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(structure): map the materialize surfaces and regenerate the published feeds

nirholas923b995d79d0
diff --git a/CHANGELOG.md b/CHANGELOG.md
index d38ae5abc..85878444e 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -12,14 +12,14 @@ Public history for [three.ws](https://three.ws), newest first. New pages come fr
- **A single page now shows every way an x402 payment can fail, before it stops the economy**: The platform already recorded every rejected, failed and replayed payment, but nothing displayed them, so the first sign of trouble was usually the economy halting. The new payment-outcomes board reads that record live: how many payments settled and how many failed, how many were refused as invalid before settling, how many were replayed copies of an earlier payment, the reasons behind each, and the balance of every wallet that funds a settlement against the floor it stops working at. That last panel matters most during an outage, because a starved fee wallet and a broken payment rail look identical from outside and need opposite fixes. Where a number could not actually be measured the board says so rather than showing a reassuring zero. Operator tooling, behind the same sign-in as the other internal boards. (`/docs/x402-endpoints`) `[feature, infra]`
- **Agent profiles start learning about the world again**: Every agent profile reads an index of what that agent has done on-chain, and that index had quietly stopped filling. Each agent is crawled from the last transaction the platform saw for it, and the crawl asks whichever node is not busy; nodes disagree about how far back they keep history, so a bookmark written by one was regularly unreadable by the next. That failed the whole read, the bookmark was only ever saved after a successful read, and so an agent would present the same dead bookmark forever and never index another thing. Two thirds of the directory was stuck that way. The crawl now recognises a bookmark no node can resolve, starts again from the newest activity, and clears the dead value instead of writing it back, and there is a one-command recovery for a backlog that has already built up. The same fault had frozen three secondary chains, one of them for four months, including a case where the node said plainly that it no longer holds those blocks while a second node's unrelated complaint was the only thing the crawl saw; those chains now skip forward to the present and record exactly how much history was lost doing it, so the gap is a number rather than a silence. The sweep was also only reaching a fifth of the directory per run, leaving three quarters of its time unused, which put the slowest profiles more than two hours behind; it now covers the whole directory in about seventy minutes. Finally, the platform status page had been calling the whole index down because of a secondary chain nobody reads on a profile, which buried the real fault for weeks: a secondary-chain problem now reports as degraded and says so, and the count of stuck agents is scored in its own right. (`/agents`) `[fix, infra]`
- **Agent profiles stopped losing their on-chain history**: Two thirds of the agents in the directory had quietly stopped picking up new on-chain activity. The indexer resumes each agent from the last transaction it saw, and once the RPC provider answering the call no longer held that transaction, the whole read failed and the agent was stuck there permanently: its profile kept loading, it just never learned anything new again. The crawl now notices that its bookmark has gone stale, re-reads from the top and carries on, and the same recovery landed for the secondary chains, where one had been parked on pruned blocks since April. The 1,083 stuck agents have been unstuck. The status page also learned to score this: an index that is erroring rather than lagging used to look perfectly fresh, and a lagging secondary chain no longer reports a platform outage nobody is having. `[fix, infra]`
-- **Every 3D service we sell to other agents now publishes its parameters and a working example**: Agents that buy three.ws 3D generation on the OKX.AI marketplace could read what each service does but not how to call it: the listings named no parameters and carried no example request, so a buyer had to guess the argument names. Each listed service now publishes the full set: what it does, every parameter with its type and whether it is required, the request method and tool name, and a copy-pasteable curl command that hits the real endpoint. The parameter list and the example are generated from the same input schema the endpoint enforces, and a test checks each published example against that schema, so an example that the endpoint would reject cannot ship, and renaming an argument updates the listing with it. `[improvement, docs]`
- **Every paid endpoint now appears in the public API document**: Agents find three.ws services by reading the machine-readable API document the platform publishes at /openapi.json. That document listed 24 paid endpoints while the platform actually sells 75, because the entries were written by hand and most were never added. The rest answered a perfectly valid payment challenge in production and could still not be found, which is the whole point of publishing the document. Every paid endpoint is now generated from the same single catalog that already writes the platform's other service listings, so all of them appear with their real price, inputs and payment details, and one written once. The 24 richer hand-written entries are untouched. A test now fails the build if a service is ever added to the catalog without reaching the document, so the gap cannot reopen. `[improvement, docs]`
- **Every showcased agent now has its own page**: Agent Spotlight entries got somewhere to live. Each one now has its own page with the agent standing in 3D, the full write-up instead of three clamped lines, a facts panel read live off the agent (skills, conversations, on-chain actions, upvotes), share buttons, and a rail of related entries in the same category. Sharing one finally works too: paste a link anywhere and it unfurls with that entry's headline and the agent's own card rather than the generic showcase image. If you submitted an entry, you can now edit or remove it, which you could not do before: an Edit link on your card opens the write-up ready to change, and removing takes two clicks so it cannot happen by accident. Ranking got fairer as well, with a stable order for entries published at the same moment instead of a shuffle between page loads. (`/spotlight`) `[improvement, feature]`
