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.

docs(seeding): record what the live cron actually accepts, and the scale numbers at 30k

nirholasf432b88af5c7
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 746f954bb..4e3f0a21c 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -23,6 +23,7 @@ Public history for [three.ws](https://three.ws), newest first. New pages come fr
- **Fact-check now reads the part of a source that is actually about your claim**: When the fact-checker fell back to Wikipedia, it only ever read each article's opening summary. An opening summary says what a page is about, so it carries the answer only when your claim is about the page's subject. Ask whether Napoleon was unusually short and it read the first lines of "Napoleon", "Napoleon III" and "Napoleonic Wars", none of which mention height at all, so every source was scored as saying nothing and the verdict came back "not enough evidence" while the passage about his height sat further down the same article. The checker now reads each article in full and picks the paragraphs that actually discuss your claim, so that question returns the "Appearance and image" and "Napoleon's height" sections instead. Reference lists are skipped, since they repeat your words without asserting anything, and the deeper read is bounded by the time the check already had, so a slow source can never cost you the evidence that already arrived. (`/docs/fact-check`) `[improvement]`
- **Fade Radar: the wallets that have never once been right, and the coins they are crowding into**: The Smart Money Radar tells you which proven winners are buying a coin. Fade Radar is the other half of that question, and on pump.fun it turns out to be the sharper one: how much of a coin's buy side is money that has never been right. A wallet joins the reverse-indicator board only after five coins it bought have a known outcome and not one of them won, which is a bar a surprising number of addresses clear. The new page ranks every launch from the last few hours by how much of its observed buy side belongs to those wallets, by wallet count and by SOL, and carries a second board of the wallets themselves with their full record. The number above the board is the part we care most about: it is measured, not asserted. We rebuild the reverse-indicator group using only coins whose outcome was known before the coins it is graded on existed, then check what actually happened. Coins where a quarter or more of the buyers were reverse indicators went on to win 1.8 percent of the time, against 22.1 percent for coins with none, across 26,626 labelled coins. Nothing here is a short and nothing signs anything: on a bonding curve the only tradeable form of a fade is not buying, so every action on the page is a link out to the coin, its intel, or a wallet's track record. Wallet links also now open directly on the Smart Money Radar, so an address cited anywhere on the platform lands on its record instead of a board you have to scan by eye. (`/fade`) `[feature]`
- **Homes connected through the relay now follow the house when you rearrange it**: If your Home Assistant is only reachable inside your house, three.ws reaches it through the three.ws integration you install there, which dials out to us. Those homes were building their room list once, when the connection opened, and then quietly ignoring changes: assign a lamp to the kitchen, rename a floor, add a device, and the 3D house kept showing the old arrangement until the connection happened to be rebuilt. Homes connected by a direct address never had this problem, and the split was in the relay, which was not carrying the four small notices Home Assistant sends when its rooms, floors, devices or entities change. It carries them now, so both kinds of connection track the house the same way and updates land in seconds. The notices only say that something changed; the relay already read the full room and floor lists, so nothing new about your house is exposed by this. (`/smart-home`) `[fix, infra]`
+- **New gallery avatars now arrive with a skeleton, so they animate properly out of the box**: Avatars generated for the gallery were published as static meshes: they still moved, but through a fallback that maps a generic body onto them rather than a skeleton of their own. Every new one now goes through our auto-rigger first and arrives with a real 52-bone humanoid skeleton, which means the whole motion library drives it directly. If the rigger fails or stalls, the avatar is still published rather than held back, so nothing is lost to a bad rig. (`/gallery`) `[improvement, feature]`
