Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.
JavaScript 3d-generationai-agentsanimationavatardigital-humangenerative-aiglbgltfimage-to-3dllmmcpmcp-servermodel-context-protocolsolanatext-to-3dthreejsweb-componentswebglwebxrx402
docs(changelog): rebuild the public feeds so they match data/changelog.json
nirholas72e8e1186861
diff --git a/CHANGELOG.md b/CHANGELOG.mdindex e9f869cce..c32a15355 100644--- a/CHANGELOG.md+++ b/CHANGELOG.md@@ -27,16 +27,18 @@ Public history for [three.ws](https://three.ws), newest first. New pages come fr- **Every Crypto Converter result now has its own link you can share**: The converter promised a shareable result and never had one: the address bar stayed on /converter no matter what you picked, so there was no way to hand someone a rate. Now the page reads and writes its own URL. Pick BTC to EUR and type an amount and the address becomes /converter?from=bitcoin&to=EUR&amount=3, a new Copy link button puts that on your clipboard, and opening the link restores both sides and the amount exactly. Alongside that: the asset picker is fully keyboard-driven (arrow keys move the highlight, Enter picks, Escape closes and returns focus to the button you came from), the amount field says what it wants when you type something it cannot use instead of showing a silent dash, a coin whose price failed to load says price unavailable rather than a bare dash, the four quick presets light up when they match what you are looking at and stay disabled until live rates arrive, and the whole widget fits a 320px phone screen, where the currency buttons used to be pushed off the right edge and the picker off the left. (`/converter`) `[feature, improvement, fix]`- **My Collection loads visibly, keeps its counts, and never mangles an agent's name**: The collection page had a loading state nobody could see: the shimmer placeholders were drawn inside a container the page kept hidden until the data had already arrived, so a slow connection got a blank stretch of nothing instead. The tabs, the placeholders and the counts now appear from the first frame. The counts next to Skills and Subscriptions used to vanish at random, because the page's own translation pass could overwrite the tab label a split second after the number was written into it; the count now lives in its own slot and survives. A skill published by an agent whose name contains a quotation mark or an angle bracket rendered as broken markup rather than as the name, and a publisher who earned a vanity wallet address had it shown as an ordinary one, because the card was reading a field the collection endpoint does not return. Both are fixed, so tipping a creator from your collection now shows the wallet they actually have. Empty states are centred under the tabs instead of squeezed into the first column, both of them now point you at the marketplace, the Retry button on a failed load is a real button rather than a raw grey box, and the whole page is keyboard-navigable: arrow keys move between tabs, Home and End jump to the ends, and every link and button shows a clear focus ring. (`/collection`) `[fix, improvement]`- **Pausing a skill bundle is no longer permanent, and every tooltip fits on a phone**: The bundle builder at /bundles was audited end to end in a real browser: signing in, picking priced skills, the price panel and its backtest, publishing, the published list, and the loading, empty and error states, at phone, tablet and desktop widths. The worst finding was that Deactivate was a one-way door. It hid the bundle from buyers, which is correct, but it also removed it from the only page that could put it back, so a misclick permanently retired a bundle you had just priced. The button is now called Pause, a paused bundle stays on the page dimmed and clearly marked, and a Reactivate button puts it back on sale. Buyers still never see a paused bundle: the public list is unchanged, and paused ones are returned only to a signed-in owner of that agent. The small question-mark tooltips that explain every number on the page were also cut off at the edge of a phone screen, so four of the seven explanations were partly unreadable at 320 pixels wide; they now shift to stay on screen at any width. The sum of the parts and the suggested price were cramped into a broken-looking two-line squeeze on a narrow screen and now stack cleanly. Finally, a pricing error raised by one agent no longer sits on screen after you switch to another agent, where it read as a verdict on an agent it had never looked at. (`/bundles`) `[fix, improvement]`-- **Scene Composer works on a phone, starts you off with real models, and every control is reachable by keyboard**: Scene Composer was unusable on anything narrower than a laptop. Below 900px the scene panel was switched off, which pushed the 3D canvas into a zero-width column, so a tablet visitor got no viewport at all and no way back to the hierarchy, inspector or bone-attach controls; at phone widths the toolbar forced the whole editor wider than the screen and pushed the right-hand panel out of view. The panel is now a drawer behind a toolbar button, the toolbar scrolls instead of shoving the layout off screen, and the canvas keeps real dimensions at 320, 768 and 1440 pixels. The Recent Creations strip also no longer greets a first-time visitor with an empty box: until you have forged something it shows the live community showcase, so there are real models to drop into the scene straight away. The avatar picker was asking the directory a question it does not answer and got a feed of on-chain