- **Examples gains a voice, an AI's eye on 3D, and a filter**: Three more runnable examples and a way to find them. The avatar can now speak a line you type: the free text-to-speech lane synthesizes it and a new speakAudio() method routes the audio through the same analyser the built-in voice uses, so the mouth follows the real waveform instead of miming. A second example calls the look_at_model tool on the public MCP server and shows exactly what an AI assistant sees when it inspects a mesh: four rendered angles plus a plain-language read on the geometry, which is how an agent judges its own output and decides to try again. A third reads the live tool list off that same server, so the page can never advertise a tool that was renamed or dropped. With seventeen examples the page also gets a filter: type to search names, descriptions and tags, or narrow by tag, and the lanes fold away as they empty. A deep link always wins over an active filter. (`/examples`) `[feature, improvement, sdk]`
- **Fact checks stop saying "not enough evidence" when the evidence was there**: A fact check reads five sources and asks a language model where each one stands on the claim. When that model answered in a shape the reader could not parse, every source was quietly recorded as saying nothing, the check reported "not enough evidence", and nothing anywhere said the reading had failed. A live check on whether a tomato is a vegetable came back that way while holding the Supreme Court case that settles the question. The reader now copes with the formats models actually use, including the thinking-out-loud preamble and code fences it used to choke on, and it no longer stops at the first square bracket it sees. When a reply genuinely cannot be read, the check now says it was degraded instead of presenting a verdict it did not earn, which also keeps that answer out of the week-long cache and out of the published accuracy score. The same fault had been collapsing the three search angles behind each check down to one. (`/fact-checker`) `[fix, improvement]`
- **Fixed a footer link that led to a dead page**: The terminal page's footer offered an Avatars link that pointed at an address the site does not serve, so clicking it returned a 404 instead of the avatar gallery. It now opens the gallery, the same place every other page sends you. On the knock page, a placeholder link that sat empty until the profile finished loading was removed, so people navigating by keyboard or screen reader no longer land on a link with no name and nowhere to go. (`/tty`) `[fix]`
- **License terms for three.ws containers on NVIDIA NGC**: three.ws is preparing to publish its image-to-3D inference server in NVIDIA's NGC catalog, so the licence that will cover those container images is now public. It states plainly what you get: Apache-2.0 source, no model weights inside the image (you bring your own, under the upstream licence), and a container that carries no analytics or telemetry and sends nothing back to us. It sits alongside the other policies on the legal page. (`/legal/nvidia-ngc-eula`) `[docs]`
-- **The 3D Studio connector for ChatGPT is documented as the eleven tools it actually serves**: The free 3D Studio connector grew to eleven tools when look_at_model landed alongside check_job, but a lot of the writing about it had not caught up: the connector page, the press kit, the partner page, the blog summary, the announcement copy, the press release and the store-submission package still said nine or ten, and the social card enumerated only nine tool names. Every one of those now matches a fresh pull from the live server, the social card lists all eleven, and the module headers inside the connector describe the real catalog including which tools spend generation quota. A test now pins the total, so the next tool to land fails the suite instead of quietly re-opening the same drift. The standalone viewer also stopped crediting the AR launch endpoint with a renderer it no longer loads. Two smaller corrections rode along: the pending-generation example now shows the field that tells a client whether the model coming back is a bare mesh or a finished rig, and the three places the platform pins the 3D viewer library are now explained in one comment, so nobody has to guess why an embed and a standalone page load different builds. (`/docs/mcp-studio`) `[docs, fix]`
+- **The 3D Studio connector for ChatGPT is documented as the eleven tools it actually serves**: The free 3D Studio connector grew to eleven tools when look_at_model landed alongside check_job, but a lot of the writing about it had not caught up: the connector page, the press kit, the partner page, the blog summary, the announcement copy, the press release and the store-submission package still said nine or ten, and the social card enumerated only nine tool names. Every one of those now matches a fresh pull from the live server, the social card lists all eleven, and the module headers inside the connector describe the real catalog including which tools spend generation quota. A test now pins the total, so the next tool to land fails the suite instead of quietly re-opening the same drift. The standalone viewer also stopped crediting the AR launch endpoint with a renderer it no longer loads. (`/docs/mcp-studio`) `[docs, fix]`
+- **The docs now agree on how many tools the free ChatGPT connector has**: The free 3D Studio connector for ChatGPT gained an inspector tool that renders a finished model from several angles, and a collector that picks up a generation which outran its tool call. The pages describing it had not all caught up: the partner page, the press boilerplate, the announcement post and the OpenAI submission pack still counted nine or ten tools between them, and the partner page's tool list was missing the inspector entirely. Every one of those now says eleven and lists the same eleven tools, checked against a live call to the connector, and a test pins the number so the next tool cannot quietly re-open the gap. Alongside it: the pending-generation example now shows the field that tells a client whether the model coming back is a bare mesh or a finished rig, and the three places the 3D viewer library is pinned are documented in one comment so nobody has to guess why a page loads a different build. (`/openai`) `[docs, fix]`
- **The treasury's self-heal plan now only promises SOL it can actually move**: When the payment rail runs short of the SOL it needs to settle, the platform first tries to heal itself: it looks across its own agent wallets for idle funds and pulls them back to the wallet that pays fees. The preview of that plan was counting wallets it could never open. Eight custodial wallets were encrypted under a key retired during the 2026-07 move to Google Cloud, so the preview kept advertising their balance as available while every real run failed on them, and the plan looked the same the next time anyone checked. The preview now opens each wallet's key before listing it, exactly as a real run does, and reports the ones it cannot open separately with the reason. A plan that reads as capital on hand when the capital is unreachable is worse than no plan at all, and it cost two separate investigations. The health check's settlement hint was corrected to match: it no longer says owner funds are needed only once every wallet reports being at its floor, because a sealed wallet never reports that. `[fix, infra]`
## 2026-09-01
diff --git a/STRUCTURE.md b/STRUCTURE.md
index 0c7a97acf..0bb1d989e 100644
--- a/STRUCTURE.md
+++ b/STRUCTURE.md
@@ -94,7 +94,7 @@ bottom.