- **News cards stop losing their preview picture to a cold cache**: About a fifth of the crypto-news feed ships text-only, so those cards start as a source tile and upgrade in place once the publisher's own preview image is found. That lookup searched all 197 publisher feeds at once and gave up after 2.5 seconds, which on a freshly started server dropped whichever feed had not answered yet, so an article the feed had just served came back as unknown. The reader kept its tile, but the failed lookup was recorded as a public answer for the next ten minutes, so every visitor to /markets and /markets/news got the same dead result and a 404 in their browser console. The lookup now asks the one publisher that could hold the article first and waits for it, and a result that could not be confirmed is held for 30 seconds instead of ten minutes. (`/markets/news`) `[fix]`
- **Oracle: a failed request now says so, instead of claiming no coins have been scored**: A full audit of /oracle, the fused conviction board for pump.fun launches. When a panel's request failed, the page told you the opposite of the truth: the live feed claimed "Oracle is warming up" and said the engine ships with its backend, the movers, proof, agent-ranking, activity and 3D-graph panels each claimed there was simply nothing to show, and the performance record said the edge was still proving itself. Every one of those is now a real error state that names what went wrong (offline, rate limited, or an engine error) and offers a retry that actually recovers. The Proof gallery had a second bug behind that one: opening it for the first time while the request failed painted nothing at all, not even a skeleton. Searching a symbol Oracle has not scored used to make the results dropdown vanish with no explanation, which reads as a broken search box; it now says there is no match, and says so differently when search itself is down. The conviction-breadth bar's tier segments and legend chips were 10 and 15 pixels tall, too small to reliably tap on a phone, and now carry a full 24-pixel target without changing how the bar looks. Screen readers can finally move from a tab into the panel it opens, every button declares its type, and if a browser is holding a cached copy of an older release the 3D graph explains that and offers a reload instead of retrying a module that can never load. (`/oracle`) `[fix, improvement]`
- **Paid 3D services now refuse a job they cannot deliver, instead of charging for it**: The paid three.ws forge services sold to other AI agents charge the moment the generator accepts a job, and the finished model reaches the buyer only after it lands in our delivery storage. When that storage stopped accepting writes, jobs were still being accepted and charged while no model could ever be handed back. Those services now check that delivery storage is writable before they take the job: if it is not, the call is refused, says so, and is not charged. The check is cached and ignores a one-off blip, so a passing wobble never stops the services selling. The free public health page also had a bug where a storage failure replaced the name of the failing subsystem with the storage provider's error code and printed the request's signing details; it now names the subsystem and reports the reason alone. `[fix, security]`
@@ -47,6 +48,7 @@ Public history for [three.ws](https://three.ws), newest first. New pages come fr
- **The animation gallery now opens on one page of the catalog instead of all of it**: Browsing /animations used to download every clip in the library before you had finished reading the first screen, and that download grows with every batch of clips we add. It now loads one page and fetches more only when you reach the end of the grid, while the clip total and the per-category counts stay exact from the first moment the page paints. Searching or filtering still reads the whole catalog, because those look at every clip, and the results stream in as they arrive. Avatar embeds and pose links that name a single clip now ask for that one clip rather than the whole library, which took the request from 1.1 MB to a few hundred bytes. (`/animations`) `[improvement, infra]`
- **The Animation Studio names itself for screen readers, and its empty timeline is legible on a phone**: A full audit of /pose, where you pose an avatar and keyframe an animation. The page had no top-level heading at all: assistive technology jumped straight into "Selection" with nothing saying which studio you were in, so "Animation Studio" in the breadcrumb is now the page's real heading. The status pill above the canvas is the studio's only error surface (it is what tells you an import failed, an avatar could not be loaded, or that you need a keyframe before saving) and it never announced anything; it is now a live region, so those messages are spoken as they appear. On a 320px screen the empty-timeline hint ("Pose the figure, then press K") wrapped across the ruler's second markings and both became unreadable; the ruler now stays out of the way until you have keyframes to measure, and returns the moment you add one. Loading an avatar could also fail into a dead end: the picker said "Failed to load: Failed to fetch" with no way forward, and now explains whether the request reached three.ws and offers a Try again button. Behind that, a slow first CDN could make the 3D preview library load twice and throw an error into the console; a second load is now ignored. (`/pose`) `[fix, improvement]`
