Skip to content

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.

fix(agent-edit): land an autopilot receipt's provenance link on the memory it cites

nirholasfe18c8a812d1
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 7f56d5655..76b5f11b0 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -17,9 +17,13 @@ Public history for [three.ws](https://three.ws), newest first. New pages come fr
- **Alpha Co-pilot counts its live launches again, and a mistyped agent ID no longer leaves the last one on stage**: Four things were wrong on the page where your agent reads a launch out loud. The badge next to Live launches, which tells you how many launches are waiting, was always blank: the counter was written inside the heading text, and translating the heading replaced the very element the page updates, so the number was thrown away every time. It now sits beside the heading and survives a switch into any of the 85 languages. Loading an agent that does not exist used to leave the previous agent standing on stage with its launch cards still on screen, and every Ask for a read button on those cards did nothing at all when clicked; the stage is now cleared back to its starting state with a prompt to pick another co-pilot. The buy button in the confirm drawer showed the amount with its unit twice over. And on a local or preview copy of the site the avatar's 3D body failed to load at all, because this page was the only one that did not route model files through the same asset path every other page uses. (`/alpha-copilot`) `[fix]`
- **An expired X connection now offers to reconnect instead of sending you to the login page**: Posting a Walk screenshot or clip to X goes through your connected X account. When that connection could no longer be renewed, the share endpoint answered with the same code it uses for a signed-out visitor, so the capture panel told an already signed-in user to sign in to three.ws and linked them to the login page, which fixed nothing. That case now comes back as its own answer with the X connect link attached, and the panel offers a Reconnect X button that takes you straight to the flow that repairs it. The share caption also lost a stray dash. (`/walk`) `[fix]`
- **An unsubscribe now sticks, even if you click an old confirm link**: Every three.ws newsletter email carries a confirm link and an unsubscribe link that share one token, so a reader who unsubscribed and later opened an older email and clicked confirm was quietly put back on the list. That link no longer signs anyone up: it says you are unsubscribed and leaves it that way, and opting back in starts from the site. Three other fixes ship alongside it. The allocation planner at /meta-allocator no longer serves an empty basket when the model declines to answer, it falls back to the allocator that always produces a real spread across verified leaders. The motion capture uploader told you an upload link lasted ten minutes when it really lasted five, so a large video queued on that promise could fail with nothing explaining why. And an agent whose stablecoin payment is rejected by the x402 Bazaar now gets the Bazaar's own details back instead of a different server's, so its retry is aimed at the right place. (`/changelog`) `[fix]`
+- **AR Studio: the language picker no longer sits on top of the dock controls**: On the AR Studio page the site-wide floating language picker landed in the bottom-right corner, directly over the right-hand end of the control dock, so the Clear button (and the buttons beside it on a phone) could not be tapped at all. The picker now lives in the studio's own top bar, where it is always reachable and leaves the dock its full width. It also shrinks instead of pushing its neighbours off screen on narrow phones, which fixes the same crowding anywhere else it appears in a tight header. The page description now also names every place models can come from: your forges, the community feed, the free CC0 object library, and any GLB link. (`/ar/studio`) `[fix, improvement]`
+- **Autopilot Activity keeps up when you switch agents, and its receipts link to the exact memory**: Autopilot Activity lists every action your agents took on your behalf, and it has an agent picker for narrowing that list to one of them. Picking a second agent while the first was still loading used to do nothing at all: the picker and the address bar moved to the new agent, but the request for it was never sent and the page kept showing the previous agent's actions, with no sign anything had gone wrong. Whichever agent you picked last now wins, however fast you switch. Two more repairs: each receipt cites the memories that motivated the action, and those links now scroll to the cited memory and highlight it instead of dropping you at the top of a long Knowledge list; and when an agent has no actions yet, the page sends you to that agent's own autopilot controls rather than the dashboard, which is where the setting you need actually lives. (`/autopilot-activity`) `[fix, improvement]`
- **Avatar link previews stop rebuilding themselves, and two feeds answer a bad page link plainly**: An audit of the chat, media and board endpoints found three real breaks. When you share an avatar link, the preview image is drawn once and then kept, so every later share is instant. A mix-up between the stored image and its address meant every freshly drawn preview was mistaken for an old broken one: the saved copy was thrown away and the picture was redrawn from scratch on every single share. The preview is now kept, as it always should have been, and genuinely broken old records are repaired the next time they are drawn. Separately, the character directory answered a hand-edited or stale "next page" link with a server error instead of saying the value was not a timestamp, and the bounty board did the same when the page size in the address was not a number. Both now answer plainly, and the board simply falls back to its normal page size. (`/characters`) `[fix]`
- **Avatar Studio trait library: the mirror can no longer be pointed at someone else's server**: Avatar Studio loads its hair, clothing and accessory traits through a three.ws address that mirrors the open-source trait library, so your browser never talks to an outside CDN. The list of where those traits live was being written using the address in the incoming request, and that address can be set by whoever sends the request. Because the mirror is cached at the edge for a day, one crafted request could have left every Avatar Studio visitor loading trait files from a stranger's server for the next 24 hours. That list is now written the same way for everybody, with no part of it taken from the request, so there is nothing left to steer. Three smaller repairs went in alongside it: asking the mirror how big a file is no longer answers with the compressed size (a trait manifest claimed 2 KB and delivered 20 KB), the mirror will not hand back a web page or script even if the upstream library ever served one, and a failure at the library now answers plainly as either not found or mirror is down instead of passing its raw status through as ours. (`/avatar-studio`) `[security, fix]`
- **Avatars no longer lose their preview picture forever because a server had a bad minute**: Every public avatar gets a preview picture rendered for it in the background, and an avatar that fails to render three times is given up on permanently. The problem was what counted as a failure. If the machine ran out of memory and could not start the renderer at all, or if the picture store was unreachable, that counted against the avatar, even though nothing was wrong with the avatar itself. Four avatars had already been written off that way for a fault that lasted seconds, and a longer outage would have written off every remaining one inside a quarter of an hour, because the whole batch fails together. A fault in the machinery is now told apart from a genuinely broken model: the avatar keeps its chances and the job simply waits for the machine to be healthy. The two background jobs also stop and do nothing when the picture store is switched off, instead of working through the queue and burning everyone's chances on an error they could not have avoided. The four avatars written off in error have been put back in the queue. (`/gallery`) `[fix]`
