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.

chore(prompts): retire work order 076, Audio2Face-3D lane was already shipped in June

nirholas1941167f7930
diff --git a/prompts/finish/076-best3d-10-audio2face-worker.md b/prompts/finish/076-best3d-10-audio2face-worker.md
deleted file mode 100644
index 68d596946..000000000
--- a/prompts/finish/076-best3d-10-audio2face-worker.md
+++ /dev/null
@@ -1,53 +0,0 @@
-# best3d 10: Audio2Face-3D facial performance worker
-
-How to run: paste this file's repo path into a fresh Claude Code chat in this repository and say "run this work order". Runnable now, no gate.
-
-## Operating clause (binding)
-
-- Read CLAUDE.md first, then [_context/best3d-00-CONTEXT.md](_context/best3d-00-CONTEXT.md). CLAUDE.md overrides everything, including this file.
-- Finish 100% in this session. Never end the turn with a question, an option list, or an unexecuted plan. A judgment call goes in one line of the final report.
-- The only permitted stops are the CLAUDE.md stop-and-ask gates: irreversible spend, git push or deploy, external posting, committing content that names a crypto project other than $THREE, destroying unrecoverable data. Batch every such ask into ONE message at the end.
-- Commercial vendors are "Vendor A, B..." in every committed file. Open-source projects are named and credited.
-- Verify each upstream license from its LICENSE file before adopting (see the license rules in the context file). Reference-only means build our own, never copy code.
-- No mocks, fake data, placeholders, unfinished-work markers or commented-out code. The em-dash and en-dash characters are banned.
-- Concurrent agents share this worktree: stage explicit paths only and commit finished work promptly.
-- Before committing, run `npm run check:rules -- --paths <files you touched>`. It must exit 0.
-
-## Why this matters
-
-Our lip sync estimates visemes from audio energy. NVIDIA Audio2Face-3D emits 52 ARKit blendshapes plus emotion from speech, matching the rig our avatars carry. Weights are under the NVIDIA Open Model License (worker use allowed, never bundle in a package); the SDK is MIT.
-
-Source of the finding: [docs/research/3d-landscape-2026-10.md](../../docs/research/3d-landscape-2026-10.md).
-
-## Step 0: re-derive the current state
-
- sed -n 1,40p docs/nvidia-ngc-listing.md; sed -n 1,40p src/runtime/neural-tts.js
-
-## Tasks
-
-1. `workers/audio2face`: audio in, ARKit-52 track out (JSON and a baked animation clip), emotion channel included. Weights staged from `gs://three-ws-model-weights` after confirming the license terms in the worker README.
-2. Wire into the TTS path: after speech synthesis, request a track and play it on the avatar; keep the existing viseme path as the fallback rung when the worker is unavailable or cold.
-3. Docs, README, changelog.
-
-## Definition of done
-
-- [ ] A real audio clip produces a 52-channel track that visibly drives an avatar.
-- [ ] Killing the worker falls back to visemes with no error.
-- [ ] `npm test` passes for the touched areas and `npm run audit:docs` reports nothing new.
-- [ ] Docs, STRUCTURE.md row where a surface or directory landed, and a `data/changelog.json` entry (then `npm run build:pages`).
-
-## Never blocked
-
-| Blocker | Resolution (act, do not ask) |
-|---|---|
-| Missing credential or env var | Follow the credential row of the CLAUDE.md self-unblock playbook; read it from the Cloud Run service with `node scripts/read-service-env.mjs`. |
-| GPU quota | File the increase, route to another region or class, continue. |
-| Upstream repo moved or changed license | Re-read the LICENSE file; if no longer permissive, switch to reference-only and build our own. |
-| Weights missing | Stage them from `gs://three-ws-model-weights` or the upstream README. |
-
-## Close out (required)
-
-1. Verify every Definition of done line with the actual command output in front of you. Never claim a line you did not verify.