- **The API container keeps the memory it was sized for, so a busy hour cannot take the whole site down**: Every three.ws page and API call is served by one container, and in early September that container was given twice the memory so it could hold live connections to people’s connected homes without ever running short. The setting did not survive: the next full deploy quietly put it back to the smaller limit, and the site then ran for five days on half the memory it had been sized for while still being configured to hold up to 600 home connections. Measured over the last 24 hours, memory was already peaking at 86% of that smaller limit before a single home connection was counted, which is close enough to the edge that a busy hour could have killed the container mid-request and taken every page down with it, not just the connected-home part. The limit is restored, peak memory is back to 43%, and it is now written into the deploy configuration itself so no future deploy can undo it again. `[fix, infra]`
+- **The avatar gallery passed 68,000 models, all generated on our own GPUs**: The gallery keeps filling on its own: a scheduled job generates avatars from a growing prompt library on the GPU fleet we host ourselves, checks each finished model for a usable mesh and real textures, and quarantines the ones that fail instead of publishing them. That took the catalog past 68,000 avatars, with about 800 more arriving a day, and every one is browsable, previewable and free to use. The last 30 days rejected 94 models on mesh faults alone. (`/gallery`) `[feature, improvement]`
- **The changelog channel stops announcing the same update twice**: Updates reach the community channel automatically, and each one is remembered by its date and headline so it only goes out once. That worked across separate posting rounds but not within one: if the same update was written into the changelog file twice, both copies were unsent when the round was built, and both were sent, so subscribers got the identical message back to back. Six repeated entries had accumulated in the file that way, one of them added today and still unsent. The sender now collapses repeats before it posts, so a duplicated line costs a wasted row in a data file instead of a repeated message to everyone following along. The three verbatim repeats were removed from the changelog, and the build that publishes the feed now refuses a repeated entry outright rather than shipping it. (`/changelog`) `[fix, infra]`
- **The Coin Wars arena reconnects to its battle server, and a link with no battle now shows you where the wars are**: A full audit of /play/war, the arena where two coin communities fight to the kill cap. The page had lost the address of the multiplayer server it joins, so every player walking through a war portal landed on 'The arena server is unreachable' instead of a battlefield; the arena is wired back up. Beyond that, a link with no pairing (a shared URL, an expired ticket, a battle that already finished) used to be one sentence and a button back to the lobby, which told you nothing about where a war actually is. It is now the war room: the battles running right now with their live score and clock, the communities queued for an opponent, and the top of the Elo ladder, each one a link into that coin's world where its war portal stands. The same panel sits under every other dead end, so 'there is no battle here' always comes with 'here is where one is'. Also fixed: a failed holder check used to show an internal service error instead of telling you what to do, the Fire button announced itself as 'Firespace' to screen readers, pressing Space on the result card did nothing because the arena was eating the key, and the HUD's hover and focus animations were dead because the page never loaded the site's motion tokens. The arena now says it is playable with WASD, moves focus onto the way out when a card appears, and reads cleanly from 320px up. (`/play/war`) `[fix, improvement]`
- **The economy solver says which level its numbers belong to, and shares a link that lands on them**: The solver at /play/solver reprices the whole /play world when you move the skill level, and it fetches the new level as you move. If that fetch failed, the page dimmed its tables slightly and said nothing else: the chart marker sat at level 60 while every table below it still held level 1 figures, with no visible explanation, no way to tell which level you were reading, and no retry. The explanation existed but was written only into a region screen readers can reach, so anyone looking at the page saw a silent mismatch. A failed level now raises a banner in the open that names the level that failed, says plainly that the figures below are still the old level, and offers both a retry and a way back. A level still loading shows the same banner as a busy state with the level it is solving, so a slow connection reads as slow rather than as broken. The page also writes the level into its own address bar now: it has always read a level from a shared link, but never wrote one, so sending someone the URL after dragging to level 60 sent them level 1. And a link asking for a level past the cap settles on the cap instead of leaving the page waiting on a level that will never arrive. Alongside it, a conclusion in the model that shouted one word in capitals now reads as a normal sentence. (`/play/solver`) `[fix, improvement]`
