dsh-tongflow
TongFlow as a DeepSeek Harness (dsh) plugin — a media studio inside your agent.
Three layers, never mixed up:
| Layer | Owns |
|---|
| dsh | the harness: sessions, model routing, tools, jobs, web UI |
| the agent | creativity: the plan, the folder structure, briefs, scripts, prompts, review notes — plain files |
| TongFlow | deterministic generation: every image / voice / music / video / 3D asset is produced by running a saved workflow file (<name>.tongflow.json) through the TongFlow engine and its plugins |
There is deliberately no "generate an image" tool and no project template. The agent studies what the user wants to make (with web research when useful), proposes a folder structure, writes it down, and then — for every asset — creates a workflow file where that asset belongs, runs it, reviews the result, and builds the next stage on it. Users open the same .tongflow.json on the embedded canvas, tweak it, and re-run; they can also reorganize the folders by hand at any time.
Install
npx @deepseek-ai/dsh@next plugin --profile web add dsh-tongflow # from npm
# or from a tarball: pnpm --filter dsh-tongflow pack → dsh plugin --profile web add ./dsh-tongflow-x.y.z.tgz
npx @deepseek-ai/dsh@next web
Requirements: dsh ≥ 0.1.2-rc.1 — the 0.1.2-rc, 0.1.3-alpha and 0.1.5 lines are supported; 0.1.1-rc and older are not (the client Runtime package they build on was removed upstream, so upgrade dsh rather than pinning an old plugin). Because npm only matches a prerelease when a range clause carries its exact major.minor.patch, each new dsh prerelease line needs its own clause in peerDependencies — a host on a newer line reports a compatibility warning until then. Also Node ≥ 22.19, Python ≥ 3.10 on PATH (or pythonPath in the plugin config), git, and ffmpeg for video contact sheets. On first use the plugin creates ~/.dsh/tongflow/venv with the tongflow SDK and shallow-clones every official TongFlow plugin into ~/.dsh/tongflow/plugins (the live list from config/official-plugins.json; set autoInstallOfficial: false to install by hand), so the canvas offers the same node/plugin catalog as the hosted app. Keys and Modal deploys are only needed when something runs.
Start a session whose first message begins with @tongflow — that session becomes a studio session: the conversation view turns into the Studio (chat column · the project's folder tree · preview / editor / canvas · a runs drawer, all in the UI language of your browser), and the agent gets the tongflow_* tools and skill. Any other session is untouched dsh. In the Studio: create a project (a title and a brief — what you want to make), install TongFlow plugins and paste API keys under Plugins & keys, then talk to the agent — or click any file to preview / edit it, click a workflow to open it on the canvas.
Chat model
Any model dsh can route works. For the agent to see generated images (tongflow_look) use a vision-capable route, e.g. in $DSH_HOME/settings.yaml:
llm-pi-ai:
providers:
google:
apiKeyEnv: GEMINI_API_KEY
my-qwen: # a self-hosted Qwen3.8-27B behind vLLM
apiKeyEnv: QWEN_API_KEY
api: openai-completions
baseURL: http://127.0.0.1:8000/v1
models:
- id: Qwen/Qwen3.8-27B
input: [text, image]
Video and audio are reviewed through TongFlow's own describe / transcribe slots (tongflow_perceive), so they work with any chat model.
The project (a plain folder)
~/.dsh/tongflow/projects/<id>/
project.json title, brief, locale — the only fixed file
README.md the agent's plan: what the folders are, in what order things get made
…whatever the work needs… e.g. characters/, ep01/sh010/, music/, export/ — designed per project
The one rule: every AI-generated asset comes from a workflow file that sits next to its outputs.
characters/mei/
mei.md what the agent wrote about her
mei_ref.tongflow.json the workflow that renders her reference sheet
mei_ref.01.png run 1
mei_ref.02.png run 2 (a run never overwrites — fix the workflow, run again)
mei_ref.runs.json provenance of every run: inputs, plugins, note, timing
- Multi-output runs keep the workflow's output names:
mei_ref.03.image.png + mei_ref.03.caption.txt; text outputs are written as .txt too.
- Workflows reference project files by path:
./mei_ref.02.png / ../style/palette.png (relative to the workflow file) or characters/mei/mei_ref.02.png (relative to the project root); URLs pass through.
- Text files can be included in prompts:
{{../style.md}} {{./mei.md}} full-body sheet — expanded at run time, so a shared style note is written once.
- Compose:
tongflow_workflow_compose({ folder }) merges the small workflows of a folder (or an explicit list) into one <folder>_all.tongflow.json — a data node that references another part's output file (./ref.01.png) becomes an edge from that part's producing node, parts are ordered by those dependencies, every stage stays an output labelled after its part (shot_all.01.i2v.mp4 via meta.outputLabels), the parts are untouched.
