Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.
JavaScript 3d-generationai-agentsanimationavatardigital-humangenerative-aiglbgltfimage-to-3dllmmcpmcp-servermodel-context-protocolsolanatext-to-3dthreejsweb-componentswebglwebxrx402
docs(changelog): publish the studio countdown fix and map the voice surface
nirholased46aee2155d
diff --git a/CHANGELOG.md b/CHANGELOG.mdindex 33df81d10..8c3b0955c 100644--- a/CHANGELOG.md+++ b/CHANGELOG.md@@ -37,6 +37,7 @@ Public history for [three.ws](https://three.ws), newest first. New pages come fr- **Smart Home: the connection pool now sizes itself to the server it is running on**: The number of houses one server holds open at once was set by hand, in a different place from the memory that has to hold them, and the two drifted apart three times. Left alone, that ends with a server running out of memory and taking the whole API down with it, which every part of three.ws would feel, not just Smart Home. The pool now reads the memory its own container actually has and refuses to hold more houses than fit, saying so in the health endpoint when it does. When there is less room it connects fewer houses per server and keeps working, instead of risking an outage for everybody. (`/docs/smart-home`) `[infra, fix]`- **The "have you tried" card no longer hangs off the side of the screen on a phone**: The small suggestion card that appears in the bottom corner of the site was drawn 34 pixels wider than the space it asked for, because its padding and border were being added on top of its width instead of counted inside it. On a phone that pushed the card 22 pixels past the left edge of the screen, so its kicker, title and description were all cut off mid-word and the button inside it ran wider than the card itself. This only happened on the pages that do not load the main site stylesheet, which includes the whole markets family: the screener, coin pages, categories, the heatmap and about thirty others. The card now sits fully on screen with an even gutter on both sides at every width, and its button fits inside it. (`/screener`) `[fix]`- **The 3D Playground recovers when a model will not load, and now works from the keyboard**: Dropping a .glb into /playground had exactly one failure mode: a dimmed screen and a spinner that never stopped. If the file was corrupt, the download stalled, or the 3D viewer script itself was blocked, nothing on the page said so and there was no way back. That spinner is now a shimmer inside the viewer frame instead of a curtain over the whole page, so the controls stay usable while a model streams, and a failure lands in a designed error card that names what went wrong and offers Retry or a file picker. A load that never finishes gives up after 25 seconds rather than hanging forever. The lighting presets no longer strand the scene on the previous preset’s exposure. The page also gained the structure it was missing: a real heading that says what it is for, headings on every control group, an accessible name on each switch, slider and field, a focus ring on everything you can tab to, and an upload button the keyboard can actually reach (it was hidden with display:none, which drops it out of the tab order entirely). On a phone the top navigation scrolls instead of vanishing. Animation names read out of an uploaded file are rendered as text, never as markup. (`/playground`) `[fix, improvement]`+- **The 3D studio stops promising "5 seconds to go" for eleven minutes**: Every surface that waits on a generation, the ChatGPT widget and the custom GPT included, showed a countdown built from the lane's own time estimate. Once a model outran that estimate the countdown stopped at five seconds and stayed there for the rest of the job: on a generation measured this week that meant reading "roughly 5s to go" for the final eleven minutes. The countdown now runs only while there is genuinely time left on the estimate, and past it the wait says the job keeps running in the background and shows how long it has actually been going, which is a number that moves. The inline 3D viewer in ChatGPT also now loads its model through three.ws itself rather than straight from the asset store, so it renders from any host that embeds it instead of only from our own pages. (`/forge`) `[fix, improvement]`- **The agent-wallet demo pays again when an endpoint asks the buyer to cover its own network fee**: On /play/agent-wallet