| Portal (any website as a walkable world) | Page [pages/portal.html](pages/portal.html) `→ /portal` + controller [src/portal/main.js](src/portal/main.js), renderer [src/portal/render.js](src/portal/render.js), styles [src/portal/portal.css](src/portal/portal.css); reader [api/_lib/portal/outline.js](api/_lib/portal/outline.js) + guarded fetch [api/_lib/portal/fetch-site.js](api/_lib/portal/fetch-site.js) + robots matcher [api/_lib/portal/robots.js](api/_lib/portal/robots.js); pure layout [packages/portal/src/layout.js](packages/portal/src/layout.js); GLB export [api/_lib/portal/world-glb.js](api/_lib/portal/world-glb.js); endpoint [api/portal.js](api/portal.js) `→ /api/portal` | Live | Paste an address and walk the page: each section becomes a building sized by its word count (log scale), each link a door (same-site doors rebuild the world there and push browser history, outside links open the page), each image a billboard loaded through `/api/img`, each code block a monolith, and the palette comes from the page's own `theme-color` or a hash of its host. The layout is a pure, seeded function so the browser, the GLB exporter, the MCP tools and the npm package all build the identical city from one link. Crawler manners are part of the product: a declared user agent, `robots.txt` honoured per RFC 9309, one fetch per page cached fleet-wide for an hour, and an IP-pinned SSRF guard with size and time ceilings. `format=glb` returns the world as glTF with named nodes. SDK + CLI + MCP server: [packages/portal/](packages/portal). Docs: [docs/portal.md](docs/portal.md). Spec: [specs/PORTAL_WORLD.md](specs/PORTAL_WORLD.md). Tests: [tests/portal-robots.test.js](tests/portal-robots.test.js), [tests/portal-outline.test.js](tests/portal-outline.test.js), [tests/portal-layout.test.js](tests/portal-layout.test.js), [tests/portal-endpoint.test.js](tests/portal-endpoint.test.js), [tests/e2e/portal.spec.js](tests/e2e/portal.spec.js) |
| IRL (agents in the real world) | [pages/irl.html](pages/irl.html) `→ /irl` + [src/irl.js](src/irl.js), WebXR placement [src/ar/webxr.js](src/ar/webxr.js) + pinch [src/ar/pinch-scale.js](src/ar/pinch-scale.js), animated Quick Look bake [src/usdz-animated.js](src/usdz-animated.js) (static fallback [src/usdz-pipeline.js](src/usdz-pipeline.js)), backend [api/irl/](api/irl) | Live | Pin an agent at a real GPS spot; discovery is presence-proof-gated ([docs/irl/THREAT-MODEL.md](docs/irl/THREAT-MODEL.md)). **Place in AR** leads the dock on AR-capable devices: WebXR hit-test anchoring on Android with two-finger pinch resize (25%..400%) persisted to the pin (`anchor_scale`; the pinch lands as a follow-up `PATCH /api/irl/pins {id, scale}` after the placement tap), and ARKit Quick Look on iOS with the idle clip baked into the USDZ so the agent moves instead of standing in a T-pose. Discovered pins are equally alive in the in-page camera view: every mounted humanoid GLB plays the retargeted idle clip via a per-pin mixer ([src/irl/pin-idle.js](src/irl/pin-idle.js), phase-randomized, LOD-bounded), camera-aware gaze composes on top of the running animation, and the scene renders with ACES tone mapping + IBL + soft shadows (e2e: [tests/e2e/irl-pin-idle.spec.js](tests/e2e/irl-pin-idle.spec.js)). Inbound `/irl?avatar=<glb>` is the living-agent handoff from `GET /api/ar?…&kind=avatar` (explicit `?avatar=` wins over a saved session). Money Drops and World Lines: [docs/irl.md](docs/irl.md). SDK: [packages/irl/](packages/irl). **Share** ([src/irl/share-frame.js](src/irl/share-frame.js), [api/irl/share.js](api/irl/share.js), [api/irl/share/[token].js](api/irl/share/%5Btoken%5D.js)) mints a permanent, unfurlable `/irl/s/:token` link from the captured AR photo for the pin you just placed - never a coordinate, owner-gated, falls back to the plain native share sheet with no active pin. **Analytics** ([api/_lib/irl-analytics.js](api/_lib/irl-analytics.js), [api/irl/analytics.js](api/irl/analytics.js)) is the first site-wide usage rollup for the feature - pins placed, nearby reads, placement method, shares, drop claims - aged out by the existing hourly reaper. |
| Verifiable 3D provenance (content credentials) | Spec [specs/PROVENANCE_3D.md](specs/PROVENANCE_3D.md), pure core [api/_lib/provenance-3d.js](api/_lib/provenance-3d.js), Solana anchor [api/_lib/provenance-anchor.js](api/_lib/provenance-anchor.js), MCP tools [api/_mcp3d/tools/provenance.js](api/_mcp3d/tools/provenance.js) (`verify_provenance` free, `anchor_provenance` paid $0.05), free REST verify [api/provenance.js](api/provenance.js) `→ /api/provenance`, viewer badge [public/viewer.html](public/viewer.html), docs [docs/provenance.md](docs/provenance.md). | Live (verify) · anchor gated on issuer key + devnet funds | C2PA-style signed content credentials for AI 3D: creator, prompt, model/provider, lineage, timestamp, and the GLB's sha256, ed25519-signed by the three.ws issuer with the credential hash anchored on Solana. Free public `verify_provenance` / `GET /api/provenance?src=…` recompute the hash and return `verified | tampered | unknown` (a single changed byte → tampered); the verify path and response are coin-clean (both stores). `anchor_provenance` signs + writes the SPL Memo anchor + stores the credential in R2. Tests: [tests/provenance-3d.test.js](tests/provenance-3d.test.js). Evidence: [prompts/store-submissions/_generated/provenance/](prompts/store-submissions/_generated/provenance/). Store-submission prompt 19. |
-| OpenAI Select Partner (partner surface + announcement pack) | [pages/openai/index.html](pages/openai/index.html) `→ /openai`, badge assets [public/partners/openai/](public/partners/openai), announcement [blog/three-ws-openai-select-partner.html](blog/three-ws-openai-select-partner.html) `→ /blog/three-ws-openai-select-partner`, directory card in [pages/partners.html](pages/partners.html) `→ /partners`, press/social pack [marketing/openai-select-partner/](marketing/openai-select-partner/README.md), original OpenAI-supplied badge files + press-release template [OAI Select Tier Assets/](OAI%20Select%20Tier%20Assets/README.md) | Live | three.ws's OpenAI Partner Network Select-tier surface. `/openai` is the partner landing: the supplied badge (used unmodified, on a white plate), the four live OpenAI surfaces (the free keyless [3D Studio MCP connector](docs/mcp-studio.md) `→ /api/mcp-studio`, the custom-GPT Actions twin `→ /api/3d/studio`, [Spatial MCP](specs/SPATIAL_MCP.md), and [AR in ChatGPT](docs/chatgpt-ar.md)), the connect table, and the nine-tool list. The marketing pack holds badge-usage rules and the press-release draft, which is gated on OpenAI's written approval before publishing (see its README). |