+- **Blockchain TVL by Chain is readable on a phone, and every chain in it is now a real link**: The chain leaderboard at /chains ranks 100 blockchains by total value locked and shows each one's dominance share of on-chain DeFi. On a phone the table was pinned to a fixed 540px width, so the only column that fit on screen was the chain name: the TVL and dominance figures, the entire point of the page, sat off the right edge behind a sideways scroll with nothing to hint they were there. The table now sizes itself to the screen, and all three columns read in full at 320px. Opening a chain's detail page was hidden in a similar way. Rows navigated on click but showed no clickable cursor, and the destination was not a link at all, so it could not be opened in a new tab, copied, followed from a screen reader's list of links, or seen by a search engine. Each chain name is now a genuine link, the row still follows it when you click anywhere along it, and the sortable column headers announce their sort state to screen readers. When the data provider fails, the page now offers a Try again button that refetches in place instead of a link that reloads everything, and it keeps its DeFiLlama attribution instead of going blank. If the provider reports nothing at all, the page says so rather than presenting a $0.00 total as though cross-chain DeFi had emptied out. (`/chains`) `[improvement, fix]`
+- **CA to x402 now shows the momentum and risk read its own paid endpoint sells**: Paste a contract address at /ca2x402 and the resolver has always returned four timeframes of price change and a 0-to-100 risk score with the reasoning behind it, the same intel an agent gets when it pays the generated endpoint $0.01 USDC. The page quietly threw both away and rendered only the 24h number, so a buyer could not see what they were about to pay for. Both now render: a momentum strip across 5m, 1h, 6h and 24h that shows whether a move is building or fading, and a risk panel with the score, its level, and the depth, age, float and flow checks that produced it, each marked by severity. Both read at 320px. The primary Resolve button was also painting white text on the white accent, so the page's main action looked like an empty slab; its label is now legible in both themes. And the error card no longer says "Could not resolve" for a mistyped address: a typo, a token with no market, and a resolver outage each say which one happened. Over on /capture, the URL and settings fields get a real focus ring for keyboard users, where an outline reset had left only a faint border tint. (`/ca2x402`) `[improvement, fix]`
- **Coin Autopilot tells you whether your rules actually saved**: Coin Autopilot is where you set the floors that decide when your coin buys itself back and when it pays creator fees out to holders, so a rule that looks set but never reached the database costs you real money. Two ways that could happen are fixed. A threshold you typed was only sent when the field lost focus, so typing a new floor and then closing the tab or hitting back threw it away silently; it now saves shortly after you stop typing, and anything still in flight is sent as the page closes. And the only confirmation was a toast that vanished after a couple of seconds, which meant a failed write left the switch showing a setting the server had never accepted. Each coin card now reports its own save: Saving, then Saved, or a Not saved marker that names the reason and stays put with a Retry button until the write lands. A coin you have never configured is marked Defaults, so the platform's inherited settings can no longer be mistaken for rules you chose. If the service goes down the page says so in both columns instead of leaving loading placeholders spinning, and it restores itself when the service returns rather than waiting for you to press Retry. Signed out, it stops retrying an endpoint that will keep refusing. Screen reader and keyboard fixes throughout: every switch and floor field is now named, save results are announced, the loading state is cleared when the data arrives, and the number fields have a visible focus ring. (`/autopilot`) `[fix, improvement]`
- **Community-voted models reach our Sketchfab showcase with a real name**: Models made from a picture rather than a written description carry no prompt to name them by, so they were labelled with the internal route name instead. Models the community voted for skip the usual title checks on purpose, since a winning model should never be dropped over its title, which meant a winner made from a picture was one run away from being published to the official three.ws Sketchfab account under the name Image-To-3d. It was, in fact, the second model queued to go out. Those models are now named from their category, or given a plain generic name when the category says nothing useful. (`/forge`) `[fix]`
- **Corrected stale numbers across the avatar and animation docs**: An audit of eight documentation pages against the code they describe turned up four claims that had drifted. The Animations page said the motion-capture library held about 2,000 clips when it actually serves 2,874, so both the intro and the gallery section now quote the real size and point at the count the API itself returns. The selfie-to-avatar page listed the wrong default for the face-morph displacement clamp and credited a single tuning sweep instead of the two that chose it. The avatar-creation walkthrough still told people to pick a body type on the selfie flow, a control that was deliberately removed because it changed nothing, and named the wrong fields on the event the page dispatches. The architecture page now says plainly that the whitelist trust mode is reserved rather than enforced. (`/docs/animations`) `[docs]`
@@ -34,6 +38,7 @@ Public history for [three.ws](https://three.ws), newest first. New pages come fr
- **Paying with $THREE works again: credits top-ups, tips and purchases stopped failing before they were signed**: Every attempt to top up credits with $THREE died in your wallet with a message about an incorrect program id, before you could approve anything and before any money moved. $THREE is issued under Solana's newer token standard, Token-2022, but the pages building the transfer were addressing the older one. That points the transfer at an account that does not exist, so the network rejected it during the safety check your wallet runs first. The credits page, agent tipping and the in-page purchase flow now read the standard directly from the coin itself and build against whichever one it actually uses, so $THREE and ordinary SPL tokens both go through. Two related touches on the credits page: it now tells you up front if the connected wallet holds no $THREE, or not enough for the amount you typed, instead of letting the wallet fail with something cryptic. The same mix-up was quietly hobbling the treasury side: the buyback and micro-buy lanes read their own $THREE balance from the wrong place, always saw zero, and so never moved what they had bought into the treasury. They now read and move it correctly. (`/credits`) `[fix]`
- **Prompt polish keeps its quotes out of your model, and a busy Forge says so plainly**: Three fixes across the Forge endpoints. The Enhance button, which rewrites your rough idea into the prompt the generator actually draws from, stopped removing the quotation marks the writing model wraps around its answer: those quotes were being handed straight to the image step as if you had asked for them to be drawn. Both plain and curly quotes are stripped again. Second, when you ask the studio to change a model you already made and the generator is saturated, you now get an honest "busy, try again in a moment" with a wait time instead of a vague failure that blamed the wording of your request. Third, the self-hosted generator panel now tells a caller which input was missing rather than reporting the whole feature as unconfigured, and its status light no longer claims nothing is set up when the only problem was the address you typed. (`/forge`) `[fix, improvement]`
- **Robinhood Chain worlds survive a hiccup in the live feed instead of going dark**: The in-world trading terminal inside a Robinhood Chain world polls the chain feed every five seconds, and the Worlds lobby lists those coins from the same feed. A single event that arrived without its payload, or with a timestamp the chain wrote in an unexpected format, took both requests down with a server error: the terminal showed a dead tape, and the lobby's Robinhood Chain tab quietly emptied with nothing on screen to say why. Bad events are now skipped and the rest of the tape still renders, a price that is genuinely missing stays blank rather than printing as a real $0 trade, and a launch with no readable contract address no longer produces a card that leads nowhere. Asking for trades with an address that is not a contract address now gets a clear error back instead of an empty list that read as "this coin has never traded". (`/worlds`) `[fix]`