agents with no 3D model attached, so most slots were silently dropped and the rest never showed their thumbnails; it now asks for public 3D avatars and renders them. The start panel finally honours what it always promised: paste an avatar ID, not just a URL, and a model that cannot be loaded leaves a written explanation instead of a toast that vanishes before you can read it. Both the creations strip and the avatar picker now offer a working Try again when a request fails. On the keyboard side: the page has a real heading, the camera-view menu and every creation and avatar tile are proper buttons, the scene rows expose their hide and remove actions on focus, the prompt and avatar fields have labels, dialogs trap focus and hand it back on Escape, and there is a visible focus ring throughout. Removing an object now frees its GPU memory once the step drops out of undo range, undo puts a bone-attached item back on its bone, and saving an outfit refuses a locally dropped file rather than storing a link that breaks on reload. (`/compose`) `[fix, improvement]`+- **Scene Composer works on a phone, starts you off with real models, and answers the keyboard**: Scene Composer was unusable on anything narrower than a laptop. Below 900px the scene panel was switched off, which pushed the 3D canvas into a zero-width column, so a tablet visitor got no viewport at all and no way back to the hierarchy, inspector or bone-attach controls; at phone widths the toolbar forced the whole editor wider than the screen and shoved the right-hand panel out of view. The panel is now a drawer behind a toolbar button, the toolbar scrolls instead of stretching the layout, and the canvas keeps real dimensions at 320, 768 and 1440 pixels. The Recent Creations strip also no longer greets a first-time visitor with an empty box: until you have forged something it shows the live community showcase, so there are real models to drop into the scene straight away. The avatar picker was asking the directory a question it does not answer and got back a feed of on-chain agents with no 3D model attached, so most slots were silently dropped and the rest never showed their thumbnails; it now asks for public 3D avatars and renders them. The start panel finally honours what it always promised: paste an avatar ID, not just a URL, and a model that cannot be loaded leaves a written explanation instead of a toast that vanishes before you can read it. Both the creations strip and the avatar picker now offer a Try again that works when a request fails. On the keyboard side: the page has a real heading, the camera-view menu and every creation and avatar tile are proper buttons, the scene rows expose their hide and remove actions on focus, the prompt and avatar fields have labels, dialogs trap focus and hand it back on Escape, and there is a visible focus ring throughout. Removing an object now frees its GPU memory once the step drops out of undo range, undo puts a bone-attached item back on its bone, and saving an outfit refuses a locally dropped file rather than storing a link that breaks on reload. (`/compose`) `[fix, improvement]`+- **Self-hosted NIM says up front that no endpoint is wired, and a finished reconstruction actually appears on screen**: The /forge-nim page drives a TRELLIS NIM you host yourself, and it was driven end to end in a real browser. It already asked the API on load whether a NIM was configured, then threw the answer away: there was nowhere on the page to show it. So a first-time visitor dropped a photo, waited, and got a failure at the end of the trip, and the failure read as the raw code nim_unconfigured because the page was reading the machine code instead of the sentence next to it. The page now carries a status pill that says whether a NIM is wired and reachable, and when none is, the stage explains that this page needs a NIM you host, offers a button that opens the endpoint field, and links to the hosted forge and the NVIDIA model docs in the meantime; the Reconstruct button stays switched off rather than promising a trip that cannot arrive. Typing your own endpoint now re-probes that box and reports what it found, and an address the server refuses comes back with the reason instead of a silent failure later. The bigger fix was invisible until a reconstruction finished: the 3D viewer had collapsed to zero height, so the model was never even loaded and a completed job showed an empty stage under its own result bar. The viewer now fills the stage. The button also never showed its loading spinner, because translating the page had quietly replaced the label it was written into. The page carries a real heading, the loading and error states announce themselves to screen readers, every control has a visible keyboard focus ring, and it holds up at phone width. (`/forge-nim`) `[fix, improvement]`- **Selfie to Avatar tells you your library is full before it spends ninety seconds building**: Five fixes to Selfie to Avatar. If your avatar library was already at the limit for your plan, nothing noticed until the very end: you waited out the full reconstruction, and then the page said the 3D engine had hit a snag and told you to try again. It was not an engine problem, trying again could never work, and every attempt spent another ninety seconds of generation to land in the same place. Your library is now checked the moment you press the button, so a full one is refused immediately with the actual reason and what to do about it, and a build that fills the last slot while it is running ends with that same plain explanation