Files
artcraft/_docs/dev_setup_local_stack.md
Tanish Manku f5b80ea982 Dev infra: local filesystem uploads + fully-local pagescene media
Closes the two remaining fully-local-stack seams found after the last round.

Uploads (previously every upload endpoint returned 500 with dummy S3 creds -
the bucket put happens before the DB insert and rust-s3 is fail-on-err):
- LegacyBucketClient gains an internal backend enum (S3 unchanged | LocalDir)
  and a create_local(root, bucket_name, root_prefix) constructor. All upload/
  download methods route LocalDir to the filesystem; bucket_name() reports the
  configured name. Objects map disk = root + rooted object path via the new
  pub local_object_disk_path (rejects ".." and backslash segments - PMX zips
  feed user-controlled names), unit-tested. The fake-generation resolver now
  uses the same function instead of its private copy so the layouts can't
  drift.
- build_dependencies: when SERVER_ENVIRONMENT=Development and
  LOCAL_MEDIA_ROOT is set, all three bucket clients are constructed in local
  mode (private/gc have zero callers in storyteller-web, so this is purely
  uniform); the server_environment read moved above client construction.
  Uploads land under .devstack/media and are immediately viewable through
  the existing /media mount + CDN_BASE_URL - no S3/R2/MinIO.

Frontend (pagescene was the only surface actually broken against a local
backend; moodboard/library/create-* already consume media_links correctly):
- The vite /__cdn dev proxy assumed https upstreams, so localhost media
  (http) died on a TLS handshake with a 502 - breaking every pagescene
  GLB/splat/animation fetch. fetchAsset no longer proxies localhost at all
  (the dev backend sends CORS for any localhost origin), and the proxy
  itself now picks http for localhost hosts as a backstop.
- GetCdnOrigin() returns the API host when it is localhost (the local
  backend serves media itself; CDN origin == API origin), production CDN
  otherwise. The pagescene adapter's getCdnUrl skips the Cloudflare
  /cdn-cgi/image resize segment on localhost - the local backend has no
  resizer, and dev serves originals anyway.
- vite /v1 fallback proxy now targets http://localhost:12345 instead of
  production, so nothing can silently fall back to prod traffic in dev;
  fixed the stale main.tsx comment claiming setDevelopment() is disabled.

Docs: seams table updated (uploads + vite proxy + GetCdnOrigin closed;
API_TARGETS/SegmentationApi noted as dead code; the earlier moodboard claim
was wrong - it never had a hardcoded host), LOCAL_MEDIA_ROOT now documents
the storage mode, verification checklist gains the upload check.

Verified E2E on this machine: POST /v1/media_files/upload/image (multipart,
real endpoint) returns success, the object lands under .devstack/media at
the canonical fan-out path, lists in the library with media_class=image,
and serves 200 image/jpeg from http://localhost:12345/media/...;
cloud_storage unit tests pass; storyteller-web + webapp dev server boot
clean with "Local filesystem storage mode is ON".
2026-07-29 21:10:42 -07:00

295 lines
19 KiB
Markdown

# Local Dev Stack — Requirements & Bootstrap
How to run the full ArtCraft dev stack **entirely locally** — the
`storyteller-web` backend plus the `artcraft-webapp` frontend — with no traffic
to production servers. Scope: webapp-first; the Tauri desktop app is a later
phase.
The automated path on **Windows** (primary platform, native — no WSL, no
Docker, no admin rights for the stack itself):
```powershell
.\script\bootstrap\windows\bootstrap_dev_stack.ps1 # one-time setup, idempotent, re-run anytime
.\script\bootstrap\windows\run_backend_dev.ps1 # terminal 1: API on http://localhost:12345
.\script\bootstrap\windows\seed_demo_user.ps1 # terminal 2: create demo login (localdev1 / localdev1pass)
cd frontend; npx nx dev artcraft-webapp # terminal 2: webapp on http://localhost:4201
.\script\bootstrap\windows\dev_stack_doctor.ps1 # health check, any time
.\script\bootstrap\windows\stop_dev_services.ps1 # stop MySQL/Redis when done
```
The equivalent on **Linux** (Ubuntu 22.04+ — native, WSL2, or CI):
```bash
./script/bootstrap/bootstrap_dev_stack.sh # one-time setup, idempotent, re-run anytime
./script/bootstrap/run_backend_dev.sh # terminal 1: API on http://localhost:12345
./script/bootstrap/seed_demo_user.sh # terminal 2: create demo login (localdev1 / localdev1pass)
cd frontend && npx nx dev artcraft-webapp # terminal 2: webapp on http://localhost:4201
./script/bootstrap/dev_stack_doctor.sh # health check, any time
```
This document records the underlying requirements so the scripts aren't a
black box, and corrects some stale claims in older docs.