+- **Scene Capture fits a phone screen, and a file that is not a point cloud now says so**: Scene Capture turns a video of a room into an explorable 3D point cloud you can orbit in the browser, and also renders any .ply you already have. On a phone the viewer stage forced the page 553px wide against a 320px screen, so the whole page scrolled sideways and the side panel sat off the edge; the stage now sizes to the screen and the page fits at 320px. Dropping in a file that is not a point cloud used to look like a success: the viewer reported 0 points over an empty black stage and offered the broken file back as a download. It now tells you the file is not a valid .ply and what one should contain. The error and progress headings were also being overwritten a second after they appeared, so every specific message (Could not fetch that point cloud, Reconstructing) reverted to the generic Something went wrong or Working; the real message now stays on screen. A reconstruction whose job stops responding no longer counts up forever: after a run of failed checks, or past twenty minutes, the page says what happened and how to restart. When the worker caps a long clip at its frame budget, the viewer says the clip was truncated instead of presenting a partial scene as the whole thing. And when reconstruction is offline, the message is written for you rather than repeating a server deployment checklist. Accessibility: the drop area is a real button you can reach with the keyboard, every field in the panel is labelled, and progress, errors and the point count are announced to screen readers. (`/capture`) `[fix, improvement]`
- **Scrolling the activity feed past a bad cursor no longer breaks the whole feed**: An audit of the account endpoints found one real break and two smaller ones. The activity feed on /community and your personal feed ask for the next page of results by passing the last item’s timestamp back; if that value arrived malformed, from a hand-edited link or a stale bookmark, the database rejected it and the entire feed came back as a server error with nothing to act on. It now answers with a plain "that is not a timestamp" so the page can recover and reload from the top. The referral code editor stated its length rule with a typographic dash that read as a minus sign on some devices, and now spells it out. The purchase history endpoint also documented the wrong default view: it always merges your purchases and your sales, which is what the dashboard shows. (`/community`) `[fix]`
- **Seeding an agent from your X account no longer asks for permission to post as you**: Teaching an agent your voice from your public X posts required connecting X first, and that connect asked X for everything the platform can do with an account, including posting and uploading media on your behalf. Reading your posts once and writing as you are two different things to agree to, and only the first one is what seeding needs. The Connect button on the X card in Settings now asks X for read access to your profile and posts and nothing else, and it says so before you click. Posting is its own connect, offered right beside it and labelled as such, and Reconnect keeps whichever of the two you already chose rather than quietly widening it. If a connection cannot read your posts at all, the card now says which permission is missing and offers a reconnect instead of showing you a consent screen for a seed that could not run, and an agent trying to post through a read-only connection is told exactly that instead of failing against X with nothing to act on. Everything else about the lane is unchanged: nothing is read until you consent on a screen that lists what is read and stored, and revoking still deletes every memory the seeding produced. (`/settings`) `[improvement, security]`
- **Shared agent links keep their preview card through a chain hiccup**: The preview card for an on-chain agent is rendered on demand from the chain and the agent's manifest. When either read was slow or briefly unavailable, the card answered as if the agent did not exist, and social sites cached that answer for up to an hour, so a link to a real agent unfurled with no image at all. The card is now served either way and only kept for a minute when a read fell short, so the next crawl picks up the real thing. Alongside it, ten agent endpoints (the preview card and page, the oEmbed provider, memory sync, the action log, ask, delegate and team tasks) now answer a mistyped agent id with a clear, specific error instead of a generic server failure. `[fix, improvement]`
@@ -51,6 +56,8 @@ Public history for [three.ws](https://three.ws), newest first. New pages come fr
- **The AGI page no longer goes quiet when its 3D body cannot load**: The AGI page is built around a live 3D body that reacts to the market, and when that model could not be fetched the page said nothing at all. The loading line disappeared on a timer whether or not anything had arrived, so the centrepiece became a silent empty panel with no explanation and no way to try again, while the decision stream and track record beside it carried on as normal. A body that fails to load now falls back to the platform's own avatar, so there is still something on the stage reacting to the agent's mood. If that cannot be fetched either, the stage says so plainly, makes clear that the agent is still trading and that the decisions and track record below are live and unaffected, and offers a button to try embodying again. The loading line now stays up until the body genuinely arrives instead of vanishing early, and the stage no longer re-requests a model it already knows is unreachable every twenty seconds. The page's description shown in search results and link previews was also carrying an older wording than the page itself. (`/agi`) `[fix, improvement]`
- **The Animation Gallery keeps its keyboard, its focus, and its promises about what will play**: The gallery of nearly 3,000 motion clips had a set of small breakages that added up. Roughly a third of the mocap catalogue has no picture made for it yet, and the page was asking the picture store for one anyway: every one of those requests came back empty, so a third of the cards spent a request to arrive at the same drawing they now show immediately. Opening a clip's details trapped nobody: pressing Tab walked out of the open window and into the page behind it, and closing the window dropped you back at the very top of the page instead of the card you came from. The playback controls were on show while a preview was still loading and stayed on show when it failed, so there were buttons that did nothing. Arrow keys jumped to the next clip even while you were dragging the speed slider, which moved the slider and changed the clip at the same time. Sharing a filtered view was misleading too: a link carrying the loops or one-shot filter applied it but left the buttons showing All, and a link with a filter the page does not recognise emptied the sort menu. All of it is fixed, along with the keyboard hint the window shows: Space really does play and pause now, from the moment the window opens. (`/animations`) `[fix, improvement]`
- **The arbitrage board says which price feeds it actually reached**: The x402 Arbitrage page compares what different providers charge for the same capability, and it builds that comparison from several facilitator catalogs at once. When one of those catalogs was unreachable, its listings simply vanished from every comparison and the page said nothing about it. A scan that had only seen half the market looked exactly like a market with fewer price gaps in it, which is the opposite of what the page is for. Each scan now names the facilitators it reached and how many listings each one contributed, and calls out any that dropped out along with the reason, so a short board is never mistaken for a quiet one. Separately, closing the payment window without paying used to leave the Pay cheapest button dead for four seconds before it would accept another click. Deciding not to pay now hands the button straight back. (`/arbitrage`) `[fix, improvement]`
+- **The artifact builder now reports what it actually measures, and stops flagging legal artifacts as over budget**: The /artifact page is the pre-flight for pasting a three.ws agent into a Claude conversation: it builds the self-contained HTML, renders it in a sandbox that mirrors Claude's, and reports what you are about to paste. Two of those numbers were misleading. The stat labelled First paint was really the time the endpoint took to hand over the bytes; nothing had been rendered yet when it was measured. It is now split into Fetch, which is that download, and Render, which times how long the sandboxed frame takes to parse the document and start the viewer. The size gauge was grading the finished artifact against the size limit for the model inside it, but inlining base64-encodes the model and costs four bytes for every three, so a perfectly acceptable 7 MB artifact was painted amber and anything built from a model over about 5.6 MB went red, contradicting the endpoint that had just accepted it. Size is now graded against the roughly 8.6 MB an accepted model can actually produce, and hovering the number tells you how much of that budget you have used. The Size budget note explained the same thing incorrectly and now spells out the base64 cost. (`/artifact`) `[fix, improvement]`
