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(openai): record why the ChatGPT plugin was rejected a third time and rewrite its test cases to match the viewer
nirholas21f78801dd83
diff --git a/CHANGELOG.md b/CHANGELOG.mdindex 5db043f58..d93d0b69a 100644--- a/CHANGELOG.md+++ b/CHANGELOG.md@@ -15,6 +15,7 @@ Public history for [three.ws](https://three.ws), newest first. New pages come fr- **Docs · Syndicates** (`/docs/syndicates`): How copy-trading syndicates work: founding around copyable leaders, joining with your own wallet and caps, how the group curve and rival standing are computed from real copies, and the full API.- **Live Trade Rooms: watch pump.fun traders trade live in 3D** (`/trade-rooms`): Step into a trader's 3D room and watch their real pump.fun buys and sells land as they fill: the avatar reacts, every trade carries its receipt, P&L and Solscan tx, and one tap forks the trade or ghost-copies the trader. Quiet rooms show the latest real trades and which rooms are live now.- **Syndicates: copy-trade as a team** (`/syndicates`): Named teams of copiers who follow the same verified pump.fun trader agents together. Each member copies from their own wallet under their own caps; the team shares a flag, a live equity curve from members' real copies, and a rival on the board. No pooled funds, no custody.+- **3D Studio in ChatGPT finishes every model, however long it takes**: ChatGPT stops waiting on a tool after one minute, and a detailed 3D model takes one to four. The 3D Studio plugin now answers inside that minute and the inline viewer finishes the job itself: it shows a live timer, drops the model in the moment it is ready, rigs an avatar's skeleton on its own when that step is still to come, and keeps a refined model's version history so you can flip back to the original. Inspecting an existing model now shows it in the viewer too, avatars always come out as a full body ready to pose even when the prompt helper is unavailable, and busy periods no longer let one ChatGPT user's generations use up everyone else's. (`/docs/mcp-studio`) `[improvement, fix]`- **Copy Coach: your first copy trade, the careful way**: Never copied a trader before? The new Copy Coach walks you through it in four small steps. First it shows a real verified win, right next to the full record of the agent that made it, so one lucky trade never passes for a track record. Then you ghost-copy a leader with fake money over their real trades. Only then can you set up one tiny real copy: capped at 0.05 SOL per trade and 0.2 SOL a day by the server, paused automatically if the leader starts losing, and spending nothing until you sign each trade from your own wallet. Ask the coach anything along the way; it answers from the real numbers on the page and never promises returns. (`/copy-coach`) `[feature]`- **Daily trading quests with XP, levels and badges**: A new quests page gives you five small trading goals every day: make three trades, fork a verified trade, ghost-copy an agent you have never tried, act on a copy intent, and have your own agent close a winning trade. Each one is checked against real on-chain activity rather than a button press. Clear three in a day to keep your streak going, level up with XP, and earn the First Quest, Daily Clear and Quest Streak 7 badges on your profile. (`/quests`) `[feature]`- **Sentiment Scout: momentum with receipts**: Coin Intelligence has a new Scout tab. It ranks the last hour of pump.fun launches by momentum and backs every claim with its receipt: the X post or paid read it came from, when it was true, and what the chain showed at that same moment. It counts distinct voices rather than posts, calls out coins where all the chatter is one account, names the main risk on every coin, and grades its own past calls against what actually happened. The same evidence now appears on trade receipts, in each coin's detail view, and through the sentiment_scout MCP tool. (`/coin-intel?tab=scout`) `[feature]`diff --git a/data/changelog.json b/data/changelog.jsonindex 35e791cca..477c72c47 100644--- a/data/changelog.json+++ b/data/changelog.json@@ -28807,6 +28807,13 @@"infra"],"link": "/status"+ },+ {+ "date": "2026-09-29",+ "title": "3D Studio in ChatGPT finishes every model, however long it takes",+ "summary": "ChatGPT stops waiting on a tool after one minute, and a detailed 3D model takes one to four. The 3D Studio plugin now answers inside that minute and the inline viewer finishes the job itself: it shows a live timer, drops the model in the moment it is ready, rigs an avatar's skeleton on its own when that step is still to come, and keeps a refined model's version history so you can flip back to the original. Inspecting an existing model now shows it in the viewer too, avatars always come out as a full body ready to pose even when the prompt helper is unavailable, and busy periods no longer let one ChatGPT user's generations use up everyone else's.",+ "tags": ["improvement", "fix"],+ "link": "/docs/mcp-studio"}]}diff --git a/docs/mcp-studio.md b/docs/mcp-studio.mdindex b06eb9043..c5987c03e 100644--- a/docs/mcp-studio.md+++ b/docs/mcp-studio.md@@ -201,10 +201,44 @@ collect the result:`etaRemainingSeconds`. Call again after the suggested wait.