## Supported platforms
| Platform | Status |
|---------------------|---------------------------------------------------------------------------|
| Windows 10/11 | Primary. `script/bootstrap/windows/*.ps1` run a **portable stack**: MySQL 8.4 (official zip) and Redis live under the gitignored `.devstack/` and run as plain user processes — no admin, no services, no WSL, no Docker. Build tools (cmake, perl, nasm, llvm) are offered via winget |
| Ubuntu 22.04+ | Supported by `script/bootstrap/*.sh` (native, WSL2, or CI runner/container); services via apt |
| macOS | Not covered by the scripts; follow `_docs/dev_setup_server.md` manually |
### Windows specifics
- Everything stateful lives in `.devstack/` (gitignored): `mysql/` (extracted
zip), `mysql-data/`, `redis/`, `logs/`, `pids/`, `downloads/` (cached zips),
`my.ini`. Delete the directory to fully reset; re-run bootstrap to rebuild.
- Pinned services: MySQL **8.4.11** from `cdn.mysql.com` (the 8.x series is
required; only some patch versions exist on the CDN) and Redis **5.0.14.1**
(tporadowski build — old but sufficient for the backend's keepalive/TTL/
rate-limit usage; override `REDIS_ZIP_URL` to use Memurai/Garnet instead).
- The portable MySQL root user has **no password** (`--initialize-insecure`) —
local dev only, bound to 127.0.0.1.
- diesel-cli is compiled against the portable MySQL's own client library
(`MYSQLCLIENT_LIB_DIR`/`MYSQLCLIENT_VERSION`); `libmysql.dll` must be on
`PATH` to *run* it — the scripts handle both.
- Compiling the backend needs cmake, Strawberry Perl, NASM, and LLVM/libclang
(the `wreq` HTTP client builds BoringSSL, and `aws-lc-sys`/`ring` need
NASM/clang on MSVC). The bootstrap checks and offers winget installs.
- The generated secrets file additionally sets `TEMP_DIR` (the upload temp-dir
helper defaults to `/tmp`, which doesn't exist on Windows).
## Requirements matrix
| Requirement | Version / detail | Needed for |
|-------------------|--------------------------------------------------|---------------------------------------------------|
| MySQL server | 8.x series (9.x unsupported by tooling) | Backend boot (eager connect) + migrations |
| Redis server | any recent | Backend boot (r2d2 pool connects eagerly) |
| Rust (stable) | no pinned MSRV; Docker builds use 1.93.0 | Building `storyteller-web`, diesel_cli |
| diesel_cli | `cargo install diesel_cli --no-default-features --features mysql,sqlite` | Running the 281 migrations |
| Node.js | 20+ (enforced by `script/common/frontend_preflight.sh`) | Frontend (Nx 21 / Vite 6) |
| npm | ships with Node; **pnpm must not be used** | Frontend install (`frontend/package-lock.json`) |
| System libs (apt) | build-essential, cmake, libssl-dev, libclang-dev, libfontconfig1-dev, fontconfig, perl, pkg-config, ffmpeg, libmysqlclient-dev, libsqlite3-dev, curl, jq | Compiling the workspace + diesel_cli; seeding scripts |
| Elasticsearch | optional (7.x) | Only 4 search endpoints; server boots without it |
| Docker | **not required** — no docker-compose exists | — |
Database identity (used consistently by docs, `.env`, and defaults):
database `storyteller`, user `storyteller`, password `password`, host
`localhost`.
## Backend: what `storyteller-web` needs to boot
Binary: `storyteller-web` in `crates/service/web/storyteller_web`
(binaries: `storyteller-web`, `docs` — Swagger UI on :8989, `docs-cli` — emits
`api.json`). Binds `0.0.0.0:12345` (`BIND_ADDRESS`), plain HTTP.
**Run it from the repo root** — the config search path and `includes/` paths
are cwd-relative.
### Config loading (corrected facts)
- Mode is selected by the `SERVER_ENVIRONMENT` **process env var only**
(default: Development). Setting it inside an env file does nothing — it is
read before any file loads.
- In Development mode, files load from
`crates/service/web/storyteller_web/config/` in this order:
`storyteller-web.common.env` → `storyteller-web.development.env` →
`storyteller-web.development-secrets.env` (gitignored; the bootstrap
generates it).
