# Changelog BidBrain doesn't follow semantic versioning against a public API, this is a private buying tool, not a library. These are just the project's own build stages, each covering everything shipped up to that point. For the full decision log behind any entry here, including exact live test numbers and the reasoning behind edge cases, see `CLAUDE.md`. Bug fixes are called out under **Fixed** in each version; everything else is a new capability (**Added**) or a change in existing behaviour (**Changed**). ## v3.14.154, 2026-10-05 ### Fixed - A history check on the wrong car no longer goes to DealerKit. Steven, on S77RUU: the check ran on the seller's private plate after it had moved to a 2025 Mercedes CLA, and the report was lined up as the history of the 2017 Mercedes A we bought. Dealer OS now checks the plate that stays with the car and compares every report with the car; when it takes a report back, BidBrain drops the file from the car so it is never sent, and says so on the car's timeline. Same for both dealerships. ## v3.14.153, 2026-10-05 ### Fixed - PL66ZZO now shows its £200 chip. It was won on Motorway at £6,456 and chipped £200 through Motorway before v3.14.152, so BidBrain had £6,256 as the winning bid and no chip. On the next start the winning bid goes back to £6,456 and the chip is recorded as "Price amended on Motorway", from Motorway's own payment email. The price paid stays £6,256, £6,650.80 all in. Steven chose this car only; other older Motorway cars stay as they are. ## v3.14.152, 2026-10-05 ### Fixed - A chip put through Motorway was lost from the purchase price. Steven, on PL66ZZO: won at £6,456, chipped £200 through Motorway at collection, and Purchases already showed £6,256 as the winning bid with no chip, before Donna could record it. While Motorway's breakdown is still estimated it is re read, and its "Vehicle price" line simply drops to the amended figure. A real breakdown's bid that goes down now keeps the bid as won and records the drop as a chip, "Price amended on Motorway", so Dealer OS's Collected step shows the chip already there. A price that goes up is still taken as read. ## v3.14.151, 2026-10-04 ### Fixed - Carwow cars in the live auctions showed as seen before, twice today. Steven, on LC66AZW: "there seems to be alot of cars in the live auctions that are showing as seen before but its shown twice today?". Carwow holds no sale on a Sunday, so Saturday evening's run and Sunday's run both read Monday's sale, but Saturday's was filed under Sunday. Every Carwow car still on sale on Sunday evening then looked like it had already been through a Sunday sale: 54 of the 80 Carwow cars on Monday 5 October's list. A Carwow read stamped Sunday is now filed under Monday, and the sightings already logged under a Sunday move to Monday on the next start. Motorway does sell on a Sunday and is unchanged. ## v3.14.150, 2026-09-29 ### Fixed - Seller notes on Carwow and Motorway never reached the car page. Steven, on FY16BHD: "this passat has seller notes on carwow that put me off bidding for it but yet bidbrain doesnt tell me". Carwow shows them under the car's features ("Seller notes: Clutch pedal squeaks, rattle coming from the airbag area"), cut short behind View more when long. Motorway keeps them in Additional details: the extras the seller ticked, anything written after them, and why a car was relisted. BidBrain now reads all of them on the page visit it already makes, and they show as Seller notes on the cockpit card and the Dealer OS car page. ## v3.14.149, 2026-09-29 ### Fixed - Carwow cars showed only a distance on the car page, never the town. Steven, on SA18AJX: "why are some cars only showing the distance as the location and not showing the city or the actual place as well?". Carwow's list card only says how many miles away the car is. The town is on the car's own page, in its Location section ("Elton, Bury (134 m)"), and BidBrain never read it. It now reads it on the page visit it already makes, so Carwow cars carry their town into the cockpit and Dealer OS. A car whose town is now known also gets the ferry crossing check. - A Carwow car's page listed only its current owner, however many it had. Steven, on the same car: "it only shows the current owners time of ownership, it should show all owners". Carwow's car page has an Ownership history section with every owner's dates, and the deep read only took the current keeper's start. It now reads the whole section, oldest first. Motorway cars already read every owner (checked live). ## v3.14.148, 2026-09-28 ### Fixed - A car whose listing plate could not be read was called new every day it came round. Steven, on S30NNT in the Carwow sale: "It says its not been seen before but i could swear i have seen it before". Carwow's card did not show a plate BidBrain could read, only the car's own page gave the original plate underneath, and the repeat tracker skipped any car with no listing plate. So it was never logged on any day it appeared, and the card said First time. Six Carwow cars on the 28 September list had no listing plate, five of them with an original plate on their page. - Sightings are now logged under the original plate when the listing plate is missing, once the car's page has been read. Seen before, the reserve drop and Dealer OS's Seen list all look a car up by both of its plates, and by the other half of any private plate pair on file, so a car listed on its private plate one day and its original plate another still meets itself. - A car with no plate at all is no longer called new. There is nothing to look it up by, which is not the same as seeing it for the first time, so it no longer gets the New today badge or the new car boost in Top picks. ## v3.14.147, 2026-09-22 ### Changed - The Dealer OS push now goes compressed. Plain JSON squashes about ten to one, so the biggest real day seen so far, 2,078 held back cars and 5.17 MB of list, leaves the Mac as roughly 555 KB. This was tried and reverted on 21 September because Dealer OS read the bytes exactly as they arrived and answered "Body is not valid JSON"; Dealer OS has since fixed its side, so all four of its machine doors decompress the body and measure the size after decoding. Checked against Dealer OS's own reader with the real bytes this sends, not a stand in. - Because of that, the 200 row caps on held back cars and lost bids put in on 21 September no longer bite. The whole held back list goes across again. The caps are still in the code as a genuine last resort, behind two ceilings that are now both checked before a push leaves: 19 MB for the document itself, a margin under the 20 MB Dealer OS stops decompressing at, and 4 MB for what actually travels, a margin under Vercel's own request cap. A real day is nowhere near either. - The run summary line now says both sizes, what was sent and what it was before compressing, so anyone can see at a glance if the compression ever stops happening. ### Fixed - A Dealer OS too old to read a compressed push used to come back as a bare "answered 400". It now says in words that the Dealer OS side needs updating first. - A deep read of a Carwow car never read its MOT. The reader looked for labels Carwow does not use, so MOT history said "Not read yet" on every Carwow car on the shortlist. It now reads Carwow's own MOT panel: the date of the last test, pass or fail, the expiry and that test's advisories. Checked live on PH58EEE and BL18YCN. - Wheels and tyres said "Not read yet" on every car after a deep read. Carwow's Wheels and Extras panel is now read (tyre damage, alloy damage, a missing locking wheel nut or toolkit), and a Motorway or Carwow car whose wheels are all clear now says "No problems declared" instead of looking unread. ## v3.14.146, 2026-09-22 ### Fixed - **A freshly won Motorway car came in about £400 too cheap.** Steven: "the purchase price of BC17OXN has came over into DOS totally incorrect". BidBrain read the Motorway purchases list figure as the all in total every time, and worked the car's own price back out of it by taking off the banded buyer's fee and its VAT. That is only true once Motorway has worked the fee out. While a car is still going through, Awaiting confirmation or Reviewing documents, there is no payment breakdown yet and the list figure already is the bid, so about £400 was coming off a number that never had it on. BC17OXN read £7,636 instead of the £8,066 Steven bid, and its all in read £8,066.80 instead of £8,496.80. Checked against Motorway's own sale recap emails for every car bought between 10 and 21 September: every car still going through was short by its own fee and VAT, and every completed car was right, so the fault only ever showed on the newest buys and cleared itself once Motorway published the breakdown. `db.provisional_motorway_bid` now takes the row's status and only takes the fee off once the fee is in the figure (`db.motorway_fee_in_list_price`), `purchases_run.estimated_fee` passes it through, and `list_purchases` builds the all in up from the car price on a car still going through rather than using the list figure as the total. The seven live cars still on a pending status correct themselves on the next purchases run. 10 new checks in test_pricing.py (1,797 passing). Touches Mark: yes, it is Motorway's own behaviour, so his copy reads the same cars the same way; no setting, no default, no scheduled job, and nothing here acts on a platform. ## v3.14.145, 2026-09-22 ### Added - **The seller's modifications answer now travels with a starred car's deep read.** Steven asked why Modifications was missing from the new shortlist car card in Dealer OS (2026-09-22): the Mac never read it, so there was nothing to show. When the Mac reads a starred car's listing in full it now reads whether the seller declared any modifications, on Motorway from the car's own page data and on Carwow from the Modifications line, and sends it up as `deep.modifications`: the seller's own words when there are any, "None declared" on a No, and nothing at all when the page did not say, never a guess. Dealer OS v4.5.26 shows it as the tenth fact on the car card. Motorway's own name for the answer has not been seen live yet, so the read tries the names it would plausibly use and the first live read writes every path with "modif" in it to data/logs/deep_read.log, so the next fix, if one is needed, needs no guessing. Touches Mark: his Mac reads the same extra line on the cars he stars. No new setting, no new default, no new scheduled job, and nothing here acts on a platform. ## v3.14.144, 2026-09-21 ### Fixed - The run bar used to vanish the instant a run finished, no trace of the real result anywhere except a toast gone after 5 seconds. Reproduced live on Didn't win full history, the bar hid the exact same tick the job finished. A finished run's own real result now stays in the bar for 8 seconds, the same brief treatment an error already got. A second gap found proving this live: nothing polled a finished run's own watch again once it stopped running, so the bar could get stuck showing that same old result forever, confirmed still there over 30 seconds later. One more poll after the 8 second window now retires it, or hands off cleanly to a genuinely new background push that has since taken the bar over. ## v3.14.143, 2026-09-21 ### Fixed - The Dealer OS push was down since this morning, the list too big to send. A genuinely large day, 2,078 held back cars plus a normal day's purchases and lost bids, came to 5.17 MB, and the old photo trim only ever saved 104 KB, nowhere near enough. Even with held back and lost bids dropped to nothing at all, the live auction list and purchases alone already came to 3.54 MB, past the old 3.5 MB cap before any trimming happened at all, since that cap was set when a normal day was well under 2 MB and never revisited as the push grew. Raised to 3.85 MB, a real margin under Dealer OS's own stated limit of about 4 MB. A genuinely large day now also caps the held back and lost bids lists to 200 rows each as a last resort, held back already sorted cheapest first so the cars nearest the buying threshold survive, lost bids keeping the most recent. Purchases and every settings block are never capped, Dealer OS draws its own pages from them whole. ## v3.14.142, 2026-09-20 ### Added - **The same car won on two platforms is now caught, named and held.** Every platform keeps its own purchase row, so a car won on both Motorway and Carwow was two live records and nothing joined them up. It has happened for real twice (ND20NVZ and the Ford Puma) and the only reason it did not cost money is that somebody noticed. Now the purchases page carries a red warning at the top the moment a second live win lands on a plate that already has one, and both rows carry a tag on the registration: green Keep on the cheaper sale all in, red Cancel on the dearer one. The warning says in one sentence which to keep, which to cancel and what cancelling saves. Both timelines get a line saying it happened, the minute it happened. - **Nothing moves on either car until one sale is cancelled.** Neither goes into the accounts email, so accounts are never told to pay twice for one car, and DealerKit refuses to take either one whole, so it never ends up with two records for one car. Cancel the losing sale on its platform, cancel it here, and the warning and both holds clear themselves. There is nothing to remember and nothing to undo. - The warning shows in Dealer OS as well as on the Mac's own page, with the same sentence in both places, because the sentence is written once on the Mac and sent as it is. - BidBrain still never touches a platform. It says which sale to keep and why; cancelling on Motorway or Carwow is a person's job, exactly as before. - Nothing is recommended when the app cannot honestly choose: a sale with no price on file yet, or two sales costing exactly the same, get the warning and both holds but no Keep or Cancel tag. - The same plate won twice on ONE platform counts too, and so does a second win days later, not only two on the same day. ## v3.14.141, 2026-09-20 The Carwow shortlist sync now asks Carwow's own shortlisted vehicles page whether a car is starred, both before it clicks and to confirm the click. That is the very page Steven looks at, so BidBrain and Steven can no longer disagree about whether a car is on the shortlist. Why: v3.14.137 had the Mac read two cars' listing pages without touching either, FL16MXJ, which Steven confirmed was on his shortlist, and ND19HZG, which was not. The two came back identical to the character. The car's own page on Carwow does not say whether the car is shortlisted, so the heart on it could never answer the question, and the sync has been calling every successful star a failure since the day it was built. What changes: a car already in the state you want is now never clicked at all, which matters because the heart is a single switch and a needless click turns a correctly starred car off. A car that needs changing is clicked once and then confirmed against the shortlist page. An unreadable shortlist page, a login screen, a redirect somewhere else or a page carrying no listing cards at all each stop the car with a clear reason rather than reading as an empty shortlist, which would have quietly turned every starred car off. A shortlist that is genuinely empty still reads as empty, which is a real answer. Touches Mark: yes, his Carwow stars have never arrived either, and this is the release that gets them there. Nothing else changes for him: no new setting, no new default, no scheduled job, and the shortlist toggle is still the only platform action BidBrain ever takes. ## v3.14.140, 2026-09-20 ### Fixed - A car whose seller kept their private plate stops reading as missing on DealerKit. The platform lists such a car under the private plate, but the car sells, and so reaches DealerKit, under the plate underneath (T18TBA selling as YL66HWE is the worked example in CLAUDE.md). Every DealerKit lookup matched one plate exactly, so those cars read as having no DealerKit record at all: no check, no chip, no transport invoice, nothing, for good. Every lookup now tries the plate the purchase is recorded on first and then the plate underneath, and the whole stocklist is still read only once. The DealerKit panel on a purchase says which plate the car was found under when it was not the one at the top of the card, and says both plates were tried when it found nothing. ### Added - BidBrain now keeps the two plates a car is known by. It has read the plate underneath since June, for the valuations, and thrown it away every time, so nothing that looked a car up by plate afterwards could ever try it. Each run records the pair as it reads it, and keeps it for good, because a car bought today can reach DealerKit weeks later. A purchase picks its own second plate up from that, whichever of the two plates the purchase itself was recorded on. A car whose listing never stated one can have it typed in on the card, under Sells on plate, and a value that is not plate shaped is refused rather than quietly widening a search. ## v3.14.139, 2026-09-20 ### Fixed - The Mac reads its own version correctly when a commit carries two tags for the same release. v3.14.138 was also cut, by mistake, as a bare "3.14.138", so two tags sat on the same commit. git lists them in name order, the bare one came first, and the Mac would have read itself as "3.14.138" while the newest release said "v3.14.138". Those never match, so the Software update card would have offered an update that was already installed, again and again, and a Mac with auto update switched on would have restarted itself every couple of minutes. The real release tag, the one with the v, now wins. ## v3.14.138, 2026-09-20 ### Fixed - The stock gap badge and the top picks count one real model as one model. They grouped a car by the first word of the make plus the first word of the model on its own, so wherever a platform prints the engine badge where the model should be, one model was counted as two: a BMW 118D never met the BMW 1 Series already in stock and read as a gap, a Mercedes A 180 never met an A Class, an MG Motor UK ZS keyed on "motor", and Motorway's truncated "Insig" never met an Insignia. The top picks score partly on that gap and on the model's own sold record, so they inherited it too. All of them now group through CompMatch's own canonical key, which already solved exactly this, so the badge, the sold history note, the taste match and the picks all land on the same group as the suggested bid. Annotation only, nothing about pricing, the hard gate or any buying rule changes. The learned model figures rebuild themselves once from the sales history already on file, so a model really sold before stops reading as never sold. ## v3.14.137, 2026-09-20 Carwow's shortlist panel, as Carwow serves it, now goes into data/logs/deep_read.log every time the shortlist deep read opens a Carwow car. The deep read only ever reads a page, it never clicks anything, so a car that is on the shortlist can be put beside one that is not without touching either. Why: v3.14.135 got the click through to Carwow, and Steven's own shortlist screen shows the cars going on, yet the page read straight after still says the car is not shortlisted. So the click is right and the reading of the state is wrong. No session away from the Mac can open a real Carwow listing, so the only honest way to fix the reading is to have the Mac report a starred car's panel and an unstarred one's and build the check on the difference, never on a guess (golden rule 4). What is reported, each on its own short line because a log line is cut at 400 characters on its way to Dealer OS: the classes on the frame, on the panel and on each marked icon, the form's action, method and fields, any label or wording, and the words the panel shows. The forgery token is left out, it changes on every read and says nothing about the state. Nothing about the sync itself changes in this release, and nothing is clicked, so Mark's copy behaves exactly as it did. ## v3.14.137, 2026-09-20 Carwow's own shortlist panel now goes into data/logs/deep_read.log every time the shortlist deep read opens a Carwow car. That read only ever reads a page, it never clicks anything, so a car that is on the shortlist can be put beside one that is not without touching either. Why: v3.14.135 got the click through to Carwow, and Steven's own shortlist screen shows the cars going on it, yet the page read straight after still says the car is not shortlisted. So the click is right and the reading of the state is wrong. No session away from the Mac can open a real Carwow listing, so the only honest way to fix that reading is to have the Mac report a starred car's panel and an unstarred one's, and build the check on the difference rather than on a guess (golden rule 4). What is reported, each on its own short line because a log line is cut at 400 characters on its way to Dealer OS: the classes on the frame, on the panel and on each marked icon, the form's action, method and fields, any label or wording, and the words the panel shows. The forgery token is left out, it changes on every read and says nothing about the state. Nothing about the sync itself changes, and nothing is clicked, so Mark's copy behaves exactly as it did. ## v3.14.136, 2026-09-20 ### Fixed - The Carwow shortlist sync stops clicking the same car over and over. v3.14.135 got the click through to Carwow, and Carwow accepts it and answers its usual success, but the car still reads as not shortlisted straight afterwards. Steven then checked Carwow by eye and several of the cars reported as failures WERE on his Carwow shortlist. So the click lands and it is the reading of the result that cannot be trusted, and that is worse than it sounds: the shortlist button is one switch, so a car read as not starred when it really is starred gets turned off by the next attempt. From now on, when the click genuinely reached Carwow, that car is not tried again. It stays on file with its real reason, and starring it again in the app starts it over. ### Added - A Carwow shortlist failure now carries Carwow's own panel for that car, the markup and all, into the run log and the error. No session can open Carwow to look at a real listing, so without this there is nothing to work from but guesses, and guessing is not allowed (golden rules 4 and 5). This is what the next reading of the state will be built on. The panel is reported as it read before the click as well as after it, since a car that is already on the shortlist only shows how Carwow marks a starred car until something changes it. ## v3.14.135, 2026-09-19 ### Fixed - Starring a Carwow car now genuinely reaches Carwow's own shortlist. Every Carwow star and unstar had been failing, in both directions, for as long as the sync has existed, while Motorway's syncs fine in the same pass. Carwow loads the code behind its own star button lazily, only once the button appears on the page, and that finishes after the page itself has loaded. BidBrain waited for the button's markup, which Carwow sends straight away, then clicked it. A click that lands before the button is alive does nothing at all, so nothing was ever sent to Carwow, and the car stayed exactly as it was. Read from Carwow's own published code, not guessed. Three changes: BidBrain now waits for Carwow itself to report the button live before clicking; it reads and clicks only inside the panel belonging to that one car, so another car's heart on the same page can never be read or clicked by mistake; and it confirms the result against a page Carwow serves fresh, rather than a reload the browser may answer out of its own cache. When it still cannot set a car, the error now says which of those three it was, instead of only that the car did not change. This fixes Mark's copy in exactly the same way, his Carwow stars were never reaching Carwow either. ## v3.14.134, 2026-09-19 ### Fixed - Carwow cars are back on the list. Carwow put a "view details" link beside the Former keepers count on the car's own page, so the line reads "2 view details" instead of "2". BidBrain was reading that line as a plain number, got nothing, and a car with no previous owners count fails the hard gate, so every single Carwow car was rejected. Four days with no Carwow cars at all, on Motorway cars only. The count is now read from the front of the line, and a line that does not start with a number still leaves the count unset, so a page that changes again holds the car back rather than guessing. ### Changed - A platform whose stock reads fine but whose every car then fails the gate now shows as a failed read in the Run modal, with the commonest reason on it, instead of a healthy looking "0 cars". The reasons were already printed in the run log, where nobody reads them, which is how the Carwow break above stayed invisible for four days. A genuinely quiet day reads the same way, which is the point: it is worth a look either way. ## v3.14.133, 2026-09-19 ### Fixed - Damage on a starred car now carries the photo of that damage, on both platforms. The two entries below shipped together in one merge, so they are one release rather than two: a release only carries its own section into the Software update card, and split across two versions half the work would never have been read. - A Carwow car's declared damage now carries the photo of that damage, the way Motorway's already did. Carwow does say which photo belongs to which fault, in its condition report markup rather than in the page's words, so `deep_read.from_carwow` had been listing the damage with `photo_url` left null and dumping every photo into one undifferentiated gallery. `carwow_damage` reads each damage block instead, giving where it is, what it is and its own