+- **The Avatar CLI page now shows real command output, and its terminal is readable by a screen reader**: The /avatar-cli page teaches the four commands of @three-ws/avatar-cli, and one of its examples was staged rather than captured: the validate card showed a seven-error failure stitched together from errors no single manifest produces. It now shows a verbatim run against a genuinely broken manifest, with the instance paths (/, /mesh, /skeleton) the CLI really prints, so what you read on the page is what your shell will print back. The replayed terminal in the hero was marked up as one flat image, which meant a screen reader announced a one-line caption and hid both the transcript and the Replay button inside it; the transcript is now a named, readable region and Replay is an ordinary button again. The live viewer's failure notice keeps a link to the web component reference, and that link now looks like a link and shows a focus ring when you tab to it, as do the scrollable code blocks. The copy button's fallback hint used to tell everyone to press Command C; it now names the shortcut for the keyboard you are actually on. (`/avatar-cli`) `[fix, improvement]`
- **The AWS page stops promising AWS invoicing three.ws does not do, and the GLB optimizer reads compressed avatars again**: Three corrections found by reading the docs pages against the code they describe. The AWS page told you a three.ws subscription would bill usage onto your AWS invoice, draw down AWS credits, and count toward Enterprise Discount Program commitments. None of that is the plan: the Marketplace subscription is free, and usage is paid per call in USDC over x402, exactly as it is for any other caller, so there is nothing to reconcile on your AWS bill at all. The page now says that, and describes the free subscription, the single key that opens every x402 endpoint, and the absence of any trial clock. Separately, the GLB optimization command documented on the 3D asset pipeline page failed on every compressed avatar, which is most of them, because it loaded the glTF extensions without the matching decoders; it now decodes meshopt and Draco and leaves the output as plain glTF the way its own contract promises. The NVIDIA models page also listed the tool-capable fallback model under the wrong vendor prefix. (`/aws`) `[fix, docs]`
- **The background agent mint says out loud whether it ran**: Every avatar that finishes a scan or a prompt is given a draft on-chain identity in the background. That step is deliberately best-effort, so it can never cost you your avatar, but it was also completely silent: whether it minted, was skipped because the deployment holds no minting key, or was switched off entirely, the logs looked identical, which is how a run of avatars can go by without anyone noticing the identity was never created. Each finished avatar now records the outcome of both chains, including the reason a leg was skipped. The documentation for the feature also gained a way to prove the whole path on a machine that cannot reach the public test-network faucet, which is routinely exhausted on shared build machines: run your own node with the real Metaplex programs, and the proof reports which node served it rather than handing back an explorer link that leads nowhere. (`/docs/draft-agent-mint`) `[improvement, docs]`
- **The BNB Chain page now says how many demos are actually live**: The BNB Chain page checks each of its three demo tracks against the running deployment and reports the tally in its header. That line was stuck: it read "Checking which tracks are live" forever, even once all three cards beside it had finished and said Live. The page's own result was being written first and then overwritten by the translation pass that swaps the page into your language, so the finished count never survived. The header now holds its result and reads "3 of 3 demo tracks live now". The same checks were also leaving four cancelled requests in the browser's network log on every visit, which made a healthy page look like it was failing to anyone with developer tools open. They now close cleanly. (`/bnb`) `[fix]`
diff --git a/data/changelog.json b/data/changelog.json
index d7340b195..5066aefe6 100644
--- a/data/changelog.json
+++ b/data/changelog.json
@@ -1,6 +1,66 @@
{
"$comment": "Curated changelog entries for three.ws, the editorial layer of the public changelog. New PAGE launches are tracked automatically via the `added` field in data/pages.json; this file is for everything else holders care about: improvements, fixes, SDK releases, trading/payments capabilities, security work. scripts/build-page-index.mjs merges both sources into CHANGELOG.md, public/changelog.json, and public/changelog.xml. Newest first. Allowed tags: feature, improvement, fix, sdk, infra, docs, security. `link` is optional and must be a live three.ws path.",
"entries": [
+ {
+ "date": "2026-08-16",
+ "title": "AR Studio: the language picker no longer sits on top of the dock controls",
+ "summary": "On the AR Studio page the site-wide floating language picker landed in the bottom-right corner, directly over the right-hand end of the control dock, so the Clear button (and the buttons beside it on a phone) could not be tapped at all. The picker now lives in the studio's own top bar, where it is always reachable and leaves the dock its full width. It also shrinks instead of pushing its neighbours off screen on narrow phones, which fixes the same crowding anywhere else it appears in a tight header. The page description now also names every place models can come from: your forges, the community feed, the free CC0 object library, and any GLB link.",
+ "tags": [
+ "fix",
+ "improvement"
+ ],
+ "link": "/ar/studio"
+ },
+ {
+ "date": "2026-08-16",
+ "title": "The artifact builder now reports what it actually measures, and stops flagging legal artifacts as over budget",
+ "summary": "The /artifact page is the pre-flight for pasting a three.ws agent into a Claude conversation: it builds the self-contained HTML, renders it in a sandbox that mirrors Claude's, and reports what you are about to paste. Two of those numbers were misleading. The stat labelled First paint was really the time the endpoint took to hand over the bytes; nothing had been rendered yet when it was measured. It is now split into Fetch, which is that download, and Render, which times how long the sandboxed frame takes to parse the document and start the viewer. The size gauge was grading the finished artifact against the size limit for the model inside it, but inlining base64-encodes the model and costs four bytes for every three, so a perfectly acceptable 7 MB artifact was painted amber and anything built from a model over about 5.6 MB went red, contradicting the endpoint that had just accepted it. Size is now graded against the roughly 8.6 MB an accepted model can actually produce, and hovering the number tells you how much of that budget you have used. The Size budget note explained the same thing incorrectly and now spells out the base64 cost.",
+ "tags": [
+ "fix",
+ "improvement"
+ ],
+ "link": "/artifact"
+ },
+ {
+ "date": "2026-08-16",
+ "title": "The Avatar CLI page now shows real command output, and its terminal is readable by a screen reader",
+ "summary": "The /avatar-cli page teaches the four commands of @three-ws/avatar-cli, and one of its examples was staged rather than captured: the validate card showed a seven-error failure stitched together from errors no single manifest produces. It now shows a verbatim run against a genuinely broken manifest, with the instance paths (/, /mesh, /skeleton) the CLI really prints, so what you read on the page is what your shell will print back. The replayed terminal in the hero was marked up as one flat image, which meant a screen reader announced a one-line caption and hid both the transcript and the Replay button inside it; the transcript is now a named, readable region and Replay is an ordinary button again. The live viewer's failure notice keeps a link to the web component reference, and that link now looks like a link and shows a focus ring when you tab to it, as do the scrollable code blocks. The copy button's fallback hint used to tell everyone to press Command C; it now names the shortcut for the keyboard you are actually on.",
