immersive-studio 0.1.6

Immersive Studio CLI and SDK: spec generation, ComfyUI textures, Unity packs, and HTTP worker API.

Immersive Studio (CLI + SDK)

Publishable PyPI package immersive-studio: the immersive-studio terminal command, importable Python SDK immersive_studio, and the same FastAPI worker used by the Video Game Generation Studio (/studio in @immersive/web). End-user orientation (deploy, ComfyUI, Blender, GCP + Cloudflare): /docs on the marketing site (apps/web/src/pages/DocsPage.tsx). Features include Ollama/mock spec generation, ComfyUI texture passes, optional Blender placeholder .glb, Unity pack layout (manifest.json, spec.json, pack.zip), and the HTTP API.

Install (PyPI — like any other CLI)

After the package is published on PyPI as immersive-studio (maintainers: see Publishing below), end users can install globally with pipx so the executable lands on PATH without touching system Python:

pipx install immersive-studio
immersive-studio doctor

Or into the active virtual environment:

pip install immersive-studio
python -m studio_worker.cli doctor

Optional extras: pip install immersive-studio[postgres], [redis], [s3], or [scale] for production-style backends (see Environment table below).

pip install -U still shows 0.1.0? PyPI may not list newer wheels yet (run pip index versions immersive-studio). Until maintainers publish a newer version (GitHub Actions PyPI publish workflow), install from this monorepo so worker_version matches apps/studio-worker/pyproject.toml:

pip install -e ./apps/studio-worker

Then restart the API process (immersive-studio serve, Docker, or systemd) so the running server loads the new code.

Note: The distribution was previously named immersive-studio-worker. Use immersive-studio going forward.

Python SDK

from immersive_studio import __version__, run_studio_job

result = run_studio_job(
    user_prompt="wooden barrel",
    category="prop",
    style_preset="toon_bold",
    use_mock=True,
    generate_textures=False,
    unity_urp_hint="6000.0.x LTS (pin when smoke-tested)",
    export_mesh=False,
)
print(result["job_id"], result["zip_path"])

Public symbols: __version__, run_studio_job, validate_asset_spec_file, generate_asset_spec_with_metadata, comfy_base_url, comfy_reachability. Advanced integrations can import from the studio_worker package directly.

Publishing to PyPI

I cannot create or “register” a PyPI project on your behalf (that requires you to sign in at pypi.org). Use an API token as below to upload from CI or your laptop.

GitHub Actions publish (this repo)

The workflow .github/workflows/publish-immersive-studio.yml uploads with twine only. It does not use OIDC or pypa/gh-action-pypi-publish, so you will not see invalid-publisher from that path.

1. Create a PyPI API token with upload scope for immersive-studio (PyPI → API tokens).
2. In GitHub: Settings → Secrets and variables → Actions → Repository secrets → New repository secret.
   • Name: PYPI_API_TOKEN (exact spelling, all caps).
   • Value: the token string (often starts with pypi-).
     Do not put it under Dependabot or Codespaces only; it must be a repository secret for this repo.
3. Run Actions → “PyPI publish (immersive-studio package)” → Run workflow on main, or publish a GitHub Release to trigger it. Use a new run after changing secrets (re-running a very old failed job can still use old workflow YAML).

If the job fails immediately with “PYPI_API_TOKEN is missing or empty”, GitHub did not inject the secret (wrong name, empty value, or workflow running from a fork without secrets).

Optional: PyPI trusted publishing (OIDC)

OIDC is not used by the publish workflow above. To use trusted publishing instead, upload from your machine with twine or configure a separate workflow; PyPI’s form needs your owner/repo, workflow filename publish-immersive-studio.yml, and an empty environment name unless you add a matching GitHub Environment.

Manual upload (API token)

From apps/studio-worker with build and twine:

pip install build twine
python -m build
twine upload dist/*

Use a PyPI API token with upload scope; the first upload can create the project if the name is free.