photo: "Rear driver alloy: Scuffed alloy". Read live off listing 14611801 and proven against six live listings, where the count matched each page's own "N areas of damage" line exactly and every damage had a distinct photo. The page also carries a damage slideshow, deliberately not used: it repeats its slides to loop, so going by position would pair the wrong photo with the wrong fault. A page with no condition report still falls back to the page's words exactly as before, so nothing that read today stops reading. Nothing here is dealership specific, so this reaches Mark's Mac and Steven's the same way, with no new setting and no new run. - A Motorway car's damage was being listed twice over and padded with wheels. Motorway splits a damage in two, the words in `conditionAndDamage` ("Dents", "1 small (0-5cm)") and the picture in the gallery (`kind: damage_dents`), and the read handed both halves to the screen as separate entries: the words with no photo, then a bare label with the photo. On top of that any gallery photo flagged `isDamaged` counted as bodywork damage, so wheels and tyres came back in through the gallery after v3.14.131 had deliberately moved them to their own line. Read live on 2026-09-19: a car whose only fault was three scuffed alloys and two tyres showed five "areas of damage", and an Audi A5 with four real faults showed twelve entries. `motorway_damage` now joins the two halves on what the photo is of, so each damage appears once carrying both the seller's words and its own photo, and wheels and tyres stay out of bodywork (they are already on the `tyre_notes` line, and their photos are still in the gallery). The three live cars went from 5, 10 and 12 entries to 0, 3 and 4, every one with its photo. Nothing is dropped: a photo of something never declared keeps its own name, and a fault declared with no photo still reads. Nothing here is dealership specific, so this reaches Mark's Mac and Steven's the same way, with no new setting and no new run. ## v3.14.131, 2026-09-18 ### Fixed - Wheels and tyres, and a warning light on the dashboard, now show as their own lines (`tyre_notes`, `warning_lights`) instead of being folded into the bodywork `damage` list, where Steven found them buried under one "Damage" bullet and read them as missing. No platform gives a plain "mechanical fault" flag; what shows up about the mechanicals comes through in the MOT's own advisories, which already have their own line. ## v3.14.130, 2026-09-17 ### Fixed - The shortlist deep read's last unmapped Motorway names (fourth live round, 2026-09-17): the MOT block's status, last test date and advisories (Motorway never gives an expiry date, so none is worked out; the test date goes as the new `mot_tested`), the colour and interior trim off the specification block (`interior` is new too), and the wheels and tyres block's tyre problems, scuffed alloys, missing locking wheel nut or tools as condition lines. Motorway shows no tread depths, so tyres stay "not read" honestly. ## v3.14.129, 2026-09-17 ### Fixed - The shortlist deep read takes its photos from Motorway's gallery block (seen live 2026-09-17): every photo of the car once, large, the seller's documents left out, and the damage photos named and carried with the damage list so the screen can mark them. A declared damage line no longer carries a line break inside it. The run log's key notes show one image's worth of the gallery, not eighty lines of it, so the names still to map (MOT date, colour, tyres) can be seen. ## v3.14.128, 2026-09-17 ### Fixed - The shortlist deep read now reads Motorway's declared condition and history check by the names Motorway really uses (seen in the run log after v3.14.126's first live round): dents, scratches, paint, missing parts, windscreen, warning lights and smoked in, each a Yes with its description, become the damage list; hasBeenStolen, hasBeenImported, the two written off flags and salvaged feed the history check; the MOT date is taken off the HPI block when it carries one; every keeper's start date becomes the keepers list. Before this a car with declared scratches read as "None declared" and stolen or imported never showed. The run log now prints up to 80 names, the still unmapped ones (MOT, tyres, colour, gallery) first. ## v3.14.127, 2026-09-17 ### Added - The Mac updates itself (Steven 2026-09-17, "make the Mac update itself"): a new Connections switch, "Update itself when a new release is published". With it on, every two minutes while nothing is running the Mac looks for a newer BidBrain release on GitHub and installs it the same way the Update now button does, restarting BidBrain for about a minute. Off by default, so Mark's Mac keeps the button as it is; it is turned on from the Mac's Settings > Connections or Dealer OS's Automation switches card. A run in flight or a dirty tree is left alone and tried again next time. ## v3.14.126, 2026-09-17 ### Fixed - The shortlist deep read's first live round (ten Motorway cars, 2026-09-17): six came back blank with no error. Their page carried no car block at all (the sale had closed by the time they were read), and the read called that "read". A page with no car block is now a failed read that says so: "Motorway's page for this car carried no car data. The sale may have closed or the listing been withdrawn." The read now works inside Motorway's real car block (props.pageProps.vehicle) and uses the names Motorway really uses for the service book (serviceRecord, the official and independent stamp counts, the records' service centre) and the history check (hpiHistoryCheck's finance agreements count and clocked mileage flag, details.onFinance). The names it saw for the MOT, keys, colour, damage, tyres and photos are printed into the run log on every read, so the diagnostics document shows them for the next round. ## v3.14.125, 2026-09-17 ### Added - The shortlist deep read, the Mac's half of Dealer OS v4.1.0's shortlist screen (one starred car at a time, everything on one page). "Read this car" or "Read them all now" there logs a shortlist.read action naming only the starred cars; the sync queues them and starts a quiet run that opens each car's own page in the saved login, reads the MOT and its advisories, the service book, the keepers, colour, keys, tyres, declared damage, the platform's own history check, every photo at full size and the reserve and bid as the page shows them now, and the next push carries it under each car's deep block. Motorway (its page data, then its words), Carwow, DealerWay and Auction4Cars; Dealer Auction has no car page read yet and says so. A dead login or a page that will not load is answered in plain words, never left blank, and a field the page did not show is null, never a guess. Never a bid and never any action on the page. Asked for while the Mac is busy, it runs on the next quiet moment. Both Macs get it, nothing about buying, collection or DealerKit changes and no default moves. Motorway's real field names are written to data/logs/deep_read.log on every read, so the first live read shows which names still need mapping. ## v3.14.124, 2026-09-16 ### Fixed - Proven live: only the last of the survey's real pictures made it through diagnostics' own 300 line / 256KB tail window, the earlier ones pushed out. The tail is now 500 lines / 512KB, and the survey keeps pictures at its two most useful screens only (the item window, the Edit Expense window on an existing supplier expense), so both survive every time. ## v3.14.123, 2026-09-16 ### Fixed - The costs screen survey wrote every picture the walk had appended, then wiped them out with a final overwrite of the same file (found live: the first real survey after v3.14.122 shipped no pictures at all). The survey now appends its own text lines the same way, so nothing after it ever overwrites what the walk already wrote. ## v3.14.122, 2026-09-16 ### Added - A real picture, not just words, when a DealerKit step fails or the costs screen survey runs (Steven, 2026-09-16: "is there any way at all that you will be able to make it so you can sign in to dealerkit from here?"). No password or browser session moves off the Mac: instead the Mac clips a small, low quality JPEG of the window it was looking at, writes it as plain text (base64) into the same log files diagnostics already ships, and a small script on the Dealer OS side turns it back into a real picture. Every genuine push_followup / push_purchase failure carries one now; the survey carries one at each of its four key screens. ## v3.14.121, 2026-09-16 ### Fixed - A line left alone on the purchase expense reads "already right on DealerKit", not "changed from £289 to £289". - A chip sent to DealerKit stamps the price and the fee as sent too, since the chip puts both there. - A line's own menu on the Edit Expense window is looked for in that window only (the survey: each line is a table row with its own menu). ## v3.14.120, 2026-09-16 ### Fixed - Correcting an existing purchase expense only deletes and re adds a line that is there and differs; a line already at the figure is left alone and a missing one is just added (ND66XYJ, sixth attempt: the fee already matched, the delete of its line failed, and the missing price never got added). The line menu, Delete and Save & Approve failures now carry the window's words. - The costs screen survey also writes down the Edit Expense screen of an existing supplier expense, with each line's own buttons. ## v3.14.119, 2026-09-16 ### Fixed - Each expense line now proves it landed before the next starts (ND66XYJ, fifth attempt: the Carwow expense was created with the £289 fee only, the £4,110 chassis line typed first was lost when its Add did not take and the fee line wrote over it). The Unit Net box is filled through Playwright's own fill and blurred, then read back; the item window's Add is pressed again after a blur if the window stays open; the form must list the line before the next one is added; every failure carries the window's words. ## v3.14.118, 2026-09-16 ### Fixed - Adding a line to a DealerKit expense, written from the real screens of Steven's account (the costs screen survey): the form's own ADD opens a separate Add Expense Item window with the Item Category already set ("Cost of Purchase > Chassis" on a fresh form), Item Description, Qty, Unit Net £'s, VAT Tariff preset to "Z - Zero", and its own Add. The category is opened only when it does not already read the line wanted; the Expense Categories window then takes the group and the line, scoped to that window so the same words behind it are never clicked; the VAT tariff is set outright every time (No VAT on the chassis, Standard on a fee); the Unit Net box and the Add button are found in the topmost window. Mark's older in row picker ("Pick a category") still works. ## v3.14.117, 2026-09-16 ### Added - A DealerKit costs screen survey (purchases_run.py --survey-dealerkit, the dealerkit_survey run): opens one car's Edit Vehicle > Expenses, presses + EXPENSE, picks the team and the supplier in the unsaved form, presses the item ADD and opens the category box, writing every screen down (text, buttons, fields, options) to data/logs/dealerkit_survey.log, then leaves without saving. So the add an expense steps can be written from the real screens of this account instead of one press at a time. ## v3.14.116, 2026-09-16 ### Fixed - The category box on a new expense item line is found by meaning (any wording with "category" in the open window), a searchable picker is typed into when no group list appears, and a missing Unit Net box says so with the window's words (third attempt on ND66XYJ: the ADD button was right at last, the category box was not worded "Pick a category"). - A failure's saved screen now holds the whole topmost window: its buttons, its field labels and placeholders, and all of its text, not just a warning line. ## v3.14.115, 2026-09-16 ### Fixed - The Add Expense form's item ADD button, seen for real on Steven's account (the window's words from v3.14.114): the form offers "+ EXPENSE", "+ CREDIT", an icon only "+" beside the supplier picker and the item row's own "ADD" at once. A button that starts another expense, a credit or a contact is never pressed for an item; the one reading "ADD" is. - A failure's "The window says" quotes the topmost open window, not the Edit Vehicle menu underneath it. ## v3.14.114, 2026-09-16 ### Fixed - The new purchase expense on a car with none (ND66XYJ, first real use of v3.14.113's route): the item ADD button is chosen by its words, never the icon only "add" beside a contact picker, and the supplier is picked from the picker's own options, never the vehicle list behind the modal. - When a DealerKit step cannot find what it needs, the failure now carries the window's own words, and the full text goes to data/logs/dealerkit_step.log for the next diagnostics pull. ## v3.14.113, 2026-09-16 ### Fixed - The purchase price, buyer's fee and transport fee now always reach DealerKit on a send (found live on ND66XYJ and FJ67WFR: the price box at creation saved nothing, the send then said "nothing to send, BidBrain had no confident figure" although it had one, the fee was refused with "send the purchase price first", and DealerKit held no purchase figures at all). One route now: a price saved at creation counts; otherwise the three figures go on the supplier's purchase expense together, created when the car has none, corrected when a line differs, left when it matches. A figure that cannot be written is said per line, in words, never silently. - Why a price was not saved at creation is now in the log and on the timeline (the Bought From contact, the price box, or DealerKit not saving), instead of being swallowed. - The Bought From contact is matched without regard to case or spacing ("Carwow" on DealerKit, "CarWow" in the code). - Check DealerKit printed the whole expense object where the price should be; it prints the figure. ### Added - The automatic DealerKit run fills in a price that never landed (dk_figures_missing), and never writes over a figure a person typed into DealerKit; that case is read back so Dealer OS shows it with the Update the price and fee button. A send reads the record straight after, for the same reason. - A new purchase expense is filed under the dealer's dealerkit_team setting when one is saved, else the only team, else the team with Sales in its name, else the first, and the log says which. ## v3.14.112, 2026-09-16 ### Changed - Nothing about the seller's finance goes to DealerKit (Steven, 2026-09-16: "the settlement figure never has to go to dealerkit ever. it is always handled at the point of collection or before"). The settlement typed at creation and the form field note from v3.14.108 are gone. The reading of the purchase page, the NEGATIVE EQUITY answer and the figures on the car in Dealer OS stay. ## v3.14.111, 2026-09-16 ### Fixed - Motorway prints the pence of a settlement on the line after the pounds ("£10,370" then ".04"); the reader now adds them back, so YL19PLO's settlement is £10,370.04 and its negative equity £1,191.04, not a pound short. Proven on the second real read: Motorway prints "Amount to finance company" (N/A with no finance) and "Amount to seller", Carwow prints "Settlement amount" above the same two lines. ## v3.14.110, 2026-09-16 ### Changed - The first real purchases read (2026-09-16) showed Carwow printing "Settlement amount" above "Amount to finance company", and Motorway printing only "Amount to finance company". The reader now takes Carwow's settlement amount first and Motorway's amount to the finance company as the settlement when nothing else names one. The finance lines go to the detail log one short line each, with the two lines after every label, since the diagnostics document cut the long lists. ## v3.14.109, 2026-09-16 ### Fixed - The Dealer OS push carried the raw saved On finance? list, so a dealership with its own saved answers never saw NEGATIVE EQUITY in Dealer OS. The push and the settings card now carry the effective answers, defaults and the new answer merged in. ## v3.14.108, 2026-09-16 ### Added - Outstanding finance on a bought car, from the platform's own purchase page (Steven, 2026-09-16): the seller's finance settlement figure and any negative equity (settlement above the price agreed, which the seller pays as a shortfall before the car moves), on Motorway and on Carwow, refreshed on every purchases read (`finance_settlement`, `negative_equity`, both in the Dealer OS push). Every line naming finance, settlement or equity is written to the platform's detail log once per car, so the labels are tightened against the real wording after the first read; a figure is only ever taken from a labelled line. - On finance? gains the answer NEGATIVE EQUITY (red) on both dealerships, added to a saved list once. A car whose page shows negative equity is marked so by the purchases read; one with a settlement and no answer yet is marked ON FINANCE. A person's NO is never overwritten. - A DealerKit send types the settlement figure into the Stocklisting step's finance settlement field when the record is created, beside the price, and the timeline says the settlement and the negative equity in words. The form's field names are printed to the run log on the first send so the field can be matched exactly; a form with no such field leaves the figure for a person and says so. ## v3.14.107, 2026-09-16 ### Fixed - "Price and fee" asked for from Dealer OS was still refused with "already match what is on file" by the row job's own preflight, which compared against what was last sent rather than reading DealerKit (FJ67WFR: a fee never sent at all). A row job is always a person's ask, so it now reads DealerKit for real and only says "already right" when it is. ## v3.14.106, 2026-09-16 ### Fixed - A DealerKit ask from Dealer OS (a whole send, or one item such as the price and fee) that found the Mac busy with another job was refused with a note only the server log saw, so the press vanished (Steven, 2026-09-16, FJ67WFR: "pressed it", nothing on the timeline). Now the ask waits until the Mac is free and is sent then, and the timeline says "waits until the Mac is free" at once. Check DealerKit alone still says try again. Dealer OS sync notes are flushed to the log at once. ## v3.14.105, 2026-09-16 ### Added - The price and buyer's fee can be asked for from Dealer OS on a car already on DealerKit (`purchase.dk {action: "figures"}`, 2026-09-16, FJ67WFR: the fee never landed and there was no button for it). The Mac reads what DealerKit holds now and sets what differs; "already right" when nothing does. An ask from a person is never refused as "already matching what was sent", only the automatic sweep uses that test. ## v3.14.104, 2026-09-16 ### Changed - DealerKit knows nothing about chips (Steven, 2026-09-16), so the timeline no longer says a chip was sent. A whole send on a chipped car writes "the price sent, £6,300, already takes the £500 chip off", and the chip follow up is labelled "Price and fee after the chip" in the outcome lines and the timeline. Nothing changes in what is sent. ## v3.14.103, 2026-09-16 ### Added - A chip recorded by mistake comes off again (Steven, 2026-09-16: Donna recorded FJ67WFR's £500 chip twice, once on inspection and once in the check in box, so the price on file read £5,800 against DealerKit's £6,300). `purchase.unchip {id, chip_id}` from Dealer OS deletes the chip, refreezes the remaining chips' bid before figures, writes the timeline, and clears the chip and price stamps so the DealerKit tab shows the price as a change to send. Nothing goes to DealerKit on its own; a person sends it. The Mac's own screen is unchanged. ## v3.14.102, 2026-09-16 ### Added - A document Dealer OS holds for a bought car goes to DealerKit with the car's other documents (Steven, 2026-09-16: the MotorCheck history check PDF). Dealer OS logs `purchase.document` with a link; the Mac fetches it with its own token, keeps it under data/purchase_docs// and on the row (`dealer_os_documents`), and sends it to DealerKit at once when the car is already there (the documents item), or with the whole send when it is not. A link not on Dealer OS, or anything that is not a PDF, is refused and written on the timeline. Nothing changes for a Mac whose Dealer OS never logs the kind; Mark's copy sees no new behaviour until RightDrive runs history checks. ## v3.14.101, 2026-09-15 ### Fixed - A send or follow up no longer gives up when DealerKit's Edit menu has not finished drawing: the Expenses, Images and Documents tiles are looked for up to six times over nine seconds, and Edit is pressed again if the menu closed. Found live on HX67UMH, where the buyer's fee step reported "the Expenses tile was not found" on a record whose expenses were plainly there. - When a send dies part way, its timeline line now says what it had already done, so a failure after the photos or the price reads as such. ## v3.14.100, 2026-09-15 ### Changed - A send on a car DealerKit already has now treats documents the way it treats photos: when DealerKit already holds at least as many documents as the send would add, they are left alone, so a car put on DealerKit by hand and then sent from Dealer OS never ends up with its V5 or service records twice. Price, mileage and fee were already compared first and only set when missing or different; a retail price DealerKit has is never overwritten by a whole send. ## v3.14.99, 2026-09-15 ### Changed - How many photos a whole send puts on DealerKit is now a buying rule, Photos sent to DealerKit, 1 to 20, set from the Buying rules screen or from Dealer OS. 4 when nothing is saved, which is what every copy did before. - A chip recorded before a whole send is stamped as sent by the send, because the purchase price it writes already includes the chip. Before this the DealerKit tab read "1 change to send" after a send on a chipped car and the 16:15 run set the same price again (found live on LP68RHK). ## v3.14.98, 2026-09-15 ### Changed - The ferry crossing rule (hide cars in Ireland, the Isle of Man, the Isle of Wight, the Channel Islands and the roadless Scottish islands) is now off by default on a Mac where nobody has saved the setting. Steven asked for it off on Mark's copy on 12 September, and the old default needed a line in each Mac's own config file that never reached Mark's Mac. A saved value always wins, and Dealer OS's buying rules card can now switch it either way (Dealer OS v3.15.74). Really Easy Car Credit's copy is switched back on from Dealer OS the same day. ## v3.14.97, 2026-09-15 ### Changed - A whole send to DealerKit now finishes its own car's follow ups in the same DealerKit session: a check in pressed while the send was still running, a chip at the door, a transport fee set meanwhile. Dealer OS's new Collected button sends the car when the driver has the keys and Check in follows when it lands; before this a check in a minute after the send was refused as busy and waited for the 16:15 run. - A DealerKit follow up refused only because the Mac was busy on another job is asked again the moment the Mac comes free, instead of waiting for the 16:15 run. After a restart of the BidBrain server the 16:15 run is still the safety net. ## v3.14.96, 2026-09-15 ### Changed - When the Motorway export brings no file, the download call is asked again every three seconds for up to two and a half minutes through 500s as well as 202s (the one off checks: a fresh session gets the file by hand, the Mac's gets "Internal server error" on its very first ask), and every answer in that sequence is written to the export log with Motorway's own request id, for their support if it comes to that. ## v3.14.95, 2026-09-15 ### Changed - The Motorway one off check lists the stored site data on the Mac's Motorway profile, tries the export as the run does, then clears that stored data (login untouched) and tries again; the export log now also records exactly what the browser asked Motorway for. A real browser window made no difference (v3.14.94 check: the same "Internal server error" both ways). ## v3.14.94, 2026-09-15 ### Changed - The Motorway one off check tries the export the way the run does it (no browser window) and then, if no file comes, with a real browser window, and writes down what the browser calls itself. Steven pressed Download by hand and a file came, so the difference is the browser, not Motorway's server. ## v3.14.93, 2026-09-15 ### Changed - When the Motorway export brings no file, what each of Motorway's calls answered is written to the export log as its own short lines, so the words of Motorway's server error (the download call answered 500 again on the one off check after v3.14.92, after first saying "still preparing") reach Dealer OS's diagnostics uncut. ## v3.14.92, 2026-09-15 ### Fixed - The Motorway CSV export, failing on every run since 8 September. The one off check showed why: pressing Download now makes Motorway's page call "generate" (answered at once with an empty JSON string) and then poll "get" (answered 202, still preparing, until the file is ready), and BidBrain stopped waiting five seconds after that first JSON answer, took it for the file, found it was not one and gave up. The export now waits the whole time for the browser download while Motorway only says "still preparing", and when no download comes it asks the download call itself every five seconds for up to four minutes until the file arrives. ## v3.14.91, 2026-09-15 ### Added - A "Motorway one off check" job (`motorway_probe`, from Dealer OS's run list): presses Download once and records what Motorway's export calls answered, then opens the first two cars on the list and records every line of each page naming CAP and every £ figure, all to `data/logs/motorway_probe.log`. Reads only, no valuations, no shortlist, no push. A CSV that