+| OpenAI Select Partner (partner surface + announcement pack) | [pages/openai/index.html](pages/openai/index.html) `→ /openai`, badge assets [public/partners/openai/](public/partners/openai), announcement [blog/three-ws-openai-select-partner.html](blog/three-ws-openai-select-partner.html) `→ /blog/three-ws-openai-select-partner`, directory card in [pages/partners.html](pages/partners.html) `→ /partners`, press/social pack [marketing/openai-select-partner/](marketing/openai-select-partner/README.md), original OpenAI-supplied badge files + press-release template [OAI Select Tier Assets/](OAI%20Select%20Tier%20Assets/README.md) | Live | three.ws's OpenAI Partner Network Select-tier surface. `/openai` is the partner landing: the supplied badge (used unmodified, on a white plate), the four live OpenAI surfaces (the free keyless [3D Studio MCP connector](docs/mcp-studio.md) `→ /api/mcp-studio`, the custom-GPT Actions twin `→ /api/3d/studio`, [Spatial MCP](specs/SPATIAL_MCP.md), and [AR in ChatGPT](docs/chatgpt-ar.md)), the connect table, and the eleven-tool list. The marketing pack holds badge-usage rules and the press-release draft, which is gated on OpenAI's written approval before publishing (see its README). |
| NVIDIA Inception (membership surface + GPU docs cluster) | Landing page [pages/nvidia/index.html](pages/nvidia/index.html) `→ /nvidia` (vercel route + vite input + dev middleware, same wiring as `/openai`), membership doc [docs/nvidia-inception.md](docs/nvidia-inception.md) `→ /docs/nvidia-inception` + overview deck [docs/nvidia-inception/index.html](docs/nvidia-inception/index.html) `→ /docs/nvidia-inception/index.html`, free hosted-model map [docs/nvidia-models.md](docs/nvidia-models.md) `→ /docs/nvidia-models`, forum write-up [docs/nvidia-nemotron-spotlight.md](docs/nvidia-nemotron-spotlight.md) `→ /docs/nvidia-nemotron-spotlight`, tutorials [docs/tutorials/nvidia-3d-free.md](docs/tutorials/nvidia-3d-free.md) `→ /tutorials/nvidia-3d-free` and [docs/tutorials/nvidia-nim-self-host.md](docs/tutorials/nvidia-nim-self-host.md) `→ /tutorials/nvidia-nim-self-host`, engineering post [blog/image-to-3d-on-nvidia-l4-and-blackwell.html](blog/image-to-3d-on-nvidia-l4-and-blackwell.html) `→ /blog/image-to-3d-on-nvidia-l4-and-blackwell`, member badge [public/marks/nvidia-inception-badge.svg](public/marks/nvidia-inception-badge.svg) in [public/footer.html](public/footer.html), directory card in [pages/partners.html](pages/partners.html) `→ /partners` | Live | Everything NVIDIA on the platform, in two layers. **Free hosted lane:** one `NVIDIA_API_KEY` behind TRELLIS text-to-3D, FLUX.1-schnell, the Llama/Nemotron chat lineup, Nemotron vision, nv-embedqa, the reranker, NemoGuard, and Magpie speech (`nvidia-models`). **Self-hosted GPU fleet:** 12 Cloud Run GPU services, 11 on `nvidia-l4` and one on `nvidia-rtx-pro-6000` (Blackwell), covering text/image-to-3D, texture, remesh, rigging, and motion (see each `workers/model-*/README.md`). Membership was accepted July 2026 and is a startup program, not a partnership or an investment; the badge links to NVIDIA's startup program and must not imply endorsement. |
| Press kit (brand marks + media assets) | [pages/press/index.html](pages/press/index.html) `→ /press`, marks [public/brand/](public/brand), layout source [marketing/brand/](marketing/brand/README.md), renderer [scripts/render-brand-assets.mjs](scripts/render-brand-assets.mjs) `npm run build:brand-assets` | Live | The self-serve press surface: transparent-PNG marks (cube, horizontal and stacked lockups for dark and light grounds), the OpenAI announcement graphics, boilerplate at four lengths (one line, short, long, and a full-length bio) with copy buttons, fast facts, and the mark-usage rules. Everything is rendered from source rather than hand-exported, so a copy change re-renders the whole set; the renderer also builds `/brand/three-ws-press-kit.zip` with a README carrying the usage rules so the files cannot travel without them. Announcement cards come from [marketing/openai-select-partner/cards/social-card.html](marketing/openai-select-partner/cards/social-card.html) via `npm run build:openai-cards`. |
| Sperax partnership + SperaxOS plugin | [pages/sperax.html](pages/sperax.html) `→ /sperax`, plugin [public/sperax/](public/sperax/README.md), manifest `→ /.well-known/sperax-plugin.json`, tools [api/chat-plugin/[tool].js](api/chat-plugin/%5Btool%5D.js) `→ /api/chat-plugin/*`, panel `→ /sperax/iframe/`, bundled SDK component [chat-plugin/](chat-plugin) (`@three-ws/chat-plugin`, for hosts that mount plugins as React components instead of framing the manifest) | In-repo · Live | Partnership landing for the SperaxOS integration: three.ws ships as a standalone `plugin.delivery` plugin that gives a SperaxOS agent an embodied 3D avatar (speak/gesture/emote/render_agent), bound to an ERC-8004 identity, settling over Base x402. Launch offer: free AI credits for three.ws users on [chat.sperax.io](https://chat.sperax.io) — page explains what it is and how to claim. Blog: [/blog/three-ws-speraxusd-integration](blog/three-ws-speraxusd-integration.html) |
diff --git a/marketing/nvidia-inception/README.md b/marketing/nvidia-inception/README.md
index 773da98ed..931a7c544 100644
--- a/marketing/nvidia-inception/README.md
+++ b/marketing/nvidia-inception/README.md
@@ -13,6 +13,7 @@ Strategy lives in docs, not here:
- [docs/nvidia-visibility-map.md](../../docs/nvidia-visibility-map.md): every NVIDIA surface, the verified intake route for each, and the member benefits still unclaimed
- [docs/nvidia-apps-catalog-listing.md](../../docs/nvidia-apps-catalog-listing.md) and [docs/nvidia-apps-catalog-request.md](../../docs/nvidia-apps-catalog-request.md): the Accelerated Apps Catalog listing kit and its outbound email
- [docs/nvidia-inception.md](../../docs/nvidia-inception.md): what the membership is, and the rule that it is a program rather than a partnership or an endorsement
+- [docs/nvidia-ngc-listing.md](../../docs/nvidia-ngc-listing.md): the NGC catalog kit, the only NVIDIA listing with a self-serve intake form
- [docs/nvidia-forum-browser-digital-human.md](../../docs/nvidia-forum-browser-digital-human.md): NVIDIA Developer Forums post 3, the browser-native Audio2Face-3D write-up, drafted and owner-gated
## Assets
diff --git a/prompts/finish/production-100-00-INDEX.md b/prompts/finish/production-100-00-INDEX.md
index 4fde96787..448c4c5da 100644
--- a/prompts/finish/production-100-00-INDEX.md
+++ b/prompts/finish/production-100-00-INDEX.md
@@ -66,9 +66,8 @@ run in parallel, H waits on owner gates, I is anytime, J is the parallel swarm.