If the project name immersive-studio is already taken on PyPI, change name in pyproject.toml to an available name (for example immersive-labs-studio) and update install docs accordingly.

When you change packages/studio-types/schema/studio-asset-spec-v0.1.schema.json, run python scripts/sync-studio-asset-schema.py from the monorepo root (or copy manually to src/studio_worker/data/studio-asset-spec-v0.1.schema.json). See docs/studio/releasing.md and root CHANGELOG.md for full release steps.

Setup (monorepo / contributors)

From repository root:

cd apps/studio-worker
python -m venv .venv
source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install -e .

Invoking the CLI: After pip install -e ., the immersive-studio command lives next to your Python interpreter (for example .venv\Scripts\immersive-studio.exe, or %APPDATA%\Python\Python3xx\Scripts\immersive-studio.exe when pip used a user install). If your shell says command not found, either add that Scripts folder to your PATH, or call the module directly (works from any directory once the package is installed):

python -m studio_worker.cli doctor

From this app directory you can also use the repo wrappers (no PATH change):

./scripts/immersive-studio.sh doctor          # Git Bash / MSYS
scripts\immersive-studio.cmd doctor            # cmd.exe / PowerShell

Ensure Ollama is running locally if you use LLM generation (default http://127.0.0.1:11434). Pull a JSON-capable model, for example:

ollama pull llama3.2

For texture generation, the worker calls ComfyUI at STUDIO_COMFY_URL. When unset, the default is https://comfy.immersivelabs.space (override with http://127.0.0.1:8188 for local ComfyUI). Match STUDIO_COMFY_CHECKPOINT to a checkpoint on that host. See comfy/README.md.

Why there is no blender.immersivelabs.space: ComfyUI exposes an HTTP API (/prompt, /system_stats), so a public URL makes sense. Mesh export runs Blender as a subprocess on the worker host (blender --background --python …); it is not a remote URL. The Docker image installs Blender so operators do not rely on end-user installs. A future “remote Blender service” would need its own HTTP API — not the same knob as STUDIO_COMFY_URL.

After installing ComfyUI and (optional, for local mesh) Blender, verify the worker can reach ComfyUI:

python -m studio_worker.cli doctor

Exit code 0 means ComfyUI answered at STUDIO_COMFY_URL (default when unset: https://comfy.immersivelabs.space). Blender is optional unless you pass doctor --strict (CI / full local toolchain). The Docker image installs Blender for mesh export without end-user installs. Use doctor --comfy-url … to probe a URL once.

To launch ComfyUI from a git checkout once dependencies are installed, set COMFYUI_ROOT and run scripts/start-comfyui-dev.sh or scripts/start-comfyui-dev.ps1 (listens on 127.0.0.1:8188).

Docker (VM / production-shaped)

The image copies packages/studio-types/schema and apps/studio-worker into a layout that matches STUDIO_REPO_ROOT (defaults to /repo in the image).

Build (from monorepo root):

docker build -f apps/studio-worker/Dockerfile -t immersive-studio-worker:local .

Run (set CORS to every browser origin that loads /studio — comma-separated, no trailing slashes):

docker run -d --name studio-worker --restart unless-stopped \
  -p 8787:8787 \
  -e STUDIO_CORS_ORIGINS='https://YOUR-APP.vercel.app,https://www.yourdomain.com,https://yourdomain.com' \
  -v studio-output:/repo/apps/studio-worker/output \
  immersive-studio-worker:local

If you only use the default Vercel hostname, a single origin is enough: https://YOUR-APP.vercel.app. Trailing slashes in the env var are ignored.

On a cloud VM behind Cloudflare Tunnel, bind Docker to localhost only: -p 127.0.0.1:8787:8787 (see deploy-gcp-free-vm.md).

Compose (dev smoke test; overrides STUDIO_CORS_ORIGINS as needed):

docker compose -f apps/studio-worker/docker-compose.yml up --build

Free GCP VM + HTTPS + Vercel: docs/studio/deploy-gcp-free-vm.md (includes scripts/studio-cloudflare-tunnel/ for gcloud + Cloudflare Tunnel).