+ "tags": [
+ "fix",
+ "improvement"
+ ],
+ "link": "/avatar-cli"
+ },
+ {
+ "date": "2026-08-16",
+ "title": "Scene Capture fits a phone screen, and a file that is not a point cloud now says so",
+ "summary": "Scene Capture turns a video of a room into an explorable 3D point cloud you can orbit in the browser, and also renders any .ply you already have. On a phone the viewer stage forced the page 553px wide against a 320px screen, so the whole page scrolled sideways and the side panel sat off the edge; the stage now sizes to the screen and the page fits at 320px. Dropping in a file that is not a point cloud used to look like a success: the viewer reported 0 points over an empty black stage and offered the broken file back as a download. It now tells you the file is not a valid .ply and what one should contain. The error and progress headings were also being overwritten a second after they appeared, so every specific message (Could not fetch that point cloud, Reconstructing) reverted to the generic Something went wrong or Working; the real message now stays on screen. A reconstruction whose job stops responding no longer counts up forever: after a run of failed checks, or past twenty minutes, the page says what happened and how to restart. When the worker caps a long clip at its frame budget, the viewer says the clip was truncated instead of presenting a partial scene as the whole thing. And when reconstruction is offline, the message is written for you rather than repeating a server deployment checklist. Accessibility: the drop area is a real button you can reach with the keyboard, every field in the panel is labelled, and progress, errors and the point count are announced to screen readers.",
+ "tags": [
+ "fix",
+ "improvement"
+ ],
+ "link": "/capture"
+ },
+ {
+ "date": "2026-08-16",
+ "title": "CA to x402 now shows the momentum and risk read its own paid endpoint sells",
+ "summary": "Paste a contract address at /ca2x402 and the resolver has always returned four timeframes of price change and a 0-to-100 risk score with the reasoning behind it, the same intel an agent gets when it pays the generated endpoint $0.01 USDC. The page quietly threw both away and rendered only the 24h number, so a buyer could not see what they were about to pay for. Both now render: a momentum strip across 5m, 1h, 6h and 24h that shows whether a move is building or fading, and a risk panel with the score, its level, and the depth, age, float and flow checks that produced it, each marked by severity. Both read at 320px. The primary Resolve button was also painting white text on the white accent, so the page's main action looked like an empty slab; its label is now legible in both themes. And the error card no longer says \"Could not resolve\" for a mistyped address: a typo, a token with no market, and a resolver outage each say which one happened. Over on /capture, the URL and settings fields get a real focus ring for keyboard users, where an outline reset had left only a faint border tint.",
+ "tags": [
+ "improvement",
+ "fix"
+ ],
+ "link": "/ca2x402"
+ },
+ {
+ "date": "2026-08-16",
+ "title": "Blockchain TVL by Chain is readable on a phone, and every chain in it is now a real link",
+ "summary": "The chain leaderboard at /chains ranks 100 blockchains by total value locked and shows each one's dominance share of on-chain DeFi. On a phone the table was pinned to a fixed 540px width, so the only column that fit on screen was the chain name: the TVL and dominance figures, the entire point of the page, sat off the right edge behind a sideways scroll with nothing to hint they were there. The table now sizes itself to the screen, and all three columns read in full at 320px. Opening a chain's detail page was hidden in a similar way. Rows navigated on click but showed no clickable cursor, and the destination was not a link at all, so it could not be opened in a new tab, copied, followed from a screen reader's list of links, or seen by a search engine. Each chain name is now a genuine link, the row still follows it when you click anywhere along it, and the sortable column headers announce their sort state to screen readers. When the data provider fails, the page now offers a Try again button that refetches in place instead of a link that reloads everything, and it keeps its DeFiLlama attribution instead of going blank. If the provider reports nothing at all, the page says so rather than presenting a $0.00 total as though cross-chain DeFi had emptied out.",
+ "tags": [
+ "improvement",
+ "fix"
+ ],
+ "link": "/chains"
+ },
{
"date": "2026-08-16",
"title": "Coin Autopilot tells you whether your rules actually saved",
@@ -10,7 +70,18 @@
"improvement"
],
"link": "/autopilot"
- }, {
+ },
+ {
+ "date": "2026-08-16",
+ "title": "Autopilot Activity keeps up when you switch agents, and its receipts link to the exact memory",
+ "summary": "Autopilot Activity lists every action your agents took on your behalf, and it has an agent picker for narrowing that list to one of them. Picking a second agent while the first was still loading used to do nothing at all: the picker and the address bar moved to the new agent, but the request for it was never sent and the page kept showing the previous agent's actions, with no sign anything had gone wrong. Whichever agent you picked last now wins, however fast you switch. Two more repairs: each receipt cites the memories that motivated the action, and those links now scroll to the cited memory and highlight it instead of dropping you at the top of a long Knowledge list; and when an agent has no actions yet, the page sends you to that agent's own autopilot controls rather than the dashboard, which is where the setting you need actually lives.",
+ "tags": [
+ "fix",
+ "improvement"
+ ],
+ "link": "/autopilot-activity"
+ },
+ {
"date": "2026-08-16",
"title": "Crypto Categories is searchable, and opens in a fraction of the time",
"summary": "The sector leaderboard at /categories renders every category the market data provider reports, and that list has grown past 750. All of them were being drawn into the page at once, roughly 2,250 coin icons in a single paint, and the only way to find a sector was to scroll past every one ranked above it. The table now opens with the top 50 and pages in 50 more at a time, and a search box narrows the whole list by name before you ever need to page. Full category names are shown too, instead of being cut off mid-word on a wide screen with half the row empty. The three things that can go wrong now read differently and each offers a way out: a failed load says the data provider did not answer and gives you a Try again button that refetches, a provider reporting nothing says so plainly, and a search that matches nothing offers to clear itself. Sortable column headers now announce themselves to screen readers, and keyboard focus stays on the column you just sorted instead of being dropped back to the top of the page.",
diff --git a/public/changelog.json b/public/changelog.json
index 2e1d38d35..dd713553f 100644
--- a/public/changelog.json
+++ b/public/changelog.json
@@ -1,5 +1,5 @@
{
- "generated_at": "2026-08-16T20:17:45.645Z",
+ "generated_at": "2026-08-16T20:23:07.210Z",
"generated_by": "scripts/build-page-index.mjs from data/pages.json + data/changelog.json",
"site": {
"name": "three.ws",
@@ -122,6 +122,28 @@
"fix"
]
},
+ {
+ "date": "2026-08-16",
+ "type": "update",
+ "title": "AR Studio: the language picker no longer sits on top of the dock controls",
+ "summary": "On the AR Studio page the site-wide floating language picker landed in the bottom-right corner, directly over the right-hand end of the control dock, so the Clear button (and the buttons beside it on a phone) could not be tapped at all. The picker now lives in the studio's own top bar, where it is always reachable and leaves the dock its full width. It also shrinks instead of pushing its neighbours off screen on narrow phones, which fixes the same crowding anywhere else it appears in a tight header. The page description now also names every place models can come from: your forges, the community feed, the free CC0 object library, and any GLB link.",
+ "link": "/ar/studio",
+ "tags": [
+ "fix",
+ "improvement"
+ ]
+ },
+ {
+ "date": "2026-08-16",
+ "type": "update",
+ "title": "Autopilot Activity keeps up when you switch agents, and its receipts link to the exact memory",