- **failed**: returns a clean, actionable error.-`check_job` is read-only and idempotent, and it does not consume generation-quota, so collecting a model can never be rate-limited by the generation that-created it. The inline widget renders the pending state as a designed-"still rendering" panel with the live countdown, not an empty state.+A check that fails while the job itself is fine (the status check timed out, or+its rate bucket is busy) comes back with `retryable: true`. The first check of a+finished job saves the model and scores it, which can take 20 to 30 seconds, so+checking again usually returns the model at once. Only an unrecognized `job_id`+is final.++A pending `refine_model` result also carries a `refine` object: the version+history the new model joins. Pass it back to `check_job` unchanged+(`{"job_id": "…", "refine": {…}}`) and the finished model comes back as a+refinement with its `lineage`, exactly as if `refine_model` had finished inline,+so the version strip survives the wait. A pending `forge_avatar` mesh carries+`"next": "rig"`: once it lands, run `rig_mesh` on it to finish the avatar.++`check_job` never modifies an existing model, but its first check of a finished+job writes one (it saves the GLB and records the creation), so it is annotated+`readOnlyHint: false`. It does not consume generation quota, so collecting a+model can never be rate-limited by the generation that created it.++**The inline widget collects pending jobs itself.** When the host exposes+`window.openai.callTool` (ChatGPT does), the model viewer polls `check_job`,+shows a live elapsed timer, runs `rig_mesh` when the envelope says+`"next": "rig"`, passes `refine` back so the version strip survives, and saves+the finished model to widget state so reopening the conversation shows it+instead of waiting again. A host without widget tool calls gets a designed+"still rendering" panel instead.++#### The ChatGPT call budget++ChatGPT ends any tool call still open at 60 seconds, and that limit is not+configurable. A generation takes one to four minutes, so on+`/api/mcp-chatgpt` every call runs under `CHATGPT_CALL_BUDGET_MS` (40 seconds,+[`api/_mcp-studio/dispatch.js`](../api/_mcp-studio/dispatch.js)): the prompt+director, the submit and the inline wait all shrink to fit it, and a model+still rendering at the deadline comes back as the pending envelope above for+the widget to finish. A submit always gets an 8-second floor, since there is+nothing to hand back without an accepted job, and the budget leaves room for it+under the host's limit. `/api/mcp-studio` keeps the long inline wait (up to+three minutes), because other MCP hosts wait for it.Any HTTP client can poll `pollUrl` directly instead; it is the same public,auth-free job handle the [3D API](/docs/3d-api) hands anonymous callers.@@ -328,6 +362,11 @@ The endpoint still enforces real per-IP abuse protection (`api/_lib/rate-limit.j- **Burst:** 4 generations / minute / IP- **Hourly:** 30 generations / hour / IP+- **ChatGPT users:** every ChatGPT user reaches `/api/mcp-chatgpt` from OpenAI's+ shared egress IPs, so there the burst and hourly caps key on the anonymized+ per-user `openai/subject` ChatGPT sends with each tool call, and one IP is+ held to 300 generations / hour in total. Without the subject the caps key on+ the IP as above.- **Persona writes:** 20 / minute / IP (`create_agent_persona`, `persona_say`;`get_agent_persona` is a read and rides the transport cap)- **Transport:** 300 requests / minute / IP (discovery, never throttled by thediff --git a/prompts/store-submissions/_generated/TRACKER.md b/prompts/store-submissions/_generated/TRACKER.mdindex c488930a7..ea525c638 100644--- a/prompts/store-submissions/_generated/TRACKER.md+++ b/prompts/store-submissions/_generated/TRACKER.md@@ -87,7 +87,7 @@ There are **six** MCP gates, not four: `audit:mcp`, `audit:mcp-golden`, `audit:m| **Claude Connectors Directory** | ready-to-submit (pending the human items) | `POST /api/mcp tools/list` unauthenticated returned **200 with 53 tools in 0.10s**; `initialize` likewise. `GET /.well-known/oauth-protected-resource` 200 with 13 scopes. `/legal/privacy`, `/legal/tos`, `/support`, `/docs/mcp` all 200, and privacy still carries the "MCP Connectors, AI Processing & Payments" section. | The 2026-06-28 blocker that discovery was auth-gated is **resolved and no longer true**: the portal can sync tools with no credential. Remaining: `[HUMAN: Team/Enterprise org + Owner role, or use the public form]` and `[HUMAN: generate a real MCP_REVIEW_SECRET and decide which vendor credentials a reviewer gets]`. Note the 147s free-lane latency above before demoing live. || **Claude plugin marketplace** | ready-to-submit | `.claude-plugin/marketplace.json` carries 4 plugins (`three-ws-core`, `three-ws-developer`, `three-ws-pump-fun`, `three-ws-3d`). The end-to-end CLI install evidence from 2026-07-08 stands in `_generated/claude-plugin-marketplace-e2e.md` and was not re-driven today. | `[HUMAN: submit via https://platform.claude.com/plugins/submit per marketplace/plugins/three-ws-3d/SUBMISSION.md]` || **Claude Agent Skills pack** | ready-to-submit | `node scripts/build-skills-pack.mjs --check` → `skills-pack manifest up to date (43 