CLI

In the examples below, immersive-studio is shorthand for python -m studio_worker.cli (or the .exe on your PATH after you add Python’s Scripts folder — see Setup).

# Generate spec via Ollama (default model from env STUDIO_OLLAMA_MODEL or llama3.2)
immersive-studio generate-spec --prompt "wooden crate with iron bands" --style-preset toon_bold --category prop

# Validated mock spec without Ollama (CI / offline)
immersive-studio generate-spec --mock --prompt "neon barrel" --style-preset anime_stylized --category prop

# Validate a JSON file against schema + studio rules
immersive-studio validate-spec --file ./spec.json

# Write pack folder: manifest.json, spec.json, UnityImportNotes.md, Models/ + Textures/ layout
immersive-studio pack --spec ./spec.json --output ./out/MyPack

# Full persisted job under output/jobs/ (spec + pack + zip + index); optional ComfyUI textures
immersive-studio run-job --mock --prompt "wooden barrel" --textures

# Optional: ComfyUI on a non-default URL for this run only
immersive-studio run-job --mock --prompt "wooden barrel" --textures --comfy-url http://127.0.0.1:8188

# Same, plus headless Blender placeholder `.glb` under Models/ (requires Blender on PATH or STUDIO_BLENDER_BIN)
immersive-studio run-job --mock --prompt "crate" --export-mesh

# Check ComfyUI + Blender before a full job
immersive-studio doctor

# HTTP API for the web UI (CORS allows Vite dev server)
immersive-studio serve --host 127.0.0.1 --port 8787

# Poll SQLite-backed job queue (run multiple terminals or processes for parallel workers)
immersive-studio queue-worker --worker-id worker-a
immersive-studio queue-worker --once --worker-id one-shot

# Multi-tenant / indie SaaS (after STUDIO_API_AUTH_REQUIRED=1 on the server)
immersive-studio tenants create --name "Pixel Friends" --tier indie
immersive-studio tenants issue-key --tenant-id <uuid-from-create> --label "lead-dev-laptop"
immersive-studio tenants list
immersive-studio tenants set-tier --tenant-id <uuid> --tier team

HTTP API (selected)

Method	Path	Purpose
GET	/api/studio/health	Worker liveness (auth_required flag for clients)
GET	/api/studio/metrics	Queue counts by status, jobs_indexed, and slo (SQLite: oldest pending/running row ages for alerting; tenant-scoped when auth on; Cache-Control: no-store)
GET	/api/studio/comfy-status	ComfyUI /system_stats probe (no API key)
GET	/api/studio/usage	Tier + monthly credits used/cap (requires key when auth on; Cache-Control: no-store)
GET	/api/studio/dashboard	One JSON: usage + billing + jobs + worker_hints + queue_slo (same shapes as /usage, /billing/status, /jobs, and /metrics → slo; Cache-Control: no-store)
POST	/api/studio/generate-spec	Prompt → validated spec
POST	/api/studio/pack	Ad-hoc pack (scoped under output/packs/… when auth on)
POST	/api/studio/jobs/run	Full job → output/jobs/job_*/ + pack.zip + index
POST	/api/studio/queue/jobs	Enqueue async job (SQLite or Postgres); optional idempotency_key dedupes per tenant; response deduplicated
GET	/api/studio/queue/jobs	Recent queue rows (tenant-scoped when auth on; Cache-Control: no-store)
GET	/api/studio/queue/jobs/{queue_id}	Single queue row (no-store on success)
GET	/api/studio/queue/jobs/{queue_id}/events	SSE (text/event-stream): pushes event: job when the row changes (same JSON as GET row); /studio uses fetch streaming with Authorization, then falls back to polling if the stream fails
GET	/api/studio/jobs	List persisted jobs + jobs_root path (tenant-scoped when auth on; Cache-Control: no-store)
GET	/api/studio/jobs/{job_id}/download	Download pack.zip (local file) or 302 to a remote URL when STUDIO_JOB_ARTIFACTS uploads zips (no-store)
GET	/api/studio/paths	Debug paths + active backends (queue_backend, redis_queue_engine when Redis, tenants_backend, …; no-store)
GET	/api/studio/billing/status	Stripe readiness + linked customer (requires key when auth on; Cache-Control: no-store)
POST	/api/studio/billing/checkout-session	Returns Stripe Checkout URL for { "tier": "indie" | "team" }
POST	/api/studio/billing/portal-session	Returns Stripe Customer Portal URL (manage/cancel)
POST	/api/studio/billing/webhook	Stripe webhooks (signing secret); no API key

