TongFlow desktop (cloud shell for https://app.tongflow.com)

TongFlow desktop (cloud shell for https://app.tongflow.com)

TongFlow desktop (cloud shell for https://app.tongflow.com)

Added

• CometAPI router plugin —
  tongflow-router-cometapi:
  one key for 500+ models behind CometAPI's OpenAI-compatible routes, covering
  19 slots — gen_text / split / combine, image / video / audio
  understanding, image gen/edit/fusion (GPT Image 2, Seedream), text / image(s)
  → video (Sora 2, Veo 3.1, Seedance, Wan, MiniMax H3, HappyHorse, Vidu), video
  edit (Omni), TTS and Whisper transcription.
• ToAPIs router plugin —
  tongflow-router-toapis:
  one key behind ToAPIs' OpenAI-compatible chat route and unified async image /
  video routes, covering 15 slots — gen_text / split / combine, image
  understanding, image gen/edit/fusion (GPT Image 2, Seedream 5, Gemini Image,
  Flux 2, Grok), text / image(s) / first-last-frame / multimodal refs → video
  (Sora 2, Veo 3.1, Seedance 2, Kling, MiniMax H3, Wan, HappyHorse, Vidu),
  video edit (HappyHorse). Inputs are hosted through ToAPIs' upload routes; the
  model picker follows the key's live model list.
• Authenticated model catalogs — TONGFLOW_MODEL_CATALOG may set
  authEnv (the env key holding a bearer token); the canvas then loads the
  catalog through GET /api/plugins/model-catalog?pluginId=…, which injects
  the key from Settings server-side, instead of fetching the URL in the
  browser. Slot rules also accept token lists and !-negated tokens. Scanner
  version 7, SDK 0.3.2.
• Live model catalogs for the per-node model picker — a plugin may declare
  TONGFLOW_MODEL_CATALOG (a public, CORS-enabled catalog URL plus per-slot
  filter rules); the canvas fetches it in the browser (10-minute cache,
  re-checked when the dropdown opens) and appends matching ids after the static
  TONGFLOW_SLOT_MODELS shortlist. Scanner version 6, SDK 0.3.1
  (modelCatalog on the plugin registry entry), filterModelCatalog in
  tongflow core.

Removed

• Agnes AI plugin (tongflow-api-agnes) — unregistered; the upstream
  gateway is superseded by CometAPI.

Added

• dsh-tongflow (packages/dsh-tongflow) — TongFlow as a DeepSeek Harness
  plugin: a film-crew studio (~/.dsh/tongflow/projects/<id>: bible entities
  CHR_/LOC_/PRP_/STY_, shot breakdown EP01_SC003_SH0010, numbered takes with
  provenance, tf:// references and {{tf://…}} prompt templates), tongflow_*
  agent tools where all media generation runs saved *.tongflow.json workflows
  through the Python engine, packaged skills (tongflow-studio,
  tongflow-manga-drama + workflow templates), and a Studio UI in dsh's web
  shell embedding tongflow/canvas (canvas-compat API under
  /tongflow/p/:pid/api/*). Published as npm dsh-tongflow (dsh-npm-v* tags).
• tongflow 0.2.0 — canvas-only dependencies moved to optional peer
  dependencies (the core entry now depends only on zod / json-schema-to-ts);
  the exporter honours data.inputName on level-0 data / add nodes so workflow
  inputs get readable names.
• Krea 2 Turbo text-to-image plugin (#143) —
  tongflow-modal-krea2
  (open-weights 12B, 8-step, up to 2K) joins the official GPU plugins.

Removed

• LTX and FastWan video plugins (#138) — tongflow-modal-ltx and
  tongflow-modal-fastwan are no longer official plugins. The preloaded
  example workflow's image-to-video step now runs on tongflow-modal-minimax-h3.
• In-app workspace agent — the chat panel that built workflows on the
  canvas (added in 0.3.0, #130) is gone, along with its /api/agent/chat
  proxy, generated docs corpus and the openai dependency. TongFlow is a
  pure workflow product again; agent-driven building moves to external hosts
  that consume the tongflow npm package + Python SDK engine (see dsh-tongflow).
  docs/agent-workflow-manual.md stays as the reference for such hosts.
• Skill packages (tongflow-package-*, added in 0.3.2, #140) — the Gen
  Text skill picker, the skills registry / /api/skills/registry route and the
  content-package install path are removed; the plugin scanner no longer
  special-cases the prefix (scanner v5). Prompt packs belong to the agent host,
  not to the workflow.