skills)` (re-run 2026-09-12); `.agents/skills/` holds 43 skill directories. The 4 cross-platform-safe 3D ones are also published as a public bundle at `public/skills/3d-studio/`, gated against drift by `npm run check:openai-skills`. | None in code. The live free-lane E2E it wanted is now green but slow (147s, see above). |-| **OpenAI App Directory** | submitted, in review (v1.0.0, 2026-09-12 00:10Z) | Whole surface re-verified live 2026-09-09: `initialize` 200 keyless, `tools/list` **11 tools** each with its four `openai/*` annotations and correct `outputTemplate`, `resources/list` **2 widget resources** with `widgetCSP`/`widgetDescription`/`widgetDomain`, `/.well-known/3d-studio-openapi.yaml` (Redocly-clean, free-only, all inner URLs 200) and `/.well-known/ai-plugin.json` both 200, `/legal/privacy` + `/legal/tos` 200, a real keyless `forge_free` returned a 5.40 MB GLB in 137s whose `/viewer?src=` returned 200, `POST/GET /api/3d/studio` polled to a real GLB, `/api/ar` correct per device class, and the widget drove through empty/loading/ready/error in a `window.openai`-less browser. Every one of those payloads scanned clean of crypto, payment and internal ids. | **B3 is narrower than it looked.** The scorer fix IS deployed (revision `three-ws-api-00420-ljh`) and re-measuring it on 40 consecutive production generations gave 4 planar flags, 4/4 escalating. The fallback rung is also **not** dead: a 15:04Z generation scored through the free NVIDIA vision lane with a full verdict, so "the NIM fallback 504s" no longer holds in general. **B3 is now closed, and NOT by clearing the billing hold.** Measured 2026-09-11 against revision `three-ws-api-00424-5p4`: ten consecutive authenticated calls to `/api/forge-quality-check` returned **10/10 real verdicts** (`provider: nvidia`, `qa_available: true`), against 3/10 on 2026-09-10 and 0/10 on 2026-09-09. The gate is deterministic again because the vision chain was fixed rather than because Vertex came back: JSON validity became part of the failover test instead of a filter applied after a winner was picked, the per-lane budget stopped being divided equally (which had starved the one healthy rung), a billing-denied Vertex is now parked by a circuit breaker instead of re-probed on every call, and the chain was widened from four providers to ten. Vertex is still denied on the hold; it no longer decides the outcome. The four fixes listed here as unshipped are shipped. Submitted 2026-09-12 at 00:10Z as version 1.0.0; the portal shows status Review. Domain verification passed on `/.well-known/openai-apps-challenge` (commit `f569ec6c4`, live 23:46Z on 2026-09-11). Went in with 11 scanned tools and all 33 annotation justifications, 3 prompts, 5 test cases and 3 negative ones, and the demo recording at `https://youtu.be/DIKBUJypXzw`. Full record of what was submitted: `_generated/openai-submission.md` §0. Remaining: `[HUMAN: wait for the review outcome]`. Clearing the GCP billing hold is still worth doing, but it never was this row's blocker. |+| **OpenAI App Directory** | rejected three times; the v1.0.1 fix is built and awaits deploy, a re-recorded demo and resubmission (2026-09-29). The third rejection's cause: every generation call outlived ChatGPT's 60-second tool-call limit. See `_generated/openai-submission.md` §0 | Whole surface re-verified live 2026-09-09: `initialize` 200 keyless, `tools/list` **11 tools** each with its four `openai/*` annotations and correct `outputTemplate`, `resources/list` **2 widget resources** with `widgetCSP`/`widgetDescription`/`widgetDomain`, `/.well-known/3d-studio-openapi.yaml` (Redocly-clean, free-only, all inner URLs 200) and `/.well-known/ai-plugin.json` both 200, `/legal/privacy` + `/legal/tos` 200, a real keyless `forge_free` returned a 5.40 MB GLB in 137s whose `/viewer?src=` returned 200, `POST/GET /api/3d/studio` polled to a real GLB, `/api/ar` correct per device class, and the widget drove through empty/loading/ready/error in a `window.openai`-less browser. Every one of those payloads scanned clean of crypto, payment and internal ids. | **B3 is narrower than it looked.** The scorer fix IS deployed (revision `three-ws-api-00420-ljh`) and re-measuring it on 40 consecutive production generations gave 4 planar flags, 4/4 escalating. The fallback rung is also **not** dead: a 15:04Z generation scored through the free NVIDIA vision lane with a full verdict, so "the NIM fallback 504s" no longer holds in general. **B3 is now closed, and NOT by clearing the billing hold.