- **Precedence is first-value-wins** (dotenv never overrides an existing key):
real process env > common.env > development.env > development-secrets.env.
The secrets file therefore cannot override anything already set in
`development.env`.
- The repo-root `.env` is **NOT loaded by the server** (it passes
`ignore_legacy_dot_env_file: true`). Root `.env`'s `DATABASE_URL` feeds
diesel-cli/sqlx-cli only; the server reads **`MYSQL_URL`** (default
`mysql://storyteller:password@localhost/storyteller`).
### Hard requirements at boot
| Requirement | Behavior if missing |
|----------------------------------|--------------------------------------------------------|
| MySQL reachable via `MYSQL_URL` | Boot fails (sqlx pool connects eagerly; the DB must exist and be migrated — health-check and ban-poller threads query immediately) |
| Redis reachable via `REDIS_0_URL` (default `redis://localhost/0`) | Boot fails (r2d2 pool connects eagerly) |
| ~30 `get_env_string_required` vars present | Boot fails on the first absent var |
Of the required-present vars, 19 are supplied by **no** checked-in file and
live in the generated secrets file: `ACCESS_KEY`, `SECRET_KEY`, `REGION_NAME`,
`W2L_PRIVATE_DOWNLOAD_BUCKET_NAME`, `W2L_PUBLIC_DOWNLOAD_BUCKET_NAME`,
`RESEND_API_KEY`, `FAL_API_KEY`, `GMICLOUD_API_KEY`, `GROK_API_KEY`,
`BEEBLE_API_KEY`, `OPENAI_API_KEY`, `WORLDLABS_API_KEY`,
`SEEDANCE2PRO_COOKIES`, `SEEDANCE2PRO_WHITELIST_COOKIES`,
`SEEDANCE2PRO_BYTEPLUS_ULTRA_COOKIES`, `STRIPE_FAKEYOU_ACCOUNT_ID`,
`STRIPE_ARTCRAFT_ACCOUNT_ID`, `STRIPE_ARTCRAFT_SECRET_KEY`,
`STRIPE_ARTCRAFT_SECRET_WEBHOOK_KEY`.
**Placeholder values are safe at boot**: none of these services is contacted
at boot (clients are constructed offline; Elasticsearch, S3/R2, Stripe,
Resend, and Google-cert fetches are all lazy). But two earlier claims here
were wrong and are corrected:
- **The omni_gen endpoints call providers synchronously in the HTTP handler**
(`/v1/omni_gen/generate/{image,video}` — the path the webapp uses). With
dummy keys the provider call fails and the submit returns HTTP 500 *after*
billing the wallet (the image pipeline has no refund; video refunds only
Kinovi). Only the legacy polling providers (GmiCloud, Grok, Kinovi orders,
WorldLabs) are worker-driven. `DEV_FAKE_GENERATION` (below) exists so local
submits never reach a real provider.
- **Uploads would fail hard with dummy S3 creds** (every media upload
endpoint uploads to the bucket *before* inserting the DB row, and `rust-s3`
is built with `fail-on-err`) — which is why `LOCAL_MEDIA_ROOT` (below) also
switches the bucket clients into a local filesystem mode: uploads write
under `.devstack/media/` and are immediately viewable via the same `/media`
mount. No S3/R2/MinIO needed.
### Fully-local generation and media (`DEV_FAKE_GENERATION`)
The generated secrets file enables three development-only settings (appended
automatically on bootstrap re-runs for older stacks):
| Env var | Effect |
|-----------------------|----------------------------------------------------------------------------|
| `DEV_FAKE_GENERATION` | omni_gen image/video submits skip the external provider and insert a normal **pending** job with a `fake_…` external id. A resolver thread inside `storyteller-web` completes each one after `DEV_FAKE_GENERATION_RESOLVE_SECS` (default 6s, poll every `DEV_FAKE_GENERATION_POLL_SECS`, default 2s) with a placeholder result copied from `test_data/`. Prompts containing `simulate_artcraft_failure` (or `test_artcraft_failure`) fail instead, optionally with a `FrontendFailureCategory` word in the prompt (e.g. `simulate_artcraft_failure face_not_detected`). Hard-disabled outside `SERVER_ENVIRONMENT=Development` |
| `CDN_BASE_URL` | Overrides the compiled-in CDN hosts in `cdn_link.rs` for all media URLs in API responses. Must be scheme+host only (no path) — set to `http://localhost:12345` so media resolves against the backend itself |
| `LOCAL_MEDIA_ROOT` | Directory the backend serves at `/media` (dev-only actix-files mount), mirroring the public bucket layout: object path `/media/{…}` lives at `{LOCAL_MEDIA_ROOT}/media/{…}`. Also switches all three `LegacyBucketClient`s (public/private/gc) into a **local filesystem mode**, so every media upload endpoint writes here instead of S3/R2 — uploads work with dummy cloud creds. The bootstrap points it at `.devstack/media/` |
With all three set, the full loop works offline: submit → pending job →
polling (`/v1/jobs/session`) → success with a `media_files` row whose URL the
backend itself serves — or a failed job with a real failure category. Credits
are genuinely deducted from the seeded wallet (no refund on fake results —
the demo wallet is large enough not to care).