Remote clients send Authorization: Bearer <api_key> or X-API-Key. The /studio UI stores an optional key in localStorage and uses it for fetch + zip download.

Environment

Variable	Default	Purpose
STUDIO_OLLAMA_URL	http://127.0.0.1:11434	Ollama base URL
STUDIO_OLLAMA_MODEL	llama3.2	Chat model name
STUDIO_OLLAMA_READ_TIMEOUT_S	3000	Base seconds for each Ollama /api/chat read (30–7200); after a read timeout the client retries once with ~1.5× this base (capped 7200/attempt). Streaming (default) resets the read timer between NDJSON chunks, so modest caps tolerate slow token generation better than STUDIO_OLLAMA_STREAM=0. See docs/studio/essentials.md (Production API stability).
STUDIO_OLLAMA_STREAM	on	stream: true (NDJSON) by default. Set 0 / false / no / off for a single blocking HTTP response if your model misbehaves with streaming.
STUDIO_OLLAMA_NUM_PREDICT	4096	Ollama options.num_predict — caps output length for the JSON spec (512–16384); lowers risk of endless generation on weak VMs. Raise if validation fails with truncated JSON.
STUDIO_OLLAMA_NUM_CTX	—	Optional Ollama options.num_ctx (e.g. 2048) to reduce RAM on small hosts when supported by the model.
STUDIO_RATE_LIMIT_ENQUEUE_PER_MINUTE	60	Sliding-window cap on POST /api/studio/queue/jobs per API key / IP bucket; 0 disables.
STUDIO_QUEUE_SSE_POLL_S	1	Server poll interval (seconds) for GET /queue/jobs/{id}/events (SSE).
STUDIO_REPO_ROOT	—	When set, writable job/queue data uses apps/studio-worker/output under this monorepo root (Docker / local dev).
STUDIO_WORKER_DATA_DIR	—	Override writable root for jobs, queue.sqlite, and tenants.sqlite (default: ~/.immersive-studio/worker when STUDIO_REPO_ROOT is unset).
STUDIO_COMFY_URL	https://comfy.immersivelabs.space (when unset)	ComfyUI HTTP API base URL; use http://127.0.0.1:8188 when ComfyUI is on the same machine
STUDIO_COMFY_IMAGE_WAIT_S	420	Wall-clock seconds (60–3600) to wait for each ComfyUI txt2img completion (history polling). Raise on very slow GPUs or heavy checkpoints.
STUDIO_COMFY_PROFILE	sd15	sd15 or sdxl (selects workflow + latent size)
STUDIO_COMFY_CHECKPOINT	profile-specific	Checkpoint file name as shown in ComfyUI
STUDIO_TEXTURE_MAX_IMAGES	32	Cap variant×slot ComfyUI renders per job (albedo/normal/orm)
STUDIO_JOBS_MAX_COUNT	200	Index cap; oldest job folders deleted when trimming
STUDIO_JOBS_MAX_TOTAL_BYTES	5368709120	Target max bytes under output/jobs/ before pruning
STUDIO_QUOTAS_DISABLED	—	Set to 1 to skip quota enforcement
STUDIO_PROMPT_BLOCKLIST	—	Comma-separated substrings to block in prompts
STUDIO_MODERATION_DISABLED	—	Set to 1 to skip moderation
STUDIO_API_AUTH_REQUIRED	off	Set to 1 to require API keys for studio routes (indie SaaS mode)
STUDIO_CORS_ORIGINS	Vite dev origins	Comma-separated browser origins, or * for any (use carefully in production)
STRIPE_SECRET_KEY	—	Secret API key (Dashboard → Developers) — required for Checkout / Portal
STRIPE_WEBHOOK_SECRET	—	Webhook signing secret for POST /api/studio/billing/webhook