| Order | State |
|---|---|
-| [../backlog/01-x402-settle-runway.md](backlog-01-x402-settle-runway.md) | Code complete; outcome false: settle 5.9%, `cause: sponsor_floor`. One code line left (the dry-run reclaim plan still counts sealed wallets). Capital is OWNER-ACTIONS row 2. |
+| [../backlog/01-x402-settle-runway.md](backlog-01-x402-settle-runway.md) | Code complete; outcome false: settle 5.9%, `cause: sponsor_floor`. The dry-run reclaim line it was waiting on is fixed (`afd349790`, pinned by tests). Capital is OWNER-ACTIONS row 2. |
| [../backlog/11-agent-index-lag.md](backlog-11-agent-index-lag.md) | New 2026-09-01: healthz reports `agent_index` down (1,092 of 1,602 Solana agents erroring, EVM cursors 3,039 hours stale) and nothing owned it. |
-| [02-stranded-wallet-reclaim.md](production-100-02-stranded-wallet-reclaim.md) | Open; nothing shipped. Its ops-card target was removed with the admin panel on 2026-08-05, so the visible surface is the JSON endpoint plus the brief. |
| [03-master-key-hygiene.md](production-100-03-master-key-hygiene.md) | Open; the runbook section was never written and the cloud half needs a live `gcloud`. |
### C. Repo and product defects
diff --git a/prompts/store-submissions/_generated/live-call-forge_free.json b/prompts/store-submissions/_generated/live-call-forge_free.json
index 860e8d267..948cf6389 100644
--- a/prompts/store-submissions/_generated/live-call-forge_free.json
+++ b/prompts/store-submissions/_generated/live-call-forge_free.json
@@ -1 +1,76 @@
-{"error":"rate_limited","error_description":"generation rate limit — slow down and try again shortly","retry_after":60,"reason":"rate_limiter_unavailable"}
\ No newline at end of file
+{
+ "captured": "2026-09-02",
+ "note": "Live, unauthenticated tools/call against https://three.ws/api/mcp-studio. No Authorization header, no API key, no account. Verbatim response body.",
+ "request": {
+ "method": "POST",
+ "url": "https://three.ws/api/mcp-studio",
+ "headers": {
+ "content-type": "application/json"
+ },
+ "body": {
+ "jsonrpc": "2.0",
+ "id": 10,
+ "method": "tools/call",
+ "params": {
+ "name": "forge_free",
+ "arguments": {
+ "prompt": "a low-poly wooden treasure chest with iron bands"
+ }
+ }
+ }
+ },
+ "http_status": 200,
+ "elapsed_seconds": 88.3,
+ "response": {
+ "jsonrpc": "2.0",
+ "id": 10,
+ "result": {
+ "content": [
+ {
+ "type": "text",
+ "text": "Generated a 3D model (GLB). View it: https://three.ws/viewer?src=https%3A%2F%2Fthree.ws%2Fcdn%2Fforge%2Fanon%2Fd36824c3-5749-40d5-a7f5-fa7e0ff159f2.glb\nPlace it in your room (AR, open on a phone): https://three.ws/api/ar?src=https%3A%2F%2Fthree.ws%2Fcdn%2Fforge%2Fanon%2Fd36824c3-5749-40d5-a7f5-fa7e0ff159f2.glb&title=a%20low-poly%20wooden%20treasure%20chest%20with%20iron%20bands\nDownload: https://three.ws/cdn/forge/anon/d36824c3-5749-40d5-a7f5-fa7e0ff159f2.glb\nConcept image it was sculpted from: https://three.ws/cdn/forge/refs/ceaf5386-7302-4f17-994d-f8677282602e.jpg"
+ }
+ ],
+ "structuredContent": {
+ "kind": "model",
+ "glbUrl": "https://three.ws/cdn/forge/anon/d36824c3-5749-40d5-a7f5-fa7e0ff159f2.glb",
+ "viewerUrl": "https://three.ws/viewer?src=https%3A%2F%2Fthree.ws%2Fcdn%2Fforge%2Fanon%2Fd36824c3-5749-40d5-a7f5-fa7e0ff159f2.glb",
+ "arUrl": "https://three.ws/api/ar?src=https%3A%2F%2Fthree.ws%2Fcdn%2Fforge%2Fanon%2Fd36824c3-5749-40d5-a7f5-fa7e0ff159f2.glb&title=a%20low-poly%20wooden%20treasure%20chest%20with%20iron%20bands",
+ "format": "glb",
+ "prompt": "a low-poly wooden treasure chest with iron bands",
+ "referenceImageUrl": "https://three.ws/cdn/forge/refs/ceaf5386-7302-4f17-994d-f8677282602e.jpg",
+ "spatial": {
+ "spatialMcpVersion": "0.1",
+ "kind": "model",
+ "scene": {
+ "glbUrl": "https://three.ws/cdn/forge/anon/d36824c3-5749-40d5-a7f5-fa7e0ff159f2.glb",
+ "format": "glb"
+ },
+ "camera": {
+ "autoRotate": true
+ },
+ "environment": {
+ "image": "neutral",
+ "exposure": 1,
+ "shadowIntensity": 1
+ },
+ "affordances": {
+ "orbit": true,
+ "zoom": true,
+ "fullscreen": true,
+ "download": true
+ },
+ "ar": {
+ "glbUrl": "https://three.ws/cdn/forge/anon/d36824c3-5749-40d5-a7f5-fa7e0ff159f2.glb",