instead of the engine message. The page also stopped losing its own error messages: the Build button floats above the page, so tapping it from higher up left the explanation somewhere below the fold while the button quietly reset itself, which read as nothing having happened. Errors now scroll into view. On the finished-avatar screen, Learn how to embed pointed at a page that did not exist and answered 404; it now opens the embedding guide. And on a narrow phone the Photoreal and Stylized choices were squeezed into two columns barely wide enough for one word per line, so they now stack and can be read. (`/create/selfie`) `[fix, improvement]`- **Talking Avatar Video stops handing the renderer a 3D file where a photo belongs, and says up front when the GPU lane is down**: The /create/video page turns one of your avatars plus an audio clip into a lip-synced video, and it was driven end to end in a real browser: every state, every button, every width. The worst finding was invisible from the page. The reference frame the GPU worker renders from was being taken from whatever the browser had handy, which for an avatar with no thumbnail meant its 3D model file. The worker saved that as the reference photo and then failed on it minutes later, after the wait had already been spent. The page now sends only the avatar, and the server resolves a real image for it: the stored thumbnail, or a fresh portrait render when there is none. Second, the GPU renderer is a separate service, and when it is not running the only signal was a failure after you had already picked a file and uploaded it. The page now asks first and, if the lane is down, says so at the top, keeps Generate switched off, and points at the Avatar Studio and prompt-to-avatar in the meantime. The rest is the ordinary polish an audit turns up: a failed avatar list used to look identical to owning no avatars, so it sent people off to create one they already had, and now it says what broke and offers a retry; having no avatars yet gets a real empty state with a link to make one; the Generate button always explains what it is still waiting for and, once audio is in, how many clips of video that audio is worth; the clip you picked can be played back with its length and size before you spend a generation on it; a render job that goes unreachable now says so after thirty seconds instead of spinning for twenty minutes; and after a spent free trial the button under the error no longer stayed stuck on Upgrade plan for every later error. Avatar thumbnails are real buttons with visible focus rings, the two form fields have proper labels, and the second animated avatar that the site companion used to drop onto this page no longer loads over the one you are working on. (`/create/video`) `[fix, improvement]`+- **Text to 3D gets its footer back, a retry button that knows when to open, and an honest name for the model that refined your prompt**: The /forge-spark page was driven end to end in a real browser. The closing note under the workspace was written as a page footer, and the site's old corner-credit styling for the 3D viewer pages had quietly claimed it: it was pulled out of the layout into a floating monospace strip that sat on top of the viewer on a desktop, and on a 320px phone it landed across the composer with its link to the full studio pushed off the edge of the screen entirely. That note is now ordinary page text, centered under the workspace where it belongs, and the real three.ws footer, with its full set of links, is on the page for the first time. When a generation fails there is now a Try again button inside the error itself, so recovery does not mean scrolling back to the top; when the generator answers that it is busy, the button counts down and opens exactly when a retry can actually get through. A stage that fails is marked in red instead of silently dropping back to grey, the wait now shows real elapsed seconds so a cold reconstruction lane no longer reads as a hang, and the first step is renamed to whatever model actually answered, because the prompt refiner is a chain of engines and Nemotron is only the first one it tries. (`/forge-spark`) `[fix, improvement]`- **The avatar SDK bundle we ship is now built from the code we deploy**: The <agent-3d> runtime that sites load from three.ws is a single bundled file, and the production build never rebuilt it. It was copied from whatever the build machine happened to have on disk, which could be days older than the code being deployed, so a fix to the avatar element could ship everywhere else on the site and quietly not reach the SDK. The deploy now rebuilds that bundle from the same commit as the rest of the release, every time. Nothing to change on your side: existing embeds keep working and simply pick up the current runtime. (`/avatar-sdk`) `[fix, sdk, infra]`- **The Character Library counts its own characters, and fills the screen while it does**: The library page had been advertising 106 rigged characters since launch. There are 107 on the shelf today, and there will be more tomorrow, so the number is no longer typed into the headline, the page description or the share card: the page asks the library how many it has and prints the answer under the hero. The grid also stopped wasting a third of a desktop screen. A layout rule left the page shrink-to-fitting to about two thirds of the window, so a wide monitor showed three columns where five fit; it now uses the full width it was designed for. Every character card previously carried a hidden second tab stop that announced