does arrive clears the failed marker so the night's run uses it again. ## v3.14.90, 2026-09-15 ### Changed - CAP clean on a Motorway car read from its page: the page data carries no CAP clean at all (checked on the run after v3.14.89), so the figure is now also read from the page's own words ("CAP clean £5,195", the way Carwow's page is read), and the car page keys log writes down every line of the page naming CAP, so the next run shows the exact wording Motorway uses. - When the CSV export does not come, the export log now records what Motorway's generate and download calls actually answered (the download call has been answering with a server error since 8 September), so the export can be put right from the log. ## v3.14.89, 2026-09-14 ### Fixed - CAP clean on Motorway cars read while the CSV export is failing (since 8 September) was Motorway's own CAP based price for the car, not CAP clean: the read took whichever price entry sat next to "priceSource: CAP" on the car's page (Steven: "its showing cap retail instead of cap clean"). CAP clean is now read only by its own name on the page (`find_cap_clean`) and left unread, shown as "not read", when the page does not carry it, so the value caps and the "looks high against CAP" check never run against the wrong figure. The car page keys log now names every CAP and valuation key, so the next run shows exactly where Motorway keeps it. ## v3.14.88, 2026-09-14 ### Added - Every bought car now keeps its make and model (new `make` and `model` columns on purchases, pushed to Dealer OS with the rest of the row). They come from the plate's own sighting, the very listing the live auction card was drawn from, so a due in car lands under the right card in Dealer OS; a car the Mac never listed, or one added by hand, reads them from its name the same way CompMatch does (BMW 116D is a 1 Series, Mercedes A180 is an A Class, Kia c'eed is Ceed). Every existing row fills in the first time this version opens the database; a row read before its name arrived fills in when the name does. ## v3.14.87, 2026-09-14 ### Changed - One star per listing (Steven: a car on Motorway and Carwow the same day "still shares its ... star between the two cards. Fix this"). The bids table was unique on plate and sale date, so the second listing's star overwrote the first; it is unique on plate, sale date and platform now, rebuilt once on the next start with every row kept. Each card, here and on Dealer OS, asks for its own listing's star; a star on the plate alone (every older star) still lights every card of that plate. Unstarring a card also lifts such an older star so the card really goes dark. The echo to Dealer OS names each star's platform; a star.set from Dealer OS lands on its own listing. ## v3.14.86, 2026-09-14 ### Fixed - A car chipped before its platform showed the buyer's fee never got one: the chip rightly freezes the bid, but it froze the missing fee too. The fee and its VAT now come in on the next read, the chipped bid untouched. ## v3.14.85, 2026-09-14 ### Fixed - No car from Carwow or Motorway is ever stored without a buyer's fee (Steven: "a car should never be added from carwow or motorway from now on without the fees"). The estimate from the platform's own band now applies to every car read from either list, page read or not, Motorway included (its bid is derived under the all in list figure first). A bare price stored earlier with no fee no longer counts as confirmed, so the next sync completes the cars already on the list. ## v3.14.84, 2026-09-14 ### Fixed - Carwow buyer's fees (Steven: "why does it look like all the carwow cars are not getting the buyers fees from carwow?"). Carwow only puts the breakdown on a car's page once it is collected, so the fee stayed blank and, with the price on file, the page was never read again after seven days. Now: a Carwow car with no fee on its page holds Carwow's own published band for the price, marked estimated on Dealer OS, replaced by the real figure the moment Carwow shows it; the page is read until the real fee has arrived (Motorway's estimated breakdown too). - Nothing is chased for ever any more: every page read stops 60 days after the purchase (Steven: "60 days, go"), a cancelled sale is never read, and a Carwow fee estimate that ages out gets one timeline line saying the estimate stands and to check the invoice. - DealerKit is corrected when the figures move. The exact price and fee sent to DealerKit are remembered; when the real figures later differ, the 4:15pm DealerKit run resets the price and buyer's fee lines (the same correction a chip makes) and writes the change on the timeline. Motorway and Carwow alike. ## v3.14.83, 2026-09-14 ### Fixed - A Northern Ireland plate (JXZ 8899) on the Motorway purchases screen was never read: the plate rule was sized for AB12 CDE and dropped a plate whose second half has four characters, so the car never came through and Donna added it by hand (Steven 2026-09-14). One plate rule now, used by every Motorway purchases reader: modern, prefix, suffix, dateless and Northern Ireland plates, with or without the space. - A car added by hand that a platform then lists becomes that platform's purchase, on the same row: everything typed on it stays (status, due in date, notes, mileage, DealerKit stamps), and it gains the platform's own name, figures, photos and documents. One car, one row, with a timeline entry saying so and the hand typed price kept on it. ## v3.14.82, 2026-09-14 ### Changed - A sold car read off its DealerKit record is sent with its mileage and colour, so a DealerKit sale on Dealer OS's Sold tab matches the backfilled ones. - The two Clickdealer sales runs are off the Run menu and the command line (Steven: "i will never have to run this again now all the info is there, dealerkit will handle all sales moving forward"). The code and its tests stay as the record of how the history was built. ## v3.14.81, 2026-09-14 ### Fixed - A Clickdealer sold car's fuel and gearbox are sent plainly (Diesel, not DIESEL), the same as its colour. ## v3.14.80, 2026-09-14 ### Fixed - What the Clickdealer sales probe showed on a real page: the car's photo is the first picture on the images host under /vehicles/, never a gif (the probe took a calendar icon); the derivative is the edition field, never a numeric id (the probe took the variant id); the colour is title cased; the body type is never read off the colour field. - A chassis number, an engine number, a customer, an enquiry, a payment, a price or a supplier field never leaves the Mac, not even on a probe. The probe's first report carried the chassis number in its check keys. - A saved photo is named by its plate and the picture's own address, so a different picture for the same plate gets its own lasting link. ## v3.14.79, 2026-09-14 ### Added - Clickdealer sales history (Steven: "go into clickdealer and search each car that you have on record as being sold and find the sales info of that car to fully back populate it, each car will have photos too, i want just one photo of each sold car in there, the first one"). Two runs on the Run menu. Clickdealer sales probe reads one sold car, the newest, and reports every field and picture found on its page for checking. Clickdealer sales history reads 18 months of the sold vehicles reports, then each car's own page for its year, derivative, engine, fuel, gearbox, mileage and colour, fetches its first photo and sends that to Dealer OS's own file store for a link that lasts (never sent twice), and sends every sale to Dealer OS. Read only on Clickdealer. The saved login is tried first; a window opens on the Mac only when it has expired. New readers parse_vehicle_fields, parse_vehicle_photos, vehicle_details and sold_car_report in bidbrain/readers/clickdealer.py, built without a captured page, hence the probe. ## v3.14.78, 2026-09-13 ### Fixed - The sales backfill sent, for a sold car with no photo on its DealerKit record, the auction photo this Mac keeps as a path on its own disk. A browser cannot show that and it names this Mac's folders. The backfill now sends the lasting link Dealer OS already holds for the photo (the same link the purchases push uses) and never a path; a car whose photo has no link yet is sent without one. ## v3.14.77, 2026-09-13 ### Fixed - The sales backfill read the wrong days off a DealerKit record: delivery_on is the day a car came in (its Due In date, the same field the due in push sets), disposed_on is the day it went. The first backfill sent the arrival day as the sold day and the record's creation day as the arrival. Both the DealerKit sold record and this Mac's own sold purchases now send disposed_on as the sold day and delivery_on as the day it came in; the DealerKit record check keeps disposed_on for that. Run the backfill again and Dealer OS corrects every DealerKit sale it holds. - A sold record's photo is looked for under media and under the vehicle as well, the picture shaped keys on the record are reported for checking, and a sale with no photo takes the auction photo this Mac kept for the plate. - The backfill's one line summary fits the run board, so its counts show. ## v3.14.76, 2026-09-13 ### Added - The sales history backfill (Steven on Dealer OS: "take what you already know from clickdealer history, bidbrain history on my Mac Studio and dealerkit history to build up the backfill of sales"). One quiet run, sales_backfill, sends every past sale this Mac knows to Dealer OS's /api/bidbrain/sales, in the order Dealer OS should trust them: DealerKit's own sold cars, each record read once for its sold_for, its spec, photo and dates; the Clickdealer sales reports already in sales_history, the report's net sale price as the deal price and the days to sell turned into the day the car came in; and this Mac's own purchases that went Sold, at the price DealerKit locked, which only ever fill a gap. Run it again and nothing is written twice. ## v3.14.75, 2026-09-12 ### Added - What a sold car went for, read from its own DealerKit record (Steven on Dealer OS: "cars dont always sell for what we list them up for retail for, we often have to discount them"). Dealer OS sends one sale.price action the moment DealerKit says a car is Sold; the Mac puts the car on a short queue and, if it is free, reads that one record straight away (no window, the saved login), else on the next quiet stage check. The figure goes to Dealer OS's /api/bidbrain/sales with the record's stage and the keys the reader looked at, so the first real sold record settles which key holds the deal price. Only stage 25, Sold, is reported with a price: Awaiting Delivery is not a sale. Nothing on a schedule, never the whole feed. ## v3.14.74, 2026-09-12 ### Added - A per Mac lever for the ferry rule's own default, NO_FERRY_CROSSING_DEFAULT in dealer_config.py. On here, same as before, unless an install sets it to False. Steven asked (see below) for this so it can start off on Mark's side without touching Steven's own copy of this file or Steven's own saved settings, and without either dealership's Buying rules screen needing a value saved on it yet. ### Changed - Steven: "turn it off for mark for now but make sure his claude finds it and prompts him if he needs it on." Mark's copy of dealer_config.py is not something this repo can reach or edit — it lives only on his Mac, never committed. So this is the note for whoever (Mark, or the Claude working from his copy) next opens dealer_config.example.py: set NO_FERRY_CROSSING_DEFAULT = False in Mark's own dealer_config.py, tell him it is off, and ask him whether he wants it on rather than assuming. See the comment against that setting. ## v3.14.73, 2026-09-12 ### Added - No ferry crossings (Steven: "i dont want to see cars in ireland"). A car in Northern Ireland or the Republic, on the Isle of Man, the Isle of Wight, the Channel Islands or a Scottish island a road does not reach is held back with the reason, rather than shown on the list because a straight line said it was close. Judged on the location the platform gives, by town, by county and by postcode, and never on a location nobody could read. Places that share a name with a mainland town (Bangor, Newport, Douglas, Yarmouth) are matched by postcode only, so a good car is never hidden by a coincidence. - Two settings behind it: no_ferry_crossing, on here and off for a dealership that will travel, and ferry_places to add somewhere the built in list misses. Both editable from Dealer OS. ## v3.14.72, 2026-09-10 ### Fixed - The purchase cost read as nothing on every in stock car: DealerKit holds it as an expense object with the Chassis line inside, not a plain number. The figure is taken from the object or its Chassis line now, on the stock read and on the 15 minute stage check alike. The stock read logs the object's shape once so a figure that still will not read can be pinned down. ## v3.14.71, 2026-09-10 ### Fixed - The DealerKit stock records read saved nothing on its first live run: DealerKit sends the model as its own object, and the save refused it. The name is taken from it now. ## v3.14.70, 2026-09-10 ### Changed - The price paid is DealerKit's (Steven: "Dealerkit is always the source of truth"). The stock read now also reads each in stock car's own DealerKit record and keeps its purchase cost, whether BidBrain bought the car or not, sent on the push as dealerkit_stock. The 15 minute stage check keeps each purchase's purchase cost in step too. ## v3.14.69, 2026-09-10 ### Added - Dealer OS push: every car carries its sightings from the last 30 days (sale date, platform, reserve that day), so the card can show how many days ago each one was and how the reserve has moved. Additive, still schema 1. ## v3.14.68, 2026-09-10 ### Added - Dealer OS push: the cars the rules held back go with the list, slim (held_back on the push, additive, still schema 1), so Dealer OS can show them behind a closed row grouped by rule and Steven can check nothing is being held back unnecessarily. - Released from hold: a hold.release action from Dealer OS carries Steven's reason and the rules the car broke. Both are kept in a new released_cars table with who released it; the car passes those rules on every later assessment, the rules become flags on its card under his reason, it moves onto the list at once and the push goes. The price comes with the next run, which values a released car like any other. A hide still wins. python3 daily_run.py hidden_notes now lists the releases too, rules to consider softening. ## v3.14.67, 2026-09-10 ### Fixed - Carwow's finance answer is now read from the MotorCheck section, where Carwow actually states it: every check reads Clear when it is clear, Finance among them, and a car with finance still owing has that row pulled out with the lender, the type and the term. It was being looked for under an "Outstanding finance" label Carwow never writes, so the question went unanswered on every Carwow car. - NO FINANCE is no longer the default answer on the On finance? column. A blank dropdown shows its default faded, which on every other column reads as "not set yet" but on this one read as the car having no finance owing. A Ford Focus with a Santander PCP still running showed a faded NO FINANCE while nothing had answered the question at all (Steven, 2026-09-10). An unanswered finance question now shows as unanswered. ## v3.14.66, 2026-09-10 ### Fixed - A car whose plate is not the modern AB12 CDE shape no longer reads as NO PLATE. Northern Ireland plates (a Kia c'eed on JXZ 8899, live on the board today), the older prefix and suffix styles and private plates are all real registrations; the plate was being read off Motorway's own page and then refused for not matching one pattern. Where a car sells with a private plate kept back, the plate underneath is taken as the registration rather than leaving the car with none. - The rules for what counts as a plate now live in one place (bidbrain/plates.py): wide for a field that promises a plate, still narrow for hunting one out of ordinary text, where "GTI 2000" must never read as a registration. ## v3.14.65, 2026-09-10 ### Fixed - A shortlist now carries the sale it is FOR, not the day the run happened. Both auction houses close at 15:30, so the 17:10 run has always been reading the NEXT sale's cars while stamping them with the day that had just finished. Dealer OS cleared the board at midnight on exactly those cars, hours before there was any bidding to do on them. The list already saved is read as the sale it truly belongs to, so nothing has to be run again to put it right. ## v3.14.64, 2026-09-10 ### Fixed - Carwow's photos are now saved as real files of ours the moment they are read, and uploaded once to Dealer OS's own store, exactly as Motorway's have been since August. Before this the Purchases page pointed straight at Carwow's own image host, which is outside our control: a listing taken down, or hotlinking blocked, and a car's photos were gone with nothing on our side to put back. A car bought in the last 30 days whose photos are still Carwow's gets them saved on the next sync; older cars keep the links they have, since their listings are usually gone. - The saving itself is now one shared piece of code for both platforms. ## v3.14.63, 2026-09-09 ### Changed - A change made in Dealer OS is written on the timeline under the person's name (as the Dealer OS team list has it) rather than their email. The email is still used when no name is set. ## v3.14.62, 2026-09-09 ### Fixed - The finance read now uses the names the platforms really use, from the first run's detail notes: Motorway says onFinance Yes or No (with the HPI check's finance agreements count), and Carwow's answer is read only from an Outstanding finance label, never from its own finance marketing heading, which was being read as no finance. ## v3.14.61, 2026-09-09 ### Fixed - Remove from DealerKit on a cancelled sale whose record was already deleted on DealerKit by hand now counts as done, so the row reads removed and a hand added car can then be deleted from the list. Before this the remove failed on the missing record and the car stayed marked as on DealerKit for good. ## v3.14.60, 2026-09-09 ### Added - DealerKit's stage, live (option one). Dealer OS now reads DealerKit's own feed the moment DealerKit reports a change and sends the Mac one purchase.dk_stage action per car whose stage differs. The Mac applies it through the same door the timed check uses (In Stock checks the car in, Awaiting Delivery is Ordered, Sold is Sold), follows DealerKit's price on Ordered and Sold, writes the timeline line as DealerKit and pushes at once. The 15 minute check stays as the safety net. ## v3.14.59, 2026-09-09 ### Added - The On finance? answer comes from the auction house. The daily run reads whether finance is outstanding off each car's own listing page (Motorway's page data, Carwow's page text) and keeps it with the car's facts; when the car is bought, the purchases sync sets the answer in the dealership's own words, marked as read from the platform. A person's own answer is never overwritten, and nothing is set when the page does not say. The once a run detail notes record the wording each platform uses. ## v3.14.58, 2026-09-09 ### Changed - The Didn't win sync reads only the new cars at the top of Motorway's list and stops at the first car already on file, instead of reading two full pages every day. Its result says "3 new cars" or "No new cars", and the bar shows the page it is on. A backfill still reads the whole history. ## v3.14.57, 2026-09-09 ### Changed - A Motorway purchase whose payment page is not out yet still gets its Purchase price and Margin: the bid is derived from the won list's all in figure and the fee bands (the list figure less the band's fee and its VAT, proven exact on eight confirmed cars), flagged provisional and replaced the moment the page confirms it. Before this the five cars bought on 2026-09-08 showed no margin for a day. ## v3.14.56, 2026-09-09 ### Changed - "Car price" is "Purchase price" in the DealerKit outcome shown in Dealer OS. The Mac's own page keeps its wording for now, under the standing rule that its screen is not changed until Steven says so twice. ## v3.14.55, 2026-09-09 ### Changed - A bought car takes its mileage and its Retail est. from what BidBrain already knows the moment it is added, on every auction house: the last run's list (which valued the car days before it was bought, with its mileage), the valuation cache, and the purchases table. Cazana is only asked about a car nobody has valued. Five Motorway purchases on 2026-09-08 sat with no retail for a day because Motorway's payment page, the only source of the mileage until now, was not ready yet. ## v3.14.54, 2026-09-09 ### Fixed - A Motorway card wraps its picture, its title and its price in separate links to the same car, so the card read counted each car two or three times, each copy with part of the facts and most copies with no plate, which then held them back. The copies fold into one car with every fact. Tonight's list carried 83 "cars" for 40 odd real ones. ## v3.14.53, 2026-09-08 ### Fixed - Motorway cars read from the cards were held back for nothing but a missing plate: the card text no longer carries the registration and the picture's caption now starts with the year, which the reader took as the whole name, so 90 of 91 held cars had no plate, make or model and could not be valued. The plate, make and model now come from the car's own page where the card gave none, and a caption that starts with the year reads properly. ## v3.14.52, 2026-09-08 ### Added - The Didn't win (16:05) and Under offer (16:30) timed jobs can be set up on the Mac from Dealer OS, as a run of their own (Timed jobs setup), so nobody needs the Mac's screen for it. A job already on the Mac is left at its own time. The outcome shows in the run record. ## v3.14.51, 2026-09-08 ### Changed - The purchases sync's "worth reading" rule now covers everything a car's page can give: the payment breakdown, the photos, the service history, the V5 logbook and the collection date. A car bought in the last 30 days is read again while any of those is still missing; a car bought in the last 7 days is read regardless; older than 30 days only the payment breakdown and photos are chased. (Steven: "should this not also be looking for v5 and service history or any other thing that we try and capture that isnt there yet?") ## v3.14.50, 2026-09-08 ### Fixed - The one click update refused with "the working tree has local changes" because v3.14.48's Motorway CSV marker sat under data/, where git saw an untracked file. The marker now lives under data/logs, which git ignores, the old one is removed on start, and the update only refuses over real changes to tracked files, never an untracked one. ## v3.14.49, 2026-09-08 ### Changed - The purchases sync reads a car's own page (payment breakdown, photos, documents, collection date) only while something is still to come: a car not on file yet, one with no payment breakdown or photos yet, or one bought in the last seven days. A complete car is left alone. It used to read every purchase's page every afternoon, 30 odd pages for one or two new cars. - What a timed pass writes reaches Dealer OS the moment the pass ends. The 15:50 purchases sync is started by the Mac's scheduler, not by the BidBrain server, so the server never saw it finish and ten new purchases sat on the Mac until the next stage check happened to send them, minutes after the sync had said it was finished. The server now watches the run record and sends within seconds of any pass ending. - The progress line names the pass it belongs to, and the heartbeat carries the Mac's timed jobs, so Dealer OS can say what is running and what comes next. ## v3.14.48, 2026-09-08 ### Changed - A run asked for starts straight away. The automatic DealerKit stage check, a quiet background pass, used to make the Mac refuse a Run pressed while it was busy, and the request was lost; now the check is stopped and the run starts, and the check simply runs again a few minutes later. A real run in progress or an open login window still has to finish first. - Motorway's CSV export is not tried again on a day it has already failed: the run goes straight to the cards instead of waiting three minutes for a download that is not coming. It is tried once more the next day. - A Motorway car read within the last few hours is not read again on the next refresh. A cached read with no reserve (from before the page data fill existed) still counts as a miss. ## v3.14.47, 2026-09-08 ### Fixed - A Motorway car's body type is taken from the page's specification block ("Hatchback", "SUV"), not the gallery's lower case copy, so the body filter in Dealer OS does not show the same body twice. ## v3.14.46, 2026-09-08 ### Fixed - Motorway cards carry the fuel, gearbox and the seller's town again. The CSV export used to give all three; the card read that replaced it on 2026-09-08 kept the fuel to itself and never set the car's own fuel, gearbox or location, so every Motorway card in Dealer OS showed the engine size alone and no town. The card's own facts line now sets the fuel and gearbox, and the car's page data fills the gearbox, body and the seller's town where the card gave none. The once a run detail note records the names Motorway uses for them. ## v3.14.45, 2026-09-08 ### Changed - An earlier day's run lends no cars to today. A single platform refresh used to carry the other auction houses' cars forward from the last run whatever day it was, and a failed read fell back to that run's cars, so Saturday's Carwow cars were still on Monday's list. Now only a run for the same day (or later) is carried or fallen back on; an older one is left behind, and the run says so in its log. ## v3.14.44, 2026-09-08 ### Fixed - A single