- The new console became the place the team actually works from. The two older separate areas, user administration and customer support, folded into the single Explorer, and its section switcher became a proper tab strip. You can open a project's uploaded files and download them. The raised-flag signal now shows as a column on the project lists, so you can see which projects have something waiting without leaving the table. Talent sub-categories read as real service names instead of internal codes, the console uses the full width of the screen rather than a narrow column, and each area got its own purpose-drawn portrait of Mira.
- Big clients got a desk, on the console that actually reaches production. Until now the one screen for "this client just committed serious money, go and talk to them yourself" lived only on the older internal console, which is published to the test environments and nowhere else. It is now its own tab: the client, the project, the committed budget, whether the assistant is still running, and when the flag went up, sortable and filterable like every other list there. Opening one composes the whole picture underneath it, the live conversation, the job description exactly as the client received it, and the talents from the latest search with their Fiverr links.
- You can now stop the assistant and take a client over, from the project page itself. A flag used to link you to a project you could only read; to actually stop anything you had to go somewhere else. Two buttons now sit in the project header. Pause parks the assistant's turns while the client's messages keep arriving, so a human answers. Stop freezes every automated step and tells the client that a human from our team will contact them in their Fiverr inbox. Only Stop asks you to confirm, and the confirmation describes what the client will experience rather than which field moves, because a dialog on the reversible, invisible one would only train people to click through the one that matters. The conversation tab gained a compose box for replying as Mira, with shared snippets and the thread updating live.
- The money lines moved, and past a point the system stops by itself. A big-client flag now goes up from $1,000 committed rather than $2,500. And there is a new, higher line: at $6,000 or more, the moment that client presses to finish their job description and start looking for talent, the search is refused and the entire take-over runs on its own, the assistant paused, the flag raised, and the client told a specialist will be in touch. Nobody at that size gets an automated search before a human has spoken to them. The manual Stop button is untouched, and all three routes into a take-over now go through one shared implementation, so a client is left in exactly the same state however it was triggered.
- A real buyer was being pushed to the cheap lane over a currency. A Tel Aviv client said "ILS 3,000" for a full website redesign, about $978, and the system had already converted and recorded that correctly. The part that decides which experience a client gets still went back and read the raw shekel figure out of the chat, ruled it under our $500 minimum, and hard-routed a genuine buyer to plain self-serve search, confidently. That minimum is now judged from the one recorded dollar figure and it works in both directions: below it, the client goes to search; at or above it, a claim that the floor applies is simply overruled.
- Every search now leaves a record, with the scout's own reasoning. The live assistant flow was writing no search record at all, so the searches page could only ever show the older guest funnel. Every hunt is now logged: the full shortlist of twelve, which of them outreach actually contacted rather than all twelve, and the scout's own explanation of why it chose them. That explanation was previously worked out and thrown away, with the page showing a mechanical "top match, nine shortlisted" headcount in its place, so we were presenting furniture as the model's thinking. Historical searches were backfilled so the page is not empty behind today.
- The Search Tier column now means the thing we actually rank by. The earlier reading of it was simply wrong, showing a grade from Fiverr that we rarely even receive. It now shows our own ranking, stamped onto each pick at the moment the sort happens rather than recalculated later, which would otherwise describe today's settings instead of the ranking that really shipped. Each search also carries one score from 0 to 100, worked out once when it is written down, so the column is comparable across searches instead of meaning something different in every row.
- A talent page, with every project we approached them for. Open any talent and see their statistics alongside the full list of projects we reached out to them about. The project's own talent tab became a real table you can open from, rather than a flat list.
- The card quotes the package the scout actually graded. On the budget path a card could show a neighbouring package at $250 while the scout had read, chosen and graded the $2,000 tier, so the price on the card was not the price of the work that had been picked. It now honours the scout's choice.