** Measured 2026-09-11 against revision `three-ws-api-00424-5p4`: ten consecutive authenticated calls to `/api/forge-quality-check` returned **10/10 real verdicts** (`provider: nvidia`, `qa_available: true`), against 3/10 on 2026-09-10 and 0/10 on 2026-09-09. The gate is deterministic again because the vision chain was fixed rather than because Vertex came back: JSON validity became part of the failover test instead of a filter applied after a winner was picked, the per-lane budget stopped being divided equally (which had starved the one healthy rung), a billing-denied Vertex is now parked by a circuit breaker instead of re-probed on every call, and the chain was widened from four providers to ten. Vertex is still denied on the hold; it no longer decides the outcome. The four fixes listed here as unshipped are shipped. Submitted 2026-09-12 at 00:10Z as version 1.0.0; the portal shows status Review. Domain verification passed on `/.well-known/openai-apps-challenge` (commit `f569ec6c4`, live 23:46Z on 2026-09-11). Went in with 11 scanned tools and all 33 annotation justifications, 3 prompts, 5 test cases and 3 negative ones, and the demo recording at `https://youtu.be/DIKBUJypXzw`. Full record of what was submitted: `_generated/openai-submission.md` §0. Remaining: `[HUMAN: wait for the review outcome]`. Clearing the GCP billing hold is still worth doing, but it never was this row's blocker. || **OpenAI GPT Store** | live, with a duplicate to remove | `https://chatgpt.com/g/g-6a563a3b49a88191abf346245491a444-three-ws-3d-studio` returns 200 logged-out. The old draft `g-6a5672fbf3f48191b559e482c7fcbf51` **301s to a 200 page titled "ChatGPT - three.ws 3D Studio"**, so the 2026-09-02 escalation stands: two public GPTs share one name. | `[HUMAN: delete the draft duplicate in My GPTs]` (no API for it) || **Official MCP Registry** | ready-to-submit | `node prompts/store-submissions/_generated/build-registry-republish.mjs` against the live registry, after the pagination fix below: **3 servers need a publish (0 stale, 3 new: `herald-mcp`, `knock-mcp`, `home-mcp`), 47 are already current**, including all 7 hosted remotes. Cross-checked with `node scripts/publish-mcp-servers.mjs --dry-run`, which agrees. The batch is staged in `_generated/mcp-registry-republish.sh` and was NOT run. | `[HUMAN: mcp-publisher login github, then review and run the staged batch]` || **Smithery** | ready-to-submit | `_generated/mcp-directories/smithery.md` regenerated today from the corrected canonical source; it now covers **50 servers (7 remote + 43 stdio)**. | `[HUMAN: create/claim the Smithery account]` |diff --git a/prompts/store-submissions/_generated/chatgpt-app-submission.json b/prompts/store-submissions/_generated/chatgpt-app-submission.jsonindex b3faa4250..38b2bec13 100644--- a/prompts/store-submissions/_generated/chatgpt-app-submission.json+++ b/prompts/store-submissions/_generated/chatgpt-app-submission.json@@ -110,24 +110,24 @@"description": "Generate a 3D prop from a text description","user_prompt": "Make a 3D model of a friendly round robot mascot, glossy white plastic.","file_attachment_urls": null,- "tools_triggered": "forge_free, then check_job if the first response returns status \"pending\"",- "expected_output": "An inline interactive 3D viewer with the textured model, plus Download, Spin, Recenter and Open in three.ws. It auto-rotates until dragged. Free, no account. Takes one to four minutes; past the inline wait it returns a job_id and check_job collects the finished model.",+ "tools_triggered": "forge_free; if it returns status \"pending\", the viewer widget collects the model with check_job",+ "expected_output": "An inline 3D viewer showing the textured robot, auto-rotating until dragged, with Download GLB, Open viewer and View in your space. Free, no account. The call answers within about 40s; if still rendering, the viewer shows a live timer and fills in by itself, usually in 1-4 minutes.","expected_output_url": null},{"description": "Generate a rigged, animation-ready character","user_prompt": "Make a rigged, animation-ready knight character I can pose.","file_attachment_urls": null,- "tools_triggered": "forge_avatar, then check_job if it returns status \"pending\", then rig_mesh if the finished job was only the mesh",- "expected_output": "A GLB in the inline viewer with a humanoid skeleton and skin weights applied, so it can be posed. One call normally does mesh and rig; if it times out at the mesh stage the pending result says to finish with rig_mesh. A humanoid rig also plays an idle clip.",+ "tools_triggered": "forge_avatar; if it returns status \"pending\", the viewer widget collects the mesh with check_job and rigs it with rig_mesh",+ "expected_output": "The knight in the inline viewer, labeled rigged model, carrying a humanoid skeleton and skin weights so it can be posed. If the call answers before the mesh is done, the viewer finishes the mesh, runs the rig step itself, then shows the rigged model.","expected_output_url": null},{"description": "Iterate on a model by describing the change in words","user_prompt": "Now make that robot's shell matte instead of glossy.","file_attachment_urls": null,- "tools_triggered": "refine_model, then check_job if it returns status \"pending\"",- "expected_output": "Run immediately after test case 1, in the same conversation. Returns a new version anchored to the previous model rather than an unrelated regeneration, and replaces the GLB in the viewer. The earlier version stays reachable, so the change can be reverted.",+ "tools_triggered": "refine_model; if it returns status \"pending\", the viewer widget collects it with check_job",+ "expected_output": "Run right after test case 1 in the same conversation. A new version anchored to the robot, not an unrelated model, appears in the viewer with a Versions strip (Original, then the change) so either version can be shown again. A still-rendering one fills in by itself.","expected_output_url": null},{@@ -135,15 +135,15 @@"user_prompt": "Look at this 3D model and tell me what it is and whether it has any obvious defects: https://three.ws/cdn/objects/polyhaven/glb/ArmChair_01.glb","file_attachment_urls": null,"tools_triggered": "look_at_model",- "expected_output": "Frames from several angles returned as MCP image content blocks, which the client renders into the conversation, plus geometry stats. The assistant describes the armchair from images it can actually see. Free, seconds not minutes. The URL is ours and needs no credentials.",+ "expected_output": "Frames from four angles returned as image content, plus geometry stats, and the armchair in the inline viewer. The assistant describes the chair from images it can actually see. Free, about 30 seconds. The URL is ours and needs no credentials.","expected_output_url": null},{"description": "Generate a model through the art-directed lane","user_prompt": "Use the art-directed generator to make a detailed vintage brass telescope on a wooden tripod.","file_attachment_urls": null,- "tools_triggered": "mesh_forge, then check_job if it returns status \"pending\"",- "expected_output": "A textured GLB in the same inline viewer, with Download and Open in three.ws. An art-direction pass may first tighten the prompt into a single-subject spec. Free, no account. One to four minutes; a pending result is collected with check_job.",+ "tools_triggered": "mesh_forge; if it returns status \"pending\", the viewer widget collects the model with check_job",+ "expected_output": "A textured telescope in the same inline viewer, with Download GLB and Open viewer. An art-direction pass may first tighten the prompt into a single-subject spec. Free, no account. If still rendering after about 40s, the viewer shows progress and fills in by itself.","expected_output_url": null}],diff --git a/prompts/store-submissions/_generated/openai-portal-fields.md b/prompts/store-submissions/_generated/openai-portal-fields.mdindex b2f15d35d..b52ec435b 100644--- a/prompts/store-submissions/_generated/openai-portal-fields.md+++ b/prompts/store-submissions/_generated/openai-portal-fields.md@@ -156,12 +156,12 @@ Make a 3D model of a friendly round robot mascot, glossy white plastic.**Tool triggered**```-forge_free, then check_job if the first response returns status "pending"+forge_free; if it returns status "pending", the viewer widget collects the model with check_job```-**Expected output** (268/300)+**Expected output** (282/300)```-An inline interactive 3D viewer with the textured model, plus Download, Spin, Recenter and Open in three.ws. It auto-rotates until dragged. Free, no account. Takes one to four minutes; past the inline wait it returns a job_id and check_job collects the finished model.+An inline 3D viewer showing the textured robot, auto-rotating until dragged, with Download GLB, Open viewer and View in your space. Free, no account. The call answers within about 40s; if still rendering, the viewer shows a live timer and fills in by itself, usually in 1-4 minutes.```### Test Case 2@@ -178,12 +178,12 @@ Make a rigged, animation-ready knight character I can pose.**Tool triggered**```-forge_avatar, then check_job if it returns status "pending", then rig_mesh if the finished job was only the mesh+forge_avatar; if it returns status "pending", the viewer widget collects the mesh with check_job and rigs it with rig_mesh```-**Expected output** (257/300)+**Expected output** (250/300)```-A GLB in the inline viewer with a humanoid skeleton and skin weights applied, so it can be posed. One call normally does mesh and rig; if it times out at the mesh stage the pending result says to finish with rig_mesh. A humanoid rig also plays an idle clip.+The knight in the inline viewer, labeled rigged model, carrying a humanoid skeleton and skin weights so it can be posed. If the call answers before the mesh is done, the viewer finishes the mesh, runs the rig step itself, then shows the rigged model.```### Test Case 3@@ -200,12 +200,12 @@ Now make that robot's shell matte instead of glossy.**Tool triggered**```-refine_model, then check_job if it returns status "pending"+refine_model; if it returns status "pending", the viewer widget collects it with check_job```-**Expected output** (256/300)+**Expected output** (266/300)```-Run immediately after test case 1, in the same conversation. Returns a new version anchored to the previous model rather than an unrelated regeneration, and replaces the GLB in the viewer. The earlier version stays reachable, so the change can be reverted.+Run right after test case 1 in the same conversation. A new version anchored to the robot, not an unrelated model, appears in the viewer with a Versions strip (Original, then the change) so either version can be shown again. A still-rendering one fills in by itself.```### Test Case 4@@ -225,9 +225,9 @@ Look at this 3D model and tell me what it is and whether it has any obvious defelook_at_model```-**Expected output** (272/300)+**Expected output** (243/300)```-Frames from several angles returned as MCP image content blocks, which the client renders into the conversation, plus geometry stats. The assistant describes the armchair from images it can actually see. Free, seconds not minutes. The URL is ours and needs no credentials.