the same name as the card itself, which made keyboard browsing twice as long as it needed to be, and thumbnails could not be dragged out of alignment by an accidental swipe anymore because the preview no longer pretends to be an interactive 3D control. The search box has a visible focus ring again instead of a one pixel border shift, and when the library cannot be reached the page now says what actually went wrong rather than reverting to a generic sentence a moment later. (`/character-library`) `[fix, improvement]`- **The Characters gallery pages past the first screen, and stops fighting dark mode**: The character gallery at /characters was audited end to end in a real browser: loading the feed, both sort orders, search, Load more, following a card through to its character page, and the loading, empty and error states, at phone, tablet and desktop widths. The worst finding was that the Top sort was quietly broken. Sorting by most-chatted read the right first screen, but pressing Load more asked the server to continue from the wrong place, so the feed ended after eight characters out of more than two thousand. It now pages all the way through, in order. The page also ignored the site theme: it painted a fixed light-blue background under the dark header no matter what you had chosen, which is now fixed in both directions, and wallet addresses on the cards were nearly invisible on a light background everywhere they appear on the site. The gallery filled only a narrow strip down the middle of a desktop screen and now uses the full width, up to four cards across. If the feed failed to load you got one line of grey text and no way to retry; there is now a proper message with a Try again button, a failed Load more keeps the characters you already had, and searching for something with no matches offers to clear the search rather than leaving you at a dead end. Pressing Top while the first screen was still loading used to light up the button and change nothing, so you were left reading New results under a Top heading. The search box had no label for screen readers, the sort buttons never showed which was selected, and neither they nor Load more showed a focus outline for anyone navigating by keyboard. Keyboard users can also now skip the header and land in the gallery. Finally, the feed itself is several times faster to answer on a cold request, because it no longer counts every conversation of every published character just to show you twenty-four of them. (`/characters`) `[fix, improvement]`- **The Community page's live activity feed renders as readable rows again**: The activity list on the Community page was building each row as a link wrapped around another link, which no browser can parse: every entry was torn into three stray fragments, so the creator, the item and its thumbnail landed on separate lines with gaps between them and the whole list was unreadable. Rows are now single, valid entries again: click anywhere on one to open the creation, or click the creator's name to open their profile. If the feed cannot load, or returns something the page cannot read, you get a clear message and a Try again button instead of skeleton placeholders that never resolve. The page also follows the light and dark theme switch now instead of staying black on a light canvas, and the newsletter section finally has the signup field its text promises, rather than sending you hunting for the one in the footer. (`/community`) `[fix, improvement]`- **The Concierge page stops cutting itself off on a phone, and its install tabs work from the keyboard**: Three fixes to the Concierge landing page. On a narrow phone screen the whole page was being sized against the longest line of the install snippet, which pushed the column roughly twice as wide as the screen; because the site clips sideways overflow, the right-hand half of every headline, paragraph and button was simply cut off with no way to scroll to it. The column now fits the screen and the code snippet scrolls inside its own box, the way it already did on a laptop. The four install tabs (one tag, web component, npm, imperative) announced themselves to a screen reader as tabs but were not wired to the panels they open and ignored the arrow keys; they now move with Left, Right, Home and End, and each tab is properly tied to its snippet. Copy buttons also say which snippet they copy, and if the browser blocks the clipboard the snippet is selected for you so Ctrl+C still works. Finally, the passive "Have you tried" suggestion card could land on top of the live concierge demo in the corner; it now stacks clear of it. (`/concierge`) `[fix, improvement]`- **The Create page no longer black-screens when the sign-in check stalls, and a broken remix link no longer kills every button on it**: Two ways the front door for creating could strand you are gone. The page opens behind a full-screen loading screen while it checks whether you are signed in; if that check got no answer at all (a stalled connection rather than a refused one) the screen stayed black forever and the page read as dead. It now hands the page back after a short wait and treats you as signed out, which costs at most one extra trip through sign-in. Separately, opening a remix link for an avatar that had been deleted or made private used to show a message for four seconds and then leave you on a page where not one of the eight creation cards did anything: the remix ran before the buttons were wired and gave up without wiring them. The cards are live first now, and a failed remix explains itself in