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(prompts): record what QB-03 finished and what a rejected storage credential blocks

nirholas4ce643fdd411
diff --git a/prompts/finish/006-quality-bar-03-gpu-fleet-scaleout.md b/prompts/finish/006-quality-bar-03-gpu-fleet-scaleout.md
index 9eadaa2c7..4cf2cd039 100644
--- a/prompts/finish/006-quality-bar-03-gpu-fleet-scaleout.md
+++ b/prompts/finish/006-quality-bar-03-gpu-fleet-scaleout.md
@@ -16,6 +16,41 @@ It is complete on its own. Also read `prompts/finish/_context/quality-bar-_share
share this worktree. Do not `git push` unless the owner's message asked for it.
4. GCP spend is pre-approved (credits). Never trade quality for cost.
+## Session status as of 2026-09-08 (read before re-running anything)
+
+A prior session executed this order. Do not redo tasks 2 and 3.
+
+- **Task 2 (cold-start honesty): DONE and verified.** `/forge` renders the boot state at
+ 320/768/1440 with zero console errors, its countdown uses the lane's real
+ `cold_start_seconds`, and its elapsed meter is now anchored to the server's
+ `elapsed_seconds` instead of a tab-local clock. `/api/gpt-forge` now returns
+ `cold_start` on POLL frames (it only did so at submit, which is why the MCP
+ surfaces, who see nothing but poll frames, could never report a boot), and the
+ `_mcp-studio` tools narrate it. The wording lives once in
+ `src/shared/forge-frames.js` (`coldStartState` / `coldStartLabel`), used by the
+ homepage chamber, the Studio, the in-world prop forge and the MCP tools.
+- **Task 3 (keep-warm): was ALREADY shipped before this order was written.**
+ `/api/cron/gpu-keepwarm` runs on exactly the `*/10 14-23,0-4` schedule this file
+ asks for, and is deliberately quota-aware rather than warming every minScale=0
+ worker (the us-central1 L4 grant is 3 and two floors already hold 2; pinning a
+ third starved `model-text2motion` on 2026-07-26). The gap that was fixed:
+ `model-hunyuan3d`, the fleet's longest spin-up, was missing from the registry
+ entirely, so no override could ever reach it. `tests/cron-forge-lane-guards.test.js`
+ now fails if a routed scale-to-zero lane is absent or carries the wrong url env.
+- **Task 1 (scale ceilings): BLOCKED on one `gcloud auth login`.** Also note its target
+ numbers conflict with a twice-learned lesson recorded in `docs/ops/gcp-credits-plan.md`:
+ at a granted L4 quota of 3, `model-trellis` min 1 + `model-hunyuan3d` min 1 + `model-rig`
+ min 1 consume the whole pool and trellis can never burst to its stated max 3. Read the
+ live grant before applying them; if it is still 3, raise the grant first.
+- **Tasks 4 and 5 (load test, prove a lane): BLOCKED by a production outage, not by capacity.**
+ Every `/api/forge` generation fails with `SignatureDoesNotMatch` from object storage: the
+ R2 credential on the Cloud Run service is rejected, so no lane can park a reference image
+ or a finished mesh. 24h metrics read 12 attempts, 0 successful. Only a pre-outage CACHED
+ result still returns. There is nothing to load test until the credential is rotated and the
+ pending commits deploy (the graceful-degradation fix for this exact class landed
+ 2026-09-07 and is not yet in production, which is why the raw AWS signing sentence is
+ currently shown to users).
+
## Step 0: re-derive current state (do this first, trust nothing below)
This file's claims rot. Measure before you plan:
diff --git a/prompts/finish/_context/quality-bar-README.md b/prompts/finish/_context/quality-bar-README.md
index 7269080f3..b4919ac13 100644
--- a/prompts/finish/_context/quality-bar-README.md
+++ b/prompts/finish/_context/quality-bar-README.md
@@ -18,7 +18,7 @@ special case.
| # | File | Owns | State |
|---|------|------|-------|
-| 03 | [03-gpu-fleet-scaleout.md](../006-quality-bar-03-gpu-fleet-scaleout.md) | Scale ceilings, cold-start honesty, keep-warm, load test | Partial. The TripoSG and text2motion deploys it opened with are shipped; scale, cold-start UX and the load test are open. |
+| 03 | [03-gpu-fleet-scaleout.md](../006-quality-bar-03-gpu-fleet-scaleout.md) | Scale ceilings, cold-start honesty, keep-warm, load test | Partial. Cold-start honesty (task 2) and keep-warm (task 3) are DONE and verified; do not redo them. Scale ceilings (task 1) need one `gcloud auth login`. The load test (tasks 4-5) is blocked by a production outage, not by capacity: every `/api/forge` generation fails on a rejected object-storage credential, so nothing can be measured until it is rotated and the pending commits deploy. |
| 04 | [04-pbr-texture-material-realism.md](../003-quality-bar-04-pbr-texture-material-realism.md) | Full PBR sets, measured presets, skin/eye/hair | Partial. Everything but the bench delta is verified: the channel matrix, the derive pass wired opt-out into `forge-store.js`, the presets rendered under all four shipped HDRIs (`docs/assets/material-presets/`), the avatar skin/eye/hair readback at 320/768/1440 (the check had been aimed at `/viewer`, which never runs that pass, and had been reporting zero meshes), and the mobile payload gate with measured 4K byte sizes (`npm run check:glb:payload`). The quality-bench comparison needs Vertex, and this Codespace's gcloud credentials have expired: `gcloud auth application-default login` unblocks it. |
| 06 | [06-forge-ux-flow.md](../004-quality-bar-06-forge-ux-flow.md) | The `/forge` flagship experience | Partial. History, compare mode and starters shipped; the result-moment click-through and the clean audit sweep are open. |
| 07 | [07-design-system-sweep.md](../002-quality-bar-07-design-system-sweep.md) | Tokens, states, microinteractions sitewide | Open |