+ "summary": "Autopilot Activity lists every action your agents took on your behalf, and it has an agent picker for narrowing that list to one of them. Picking a second agent while the first was still loading used to do nothing at all: the picker and the address bar moved to the new agent, but the request for it was never sent and the page kept showing the previous agent's actions, with no sign anything had gone wrong. Whichever agent you picked last now wins, however fast you switch. Two more repairs: each receipt cites the memories that motivated the action, and those links now scroll to the cited memory and highlight it instead of dropping you at the top of a long Knowledge list; and when an agent has no actions yet, the page sends you to that agent's own autopilot controls rather than the dashboard, which is where the setting you need actually lives.",
+ "link": "/autopilot-activity",
+ "tags": [
+ "fix",
+ "improvement"
+ ]
+ },
{
"date": "2026-08-16",
"type": "update",
@@ -153,6 +175,28 @@
"fix"
]
},
+ {
+ "date": "2026-08-16",
+ "type": "update",
+ "title": "Blockchain TVL by Chain is readable on a phone, and every chain in it is now a real link",
+ "summary": "The chain leaderboard at /chains ranks 100 blockchains by total value locked and shows each one's dominance share of on-chain DeFi. On a phone the table was pinned to a fixed 540px width, so the only column that fit on screen was the chain name: the TVL and dominance figures, the entire point of the page, sat off the right edge behind a sideways scroll with nothing to hint they were there. The table now sizes itself to the screen, and all three columns read in full at 320px. Opening a chain's detail page was hidden in a similar way. Rows navigated on click but showed no clickable cursor, and the destination was not a link at all, so it could not be opened in a new tab, copied, followed from a screen reader's list of links, or seen by a search engine. Each chain name is now a genuine link, the row still follows it when you click anywhere along it, and the sortable column headers announce their sort state to screen readers. When the data provider fails, the page now offers a Try again button that refetches in place instead of a link that reloads everything, and it keeps its DeFiLlama attribution instead of going blank. If the provider reports nothing at all, the page says so rather than presenting a $0.00 total as though cross-chain DeFi had emptied out.",
+ "link": "/chains",
+ "tags": [
+ "improvement",
+ "fix"
+ ]
+ },
+ {
+ "date": "2026-08-16",
+ "type": "update",
+ "title": "CA to x402 now shows the momentum and risk read its own paid endpoint sells",
+ "summary": "Paste a contract address at /ca2x402 and the resolver has always returned four timeframes of price change and a 0-to-100 risk score with the reasoning behind it, the same intel an agent gets when it pays the generated endpoint $0.01 USDC. The page quietly threw both away and rendered only the 24h number, so a buyer could not see what they were about to pay for. Both now render: a momentum strip across 5m, 1h, 6h and 24h that shows whether a move is building or fading, and a risk panel with the score, its level, and the depth, age, float and flow checks that produced it, each marked by severity. Both read at 320px. The primary Resolve button was also painting white text on the white accent, so the page's main action looked like an empty slab; its label is now legible in both themes. And the error card no longer says \"Could not resolve\" for a mistyped address: a typo, a token with no market, and a resolver outage each say which one happened. Over on /capture, the URL and settings fields get a real focus ring for keyboard users, where an outline reset had left only a faint border tint.",
+ "link": "/ca2x402",
+ "tags": [
+ "improvement",
+ "fix"
+ ]
+ },
{
"date": "2026-08-16",
"type": "update",
@@ -303,6 +347,17 @@
"fix"
]
},
+ {
+ "date": "2026-08-16",
+ "type": "update",
+ "title": "Scene Capture fits a phone screen, and a file that is not a point cloud now says so",
+ "summary": "Scene Capture turns a video of a room into an explorable 3D point cloud you can orbit in the browser, and also renders any .ply you already have. On a phone the viewer stage forced the page 553px wide against a 320px screen, so the whole page scrolled sideways and the side panel sat off the edge; the stage now sizes to the screen and the page fits at 320px. Dropping in a file that is not a point cloud used to look like a success: the viewer reported 0 points over an empty black stage and offered the broken file back as a download. It now tells you the file is not a valid .ply and what one should contain. The error and progress headings were also being overwritten a second after they appeared, so every specific message (Could not fetch that point cloud, Reconstructing) reverted to the generic Something went wrong or Working; the real message now stays on screen. A reconstruction whose job stops responding no longer counts up forever: after a run of failed checks, or past twenty minutes, the page says what happened and how to restart. When the worker caps a long clip at its frame budget, the viewer says the clip was truncated instead of presenting a partial scene as the whole thing. And when reconstruction is offline, the message is written for you rather than repeating a server deployment checklist. Accessibility: the drop area is a real button you can reach with the keyboard, every field in the panel is labelled, and progress, errors and the point count are announced to screen readers.",
+ "link": "/capture",
+ "tags": [
+ "fix",
+ "improvement"
+ ]
+ },
{
"date": "2026-08-16",
"type": "update",
@@ -482,6 +537,28 @@
"improvement"
]
},
+ {
+ "date": "2026-08-16",
+ "type": "update",
+ "title": "The artifact builder now reports what it actually measures, and stops flagging legal artifacts as over budget",
+ "summary": "The /artifact page is the pre-flight for pasting a three.ws agent into a Claude conversation: it builds the self-contained HTML, renders it in a sandbox that mirrors Claude's, and reports what you are about to paste. Two of those numbers were misleading. The stat labelled First paint was really the time the endpoint took to hand over the bytes; nothing had been rendered yet when it was measured. It is now split into Fetch, which is that download, and Render, which times how long the sandboxed frame takes to parse the document and start the viewer. The size gauge was grading the finished artifact against the size limit for the model inside it, but inlining base64-encodes the model and costs four bytes for every three, so a perfectly acceptable 7 MB artifact was painted amber and anything built from a model over about 5.6 MB went red, contradicting the endpoint that had just accepted it. Size is now graded against the roughly 8.6 MB an accepted model can actually produce, and hovering the number tells you how much of that budget you have used. The Size budget note explained the same thing incorrectly and now spells out the base64 cost.",
+ "link": "/artifact",
+ "tags": [
+ "fix",
+ "improvement"
+ ]
+ },
+ {
+ "date": "2026-08-16",
+ "type": "update",
+ "title": "The Avatar CLI page now shows real command output, and its terminal is readable by a screen reader",
+ "summary": "The /avatar-cli page teaches the four commands of @three-ws/avatar-cli, and one of its examples was staged rather than captured: the validate card showed a seven-error failure stitched together from errors no single manifest produces. It now shows a verbatim run against a genuinely broken manifest, with the instance paths (/, /mesh, /skeleton) the CLI really prints, so what you read on the page is what your shell will print back. The replayed terminal in the hero was marked up as one flat image, which meant a screen reader announced a one-line caption and hid both the transcript and the Replay button inside it; the transcript is now a named, readable region and Replay is an ordinary button again. The live viewer's failure notice keeps a link to the web component reference, and that link now looks like a link and shows a focus ring when you tab to it, as do the scrollable code blocks. The copy button's fallback hint used to tell everyone to press Command C; it now names the shortcut for the keyboard you are actually on.",
+ "link": "/avatar-cli",
+ "tags": [
+ "fix",
+ "improvement"
+ ]
+ },
{
"date": "2026-08-16",
"type": "update",
diff --git a/public/changelog.xml b/public/changelog.xml
index 1984b4fa2..184be19ef 100644
--- a/public/changelog.xml
+++ b/public/changelog.xml
@@ -94,6 +94,22 @@
<category>fix</category>