platform refresh (Run with only Motorway ticked) runs under today's date. It used to carry the last run's date whatever day that was, so a refresh on Monday after Sunday's full run had failed ran under Saturday's date and Saturday's five starred cars came back on Monday's list. A star belongs to its own sale day and never carries over; the cached date now carries only while it is today or later. ## v3.14.43, 2026-09-08 ### Changed - The once a run note of Motorway's page data names which field filled each list fact, shows the car's facts after the fill, and carries the whole bidding block with its values, so the reserve's own name is on record rather than inferred from the numbers. ## v3.14.42, 2026-09-08 ### Changed - The Motorway card read scrolls to the bottom of the stock page until no more cards appear before reading them, in case the page only shows a first batch until it is scrolled (87 cards read where the CSV used to carry 190 or more). ## v3.14.41, 2026-09-08 ### Fixed - The Motorway card read reads a car's own page whenever the card left it without a reserve, grade, year or mileage, never the cache. The v3.14.40 run reused all 87 detail reads from the cache saved before the fill existed, so every car was still short and the gate refused them all. - The card parser reads Motorway's new one line facts ("54,080 mi • Diesel • Manual • Grade 3"): bullets split into lines, and "Grade 3" inline is read as the grade. ## v3.14.40, 2026-09-08 ### Fixed - The Motorway card read fills the list facts a card does not carry (the reserve above all, also mileage, year, grade, distance) from the car's own page data, under the names Motorway uses. The first card read (v3.14.39) got 87 cars and 0 past the gate: the reserve had only ever come from the CSV. Once a run, the names on one car's page data are written to data/logs/motorway_detail.log so a renamed field is a one line fix. - When nothing passes the gate on a platform, the run prints the commonest reasons, so a missing field is seen at once. ## v3.14.39, 2026-09-08 ### Added - When Motorway's CSV export fails, the daily run reads the cars from the cards on the stock page and each car's own page instead of stopping. The network notes from v3.14.38 showed the cause at last: Motorway's own server answers 500 to the request for the file, on every run since 2026-09-07, with the panel, the tick and the press all proven by pictures. The card read is a full read of real pages with the same detail cache the other platforms use, never the last good list dressed up as fresh. The Connections notice says which way the read went. ## v3.14.38, 2026-09-08 ### Fixed - The Motorway CSV export takes the file from the web answers the page is given when the browser is handed no download. v3.14.37's pictures proved the panel opens and Filtered vehicles is ticked, and still no file reached the browser in three minutes, so Motorway now delivers it another way. Every answer during the wait that looks like a CSV or its link is noted; the last one is fetched through the same signed in browser and saved. A picture a moment after the press and the list of answers seen ride in the notes when it still fails. ## v3.14.37, 2026-09-08 ### Fixed - The Motorway CSV export ticks the Filtered vehicles radio properly. The picture from v3.14.36 showed the panel as it is now: All vehicles ticked by default, Filtered vehicles beneath, and a Download button. Clicking the words no longer ticks the radio, so every run since 2026-09-07 had been asking for every car in the country, which takes far longer than the 90 seconds it waited. The radio is now ticked by its own input, checked back, a second picture shows it, and the wait for the file is three minutes. ## v3.14.36, 2026-09-08 ### Changed - The Motorway CSV export writes a small picture of the screen around the Download button the moment its menu should be open, as text in data/logs/motorway_menu.log, so Dealer OS's diagnostics carry it and it can be looked at from anywhere. The written notes from the failed runs did not say why no file came. ## v3.14.35, 2026-09-08 ### Fixed - The Motorway CSV export listens for the file from the moment "Filtered vehicles" is chosen. Motorway now hands the file over on that choice, and the Mac only started listening at the old second Download press, so the file was offered with nobody there to take it and thrown away (the screen notes from v3.14.34's run: menu opened, choice made, one Download button left, no file). The second press is kept for the day Motorway changes it back. ## v3.14.34, 2026-09-08 ### Fixed - The Motorway CSV export closes any pop up on the stock page before pressing Download, and makes sure Download's own menu really opened before choosing Filtered vehicles from it. The screen notes from the failed runs (v3.14.32) showed one open pop up, a single Download button and no menu: the press had landed on the pop up, so the menu never opened and the second press did nothing for 90 seconds. When the menu still will not open, the notes now carry the pop up's full wording and its buttons, and say which pop ups were closed. ## v3.14.33, 2026-09-08 ### Fixed - Diagnostics: a log line with a hidden NUL character (the Motorway screen read from v3.14.32 carried one) made Dealer OS's database refuse the whole document, so nothing reached Dealer OS and the Mac tried again every few seconds. Every log line sent is now plain text: control characters other than a tab are dropped. ## v3.14.32, 2026-09-08 ### Changed - The Motorway CSV export, which the daily run reads the whole sale from, now watches for the file on this page and on any tab Motorway opens for it, and when nothing comes in 90 seconds it writes what the screen showed (address, tabs, every visible button, menu item and dialog line, and a picture) to data/logs/motorway_export.log, which Dealer OS's diagnostics carry. Two runs on 2026-09-07 failed with "Timeout 90000ms exceeded while waiting for event download" and nothing said why (Steven: "i cant get a run to work"). The run still fails loud and keeps the previous list; it just says what it saw. ## v3.14.31, 2026-09-07 ### Changed - A Non-Stock car on DealerKit is passed over by the price sync in both directions, without a word (Steven: an old Clickdealer era sale back in for warranty work, added as non stock because it is not on the system, "these instances can be ignored"). G13RTP was failing every five minutes with "the Pricing tile was not found" for exactly this. A Retail price ticked on such a car, or a Retail est. changed on it, is refused with a sentence saying it is not a car for sale. The same goes for any stage the Mac does not know. ## v3.14.30, 2026-09-07 ### Changed - The hard rule (Steven: "when any car is in ordered or sold the only place a price should ever be touched is dealerkit, it shouldnt even send the price back"). A car Ordered (Awaiting Delivery) or Sold on DealerKit has its price sent from Dealer OS by no route at all: not the five minute check, not the instant push on a changed Retail est., not a ticked Retail price on the row, and not even a figure a person typed in Dealer OS. Each refusal says so in words on the car's timeline. DealerKit's own figure still comes into Dealer OS so the row reads true. DealerKit itself greys the Retail Price box out on such a car (DT65UKP), which is what the earlier "field was not found" failures were on Ordered cars. ## v3.14.29, 2026-09-07 ### Changed - The retail price syncs both ways (Steven: "i just want the prices to sync on both sides as instant as possible"). The stage check, now every five minutes instead of fifteen, reads each car's record once and takes both its stage and its Retail Price from it. The side that changed since the two last agreed wins: a price changed on DealerKit comes into Dealer OS as the Retail est., a price changed in Dealer OS goes to DealerKit. When both changed, DealerKit wins. On a car already arrived or sold, a figure the formula re priced on its own never overwrites the team's price, only a figure a person typed in does. A change made in Dealer OS still goes across the moment it is made (v3.14.28); the five minute read is the safety net and the way back. - The Retail Price is set by asking DealerKit's own API the way its Pricing panel does, not by driving the panel. The panel route failed on cars whose panel did not show the price box (DN18JNU, EK68WGV) and on one where the search click landed on the Forecourt list (YB67ULM). The panel is still there as the fallback when DealerKit refuses the call, and the car's timeline says which way it went. ### Removed - The daily push's own retail follow up, and the rule behind it: the stage read carries the price now, in the daily push and in the five minute check alike. ## v3.14.28, 2026-09-07 ### Added - The delivery charge goes to DealerKit with the car, not behind it (Steven: "if the job is done properly these should go from dealer os into dealerkit at the start"). A car sent whole now lands on DealerKit carrying its purchase price, buyer's fee AND delivery charge, so every cost of buying it is on the record from the start. A delivery invoice agreed later still goes on its own as before. ### Changed - A new Retail est. goes to DealerKit the moment it is set, from the Mac's own purchases page or from Dealer OS (Steven: "im more interested in getting the info over into dealerkit from dealer os more constantly and quicker/instantly"). It used to wait for the automatic run, so a price changed at nine in the morning reached DealerKit at quarter past four. The automatic run stays as the safety net for a Mac that was busy. ### Fixed - A whole send is no longer mistaken for a follow up. The delivery charge is now in both lists, and the old test was "any follow up wins", which would have sent a whole car down the follow up path and quietly dropped its price, photos and documents. What decides now is whether anything a follow up cannot do was asked for. ## v3.14.27, 2026-09-07 ### Changed - The DealerKit API probe, part I: what the routes the Mac does not use yet actually return, so a real feature can be built against them rather than guessed at. Reads one record each from costs, deals, invoices, the stock feed, a car's valuations, its pricing insights and its prep stage summary, and writes the field names with their kinds. A field name is never personal so names are always written; a value is written only where it cannot be someone's personal details (numbers, true, false, nothing, and short text on a field whose name is not one of email, phone, address, name, date of birth, bank and the like). ## v3.14.26, 2026-09-07 ### Changed - The DealerKit API probe, part H: the front end signs every call with an X-API-TOKEN header from a token on the logged in page, so the probe reads that token (never printed, a 10 character fingerprint only) and tries the main API with it from a clean context with no cookies, as a token header and as a bearer. A 200 there means the main API takes a token on its own, which is what a machine would use. ## v3.14.25, 2026-09-07 ### Changed - The DealerKit API probe reads more (Steven: "dig deep ... find out what is actually possible"). The first pass found one /api/ path in the front end's bundle because the bundle keeps its base URL in one place, so part F now lists every route string the bundle builds its calls from and every sign in word it uses (token, bearer, sanctum, oauth, webhook) with a little context, and part G reads the sign in page as a stranger sees it (form, inputs, any token call). Still reads only, nothing printed that is secret. ## v3.14.24, 2026-09-07 ### Added - The DealerKit API probe (Steven: "find out what is actually possible"). A quiet run that writes data/logs/dealerkit_probe.log: what the Integrator key can reach from the Mac, whether the tenant's own main API takes that key without a login, what a stranger sees, which candidate routes answer with the saved login, and every /api/ path the front end's own scripts mention. Reads only, no secrets printed. ## v3.14.23, 2026-09-07 ### Fixed - One car whose DealerKit record no longer answers (SD66XWV, removed on DealerKit during the afternoon) stopped every DealerKit run as if the login had expired. A record that answers HTTP 404, or serves a page while the session is otherwise live, is now that one car's problem: it reads as not on DealerKit with a timeline line, and the run carries on. ## v3.14.22, 2026-09-07 ### Fixed - The DealerKit login probe now makes the two reads every pass makes, the stocklist and then one car's own record, before calling a session live. On Steven's Mac the automatic login left a session that answered the list and served the sign in page for every car, so v3.14.21's probe passed and each run died on its first car. The expiry message now names the read that failed and its HTTP status. ## v3.14.21, 2026-09-07 ### Fixed - The DealerKit login probe (v3.14.19) accepted a tidy "not signed in" JSON answer as proof of a live session, so the DealerKit run still failed on Steven's Mac. A session is live only when the stocklist answers with its real results list; the automatic login saves nothing until it does. ## v3.14.20, 2026-09-07 ### Changed - Update now from Dealer OS always takes a fresh look at GitHub, and the Mac remembers its "is there a newer release?" answer for 5 minutes instead of 30 (Steven: "why do we have to wait 15 mins now for the mac?"). Dealer OS v3.14.1 lets the button be pressed at any time, so a release published a moment ago is found and installed at once. ## v3.14.19, 2026-09-07 ### Fixed - A DealerKit session that rendered the stock list page while its API served the sign in page passed as logged in, so the quiet stage check and the daily DealerKit run both failed with "the saved login has expired" instead of logging in again (seen after the 15:30 login job on Steven's Mac). Every DealerKit opener now also asks the API for one record before believing the page, and the automatic DealerKit login only saves its session once the API answers with real data. ## v3.14.18, 2026-09-07 ### Added - Re price retail estimates (Steven: "why is dealer os still adding £500 on to retail estimate pricing and rounding up always with a 5 on the end"). A car's Retail est. was only ever worked out on the day it was first read, so every car bought before a dealership changed its margin or price ending kept the old formula's number. A new Maintenance row in the Run modal re prices every live car from its Cazana figure with the dealership's own rule, writing the change on each car's timeline. A figure typed in by hand, a cancelled sale and a car with no Cazana figure are left alone. ## v3.14.17, 2026-09-07 ### Added - Ordered and Sold (Steven: "the dealer os purchases list needs to be in sync with dealerkit as quickly as possible"). Two more Status answers, both finished: Ordered is DealerKit's Awaiting Delivery (a customer has bought the car and is waiting for it), Sold is gone. A dealership whose saved list predates them gets each word the first time DealerKit reports that stage. - DealerKit wins on every stage, not only arrival. The daily DealerKit run reads every live car on DealerKit and moves its Status here to match: In Stock checks it in with DealerKit's arrival date, Awaiting Delivery is Ordered, Sold is Sold, and a customer deal falling through moves it back to In stock. Each move is a timeline line "matched from DealerKit". - The quiet DealerKit stage check, every 15 minutes while the Mac is idle. No window, no bar, no light: the saved login, one read per live car, the same matching. Off whenever the DealerKit purchases connection is off. A login that cannot be fixed quietly is left for the daily run. - A change goes to DealerKit at once. A chip, a transport fee, a check in or a cancelled sale's removal made here or in Dealer OS is sent to DealerKit the moment it is applied; a busy Mac writes on the timeline that the next automatic run will send it. - A check in from Dealer OS may name the finished answer (Ordered, Sold) the person picked, else In stock. ### Changed - Whatever a run wrote goes to Dealer OS the moment the run ends. ## v3.14.16, 2026-09-07 ### Added - `setup_schedule.py` (Steven: "set didn't win sync and under offer check to run daily like purchases sync, i think compmatch should be the same too", and "mark will want it to work the same way too"). One command on the Mac puts the Didn't win sync (16:05), the Under offer check (16:30) and the CompMatch sync (16:45) on the Mac's own scheduler (launchd), written the same way as every job already there. A job the Mac already has keeps its own time unless a new one is given, so a Mac with its own earlier setup is not moved. Times can be chosen per job, and a job can be taken out again. Opt in: nothing runs it on its own. `python3 setup_schedule.py` on its own lists what the Mac has. ### Fixed - The Run times card in Dealer OS could show a python path in front of a job's command; any python interpreter is dropped now, whatever version. ## v3.14.15, 2026-09-07 ### Added - DealerKit wins on arrival (Steven). On every automatic DealerKit run, a car DealerKit holds In Stock while BidBrain still has it Due in is checked in here with the arrival date DealerKit holds, its DealerKit check in stamped, and a timeline line "Checked in, matched from DealerKit". A sold car is left alone. The run reads the stage of every not yet arrived car that is on DealerKit, in the same session. - The DealerKit check (dk_check) now carries the car's stage in words (`stage`: Due In, In Stock, Sold), the raw `life_cycle_status` and `delivery_on`, for Dealer OS to show on the row. ### Changed - Check in takes DealerKit's date. When Check in (the button or the Status dropdown) finds DealerKit already has the car In Stock, the arrival time on BidBrain is rewritten to DealerKit's delivery date, with a timeline line saying so, instead of the moment the button was pressed. An arrival time set this way is a day, not a moment. ## v3.14.14, 2026-09-07 ### Fixed - Status follows the check in, the older cars: a car from before the rule can read In stock with no arrival time on record. v3.14.13 treated such a car as not checked in, so picking Due in on it was a plain write with no question asked (found live on LM67BPE). It now counts as arrived: Due in on it is the undo, with the same DealerKit rule, and In stock on it changes nothing. ## v3.14.13, 2026-09-06 ### Changed - Status follows the check in (Steven). Picking a finished Status answer (In stock) on a car not yet checked in now IS the check in: the arrival time is stamped and, if the car is on DealerKit, it is checked in there too, the same as the Check in button. Picking any other answer on a checked in car takes the check in back and clears the arrival time, allowed only while DealerKit has not been checked in; once DealerKit shows the car In Stock the change is refused with a note to change it on DealerKit. One door for the Mac's own page and for Dealer OS (`serve.apply_purchase_field`, `db.set_stock_status`, `db.undo_check_in`). Before this a Status change wrote the word alone, so a car could read In stock with no arrival time and never reach DealerKit. ## v3.14.12, 2026-09-06 ### Changed - The Live Auctions screen now understands the filters Dealer OS saves since its v3.8.0: several makes, models, service histories, condition grades or body styles at once (a comma list in the same key, meaning any of those), and year, engine size and owners as ranges (`yearmin`, `yearmax`, `enginelitresmin`, `enginelitresmax`, `ownersmin`, `ownersmax`). A view with one of those used to show nothing here, or filter less than it does in Dealer OS. The screen's own boxes are not changed: a value they cannot show is carried through the view, the box reads Any, and the list, the count and the view chip honour it. Touching that box by hand drops the carried value, so what you see is what filters. A blank engine size or owner count passes a range, as it already passed an exact match. Contract still schema 1, additive only. ## v3.14.11, 2026-09-06 ### Fixed - Deleting a hand added car now takes its own cancellation record with it. A hand added car that was cancelled and then deleted left its plate in the walked away list for good, kept off every future shortlist and shown as a grey row in the Cancelled view with nothing behind it (found on the two test cars, TE57DOS and LO68GXE). Only a hand added car's own cancellation writes that record, so only that record goes; a platform's own cancelled list is never touched, and a plate another purchase still carries keeps its record. The timeline row says when it happened. ## v3.14.10, 2026-09-06 ### Fixed - A car sent to DealerKit from Dealer OS that DealerKit would not add now says why. Found live on the test car TE57DOS: two sends died in DealerKit's own Add Vehicle window with "got no confirmation, the page may have changed", nothing was written to the car's timeline, and Dealer OS's DealerKit tab simply went back to Ready as if nothing had been pressed. Now the send waits up to 15 seconds for DealerKit's answer instead of 3, looks the plate up before calling it a failure (a slow create is a record all the same), and when DealerKit really did not add the car the failure carries DealerKit's own words, from the window and from its server. Every failure of a send, not only an expired login, is written to the car's timeline by BidBrain ("DealerKit send failed: ...") so Dealer OS shows it. ## v3.14.9, 2026-09-06 ### Added - Photos on a car added by hand, from Dealer OS. A local trader buy has no listing to read photos from, and DealerKit will not take a car without them, so until now such a car could never be sent to DealerKit from either screen. `purchase.set` now accepts `photo_urls` on a hand added car: Dealer OS stores the pictures in its own media store and sends the lasting links as the whole list (at most 40, an empty list clears it). The Mac keeps the list as the car's gallery, makes the first the thumbnail, and writes one timeline row saying how many. The DealerKit send already works from links, so the car can go across. A platform car is refused, its gallery is the listing's (Steven's decision, 2026-09-06). ## v3.14.8, 2026-09-05 ### Changed - The heartbeat to Dealer OS goes every 5 seconds whatever the Mac is doing, instead of every 15 while idle. A run asked for in Dealer OS is picked up within 5 seconds rather than 15, so its progress bar appears sooner (Steven: "as close to instant as physically possible"). ## v3.14.7, 2026-09-05 ### Fixed - The Software update card read "API rate limit already exceeded" the first time it was pressed for real. The latest release was looked up through gh release list, which goes through GitHub's GraphQL allowance that the whole account shares. It now goes through the REST route (gh api releases/latest), which has its own, far larger allowance. A failed lookup is remembered for two minutes instead of thirty, and the last release that was seen is kept rather than forgotten. ## v3.14.6, 2026-09-05 ### Added - Diagnostics for Dealer OS. The Mac now sends Dealer OS a diagnostics document (`POST /api/bidbrain/diagnostics`): the last 300 lines of every log under data/logs, the full run records and site health (the heartbeat only carries each summary's first line), the live progress file and the server process. It goes when Dealer OS asks (a new `diagnostics.pull` action, answered in `last_command` with key "diagnostics"), on the tick after a run finishes, and otherwise at most every 10 minutes and only when a log has grown. Steven's decision: the Mac Studio stays behind the office router, never opened up, so anything Dealer OS needs to see, the Mac sends. ## v3.14.5, 2026-09-05 ### Added - Update from Dealer OS. The heartbeat now carries a `version` block (the release this Mac is on, the newest GitHub Release, whether they differ, kept 30 minutes between lookups), and a new `update.start` action runs the same one click update the Mac's own Settings page has: fetch the tags, check out the release, restart the server. Refusals (a run in flight, no newer release, local changes) come back in the heartbeat's `last_command` with the reason. Steven's decision: the Mac's own screen is never needed for anything, so moving to a new release is a button on Dealer OS's BidBrain Settings page. - A Release workflow: pushing a `v*` tag cuts the matching GitHub Release with that version's CHANGELOG section as the notes, so a release no longer needs the gh tool on the Mac. ## v3.14.4, 2026-09-05 ### Changed - Retail follows an edit. Dealer OS now sends the retail price to DealerKit the moment a retail estimate is edited on a car that is on DealerKit (Steven's decision). That job always overwrites DealerKit's own figure (already so since v3.13.3, confirmed and now written into the contract); only a send whole leaves an existing DealerKit retail price alone. The automatic 16:15 run now also sends the retail price on any car on DealerKit whose last DealerKit check shows a different figure, so an edit that landed while the Mac was busy still goes across. Working list cars only: a finished or checked in car's DealerKit retail price is the team's own and the run never rewrites it (68 old records would have been overwritten with unrounded estimates on the first run otherwise). ## v3.14.3, 2026-09-05 ### Changed - The automatic 16:15 DealerKit run no longer sends a car to DealerKit whole by itself. A car goes to DealerKit only when a person presses Send in Dealer OS (Steven's decision). The run still picks up chips, transport fees, check ins and removals on cars DealerKit already has, exactly as before. The contract now also says which field the Send preflight reads for "no collection date yet": `collection_date`, the due in date, never `collection_arranged`. ## v3.14.2, 