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.

probe(forge): add the generation-suite production probe and stage the texture weights

nirholas376ec301cd1a
diff --git a/scripts/probe-generation-suite.mjs b/scripts/probe-generation-suite.mjs
index 5199e4160..5457ed3a0 100644
--- a/scripts/probe-generation-suite.mjs
+++ b/scripts/probe-generation-suite.mjs
@@ -466,9 +466,12 @@ async function probeMcp(path, label) {
const tools = list.json?.result?.tools || [];
record(`${label} tools/list`, `POST ${path}`, list.status === 200 && tools.length > 0, list.status === 200 ? `${tools.length} tools` : `HTTP ${list.status}: ${list.text.slice(0, 140)}`);
if (!tools.length) return;
- // Call the cheapest read-only tool the server advertises, so a probe never
- // bills a generation just to prove the JSON-RPC surface answers.
- const readOnly = tools.find((t) => /catalog|list|capabilit|health|search|browse/i.test(t.name));
+ // Call a tool the server itself advertises as FREE, so a probe never bills a
+ // generation just to prove the JSON-RPC surface answers. mcp-3d marks its
+ // free tools in the description; fall back to a read-only-sounding name.
+ const readOnly =
+ tools.find((t) => /\bFREE\b/.test(t.description || '')) ||
+ tools.find((t) => /catalog|list|capabilit|health|search|browse/i.test(t.name));
if (!readOnly) return record(`${label} tools/call`, `POST ${path}`, 'skip', 'no read-only tool advertised');
const call = await req(path, {
method: 'POST',
@@ -479,6 +482,36 @@ async function probeMcp(path, label) {
record(`${label} tools/call`, `${readOnly.name}`, call.status === 200 && content.length > 0 && !call.json?.result?.isError, call.status === 200 ? `${readOnly.name} → ${content.length} content block(s)` : `HTTP ${call.status}: ${call.text.slice(0, 140)}`);
}
+// The talking-avatar lane (workers/longcat). Its own README records the worker
+// as built but not deployed, so the bar this leg holds production to is the one
+// that matters to a user: either a real MP4 comes back, or the site refuses
+// cleanly with 503 worker_unconfigured and burns none of the caller's free
+// trial. A 500, or a 202 that never produces a video, is a failure.
+async function probeTalkingAvatar() {
+ if (!selected('talking-avatar')) return;
+ const cookie = await qaSession();
+ const headers = cookie ? { cookie } : {};
+ const post = await req('/api/avatar/video-generate', {
+ method: 'POST',
+ headers,
+ body: { image_url: REF_IMAGE, audio_url: `${BASE}/audio/notification.mp3` },
+ timeoutMs: 60_000,
+ });
+ if (post.status === 503) {
+ return record('talking-avatar', 'POST /api/avatar/video-generate', false, `worker not deployed: ${post.json?.error || post.text.slice(0, 120)} (LONGCAT_WORKER_URL unset)`);
+ }
+ if (post.status !== 202 && post.status !== 200) {
+ return record('talking-avatar', 'POST /api/avatar/video-generate', false, `HTTP ${post.status}: ${post.text.slice(0, 160)}`);
+ }
+ const jobId = post.json?.job_id;
+ if (!jobId) return record('talking-avatar', 'POST /api/avatar/video-generate', false, `no job_id: ${post.text.slice(0, 160)}`);
+ const done = await pollJob(`/api/avatar/video-status?job=${encodeURIComponent(jobId)}`, { jobKey: 'talking-avatar', headers, budgetMs: 600_000 });
+ const url = done.json?.video_url;
+ if (!url) return record('talking-avatar', 'POST /api/avatar/video-generate', false, `job ${done.json?.status || done.status}: ${done.json?.error || done.text.slice(0, 160)}`);
+ const s2 = await sniff(url);
+ record('talking-avatar', 'POST /api/avatar/video-generate', s2.ok && s2.kind === 'mp4', s2.ok ? `${s2.bytes} B ${s2.kind}` : s2.detail, { artifact: url });
+}
+
async function main() {
console.log(`Generation-suite probe against ${BASE}\n`);
await probeCatalogHealth();
@@ -491,6 +524,7 @@ async function main() {
if (selected('rembg')) await meshJob('rembg', '/api/forge-rembg', { image_url: REF_IMAGE }, 'result_url', 'png');
await probeMotion();
await probeRetexture();
+ await probeTalkingAvatar();
await probeX402();
await probeMcp('/api/mcp-3d', 'mcp-3d');
await probeMcp('/api/mcp-studio', 'mcp-studio');
diff --git a/workers/texture/README.md b/workers/texture/README.md
index 5af13942d..c5a793804 100644
--- a/workers/texture/README.md
+++ b/workers/texture/README.md
@@ -32,11 +32,20 @@ other half of "the surface reads as real" this campaign targets.
## Status: built, not currently deployed
There is **no `texture-service` running in `aerial-vehicle-466722-p5`** (checked
-in every region on 2026-08-11: no Cloud Run service, no Artifact Registry repo),
-and `three-ws-api` carries no `GCP_TEXTURE_URL`. Both callers below therefore
-take their designed missing-lane path today: the HTTP gateway answers `501
-region_retex_unconfigured` and the MCP tool raises a JSON-RPC error naming the
-two env vars. Nothing is faked and no other lane is affected. Re-check with:
+in every region on 2026-08-11 and again on 2026-09-04: no Cloud Run service, no
+Artifact Registry repo), and `three-ws-api` carries no `GCP_TEXTURE_URL`. Both
+callers below therefore take their designed missing-lane path today: the HTTP
+gateway answers `501 region_retex_unconfigured` (verified live against
+three.ws on 2026-09-04) and the MCP tool raises a JSON-RPC error naming the two
+env vars. Nothing is faked and no other lane is affected.
+
+**Step 1 of "Bringing it up" is already done.** The checkpoints were staged on
+2026-09-04: `gs://three-ws-model-weights/sdxl-texture` holds 15.7 GiB across
+the four repos below, so a deploy no longer has to download SDXL inside the
+first request. Only the two `gcloud` commands (build/deploy, then point the site
+at the service) remain, and both need owner approval per the deploy gate in
+CLAUDE.md. Re-check the staged tree with `gsutil du -s
+gs://three-ws-model-weights/sdxl-texture`; re-check the service with:
```bash
gcloud run services list --project aerial-vehicle-466722-p5 --region us-central1 | grep texture
@@ -59,7 +68,7 @@ Bringing it up is three commands, in this order:
# 1. stage the checkpoints (~15 GiB, once, and again on any checkpoint change).
# Needs huggingface_hub and google-cloud-storage on the host, and credentials
# the storage client can find (a service account, or gcloud application
-# default credentials).
+# default credentials). DONE on 2026-09-04; re-run only on a checkpoint change.
python3 workers/texture/stage_weights.py --prefix sdxl-texture
# 2. build + deploy from the repo root (the build step declares dir: workers/texture,