<description>Every three.ws newsletter email carries a confirm link and an unsubscribe link that share one token, so a reader who unsubscribed and later opened an older email and clicked confirm was quietly put back on the list. That link no longer signs anyone up: it says you are unsubscribed and leaves it that way, and opting back in starts from the site. Three other fixes ship alongside it. The allocation planner at /meta-allocator no longer serves an empty basket when the model declines to answer, it falls back to the allocator that always produces a real spread across verified leaders. The motion capture uploader told you an upload link lasted ten minutes when it really lasted five, so a large video queued on that promise could fail with nothing explaining why. And an agent whose stablecoin payment is rejected by the x402 Bazaar now gets the Bazaar's own details back instead of a different server's, so its retry is aimed at the right place.</description>
</item>
+ <item>
+ <title>AR Studio: the language picker no longer sits on top of the dock controls</title>
+ <link>https://three.ws/ar/studio</link>
+ <guid isPermaLink="false">2026-08-16:AR Studio: the language picker no longer sits on top of the dock controls</guid>
+ <pubDate>Sun, 16 Aug 2026 12:00:00 GMT</pubDate>
+ <category>fix</category>
+ <description>On the AR Studio page the site-wide floating language picker landed in the bottom-right corner, directly over the right-hand end of the control dock, so the Clear button (and the buttons beside it on a phone) could not be tapped at all. The picker now lives in the studio's own top bar, where it is always reachable and leaves the dock its full width. It also shrinks instead of pushing its neighbours off screen on narrow phones, which fixes the same crowding anywhere else it appears in a tight header. The page description now also names every place models can come from: your forges, the community feed, the free CC0 object library, and any GLB link.</description>
+ </item>
+ <item>
+ <title>Autopilot Activity keeps up when you switch agents, and its receipts link to the exact memory</title>
+ <link>https://three.ws/autopilot-activity</link>
+ <guid isPermaLink="false">2026-08-16:Autopilot Activity keeps up when you switch agents, and its receipts link to the exact memory</guid>
+ <pubDate>Sun, 16 Aug 2026 12:00:00 GMT</pubDate>
+ <category>fix</category>
+ <description>Autopilot Activity lists every action your agents took on your behalf, and it has an agent picker for narrowing that list to one of them. Picking a second agent while the first was still loading used to do nothing at all: the picker and the address bar moved to the new agent, but the request for it was never sent and the page kept showing the previous agent's actions, with no sign anything had gone wrong. Whichever agent you picked last now wins, however fast you switch. Two more repairs: each receipt cites the memories that motivated the action, and those links now scroll to the cited memory and highlight it instead of dropping you at the top of a long Knowledge list; and when an agent has no actions yet, the page sends you to that agent's own autopilot controls rather than the dashboard, which is where the setting you need actually lives.</description>
+ </item>
<item>
<title>Avatar link previews stop rebuilding themselves, and two feeds answer a bad page link plainly</title>
<link>https://three.ws/characters</link>
@@ -118,6 +134,22 @@
<category>fix</category>
<description>Every public avatar gets a preview picture rendered for it in the background, and an avatar that fails to render three times is given up on permanently. The problem was what counted as a failure. If the machine ran out of memory and could not start the renderer at all, or if the picture store was unreachable, that counted against the avatar, even though nothing was wrong with the avatar itself. Four avatars had already been written off that way for a fault that lasted seconds, and a longer outage would have written off every remaining one inside a quarter of an hour, because the whole batch fails together. A fault in the machinery is now told apart from a genuinely broken model: the avatar keeps its chances and the job simply waits for the machine to be healthy. The two background jobs also stop and do nothing when the picture store is switched off, instead of working through the queue and burning everyone's chances on an error they could not have avoided. The four avatars written off in error have been put back in the queue.</description>
</item>
+ <item>
+ <title>Blockchain TVL by Chain is readable on a phone, and every chain in it is now a real link</title>
+ <link>https://three.ws/chains</link>
+ <guid isPermaLink="false">2026-08-16:Blockchain TVL by Chain is readable on a phone, and every chain in it is now a real link</guid>
+ <pubDate>Sun, 16 Aug 2026 12:00:00 GMT</pubDate>
+ <category>improvement</category>
+ <description>The chain leaderboard at /chains ranks 100 blockchains by total value locked and shows each one's dominance share of on-chain DeFi. On a phone the table was pinned to a fixed 540px width, so the only column that fit on screen was the chain name: the TVL and dominance figures, the entire point of the page, sat off the right edge behind a sideways scroll with nothing to hint they were there. The table now sizes itself to the screen, and all three columns read in full at 320px. Opening a chain's detail page was hidden in a similar way. Rows navigated on click but showed no clickable cursor, and the destination was not a link at all, so it could not be opened in a new tab, copied, followed from a screen reader's list of links, or seen by a search engine. Each chain name is now a genuine link, the row still follows it when you click anywhere along it, and the sortable column headers announce their sort state to screen readers. When the data provider fails, the page now offers a Try again button that refetches in place instead of a link that reloads everything, and it keeps its DeFiLlama attribution instead of going blank. If the provider reports nothing at all, the page says so rather than presenting a $0.00 total as though cross-chain DeFi had emptied out.</description>
+ </item>
+ <item>
+ <title>CA to x402 now shows the momentum and risk read its own paid endpoint sells</title>
+ <link>https://three.ws/ca2x402</link>
+ <guid isPermaLink="false">2026-08-16:CA to x402 now shows the momentum and risk read its own paid endpoint sells</guid>
+ <pubDate>Sun, 16 Aug 2026 12:00:00 GMT</pubDate>
+ <category>improvement</category>
+ <description>Paste a contract address at /ca2x402 and the resolver has always returned four timeframes of price change and a 0-to-100 risk score with the reasoning behind it, the same intel an agent gets when it pays the generated endpoint $0.01 USDC. The page quietly threw both away and rendered only the 24h number, so a buyer could not see what they were about to pay for. Both now render: a momentum strip across 5m, 1h, 6h and 24h that shows whether a move is building or fading, and a risk panel with the score, its level, and the depth, age, float and flow checks that produced it, each marked by severity. Both read at 320px. The primary Resolve button was also painting white text on the white accent, so the page's main action looked like an empty slab; its label is now legible in both themes. And the error card no longer says &quot;Could not resolve&quot; for a mistyped address: a typo, a token with no market, and a resolver outage each say which one happened. Over on /capture, the URL and settings fields get a real focus ring for keyboard users, where an outline reset had left only a faint border tint.</description>
+ </item>
<item>
<title>Coin Autopilot tells you whether your rules actually saved</title>
<link>https://three.ws/autopilot</link>
@@ -230,6 +262,14 @@
<category>fix</category>
<description>The in-world trading terminal inside a Robinhood Chain world polls the chain feed every five seconds, and the Worlds lobby lists those coins from the same feed. A single event that arrived without its payload, or with a timestamp the chain wrote in an unexpected format, took both requests down with a server error: the terminal showed a dead tape, and the lobby's Robinhood Chain tab quietly emptied with nothing on screen to say why. Bad events are now skipped and the rest of the tape still renders, a price that is genuinely missing stays blank rather than printing as a real $0 trade, and a launch with no readable contract address no longer produces a card that leads nowhere. Asking for trades with an address that is not a contract address now gets a clear error back instead of an empty list that read as &quot;this coin has never traded&quot;.</description>
</item>
+ <item>
+ <title>Scene Capture fits a phone screen, and a file that is not a point cloud now says so</title>
+ <link>https://three.ws/capture</link>
+ <guid isPermaLink="false">2026-08-16:Scene Capture fits a phone screen, and a file that is not a point cloud now says so</guid>
+ <pubDate>Sun, 16 Aug 2026 12:00:00 GMT</pubDate>