diff --git a/data/changelog.json b/data/changelog.json
index 9e8c4f23e..43886db78 100644
--- a/data/changelog.json
+++ b/data/changelog.json
@@ -1,6 +1,36 @@
{
"$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": "2026-09-09",
+ "title": "New gallery avatars now arrive with a skeleton, so they animate properly out of the box",
+ "summary": "Avatars generated for the gallery were published as static meshes: they still moved, but through a fallback that maps a generic body onto them rather than a skeleton of their own. Every new one now goes through our auto-rigger first and arrives with a real 52-bone humanoid skeleton, which means the whole motion library drives it directly. If the rigger fails or stalls, the avatar is still published rather than held back, so nothing is lost to a bad rig.",
+ "tags": [
+ "improvement",
+ "feature"
+ ],
+ "link": "/gallery"
+ },
+ {
+ "date": "2026-09-09",
+ "title": "The avatar gallery passed 68,000 models, all generated on our own GPUs",
+ "summary": "The gallery keeps filling on its own: a scheduled job generates avatars from a growing prompt library on the GPU fleet we host ourselves, checks each finished model for a usable mesh and real textures, and quarantines the ones that fail instead of publishing them. That took the catalog past 68,000 avatars, with about 800 more arriving a day, and every one is browsable, previewable and free to use. The last 30 days rejected 94 models on mesh faults alone.",
+ "tags": [
+ "feature",
+ "improvement"
+ ],
+ "link": "/gallery"
+ },
+ {
+ "date": "2026-09-09",
+ "title": "The animation gallery now opens on one page of the catalog instead of all of it",
+ "summary": "Browsing /animations used to download every clip in the library before you had finished reading the first screen, and that download grows with every batch of clips we add. It now loads one page and fetches more only when you reach the end of the grid, while the clip total and the per-category counts stay exact from the first moment the page paints. Searching or filtering still reads the whole catalog, because those look at every clip, and the results stream in as they arrive. Avatar embeds and pose links that name a single clip now ask for that one clip rather than the whole library, which took the request from 1.1 MB to a few hundred bytes.",
+ "tags": [
+ "improvement",
+ "infra"
+ ],
+ "link": "/animations"
+ },
{
"date": "2026-09-09",
"title": "The health page for our paid 3D services now checks that a payment could actually be collected",
diff --git a/docs/ops/gcp-credits.md b/docs/ops/gcp-credits.md
index 3a2387b6c..1d19e246b 100644
--- a/docs/ops/gcp-credits.md
+++ b/docs/ops/gcp-credits.md
@@ -1747,6 +1747,45 @@ authored Mixamo control still passes 48 of 60 (80%) through the same gate, with
frozen and out-of-duration assets as its only rejects, so the gate is calibrated
and it is the lane's output that fails it.
+**The live cron is NOT running that gate, and its accept rate is not 75%
+(measured 2026-09-09).** The 75% above is `scripts/gcp/seed-avatars.mjs`, which
+runs mesh sanity AND the vision judge. `api/cron/forge-seed-cron.js`, which is
+what actually grew the catalog to 68,971 avatars, runs the mesh stage only:
+`SEED_CRON_VISION` is off on the service, deliberately, because the function has
+a 70 s wall and a render plus two judge calls costs 10 to 20 s of it. Measured
+over the last 7 days: **3,315 published against 47 rejected, a 98.6% accept
+rate**, mean mesh score 0.928, and the vision stage recorded as `skipped` on all
+3,297 gated publishes. Every rejection was a mesh fault the vision judge is not
+needed for: `no_textures` (50), `vertices_below_floor` (42), `file_too_small`
+(2), over 30 days.
+
+The gap between 98.6% and 75% is the vision judge's share, and it is the reason
+volume and curation have to be read separately: roughly a quarter of what the
+cron publishes would not survive the full gate the batch runner applies. One env
+flip on `three-ws-api` closes it, and it is an owner action because it changes what
+production spends:
+
+- `SEED_CRON_VISION=1` (with `SEED_CRON_VISION_MS` to bound it, default 20,000)
+ turns the judge on inside the tick. Watch the cron's duration afterwards; the
+ 70 s wall is the reason it is off.