### Auth on localhost
- `POST /v1/login` takes `{"username_or_email": "...", "password": "..."}`;
success returns `{"success": true, "signed_session": "<jwt>"}` **and** a
`session` cookie.
- With the checked-in dev config (`COOKIE_DOMAIN=.jungle.horse`), the session
cookie is issued non-Secure and host-only (the Domain attribute is never
set), so plain-HTTP localhost login works as-is. The frontend additionally
stores `signed_session` in localStorage and sends it as a `session` header.
- Dev-mode CORS allows any `localhost`/`127.0.0.1` origin on any port, with
credentials.
- Health endpoint for scripts/CI: `GET /_status`.
### Build
- Compile with `SQLX_OFFLINE=true` (uses the checked-in `.sqlx/` cache; 483
queries). Without it, the sqlx macros open a **live DB connection at compile
time** using `DATABASE_URL` from the root `.env`. `SQLX_OFFLINE` has no
runtime effect — ignore older notes saying not to set it when running.
- First build takes several minutes; the crate has ~370 pre-existing warnings.
## Database: migrations and seed data
- 281 Diesel migrations in `_database/sql/migrations/` (`diesel.toml` at repo
root points there). Run from the repo root: `diesel migration run`.
`"Encountered unknown type for Mysql: enum"` warnings are harmless.
`migrations_squashed/` is documentation only — not executable.
- Seed data: `_database/sql/seed/sql/system_roles.sql` (roles `user`, `mod`,
`admin`) and `user_badges.sql`. **The `user` role is mandatory** — account
creation hardcodes `user_role_slug='user'`. The legacy wrapper
`bootstrap_inserts_roles_etc.sh` must run with cwd `_database/sql/` and is
not idempotent; the bootstrap script applies the same SQL files with
idempotency guards instead.
- Demo user: created through **`POST /v1/create_account`** (no captcha, no
invite code, no email-verification gate; password just needs 6+ chars). This
exercises the real signup path instead of hand-rolled SQL. Gotcha: usernames
like `demo`, `admin`, `test`, `dev`, `artcraft` are reserved
(`includes/binary_includes/usernames/reserved_usernames.txt`) — the default
demo user is `localdev1`. The Rust seeder
(`cargo run --bin dev-database-seed`) exists but hard-requires a
`.env-secrets` file and also seeds model weights, so the HTTP path is
preferred.
- Demo credits: signup creates **no wallet**, and the webapp gates the
generate button on wallet balance — so `seed_demo_user` also seeds an
`artcraft`-namespace wallet with `DEMO_CREDITS` (default 100,000) banked
credits, plus matching ledger rows. Idempotent: re-runs top the balance
back up to `DEMO_CREDITS` but never lower it. (Raw SQL, because the only
real credit-granting paths are Stripe webhooks and moderator tooling.)
- Demo gallery: `seed_demo_user` also seeds a handful of `media_files` image
rows (from `test_data/image/`) with their objects copied under
`.devstack/media/`, so the library isn't empty on a fresh stack. Rows mirror
what a real image upload writes (`media_class='image'` matters — the
library filters on it, and the column default `'unknown'` is invisible in
every tab).
## Frontend: `artcraft-webapp`
- Nx workspace at repo-root `frontend/` (npm workspaces; `npm install` from
`frontend/`). The FontAwesome Pro registry token is committed in
`frontend/.npmrc` — no manual auth step.
- Serve: `npx nx dev artcraft-webapp` → `http://localhost:4201`. No lib
pre-build is needed in dev (libs resolve to TS source via the
`"development"` export condition); production `nx build` does need lib
builds.
- **In dev the webapp already targets the local backend**: `src/main.tsx`
calls `StorytellerApiHostStore.setDevelopment()` under
`import.meta.env.DEV`, pointing all API traffic at
`http://localhost:12345`. (The adjacent comment claiming this is disabled is
stale — the call is live, and it bypasses the Vite `/v1` →
`api.storyteller.ai` proxy.) So: local backend running → fully local stack;
local backend **not** running → dev webapp is broken, not silently on prod.