+ <category>fix</category>
+ <description>Scene Capture turns a video of a room into an explorable 3D point cloud you can orbit in the browser, and also renders any .ply you already have. On a phone the viewer stage forced the page 553px wide against a 320px screen, so the whole page scrolled sideways and the side panel sat off the edge; the stage now sizes to the screen and the page fits at 320px. Dropping in a file that is not a point cloud used to look like a success: the viewer reported 0 points over an empty black stage and offered the broken file back as a download. It now tells you the file is not a valid .ply and what one should contain. The error and progress headings were also being overwritten a second after they appeared, so every specific message (Could not fetch that point cloud, Reconstructing) reverted to the generic Something went wrong or Working; the real message now stays on screen. A reconstruction whose job stops responding no longer counts up forever: after a run of failed checks, or past twenty minutes, the page says what happened and how to restart. When the worker caps a long clip at its frame budget, the viewer says the clip was truncated instead of presenting a partial scene as the whole thing. And when reconstruction is offline, the message is written for you rather than repeating a server deployment checklist. Accessibility: the drop area is a real button you can reach with the keyboard, every field in the panel is labelled, and progress, errors and the point count are announced to screen readers.</description>
+ </item>
<item>
<title>Scrolling the activity feed past a bad cursor no longer breaks the whole feed</title>
<link>https://three.ws/community</link>
@@ -366,6 +406,22 @@
<category>fix</category>
<description>The x402 Arbitrage page compares what different providers charge for the same capability, and it builds that comparison from several facilitator catalogs at once. When one of those catalogs was unreachable, its listings simply vanished from every comparison and the page said nothing about it. A scan that had only seen half the market looked exactly like a market with fewer price gaps in it, which is the opposite of what the page is for. Each scan now names the facilitators it reached and how many listings each one contributed, and calls out any that dropped out along with the reason, so a short board is never mistaken for a quiet one. Separately, closing the payment window without paying used to leave the Pay cheapest button dead for four seconds before it would accept another click. Deciding not to pay now hands the button straight back.</description>
</item>
+ <item>
+ <title>The artifact builder now reports what it actually measures, and stops flagging legal artifacts as over budget</title>
+ <link>https://three.ws/artifact</link>
+ <guid isPermaLink="false">2026-08-16:The artifact builder now reports what it actually measures, and stops flagging legal artifacts as over budget</guid>
+ <pubDate>Sun, 16 Aug 2026 12:00:00 GMT</pubDate>
+ <category>fix</category>
+ <description>The /artifact page is the pre-flight for pasting a three.ws agent into a Claude conversation: it builds the self-contained HTML, renders it in a sandbox that mirrors Claude's, and reports what you are about to paste. Two of those numbers were misleading. The stat labelled First paint was really the time the endpoint took to hand over the bytes; nothing had been rendered yet when it was measured. It is now split into Fetch, which is that download, and Render, which times how long the sandboxed frame takes to parse the document and start the viewer. The size gauge was grading the finished artifact against the size limit for the model inside it, but inlining base64-encodes the model and costs four bytes for every three, so a perfectly acceptable 7 MB artifact was painted amber and anything built from a model over about 5.6 MB went red, contradicting the endpoint that had just accepted it. Size is now graded against the roughly 8.6 MB an accepted model can actually produce, and hovering the number tells you how much of that budget you have used. The Size budget note explained the same thing incorrectly and now spells out the base64 cost.</description>
+ </item>
+ <item>
+ <title>The Avatar CLI page now shows real command output, and its terminal is readable by a screen reader</title>
+ <link>https://three.ws/avatar-cli</link>
+ <guid isPermaLink="false">2026-08-16:The Avatar CLI page now shows real command output, and its terminal is readable by a screen reader</guid>
+ <pubDate>Sun, 16 Aug 2026 12:00:00 GMT</pubDate>
+ <category>fix</category>
+ <description>The /avatar-cli page teaches the four commands of @three-ws/avatar-cli, and one of its examples was staged rather than captured: the validate card showed a seven-error failure stitched together from errors no single manifest produces. It now shows a verbatim run against a genuinely broken manifest, with the instance paths (/, /mesh, /skeleton) the CLI really prints, so what you read on the page is what your shell will print back. The replayed terminal in the hero was marked up as one flat image, which meant a screen reader announced a one-line caption and hid both the transcript and the Replay button inside it; the transcript is now a named, readable region and Replay is an ordinary button again. The live viewer's failure notice keeps a link to the web component reference, and that link now looks like a link and shows a focus ring when you tab to it, as do the scrollable code blocks. The copy button's fallback hint used to tell everyone to press Command C; it now names the shortcut for the keyboard you are actually on.</description>
+ </item>
<item>
<title>The AWS page stops promising AWS invoicing three.ws does not do, and the GLB optimizer reads compressed avatars again</title>
<link>https://three.ws/aws</link>
diff --git a/public/features.json b/public/features.json
index d732e9338..0cd49f8a1 100644
--- a/public/features.json
+++ b/public/features.json
@@ -1,5 +1,5 @@
{
- "generated_at": "2026-08-16T20:17:45.173Z",
+ "generated_at": "2026-08-16T20:23:06.967Z",
"generated_by": "scripts/build-page-index.mjs from data/pages.json",
"site": {
"name": "three.ws",
diff --git a/src/agent-edit.js b/src/agent-edit.js
index 94801aa53..7c674785a 100644
--- a/src/agent-edit.js
+++ b/src/agent-edit.js
@@ -1934,12 +1934,26 @@ function renderMemories(rows) {
// browser has already resolved the fragment, so nothing scrolled and the cited
// memory was indistinguishable from the rest: bring it into view and mark it.
function revealCitedMemory() {
- const id = decodeURIComponent(location.hash.replace(/^#/, ''));
- if (!id.startsWith('mem-')) return;
- const row = document.getElementById(id);
+ if (!hasCitedMemoryTarget()) return;
+ const row = document.getElementById(decodeURIComponent(location.hash.slice(1)));
if (!row) return;
- row.scrollIntoView({ behavior: 'smooth', block: 'center' });
row.classList.add('mem-cited');
+
+ // The sections above this one mount asynchronously (avatar preview, mind
+ // palace, analytics) and every one that lands grows the column above the
+ // memory list, pushing the cited row back off screen after a single scroll.
+ // Watch the row itself and re-centre whenever it drifts out of view, then
+ // stand down the moment the reader starts scrolling for themselves. Nothing
+ // polls: if the layout stops moving, the observer simply stops firing.
+ const col = document.querySelector('.content-col');
+ const io = new IntersectionObserver((entries) => {
+ for (const e of entries) if (!e.isIntersecting) row.scrollIntoView({ block: 'center' });
+ }, { root: col && getComputedStyle(col).overflowY !== 'visible' ? col : null, threshold: 0.5 });
+ for (const evt of ['wheel', 'touchstart', 'keydown']) {
+ window.addEventListener(evt, () => io.disconnect(), { once: true, passive: true });
+ }
+ row.scrollIntoView({ behavior: 'smooth', block: 'center' });
+ io.observe(row);
}
async function toggleMemoryVisibility(btn) {
@@ -3297,8 +3311,17 @@ function ensureAutopilotTab() {
mountAutopilotMind(host, { agentId });
}
+// True when the URL fragment names a specific memory (an autopilot receipt's
+// provenance link). That target is more precise than the ?tab= section, so it
+// owns the scroll and the section scroll stands down: two smooth scrolls racing
+// each other landed the reader wherever the slower one finished.
+function hasCitedMemoryTarget() {
+ return /^#mem-/.test(location.hash);
+}
+
// Scroll to a named section in the unified layout (used by OAuth callback redirects via ?tab=…)
function scrollToSection(tabId) {
+ if (hasCitedMemoryTarget()) return;
const sectionMap = {
persona: 'section-identity',
outfit: 'section-identity',