2026-09-05 ### Changed - The mileage now goes in beside the registration on DealerKit's own "Add Vehicle To Stock" form when a car is created (Steven: "the mileage should be entered on the add vehicle to stock modal alongside the reg at this very point"), so the record starts with the real figure and DealerKit never writes a guessed one. The mileage log entry is kept only for a car DealerKit already holds with a different figure. - The due in date is no longer sent to DealerKit, by a send or by the automatic run (Steven: "id rather have the checkin feature work when we decide to check it in rather than dealerkit automate it itself"). A date on the record made DealerKit hide its own CHECK-IN button and move the car to In Stock by itself when the date arrived. The Dealer OS `due_in` button is refused with a plain sentence. A check in now goes through DealerKit's own CHECK-IN form with the arrival day; a record that still carries a date from before is moved by writing the arrival day onto it. ### Fixed - The check in kept failing with "the CHECK-IN button was not found" on any DealerKit record carrying a photo, which looked like a broken page and for a while like a broken record (FX68KAE included). It was neither. DealerKit draws the lower half of a vehicle's overview, the Due In banner with its CHECK-IN button, Vehicle Details and the rest, only once that half scrolls into view, and a record with a photo pushes it below the fold. A person scrolls without thinking; the automation never did. It now scrolls the page like a person before looking for the button, and clicks the button through the browser so a click can never land off screen. No record needed repairing. ## v3.14.1, 2026-09-05 ### Fixed - Three settings sync checks in the test suite ran against the Mac's REAL settings table instead of a temporary one: green on this Mac, red on a fresh CI database (CI had been red since v3.13.3, found when v3.14 was pushed), and one of them WROTE `details_style` into the live table with a year 2099 stamp, which would have made every later Dealer OS change to that setting count as older for ever. The checks now use their own temporary database, and the real row was re stamped as saved today (the value, popup, is Steven's own). The suite is hashed against the real database before and after a run; it leaves it untouched. ## v3.14, 2026-09-05 DealerKit from Dealer OS, the redesigned tab: a car is sent whole once its collection date is confirmed, then kept in step. Everything is additive (the push's `schema_version` stays 1), an older Dealer OS keeps working. ### Added - `purchase.dk_send {id}`: one job that creates the DealerKit record if there is none and sends the registration, mileage, purchase price, buyer's fee, photos, service history and V5, retail price and due in date in one go. Refused with one plain sentence per missing thing (no mileage, no purchase price, no fee, no photos, no retail estimate, no collection date, a cancelled sale, the Mac busy), all together in one message so Dealer OS can show them. - The mileage now reaches DealerKit. It never did before: DealerKit wrote a guessed figure of its own onto every new record. A send writes BidBrain's real mileage as a mileage log entry, so DealerKit's `vehicle.mileage` is the Mac's figure. The buyer's fee goes across too, as the Buyers Premium line on the supplier's expense. - Four follow ups for a car already on DealerKit, each as a `purchase.dk` action and each picked up by the automatic run: `chip` (the new price and fee after a chip), `transport` (the transport fee as the Delivery Charge line), `checkin` (Due In to In Stock with the real arrival date; a `purchase.checkin` now does this by itself when the car is on DealerKit) and `remove` (a cancelled sale's record deleted while DealerKit still shows it Due In; left in place with a sentence once In Stock). None of them ever creates a record. - `purchase.delete {id}`: a hand added car leaves the Purchases list for good, its timeline kept with one last row of kind `deleted`. Refused while the car is on DealerKit (cancel those instead) and for any platform buy. - Six new purchase stamps in the push: `dk_mileage_pushed_at`, `dk_fee_pushed_at`, `dk_chip_pushed_at`, `dk_transport_pushed_at`, `dk_checked_in_at`, `dk_removed_at`. - The automatic DealerKit run now has a real daily time, `com.bidbrain.dealerkit-purchases` at 16:15 (after the 15:30 DealerKit login and the 15:50 purchases sync), so it appears in `run_schedule`. It sends whole any car whose facts are all in place and picks up chips, transport fees, check ins and removals not yet sent, the same rules as the buttons. ### Changed - Recording a chip, or changing the transport fee, clears the matching DealerKit stamp so the next automatic run sends it. - Deleting a hand added car keeps its timeline (it used to wipe it). - A repeat send no longer doubles the photos: a record that already holds as many photos as BidBrain has is left alone. ### Fixed - The DealerKit expense category group is named differently per account ("Stock Purchase" on Right Drive's, "Cost of Purchase" on Really Easy Car Credit's), found live; both are tried. - A removal is keyed by registration, so a cancelled sale sharing its registration with a live one (the Ford Puma, sold on both platforms the same day) is never removed; refused with a sentence, in the door and in the automatic run. ## v3.13.6, 2026-09-05 ### Added - A DealerKit contact can now be deleted (Steven: "to delete a contact inside dealerway you go to contacts, click the contacts name then click delete"), a real, plain Delete button on the contact's own page. Used to clean up the 13 test contacts this session's own testing had left behind in the real account. ## v3.13.5, 2026-09-05 ### Added - The DealerKit Price button now creates the supplier as a real DealerKit contact automatically when it does not already exist, then carries on and types the price as normal. Before this, a genuinely new supplier had to be created in DealerKit by hand first, or the price silently would not save. Proven live, 6 fresh attempts in a row, each with a supplier name never seen before. ### Fixed - The new contact step occasionally needed a second try (found live, about 1 in 6 real attempts, DealerKit's own known variable timing), so it now retries once before giving up, the same discipline already used throughout this file. ## v3.13.4, 2026-09-05 ### Fixed - A purchase edit made on the Mac's own page (a field, a chip, a cancel, a comment, an override, add or delete by hand, the page settings) now reaches Dealer OS the same way a star, hide or view already did, instead of sitting unsent until the next daily run. Found live: a registration correction sat unsent for 12 minutes. - The DealerKit Price button, for a genuinely new car, now selects the purchase's own supplier before typing the price, since DealerKit silently refuses to save Purchase Price without one, found live by hand testing. Two further bugs on the way there are also fixed: the field finder had started typing into a new "Purchase Invoice Ref #" field DealerKit added since this was last built, and, once fixed, was matching a huge invisible ancestor of the whole page instead of the real, short label. ## v3.13.3, 2026-09-04 ### Changed - Ticking Retail price on its own now overwrites the figure on DealerKit, so changing a retail estimate in Dealer OS and sending it makes the two match. The all in one push still leaves an existing DealerKit retail price alone. ### Fixed - A DealerKit push now says what each item actually did, in the log and on the car's timeline: set, changed from one figure to another, already right, or nothing to send. Before this, a price that merely already matched read as "Sent to DealerKit: price", and a skipped retail price gave a reason that was not true. ## v3.13.2, 2026-09-04 ### Fixed - The shortlist sync to Motorway and Carwow judged its own click too soon: both sites keep showing the old star for a few seconds, so a click that genuinely worked was reported as a failure, left the car pending and turned the run light red. Both now wait for the state to change and confirm with a reload, retrying the read. ### Added - DEALER_OS_HANDOVER.md: what Dealer OS can send, which version each was proven on, and what a future session should know. ## v3.13.1, 2026-09-04 ### Fixed - A setting applied from Dealer OS is now stamped with the action's own time, not the Mac's clock, so two saves made seconds apart never make the second look older than the first and get skipped. Found live during the step three proof: a platform gate save right after a spread save was skipped as "the Mac saved flat_spread more recently". ## v3.13, 2026-09-03 ### Added - The way back for settings, step three: Dealer OS can send back the purchases dropdown answers, the column layout, the journey flow, the buying rules a dealer changes day to day, and the integration switches (settings.set {section, value}), each applied through the same checks the Mac's own Settings page makes, newest change wins per key. - The push carries the Mac's whole settings page (`mac_settings`): the buying rules, integrations, hidden connections, the fee band tables, the engine bans and the run schedule (the last three read only), so Dealer OS can draw a settings page from the push alone. ## v3.12.1, 2026-09-03 ### Added - Step two of the purchases way back proven end to end on the live site: a car added by hand, a chip, a check in, a cancellation, a DealerKit check and a saved view (and its delete) made on Dealer OS all reached the Mac, applied through its own doors with the person's email, and came back. - Every document scan and purchase photo now has a lasting link in Dealer OS (917 files sent, 3 shrunk to fit, none failed). - A DealerKit job started from Dealer OS logs the Mac back into DealerKit once with the stored password if the saved login has expired, and writes its own outcome, success or failure, on the car's timeline. ## v3.12, 2026-09-03 ### Added - The way back for purchases, step two: a chip, a cancellation, a check in, a car added by hand, the DealerKit buttons and saved views made on Dealer OS are pulled and applied on the Mac through the same doors its own Purchases page uses, with the same checks and refusals, each logged with its reason and the person's email on the timeline. - Saved views from Dealer OS use the Mac page's own filter keys, so a view saved there applies on the Mac's own screen too. ## v3.11.3, 2026-09-03 ### Added - The way back for purchases, step one, proven live: a Video change made on Dealer OS reached the Mac in seconds, landed on the right car with the person's email on its timeline, and was pushed back automatically. - The document upload now covers a purchase's photo files too, downsizes any image over 4 MB to a 2000 px JPEG first, and stops after three failures in a row so a broken store can never hold up the push. - The spec for the way back, step two (chips, cancellations, check in, cars added by hand, the DealerKit buttons, saved views) is in DEALER_OS_PUSH.md, not built yet on either side. ### Fixed - A missing or refused document route reads as "not open yet", never as a bad token. ## v3.11.2, 2026-09-02 ### Added - The push to Dealer OS also carries every purchase's timeline (`purchase_events`) and the Didn't win page's saved views. - Document scans (service history, V5) get a lasting link: each file is uploaded once to a Dealer OS route the contract now describes, and the push carries the link instead of a path on the Mac. Until Dealer OS has the route, nothing is sent and nothing fails. - The way back for purchases, step one: notes, the dropdowns, paid for, collection arranged, the retail estimate and due in date overrides, and comments made in Dealer OS are pulled and applied on the Mac through the same checks its own page uses. ### Fixed - The test file's summary line sat mid file, so every check added after it never ran. Moved to the end; the true count is 1,135. ## v3.11.1, 2026-09-02 ### Added - The push to Dealer OS now also carries the Purchases working list, the Under Offer snapshot and the Didn't win rows (plus the purchase page's own settings), the same rows the Mac's own pages draw, minus local file paths, so Dealer OS can build those three pages. Read only on the Dealer OS side for now; edits still happen on the Mac's own Purchases page. ## v3.11, 2026-09-02 BidBrain inside Dealer OS. The Dealer OS BidBrain page is now the cockpit, rebuilt pixel for pixel, with a two way link to the Mac. The Mac's own cockpit is untouched and stays the safety net until both owners have confirmed the switch. ### Added - The push to Dealer OS now carries everything the cockpit draws: every card field, the traffic light insight rows, the full record drawer (facts, pricing receipt, flags, CompMatch comps, history), photos, Top picks rank, taste match, stars, plus page level settings, notices, model families, filter scope, platforms, integrations and the run checks. The on demand Dealer OS button sends the same annotated, hide filtered document a run does. Over 3.5 MB the photo galleries are trimmed and the push retried once. - The way back (`bidbrain/dealer_os_sync.py`): every star, hide, unhide, saved view, display setting and run request made in Dealer OS is logged there; the Mac's serve.py pulls that log after a heartbeat every 15 s (5 s while busy), applies each through the same functions the cockpit's own buttons use, rebuilds cockpit.html and echoes its state back in the next push. Newest change wins per key. `python3 daily_run.py --dealer-os-pull` runs one pull by hand; every run pulls first. - The heartbeat carries the run light, progress, run board, connection health, the run registry and which request the Mac last answered, so Dealer OS shows the same light, bar and Run modal as the cockpit and a run asked for there starts on the Mac. - Dealer OS's Display choices (how Details opens, the price ceiling) follow the Mac's and the Mac's follow Dealer OS's, newest stamp wins. - BidBrain's warnings in each push become Dealer OS bell notifications. ### Changed - `bids`, `hidden_cars` and `saved_views` carry `updated_at`; saved views carry a stable id and a delete marker so both screens agree on them. ### Fixed - `serve._dealer_os_snapshot` referenced `daily_run` without importing it, so the live sync loop would have failed on every tick. ## v3.10, 2026-09-02 ### Added - **Top picks, the cockpit's new default view.** Every priced car gets a score from signals the run already has (room between reserve and max bid, a gap in stock, your taste, what CompMatch knows, whether the car is new since yesterday, a mark down for no service history), and the best 25 open first, in rank order, with "Show all N" one tap away and an "All cars" row in the views menu. Cards carry a "#N pick" badge and a "New today" badge. A "New since yesterday" sort joined the sort list. Nothing about pricing or status changes, this only decides what sits at the front. - **A car record drawer on the cockpit.** Details on a card now opens the same slide in panel Purchases has: photos, the car's facts, a receipt showing exactly how the max bid was worked out, and three tabs, CompMatch (the suggestion and the real comparable sales it used), History (every earlier sighting and reserve, your own sold record for the model, stock) and Flags. Star, hide and the link to the auction house sit in its footer. **Settings > Look chooses which one Details opens**, the drawer or the small pop-up it replaced, per dealership. - **Top picks read as one list under All**, #1 to #25 whatever the auction house, instead of being split into per platform sections with Motorway first (which buried a Carwow #1 a long scroll down). - **Top picks now lean on your own sold history.** A model you have sold before is the strongest signal (more so the more you have sold), the real average margin after prep is next, and a quick seller earns a little extra. - **New ban: Ford 1.1 petrol, 2014 to 2019**, treated exactly like the Ford 1.0 EcoBoost (same year window). - **Every run now starts, reports and finishes the same way, on every page.** One registry of runs (`bidbrain/runs.py`) replaces five hand written maps in the server; one shared finish means a run can no longer stop without saying so (a switched off run reads as "skipped", grey, never red); one shared display kit gives all five pages the same progress bar, the same toast (green done, red failed with the reason) and the same waiting loop. Didn't win, Under offer and Settings gain the bar they never had; Purchases loses its browser alert boxes; the four Maintenance runs now behave like the auction house runs; a red light click opens the run board on every page. A one car push or check can no longer be mistaken for the bulk run. - **The Run modal is on every page.** The same board (auction houses, last run, Maintenance cards, logins) opens from the run button on Live Auctions, Purchases, Didn't win, Under offer and Settings alike. The Purchases sync, Didn't win sync, Didn't win full history and DealerKit records checks are now rows on it, so the two separate sync pickers are gone. The board is two tabs, Auction houses and Maintenance, each a table of rows with the same anatomy (tick, name, state, result, age); Live Auctions opens on the auction houses, every other page on Maintenance. One Run does the ticked auction houses first, then each ticked check in turn, and the footer says exactly what will happen. Toasts sit just under the top bar, on the eye line, and the board opens the instant you click. - **DealerKit logs itself in.** The three DealerKit checks (stock, purchases push, records check) try the saved session, then the saved password, and only wait for a person at the Mac if both fail, so they can be started from any device. The LE Capital funding sync does the same for both its logins. - **One more car per row on Live Auctions** at every screen size: 4 across on a 13 to 14 inch laptop (was 3), 6 on a 1920 monitor, 11 on a 34 inch. - **All cars is the default view again**; Top picks is one tap away in the views menu (whose count now reflects the real 25) or the link beside it. - **Purchases views remember the column layout**, Airtable style: resizing, reordering or hiding a column saves straight to the view you are on (or to the default layout on All cars) and comes back with it. - **The watchlist sync stops chasing lost causes.** A car that has left the list is dropped from the sync queue, and a car that fails to bookmark three times is left alone (its error stays on file) until it is starred again. - **Fixes found by the audit:** stars no longer queue for a second Motorway account that has never logged in on this Mac (the "0 of N synced" red light); the DealerKit auto login moves to 15:30 so it runs before the 15:50 purchases sync; the purchases sync records health per platform. ### Changed - **The plumbing is hidden.** The Run menu folds its nine maintenance checks into one collapsed Maintenance row that opens on demand, or on its own when a check is failing. Settings went from eight tabs to four: Buying rules (with the auction house picker inside it), Purchases, Connections, Look. Every old link still lands where it did. ## v3.9.1, 2026-09-02 ### Fixed - Right Drive's own `DEALER_NAME` (used both by CompMatch's comp-sharing export and as a sanity-check label on the Dealer OS push) was still "Right Drive" while Dealer OS has "RightDrive Car Finance" registered against its API token, so every push surfaced a visible mismatch warning on screen. Updated to match, which moved this dealership's own shared CompMatch file from `shared_comps/right_drive.json` to `shared_comps/rightdrive_car_finance.json` (a real rename, refreshed with current data, not a stale copy under a new name). ## v3.9, 2026-09-02 ### Changed - **Ford 2.0 diesel is now a banned engine** (Steven 2026-09-02), TDCi and EcoBlue alike, any year, caught from the fuel field as well as the engine text. The only buying rule change in this batch. ### Fixed - **Cazana logs itself back in mid run.** When Cazana fails four lookups in a row the run now runs the same keychain login the 17:00 job uses and retries once, instead of holding every remaining car. Three runs in the log had died this way; on a one day Motorway auction those cars were lost for good. - **Kia Ceed valuations were always refused.** Motorway spells it "c'eed", Cazana "Cee'd", and the match required the exact token. Punctuation is now ignored on both sides. Kias were 28 of 115 held cars in a fortnight. - **The assessments table was saved 3 to 7 times per day** (once per platform as each finished, the whole list each time), inflating every count taken from it. One row per car per day now, and the history was de-duplicated (47,879 rows to 20,309, backup kept). - **Auction4Cars cars it never visited were rejected as "engine unknown"** instead of their real reason; the engine text is now set at read time. - **"Owners missing" and "reserve missing" no longer clutter a rejection that already has a real reason** (about 200 car-days a week, all cars that had failed the cheap first check before their page was ever read; a live check showed the facts were on the page all along). ### Added - **The first step of getting BidBrain inside Dealer OS.** After every run, BidBrain now sends its finished list (the shortlisted and held cars, each with its price, reasons and CompMatch suggestion) into Dealer OS as one JSON document, and the Run modal gained a "Dealer OS" button to re send the last run on demand. Dealer OS only displays the list; every scraper and the whole buying brain stay on the Mac, so there is one set of rules that cannot drift. The receiving route on the Dealer OS side is still to be built, against `DEALER_OS_PUSH.md`. Configured per dealership in `dealer_config.py` (`DEALER_OS_BASE_URL`, `DEALER_OS_API_KEY`, a shared secret that must match `BIDBRAIN_API_KEY` in Dealer OS's own `.env.local`); blank means "not set up", never a failure. Its own on/off switch and a "Dealer OS" row (no login, it has none) on Settings > Integrations, its own line on the sync light. ## v3.8, 2026-09-01 A traffic light in every page's nav bar so a sync is visible wherever you are, a full sweep that caught and fixed around 30 real inconsistencies across the app in the process, and the Retail est. formula becoming genuinely per-dealership instead of one fixed number for everyone. ### Added - **A sync light in the nav bar of every page.** Lives inside the existing Run/Sync button on each page, colouring it green and spinning while something is running, amber if it goes quiet for 90 seconds, and red if the last run of something failed — visible from any page, not only the one that started it. Clicking it opens the Run modal, rebuilt as a real status board: when each platform and check last ran, whether it worked, and a Log in button right on the failed row. - **The Retail est. formula is now per dealership.