a panel that stays put, with a button to try again and a button to move on. Also: the talking-avatar-video card now says "Coming soon" on the card itself instead of only dimming, answers the keyboard the same way it answers a mouse, and stops calling itself unavailable when the check simply could not reach us (it asks again when you click). A saved avatar whose thumbnail has been deleted falls back to a silhouette instead of a broken image. (`/create`) `[fix, improvement]`-- **The Pole Club audio controls no longer fall off the edge of a small phone**: On a narrow phone the club's top bar ran out of room and simply cut off whatever did not fit, and the page cannot scroll sideways, so the mute button and the 8D audio toggle were gone with no way to reach them. The bar now gives up the parts that are only there to read (the tagline and the price and rail labels, all repeated in the panel's How it works section) and keeps both controls at full size at every width, down to a 320px screen. (`/club`) `[fix]`+- **The Pole Club audio controls no longer fall off the edge of a small phone**: On a narrow phone the club's top bar ran out of room and simply cut off whatever did not fit, and the page cannot scroll sideways, so the mute button and the 8D audio toggle were gone with no way to reach them. The bar now gives up the parts that are only there to read (the tagline and the price and rail labels, all of which the panel's How it works section repeats) and keeps both controls at full size at every width, down to a 320px screen. (`/club`) `[fix]`- **The Pole Club leaderboard now ranks the dancers who are actually on the poles**: The club's leaderboard was reading names out of the payout registry instead of off the stage, so it listed four dancers under names matching none of the three at the poles, every one of them sitting at zero. It now names each dancer exactly as her pole card does, drops entries for dancers who are not in tonight's lineup and have never been tipped, and says what a quiet board actually means: the 24h and 1h views tell you the club has been quiet rather than claiming nobody has ever tipped, which was never true while the all-time board was full. Keyboard and screen-reader visitors get the rest of the pass: the page carries one real heading instead of five competing ones, Skip to stage now lands you on the stage, and the controls belonging to the walk-in (the agent picker, the Enter the club prompt) no longer sit invisibly in the tab order once you are inside the room. (`/club`) `[fix, improvement]`- **The security agent in the world no longer opens its scanner in your face**: zauth, the security agent standing in the town plaza, used to throw its RepoScan panel over the screen the moment you walked within a few metres of it, interrupting whatever you were doing and hiding the world behind a modal you never asked for. Walking up now only gets you the agent's line and the usual E prompt above its head. The scanner opens when you choose to open it: press E, click the agent, or click its nameplate. Nothing else changed about the scan itself, and the price is still fetched live from zauth in the background as you approach, so the panel shows the real cost the instant it opens instead of a loading shimmer. (`/play`) `[fix, improvement]`- **The viewer's backdrop picker stops vanishing while you read it, and sharing no longer posts "Meet Agent"**: Five fixes to the 3D model viewer at /app. Opening the Backdrop picker and taking a few seconds to read the six choices used to make the picker fade out from under the cursor: the viewer hides its controls after five seconds of stillness, and the picker was not on the list of surfaces that keep them awake, so an open picker went invisible and unclickable while still counting as open. It now stays put until you choose a backdrop or close it, and the controls resume fading once it is shut. Typing in the animation search while the clips were still loading stacked a second, contradictory message under the first, so the panel read "Loading clips" and "No clips matching" at the same time, and clearing the box left behind the nonsense line "No clips matching" with nothing in the quotes; the search now stays quiet until there are clips to search. Sharing the viewer to X, Farcaster or Telegram published "Meet Agent", an internal placeholder name, for anyone who had not loaded an agent of their own, and now uses the generic line instead. The page also shipped a sign-in panel and a welcome card that its own stylesheet blanked unconditionally, four links nobody could ever reach, and the preview card shown when the page is shared described the platform instead of the viewer. The avatar gallery behind the Avatar button is now announced to screen readers as a named dialog. (`/app`) `[fix, improvement]`diff --git a/public/changelog.json b/public/changelog.jsonindex d92b5e496..6dada7d70 100644--- a/public/changelog.json+++ b/public/changelog.json@@ -1,5 +1,5 @@{- "generated_at": "2026-08-17T05:04:17.067Z",+ "generated_at": "2026-08-17T05:22:51.398Z","generated_by": "scripts/build-page-index.mjs from data/pages.json + data/changelog.json","site": {"name": "three.ws",@@ -246,6 +246,17 @@"improvement"]},+ {+ "date": "2026-08-17",+ "type": "update",+ "title": "Self-hosted NIM says up front that no endpoint is wired, and a finished reconstruction actually appears on screen",+ "summary": "The /forge-nim page drives a TRELLIS NIM you host yourself, and it was driven end to end in a real browser. It already asked the API on load whether a NIM was configured, then