+ "launchUrl": "https://three.ws/api/ar?src=https%3A%2F%2Fthree.ws%2Fcdn%2Fforge%2Fanon%2Fd36824c3-5749-40d5-a7f5-fa7e0ff159f2.glb&title=a%20low-poly%20wooden%20treasure%20chest%20with%20iron%20bands",
+ "supported": true
+ },
+ "meta": {
+ "prompt": "a low-poly wooden treasure chest with iron bands",
+ "viewerUrl": "https://three.ws/viewer?src=https%3A%2F%2Fthree.ws%2Fcdn%2Fforge%2Fanon%2Fd36824c3-5749-40d5-a7f5-fa7e0ff159f2.glb"
+ }
+ }
+ }
+ }
+ }
+}
diff --git a/prompts/store-submissions/_generated/openai-submission.md b/prompts/store-submissions/_generated/openai-submission.md
index e9da4e568..8985970c2 100644
--- a/prompts/store-submissions/_generated/openai-submission.md
+++ b/prompts/store-submissions/_generated/openai-submission.md
@@ -1,16 +1,29 @@
# three.ws 3D Studio — OpenAI ChatGPT App Directory submission package
-**Prepared:** 2026-07-07 · **Owning prompt:** 06 · **Endpoint:** `https://three.ws/api/mcp-studio`
+**Prepared:** 2026-07-07 · **Last re-verified live:** 2026-09-02 · **Owning prompt:** 06
+**Endpoint:** `https://three.ws/api/mcp-studio`
**Prereqs verified live:** prompt 04 (`/api/mcp-studio` deployed), prompt 05 (widget renders real GLBs).
This is the copy-paste-ready answer sheet for submitting the free three.ws 3D Studio to the
-OpenAI ChatGPT App Directory, plus an evidence-backed compliance audit. **Every field is now
+OpenAI ChatGPT App Directory, plus an evidence-backed compliance audit. **Every field is
filled** (2026-07-14: organization verified on platform.openai.com, support contact and privacy
-policy confirmed live). The only remaining step is the owner's final submit in the portal.
+policy confirmed live). Submission is currently held on the §0 output-quality blocker; once that
+fix is deployed and re-verified, the remaining step is the owner's final submit in the portal.
---
-## 0. Submission verdict — READY (blockers cleared 2026-07-14)
+## 0. Submission verdict: NOT READY (output-quality blocker found 2026-09-02)
+
+**Every connector, manifest, widget and Actions surface re-verified live on 2026-09-02 and all
+pass** (see the §7 checklist). Submission is held on one product defect, not a protocol one:
+
+| # | Blocker | State |
+|---|---------|-------|
+| **B3** | The free lane intermittently ships a **degenerate slab**: a reconstruction that lost its image conditioning returns a full-footprint relief (a 2.0 x 0.31 x 2.0 pancake) instead of the object. It carries 30k triangles and PBR textures, so `scoreGlbQuality` rated it **0.976 / flag `ok`** and skipped vision QA entirely, shipping it as a top-confidence result. Measured on 40 consecutive live free-lane generations: **6 affected**. Both first-try verification generations run for this checklist hit it, so a reviewer's first prompt has a material chance of returning a flat slab. | Cheap-scorer fix landed in `api/_lib/glb-quality.js` (flatness metric + `planar` signal that forces vision-QA escalation; validated to flag 6/6 known-bad and 0/34 healthy). **Needs a production deploy, then a fresh re-verification generation, before submitting.** |
+
+The historical 2026-07-14 blockers below stay cleared.
+
+## 0a. Prior verdict: READY (blockers cleared 2026-07-14)
The app **passes every OpenAI content/privacy/annotation policy** (§2 audit: all PASS). The two
production defects found on 2026-07-07 are both **fixed and re-verified live on 2026-07-14**:
@@ -406,8 +419,14 @@ curl -s -X POST https://three.ws/api/mcp-studio -H 'content-type: application/js
## 7. Pre-submit checklist
-- [x] **B1** cleared — `tools/call forge_free` returns 200 with a GLB (re-verified 2026-07-14, 1.45 MB GLB).
-- [x] **B2** cleared — `/viewer?src=<glb>` returns 200 and renders the model (re-verified 2026-07-14).
+- [ ] **B3** open: the free lane can ship a degenerate slab as a top-confidence result (§0). Cheap-scorer
+ fix is in the tree (`api/_lib/glb-quality.js`); **deploy it, then re-run a `forge_free` generation and
+ confirm the returned GLB is a solid object** before submitting.
+- [x] **B1** cleared: `tools/call forge_free` returns 200 with a GLB (re-verified live 2026-09-02: HTTP 200
+ in 88s, real 2,256,144-byte GLB with `glTF` magic, `model/gltf-binary`; verbatim response saved to
+ `_generated/live-call-forge_free.json`, which previously held only a stale rate-limit error).