a 3D avatar walks up to a paid x402 endpoint and buys a live market signal with real USDC on Solana. The page had gone quiet. three.ws's own paid endpoints switch to a self-pay payment challenge whenever the wallet that normally sponsors the Solana network fee cannot co-sign, and the bridge behind the page only knew how to build a sponsored payment, so every price quote came back as a server error and the demo never got past its connecting state. The bridge now signs both kinds: when no sponsor is offered, the agent wallet pays its own network fee, exactly as the platform's own payment ring already does. Four smaller things came with it. The page no longer offers a Pay button it knows will fail: when the agent wallet cannot cover the quoted price it says so and leaves the deposit address in front of you. Switching between BTC, ETH and SOL re-prices immediately instead of waiting for the next payment. The wallet balance and the Pay button stop reverting to placeholder text a few seconds after they load, which the translation pass had been overwriting. And the page carries a real page heading for screen readers and search engines, which it had been missing entirely. (`/play/agent-wallet`) `[fix, improvement]`- **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]`diff --git a/STRUCTURE.md b/STRUCTURE.mdindex 9108da488..f09d03d2b 100644--- a/STRUCTURE.md+++ b/STRUCTURE.md@@ -69,7 +69,7 @@ bottom.| Notification bell & preferences | Event vocabulary [api/_lib/feed.js](api/_lib/feed.js) (`USER_EVENT_TYPES` → `publishUserEvent()` → `user_notifications`), endpoints [api/notifications/](api/notifications), preference matrix [api/_lib/notify-prefs.js](api/_lib/notify-prefs.js), bell client [src/notifications.js](src/notifications.js), avatar announcements [src/notification-herald.js](src/notification-herald.js), settings panel [src/dashboard-next/pages/settings.js](src/dashboard-next/pages/settings.js) | Live | In-app inbox covering remixes ([api/x402/remix-asset.js](api/x402/remix-asset.js)), DMs ([api/friends/messages.js](api/friends/messages.js)), coin bonding-curve graduations (pump cron in [api/cron/[name].js](api/cron/%5Bname%5D.js)), follows, sales, purchases, IRL, market, and account events. Preferences are a category × channel matrix (`in_app` always on; push/email/telegram/avatar are per-category opt-outs) edited from `/dashboard/settings`. **The `avatar` channel delivers in person:** for a fresh, unread notification in an opted-in category the corner companion walks on screen, gestures, and says the line with a link straight to it, summoning itself for that one message and retreating if it was off ([walk-sdk](walk-sdk) `control.announce`). Rules (15-minute freshness, once per browser, two per batch, visible tab only, device-level mute) live in [src/notification-herald.js](src/notification-herald.js). Docs: [docs/notifications.md](docs/notifications.md), [docs/social-layer.md](docs/social-layer.md). Tests: [tests/notification-herald.test.js](tests/notification-herald.test.js), [tests/notify-prefs.test.js](tests/notify-prefs.test.js) || Personal companion (`/companion`) | Page [pages/companion.html](pages/companion.html) `→ /companion` + [src/companion/](src/companion) (page, source cards, voice), API [api/companion/](api/companion) (settings, sources, contacts, events, `ingest`, `stream`, `poll`), engine [api/_lib/companion/](api/_lib/companion) (store, encrypted BYOK configs, triage, lanes for Telegram / iCal / IMAP), cron [api/cron/companion-poll.js](api/cron/companion-poll.js), migration [api/_lib/migrations/20260828140000_companion_harness.sql](api/_lib/migrations/20260828140000_companion_harness.sql) | Live | Turns a person's own inbox, chats, calendar and phone into deliveries a 3D character makes **in person**. BYOK per source (a Telegram bot the user made, a private iCal URL, IMAP with an app password) plus a universal bridge token anything can POST to. Every message is scored 0-100 by the shared deterministic rules ([@three-ws/companion/triage](packages/companion-sdk/src/triage-rules.js), the same module the CLI runs locally) and optionally refined by the user's own Anthropic key; above their threshold and outside quiet hours it is delivered through `companion_delivery` notifications, so the bell, Web Push, and the avatar channel all carry it. A contact's message is spoken by **that contact's avatar and voice** (`control.announce({ avatar })` in [walk-sdk](walk-sdk)). A delivery whose lane can carry an answer (Telegram, through the user's own bot) offers a reply box in place, and the reply is stored with the message. Live SSE at `/api/companion/stream` feeds the desktop app and any third-party body. Docs: [docs/companion.md](docs/companion.md). Tests: [tests/companion-triage.test.js](tests/companion-triage.test.js), [tests/companion-calendar.test.js](tests/companion-calendar.test.js), [packages/companion-sdk/tests/](packages/companion-sdk/tests) || three.ws Drive: the agent in the car (`/drive`) | Page [pages/drive.html](pages/drive.html) `→ /drive` + [src/drive/](src/drive) (page controller, surface presets, native bridge, local commands, motion gate); native CarPlay scene [ios/native/App/App/CarPlaySceneDelegate.swift](ios/native/App/App/CarPlaySceneDelegate.swift) + [DriveLink.swift](ios/native/App/App/DriveLink.swift); Android Auto car app [solana-mobile/android-overlay/app/src/main/java/ws/three/app/drive/](solana-mobile/android-overlay/app/src/main/java/ws/three/app/drive) | Live (web) / built and compiled (Android Auto) / built, entitlement pending (CarPlay) | The existing talk loop (`/api/asr` → `/api/chat` → `/api/tts` → lipsync on a real rigged avatar) rebuilt for a car panel: one glanceable two-line reply instead of a transcript, exactly four controls, night by default, and a keyboard that hides itself while the wheels turn. Speed comes from `navigator.geolocation`'s real ground speed, so moving above walking pace locks text entry back and a sustained stop is what unlocks it. Control phrases ("repeat that", "stop talking", "louder", "night mode") are answered locally with no round trip, matched against the whole utterance so "stop at the next charger" still reaches the agent. Hands free ends a turn from the real captured mic level on the Riva lane. **The car screen is never ours to draw on, on either platform:** Apple's voice-based conversational category (iOS 26.4) grants templates and an audio session and no drawing surface, and Android grants a canvas only to navigation apps and has no conversational category at all, so both car screens render the same four rows and this page runs beside them, joined by the versioned `threeWsDrive` channel (a WebKit message handler on iOS, a JavaScript interface on Android). Android Auto declares `androidx.car.app.category.IOT` because that is accurate: the home tools in `api/chat.js` reach a real house, so "turn the porch light on" from the car is the product. Its web view has no window and therefore no animation frames, so `?surface=androidauto` runs the loop as audio with the stage off, and refuses tap-to-approve for guarded actions because a person who cannot see what they are approving must not be asked. A guarded action while moving is refused on every surface and offered for when the car stops. Docs: [docs/carplay.md](docs/carplay.md), [ios/docs/CARPLAY.md](ios/docs/CARPLAY.md), [solana-mobile/docs/ANDROID-AUTO.md](solana-mobile/docs/ANDROID-AUTO.md). Tests: [tests/drive-surface.test.js](tests/drive-surface.test.js) |-| Hands-free voice for the house (`/voice/home`) | Loop [src/voice/home-voice.js](src/voice/home-voice.js), VAD [src/voice/vad.js](src/voice/vad.js), wake word [src/voice/wake-word.js](src/voice/wake-word.js) + catalog [src/voice/wake-words.js](src/voice/wake-words.js), panel [src/voice/home-voice-ui.js](src/voice/home-voice-ui.js) + [public/home-voice.css](public/home-voice.css); page [pages/voice-home.html](pages/voice-home.html) `→ /voice/home` + [src/voice-home.js](src/voice-home.js); models committed under [public/models/voice/wake-word/](public/models/voice/wake-word) with the silero + onnxruntime runtime staged by [scripts/copy-voice-models.mjs](scripts/copy-voice-models.mjs) | Live | The browser half of three.ws Home: mic → silero VAD → openWakeWord → `/api/asr` → `/api/chat` (home tools) → `/api/tts/speak` → lipsync, with one `getUserMedia` shared by all three consumers. Wake-word matching runs entirely in the tab on committed openWakeWord v0.5.1 weights served from three.ws, so nothing is uploaded until after the wake word; nothing about detection is even downloaded until the user opts in. Barge-in cuts playback ~110 ms after the user starts talking, and a playback-aware guard means the agent saying its own wake word (score 0.99) wakes it zero times. A pause is not an ending: a closed segment stays open 450 ms so "Hey Jarvis… turn the kitchen light off" is one command. Guarded actions are spoken in full, shown on screen by name and entity id, and satisfied only by the token `confirm` redeemed at `POST /api/home/:id/confirm`; a general "yeah" is refused by name and a screenless surface is refused outright. Twelve designed states, mute that ends the track rather than hiding a dot, and every leg timed live. Docs: [docs/home-voice.md](docs/home-voice.md). Proof: [scripts/check-home-voice.mjs](scripts/check-home-voice.mjs) (36 assertions in a real Chromium, plus `--live` against a real Home Assistant that proves the spoken command really turns the light off). Tests: [tests/home-voice.test.js](tests/home-voice.test.js) |+| Hands-free voice for the house (`/voice/home`) | Loop [src/voice/home-voice.js](src/voice/home-voice.js), VAD [src/voice/vad.js](src/voice/vad.js), wake word [src/voice/wake-word.js](src/voice/wake-word.js) + catalog [src/voice/wake-words.js](src/voice/wake-words.js), panel [src/voice/home-voice-ui.js](src/voice/home-voice-ui.js) + [public/home-voice.css](public/home-voice.css); page [pages/voice-home.html](pages/voice-home.html) `→ /voice/home` + [src/voice-home.js](src/voice-home.js); models committed under [public/models/voice/wake-word/](public/models/voice/wake-word) with the silero + onnxruntime runtime staged by [scripts/copy-voice-models.mjs](scripts/copy-voice-models.mjs) | Live | The browser half of three.ws Home: mic → silero VAD → openWakeWord → `/api/asr` → `/api/chat` (home tools) → `/api/tts/speak` → lipsync, with one `getUserMedia` shared by all three consumers. Wake-word matching runs entirely in the tab on committed openWakeWord v0.5.1 weights served from three.ws, so nothing is uploaded until after the wake word; nothing about detection is even downloaded until the user opts in. Barge-in cuts playback ~110 ms after the user starts talking, and a playback-aware guard means the agent saying its own wake word (score 0.99) wakes it zero times. A pause is not an ending: a closed segment stays open 450 ms so "Hey Jarvis… turn the kitchen light off" is one command. Guarded actions are spoken in full, shown on screen by name and entity id, and satisfied only by the token `confirm` redeemed at `POST /api/home/:id/confirm`; a general "yeah" is refused by name and a screenless surface is refused outright. Twelve designed states, mute that ends the track rather than hiding a dot, and every leg timed live. Docs: [docs/home-voice.md](docs/home-voice.md). Proof: [scripts/check-home-voice.mjs](scripts/check-home-voice.mjs) (36 assertions in a real Chromium, plus two `--live` scenarios against a real Home Assistant: `live-confirm` mints a real confirmation and proves redeeming it really unlocks a real deadbolt and cannot be replayed, and `live` proves the spoken command really turns the light off wherever a tool-calling provider key is configured, skipping with its reason where none is). Tests: [tests/home-voice.test.js](tests/home-voice.test.js) || Knock: a priced door to a person (`/knock`) | Hub [pages/knock.html](pages/knock.html) `→ /knock` + [src/knock/page.js](src/knock/page.js), public door [pages/knock-door.html](pages/knock-door.html) `→ /knock/<handle>` + [src/knock/door.js](src/knock/door.js), API [api/knock/](api/knock) (door, directory, send, reply, settings, inbox, escrowed, escrow-sync), paid lane [api/x402/knock.js](api/x402/knock.js), escrowed lane [api/_lib/knock/escrow.js](api/_lib/knock/escrow.js) + browser client [src/knock/escrow-program.js](src/knock/escrow-program.js) / [src/knock/escrow-checkout.js](src/knock/escrow-checkout.js), engine [api/_lib/knock/](api/_lib/knock) (pure policy, store, delivery, receipt tokens), migration [api/_lib/migrations/20260828210000_knock_doors.sql](api/_lib/migrations/20260828210000_knock_doors.sql), SDK + CLI [packages/knock-sdk/](packages/knock-sdk) (`@three-ws/knock`), MCP [packages/knock-mcp/](packages/knock-mcp) (`@three-ws/knock-mcp`) | Live | Every account can publish a door at `/knock/<username>` and set what one message from a stranger costs. Paying it buys exactly one message through, and the delivery is **in person**: the knock becomes a `companion_events` row whose importance is derived from the amount paid, so the recipient's corner avatar walks on wherever they are and says who is at the door and what they paid. **The