threw the answer away: there was nowhere on the page to show it. So a first-time visitor dropped a photo, waited, and got a failure at the end of the trip, and the failure read as the raw code nim_unconfigured because the page was reading the machine code instead of the sentence next to it. The page now carries a status pill that says whether a NIM is wired and reachable, and when none is, the stage explains that this page needs a NIM you host, offers a button that opens the endpoint field, and links to the hosted forge and the NVIDIA model docs in the meantime; the Reconstruct button stays switched off rather than promising a trip that cannot arrive. Typing your own endpoint now re-probes that box and reports what it found, and an address the server refuses comes back with the reason instead of a silent failure later. The bigger fix was invisible until a reconstruction finished: the 3D viewer had collapsed to zero height, so the model was never even loaded and a completed job showed an empty stage under its own result bar. The viewer now fills the stage. The button also never showed its loading spinner, because translating the page had quietly replaced the label it was written into. The page carries a real heading, the loading and error states announce themselves to screen readers, every control has a visible keyboard focus ring, and it holds up at phone width.",+ "link": "/forge-nim",+ "tags": [+ "fix",+ "improvement"+ ]+ },{"date": "2026-08-17","type": "update",@@ -268,6 +279,17 @@"improvement"]},+ {+ "date": "2026-08-17",+ "type": "update",+ "title": "Text to 3D gets its footer back, a retry button that knows when to open, and an honest name for the model that refined your prompt",+ "summary": "The /forge-spark page was driven end to end in a real browser. The closing note under the workspace was written as a page footer, and the site's old corner-credit styling for the 3D viewer pages had quietly claimed it: it was pulled out of the layout into a floating monospace strip that sat on top of the viewer on a desktop, and on a 320px phone it landed across the composer with its link to the full studio pushed off the edge of the screen entirely. That note is now ordinary page text, centered under the workspace where it belongs, and the real three.ws footer, with its full set of links, is on the page for the first time. When a generation fails there is now a Try again button inside the error itself, so recovery does not mean scrolling back to the top; when the generator answers that it is busy, the button counts down and opens exactly when a retry can actually get through. A stage that fails is marked in red instead of silently dropping back to grey, the wait now shows real elapsed seconds so a cold reconstruction lane no longer reads as a hang, and the first step is renamed to whatever model actually answered, because the prompt refiner is a chain of engines and Nemotron is only the first one it tries.",+ "link": "/forge-spark",+ "tags": [+ "fix",+ "improvement"+ ]+ },{"date": "2026-08-17","type": "update",diff --git a/public/changelog.xml b/public/changelog.xmlindex 6d66b8a9e..760a47df9 100644--- a/public/changelog.xml+++ b/public/changelog.xml@@ -175,12 +175,20 @@<description>The bundle builder at /bundles was audited end to end in a real browser: signing in, picking priced skills, the price panel and its backtest, publishing, the published list, and the loading, empty and error states, at phone, tablet and desktop widths. The worst finding was that Deactivate was a one-way door. It hid the bundle from buyers, which is correct, but it also removed it from the only page that could put it back, so a misclick permanently retired a bundle you had just priced. The button is now called Pause, a paused bundle stays on the page dimmed and clearly marked, and a Reactivate button puts it back on sale. Buyers still never see a paused bundle: the public list is unchanged, and paused ones are returned only to a signed-in owner of that agent. The small question-mark tooltips that explain every number on the page were also cut off at the edge of a phone screen, so four of the seven explanations were partly unreadable at 320 pixels wide; they now shift to stay on screen at any width. The sum of the parts and the suggested price were cramped into a broken-looking two-line squeeze on a narrow screen and now stack cleanly. Finally, a pricing error raised by one agent no longer sits on screen after you switch to another agent, where it read as a verdict on an agent it had never looked at.</description></item><item>- <title>Scene Composer works on a phone, starts you off with real models, and every control is reachable by keyboard</title>+ <title>Scene Composer works on a phone, starts you off with real models, and answers the keyboard</title><link>https://three.ws/compose</link>- <guid isPermaLink="false">2026-08-17:Scene Composer works on a phone, starts you off with real models, and every control is reachable by keyboard</guid>+ <guid isPermaLink="false">2026-08-17:Scene Composer works on a phone, starts you off with real models, and answers the keyboard</guid><pubDate>Mon, 17 Aug 2026 12:00:00 GMT</pubDate><category>fix</category>- <description>Scene Composer was unusable on anything narrower than a laptop. Below 900px the scene panel was switched off, which pushed the 3D canvas into a zero-width column, so a tablet visitor got no viewport at all and no way back to the hierarchy, inspector or bone-attach controls; at phone widths the toolbar forced the whole editor wider than the screen and