mistyped time zone on the spot, so you correct it there instead of hitting a vague error later.
• Fixed a Tomorrow button greyed out on Sundays, unavailable days shown too faint, and a line above the skip button that just repeated it.

The page showing how this workspace is put together now matches the current tooling, after drifting six days behind.

• Every skill on the page now opens on its current text, after most were cut back hard since the last build.
• Five components and eleven notes added since then now appear, sitting in the groups they belong to rather than nowhere.
• Where a skill’s text moved into files beside it, the page now follows that content instead of showing the skill as shorter than it is.
• The chart-building tools now have their own section rather than crowding the general layout group.
• Counts stated on the page now come from the files they describe, so they can no longer quietly go wrong.
• Corrected a figure for how much of one skill moved into companion files rather than being deleted.

A couple of things that used to fail without explanation now work properly.

• You can now delete your account after adding an update to a story, where before the deletion failed and left everything in place.
• Fixed a weekly reminder time that could be set out of range, which silently stopped your weekly prompt from ever arriving.

This is an internal tool for judging which model or prompt gives better answers, and it got easier to set up and read.

• You can now run one model against two versions of a prompt, so the result names which wording wins rather than which model.
• The example test setups now group their cases, so a long run reads as a few named headings instead of many single ones.
• Every setting a run can carry is now documented, so building one no longer means reading the code to learn what a field does.