STRIPE_API_VERSION	account default	Optional; pin Stripe API version string if you need stability
STUDIO_STRIPE_PRICE_INDIE	—	Stripe Price id for Indie recurring plan
STUDIO_STRIPE_PRICE_TEAM	—	Stripe Price id for Small team recurring plan
STUDIO_STRIPE_PRICE_MAP	—	Optional price_abc:indie,price_def:team overrides / extra prices
STUDIO_BILLING_SUCCESS_URL	/studio?billing=success	Checkout success redirect (optional: add {CHECKOUT_SESSION_ID} per Stripe docs)
STUDIO_BILLING_CANCEL_URL	/studio?billing=cancel	Checkout cancel redirect
STUDIO_BILLING_PORTAL_RETURN_URL	/studio	Customer Portal return URL
STUDIO_BLENDER_BIN	—	Path to blender executable if not on PATH
STUDIO_BLENDER_TIMEOUT_S	180	Max seconds for headless Blender mesh export
STUDIO_EXPORT_MESH_DEFAULT	off	Set to 1 / true to run mesh export on every run-job when the client does not pass export_mesh
DATABASE_URL	—	Postgres connection string (e.g. Vercel Postgres, Neon). Required when using Postgres backends below.
STUDIO_QUEUE_BACKEND	sqlite	postgres (shared rows in Postgres), redis (Redis broker + hashes), or sqs (SQS long-poll + Postgres rows — needs DATABASE_URL).
STUDIO_REDIS_URL / REDIS_URL	—	Broker URL when STUDIO_QUEUE_BACKEND=redis.
STUDIO_REDIS_QUEUE_ENGINE	zset	With Redis queue: zset (sorted-set dispatch) or streams (XADD / XREADGROUP / XACK).
STUDIO_REDIS_STREAMS_PREFIX	{STUDIO_REDIS_QUEUE_PREFIX}:sx	Key namespace for Streams mode (isolated from zset keys).
STUDIO_REDIS_STREAM_GROUP	studio_workers	Redis consumer group name for Streams mode.
STUDIO_REDIS_STREAM_BLOCK_MS	5000	XREADGROUP block timeout (ms).
STUDIO_REDIS_STREAM_MAXLEN	50000	Approximate MAXLEN on the dispatch stream (~).
STUDIO_REDIS_QUEUE_PREFIX	studio:q	Key namespace for the Redis queue.
STUDIO_REDIS_QUEUE_MAX_TIMELINE	5000	Cap on timeline index entries (oldest trimmed).
STUDIO_SQS_QUEUE_URL	—	Standard or FIFO SQS queue URL when STUDIO_QUEUE_BACKEND=sqs (Postgres holds job metadata).
STUDIO_SQS_WAIT_SECONDS	20	SQS ReceiveMessage long-poll wait (max 20).
STUDIO_SQS_VISIBILITY_TIMEOUT	900	Visibility timeout for in-flight SQS messages (seconds).
STUDIO_SQS_ENDPOINT_URL	—	Optional (e.g. LocalStack).
AWS_REGION / AWS_DEFAULT_REGION	us-east-1	SQS client region when using sqs backend.
STUDIO_TENANTS_BACKEND	sqlite	Set postgres with DATABASE_URL for shared tenants / API keys / usage.
STUDIO_JOB_ARTIFACTS	local	s3 (S3-compatible, incl. R2 via endpoint), or vercel_blob to upload pack.zip after each job.
STUDIO_S3_BUCKET	—	Bucket name when STUDIO_JOB_ARTIFACTS=s3 (or r2).
STUDIO_S3_REGION	auto	AWS region; use auto for Cloudflare R2.
STUDIO_S3_ENDPOINT_URL	—	Custom endpoint URL (set for R2 / MinIO).
STUDIO_S3_KEY_PREFIX	immersive-studio/packs	Prefix for object keys ({prefix}/{folder}/pack.zip).
AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY	—	Standard boto3 credentials for S3/R2 (or instance role).
BLOB_READ_WRITE_TOKEN	—	Vercel Blob store read-write token when STUDIO_JOB_ARTIFACTS=vercel_blob.