USDC settles straight to the recipient** (`payTo` on the 402 challenge is their own address, and a priced door cannot be opened without one), so there is no platform wallet in the path and no cut. Solana leads; Base is advertised only when the owner sets an address for it. **Every refusal happens before the 402** (shut door, daily cap, block list, length limits) and again before settlement, so a knock that was never going to land is never a knock somebody paid for. Only the subject line is spoken; an untrusted body never uses the recipient's avatar as a speaker. The sender gets a derived HMAC receipt URL and reads the reply through it with no account, which is why answering someone never hands out an address. Free doors take a plain POST on a tight per-IP bucket. A door can also take **escrowed** knocks (row below), which the door page offers beside paying up front and the owner runs from their own wallet in the settings panel; nothing on this server can move an escrowed payment. Spec: [specs/KNOCK_PROTOCOL.md](specs/KNOCK_PROTOCOL.md). Docs: [docs/knock.md](docs/knock.md). Tests: [tests/knock-policy.test.js](tests/knock-policy.test.js), [packages/knock-sdk/test/](packages/knock-sdk/test), [packages/knock-mcp/test/](packages/knock-mcp/test) || Visitor feedback loop | Recorder [@three-ws/witness](packages/witness), capture + panel [src/feedback-companion.js](src/feedback-companion.js) (mounted by [src/walk-companion.js](src/walk-companion.js)), intake [api/feedback/report.js](api/feedback/report.js), queue API [api/feedback/index.js](api/feedback/index.js), scoring [api/_lib/feedback/triage.js](api/_lib/feedback/triage.js), store [api/_lib/feedback/store.js](api/_lib/feedback/store.js), cron [api/cron/feedback-triage.js](api/cron/feedback-triage.js), review page [pages/feedback.html](pages/feedback.html) `→ /feedback` + [src/feedback-page.js](src/feedback-page.js) | Live | The return channel on the corner companion: the visitor tells the avatar what broke and the report carries the route, build SHA, viewport, and the last console errors and failed requests the browser captured, so a vague sentence lands as an actionable row. The companion offers proactively at the one moment worth asking (right after the page actually threw), once per route per session, never in a background tab. **Feedback is data, never an instruction:** the model that reads untrusted visitor text writes six scalars to that report's own row and has no path to the repo, a deploy, or any other write; the admin-only `/feedback` queue groups reports into one row per problem and a person decides. **A report arrives as a runnable failing test:** the recorded session compiles into a Playwright spec asserting the failure is GONE (red until fixed, green after), reachable from the queue's Copy test button, `GET /api/feedback/repro`, `npm run feedback:repro`, and the `get_feedback_repro` MCP tool. Docs: [docs/feedback.md](docs/feedback.md), [docs/witness.md](docs/witness.md). Tests: [tests/feedback-triage.test.js](tests/feedback-triage.test.js) || Session-to-test recorder | [packages/witness/](packages/witness) `→ @three-ws/witness`: recorder [recorder.js](packages/witness/src/recorder.js), selector synthesis [selector.js](packages/witness/src/selector.js), capture-time privacy [redact.js](packages/witness/src/redact.js), spec compiler [compile.js](packages/witness/src/compile.js); compiled repros land in [tests/repros/](tests/repros) with [their own Playwright config](tests/repros/playwright.config.mjs) | Published as `@three-ws/witness` (Apache-2.0, zero deps) | Turns a real user session into a runnable failing test. Records INTENT (about 2 KB) rather than DOM mutations (megabytes), so it can be always-on, which is the point: the trace already exists when somebody decides to complain. Every element gets the strongest handle that is unique in the live document (test id > id > aria-label > role+name > text > name > placeholder > class > structural path), framework-generated ids and utility classes refused, and each trace carries a replay confidence equal to its WEAKEST link. The compiled spec asserts the failure is gone, never that it happened, so it locks in the fix rather than the bug; a session with no machine-visible failure emits `test.fail()` and says which assertion a human must add rather than faking green. Privacy is enforced at capture: typed values are never held (only a