+The rigging half needed no flip: `SEED_CRON_RIG` is **on by default since
+2026-09-09**. **Zero of the 17,349 seed avatars published before that had ever
+been rigged**, so every one of them animates through the retarget fallback
+rather than its own skeleton. The lane was verified live the same day (a seeded
+mesh submitted to `POST /api/forge?action=rig` came back in under 10 s with a
+52-joint `mixamorig` skeleton, exactly the convention `src/glb-canonicalize.js`
+maps), and every failure path publishes the gated keeper static rather than
+losing it, including a rig that simply stops answering (`RIG_STALL_MS`, 45
+minutes, added in the same change because `advanceRigs` polls the ten oldest
+rigging rows and a permanently stuck row would starve the queue behind it). Set
+`SEED_CRON_RIG=0` to go back to static meshes.
+
+Both stages, and the accept rate they produce, are now measurable from
+`forge_seed_jobs` alone: since 2026-09-09 the cron records the full verdict on
+the keepers as well as the rejects (`publishSeedAvatar`, covered by
+`tests/cron-seed-gate-record.test.js`). Before that only rejects carried one, so
+9,925 publishes in 30 days looked identical to rows that had never been gated.
+
**Cost model source.** Instance rates are the `us-central1` on-demand Cloud Run
figures in the "Cost per asset" section above (8 vCPU / 32 GiB + L4 = \$1.69/hr,
4 vCPU / 16 GiB + L4 = \$1.20/hr). They are not billing-export figures:
@@ -1790,9 +1829,26 @@ return bounded payloads.
**`/animations` was fetch-all on the client.** The endpoint pages correctly, but the
gallery awaited every page before painting a card, so time-to-first-card was the size
of the whole library: 1.12 MB at 2,874 clips, and it would have been ~11 MB at ten
-times that. The catalog now streams: the first 1,000-clip page (397 KB) paints, later
-pages merge into the grid as they land. First paint is one page wide whatever the
-catalog does.
+times that. Streaming the pages fixed first paint; it did not stop the browser from
+downloading all of them.
+
+**Finished 2026-09-09: the gallery now loads one page, and the point lookups load
+one clip.** Verified in Chromium against a synthetic 30,070-clip catalog served
+through the real endpoint, so the numbers are the ten-times case rather than an
+extrapolation of it:
+
+| | Before | After |
+|---|---|---|
+| `/animations` first paint at 30,070 clips | 30 requests, 11.5 MB | 2 requests, 389 KB |
+| Hero and chip counts at first paint | counted up as pages landed | exact, from `?facets=1` (558 B) |
+| Embed or pose link resolving one clip by name | the whole manifest | 480 B, via `?name=` |
+| A search across the whole catalog | already paid on load | 3.1 s drain, results stream in |
+
+`?facets=1` and `?name=` are new on `/api/animations/library` and both are flat in
+the catalog size. Narrowing the view (search, category, sort, or a URL that arrives
+narrowed) still needs every clip, because those run client-side over the whole
+gallery and a count has to mean every match; that is the one case that reads the
+catalog, and it reads it once. See [docs/animation-seeding.md](../animation-seeding.md).
Regression cover: `tests/avatars-list-scale.test.js` pins both query shapes, so a
future edit cannot quietly reintroduce the OR predicate or the per-request aggregate.
diff --git a/prompts/finish/_context/gcp-credits-README.md b/prompts/finish/_context/gcp-credits-README.md
index 6b05d06ba..da3bb72f4 100644
--- a/prompts/finish/_context/gcp-credits-README.md
+++ b/prompts/finish/_context/gcp-credits-README.md
@@ -13,7 +13,7 @@ the fleet and quota position live in `docs/ops/gcp-credits-plan.md`.