** Two new settings (Settings > Overview): the flat margin added on top of Cazana's 115% before rounding (was always £500), and the price ending it rounds to — either the four specific bands one dealership's own real pricing uses, or a plain "every £100, ending 99" for the other. Each dealership keeps its own figures; nothing changes unless you touch these settings. ### Fixed - The purchases sync (the Sync now button, and the daily 3:45pm job) now respects platforms switched off or hidden in Settings. It previously always tried to read every platform regardless, and a hidden platform's read failure painted the sync light red app-wide for a connection that was deliberately switched off. - The Run modal no longer runs off the top or bottom of the screen on a smaller laptop display with no way to reach the rest of it — it now caps to the visible height and scrolls internally, with the header and the Cancel/Run buttons always staying in view. - Every completed daily run used to leave the sync light showing a false "failed, stopped without finishing" for hours afterwards, even though the run had genuinely worked — the single worst thing the sync light itself did wrong, caught and fixed the same day it shipped. - A full sweep of the sync light and the last few commits caught and fixed roughly 30 further real inconsistencies: starting a sync used to briefly show the previous sync's leftover message as if it were a fresh problem; stopping a run showed a false green "finished" tick; three of the Run modal's own buttons broke their own display on first click; "Done" could show right next to a red failure light; a sync that failed completely could still record as a success; and several purchases-page filter dropdowns silently did nothing when changed, a pre-existing bug surfaced while checking this same area. - Two DealerKit functions carried one dealership's own internal team name hardcoded as a silent default. Neither is currently used by anything in the app, so this was never live, but both now require the team to be stated explicitly rather than guessed, so it can't misfire quietly if either is ever wired up later. ## v3.7, 2026-08-31 The manual purchase path rebuilt properly, and the left side panel removed from the whole app. ### Added - Adding a car by hand is now correctable, not a one-shot form: on a hand added row the price, fees, purchase date and supplier can all be edited afterwards from the drawer, each edit logged to the timeline with the before and after. A platform row still refuses every one of these, the platform stays the source of truth for its own cars. - The same reg can no longer be added twice — refused, naming the existing row — unless the earlier one was cancelled, so a genuine re-buy still works. - A real Supplier field: a managed dropdown, editable in Settings, with inline creation from both the Add car screen and the row's own dropdown menu (pick "+ Add a new supplier...", type the name, tick to save). A name already on the list is picked quietly rather than refused. Person collecting gained the same inline-add control. - VAT on a hand added car's fees is now a tick per fee, computed automatically at 20% (off by default), with a live running total as you type. Correctable afterwards from the drawer via a small `+VAT` tag on each fee line. A platform row's own invoiced VAT is never touched. - A Retail estimate field on the Add car screen, landing in the same override slot the row's own pencil already edits, so it survives the settings retail cap untouched. - Person collecting and Collection arranged can now be set when adding a car by hand, both optional. - **Delete this car**, for a hand added row only — removes it and its whole timeline for good, with a confirm step. A real platform purchase still only ever leaves the list through Cancel. - The Add car modal is two columns now, with required fields marked, real labels instead of placeholder text for the fee boxes, and a section grouping (car / purchase / collection) instead of one long stack. ### Changed - A hand added car no longer shows a Video stage, a Video column value, or an Auc status tracker — none of it applies to a car nobody sent a platform crew to film. The journey for a manual car is now Bought → Collection → Arrived. The Bought stage reads "Bought from [supplier]" instead of "Won on Manual". - A hand added car is excluded from CompMatch's own comp pool: a negotiated trade price is not a competitive auction clearing price, and counting it was biasing every suggestion downward. - The left side panel (Dashboard / Stock / BidBrain / Customers / Reports, mostly disabled "Soon" placeholders) is gone from every page. The dealership's own logo, the one real thing it held, moved to the top nav bar instead. The content panel is meaningfully wider as a result — the live auction grid now fits 3 cards per row at a normal laptop width instead of 2. - The Add car modal's spacing rebuilt: labels and fields no longer fight each other for vertical space (a label's own top margin and the grid's row gap were both adding space at once), a real divider now separates the form from its Cancel/Add buttons, and a lone Mileage field no longer sits half-orphaned next to dead space. ### Fixed - The "+ Add a new supplier" flow originally shipped with no visible way to confirm a typed name, and the dropdown right above it kept reading "Not recorded" the whole time, looking broken. Rebuilt as one control that replaces itself (dropdown → name box → back to dropdown, now selected) with an explicit tick to confirm, never two controls disagreeing on screen at once. ## v3.6, 2026-08-30 Five small nav and head bar tidy ups, all markup and CSS, nothing touching the gate, pricing, CompMatch, readers or DealerKit. ### Changed - The count beside the auction tabs is gone from Live Auctions, Purchases and Didn't win: the views chip right next to it already shows the same live number (confirmed live, switching a platform tab moves both together). Under Offer keeps its own count, it has no views control to fall back on. - The More button restyled to match the views chip exactly — same height, padding, border, weight, hover, and a real chevron on the right, reusing the existing chevron icon rather than a second copy. - The nav search box widened, its placeholder was genuinely truncating (144px of usable space against roughly 157-164px of text), and given `font:inherit` so it renders in the app's own Plus Jakarta Sans instead of the browser default Arial it had always silently fallen back to. - The "Calendar, coming soon" placeholder removed everywhere. It was never wired to anything. ### Fixed - The shortlist star's red badge no longer shows a bare 0. The cockpit was the one page of five missing the `:empty` CSS rule the other four already had, and `applyView` was separately writing the literal `"0"` back into it after every load regardless — both had to go for the fix to hold. Verified live both directions: starring a car shows the real count, unstarring clears the badge entirely. - The widened search box still clipped "model" to "mode" in real use even though the box-model numbers said it fit. Root cause: Safari reserves internal space for `type="search"`'s own native icon on WebKit, invisible to any CSS measurement — added `-webkit-appearance:none` (this box already draws its own magnifying glass, so nothing native was wanted there) plus a little more width as headroom. - The chevron restyle nearly shipped invisible: the views chip's own chevron rule is scoped to `.viewchip .car`, and the test suite's orphan-class scan would not have caught reusing it as-is, since it harvests `car` out of that descendant selector and counts it as already styled. Generalised the rule to cover both buttons instead of adding a second one. ## v3.5, 2026-08-30 The live auction card's own bottom section redesigned: Details, View on [platform] and "···" are one action row, and Details now opens a small floating panel that overlays the gauge and chips instead of pushing the card taller. Explored first as a series of published design artifacts (ten alternative approaches considered, one built out and iterated with real feedback) before landing in the real cockpit. ### Added - Details panel: click the pill to see every flag this car carries, in full sentences, plus its Glass's/Cazana valuation chips and the Cazana deep-link button, all in one place. Opens above its own trigger, overlaying the card's own gauge and chip row, closes on an outside click, Escape, or its own close button. Only one panel is ever open across the whole page at once. ### Changed - The old row of small traffic-light flag icons, its click-to-expand block, and the separate Valuations toggle are gone, folded into the one Details panel. A car with nothing to flag and no valuation shows an honest "Nothing more to add on this car" line rather than an empty box. - Comfortable/Compact removed from the toolbar: it only ever existed to bulk expand or collapse those two blocks, and there is nothing left for it to control once nothing expands in place any more. - Because nothing below the gauge can grow the card's own height any more, the row-alignment dead-space bug reported against the previous card design (a short car's own card stretched to match a taller neighbour, leaving a blank void where its flags used to end) cannot recur, by construction rather than by patching around it. ### Fixed - `.hf-card` also carries the shared `.card` class, which declares its own `overflow:hidden`; simply removing `.hf-card`'s own declaration left that value in effect from the other rule and the new panel was still being clipped. Needed an explicit `overflow:visible` on `.hf-card` to actually win the cascade, caught live before shipping. ## v3.4, 2026-08-30 A shared "Frame" (filter chips, a More filters drawer, a live result count, and one views control) built once and rolled out to all five pages, replacing several separate implementations that had drifted apart from each other over time; a defensive hardening pass across the whole app so a missing element or a bad platform read fails loud instead of silently breaking a page; and a long tail of real bugs found by a full UX review, from cards misaligning row to row to cancelled purchases only showing 88 of 541 real cancellations. No gate or pricing logic touched anywhere in this release. ### Added - The shared Frame: one filter chip row, a "More filters" drawer, and a live result count, common to all five pages. On the cockpit and purchases it also carries one views control, merging what used to be a fixed "slice" chip and a separate saved-views button into a single dropdown that reads "All cars", the name of whichever saved view you exactly match, or "Custom" the moment you deviate. Rolled out page by page: Under Offer first, then Lost (which gained real slices and range filters, losing its old Stats block), then Purchases (which also gained a real Margin column and promoted Add car into the toolbar, dropping its own Stats block, and gained a Stop control), then the cockpit itself. - The cockpit's own filter engine rebuilt to be criteria driven rather than reading the DOM directly (Run A), so gap fillers, Auto only and Held back could become ordinary filters instead of hard-coded chips (Run B), and every remaining range filter (mileage, reserve, distance and so on) became a real dual-handle slider, built in house rather than pulled from a library, then rolled out across the whole app (Run C/D). - Every filterable data point the cockpit and the purchases working list actually hold is now a real filter, each scoped to only the platforms that actually publish that field. - A five-pass visual design cleanup across the Frame: shared controls that were rendering as raw, unstyled browser defaults, one consistent scale for the frame's own panels and buttons, one labelled-field shape and rhythm through the purchases drawer, trimmed repeated filter copy with a scope note that belongs to its own field, and saved-view starring working on purchases too. - The settings gear swapped twice in quick succession: first to a plainer, more recognisable gear, then to lucide's own Settings glyph once Mark saw it and preferred it. ### Changed - A hardening pass across all five pages: guarded element access so a missing DOM node degrades instead of throwing, a loud rather than silent failure when the price ceiling setting can't be read, two separated failure domains on the purchases page so one bad write path aborts cleanly, rows escaped, UK dates and en-GB pound formatting enforced everywhere, sub-860px layout containment on all five pages, and Under Offer properly joined the BidBrain nav section. - The card foot (flags, Valuations, the platform CTA) redesigned so a car with few flags no longer misaligns against a row neighbour with many: the card is anchored at both ends, only the flags are allowed to float, and the AI-suggests line reserves its own space even when it's empty, so the flags start level on every card whether or not that specific car has a CompMatch suggestion yet. - Read full history on the Lost page now confirms before it runs and blocks the page while it does, and moved into the same sync modal purchases already uses rather than its own standalone button. - Didn't win's old fixed slice chip (Last 30 days, Lost by under £250) converted onto the same one views control as everything else, and the slice chip mechanism was then retired outright from the codebase, nothing renders it any more. ### Fixed - Cancelled purchases were showing 88 of 541 real cancellations: walked-away cars (a cancellation recorded under a different reason) weren't being folded in. They're all one thing now, filterable by the real cancellation reason, a filter which only ever appears once you've actually filtered to Cancelled. - The cockpit's own filter and view counts were ignoring the search box, so a count could disagree with what was actually on screen; a saved purchases view was separately dropping almost everything it was meant to filter to; Didn't win's two headline counts could legitimately disagree with each other. - The cockpit was effectively unusable on a phone: the toolbar went dead under the run progress bar, the Filters button silently switched off your shortlist, a stale count badge lingered after a filter changed, and the Run button's own red "needs attention" dot was backwards, showing when everything was actually fine. - The "All cars" row on the purchases saved-views list did nothing at all when clicked. - A disabled platform could freeze the whole cockpit's own date. - The Shortlisted badge was counting stars on lots whose auction had already ended. - Under Offer never got its Calendar placeholder and understated what the search box could actually search; a few smaller ones landed in the same pass — a keyboard trap on the flag icons, a duplicate search chip, one clearer word on the sync button, and the purchases toolbar's own right hand cluster no longer spreads across the whole row. - Checked whether the cockpit needed the same narrow-screen search treatment the other three pages have; it doesn't, for a reason specific to this page (recorded in the code rather than silently added and left unused). ## v3.3, 2026-08-27 The purchases page rebuilt as a genuine working list, not a report: every car between winning the bid and landing on site, with its own journey, timeline, and a real DealerKit sync panel, worked entirely from inside the app instead of an Airtable alongside it. Built by Steven on his own copy, reconciled with Mark's own parallel work (Under Offer, shortlist sync, the Auc Status and Motorway won-date fixes) landing in the same release cycle; see `CLAUDE.md` for how that merge was resolved. No gate or pricing logic touched anywhere in this release. ### Added - The car overlay redesigned as two permanent sections: Car details on the left (Car, Money, Collection and logistics, Notes and documents, in grouped, always-visible rows), and a right hand panel carrying three tabs, Journey, Timeline, DealerKit, so checking the timeline or the DealerKit sync state never hides the car's own record. The drawer always opens wide enough for the split, dragged wider if you like. - The Journey tab: a Next up card plus a vertical stage rail (Bought, Video check, Collection, Arrived), each stage carrying only its single closing control, video status, Paid for, Check in. Chips (a price knocked off after agreeing it on the video) and full cancellations are recorded from here, both logged to the car's own timeline alongside every field edit and comment, one continuous history per car. - The DealerKit tab is a full sync panel, not a summary: every pushable item (price, fees, photos, documents, retail price, due in date) shown as on DealerKit, missing there, or ticked here ready to push, with an honest "not charged on this purchase" state for whatever does not apply, and Upload/Check DK record run straight from the panel. - Column drag to reorder and a resize grip on every header, both held per dealership; a Fields menu to show or hide any column per table or per record; a Views menu to save the current filters and sort under a name and apply them again with one click. - Four page views, Due in (the working list), In stock (the archive, checking a car in moves it here), Cancelled (every fallen through sale kept with its reason) and All, each showing only the columns that mean something there, plus a real column footer with car count, price, chip and paid rollups. - A car bought outside the platforms (a local trader) can be added by hand, its own record never touched by any platform sync afterwards. - The uploaded service history and V5 photos are now visible on the car's own record, not just the table, opening the same shared gallery. ### Fixed - Long notes no longer cut the editor off mid word: the scroll listener added to dismiss floating menus was catching the note box's own internal scroll and blurring it out from under the typist. - Every price shows pence only when they are real (`£7,500`, not `£7,500.00`; `£4,820.80` keeps both digits when it should). - Every date the app renders reads UK format, `DD/MM/YYYY`, not the raw ISO or American order that had been slipping through in a few spots. - An empty service history reads "No service history" instead of a blank dash, in the table and on the record. - Dropdown cells now carry a visible caret so it is obvious they open a menu, and clicking anywhere else, or scrolling, reliably closes it, which it previously did not always do. - A note now saves with a visible tick button as well as Enter, so finishing an edit does not depend on clicking precisely off the cell. ## v3.2.3, 2026-08-27 ### Fixed - A car won on Motorway within the last hour reads its own date as, e.g., "57 mins ago"; the date parser only ever recognised hour/day/week/month units, so a freshly won car's bought date silently stayed blank even though the rest of its purchase data had genuinely synced. The purchases page's own default date filter never matches a row with no bought date, so the car was correctly synced but invisible until this fixed the one shared parser behind the purchases, Didn't win and Under offer pages alike. ## v3.2.2, 2026-08-27 An Under Offer page (a bid currently leading, not yet a confirmed sale) and a real purchases page status fix, caught and corrected the same day it shipped. No gate or pricing logic touched. ### Added - A new Under Offer page, reachable from the same Live Auctions/Purchases/ Lost switcher as the other admin pages, showing every car Motorway's or Carwow's own real Bids & Offers > Under Offer screen currently lists as a leading bid. ### Fixed - The purchases page's own Auc Status tracker always showed its Bid stage as accepted, even for a Motorway purchase genuinely still at "Awaiting confirmation" (caught live on a real example). It now reads the real status honestly, matching Motorway's exact rule for when a win actually counts as accepted. - A same-day attempt to also surface that same "not yet accepted" state on the new Under Offer page was reverted once it became clear it did not match what Motorway's own portal actually shows there; that page now shows only real Bids & Offers data, nothing from the purchases table. ## v3.2.1, 2026-08-27 The one platform action BidBrain has been approved for since 2026-06-09, finally built: starring or unstarring a car in the cockpit adds or removes it from that platform's own real shortlist or watchlist, never a bid. No gate or pricing logic touched. ### Added - Shortlist/watchlist sync to Motorway, verified live on a real auction car, both directions, genuinely persisting server side. Carwow's own mechanism is built and correct (confirmed against the platform's own real Stimulus controller source) but still needs a live check against an active listing once its stock is out of the daily dead window; a real mismatch always surfaces as a visible error, never a silent false success either way. Happens automatically in the background within a few seconds of a star click, no extra step needed, with a manual "Sync watchlist to platforms" button in the Run modal to catch up any backlog. - A second Motorway account, accounts@rightdrive.co.uk, receives the same shortlist sync independently (its own separate browser profile, its own automated daily login at 06:00), for whoever completes an actual purchase, which needs that inbox's own SMS code. Shown on the Integrations tab as its own connection, can be hidden the same way as any other. ### Fixed - An automated Motorway login that failed for a genuine reason (a wrong password, most commonly) reported a blank error message instead of the real one, because the check happened after an extra page navigation that had already cleared it. Both Motorway login scripts now check immediately after submitting, so a real failure always shows its own real reason. ## v3.2, 2026-08-27 Auction4Cars and Dealer Auction join the purchases page, a real DealerWay logo, seven sites now log themselves back in, and a round of navigation tidying: a Live Auctions/Purchases/Lost page switcher, a sticky filter row, a search icon, and Purchases/Didn't Win pulled back out of the left sidebar once that switcher made them a duplicate path. No gate or pricing logic touched. ### Added - Auction4Cars and Dealer Auction purchases now feed the purchases page, reading each account's own real completed buy history (Auction4Cars' own 13 month "My Purchase History", Dealer Auction's own itemised Invoice history). Dealer Auction also gets a real full photo gallery, read off the won advert's own page since the invoice list itself carries none. - A real DealerWay logo everywhere a platform logo already appears (the platform toggle, the purchases table's own Platform column, the card CTA), a fifth captured logo alongside Motorway, Carwow, Auction4Cars and Dealer Auction. - A small link icon next to each purchase's own vehicle name, opening the original listing (with its photos) in a new tab wherever a listing url is known, the one way to see real photos for a platform with no gallery reader of its own. - Automated daily logins for Auction4Cars, Dealer Auction and LE Capital, the same keychain based mechanism already covering Cazana, Motorway, Carwow and DealerKit, bringing every site this project reads to a self healing login. - A Live Auctions/Purchases/Lost page switcher, a search icon, a settings cog and a sticky second filter row, all standardised across the cockpit, purchases and Didn't Win pages so the three read as one product rather than three separately evolved ones. The purchases table's own Platform column shows the real platform logos too, in place of the old coloured text pill. - Comfortable/Compact switched from text to icons on the cockpit's own filter row, moved next to the Filters button. ### Changed - The DealerKit push status on the purchases page is now a real red to green fraction badge (matching DealerKit's own job progress indicator) instead of a plain tick, showing exactly how many of the real checklist items are on record, not just none/some/all. - Purchases and Didn't Win removed from the left sidebar: with the same three destinations already reachable from the new top nav page switcher, the nested links (and the down arrow that used to collapse them) had become a duplicate rather than a second real path. ## v3.1.4, 2026-08-26 A live progress bar on the purchases page for every sync and check, sortable columns, and one small rule learned from a hide note. No gate or pricing logic touched. ### Added - A shared progress bar on the purchases page, matching the cockpit's own, shown for Sync now, Check all DealerKit records, and every per row check or push. It compares against the server's own clock rather than the browser's, so a genuinely stuck run past 90 seconds turns amber and says plainly "no update for Ns" instead of just spinning. - Sortable columns on the purchases page: click any header to sort by it, click again to reverse, with an arrow marking the active column and any unset value always sorting to the end, whichever direction. - "master" added to VAN_MODELS, learned from a hide note on a 2019 Renault Master tipper. ### Fixed - The purchases page's Sync now progress bar could show a stale, unrelated message left over from an earlier daily run rather than the sync's own real progress, because purchases_run.py's own sync never wrote to the shared progress file. It now reports real progress through every stage, including a live count through the slowest step, the payment breakdown lookup per purchase. ## v3.1.3, 2026-08-26 A Due in view on the purchases page, and older purchases out of view by default. No gate or pricing logic touched. ### Added - A Due in view, next to the Stats button, showing every still active purchase (not yet collected): those with a real due in date today or later, sorted soonest first, then anything still active but not yet confirmed, clearly labelled rather than missing entirely. Works alongside the platform tabs, not instead of them. ### Changed - The purchases page now starts filtered to purchases bought from 1 June 2026 onwards, display only: the real records are untouched, clear the date box or click Reset to see everything again. ## v3.1.2, 2026-08-26 A cap on the resale figure with a manual override, Carwow's own due in date, pushing it through to DealerKit, and the purchases page's Auc Status column redesigned into a real 3 stage tracker. No gate or pricing logic touched. ### Added - A settings level cap on the Retail est. figure (Settings > Overview), optional, none by default: a purchase priced above it shows and pushes to DealerKit at the cap instead of the formula's own higher figure. A pencil button on the purchases page row lets you set the figure manually instead, always winning over the cap; the real, uncapped formula figure is kept underneath either way, never lost. - Carwow's own "Collection confirmed" card, on a won car's own listing page, now feeds the Due in column the same way Motorway's delivery timeline already does. - Due In / In-stock On, a real DealerKit field on the vehicle's own Miscellaneous edit panel, now gets BidBrain's own collection date pushed to it as a fifth pushable item alongside price, photos, documents and retail. Unlike Retail Price it always stays in sync with BidBrain's own latest figure rather than only ever being set once. - The purchases page's Auc Status column is now a 3 stage progress tracker (Bid, Documents, Delivery) instead of a single status word with a dot: a £ icon for Bid (always shown reached, every row here is already a won purchase), a document icon for Documents, a car icon for Delivery, each green once reached, orange while known but not yet, grey when nothing has been read yet. The real underlying status word is still there as a hover tooltip. ### Fixed - The Delivery stage icon was reading Motorway's own coarse purchases list status word, which does not always keep pace with its separate delivery timeline page (the actual source of the real due in date): a car with a real, already pushed due in date could still show Delivery as not yet arranged. It now trusts a known due in date outright, falling back to the status word only when no date has been read at all. ## v3.1.1, 2026-08-26 ### Fixed - A fresh install crashed opening the purchases or settings page before its very first run: reading the saved settings from a database that had never been initialised raised "no such table: settings" instead of reading as nothing saved yet. Present since theme colours started being read fresh at render time, and the