- When a talent answers a message sent in the client's name, the client can now read it, and reply, inside the app. Twice in the product we write to a talent as the client: the brief sent from search results, and the closing push after a pick. Those open a separate conversation on Fiverr, apart from Mira's own, and until today nobody here ever saw the answer: it landed in the client's Fiverr inbox and stopped. There is now a small in-app inbox for exactly that conversation. Files travel both ways, with the same size and type limits as the main chat; a file a talent sends shows as a link once Fiverr's virus check clears it, as "blocked" if it fails, and by name while it waits. Fiverr's own message types are labelled too, so a custom offer reads CUSTOM OFFER above its line instead of looking like ordinary chat, and a message that is only an offer no longer shows as an empty bubble. None of it moves Mira's side of the hunt: a talent answering the client directly does not count as answering Mira.
- A talent can no longer set a rule that picks clients by nationality, race, religion, gender, age or disability, and old ones stop working today. In early September Mira was told in her instructions to refuse these. The instruction was live and the same thing was captured again two days later on another talent. So the refusal moved out of the instructions and into the system: one check now runs when the rule is captured, whenever one is saved, in the admin editor, and when search reads the rules it filters by. "No Asian companies" is refused; "no German to English translation" is a kind of work and stays. Mira's "noted" is replaced by a polite refusal, and any rule of that kind already stored stops filtering the day this ships, with a counter that should fall to zero.
- A full run of the live behaviour tests found six real bugs, and all six are fixed. The run scored 286 out of 309. Among the failures: a talent who said "delivered in 21 days, can start next week" had their delivery recorded a week early, because the model reported the working time and not the date; it now reports the two numbers it was actually given and the system adds them. A talent who declined a brief and added "please stop sending me logo work" got a warm "noted" and nothing was saved; declining now keeps the preferences tools open. Mira agreed to "three identical seats" on a three person hire and split the client's money for them; she now says the search fills one seat at a time and asks for the per person budget. A talent's "ok" while a question was waiting on the client no longer gets a made up next question. And "$100 for the bug and $100 for the feature" is now a budget of two hundred, not a range from one hundred to two hundred. Seven more failures were tests still scoring rules the product had retired, and were brought up to date.
- The newest, much cheaper model can now be tried on real projects, one experiment per group of clients, both switched off. One experiment covers projects on the main model, where the new one costs about a twentieth as much to write with; the other covers the lighter regional lane, where it is roughly half the price. Each project gets one model for its whole life, and an operator turns either up from the admin console. Getting ready for it also fixed the test suite's own grader so it never switches model along with the thing it is grading, priced the new model family so its spend is never reported as zero, and corrected the suite's cost to about fifty dollars for the current 309 scenarios.
- An experiment on how search reads the brief is wired in and switched off. Fiverr's search team can now read the client's full brief itself instead of the short summary we send. Half the question is which reading finds better talents, so the choice is made once per client and applies to every search in a hunt. The first version compared two settings that, tested the same day, returned exactly the same talents on three unrelated briefs, so it was corrected by the afternoon: the two arms now differ by whether the brief is sent at all, which is the thing that actually changes the results.
- A talent reading back their preferences sees only the ones they set. A talent who had set only a minimum price was told "Work to skip: nothing set" two messages after answering one question about money. The readback now lists what they chose and nothing else.
- Pressing Back from the Fiverr login no longer leaves Mira stuck on "Checking your session". On an iPhone, a sign up that did not go through and a tap on Back brought back a frozen copy of the page with the loading screen still up, forever. Returning to the page that way now re-checks who is there, and the workspace opens as a guest.
- Mira inside Claude and ChatGPT stopped refusing a first message on stricter hosts. Some hosts insist every optional field be sent, and ours were typed so that "nothing" could only be sent as a null, which one host's own check rejected before the message reached us. Every optional field is now a plain text field where blank, missing and null all mean "not supplied", and the check that guards this now covers every group of tools, not three of five.
- Thirteen thousand new accounts got their Fiverr number back, and new ones keep it, without waiting for a release. A database change reached production two days ahead of the code that fills the column, so about ten new accounts a minute had no Fiverr number for forty five hours. A one time repair filled the existing rows, and a small database rule now fills the number on every new account until the code catches up, after which it will be removed.
- Housekeeping. A data analyst was given read access to the warehouse on the same terms as the analytics group, and the previous day's status page and the architecture page's note on the price judge were published.
A Thursday of 15 commits whose centre is a conversation we had been starting and never hearing the end of: messages sent in the client's name now come back to the client, with files and with Fiverr's offers labelled for what they are. Around it, one rule moved from a polite instruction into the system itself, because instructions alone had already failed once in production, and a fifty dollar run of the live behaviour tests turned up six genuine bugs, from a delivery date a week early to a budget shown as half of what the client said. Two experiments were laid down and left switched off, one on a far cheaper model and one on how search reads the brief, and the second was corrected within hours when a same day test showed its two sides were identical. Underneath, a database rule plugged a production gap without waiting for a release, and the Claude and ChatGPT connector learned to accept the stricter way some hosts send a message.