pushed the right-hand panel out of view. The panel is now a drawer behind a toolbar button, the toolbar scrolls instead of shoving the layout off screen, and the canvas keeps real dimensions at 320, 768 and 1440 pixels. The Recent Creations strip also no longer greets a first-time visitor with an empty box: until you have forged something it shows the live community showcase, so there are real models to drop into the scene straight away. The avatar picker was asking the directory a question it does not answer and got a feed of on-chain agents with no 3D model attached, so most slots were silently dropped and the rest never showed their thumbnails; it now asks for public 3D avatars and renders them. The start panel finally honours what it always promised: paste an avatar ID, not just a URL, and a model that cannot be loaded leaves a written explanation instead of a toast that vanishes before you can read it. Both the creations strip and the avatar picker now offer a working Try again when a request fails. On the keyboard side: the page has a real heading, the camera-view menu and every creation and avatar tile are proper buttons, the scene rows expose their hide and remove actions on focus, the prompt and avatar fields have labels, dialogs trap focus and hand it back on Escape, and there is a visible focus ring throughout. Removing an object now frees its GPU memory once the step drops out of undo range, undo puts a bone-attached item back on its bone, and saving an outfit refuses a locally dropped file rather than storing a link that breaks on reload.</description>+ <description>Scene Composer was unusable on anything narrower than a laptop. Below 900px the scene panel was switched off, which pushed the 3D canvas into a zero-width column, so a tablet visitor got no viewport at all and no way back to the hierarchy, inspector or bone-attach controls; at phone widths the toolbar forced the whole editor wider than the screen and shoved the right-hand panel out of view. The panel is now a drawer behind a toolbar button, the toolbar scrolls instead of stretching the layout, and the canvas keeps real dimensions at 320, 768 and 1440 pixels. The Recent Creations strip also no longer greets a first-time visitor with an empty box: until you have forged something it shows the live community showcase, so there are real models to drop into the scene straight away. The avatar picker was asking the directory a question it does not answer and got back a feed of on-chain agents with no 3D model attached, so most slots were silently dropped and the rest never showed their thumbnails; it now asks for public 3D avatars and renders them. The start panel finally honours what it always promised: paste an avatar ID, not just a URL, and a model that cannot be loaded leaves a written explanation instead of a toast that vanishes before you can read it. Both the creations strip and the avatar picker now offer a Try again that works when a request fails. On the keyboard side: the page has a real heading, the camera-view menu and every creation and avatar tile are proper buttons, the scene rows expose their hide and remove actions on focus, the prompt and avatar fields have labels, dialogs trap focus and hand it back on Escape, and there is a visible focus ring throughout. Removing an object now frees its GPU memory once the step drops out of undo range, undo puts a bone-attached item back on its bone, and saving an outfit refuses a locally dropped file rather than storing a link that breaks on reload.</description>+ </item>+ <item>+ <title>Self-hosted NIM says up front that no endpoint is wired, and a finished reconstruction actually appears on screen</title>+ <link>https://three.ws/forge-nim</link>+ <guid isPermaLink="false">2026-08-17:Self-hosted NIM says up front that no endpoint is wired, and a finished reconstruction actually appears on screen</guid>+ <pubDate>Mon, 17 Aug 2026 12:00:00 GMT</pubDate>+ <category>fix</category>+ <description>The /forge-nim page drives a TRELLIS NIM you host yourself, and it was driven end to end in a real browser. It already asked the API on load whether a NIM was configured, then threw the answer away: there was nowhere on the page to show it. So a first-time visitor dropped a photo, waited, and got a failure at the end of the trip, and the failure read as the raw code nim_unconfigured because the page was reading the machine code instead of the sentence next to it. The page now carries a status pill that says whether a NIM is wired and reachable, and when none is, the stage explains that this page needs a NIM you host, offers a button that opens the endpoint field, and links to the hosted forge and the NVIDIA model docs in the meantime; the Reconstruct button stays switched off rather than promising a trip that cannot arrive. Typing your own endpoint now re-probes that box and reports what it found, and an address the server refuses comes back with the reason instead of a silent failure later. The bigger fix was invisible until a reconstruction finished: the 3D viewer had collapsed to zero height, so the model was never even loaded and a completed job showed an empty stage under its own result bar. The viewer now fills the stage. The button also never showed its loading spinner, because translating the page had quietly replaced the label it was written into. The page carries a real heading, the loading and error states announce themselves to screen readers, every control has a visible keyboard focus ring, and it holds up at phone width.