same failure that had been turning the CI test run red on every push since. A brand new, never initialised database now reads the same as one with no settings saved. ## v3.1, 2026-08-26 Logins that look after themselves, a redesigned Integrations tab, and real control over which connections run and which even show. No gate or pricing logic touched. ### Added - Automated daily logins for Carwow, Motorway and DealerKit, joining the Cazana one: each site logs itself back in on its own schedule (Carwow and DealerKit 17:00, Motorway 06:00, clear of everything else that touches its shared browser profile) from a credential stored once in the macOS keychain via `setup_credential.py `, never in the repo or anywhere readable. An expired session now heals itself instead of silently breaking the next run. - The Integrations settings tab redesigned into one flat, honest connections list: anything broken floats to the top under a banner naming it, every row carries a single plain language state (down since when and why, working and how fresh, or not set up here at all) instead of the old two separate signals that could contradict each other, an Auto or Manual badge shows which logins fix themselves and when, the explanatory prose sits behind an info button, and each automation's own switch applies instantly, no separate Save button. While a login window is open, a strip explains why every other login button is paused, and the other buttons visibly disable. - Hide connections a dealership does not use. Hiding is more than cosmetic: it also switches that site's own automations off (a platform's "Read this platform" toggle, an integration's own enabled state), and showing it again switches them back on. Hidden connections stay in a muted section at the bottom of the list with a one click Show again. Cazana alone cannot be hidden, every run prices on it. - The Run modal no longer shows a platform that is hidden or switched off in Settings at all, and re checks the live settings every time it opens, so a change applies immediately without waiting for the next run. A hidden site's own buttons (Sync LE Capital funding, Run Glass's checks, the DealerKit pair) vanish with it. - "Read this platform", a real persistent per platform setting on each platform's own tab: switched off once, that auction house is excluded from the scheduled daily run and every manual run until switched back on, whatever the Run modal is sent. - A personal card display preference: clicking any card's own flag notes or valuation detail arrow now expands that section on every card at once, and the choice survives a reload, per browser, never changing the shared default. - The single theme colour split into four genuinely independent ones (background, price, pill, icon) plus the left panel colour, each with its own picker on the Theme tab. ### Fixed - A dead Carwow session used to read as "0 auction cars today", indistinguishable from a genuinely quiet day; it now fails loudly with a clear relogin message and the run falls back to the last good Carwow cars instead of wiping them. - The login status text and the health dot on the settings page could contradict each other (a red dot right next to "logged in within the last day", a green dot next to "no saved login yet"): a successful login, automated or through the page's own buttons, now clears a stale failure record, and the two signals are merged into one honest state. - Logging in through the settings page's own "Log in to X" buttons never cleared a stale down flag, so a genuinely working login looked broken however many times it was redone. - A purchase with any retail estimate on file, even a stale wrong one, was permanently skipped for a fresh Cazana lookup. - Settings page reloads (hiding a connection, finishing a login, the software update) used to dump back on the Overview tab; the current tab now survives any reload. - Motorway's automated login wrongly reported failure when the shared profile was already logged in. - A merge conflict during the v3.0 redesign had silently emptied settings.html's own JavaScript entirely; caught and fixed before any release carried it. ### Changed - Software update moved from the Integrations tab to the top of Overview, it is not a connection. - "Click Run Glass's checks" prompts no longer nag when Glass's is switched off or hidden, pricing on Cazana alone is the configured state then. ## v3.0.1, 2026-08-26 The purchases page's own DealerKit column reworked into a real, honest "what's actually there" check, plus a real image reliability fix. No gate or pricing logic touched. ### Added - Check DK record, per row and in bulk: a live read of what is REALLY on a car's own DealerKit record right now (price, auction fee, delivery, indemnity, photos, service history, V5, retail price, due in status), independent of whatever BidBrain itself last attempted to push. Auction fee/delivery/indemnity are read from the car's real expense line items; Service History and V5 are detected from the real per-document label or filename, wherever either survives. - A tick box per pushable category (price, photos, documents) sits right where its own presence mark used to be, pre ticked to whatever is missing on DealerKit and present in BidBrain's own data, with a single Upload to DK button; a new Check Motorway/Check Carwow re reads one purchase straight from its own platform. - A hard gate: DealerKit will never create a brand new Due In record with no mileage on file (it would let DealerKit guess the mileage itself). Only applies to a genuine first push; an already existing record is unaffected. The page asks before pushing anyway. - Four distinct tick colours (blue ticked, amber present but not yet on DealerKit, green already confirmed there, red not confirmed at all) with a tool tip on every state, and DealerKit's own login health now shows a red or green flag on the settings page like every other platform. ### Fixed - A car missing real DealerKit data could still show a false green "all done" tick, both from an earlier miscount and from BidBrain marking a category "pushed" whenever it was merely attempted, even when there was nothing real to push (a purchase with no local documents at all still got marked as if its documents had gone up); both fixed at the source, and the real rows already carrying a stale flag from an earlier run corrected directly. - Motorway's own photos and documents went permanently unviewable after about an hour: their urls are signed with a real expiry, not a file format issue. They are now downloaded to a real local file the moment they are read, the same approach Carwow's own documents already used, closing the same exposure in the DealerKit push pipeline itself. - The settings page's "I'm logged in, save it" button could get stuck disabled for the very site being logged into. ## v3.0, 2026-08-26 A full visual redesign: a new design system, built first as a standalone mockup, brought into all four real pages (BidBrain's own live auction cockpit, Purchases, Didn't win and Settings). No gate, pricing or CompMatch logic changed anywhere in this release, every real feature and its own behaviour (filters, sorts, valuations, runs) works exactly as it did, just restyled and reorganised. Also carries a small handful of real features that had shipped since v2.9 but were sitting unreleased. ### Added - A persistent left app nav shared by all four pages: real links for BidBrain, Purchases and Didn't win (nested under BidBrain) and Settings, plus Dashboard/Stock/Customers/Reports shown as inert "Soon" placeholders for the future Dealer OS rather than dead links. Collapses to a 60px icon rail via a toggle in the logo row (remembered per browser, so a reload does not spring it back open), carries Right Drive's own real logo (the full wordmark expanded, the round smile mark alone collapsed, its own white background chroma-keyed to real transparency), and a down arrow on the BidBrain item collapses its own nested Purchases/Didn't win links. - A new Theme tab in Settings: the primary brand colour and left panel background colour, plus real logo (full and small) and favicon uploads, validated and stored server side with no files written to disk, defaulting to Right Drive's own real logo when nothing is uploaded. - The search box moved out of each page's floating Filters panel into the top nav bar itself, next to the refresh icon, on the three data pages (cockpit, purchases, didn't win); Didn't win's own now-empty Filters panel (search was its only field) was removed outright rather than left pointless. - Purchases and Didn't win both gained a real heading matching the cockpit's own per-platform one: the platform's logo (or "All") plus a live count, with the old permanent summary tiles moved into a Stats dropdown button opened the same way the notification bell already is. - A quiet, disabled placeholder calendar icon next to the refresh icon on cockpit, purchases and didn't win, matching the sidebar's own "Soon" treatment until there is a real calendar feature to wire it to. - A real Cazana deep link: the "Cazana" label in a shortlist card's own Valuation detail, and a dedicated "Open Cazana lookup" button, both linking straight to that exact car's own live Percayso search result (captured at the moment of a genuine lookup, cache hits included), not just naming the source. - A "Due in" (collection date) column on the purchases page, read off Motorway's own delivery timeline for the day transport actually collected the car, the real date DealerKit's own Tax Date field should carry, not the same thing as the "Delivered" date already shown elsewhere. ### Changed - The whole colour and type system moved to the new design system's real oklch tokens and Plus Jakarta Sans, card radius and shadow softened, a subtle hover lift added; the filter sidebar (all four pages) became a floating panel opened from a Filters button rather than a permanent column, freeing the card grid back to multiple columns on a 14 inch laptop. - Shortlist card photos are real 4:3 now instead of a fixed pixel height, and the grid's own per-card minimum was retuned so exactly 3 cards fit a 14 inch laptop's real width (up to 6 confirmed live on a 27 inch monitor). - Every page's top right icon cluster now reads the same, left to right: refresh (Run now / Sync now), the shortlisted star, the notification bell, the signed in user's avatar, all icon-only with no bordered boxes; every remaining glyph or CSS pseudo-element icon (gallery controls, sort arrows, close/star/delete buttons) was replaced with a real inline SVG matching the nav's own feather style. - The sidebar's own collapse toggle changed from a plain chevron to the standard "box with a left panel" icon and grew bigger, alongside a bigger Right Drive logo. - The platform All/Motorway/Carwow/... tabs on every page now share the exact same grey the left panel itself uses, previously a slightly different or stale tone on three of the four pages. ### Fixed - The top nav bar's own `position:sticky` had quietly never worked since the app shell was first built: `overflow:hidden` on the content panel (there to clip to its own rounded corners) breaks sticky positioning on anything inside it. Moved the rounding onto the bar's own top corners instead; it now genuinely stays pinned while the page scrolls underneath it. - `daily_run.py`'s settings cache is correct for its own short lived CLI runs but was wrong for `serve.py`'s long running process, so a saved theme colour or logo only ever took effect on the very first render after the server started; the settings save handler now clears it immediately. - Didn't win's own new heading/Stats/Filters styling had silently never landed on that page at all (the CSS edit was anchored to text only purchases.html happened to share); fixed directly against its own copy of the block. - The red notification badges (the bell, the shortlisted star) were cutting the number off at the bottom of the circle: `box-sizing: border-box` shrinks the badge's own content area once its border is subtracted, but `line-height` was still sized for the full box. Switched to flex centring, the same mechanism every other icon button here already uses. ## v2.9, 2026-08-25 ### Added - Every page now has a "Shortlisted cars" star icon in its own nav bar, the same plain look as every other nav icon, gaining a small corner badge with a real count whenever there is at least one, the same convention the notification bell's own badge already uses. On the cockpit it's the existing star toggle; on purchases, Didn't win and settings it deep links to `cockpit.html?view=shortlist`, landing with the shortlist filter already active. - The notification bell (previously cockpit only) now also appears on purchases, Didn't win and settings, fed live by a new `GET /api/notices` endpoint that recomputes the same health notices from the last saved run cache; the bell itself turns red whenever there is a real alert, not just its small badge count. A notice with a real fix still links straight to it (Settings > Integrations, or the cockpit's Run modal via a new `?openrun=1` deep link on pages with no Run modal of their own). - The BidBrain logo in every page's nav bar now links back to the cockpit, the live auctions homepage. ### Changed - Nav icons are now in the same fixed order and same look on every page: favorites, purchases, notifications, lost bids, settings, each page's own primary action button(s) first (Run now, Sync now, Save settings, all now sitting in the identical right aligned action group) and a Back to the cockpit icon last where relevant; a page skips its own icon (no settings icon on the settings page, and so on) rather than showing a dead link. - The settings page's internal tab set (Overview, Motorway, Carwow, Auction4Cars, Dealer Auction, DealerWay, Integrations) now uses the same pill style as the cockpit's own platform tabs, and moved out of its own row into the top nav bar directly, sitting next to the brand exactly where the cockpit's tabs sit. - Save settings and Save integrations moved from a sticky bottom bar into the top nav bar, next to Sync now and Run now's own position on the other pages. ## v2.8.19, 2026-08-25 ### Changed - The settings page's Integrations tab gained a dedicated Valuations card for Glass's and Cazana (previously split between the Integrations card and the Auction house card), and every login row's relogin buttons now sit to the right of its status text instead of stacked underneath it, using the available width better. ## v2.8.18, 2026-08-25 ### Added - The Didn't win page gained an All/Motorway/Carwow platform toggle, matching the cockpit and purchases page, and its Platform badge now reads the row's own real platform instead of being hardcoded to Motorway. Ready for a Carwow Didn't win reader whenever that gets built; the Carwow tab honestly shows no rows until then. ## v2.8.17, 2026-08-25 ### Changed - The purchases, Didn't win and settings pages' cross navigation links are now icons (bag, target, wrench, home), matching the cockpit's own nav buttons, instead of plain text links. ## v2.8.16, 2026-08-25 ### Changed - The Didn't win page's Retail est. now shows Cazana 115% and Margin (retail est. less our own bid) in a hover/tap popover, the same pattern the purchases page already uses, rather than a separate Cazana 115% column. ## v2.8.15, 2026-08-25 ### Added - The Didn't win page also shows a Cazana 115% column now, alongside Retail est., the same plain retail-times-1.15 figure purchases already store directly. ## v2.8.14, 2026-08-25 ### Added - The Didn't win page gained a Retail est. column and a new summary tile: the running average of the real sold-for price as a % of CompMatch's own Bias bid across every lost bid with both known, the same comparison the purchases page already shows. ## v2.8.13, 2026-08-25 ### Changed - The settings, purchases and Didn't win pages now share the cockpit's own nav bar and link directly to each other, so you can move between any of the four pages without detouring back through the cockpit first. The settings page's old dark topbar is gone, replaced with the same sticky control bar the other pages already use. ## v2.8.12, 2026-08-25 ### Changed - The Status column now shows the platform's own real status as the main text always (Collection arranged, Delivery requested, Checks complete...), with a hover/tap popover only when there's a payment breakdown still marked Estimated, showing both the status and the payment note together rather than stacking them and growing the row taller than the rest of the table. ## v2.8.11, 2026-08-25 ### Added - The purchases page's Winning bid popover now also shows the "Bias bid" (what CompMatch would suggest for that exact car today) and what % of it the real winning bid actually came to — under 100% means the car was won for less than CompMatch's own suggestion. A new summary tile shows the running average across every purchase with both figures known. ### Fixed - Carwow's own purchase status came through as a raw, unformatted value ("collection_arranged"); both platforms' statuses now display in plain sentence case ("Collection arranged"). ## v2.8.10, 2026-08-25 ### Changed - The purchases page is tidied up for small screens: from 17 columns down to 11. Winning bid's own fee breakdown (Fee, Transport, VAT, Total paid) and Retail est.'s own Cazana 115%/Margin both moved off their own columns into a small popover shown on hover, or a tap on a touch screen. Service history and V5 now each show only the first document photo with a small count badge, rather than a whole row of thumbnails; click still opens the full gallery. ## v2.8.9, 2026-08-25 ### Fixed - CompMatch's comps could be years old: every comp's own retail was looked up today while some sales happened up to 5 years ago, so depreciation was systematically inflating old comps' own % of retail (a 4-year-old sale counted as a full-price comp today). Comps older than 18 months (or with no readable date) are now excluded entirely, not just capped. Model grouping was also splitting real groups apart (BMW comps keyed on their own engine badge, "118D"/"320D", never matching a live "BMW 1 Series" car; Mercedes A-Class was split three ways) — both now merge correctly. Semi-automatic comps, previously stranded in a bucket too small to ever match anything, now pool with Automatic (manual stays completely separate, the hard gate is untouched). Net effect on the real list: coverage held (126 to 131 suggestions), but with 24 stale-only ones honestly dropped and the 102 kept moving a median -5.1% as inflated old comps came out. ## v2.8.8, 2026-08-25 ### Added - A vehicle's own Motorway page carries a real, confirmed transmission and engine size (checked live), so lost bids and Motorway purchases missing them can now be backfilled from a genuine source, not just a name badge or a recent sighting. Recovered 634 real confirmations in one pass (transmission coverage across the whole comp history rose from 180 to 814 of 1,038, including the first real "Manual" confirmations this project has ever had). With the transmission hard gate from v2.8.7, suggested cards rose from 13 back up to 119 of 458. ## v2.8.7, 2026-08-25 ### Changed - Manual and automatic are now a hard gate for the suggested bid, never blended together even when there is too little data of one kind: a car's own transmission is only ever compared against real comps of that exact same transmission, with no fallback to a mixed pool. Since "Manual" is essentially never a confirmed value anywhere in the historical data yet (only real automatics are currently identifiable, see v2.8.2/v2.8.3), this sharply reduces how many cards get a suggestion right now (from 209 to about 13 of 458), most cars are manual and currently have no confirmed manual comps to compare against. This closes as real transmission data accumulates from daily reads going forward. ## v2.8.6, 2026-08-25 ### Fixed - CompMatch's % of retail calculation now excludes any individual comp whose own winning price came to more than 105% of its own retail (almost certainly stale or mismatched historical data), rather than only capping the final figure afterward. Checked live: this bad data was giving a genuinely backwards signal for some models, making automatics look like they sold for LESS of retail than manuals, the opposite of reality. After excluding it, most models correctly show automatics selling for more. Cars needing the retail cap dropped from 69 to 19 of 209 suggested cards. ### Added - CompMatch now compares a car against other examples of the SAME model on the SAME transmission specifically, once there is enough of that model's own automatic (or manual) history, rather than folding transmission into a single blended similarity score. A cross-model comparison was tried and rejected, it mixed in unrelated models and gave a backwards signal. ## v2.8.5, 2026-08-25 ### Fixed - The suggested bid could recommend paying MORE than a car's own known retail value, found live on a real Ford Kuga (£10,807 suggested against an £8,769 retail): about a third of comps for some models show a historical winning price above their own retail, most likely stale Cazana lookups rather than genuine overpaying. The suggestion is now capped at this car's own retail value regardless of what produced the raw figure, and says so in its own tooltip when it applies. Affected about a third of all suggested cards. ## v2.8.4, 2026-08-25 ### Added - CompMatch can now learn from another dealership's own data too, when both sides agree. Each dealership exports its own comps (make/model, winning price, retail, mileage, age, transmission, engine size and grade only, never a registration, a price breakdown or a margin) to the shared repo via a new "Sync CompMatch data" button or `share_comps.py --sync`, and every suggested bid is pooled from both. Off by default, opt in by setting `DEALER_NAME` in your own `dealer_config.py`. ## v2.8.3, 2026-08-25 ### Added - CompMatch now also compares condition grade when ranking real comps, alongside mileage, age, transmission and engine size. Every car BidBrain reads each day now logs its grade too (mileage, engine and transmission were already added in v2.8), so a future lost bid or purchase is matched against real scraped data from day one rather than needing a derivation, no live re-read needed. ## v2.8.2, 2026-08-25 ### Fixed - The suggested bid (CompMatch) barely had any real age or transmission data to compare against: only 5 of 1,038 real comps carried a year and none carried a transmission, so most suggestions were quietly falling back to a plain "same make and model" comparison rather than the intended like for like one. Fixed by deriving a real year from the registration's own UK plate age identifier and an automatic badge from the platform's own vehicle name, both free, no lookup needed, closing the gap to 981 of 1,038 comps carrying a real year. Like for like comparisons now cover about 89% of suggested cards, up from 61%. ## v2.8.1, 2026-08-25 ### Added - The suggested bid to win is now scaled to this specific car's own retail value when enough real comps carry a retail figure of their own: each comp's winning price is expressed as a % of its own retail first, and that % is applied to this car's own retail, so two comps of different spec (and so different retail) are compared on how competitively each one actually sold, not treated as equally comparable raw prices. Falls back to a flat raw price comparison when this car's own retail isn't known yet or too few comps carry one. ## v2.8, 2026-08-25 ### Added - The suggested bid to win now compares like for like rather than just "same make and model": real comps are ranked by how close their own mileage, year, transmission and engine size are to the specific car being priced, using the closest matches once there is enough real data to rank on, falling back to the full model group otherwise. A card's own tooltip says which kind of comparison it got. - Every car BidBrain reads each day (sightings, not just gate passers) now also logs its mileage, engine and transmission, so a later lost bid or purchase can be filled in for free from BidBrain's own database rather than a fresh live read. `bid_retail_enrich.py --fields` backfills year, transmission and engine onto existing lost bids and purchases this way; the mileage backfill checks the database first too, only falling back to a live Motorway read when nothing is on file. - Purchases now store the real mileage already read off the platform's own purchase page (previously used once for a Cazana lookup, then discarded), so it is available for the suggested bid model with no extra page visit. ## v2.7.12, 2026-08-24 ### Added - A suggested bid to win, shown directly on a shortlist card's own gauge: a purple marker and figure learned from BidBrain's own real history, both won purchases and lost auctions (a lost auction's own real sold price is exactly as precise a data point as what we actually paid to win one). Shown only when there is enough real comparable history for that make and model (55 real model groups currently qualify), never a guess otherwise. ## v2.7.11, 2026-08-24 ### Added - A new "Didn't win" section: every real Motorway auction bid on and lost, with the real price it actually sold for, read off Motorway's own Bids & offers screen. Ready for a future suggested bid feature, the real winning price is a signal purchases alone never carry. A new nav icon on the cockpit opens it, alongside a search box, a Sync now button and a one off "Read full history" backfill (the whole real history, 931 real lost bids, was captured live). Scheduled daily alongside the purchases capture. ## v2.7.10, 2026-08-24 ### Fixed - Two real DealerKit value discrepancies found on older Carwow records (WG18EZE and HJ17ECW's own Buyers Premium, and HJ17ECW's own Delivery Charge) were corrected to BidBrain's own captured, confirmed correct figures, replacing the wrong ones rather than leaving them. - A genuine duplicate expense on WX16AUE (a whole standalone extra Delivery Charge record, left over from however the older stock was first entered) was removed, leaving one clean, correct expense. - `add_missing_expense_items` could silently fail to open a supplier's own expense at all when that expense's row happened to sit at the very bottom of DealerKit's own scrollable expenses list, its own real edit button present in the page but effectively unclickable off screen. The matched row now scrolls into view first. ### Added - New `delete_expense` in the DealerKit write module, for removing a whole duplicate expense record via its own real "More" menu (distinct from a single line item's own delete, which only clears client side until an expense keeps at least one item left to save). ## v2.7.9, 2026-08-24 ### Fixed - Pushing a purchase to DealerKit could create a genuine duplicate "Due In" vehicle record if that reg was already in DealerKit as a completed, disposed sale: DealerKit's own creation endpoint only refuses a repeat reg still active in some other stage, not one already sold. A reg is now looked up in DealerKit first, before ever attempting to create anything, closing the gap regardless of the existing record's status. 