length and a shape), password/hidden/email/tel/date inputs and `data-witness="off"` subtrees record nothing, URLs lose their credentials before storage. Docs: [docs/witness.md](docs/witness.md). Tests: [packages/witness/test/](packages/witness/test) (36, `npm run test:witness`) |diff --git a/data/changelog.json b/data/changelog.jsonindex 32a2bfb88..a62a43b52 100644--- a/data/changelog.json+++ b/data/changelog.json@@ -1,6 +1,16 @@{"$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": "The 3D studio stops promising \"5 seconds to go\" for eleven minutes",+ "summary": "Every surface that waits on a generation, the ChatGPT widget and the custom GPT included, showed a countdown built from the lane's own time estimate. Once a model outran that estimate the countdown stopped at five seconds and stayed there for the rest of the job: on a generation measured this week that meant reading \"roughly 5s to go\" for the final eleven minutes. The countdown now runs only while there is genuinely time left on the estimate, and past it the wait says the job keeps running in the background and shows how long it has actually been going, which is a number that moves. The inline 3D viewer in ChatGPT also now loads its model through three.ws itself rather than straight from the asset store, so it renders from any host that embeds it instead of only from our own pages.",+ "tags": [+ "fix",+ "improvement"+ ],+ "link": "/forge"+ },{"date": "2026-09-09","title": "News cards stop losing their preview picture to a cold cache",diff --git a/public/changelog-recent.json b/public/changelog-recent.jsonindex 08a24ea96..7fed5092c 100644--- a/public/changelog-recent.json+++ b/public/changelog-recent.json@@ -1,11 +1,11 @@{- "generated_at": "2026-09-09T15:18:36.394Z",+ "generated_at": "2026-09-09T15:24:45.853Z","generated_by": "scripts/build-page-index.mjs from data/pages.json + data/changelog.json","site": {"name": "three.ws","url": "https://three.ws"},- "total": 3642,+ "total": 3643,"complete": false,"full_feed": "/changelog.json","entries": [@@ -341,6 +341,17 @@"improvement"]},+ {+ "date": "2026-09-09",+ "type": "update",+ "title": "The 3D studio stops promising \"5 seconds to go\" for eleven minutes",+ "summary": "Every surface that waits on a generation, the ChatGPT widget and the custom GPT included, showed a countdown built from the lane's own time estimate. Once a model outran that estimate the countdown stopped at five seconds and stayed there for the rest of the job: on a generation measured this week that meant reading \"roughly 5s to go\" for the final eleven minutes. The countdown now runs only while there is genuinely time left on the estimate, and past it the wait says the job keeps running in the background and shows how long it has actually been going, which is a number that moves. The inline 3D viewer in ChatGPT also now loads its model through three.ws itself rather than straight from the asset store, so it renders from any host that embeds it instead of only from our own pages.",+ "link": "/forge",+ "tags": [+ "fix",+ "improvement"+ ]+ },{"date": "2026-09-09","type": "update",@@ -1600,17 +1611,6 @@"feature","security"]- },- {- "date": "2026-09-04",- "type": "update",- "title": "A second opinion now reviews every trade a sniper agent wants to make",- "summary": "Trading agents already pass a stack of mechanical checks before they buy: the market-cap band, the price-impact ceiling, budget and slot caps, and an on-chain honeypot check that simulates a real buy and a real sell. All of those answer a question with a number. A new Risk Officer asks the one they cannot: knowing everything we know, is this specific trade still a good idea? It is shown the proposed trade plus the agent's own reason for wanting it, and is told to look for what the agent missed, so it argues with the trade instead of agreeing with it. It starts in shadow mode on every strategy: it reviews and records, and changes nothing, so its calls can be scored against what the trades actually did before anyone lets it veto one. When it is armed, a veto has to name a specific fact, it can shrink a trade instead of killing it, it can only ever reduce risk and never raise it, and if the reviewer is unavailable the trade goes ahead unchanged rather than the fleet stalling.",- "link": null,- "tags": [- "feature",- "security"- ]}]}diff --git a/public/changelog.json b/public/changelog.jsonindex 