+Frames from four angles returned as image content, plus geometry stats, and the armchair in the inline viewer. The assistant describes the chair from images it can actually see. Free, about 30 seconds. The URL is ours and needs no credentials.```### Test Case 5@@ -244,12 +244,12 @@ Use the art-directed generator to make a detailed vintage brass telescope on a w**Tool triggered**```-mesh_forge, then check_job if it returns status "pending"+mesh_forge; if it returns status "pending", the viewer widget collects the model with check_job```-**Expected output** (241/300)+**Expected output** (265/300)```-A textured GLB in the same inline viewer, with Download and Open in three.ws. An art-direction pass may first tighten the prompt into a single-subject spec. Free, no account. One to four minutes; a pending result is collected with check_job.+A textured telescope in the same inline viewer, with Download GLB and Open viewer. An art-direction pass may first tighten the prompt into a single-subject spec. Free, no account. If still rendering after about 40s, the viewer shows progress and fills in by itself.```## Testing tab: negative casesdiff --git a/prompts/store-submissions/_generated/openai-submission.md b/prompts/store-submissions/_generated/openai-submission.mdindex c99946175..c5128a0b5 100644--- a/prompts/store-submissions/_generated/openai-submission.md+++ b/prompts/store-submissions/_generated/openai-submission.md@@ -12,7 +12,50 @@ from this file.----## 0. Submission verdict: RESUBMITTING on the ChatGPT surface+## 0. Submission verdict: RESUBMITTING with every call inside ChatGPT's 60-second limit++**The resubmission on `/api/mcp-chatgpt` was rejected too (third rejection, reported 2026-09-29).**+The five fixes below were real, but none of them was the blocker. Measured on 2026-09-29 against+live revision `three-ws-api-00460-4j6`:++1. **Four of the five positive test cases could not complete inside ChatGPT.** ChatGPT ends any+ tool call still open at 60 seconds; an OpenAI staff member confirmed on the developer forum that+ "the hard limit on any tool call is 1 minute" and that it is not configurable. `forge_free`,+ `forge_avatar`, `refine_model` and `mesh_forge` each held the call open for up to 180 seconds+ of polling after a submit of up to 90. A live `forge_free` call to the submitted endpoint took+ **200 seconds** and returned a pending job. In ChatGPT the reviewer saw a failed tool call on+ test cases 1, 2, 3 and 5; only `look_at_model` (28 seconds) passed. That is the "stable,+ responsive, and complete" requirement, failed on the reviewer's own test plan.+ **Fixed:** every call on the ChatGPT surface now answers within a 40-second budget+ (`CHATGPT_CALL_BUDGET_MS`), measured at 40.0 seconds on real `forge_free`, `forge_avatar` and+ `refine_model` calls against the production generator.+2. **A pending result was a dead end.** The widget said "Ask to check the job to collect the+ finished model." **Fixed:** the viewer polls `check_job` itself through+ `window.openai.callTool`, shows a live elapsed timer, rigs a pending avatar mesh with+ `rig_mesh`, keeps a pending refinement's version strip, and retries a status check that timed+ out. Driven end to end in Chromium against real production jobs: the telescope and the rigged+ knight (one skin, 52 joints) both landed in the viewer with no user action.+3. **Test case 4 rendered "No model yet" under a working answer.** `look_at_model` returns+ `model_url`, the widget read only `glbUrl`, and fell through to its empty state. **Fixed:** the+ viewer shows the inspected model.+4. **The test cases promised controls that do not exist.** Case 1 listed "Spin" and "Recenter"+ buttons and case 2 an idle clip; the widget has Download GLB, Open viewer and View in your+ space, and does not autoplay. **Fixed:** every expectation now describes what the reviewer+ actually sees.+5. **A failed prompt director turned the knight into a bust.** With the LLM chain exhausted, the+ raw prompt reached a portrait-leaning lane and came back as head and shoulders. **Fixed:**+ the fallback brief still asks for a full-body figure.+6. **Every ChatGPT user shared one rate-limit bucket.** ChatGPT calls arrive from OpenAI's egress+ IPs, and the caps were 4 per minute and 30 per hour per IP. **Fixed:** on the ChatGPT surface+ they key on the per-user `openai/subject`, with a 300-per-hour ceiling per IP.++**Before resubmitting:** deploy, re-run the five test cases in ChatGPT developer mode against+`https://three.ws/api/mcp-chatgpt`, then **re-record the demo video** from that session. The+current video predates the fix, and the reviewer validates test cases against it. Upload the+regenerated `chatgpt-app-submission.json` (`node scripts/build-openai-portal-fields.mjs --json`)+and bump the version to 1.0.1.++### The first two rejections (2026-09-12)**Version 1.0.0 was submitted on 2026-09-12 at 00:10Z and rejected twice that day.