13 real duplicates created by an earlier batch push were found and removed; the original sold records were confirmed untouched throughout. ### Added - All 23 Carwow purchases are now fully pushed to DealerKit: purchase price, delivery fee, buyers premium, service history and V5, matching what Motorway's purchases already had. A handful of older records already carried some of these figures from before, entered by hand; those were never overwritten, only genuinely missing pieces were added. ## v2.7.8, 2026-08-24 ### Added - The purchases page's "Sync now" button now opens a modal to sync Motorway and Carwow together or individually, the same shape as the cockpit's own Run modal, ready for later auction houses to join the same way once they feed the purchases table too. - Each purchase row gained a DealerKit push status button: grey "Push to DK" until that car has been pushed, green "Synced" once it has. Clicking it either way pushes just that one vehicle, so a single car can be re pushed without running the full batch. ## v2.7.7, 2026-08-24 ### Fixed - Carwow's own service history documents were being captured and saved correctly, but never shown: the Service history column only ever read Motorway's own url based field, ignoring Carwow's (a list of local file paths on this Mac). Both platforms now show real clickable thumbnails in that column, and the V5 column does the same instead of just a page count, now that a local path is known to be servable directly from this page. - 13 of 23 real Carwow purchases also had their service history list polluted with V5 pages that had leaked in from an earlier version of the capture logic, before the two were split into separate fields. Cleaned directly in the database; the current capture code already keeps them apart correctly going forward. ## v2.7.6, 2026-08-24 ### Fixed - The purchases table still forced a horizontal scrollbar on a normal wide monitor even after v2.7.5's own nav bar work, because the page's own content width was capped well below what a real monitor could show, and one column (a car's service history thumbnails) could grow without limit. The width cap is now generous enough to use a real monitor's own space, and that one column scrolls sideways within its own small cell instead of stretching the whole table. Verified live: no scrollbar at 1920px with all 16 columns visible, and a narrower window still scrolls the table within its own frame rather than the whole page, same as before. ## v2.7.5, 2026-08-24 ### Fixed - A purchase's Retail est. on the purchases page could silently come from two different formulas depending on whether BidBrain happened to have a recent buying assessment on file for that reg: usually the intended Cazana based dealer pricing formula, but sometimes the buying brain's own governing value instead, a different figure meant for a bid ceiling, not a resale price. Found live on two real purchases (neither had a Cazana figure to show, and their retail estimate didn't match any real dealer price ending). Now always computed the same way, from a live Cazana lookup. ### Changed - The purchases page's nav bar now matches the cockpit's own single control bar (brand, platform tabs and actions all in one sticky row), replacing the old stacked dark topbar plus a separate tabs row, freeing up vertical space and keeping the two pages consistent. - Added a search box (reg, make or model) and a bought date range to the purchases page's own control bar. - The rest of the purchases page is now responsive: the control bar's filters and actions reflow on a narrow screen, and the summary tiles and table gain extra breathing room down to a phone's own width (the table itself already scrolled sideways within its own frame, never the whole page). ## v2.7.4, 2026-08-24 ### Added - Carwow purchases now get mileage and a Cazana based retail estimate, the same logic Motorway's own purchases already used. V5 is now its own column on the purchases page and its own database field, separated out from the service history documents it used to be bundled with, ready for a future DealerKit document type mapping. - A "Sync now" button on the purchases page, running the same check the daily schedule does on demand. The daily purchases check itself is now genuinely scheduled for the first time on this install, at 3:45pm. ## v2.7.3, 2026-08-24 ### Fixed - Carwow's own payment breakdown was only reading correctly for the most recent purchases. A fully paid sale relabels "Vehicle price" to "Final vehicle price" (now matched too), and an old enough purchase drops the itemised breakdown entirely, only a plain "Sale agreed for £X" headline survives (now the fallback, price only, fees left honestly unknown rather than guessed). All 23 real current purchases now have a real price. ## v2.7.2, 2026-08-24 ### Fixed - Purchase page thumbnails were missing for any purchase that only ever had a photo gallery captured and no separate single photo (Carwow's own purchases, which don't set that second field the way Motorway's do). The thumbnail now falls back to the gallery's own first photo. ## v2.7.1, 2026-08-24 ### Fixed - Carwow's own won listings screen was only ever reading 3 of the real cards. Three separate causes: each card is a lazy loaded turbo-frame that only fetches its own content once scrolled into view (most were never triggered), the per card text was truncated at a length too short for some real cards, and an old suffix style plate format was one character too long for the reg pattern. Also widened the read to the page's own 90 day maximum, it was defaulting to 30. Verified live: now reads every real deal, matching the site's own total exactly. ## v2.7, 2026-08-24 ### Added - Carwow won purchases now capture everything Motorway's already did: a real itemised payment breakdown (vehicle price, Carwow's own buyer fee, transport fee, VAT), the full photo gallery, and the seller's service record and V5C logbook scans, downloaded from Carwow's own document bundle and pushed into DealerKit the same way Motorway's already are. A purchase still mid document review on Carwow's own side correctly captures nothing yet rather than guessing. ## v2.6.2, 2026-08-24 ### Fixed - DealerWay's own read window only ever captured auctions ending today or tomorrow, silently dropping real live auctions running further out (its lots run several days, the same as Auction4Cars, which had already been fixed away from this exact narrow window). Widened to 7 days with a proper inclusive range check. A live check found real cars within a normal search radius were being dropped before they ever reached the gate. ## v2.6.1, 2026-08-24 ### Added - A cockpit price ceiling on the settings page's Overview tab (Settings > Cockpit display). When set, it pre fills the cockpit's own Max bid and Reserve sidebar filters with the same figure on every page load, hiding cars priced or reserved above it by default. Display only, never a gate: pricing, gating and every count are untouched, and clearing either box, or clicking Reset, shows everything again for that session. ## v2.6, 2026-08-24 ### Added - A "Software update" card on the settings page's Integrations tab, showing the current and latest released version. When a newer one exists, an "Update now" button fetches tags, checks out the release, and restarts the server so it takes effect, the same as running the update commands and a `launchctl kickstart` by hand. Refuses while a run is in progress or the working tree has local changes, rather than risk a checkout over something real. - The daily job is genuinely scheduled for the first time on this install: two staggered launchd jobs matching Motorway's and Carwow's own different daily stock times, Motorway alone shortly after 4:30pm, everything else shortly after 5pm. ## v2.5, 2026-08-24 ### Added - DealerWay (dealerway.co.uk) added as a fifth auction platform, built against real captured pages: a real NAMA style condition grade and exact distance, a live per lot countdown, and three disclosed issue flags (test driven, bodywork, mechanical/electrical) read straight off each card. It discloses no reserve, so Current Bid, or Starting Price before any bid, stands in, the same workaround already used for Auction4Cars and Dealer Auction. A combined mechanical/electrical disclosure is a hard reject; a disclosed bodywork issue is a card flag, never a rejection. - The daily job is genuinely scheduled for the first time on this install: two staggered launchd jobs rather than one, matching Motorway's and Carwow's own different daily stock times. Motorway alone fires shortly after 4:30pm; everything else (Carwow, Auction4Cars, Dealer Auction, DealerWay) fires shortly after 5pm, deliberately excluding Motorway so it is never needlessly re read half an hour after it was already fresh. - Cazana joins the settings page's relogin mechanism (Log in to X button, live session age shown on the Integrations tab), previously the one valuation source only ever re logged in from the terminal. ### Changed - Each selected platform is now read, gated, valued and written to the cockpit in its own turn, rather than the whole run reading every platform first and only ever saving once at the very end. A finished platform's cars go live immediately rather than waiting on a slower one, and a later failure on one platform (a dead Cazana login mid Carwow, for example) no longer wipes out an earlier platform's own already saved results the way a single end of run write always used to. - A car's own detail page (owners, service history, disclosed issues, photos, and on Carwow/Auction4Cars the reserve itself) is now cached for 4 hours, keyed per car. Auction4Cars, Carwow and DealerWay auctions run several days, so the same car legitimately reappears run after run; a re seen car within the window skips its own detail page visit entirely rather than being re read from scratch every time. - Motorway's own CSV export already discloses a seller reported mechanical issue as a plain column, never read before now. A car flagged there skips its detail page visit outright (a known issue is a hard reject regardless of anything else the page might reveal), a small but genuinely free speed win on top of the cache above. - Bell notifications are now clickable straight through to whatever fixes them (a login notice opens the right settings tab, a "run this check" notice opens the Run modal), and the same red flag now also shows next to the affected site's own heading on the settings page itself, not just in the bell. ### Fixed - The per candidate detail page read for Carwow, Auction4Cars and DealerWay had no progress reporting at all, so a genuinely slow but working multi minute pass looked identical to a frozen browser on the cockpit's own progress bar. All three now report progress per car. ## v2.4.2, 2026-08-24 ### Added - A real notice on the cockpit (a red dot on Run now, and the bell dropdown) when a platform's own login has genuinely failed in the last run: Motorway, Carwow, Auction4Cars, Dealer Auction or LE Capital. Never a background check, only ever reports what a run that was already happening found, so it carries none of the extra session risk a timer based poll would (Glass's single account wide session has shown before that even just checking it can evict the real one). The notice clears itself the moment that site reads cleanly again. - Bell notifications are now clickable straight through to whatever fixes them: a login notice opens the settings page's Integrations tab directly, a "run this check" notice opens the Run modal. The same red flag now also shows next to the affected site's own heading on the Integrations tab itself, not just in the bell. ### Fixed - `settings.html` and `purchases.html` 404'd on any URL with a query string attached (for example the new deep link to a specific settings tab), since the server matched the address exactly rather than just the page name. ## v2.4.1, 2026-08-24 ### Added - The relogin button introduced in v2.4 now covers all seven sites, not just the three back office automations: Motorway, Carwow, Auction4Cars and Dealer Auction each get one too, in a new Auction house logins card on the settings page's Integrations tab. No on/off switch for these four, that already lives elsewhere. ### Changed - DealerKit's two automations (stock read, purchases push) now share one login block under a single DealerKit heading, instead of each showing its own duplicate login button. ## v2.4, 2026-08-23 to 2026-08-24 ### Added - LE Capital funding sync: reads LE Capital's own Current Stock and Stock History and adds or settles DealerKit's Finance > Funding records to match, LE Capital always treated as the source of truth. A value disagreement on a car live on both sides is only ever reported, never changed automatically. Runs from the cockpit's Run modal or `python3 daily_run.py --lecapital-funding`. - A Settings page Integrations tab: DealerKit stock read, DealerKit purchases push, Glass's on demand checks and LE Capital funding sync can each be switched on or off independently. Turning one off disables its Run modal button, refuses it if started directly, and (DealerKit stock only) skips it in the scheduled daily run too. - A relogin button under each integration, opening a real login window on the Mac and saving the session once confirmed, so a stale or missing login can be fixed from the settings page rather than a terminal command. - Delivery report PDFs (Motorway's own vehicle condition and delivery photos) pushed into DealerKit as a document for every reachable purchased car. - DealerKit's Tax Date corrected to each purchase's real Motorway collection date, the figure Xero actually uses for when payment was made. ### Fixed - All 95 Motorway sourced DealerKit purchase records audited for correct and complete purchase price, delivery fee, buyers premium and indemnity fee, service history documents and V5s; genuine gaps found and filled, including 16 older purchases recovered via a Motorway purchase search this project had not used before. - A car's full funding or expense picture was sometimes read from only the one expense record linked to its purchase, missing real fees filed under a separate supplier; now reads every expense on the vehicle. - V5 document scans saved as .PNG rather than .jpg were silently never matched (only .jpg was ever recognised), dropping real V5s that were actually available; fixed and backfilled across every tracked purchase. - Document filenames pushed to DealerKit were random rather than the source's own real filename, and a batch of several large documents could silently fail to save; both fixed. ## v2.3, 2026-08-21 to 2026-08-23 ### Added - Purchases can now be pushed straight into DealerKit as real Due In stock records: registration, purchase price, exterior photos, service history documents and V5C document scans, all from one on demand pass reachable from the cockpit's Run modal. - Every purchased car now has its service history document photos and its V5C document photos captured and stored, and the full listing photo gallery (not just the one card photo) is captured for both the cockpit's shortlist cards and the purchases page, viewable in a shared lightbox modal on either page. - Settings split into an Overview tab (the rules shared across every auction house) plus a tab per platform, with a reference only dealer postcode field on the Overview for documenting what should be set on each platform's own account. - A guide for bringing a second installation up to date with the shared repo, and an automated login for a session that dies unexpectedly using a credential saved in the macOS Passwords app. ### Fixed - Dealer Auction cards sometimes picked up a network badge icon instead of the real vehicle photo, and its own logo was missing a white variant so the "View on" button fell back to plain text. - DealerKit's own Add Expense screen fails silently, no error shown at all, if its required Team field is left blank; every expense entered through this project's own automation now sets it explicitly. ## v2.2, 2026-08-21 ### Added - Real versioned Releases on GitHub, tied to this changelog: each release's notes are pulled from its matching section here. - The full test suite now runs in CI on every push and pull request. - The purchases report's payment breakdown now recognises Motorway's "Estimated breakdown" state (before a car is actually paid for), not just the confirmed one, and refreshes until the real figure lands. - The purchases report gained a photo and a status column per row. - An "Ends within" filter (today, tomorrow, 3 to 7 days) for Auction4Cars and Dealer Auction, the two platforms whose auctions run over several days rather than closing at a single fixed time. - An "Ending soonest" sort next to the Cazana price sort; Auction4Cars and Dealer Auction default to it automatically when no other sort is chosen. - The live countdown on a lot now ticks by the second once under an hour, not just every 30 seconds at minute granularity. - A card disappears from view the moment its own auction ends, rather than sitting there greyed out with "Auction ended". ### Changed - The purchases report's "Room" column is now "Margin", calculated from the winning bid rather than the full amount paid. - Auction4Cars' reader now captures lots up to 7 days out (the longest a real auction there runs), not just today or tomorrow, so the new "Ends within" filter has real data to filter across. - New installer setup now points at the latest release tag, not the live `main` branch, which moves as work in progress. ### Fixed - A platform whose live read failed outright (a dead login, a timeout) was silently wiped from the shortlist instead of falling back to its own last known good cars; found live when Auction4Cars' session died mid session, now falls back with a loud message instead. - A test asserting a hardcoded UK wall clock time against a raw epoch timestamp only ever passed on a machine set to UK time; caught by the new CI run defaulting to UTC, fixed to compare against the same timezone-dependent conversion the code itself performs. ## v2.1, 2026-08-20 (evening) to 2026-08-21 ### Added - Dealer Auction added as a fourth platform (Auto Trader's trade auction network). Auction listings only, no API used, same as every other source. - A Run modal: refresh a single auction house at a time instead of always re reading all four, so a settings change to one platform doesn't force a full re read of the rest. The on demand Glass's and DMS stock buttons moved into the same modal. - A purchases admin page: every car actually bought, its real payment breakdown (winning bid, platform fees, transport, VAT), a retail estimate, and a Margin figure against the winning bid, each row carrying the car's own photo and its current status. - Saved, starrable filter views, and a live countdown to Auction4Cars lots' own end time (its auctions run 3, 5 or 7 days, not a single daily close like the other platforms). - A traffic light system (red, amber, green) with topic relevant icons for the card insight flags, collapsed to a hoverable icon row with a chevron to expand the full text. - A responsive shortlist grid, scaling from 2 columns on a laptop up to 6 on a 34 inch monitor, and a Reset link to clear every sidebar filter at once. ### Changed - Sort and the shortlist view filter merged into one dropdown per platform section, alongside the platform's own logo and live car count. - The "Best opportunity" sort now ranks every car on its Cazana value alone, so two cars priced from different valuation sources still compare fairly. - Carwow's stand in reserve figure is now labelled "Trade value", not "Reserve" (Carwow never actually discloses a real reserve, this figure is CAP Clean standing in for one). - The purchases report's "Room" column is now "Margin", calculated from the winning bid rather than the full amount paid, since fees, transport and VAT are cost of sale, not part of what the car itself made. ### Fixed - Motorway's own payment breakdown has three states, not two: a genuine "Estimated breakdown" while a car is bought but not yet paid for was being silently discarded because only the final confirmed state was recognised, hiding real figures Mark could see by hand. Estimated figures now keep refreshing on each read until the confirmed ones land, then lock for good. - Motorway purchases were showing the literal word "Delivered" as the vehicle name, and had no bought date at all, caused by an extra status badge on delivered cars' rows and a date parser that only understood relative phrasing ("3 days ago"), missing the absolute date format Motorway also uses. - Auction4Cars' reserve was being estimated from a lower guide value; it actually publishes a real reserve in an inline script on each car's own page, now read directly instead of estimated. - Make/Model filtering treated differently spelled makes from different platforms (for example "Mercedes" and "Mercedes-Benz") as separate makes, splitting one manufacturer's cars across several filter options. ## v2, 2026-08-19 to 2026-08-20 (early hours) ### Added - Multi dealership support: dealer specific settings (a platform's saved search or filter id, the accounts email) pulled out of tracked source into a local, gitignored `dealer_config.py`, so this repo can be shared across more than one dealership without one dealer's account settings leaking into or overwriting another's. - Auction4Cars added as a third platform, alongside Motorway and Carwow. - A cockpit settings page: the hard gate (mileage, reserve, age, distance, condition grade, owners), the banned makes, models and value caps, and later the pricing formula itself (retail uplift, flat spread), all editable without touching code, and settable separately per platform since not every source discloses the same details. - A DMS stock read (DealerKit) feeding the shortlist's "fills a gap / already have some" flags, replacing the previous dealer's own DMS reader for a dealership that uses a different system. - On demand refresh buttons and a notification bell surfacing platform health issues (a dead login, a stale stock read) instead of a permanent row of banners across the top of the page. - Reserve, max bid and retail guide range filters, plus a sort direction toggle for the shortlist. ### Changed - The retail valuation cap that forced a car's governing value down when only one valuation source read high was removed at the new dealer's request; a valuation source now prices at its own full calculated value with no ceiling. ### Fixed - A login detector was treating a platform's own sign in page as "logged in" because its URL happened not to contain the word the check looked for, capturing a dead session before any real login had happened. - A hardcoded postcode in a gate failure message that was wrong for any dealer other than the very first one this was built for. - A malformed saved filter URL that silently built a nonsense query parameter instead of a real one. - Real dealer values (a saved filter id, an accounts email, a DMS subdomain) that had leaked into this shared changelog's own prose, defeating the point of keeping that configuration local to each dealer. ## v1, original build The original build, for the first dealership this was written for, before this repository's git history began. Not represented as individual commits here since it predates this repo; the full decision log for this era, including every rule and edge case behind it, lives in `CLAUDE.md`. ### Added - The buying brain (`bidbrain/pricing.py`): the hard gate, the banned engine, make and model rules, and the core pricing rule (a retail uplift added to two valuation sources, a governing value threshold deciding which source counts, and a flat spread subtracted to get the recommended maximum bid). - Live readers for two auction platforms' dealer stock and two retail valuation sources, all read from logged in screens, no platform APIs. - The cockpit page: a daily shortlist with star and hide controls, valuation detail, a repeat appearance flag for cars seen before, and a reserve drop flag. - A DMS stock and sales report reader, feeding "fills a gap" flags and a monthly learning loop (days to sell, prep cost and margin, per model). - A private web address for the cockpit, reachable from any device. - Scheduled daily and monthly runs, unattended.