Optional pip extras: pip install -e ".[postgres]" (psycopg), pip install -e ".[redis]" (redis client), pip install -e ".[s3]" (boto3), or pip install -e ".[scale]" (Postgres + Redis + S3 deps together).

Stripe Billing (subscriptions → tiers)

Operations (monitoring, trial emails, optional notify URL): docs/studio/stripe-monitoring-and-emails.md.

Python dependency: stripe (declared in pyproject.toml). Flow:

1. Stripe Dashboard: create Products with recurring Prices (e.g. monthly). Copy each Price id (price_…) into STUDIO_STRIPE_PRICE_INDIE / STUDIO_STRIPE_PRICE_TEAM, or set metadata.tier on the Price to indie / team / free (overrides id mapping).
2. Enable Customer portal (Billing → Customer portal) so POST /api/studio/billing/portal-session works.
3. Webhooks: add endpoint https://<your-host>/api/studio/billing/webhook and subscribe at minimum to:
   • checkout.session.completed
   • customer.subscription.created
   • customer.subscription.updated
   • customer.subscription.deleted
     Paste the webhook signing secret into STRIPE_WEBHOOK_SECRET.
4. Runtime env: STRIPE_SECRET_KEY, STUDIO_API_AUTH_REQUIRED=1, tenant + API key as today.
5. Indie group: tenants create → tenants issue-key → open /studio, paste key → Subscribe (Checkout). After payment, webhooks link stripe_customer_id, set stripe_subscription_id, and call set_tenant_tier from the subscription’s price/metadata.

Tier mapping rules: For each subscription item’s Price, the worker uses (1) metadata.tier if valid, else (2) STUDIO_STRIPE_PRICE_MAP / env price ids, else (3) stays conservative (free if unknown). Canceled / unpaid / incomplete_expired subscriptions downgrade the tenant to free. past_due keeps the paid tier until Stripe resolves or cancels (adjust in stripe_billing.py if you prefer stricter cutoffs).

Local testing: Stripe CLI stripe listen --forward-to localhost:8787/api/studio/billing/webhook and use the printed whsec_… as STRIPE_WEBHOOK_SECRET.

Manual override: immersive-studio tenants set-tier still works for support / migrations; the next subscription webhook may overwrite tier until billing state matches.

Indie / multi-tenant tiers (SaaS mode)

When STUDIO_API_AUTH_REQUIRED=1, each workspace (small studio / customer group) is a tenant with:

• API keys (hashed in SQLite — only shown once at tenants issue-key)
• Monthly credits (UTC calendar month), consumed by:
  • generate-spec: 1 credit
  • jobs/run or queue enqueue (sync cost up front): 2 credits, or 6 with textures
• Concurrent job cap (SQLite counter per tenant while run_studio_job is in flight)
• Texture jobs blocked on Free tier

Tiers are defined in src/studio_worker/tiers.py (free, indie, team). Stripe webhooks update tier_id from the active subscription; tenants set-tier remains a manual admin override.

Per-tenant isolation: job index entries, queue rows, and ad-hoc pack paths are tagged with tenant_id so one API key cannot list or download another workspace’s jobs.

Default without STUDIO_API_AUTH_REQUIRED: single-machine dev — unlimited dev tier, no keys (same as before).