** The portalnever displayed a reason, so every cause below was found by measuring the submission against thediff --git a/public/changelog-recent.json b/public/changelog-recent.jsonindex d2a54574b..a55e88017 100644--- a/public/changelog-recent.json+++ b/public/changelog-recent.json@@ -1,11 +1,11 @@{- "generated_at": "2026-09-29T07:53:23.527Z",+ "generated_at": "2026-09-29T19:55:19.816Z","generated_by": "scripts/build-page-index.mjs from data/pages.json + data/changelog.json","site": {"name": "three.ws","url": "https://three.ws"},- "total": 3773,+ "total": 3774,"complete": false,"full_feed": "/changelog.json","entries": [@@ -99,6 +99,17 @@"launch"]},+ {+ "date": "2026-09-29",+ "type": "update",+ "title": "3D Studio in ChatGPT finishes every model, however long it takes",+ "summary": "ChatGPT stops waiting on a tool after one minute, and a detailed 3D model takes one to four. The 3D Studio plugin now answers inside that minute and the inline viewer finishes the job itself: it shows a live timer, drops the model in the moment it is ready, rigs an avatar's skeleton on its own when that step is still to come, and keeps a refined model's version history so you can flip back to the original. Inspecting an existing model now shows it in the viewer too, avatars always come out as a full body ready to pose even when the prompt helper is unavailable, and busy periods no longer let one ChatGPT user's generations use up everyone else's.",+ "link": "/docs/mcp-studio",+ "tags": [+ "improvement",+ "fix"+ ]+ },{"date": "2026-09-29","type": "update",@@ -1584,17 +1595,6 @@"fix","improvement"]- },- {- "date": "2026-09-09",- "type": "update",- "title": "Smart Home reads and works without a mouse, and now speaks past English",- "summary": "Three fixes to the connected house, all of them things a person hits rather than reads about. The room you have selected states what is on in it at full contrast: on the highlighted row that line sat below the readable minimum, which is exactly the row you are looking at. Answering the safety question that stands between you and an unlocked door now hands the keyboard back to the control you pressed, whether you answer with Yes, with Cancel or with Escape; the list rebuilds itself while the question is up, so the keyboard used to land back at the top of the page and you had to tab the whole house again. And the connect flow, the manage view and the floorplan editor now read from the same translation catalog the rest of the site does, so they are ready for the 86 languages three.ws ships in. Your own words are never translated: the name you gave your house, your rooms, your devices and anything your Home Assistant says come back exactly as you wrote them, in every language.",- "link": "/docs/smart-home",- "tags": [- "improvement",- "fix"- ]}]}diff --git a/public/changelog.json b/public/changelog.jsonindex 9cddb2c27..d1e24556e 100644--- a/public/changelog.json+++ b/public/changelog.json@@ -1,5 +1,5 @@{- "generated_at": "2026-09-29T07:53:23.497Z",+ "generated_at": "2026-09-29T19:55:19.795Z","generated_by": "scripts/build-page-index.mjs from data/pages.json + data/changelog.json","site": {"name": "three.ws",@@ -96,6 +96,17 @@"launch"]},+ {+ "date": "2026-09-29",+ "type": "update",+ "title": "3D Studio in ChatGPT finishes every model, however long it takes",+ "summary": "ChatGPT stops waiting on a tool after one minute, and a detailed 3D model takes one to four. The 3D Studio plugin now answers inside that minute and the inline viewer finishes the job itself: it shows a live timer, drops the model in the moment it is ready, rigs an avatar's skeleton on its own when that step is still to come, and keeps a refined model's version history so you can flip back to the original. Inspecting an existing model now shows it in the viewer too, avatars always come out as a full body ready to pose even when the prompt helper is unavailable, and busy periods no longer let one ChatGPT user's generations use up everyone else's.",+ "link": "/docs/mcp-studio",+ "tags": [+ "improvement",+ "fix"+ ]+ },{"date": "2026-09-29","type": "update",diff --git a/public/changelog.xml b/public/changelog.xmlindex bc5464599..5b16d2bf4 100644--- a/public/changelog.xml+++ b/public/changelog.xml@@ -78,6 +78,14 @@<category>launch</category><description>Named teams of copiers who follow the same verified pump.fun trader agents together. Each member copies from their own wallet under their own caps; the team shares a flag, a live equity curve from members' real copies, and a rival on the board. No pooled funds, no custody.</description></item>+ <item>+ <title>3D Studio in ChatGPT finishes every model, however long it takes</title>+ <link>https://three.ws/docs/mcp-studio</link>+ <guid isPermaLink="false">2026-09-29:3D Studio in ChatGPT finishes every model, however long it takes</guid>+ <pubDate>Tue, 29 Sep 2026 12:00:00 GMT</pubDate>+ <category>improvement</category>+ <description>ChatGPT stops waiting on a tool after one minute, and a detailed 3D model takes one to four. The 3D Studio plugin now answers inside that minute and the inline viewer finishes the job itself: it shows a live timer, drops the model in the moment it is ready, rigs an avatar's skeleton on its own when that step is still to come, and keeps a refined model's version history so you can flip back to the original. Inspecting an existing model now shows it in the viewer too, avatars always come out as a full body ready to pose even when the prompt helper is unavailable, and busy periods no longer let one ChatGPT user's generations use up everyone else's.