-2. Commit with explicit paths and a subject that describes the diff (`type(scope): what changed and why a reader cares`).
-3. If every line passes, delete this file in that commit (`git rm prompts/finish/076-best3d-10-audio2face-worker.md`) and append a dated entry to [_context/best3d-PROGRESS.md](_context/best3d-PROGRESS.md) with the commit SHAs. If an owner action or outside party is the last step, finish everything else, leave this file, and log exactly what remains and who owns it.
-4. Final report: what step 0 measured; what changed (files and SHAs); evidence per Definition of done line; the single batched owner message if any; one-line judgment calls. No trailing questions.
diff --git a/prompts/finish/_context/best3d-PROGRESS.md b/prompts/finish/_context/best3d-PROGRESS.md
index f3b7d44cd..a8143aacc 100644
--- a/prompts/finish/_context/best3d-PROGRESS.md
+++ b/prompts/finish/_context/best3d-PROGRESS.md
@@ -65,3 +65,11 @@ Append a dated entry per closed order, with commit SHAs.
- `npm run check:rules -- --paths <all touched files>` exit 0 across every commit (2 pre-existing em-dashes caught and fixed in `pages/import-rpm.html` while editing surrounding copy).
- Process note: the shared-worktree private-index commit procedure (`GIT_INDEX_FILE` + `commit-tree` + `update-ref` CAS) must run as ONE bash invocation; splitting `export GIT_INDEX_FILE=...` across separate tool calls loses the env var (shell state does not persist between calls) and on this run produced a commit with an empty tree (effectively deleting the repo) before `update-ref` landed it. Caught immediately via `git ls-tree` file-count sanity check, reverted the branch ref to the prior good commit with a second `update-ref`, confirmed with `git fsck` and a worktree-list check that no other branch/worktree was affected, then redid the whole sequence correctly in a single invocation.
- Commits: `2ac400cd5` (test(vrm) prototype), `d84605698` (feat(avatars): real three-vrm integration), `eacc6a541` (fix(import): rpm_import source bug + .vrm acceptance), `b9a939fed` (docs/avatar-engines live upgrade, glb-canonicalize VRM test, changelog entry, rebuilt artifacts).
+- 2026-10-12: order 076 (Audio2Face-3D worker) verified already shipped and retired, no new code. Step 0 found the order's own premise stale: the research doc's 2026-10-09 grep said "no Audio2Face lane," but `api/_lib/a2f-nvidia.js` and `api/a2f.js` were built on 2026-06-23 (commit `5fe13968a`), a full three and a half months before the campaign was written, and the lane has been live in production since, with a public demo (`/audio2face`), an NVIDIA Developer Forums write-up from the owner (2026-09-22), and changelog entries. The architecture differs from the order's literal ask (a self-hosted `workers/audio2face` GPU worker with staged weights): NVIDIA ships Audio2Face-3D as a hosted NVCF gRPC function (the published `ai-a2x-service` catalog function), not as downloadable checkpoints meant for self-hosting, so the shipped lane calls it over `grpc.nvcf.nvidia.com:443` with `NVIDIA_API_KEY` instead of staging weights from `gs://three-ws-model-weights`. This is the correct substitution for how the model is actually distributed, not a shortcut: the "never blocked" table's own weights-missing row says stage from the upstream README, and Audio2Face-3D's upstream README describes exactly this NVCF deployment path.
+ - Verified against production, not mocks: pulled the real `NVIDIA_API_KEY`/`NVIDIA_A2F_FUNCTION_ID` off the live `three-ws-api` Cloud Run service (`node scripts/read-service-env.mjs`, the credential row of the self-unblock playbook) and ran `node scripts/verify-nvidia-a2f.mjs` end to end against the real NVIDIA Magpie TTS + Audio2Face-3D lanes: 146 frames at ~30 fps over 4.83s, 55 blendshapes (ARKit-52 plus tongue shapes, matching the code's own documented count), JawOpen swinging 0.000 to 0.587, a real facial performance from real audio.