</description></item><item><title>Selfie to Avatar tells you your library is full before it spends ninety seconds building</title>@@ -198,6 +206,14 @@<category>fix</category><description>The /create/video page turns one of your avatars plus an audio clip into a lip-synced video, and it was driven end to end in a real browser: every state, every button, every width. The worst finding was invisible from the page. The reference frame the GPU worker renders from was being taken from whatever the browser had handy, which for an avatar with no thumbnail meant its 3D model file. The worker saved that as the reference photo and then failed on it minutes later, after the wait had already been spent. The page now sends only the avatar, and the server resolves a real image for it: the stored thumbnail, or a fresh portrait render when there is none. Second, the GPU renderer is a separate service, and when it is not running the only signal was a failure after you had already picked a file and uploaded it. The page now asks first and, if the lane is down, says so at the top, keeps Generate switched off, and points at the Avatar Studio and prompt-to-avatar in the meantime. The rest is the ordinary polish an audit turns up: a failed avatar list used to look identical to owning no avatars, so it sent people off to create one they already had, and now it says what broke and offers a retry; having no avatars yet gets a real empty state with a link to make one; the Generate button always explains what it is still waiting for and, once audio is in, how many clips of video that audio is worth; the clip you picked can be played back with its length and size before you spend a generation on it; a render job that goes unreachable now says so after thirty seconds instead of spinning for twenty minutes; and after a spent free trial the button under the error no longer stayed stuck on Upgrade plan for every later error. Avatar thumbnails are real buttons with visible focus rings, the two form fields have proper labels, and the second animated avatar that the site companion used to drop onto this page no longer loads over the one you are working on.</description></item>+ <item>+ <title>Text to 3D gets its footer back, a retry button that knows when to open, and an honest name for the model that refined your prompt</title>+ <link>https://three.ws/forge-spark</link>+ <guid isPermaLink="false">2026-08-17:Text to 3D gets its footer back, a retry button that knows when to open, and an honest name for the model that refined your prompt</guid>+ <pubDate>Mon, 17 Aug 2026 12:00:00 GMT</pubDate>+ <category>fix</category>+ <description>The /forge-spark page was driven end to end in a real browser. The closing note under the workspace was written as a page footer, and the site's old corner-credit styling for the 3D viewer pages had quietly claimed it: it was pulled out of the layout into a floating monospace strip that sat on top of the viewer on a desktop, and on a 320px phone it landed across the composer with its link to the full studio pushed off the edge of the screen entirely. That note is now ordinary page text, centered under the workspace where it belongs, and the real three.ws footer, with its full set of links, is on the page for the first time. When a generation fails there is now a Try again button inside the error itself, so recovery does not mean scrolling back to the top; when the generator answers that it is busy, the button counts down and opens exactly when a retry can actually get through. A stage that fails is marked in red instead of silently dropping back to grey, the wait now shows real elapsed seconds so a cold reconstruction lane no longer reads as a hang, and the first step is renamed to whatever model actually answered, because the prompt refiner is a chain of engines and Nemotron is only the first one it tries.</description>+ </item><item><title>The avatar SDK bundle we ship is now built from the code we deploy</title><link>https://three.ws/avatar-sdk</link>@@ -252,7 +268,7 @@<guid isPermaLink="false">2026-08-17:The Pole Club audio controls no longer fall off the edge of a small phone</guid><pubDate>Mon, 17 Aug 2026 12:00:00 GMT</pubDate><category>fix</category>- <description>On a narrow phone the club's top bar ran out of room and simply cut off whatever did not fit, and the page cannot scroll sideways, so the mute button and the 8D audio toggle were gone with no way to reach them. The bar now gives up the parts that are only there to read (the tagline and the price and rail labels, all repeated in the panel's How it works section) and keeps both controls at full size at every width, down to a 320px screen.</description>+ <description>On a narrow phone the club's top bar ran out of room and simply cut off whatever did not fit, and the page cannot scroll sideways, so the mute button and the 8D audio toggle were gone with no way to reach them. The bar now gives up the parts that are only there to read (the tagline and the price and rail labels, all of which the panel's How it works section repeats) and keeps both controls at full size at every width, down to a 320px screen.</description></item><item><title>The Pole Club leaderboard now ranks the dancers who are actually on the poles</title>diff --git a/public/features.json b/public/features.jsonindex a8a22e668..fc488fb95 100644--- a/public/features.json+++ b/public/features.json@@ -1,5 +1,5 @@{- "generated_at": "2026-08-17T05:04:16.973Z",+ "generated_at": "2026-08-17T05:22:51.107Z","generated_by": "scripts/build-page-index.mjs from data/pages.json","site": {"name": "three.ws",