Operations inventory (as of 2026-07-03)

Live deployments (billing)

App	Hardware	Why	URL
`higgs-avatar-v1`	1× A100-80GB, hot 24/7 (`min_containers=1`)	THE avatar endpoint (official v1.0.0)	https://bosonai--avatar.modal.run
`wan-demo-api`, `elastic-demo`	CPU, warm control planes (cheap)	scale-up demo site backends	demo: https://elastic-scaleup-demo.pages.dev
`wan-avatar-shim`	A100 on demand (scales to zero)	diffsynth fallback backend	bosonai--wan-avatar-shim-shim-web.modal.run

Stopped (redeployable from milestones/): wan-web-stream (74-API dev-image endpoint, superseded by v1.0.0), wan-web, wan-serve-render, all test/introspection apps.

Hidden dependencies

Docker Hub mirror orzeai/higgs-avatar-inference:a100-v1.0.0-5c0cc03 (PRIVATE) — the only registry copy Modal can pull. Owned by the orzeai Hub account (creds from the build host's keychain; also in Modal secret dockerhub-orzeai). If that account/repo goes away, redeploys break — re-run the transfer (/tmp/higgs.tar.gz on the build host, or re-download from totoro: s3://avatar-junming/sglang-offline-single/images/…5c0cc03.tar.gz) or wire up GHCR per the release DEPLOY.md.
totoro S3 (http://totoro.canada.boson.ai) is VPN-internal — unreachable from Modal; any image/weight pull must route through a boson-network host.

Secret	Purpose	Status
`dockerhub-orzeai`	pull the private v1.0.0 mirror	required by `wan_higgs_v1.py`
`boson-s3`	totoro S3 (weights/images)	unused at runtime (VPN-blocked from Modal anyway)
~~`avatar-realtime-s3`~~	one-time weight provisioning	deleted 2026-07-03 (contained leaked keys)

wan-upstream (17 GB DiffSynth weights — needed to rebuild the shim/dev images). wan-out and wan-compile-cache were deleted 2026-07-03.

🔐 Credential rotation checklist (keys passed through chat/terminal — rotate all)

avatar-realtime AWS key AKIAYGZJ74QIJZUX22DQ
boson totoro key EQD0NYBEGA1AJMI29VF6 (also embedded in the dev image at /opt/wan-code/docs/superpowers/plans/2026-05-26-boson-ai-mirror.md — tell the image owner)
consider rotating the orzeai Docker Hub password after the mirror is moved to GHCR

Build-host artifacts

/tmp/higgs.tar.gz (39 GB release tarball — cold archive, deletable), /tmp/animate0.b64 (sample ref image for the e2e client).

Acceptance / regression

Avatar endpoint: python3 milestones/v1_e2e_client.py (PASS/FAIL, exit code).
Readiness is enforced at the infrastructure level (hardening 2026-07-03): the web port is a warmup-gated forwarder, so Modal never routes to a replica until accepts_new_session:true. Clients need NO readiness gate — if a request is routed, the replica is session-ready. A watchdog replaces dead/unhealthy containers; a 5-min scheduled probe fails loudly in the dashboard.
Access decision (2026-07-06): endpoint stays PUBLIC, no auth (debug/integration phase; owner-accepted risk: anyone with the URL can occupy the single render lane). To lock down later: requires_proxy_auth=True.
Redeploy cold start ≈ 12 min (single-bucket warmup); URL is label-pinned and never changes.

Operations — live inventory, secrets, rotation