Blender placeholder mesh

Full pipeline reference (env, CI, Unity): docs/studio/essentials.md.

Exports procedural placeholder geometry as .glb from a spec (placeholder until full kitbash lands):

blender --background --python path/to/site-packages/studio_worker/blender/export_mesh.py -- --spec path/to/spec.json --output path/to/out.glb

Use pack/spec.json as the --spec argument when generating from a worker pack. Requires Blender 4.x on your PATH, or set STUDIO_BLENDER_BIN.

Integrated path: immersive-studio run-job … --export-mesh, POST /api/studio/jobs/run / queue with export_mesh: true, or the Export placeholder mesh checkbox on /studio. Successful runs place Models/<asset_id>/<asset_id>.glb and set manifest.toolchain.mesh_pipeline (blender:export_mesh.py+ok or +error). Failures are non-fatal: the job still completes with textures/spec; check mesh_logs in the API response or job metadata.

Profiles: studio_worker/blender/export_mesh.py picks a coarse shape from category: prop — stacked boxes with deterministic jitter; environment_piece — wide platform; character_base — two-block silhouette; material_library — small swatch cube; other categories fall back to a beveled cube scaled by target_height_m. PBR material names come from studio_worker/pbr_keys.py (same ordering as ComfyUI filenames and the Unity importer). The exported glTF material names are {variant_id}_{slot_id} per part so Unity can match them to {base}_Lit materials from textures.

Unity import

Add the UPM package under ../../packages/studio-unity to a Unity 2022.3+ URP project, then Immersive Labs → Import Studio Pack... and select the folder that contains manifest.json.

Layout

Path	Role
src/studio_worker/	CLI, API, validation, jobs, ComfyUI client, texture pipeline, Stripe billing
src/studio_worker/billing_config.py	Env mapping: Stripe Price id → entitlement tier
src/studio_worker/stripe_billing.py	Webhook handlers, Checkout / Portal session builders
src/studio_worker/billing_routes.py	FastAPI routes under /api/studio/billing/*
studio_worker/comfy_workflows/*.api.json	ComfyUI /prompt graphs (SD1.5 + SDXL albedo v1)
studio_worker/blender/export_mesh.py	Headless export stub
jobs/, queue.sqlite, tenants.sqlite, packs/…	Writable state (default ~/.immersive-studio/worker when STUDIO_REPO_ROOT is unset; otherwise apps/studio-worker/output/ — gitignored in the monorepo)

JSON Schema (runtime): studio_worker/data/studio-asset-spec-v0.1.schema.json (copy kept in sync with packages/studio-types/schema/ in the monorepo).

GPU / Comfy CI smoke test

Opt-in pytest marker gpu_comfy runs a minimal SD1.5 txt2img graph against live ComfyUI when:

• STUDIO_RUN_GPU_COMFY_TESTS=1
• STUDIO_COMFY_URL — base URL of ComfyUI
• STUDIO_COMFY_CHECKPOINT — checkpoint filename as shown in ComfyUI (SD1.5-style graph)

Optional: STUDIO_COMFY_STEPS (default 4), STUDIO_COMFY_TIMEOUT_S (default 420). Workflow: .github/workflows/studio-worker-gpu-comfy.yml.

GitHub Actions — I can’t set these for you (they live in your repo/org settings). In GitHub: Settings → Secrets and variables → Actions → New repository secret, create:

Secret	Example value	Purpose
STUDIO_RUN_GPU_COMFY_TESTS	1	Enables the live pytest (otherwise it skips)
STUDIO_COMFY_URL	http://127.0.0.1:8188	Base URL of ComfyUI as reachable from the workflow runner
STUDIO_COMFY_CHECKPOINT	your_model.safetensors	Exact checkpoint name in ComfyUI’s loader

workflow_dispatch can override the URL via the comfy_url input when non-empty. GitHub-hosted runners cannot see your LAN GPU; point runs-on at a self-hosted runner (or a URL the runner can reach) or run the test locally with the same env vars.