+- [x] **B2** cleared: `/viewer?src=<glb>` returns 200 and renders the model (re-verified live 2026-09-02:
+ `<model-viewer>` reports `loaded === true` for the generated GLB, with no page errors).
- [x] Developer identity verified on platform.openai.com (verified organization, 2026-07-14).
- [x] Support contact + privacy policy confirmed live (2026-07-14); both legal URLs return 200 in the
canonical no-`.html` form (`/legal/privacy`, `/legal/tos`), matching the served OpenAPI (2026-07-18).
@@ -423,6 +442,19 @@ curl -s -X POST https://three.ws/api/mcp-studio -H 'content-type: application/js
- [ ] Screenshots — **re-capture the landscape widget shots against the shipped inline widget**
(`api/_mcp-studio/component.js`); the earlier standalone-viewer captures were removed as part of
collapsing the duplicate viewers. Confirm the form's required dimensions. `[HUMAN]`
+- [x] Inline widget re-verified live 2026-09-02 in a `window.openai`-less Chromium: the skybridge
+ resource paints its empty state ("No model yet"), accepts a real tool payload over `postMessage`,
+ shows the generating state, and reaches ready with `<model-viewer>.loaded === true` on the real
+ generated GLB. Error state reached with an unresolvable GLB ("Couldn't load the model"). No console
+ or page errors in any state.
+- [x] `/api/ar` re-verified live 2026-09-02 per device class: Android UA gets a 302 to a Google Scene
+ Viewer `intent://` URL with a `browser_fallback_url`; iOS and desktop UAs get the 200 launch page
+ that carries the real `og:image`/`og:title` and hands off to `/ar/view`; `kind=avatar` adds the
+ `irl=` hand-off and `/ar/view` renders its "Bring it to life" control; a bad `src` returns a
+ designed 400 page ("Provide a valid https URL to a .glb model.").
+- [x] Custom-GPT Actions lane re-verified live 2026-09-02: `POST /api/3d/studio` returned the documented
+ pending shape in 15s, `GET ?job=&title=` polled to `done` with a real 2,922,332-byte GLB, and both
+ responses validate against the served OpenAPI's own response schemas.
- [x] Compliance audit: 7/7 policy items PASS (§2), with the review-surface separation documented (§2.1a).
- [x] Listing metadata drafted (§1): tool list is the live 11-tool surface (re-pulled 2026-09-02);
`forge_free` tier note corrected to the honest standard default (2026-07-18).
diff --git a/public/changelog.json b/public/changelog.json
index 7a7184955..7b69880c0 100644
--- a/public/changelog.json
+++ b/public/changelog.json
@@ -1,5 +1,5 @@
{
- "generated_at": "2026-09-02T19:23:04.763Z",
+ "generated_at": "2026-09-02T19:23:44.983Z",
"generated_by": "scripts/build-page-index.mjs from data/pages.json + data/changelog.json",
"site": {
"name": "three.ws",
@@ -69,17 +69,6 @@
"infra"
]
},
- {
- "date": "2026-09-02",
- "type": "update",
- "title": "Every 3D service we sell to other agents now publishes its parameters and a working example",
- "summary": "Agents that buy three.ws 3D generation on the OKX.AI marketplace could read what each service does but not how to call it: the listings named no parameters and carried no example request, so a buyer had to guess the argument names. Each listed service now publishes the full set: what it does, every parameter with its type and whether it is required, the request method and tool name, and a copy-pasteable curl command that hits the real endpoint. The parameter list and the example are generated from the same input schema the endpoint enforces, and a test checks each published example against that schema, so an example that the endpoint would reject cannot ship, and renaming an argument updates the listing with it.",
- "link": null,
- "tags": [
- "improvement",
- "docs"
- ]
- },
{
"date": "2026-09-02",
"type": "update",
@@ -149,13 +138,24 @@
"date": "2026-09-02",
"type": "update",
"title": "The 3D Studio connector for ChatGPT is documented as the eleven tools it actually serves",
- "summary": "The free 3D Studio connector grew to eleven tools when look_at_model landed alongside check_job, but a lot of the writing about it had not caught up: the connector page, the press kit, the partner page, the blog summary, the announcement copy, the press release and the store-submission package still said nine or ten, and the social card enumerated only nine tool names. Every one of those now matches a fresh pull from the live server, the social card lists all eleven, and the module headers inside the connector describe the real catalog including which tools spend generation quota. A test now pins the total, so the next tool to land fails the suite instead of quietly re-opening the same drift. The standalone viewer also stopped crediting the AR launch endpoint with a renderer it no longer loads. Two smaller corrections rode along: the pending-generation example now shows the field that tells a client whether the model coming back is a bare mesh or a finished rig, and the three places the platform pins the 3D viewer library are now explained in one comment, so nobody has to guess why an embed and a standalone page load different builds.",
+ "summary": "The free 3D Studio connector grew to eleven tools when look_at_model landed alongside check_job, but a lot of the writing about it had not caught up: the connector page, the press kit, the partner page, the blog summary, the announcement copy, the press release and the store-submission package still said nine or ten, and the social card enumerated only nine tool names. Every one of those now matches a fresh pull from the live server, the social card lists all eleven, and the module headers inside the connector describe the real catalog including which tools spend generation quota. A test now pins the total, so the next tool to land fails the suite instead of quietly re-opening the same drift. The standalone viewer also stopped crediting the AR launch endpoint with a renderer it no longer loads.",
"link": "/docs/mcp-studio",
"tags": [
"docs",
"fix"
]
},
+ {
+ "date": "2026-09-02",
+ "type": "update",
+ "title": "The docs now agree on how many tools the free ChatGPT connector has",
+ "summary": "The free 3D Studio connector for ChatGPT gained an inspector tool that renders a finished model from several angles, and a collector that picks up a generation which outran its tool call. The pages describing it had not all caught up: the partner page, the press boilerplate, the announcement post and the OpenAI submission pack still counted nine or ten tools between them, and the partner page's tool list was missing the inspector entirely. Every one of those now says eleven and lists the same eleven tools, checked against a live call to the connector, and a test pins the number so the next tool cannot quietly re-open the gap. Alongside it: the pending-generation example now shows the field that tells a client whether the model coming back is a bare mesh or a finished rig, and the three places the 3D viewer library is pinned are documented in one comment so nobody has to guess why a page loads a different build.",