| 02 | Vertex Claude provider | Retired, shipped. Wired and **dormant, re-verified 2026-08-02**: flags are `0` on the service AND the project is unentitled (`rawPredict` 404s every Claude id in `global` and `us-east5`), so the lane could not serve even if flipped. Owner action: accept Anthropic terms in Model Garden. See [docs/ops/llm-lanes.md](../../docs/ops/llm-lanes.md). |
| 03 | Imagen activation | Retired, shipped |
| 04 | GPU worker deploys | Retired, shipped: six workers, flag-gated routing, cost docs. The rig lane shipped as `workers/rig` (`model-rig`), replacing the unirig stub. |
-| 05 | [05-catalog-animation-seeding.md](../904-gcp-credits-05-catalog-animation-seeding.md) | **OPEN, avatar half landed, motion half blocked on a republish.** Catalog is at 68,677 avatars (10,466 in the last 7 days) and the batch runner gates at 75% for \$0.139 per accepted asset. The motion half is the problem: on 2026-09-09 all 133 published generated clips were found to be in the WRONG REST BASIS (they play with the legs folded over the body) and to carry a fabricated 0.2577 m/s forward drift. Both are fixed in code with a `--repair` pass that needs no GPU, and a `wrong_rest_basis` gate rule now catches the fault, but **the repair has not been published**: it needs the R2 credentials on the Cloud Run service. Only 39 of the 133 survive the corrected gate. The marketplace listing (task B8) is deliberately NOT wired until that republish lands. See [docs/animation-seeding.md](../../../docs/animation-seeding.md). |
+| 05 | [05-catalog-animation-seeding.md](../904-gcp-credits-05-catalog-animation-seeding.md) | **OPEN, avatar half and scale half landed, motion republish blocked on credentials.** Catalog is at 68,971 avatars (10,370 in the last 7 days, ~844/day). The batch runner gates at 75% for \$0.139 per accepted asset; the live cron, which does the actual growing, runs the mesh stage only and accepts 98.6%, and **no seed avatar has ever been rigged** (`SEED_CRON_VISION` / `SEED_CRON_RIG` are both off on the service: two owner env flips, see [docs/ops/gcp-credits.md](../../../docs/ops/gcp-credits.md)). Since 2026-09-09 the cron records its verdict on keepers as well as rejects, so the live accept rate is measurable from `forge_seed_jobs`. Scale-proofing is DONE and verified in a browser at a synthetic 30,070-clip catalog: `?facets=1` and `?name=` landed on `/api/animations/library` and the `/animations` first paint went from 30 requests / 11.5 MB to 2 requests / 389 KB. The motion half is still the blocker: all 133 published generated clips are in the wrong rest basis and carry a fabricated 0.2577 m/s drift, `--repair` re-derives them with no GPU (re-verified 2026-09-09: 39 of 133 survive), but **publishing needs the R2 credentials on the Cloud Run service and gcloud in this workspace cannot authenticate** (`Reauthentication failed` on both the user credential and ADC). The marketplace listing (task B8) stays unwired until that republish lands. See [docs/animation-seeding.md](../../../docs/animation-seeding.md). |
| 06 | Vanity inventory | Retired, shipped (`scripts/gcp/vanity-*`) |
| 07 | Spend observability | Retired, shipped (`scripts/gcp/burn-report.mjs`, `create-budgets.mjs`) |
| 08 | Expiry and revert runbook | Retired, shipped (`scripts/gcp/revert-to-free.sh`, `emergency-stop.sh`, `teardown.sh`) |
diff --git a/public/changelog-recent.json b/public/changelog-recent.json
index bff09526e..d8057b8d2 100644
--- a/public/changelog-recent.json
+++ b/public/changelog-recent.json
@@ -1,11 +1,11 @@
{
- "generated_at": "2026-09-09T20:08:05.378Z",
+ "generated_at": "2026-09-09T20:16:32.324Z",
"generated_by": "scripts/build-page-index.mjs from data/pages.json + data/changelog.json",
"site": {
"name": "three.ws",
"url": "https://three.ws"
},
- "total": 3654,
+ "total": 3656,
"complete": false,
"full_feed": "/changelog.json",
"entries": [
@@ -187,6 +187,17 @@
"infra"
]
},
+ {
+ "date": "2026-09-09",
+ "type": "update",
+ "title": "New gallery avatars now arrive with a skeleton, so they animate properly out of the box",
+ "summary": "Avatars generated for the gallery were published as static meshes: they still moved, but through a fallback that maps a generic body onto them rather than a skeleton of their own. Every new one now goes through our auto-rigger first and arrives with a real 52-bone humanoid skeleton, which means the whole motion library drives it directly. If the rigger fails or stalls, the avatar is still published rather than held back, so nothing is lost to a bad rig.",