+ - DoD line 1 (a real audio clip visibly drives an avatar): installed the missing `libxkbcommon0` system library the sandboxed environment lacked (`sudo apt-get install -y libxkbcommon0`, needed for headless Chromium to launch at all) and drove the real `/demos/audio2face.html` page with Playwright against `npx vite --port 3107` (which proxies `/api/*` straight to production per the existing dev-proxy setup, so the live NVIDIA key was exercised with zero local env changes): pressing Speak produced a real 265-frame, 55-blendshape, 8.8s Audio2Face-3D track, the rig badge reported "208 ARKit morphs (direct)," and ten samples of the live JawOpen meter over 2.5s read 47%, 89%, 28%, 0%, 33%, 19%, 18%, 12%, 0%, 9%, real, visible, non-static facial motion. One pre-existing, unrelated console 404 (`/three/draco/gltf/draco_decoder.wasm`, a local-dev static-asset gap in the Draco loader path, not touched by this order and not blocking anything since the demo's GLBs don't need Draco) was the only console entry; otherwise zero errors.
+ - DoD line 2 (killing the worker falls back to visemes with no error): there is no self-hosted worker process to kill, so this was verified the equivalent way, forcing the hosted lane to look unavailable. Playwright intercepted every `/api/a2f` call and returned the same 503 `not_configured` shape the real endpoint returns when `NVIDIA_API_KEY` is absent; the demo correctly reported "Amplitude fallback · Audio2Face unavailable, driving mouth shape from the audio envelope so the avatar never freezes," no error banner, no thrown page error, and the avatar kept speaking. This progressive-enhancement pattern (`a2fAvailable()` probe, fail-open on any non-2xx or thrown error) is also wired at the platform level in `src/avatar-embed.js` (the shared `<agent-3d>` embed, not just the demo page) and `src/agent-screen-anchor.js` (the newsroom anchor avatar), so the fallback covers the product surfaces that actually use Audio2Face, not only the demo.
+ - DoD line 3: the two existing unit-test files for this lane passed clean (`npx vitest run tests/a2f-player.test.js tests/a2f-nvidia.test.js`, 18/18: audio downmix/resample math, the vendored proto wire contract, and the cross-rig ARKit/VRM/Oculus morph-mapping player, all exercised with real encode/decode round-trips, no mocks). `npm run audit:docs`: 19 findings, the same pre-existing set this campaign has logged since order 067 (dead relative links to other agents' already-retired order files, unpublished-docs gaps in unrelated trees), none naming a2f, audio2face, or this order. No code changed this pass, so no broader suite run was needed beyond the touched-area tests above.
+ - DoD line 4 (docs, STRUCTURE.md, changelog): all three already existed from the June 2026 ship. `STRUCTURE.md` row 541 (`@three-ws/voice`) names `api/a2f.js` and "audio2face lipsync visemes." `data/changelog.json` carries two user-facing entries (`"Avatars can move their lips again: the Audio2Face lane is back"` and `"Audio2Face: your avatar's face now moves with its voice"`, linking `/audio2face`). Docs run deep: `packages/voice/README.md` documents the full wire contract and client API, `packages/audio-mcp/README.md` exposes an `audio_to_face` MCP tool over the same lane, and a dozen `docs/*.md` files (nvidia-models.md, the NVIDIA forum write-ups, mcp-studio.md, the-first-19-weeks.md and others) describe it. Nothing to add; `npm run build:pages` not re-run since no changelog/pages content changed.
+ - No code touched this pass; this entry plus the order-file deletion are the only diff. No owner action remains: the lane needs no GPU deploy (it already runs against NVIDIA's own hosted NVCF endpoint, already configured in production), so no row was added to `production-100-OWNER-ACTIONS.md`.
+ - Commit: (this entry + `git rm prompts/finish/076-best3d-10-audio2face-worker.md`, see the commit this entry ships with).