- The Studio tree nests a workflow's outputs under it; the user may rename / move / delete anything by hand and upload files (header button → the selected folder, or drag & drop onto a folder view; default
uploads/) — the agent re-reads the tree (tongflow_project_status) before acting.
Billing checkpoint
A run that uses a paid plugin spends the user's money — a paid API key, or GPU seconds on their Modal account (a Modal plugin also deploys on first use). So tongflow_workflow_run without user_confirmed: true does not run: it returns needs_confirmation with the plugins involved, how each is billed (api / modal), whether its API keys are set, the models it offers and installed alternatives. The agent puts that to the user and calls again with user_confirmed only after an explicit yes — for every paid run; nothing is remembered. Runs that use only local plugins are free and start directly. The Studio's own Run drawer shows the same notice and a Confirm & run button.
Workflows follow TongFlow's grammar
tongflow_node_catalog opens with the node grammar — add/ widgets (canvas only), modality/ data nodes, and the four executable categories transfer/ (1 → 1), compose/ (N → 1), decompose/ (1 → N), batch/ (N → 1) — then lists every node type by category with its ABI slot, wires (batch / collect flags), config fields, outputs and installed plugins, all read from the ABI registry. The patch tool (apply_graph_patch from the tongflow package) validates each step against the same registry, so a workflow the agent saves is one the exporter and the canvas accept. The category table lives in src/engine/node-categories.ts and a test keeps it in step with packages/tongflow/src/canvas/node-types.tsx.
Agent tools
tongflow_project_create / _open / _list / _status · tongflow_workflow_new / _patch / _read / _list / _validate / _run / _compose · tongflow_node_catalog / _describe · tongflow_look (images / video contact sheets, returned as an image block — or described through a slot when the session's model takes no images) · tongflow_perceive (video/audio/image understanding via TongFlow slots; billing plugins need user_confirmed) · tongflow_plugins_list / _install / _uninstall · tongflow_run_status. Folder structure and text files are made with dsh's ordinary file tools. Long runs go through dsh background jobs (run_in_background).
Skill shipped: tongflow-studio (the working method: research → propose a structure → one workflow per asset next to its outputs → run → review → next stage), with four method references under skills/references/ that the agent loads only when the step needs them:
| Reference | Read before |
|---|
prompt-layers.md | writing any non-trivial prompt — the seven layers, and what belongs in the prompt text vs. node config vs. a wired reference file |
shot-contract.md | a video shot — open/close state, beat timeline, camera start-path-end, audio, continuity across shots |
failure-codes.md | a result came back wrong — locate the responsible layer, make the smallest fix |
iteration.md | running the same asset again — one variable at a time, and when to stop rewriting the prompt |
Genre knowledge is not packaged; the agent researches or the user installs a skill of their own.
HTTP (same origin as dsh)
/tongflow/projects, /tongflow/p/:pid/{tree,status,workflows,workflow[/summary|/outputs|/describe|/patch],runs,files/*}, /tongflow/runs/:id[/events|/cancel], /tongflow/plugins, /tongflow/env, /tongflow/health, plus the canvas-compat API under /tongflow/p/:pid/api/* that tongflow/canvas talks to.
Configuration (cordis row tongflow)
| key | default | |
|---|
studioRoot | <DSH_HOME>/tongflow | projects, venv, plugins, data |
pythonPath | auto-detect | Python ≥ 3.10 used to create the venv |
sdkSpec | tongflow==0.3.0 | pip spec installed into the venv (-e /path/to/sdk for development) |
pluginOrg | https://github.com/tong-io | where official plugins are cloned from |
pluginGitUrls | {} | plugin id → git URL overrides |
env | {} | environment for plugin processes (API keys); the Studio's key store (env.json) is merged over it |
maxConcurrentRuns | 2 | |
httpPrefix | /tongflow | |
locale | en | canvas UI locale (en / zh / ja / ko) |
autoInstallOfficial | true | at start, shallow-clone every official plugin that is missing (a few hundred KB each) so the canvas offers the full catalog; API keys / Modal deploys are only needed when a workflow runs |
Development
pnpm install
pnpm --filter dsh-tongflow build # host lib/index.js + browser lib/client.js
pnpm --filter dsh-tongflow test
npx @deepseek-ai/dsh@next plugin --profile web add ./packages/dsh-tongflow # link: install for hacking
The browser half is a single CJS bundle in dsh's window.__ModuleLoader__ shape: only dsh's platform modules (react, cordis, slot kits) stay external; tongflow/canvas, @xyflow/react, zustand and use-intl are inlined (and deduplicated so React contexts match). See docs/design.md.
License: AGPL-3.0-only (same as TongFlow).