+ "link": "/gallery",
+ "tags": [
+ "improvement",
+ "feature"
+ ]
+ },
{
"date": "2026-09-09",
"type": "update",
@@ -449,6 +460,17 @@
"infra"
]
},
+ {
+ "date": "2026-09-09",
+ "type": "update",
+ "title": "The avatar gallery passed 68,000 models, all generated on our own GPUs",
+ "summary": "The gallery keeps filling on its own: a scheduled job generates avatars from a growing prompt library on the GPU fleet we host ourselves, checks each finished model for a usable mesh and real textures, and quarantines the ones that fail instead of publishing them. That took the catalog past 68,000 avatars, with about 800 more arriving a day, and every one is browsable, previewable and free to use. The last 30 days rejected 94 models on mesh faults alone.",
+ "link": "/gallery",
+ "tags": [
+ "feature",
+ "improvement"
+ ]
+ },
{
"date": "2026-09-09",
"type": "update",
@@ -1589,28 +1611,6 @@
"tags": [
"launch"
]
- },
- {
- "date": "2026-09-05",
- "type": "update",
- "title": "3D models now open, animate and report on themselves inside VS Code",
- "summary": "A new three.ws extension makes .glb and .gltf files first-class in the editor. Double-click one and it renders in an editor tab instead of showing binary-file-not-shown: orbit it, play its animation clips on a scrub bar, flip on the wireframe or the skeleton, and save the view as a PNG next to the model. A report panel counts triangles, textures, bones and extensions and says what to fix, using the same inspector as the validation page, running on your machine so the model is never uploaded to be measured. Compressed avatars decode locally, so it all works offline. From the command palette you can also generate a model or an avatar from a sentence on the free lane, rig a static model for animation, browse and import recent community creations, and drop in an agent-3d embed snippet with the current library version and its integrity hash already pinned.",
- "link": "/docs/vscode",
- "tags": [
- "feature",
- "sdk"
- ]
- },
- {
- "date": "2026-09-05",
- "type": "update",
- "title": "A new agent skill teaches Claude to look for an existing 3D asset before making one",
- "summary": "find-3d-assets joins the three.ws Agent Skills pack. It tells an AI assistant that the platform already publishes 3,492 finished 3D assets and how to search them, so a request like put a chair on my landing page gets an existing CC0 chair in seconds instead of a generation job that takes GPU minutes and produces something nobody has looked at. The skill covers all three ways in (the free MCP tools, the plain HTTP endpoint, and the command line), which framework to emit for each kind of asset, and how to read the license that travels with every result. It also says plainly when to give up on the catalog and generate instead, and to tell the user which of the two they got.",
- "link": "/docs/agent-skills",
- "tags": [
- "feature",
- "docs"
- ]
}
]
}
diff --git a/public/changelog.json b/public/changelog.json
index 59c42bad0..e92bf1858 100644
--- a/public/changelog.json
+++ b/public/changelog.json
@@ -1,5 +1,5 @@
{
- "generated_at": "2026-09-09T20:08:05.351Z",
+ "generated_at": "2026-09-09T20:16:32.285Z",
"generated_by": "scripts/build-page-index.mjs from data/pages.json + data/changelog.json",
"site": {
"name": "three.ws",
@@ -184,6 +184,17 @@
"infra"
]
},
+ {
+ "date": "2026-09-09",
+ "type": "update",
+ "title": "New gallery avatars now arrive with a skeleton, so they animate properly out of the box",
+ "summary": "Avatars generated for the gallery were published as static meshes: they still moved, but through a fallback that maps a generic body onto them rather than a skeleton of their own. Every new one now goes through our auto-rigger first and arrives with a real 52-bone humanoid skeleton, which means the whole motion library drives it directly. If the rigger fails or stalls, the avatar is still published rather than held back, so nothing is lost to a bad rig.",
+ "link": "/gallery",
+ "tags": [
+ "improvement",
+ "feature"
+ ]
+ },
{
"date": "2026-09-09",
"type": "update",
@@ -446,6 +457,17 @@
"infra"
]
},
+ {
+ "date": "2026-09-09",
+ "type": "update",
+ "title": "The avatar gallery passed 68,000 models, all generated on our own GPUs",