</description>+ </item><item><title>Copy Coach: your first copy trade, the careful way</title><link>https://three.ws/copy-coach</link>diff --git a/scripts/build-openai-portal-fields.mjs b/scripts/build-openai-portal-fields.mjsindex 92b0a8047..a75ca3c8d 100644--- a/scripts/build-openai-portal-fields.mjs+++ b/scripts/build-openai-portal-fields.mjs@@ -46,27 +46,32 @@ const JUSTIFICATIONS = {},};+// Every expectation describes what the reviewer will actually see in ChatGPT.+// ChatGPT ends a tool call at 60 s, so each generation call answers within about+// 40 s (CHATGPT_CALL_BUDGET_MS in api/_mcp-studio/dispatch.js) and a model still+// rendering is collected by the viewer widget itself through check_job. The+// controls named here are the widget's real buttons (api/_mcp-studio/component.js).const TEST_CASES = [{scenario: 'Generate a 3D prop from a text description',prompt: 'Make a 3D model of a friendly round robot mascot, glossy white plastic.',- tools: 'forge_free, then check_job if the first response returns status "pending"',+ tools: 'forge_free; if it returns status "pending", the viewer widget collects the model with check_job',expected:- 'An inline interactive 3D viewer with the textured model, plus Download, Spin, Recenter and Open in three.ws. It auto-rotates until dragged. Free, no account. Takes one to four minutes; past the inline wait it returns a job_id and check_job collects the finished model.',+ 'An inline 3D viewer showing the textured robot, auto-rotating until dragged, with Download GLB, Open viewer and View in your space. Free, no account. The call answers within about 40s; if still rendering, the viewer shows a live timer and fills in by itself, usually in 1-4 minutes.',},{scenario: 'Generate a rigged, animation-ready character',prompt: 'Make a rigged, animation-ready knight character I can pose.',- tools: 'forge_avatar, then check_job if it returns status "pending", then rig_mesh if the finished job was only the mesh',+ tools: 'forge_avatar; if it returns status "pending", the viewer widget collects the mesh with check_job and rigs it with rig_mesh',expected:- 'A GLB in the inline viewer with a humanoid skeleton and skin weights applied, so it can be posed. One call normally does mesh and rig; if it times out at the mesh stage the pending result says to finish with rig_mesh. A humanoid rig also plays an idle clip.',+ 'The knight in the inline viewer, labeled rigged model, carrying a humanoid skeleton and skin weights so it can be posed. If the call answers before the mesh is done, the viewer finishes the mesh, runs the rig step itself, then shows the rigged model.',},{scenario: 'Iterate on a model by describing the change in words',prompt: "Now make that robot's shell matte instead of glossy.",- tools: 'refine_model, then check_job if it returns status "pending"',+ tools: 'refine_model; if it returns status "pending", the viewer widget collects it with check_job',expected:- 'Run immediately after test case 1, in the same conversation. Returns a new version anchored to the previous model rather than an unrelated regeneration, and replaces the GLB in the viewer. The earlier version stays reachable, so the change can be reverted.',+ 'Run right after test case 1 in the same conversation. A new version anchored to the robot, not an unrelated model, appears in the viewer with a Versions strip (Original, then the change) so either version can be shown again. A still-rendering one fills in by itself.',},{scenario: 'Let the assistant see a 3D model and report on it',@@ -74,14 +79,14 @@ const TEST_CASES = ['Look at this 3D model and tell me what it is and whether it has any obvious defects: https://three.ws/cdn/objects/polyhaven/glb/ArmChair_01.glb',tools: 'look_at_model',expected:- 'Frames from several angles returned as MCP image content blocks, which the client renders into the conversation, plus geometry stats. The assistant describes the armchair from images it can actually see. Free, seconds not minutes. The URL is ours and needs no credentials.',+ 'Frames from four angles returned as image content, plus geometry stats, and the armchair in the inline viewer. The assistant describes the chair from images it can actually see. Free, about 30 seconds. The URL is ours and needs no credentials.',},{scenario: 'Generate a model through the art-directed lane',prompt: 'Use the art-directed generator to make a detailed vintage brass telescope on a wooden tripod.',- tools: 'mesh_forge, then check_job if it returns status "pending"',+ tools: 'mesh_forge; if it returns status "pending", the viewer widget collects the model with check_job',expected:- 'A textured GLB in the same inline viewer, with Download and Open in three.ws. An art-direction pass may first tighten the prompt into a single-subject spec. Free, no account. One to four minutes; a pending result is collected with check_job.',+ 'A textured telescope in the same inline viewer, with Download GLB and Open viewer. An art-direction pass may first tighten the prompt into a single-subject spec. Free, no account. If still rendering after about 40s, the viewer shows progress and fills in by itself.',},];