Added

• Manual edge connections (#145) — drag a connection between any two node
  handles; the ABI contract is validated live during the drag (modality match,
  single-value handle occupancy, add-node fan-out), so only valid edges can
  land. Nodes display newly connected upstream data immediately, and edges can
  be dragged from either end.
• Skill packages (#140) — tongflow-package-* content packages ship
  reusable prompt packs (skills) instead of executable code. The Gen Text node
  gains a skill picker; the selected skill's body is prepended to your prompt
  at execution time. The official
  tongflow-package-skills
  starter pack (12 skills) installs like any plugin.
• IndexTTS-2.5 emotive speech plugin
  (tongflow-modal-indextts2)
  — zero-shot voice cloning and emotion-controlled speech from a voice
  reference, adding the "Emotive speech" capability.
• Local plugin prefix (#141) — tongflow-local-* registers on-device
  engine plugins that run outside Modal, e.g. MiniMax-H3 through the native
  h3.c Metal binary on Apple silicon.
• Omni-reference in more selections (#137) — the multi-select smart island
  now offers refs-gen-video for multiple videos (≤3), multiple images (≤9),
  and text + single media combos.

Fixed

• Official router plugins failed to uninstall with a 400 — the installer
  now accepts the tongflow-router-* (and new package/local) prefixes (#140,
  #141).

Added

• Omni-reference video generation (#135) — new refs-gen-video node:
  mix up to 9 reference images, 3 video clips and 3 audio clips (12 files
  max) with a prompt to generate one video. Select any mix of image / video /
  audio nodes and the smart island offers it in one click. Address references
  in the prompt as <Picture 1> / <Video 1> / <Audio 1>.
• MiniMax-H3 official plugin
  (tongflow-modal-minimax-h3)
  — self-hosted 33B video generation with native stereo audio (dialogue,
  SFX and music in one pass; 24 fps, 768p, ~5–15 s) on Modal via headless
  ComfyUI, using the Comfy-Org optimized weights (~63 GB instead of 498 GB).
  Serves six video slots including the new omni-reference; measured 4 min 17 s
  per 5 s clip on B200 (≈ $0.45).
• Seedance omni-reference — the Doubao (ByteDance) plugin implements the
  new refs-gen-video slot via Seedance 2.0 multimodal references.
• Consistent node spacing + one-click auto-layout (#134) — spawn gaps are
  now uniform across agent builds, manual adds and run-result spawning, and a
  toolbar button tidies the whole canvas into layered columns.

Changed

• Python SDK 0.2.21 — generated models for the new refs-gen-video slot.
  All official Modal plugins repinned to 0.2.21.

Fixed

• Docker image build keeps the agent manual in context (#132).

Meet the workspace agent

TongFlow now has an agent. Open the panel on the right of the canvas, say what you want in plain language — attach images, video, audio or documents straight into the chat — and watch it build the workflow on the canvas, node by node, camera following along, prompts and parameters filling into the forms as it goes.

It works the way a collaborator would, not a code generator:

• It edits, it doesn't rebuild. "Make the video 10 seconds" changes one parameter and leaves your uploads and layout untouched.
• It knows what your install can actually do. The node catalog is generated from the running installation, so it only offers what you have plugins for — and says so when you don't.
• It can't hand you a broken graph. Every step is validated against the ABI before it touches the canvas; structurally invalid wiring is rejected and corrected.
• It can debug with you. Ask why a run failed and it reads the failure detail, explains it, and proposes the fix.
• It answers product questions too — how modes differ, how to install a plugin, how to connect Modal — from the bundled docs.
• One Cmd+Z undoes a whole turn.

Bring your own brain: the agent runs on any installed OpenAI-compatible text plugin — OpenRouter, OpenAI, Gemini, DeepSeek, xAI, Doubao, APIMart, Agnes — with a two-level plugin → model picker in the panel header.

App Mode

Any workflow can now be presented as a plain form app: inputs on top, results below, canvas hidden. Assets fed through add-nodes become replaceable inputs, so a workflow you built once doubles as a reusable tool — for you or for someone you share it with.

Undo/redo on the canvas

Cmd+Z / Cmd+Shift+Z, plus toolbar buttons. A burst of typing collapses into a single undo step rather than one per keystroke.

Settings that explain themselves

Settings are now human-readable cards with a guided first-run banner. Connecting a service is its own flow: Modal (paste both tokens at once — it splits them for you), Hugging Face, and every provider API key.

Failures you can act on

Task failures carry stable error codes with localized messages and support links, and preserve the upstream Modal/deployer error body instead of swallowing it. A missing API key now opens a guided dialog that deep-links to the exact settings field.

Also in this release

• Python SDK 0.2.20 — direct-stream terminal callback so browser-direct streaming runs report their terminal state reliably, plus per-call model passthrough. All official Modal plugins repinned.
• Fixed: reconnecting to a finished task replays its stored result instead of re-running the slot.
• Fixed: mount-time defaults in six more node types no longer pollute the undo stack.

Install: macOS TongFlow-mac-universal.dmg · Windows TongFlow-win-x64.msi · or use the cloud version at app.tongflow.com.