+ "link": "/openai",
+ "tags": [
+ "docs",
+ "fix"
+ ]
+ },
{
"date": "2026-09-02",
"type": "update",
diff --git a/public/changelog.xml b/public/changelog.xml
index c8018c60c..678f1115b 100644
--- a/public/changelog.xml
+++ b/public/changelog.xml
@@ -54,14 +54,6 @@
<category>fix</category>
<description>Two thirds of the agents in the directory had quietly stopped picking up new on-chain activity. The indexer resumes each agent from the last transaction it saw, and once the RPC provider answering the call no longer held that transaction, the whole read failed and the agent was stuck there permanently: its profile kept loading, it just never learned anything new again. The crawl now notices that its bookmark has gone stale, re-reads from the top and carries on, and the same recovery landed for the secondary chains, where one had been parked on pruned blocks since April. The 1,083 stuck agents have been unstuck. The status page also learned to score this: an index that is erroring rather than lagging used to look perfectly fresh, and a lagging secondary chain no longer reports a platform outage nobody is having.</description>
</item>
- <item>
- <title>Every 3D service we sell to other agents now publishes its parameters and a working example</title>
- <link>https://three.ws/changelog</link>
- <guid isPermaLink="false">2026-09-02:Every 3D service we sell to other agents now publishes its parameters and a working example</guid>
- <pubDate>Wed, 02 Sep 2026 12:00:00 GMT</pubDate>
- <category>improvement</category>
- <description>Agents that buy three.ws 3D generation on the OKX.AI marketplace could read what each service does but not how to call it: the listings named no parameters and carried no example request, so a buyer had to guess the argument names. Each listed service now publishes the full set: what it does, every parameter with its type and whether it is required, the request method and tool name, and a copy-pasteable curl command that hits the real endpoint. The parameter list and the example are generated from the same input schema the endpoint enforces, and a test checks each published example against that schema, so an example that the endpoint would reject cannot ship, and renaming an argument updates the listing with it.</description>
- </item>
<item>
<title>Every paid endpoint now appears in the public API document</title>
<link>https://three.ws/changelog</link>
@@ -116,7 +108,15 @@
<guid isPermaLink="false">2026-09-02:The 3D Studio connector for ChatGPT is documented as the eleven tools it actually serves</guid>
<pubDate>Wed, 02 Sep 2026 12:00:00 GMT</pubDate>
<category>docs</category>
- <description>The free 3D Studio connector grew to eleven tools when look_at_model landed alongside check_job, but a lot of the writing about it had not caught up: the connector page, the press kit, the partner page, the blog summary, the announcement copy, the press release and the store-submission package still said nine or ten, and the social card enumerated only nine tool names. Every one of those now matches a fresh pull from the live server, the social card lists all eleven, and the module headers inside the connector describe the real catalog including which tools spend generation quota. A test now pins the total, so the next tool to land fails the suite instead of quietly re-opening the same drift. The standalone viewer also stopped crediting the AR launch endpoint with a renderer it no longer loads. Two smaller corrections rode along: the pending-generation example now shows the field that tells a client whether the model coming back is a bare mesh or a finished rig, and the three places the platform pins the 3D viewer library are now explained in one comment, so nobody has to guess why an embed and a standalone page load different builds.</description>
+ <description>The free 3D Studio connector grew to eleven tools when look_at_model landed alongside check_job, but a lot of the writing about it had not caught up: the connector page, the press kit, the partner page, the blog summary, the announcement copy, the press release and the store-submission package still said nine or ten, and the social card enumerated only nine tool names. Every one of those now matches a fresh pull from the live server, the social card lists all eleven, and the module headers inside the connector describe the real catalog including which tools spend generation quota. A test now pins the total, so the next tool to land fails the suite instead of quietly re-opening the same drift. The standalone viewer also stopped crediting the AR launch endpoint with a renderer it no longer loads.</description>
+ </item>
+ <item>
+ <title>The docs now agree on how many tools the free ChatGPT connector has</title>
+ <link>https://three.ws/openai</link>
+ <guid isPermaLink="false">2026-09-02:The docs now agree on how many tools the free ChatGPT connector has</guid>
+ <pubDate>Wed, 02 Sep 2026 12:00:00 GMT</pubDate>
+ <category>docs</category>
+ <description>The free 3D Studio connector for ChatGPT gained an inspector tool that renders a finished model from several angles, and a collector that picks up a generation which outran its tool call. The pages describing it had not all caught up: the partner page, the press boilerplate, the announcement post and the OpenAI submission pack still counted nine or ten tools between them, and the partner page's tool list was missing the inspector entirely. Every one of those now says eleven and lists the same eleven tools, checked against a live call to the connector, and a test pins the number so the next tool cannot quietly re-open the gap. Alongside it: the pending-generation example now shows the field that tells a client whether the model coming back is a bare mesh or a finished rig, and the three places the 3D viewer library is pinned are documented in one comment so nobody has to guess why a page loads a different build.</description>
</item>
<item>
<title>The treasury's self-heal plan now only promises SOL it can actually move</title>
diff --git a/public/features.json b/public/features.json
index 9fa4d5fb6..d653d2916 100644
--- a/public/features.json
+++ b/public/features.json
@@ -1,5 +1,5 @@
{
- "generated_at": "2026-09-02T19:23:04.330Z",
+ "generated_at": "2026-09-02T19:23:44.561Z",
"generated_by": "scripts/build-page-index.mjs from data/pages.json",
"site": {
"name": "three.ws",