- The scout admits when nothing fits, and stops overstating what it enforces. When every requirement came back empty-handed it used to quietly fall back to a general pool instead of saying so; an all-off finish is now an honest no-match. And the label describing a list of required languages called it a soft preference when every one of them is in fact enforced as a hard requirement, so it now says what it does.
- The brief writer was told the wrong industry for six of eleven categories. The file that names each service category had two of them swapped outright, so a writing and translation project was introduced as business and a business project as writing and translation. Two more were plainly wrong, and two shipped as literal placeholder text that flowed through as though it were a real industry, which is worse than a blank because a blank is simply left out. Only the names were wrong; every number behind them was always attached to the right category. The names are now pinned by a test against the official Fiverr taxonomy, because nothing in the codebase regenerates that file and so nothing else would ever catch a mislabel.
- The reporting warehouse grew up, after Wednesday's outage. It used to follow the newest reporting code every hour regardless of which version of the product an environment was actually running, so a report could arrive expecting a database change that environment did not have yet. That is exactly what happened: one core report failed every hour for eighteen hours. Each environment is now pinned to the version of the product it really runs, refreshed at deploy time rather than up to an hour later, and a failure now reaches a human instead of sitting on a dashboard nobody watches. Alongside it, the first real set of product measurements for Mira went in.
- The app took on the Fiverr brand faces, then simplified back to one. The brand typefaces landed first, a sans and a display serif together. Product's read was that the app looked less clear than before and was missing something, and design pointed at the mock, where the headline is a light weight of the single family rather than a heavy serif line. So there is one family again, and the whole set of text sizes is now lifted from five approved screens instead of guessed, with hierarchy coming from weight, size and spacing. The buyer app also got the Fiverr favicon and a proper icon for when you save it to an iPhone home screen.
- Guided onboarding was finished properly. The install had loaded the product-tour software and stopped there, which can never show anything: without being told who the user is, no session ever begins, so nothing could have fired. It now identifies the signed-in client with exactly the same id our own events already use, which is what pairs the two together; tells the tour about every screen change, because this app has no page addresses of its own and would otherwise almost never notice one; and clears the session on sign-out. Guests are refused on both sides, so a throwaway browser label can never mint a junk profile we then pay for. A production check now exercises the whole delivery path end to end.
- A run of small client-app fixes. A link that names a project now lands straight in that conversation instead of flashing the welcome screen first, a link with no prompt lands on home rather than opening an empty chat, and reloading an existing chat no longer flashes the welcome hero. The waiting headline was shortened to what actually fits, the answer acknowledgement card became one white headline per the reviewed copy, an exported job description no longer carries a draft or ready badge, and the delete warning on your profile speaks about your conversations rather than your account.
- Quiet groundwork. Production admin sign-in had its cookie scope corrected, a deploy now points traffic at the exact revision it just created rather than whatever happens to be newest, and both pre-production environments roll through one shared promoter so they cannot drift apart. The alerts we post to chat lost their trailing platform footer, a blocked client is handed the Fiverr support link as a real link, and comments are now stripped out of built pages.
The biggest day yet, 89 commits across the whole crew, and mostly about handing control to a human. The new console became the place the team works from, with user administration and support folded into one Explorer, and it gained a big-clients desk plus the ability to pause or stop the assistant from the project page and answer the client yourself. The money lines moved with it: a flag from $1,000, and at $6,000 the system now refuses to search and hands the client over by itself. A real buyer who said "ILS 3,000" had been pushed to the cheap lane because that figure was re-read out of the chat rather than taken from the converted number already on file. Every search the assistant runs is now written down with the scout's own reasoning, the tier column finally means the ranking we actually sort by, and there is a talent page showing every project we approached someone for. The brief writer had been told the wrong industry for six of eleven categories. The reporting warehouse was pinned to the version of the product each environment really runs, after a report failed hourly for eighteen hours on Wednesday. And the app took on the Fiverr brand faces, then simplified back to a single family with its type scale lifted from the approved screens.