Full details in CHANGELOG.md.

TongFlow desktop (cloud shell for https://app.tongflow.com)

Fixed

• One-click execution now handles dynamic split counts end-to-end (#116) —
  a split's runtime item count drives downstream batch nodes: one plugin call
  per item (batchField fan-out in the engine), outputs collected in batch
  order. Previously only the first item was processed, silently.
• Intermediate data nodes are pure channels in one-click runs — consumers
  bind straight to the producer's output, so stale edit-time values of
  materialized nodes no longer leak into execution. Re-save workflows exported
  before this version to pick up the new bindings.
• Loud failure instead of silent data loss — a single-value input that
  receives multiple upstream values now fails the node with a clear message.

TongFlow desktop — a ~10 MB Pake (Tauri) cloud shell for https://app.tongflow.com.

Added

• DeepSeek plugin (tongflow-api-deepseek) — DeepSeek V4 as a text
  provider for gen-text / combine-text / split-text, with a per-node
  model dropdown offering deepseek-v4-flash / deepseek-v4-pro each
  with thinking on/off (four choices).
• Streaming "thinking" bubble — when a plugin streams its reasoning
  (DeepSeek thinking mode), it now appears live in an auto-scrolling bubble
  anchored beside the node while it runs, separate from the status label and
  non-occluding. Backed by progress(..., thinking=True) in the SDK.
• SenseNova-Vision plugin (tongflow-modal-sensenova-vision) — SenseTime's
  unified vision model: image understanding / visual QA, detection & OCR
  structured text, full-scene surface normals, salient-object matting, and a
  human-pose overlay (an alternative implementation of those slots).
• Open-vocabulary sound separation (separate-sound node) — describe any
  sound in words ("dog barking") and split the audio into that sound and
  everything else.

Changed

• Cloud plugin execution path (Python SDK tongflow 0.2.7 → 0.2.17) — a
  new single-container cloud execution surface so plugins can run in the
  user's own Modal without a per-call venv + subprocess:
  • serve_slot / run_and_report — single-slot execution and a background
    single-node runner that reports back over an HTTP callback;
  • serve_stream / serve_stream_from_spec — single-container streaming
    slot execution, including a browser-direct stream entry;
  • self-reported progress (and streamed reasoning) over an HTTP sink, so
    remote plugins drive the node status and thinking bubble;
  • an injectable workflow invoker that skips the venv + subprocess hop;
  • default-slot claims (TONGFLOW_DEFAULT_SLOTS, @node_slot(default=True))
    so a plugin can declare the default implementation for a slot while still
    importing under older, already-deployed runtimes.

Fixed

• SDK HttpStore now sends a User-Agent (works around a Cloudflare 403) and
  emits a stable event id in serve_stream (tongflow 0.2.13).

Installers: TongFlow-mac-universal.dmg (Apple Silicon + Intel) and TongFlow-win-x64.msi.

[0.2.1] - 2026-07-12

Added

• Meta Sapiens2 human suite (tongflow-modal-sapiens2) — five new ABI
  slots with canvas nodes and smart-island actions, backed by the Sapiens2-1B
  checkpoints on one Modal L40S (bf16, batched):
  • Pose detection (image-pose) — 308-keypoint whole-body skeleton
    (body, hands, face) on a clean black background;
  • Body-part segmentation (image-body-seg) — 29-class human parsing
    as a pure class-color map;
  • Surface normals (image-normal) — per-pixel normal map, background
    masked;
  • Human matting (image-matting) — straight-alpha transparent PNG;
  • additionally the existing Image → 3D slot gains a Sapiens2 pointmap
    implementation (colored human point cloud GLB).
• Video motion capture (video-gen-model) — monocular video → animated
  3D human GLB on Meta's MHR character (Momentum Human Rig, Apache-2.0),
  playable directly in the model node and importable into Blender as a
  skinned armature. Two competing implementations of the same node slot:
  • tongflow-modal-sam-3d-body — per-frame MHR regression (learned
    prior; body + hands, experimental face/jaw channels driving 72 sparse
    glTF morph targets);
  • tongflow-modal-sapiens2 — geometric engine (308-keypoint pose +
    pointmap 3D lift, One-Euro smoothing, rest-pose hold for out-of-frame
    body parts).
• Model node: animation playback — models carrying animation clips now
  auto-play (30 fps cap, pauses off-viewport and in background tabs) with a
  play/pause control in the node preview and the fullscreen viewer.

Changed

• Model node initial framing — the camera now homes straight onto the
  model's front (glTF +z convention) and maximizes the model in the frame
  (aspect-aware fit); fixed a centering bug that framed tall/offset models
  poorly and mixer leaks that kept animating after unmount.
• Python SDK tongflow 0.2.6 — generated models and NodeSlots for the
  five new ABI slots.