- Optional: `VITE_GOOGLE_CLIENT_ID` in `frontend/apps/artcraft-webapp/.env`
enables the Google SSO button; email/password login needs nothing.
- Other dev ports if you run more apps: website 4200, dashboard 4202 (separate
pnpm app, excluded from Nx), video-info-website 4250, genhub 4300, Tauri
frontend 5173.
## What still touches production (known, deliberate seams)
Cleaning these up is the next phase of the infra work; the bootstrap does not
change runtime behavior:
| Seam | Today |
|-----------------------------------------------|--------------------------------------------------------------|
| Media/CDN URLs in API responses | **Closed** when `CDN_BASE_URL` + `LOCAL_MEDIA_ROOT` are set (the bootstrap default): media URLs point at the backend's own `/media` mount. Unset, the compiled-in `cdn_link.rs` constants still point at the public `pub-….r2.dev` bucket |
| Media uploads | **Closed** when `LOCAL_MEDIA_ROOT` is set: the bucket clients run in local filesystem mode and uploads land under `.devstack/media/`. Unset (or non-Development), uploads still 500 with dummy creds |
| Vite `/v1` proxy in webapp | **Closed**: now targets `http://localhost:12345` (it was already bypassed in dev by `setDevelopment()`, but the fallback could silently hit production). The `/__cdn` dev proxy also no longer intercepts localhost media (it assumed https and broke pagescene GLB/splat loads against a local backend) |
| `GetCdnOrigin` (pagescene asset thumbnails) | **Closed** for local dev: returns the API host when it's localhost (the local backend serves media itself), production CDN otherwise; the pagescene adapter also skips the Cloudflare `/cdn-cgi/image` resize path on localhost. `enums/Api.ts` `API_TARGETS` and `SegmentationApi` still hardcode production hosts but are dead code (not exported/reachable from the webapp); moodboard never had a hardcoded host — it consumes `media_links` directly |
| Elasticsearch-backed search | `search_featured` / `search_session` media files, `weights/search`, legacy `tts/search` fail without a local ES + `elasticsearch-cli` reindex (optional) |
## CI usage
The same bootstrap works on a fresh Ubuntu runner/container:
```bash
export MYSQL_ROOT_PASSWORD=root # GitHub Actions' preinstalled MySQL uses password auth
./script/bootstrap/bootstrap_dev_stack.sh --yes
./script/bootstrap/run_backend_dev.sh &
./script/bootstrap/seed_demo_user.sh # waits for /_status
./script/bootstrap/dev_stack_doctor.sh
```
`--yes` answers all prompts; `--skip-frontend` / `--skip-rust-build` /
`--skip-packages` slice the work for focused jobs.
## Verification checklist
- `./script/bootstrap/dev_stack_doctor.sh` exits 0 with no FAILs.
- `POST /v1/login` with the demo user returns `success: true` plus a
`signed_session` (done automatically by `seed_demo_user.sh`).
- Webapp at `http://localhost:4201` can log in as `localdev1`, has credits in
the top bar, and sees the seeded images in the library.
- A generation submit returns `{"success": true, "inference_job_token": …}`,
shows as pending for a few seconds, then completes with a placeholder
result (`DEV_FAKE_GENERATION`) — no external provider is contacted. A
prompt containing `simulate_artcraft_failure` fails with a real error
toast/category instead.
- Uploading an image (e.g. a reference image in the prompt box) succeeds and
the file lands under `.devstack/media/` — no S3/R2 credentials involved.
- Nothing hits `api.storyteller.ai` or the r2.dev CDN during the above
(verify in the browser network tab — media URLs should point at
`http://localhost:12345/media/…`).
## Corrections to older docs
- `crates/service/web/storyteller_web/CLAUDE.md`: the repo-root `.env` is
*not* auto-loaded by the server, and `SQLX_OFFLINE=true` is fine (and
recommended) when running — it only affects compile-time query checking.
- `_docs/dev_setup_server.md`: `./script/dev_mysql_connect.sh` doesn't exist —
it's `./script/dev_database_connect.sh`.
- `diesel.toml`'s comment about needing an unreleased diesel CLI is stale —
any current `diesel_cli` works.
- `TESTING_INFRA_HANDOFF.md` (untracked): the server env var is `MYSQL_URL`,
not `DATABASE_URL`; migration count is 281; crate path is
`crates/service/web/storyteller_web`.