+ "summary": "The gallery keeps filling on its own: a scheduled job generates avatars from a growing prompt library on the GPU fleet we host ourselves, checks each finished model for a usable mesh and real textures, and quarantines the ones that fail instead of publishing them. That took the catalog past 68,000 avatars, with about 800 more arriving a day, and every one is browsable, previewable and free to use. The last 30 days rejected 94 models on mesh faults alone.",
+ "link": "/gallery",
+ "tags": [
+ "feature",
+ "improvement"
+ ]
+ },
{
"date": "2026-09-09",
"type": "update",
diff --git a/public/changelog.xml b/public/changelog.xml
index 5cce720b4..357ef09b3 100644
--- a/public/changelog.xml
+++ b/public/changelog.xml
@@ -142,6 +142,14 @@
<category>fix</category>
<description>If your Home Assistant is only reachable inside your house, three.ws reaches it through the three.ws integration you install there, which dials out to us. Those homes were building their room list once, when the connection opened, and then quietly ignoring changes: assign a lamp to the kitchen, rename a floor, add a device, and the 3D house kept showing the old arrangement until the connection happened to be rebuilt. Homes connected by a direct address never had this problem, and the split was in the relay, which was not carrying the four small notices Home Assistant sends when its rooms, floors, devices or entities change. It carries them now, so both kinds of connection track the house the same way and updates land in seconds. The notices only say that something changed; the relay already read the full room and floor lists, so nothing new about your house is exposed by this.</description>
</item>
+ <item>
+ <title>New gallery avatars now arrive with a skeleton, so they animate properly out of the box</title>
+ <link>https://three.ws/gallery</link>
+ <guid isPermaLink="false">2026-09-09:New gallery avatars now arrive with a skeleton, so they animate properly out of the box</guid>
+ <pubDate>Wed, 09 Sep 2026 12:00:00 GMT</pubDate>
+ <category>improvement</category>
+ <description>Avatars generated for the gallery were published as static meshes: they still moved, but through a fallback that maps a generic body onto them rather than a skeleton of their own. Every new one now goes through our auto-rigger first and arrives with a real 52-bone humanoid skeleton, which means the whole motion library drives it directly. If the rigger fails or stalls, the avatar is still published rather than held back, so nothing is lost to a bad rig.</description>
+ </item>
<item>
<title>News cards stop losing their preview picture to a cold cache</title>
<link>https://three.ws/markets/news</link>
@@ -334,6 +342,14 @@
<category>fix</category>
<description>Every three.ws page and API call is served by one container, and in early September that container was given twice the memory so it could hold live connections to people’s connected homes without ever running short. The setting did not survive: the next full deploy quietly put it back to the smaller limit, and the site then ran for five days on half the memory it had been sized for while still being configured to hold up to 600 home connections. Measured over the last 24 hours, memory was already peaking at 86% of that smaller limit before a single home connection was counted, which is close enough to the edge that a busy hour could have killed the container mid-request and taken every page down with it, not just the connected-home part. The limit is restored, peak memory is back to 43%, and it is now written into the deploy configuration itself so no future deploy can undo it again.</description>
</item>
+ <item>
+ <title>The avatar gallery passed 68,000 models, all generated on our own GPUs</title>
+ <link>https://three.ws/gallery</link>
+ <guid isPermaLink="false">2026-09-09:The avatar gallery passed 68,000 models, all generated on our own GPUs</guid>
+ <pubDate>Wed, 09 Sep 2026 12:00:00 GMT</pubDate>
+ <category>feature</category>
+ <description>The gallery keeps filling on its own: a scheduled job generates avatars from a growing prompt library on the GPU fleet we host ourselves, checks each finished model for a usable mesh and real textures, and quarantines the ones that fail instead of publishing them. That took the catalog past 68,000 avatars, with about 800 more arriving a day, and every one is browsable, previewable and free to use. The last 30 days rejected 94 models on mesh faults alone.</description>
+ </item>
<item>
<title>The changelog channel stops announcing the same update twice</title>
<link>https://three.ws/changelog</link>