3abf705b7..54e07856b 100644--- a/public/changelog.json+++ b/public/changelog.json@@ -1,5 +1,5 @@{- "generated_at": "2026-09-09T15:18:36.344Z",+ "generated_at": "2026-09-09T15:24:45.795Z","generated_by": "scripts/build-page-index.mjs from data/pages.json + data/changelog.json","site": {"name": "three.ws",@@ -338,6 +338,17 @@"improvement"]},+ {+ "date": "2026-09-09",+ "type": "update",+ "title": "The 3D studio stops promising \"5 seconds to go\" for eleven minutes",+ "summary": "Every surface that waits on a generation, the ChatGPT widget and the custom GPT included, showed a countdown built from the lane's own time estimate. Once a model outran that estimate the countdown stopped at five seconds and stayed there for the rest of the job: on a generation measured this week that meant reading \"roughly 5s to go\" for the final eleven minutes. The countdown now runs only while there is genuinely time left on the estimate, and past it the wait says the job keeps running in the background and shows how long it has actually been going, which is a number that moves. The inline 3D viewer in ChatGPT also now loads its model through three.ws itself rather than straight from the asset store, so it renders from any host that embeds it instead of only from our own pages.",+ "link": "/forge",+ "tags": [+ "fix",+ "improvement"+ ]+ },{"date": "2026-09-09","type": "update",diff --git a/public/changelog.xml b/public/changelog.xmlindex 19fd0eb56..9d3b47c97 100644--- a/public/changelog.xml+++ b/public/changelog.xml@@ -254,6 +254,14 @@<category>fix</category><description>Dropping a .glb into /playground had exactly one failure mode: a dimmed screen and a spinner that never stopped. If the file was corrupt, the download stalled, or the 3D viewer script itself was blocked, nothing on the page said so and there was no way back. That spinner is now a shimmer inside the viewer frame instead of a curtain over the whole page, so the controls stay usable while a model streams, and a failure lands in a designed error card that names what went wrong and offers Retry or a file picker. A load that never finishes gives up after 25 seconds rather than hanging forever. The lighting presets no longer strand the scene on the previous preset’s exposure. The page also gained the structure it was missing: a real heading that says what it is for, headings on every control group, an accessible name on each switch, slider and field, a focus ring on everything you can tab to, and an upload button the keyboard can actually reach (it was hidden with display:none, which drops it out of the tab order entirely). On a phone the top navigation scrolls instead of vanishing. Animation names read out of an uploaded file are rendered as text, never as markup.</description></item>+ <item>+ <title>The 3D studio stops promising "5 seconds to go" for eleven minutes</title>+ <link>https://three.ws/forge</link>+ <guid isPermaLink="false">2026-09-09:The 3D studio stops promising "5 seconds to go" for eleven minutes</guid>+ <pubDate>Wed, 09 Sep 2026 12:00:00 GMT</pubDate>+ <category>fix</category>+ <description>Every surface that waits on a generation, the ChatGPT widget and the custom GPT included, showed a countdown built from the lane's own time estimate. Once a model outran that estimate the countdown stopped at five seconds and stayed there for the rest of the job: on a generation measured this week that meant reading "roughly 5s to go" for the final eleven minutes. The countdown now runs only while there is genuinely time left on the estimate, and past it the wait says the job keeps running in the background and shows how long it has actually been going, which is a number that moves. The inline 3D viewer in ChatGPT also now loads its model through three.ws itself rather than straight from the asset store, so it renders from any host that embeds it instead of only from our own pages.</description>+ </item><item><title>The agent-wallet demo pays again when an endpoint asks the buyer to cover its own network fee</title><link>https://three.ws/play/agent-wallet</link>diff --git a/public/features.json b/public/features.jsonindex ce07536dc..8dd135d1d 100644--- a/public/features.json+++ b/public/features.json@@ -1,5 +1,5 @@{- "generated_at": "2026-09-09T15:18:36.257Z",+ "generated_at": "2026-09-09T15:24:45.754Z","generated_by": "scripts/build-page-index.mjs from data/pages.json","site": {"name": "three.ws",