Video Flow — SenClaw Space App

Flow Kit ported as a SenClaw Space App. An AI video production factory: a multi-agent DAG turns a raw concept or a finished screenplay into a complete video — driving Google Flow (Veo3 video / Imagen images) through a Chrome extension bridge. Rust + axum backend, React dashboard, all LLM calls routed through the SenClaw daemon's shared LLM bridge (no provider keys of its own).

Projects → videos → scenes → characters, with reference images, per-scene frame images, per-scene Veo3 clips (vertical + horizontal state fully independent), 4K upscales and ffmpeg concat. Every sub-agent (director, screenwriter, shot design, image/video gen, critic, …) has its own editable prompt in souls/.

Architecture

┌────────────────────┐   WS ws://127.0.0.1:9222   ┌───────────────────────┐
│  Chrome            │◄──────────────────────────►│  video-flow app :4460 │
│  extension/        │                            │  (axum)               │
│  (Video Flow bridge)│─ HTTP /api/ext/callback ─►│   ├─ DAG engine       │
│  labs.google tab:  │                            │   ├─ request worker   │
│  token + reCAPTCHA │                            │   ├─ REST /api/*      │
└────────────────────┘                            │   ├─ MCP /api/mcp/sse │
        ▲ calls aisandbox-pa.googleapis.com       │   └─ React UI (/)     │
        │ (Veo3 / Imagen)                         └──────────┬────────────┘
                                                             │ LLM bridge
                                                             ▼
                                                  ┌───────────────────────┐
                                                  │ SenClaw daemon :18788 │
                                                  │ (shared LLM + agents) │
                                                  └───────────────────────┘

• The extension runs in the browser because Google Flow needs the user's ya29.* OAuth token, per-call reCAPTCHA Enterprise solving, and requests originating from labs.google.
• The app never talks to an LLM provider directly — src/llm.rs bridges to the SenClaw daemon (SENCLAW_BASE_URL, default http://127.0.0.1:18788).
• Two pollers over SQLite: the DAG engine (multi-agent pipeline, 500ms tick, max 5 concurrent) and the request worker (video/upscale queue via the extension).

Dev run

# backend (from the SemaClaw repo root; serves UI + API on :4460, ext WS :9222)
cargo run -p video-flow

# web UI dev server (hot reload)
cd apps/video-flow/web && npm install && npm run dev

The SenClaw daemon must be running (LLM bridge on :18788). Health check: curl http://127.0.0.1:4460/api/status.

Install as a Space App

• Dev / from source: register the app dir directly with the daemon (register-local) so it starts ./video-flow per senclaw-manifest.json.
• Packaged: apps/video-flow/scripts/pack.sh builds web + release binary and stages a flat release/ (binary, manifest, skills/, personas/, souls/, playbooks/, extension/, web_dist/), zipped as video-flow-app.zip — install that zip via SenClaw's install-zip flow. The manifest auto-registers the MCP server video-flow-mcp at /api/mcp/sse.

Chrome extension setup (required for any image/video generation)

1. chrome://extensions → Developer mode → Load unpacked → select the app's extension/ folder.
2. Open labs.google (Google Flow) and sign in — the extension captures the Bearer token and connects to the app WS on :9222.
3. Verify: the popup shows three green dots (app WS / app HTTP / Flow token), and vf_status (or GET /api/status) shows extension_connected: true.

If SenClaw assigns the app a port other than 4460, set it in the popup under Kết nối · cổng của app — the ports live in chrome.storage.local, so no rebuild is needed. See EXTENSION_SETUP.md.

Planning/parsing stages (pure LLM) work without the extension; image, video and upscale stages require it. Upscale additionally needs a Flow TIER_TWO account.

Environment variables

Chi tiết cài extension + đổi cổng: EXTENSION_SETUP.md.

MCP tools (video-flow-mcp, full names mcp__video-flow-mcp__vf_*)

Data location (survives reinstall)

Installing a Space App zip runs remove_dir_all(<app_dir>) before extracting, so anything stored next to the binary is destroyed on every update. The DB and downloaded media therefore live in **~/.senclaw/space-app-data/<app_id>/** (override with FLOWKIT_DATA_DIR). On first boot the app migrates a DB left in the old in-app-dir location — including the -wal/-shm sidecars and media/ — so upgrading from an older build keeps existing projects.

Local media & request lifecycle

Google Flow serves generated assets from short-lived signed URLs, so every successful generation is mirrored into media/ immediately and the DB stores /api/media/{id}/file. Downloads dedupe on original_url, and a download failure never loses the asset — the remote URL is kept and the generation still counts as done.

• Repair an older project whose thumbnails went blank: POST /api/media/localize {"project_id": "..."} or MCP vf_media_localize.
• A request row in PROCESSING cannot survive a restart, so at boot the app reconciles leftovers: if the asset it was producing exists the row becomes COMPLETED, otherwise FAILED with an explicit reason. Image requests are closed out by the image agent (video/upscale by the worker) — nothing stays spinning.

Souls editing

Each pipeline stage's behaviour is its soul — a markdown file in souls/ (e.g. director.md, image-gen.md, critic.md), frontmatter stripped at load. Edit via the UI, PUT /api/agents/{type}/soul, or the MCP pair vf_soul_get / vf_soul_set; changes apply on the agent's next execution. playbooks/ holds the internal prompt playbooks used by DB-backed skill agents (ReAct tool loop) — not to be confused with the SenClaw skills in skills/.

Credits

Port of Flow Agent Video ("Flow Kit") — the original Go backend, React dashboard and Chrome extension — onto the SenClaw Space App runtime (shared daemon LLM, MCP, skills and personas).