comfy-ts 0.3.0

Everything user-facing that landed since 0.2.0. Breaking changes are marked 💥.

The re-run contract: defineWorkflow + vars

• host.defineWorkflow({ id, vars, build }) declares the knobs once; build
  re-executes against the current values on every run(), producing a fresh
  graph each time. Var kinds: v.text v.int v.float v.seed v.toggle
v.choice v.size v.loras v.prompt.
• vars may be a LAMBDA receiving v: vars: (v) => ({ loras: v.loras(/krea/i), … })
  so cross-referencing vars share one scope, with the host's generated types
  injected (no import, no user-side cast).
• v.loras accepts a RegExp directly (resolved against the host's real lora
  list at define time) or any dynamic list such as host.schema.getLoras(/xl/i).
  activeLoras(vars.loras) normalizes to { lora_name, strength_model, strength_clip }[] for a standard LoraLoader chain.
• v.prompt yields a structured { positive, negative } at build time: //
  lines are comments, - lines are negative prompt lines, and with
  { loraKeywordsFrom: lorasVar } the active loras' trigger words prefix the
  positive prompt.
• v.seed is a MODE plus a number (+ N increment, - N decrement, = N
  fixed, ? N reroll), advanced after every run, so a queued batch gets
  distinct seeds.

Execution

• workflow.run({ log, onProgress }) reports live progress; run({ log: true })
  renders a single updating console line. execution.done resolves on success
  AND failure (inspect status), and execution.images are MediaImages on
  disk.
• start() freezes an ExecutionSnapshot (api json + workflow json) at send
  time, so the live workflow stays editable while a run is in flight.
• Failed image retrievals land in execution.imageErrors instead of hanging
  done.

The TUI (bunx comfy-ts tui)

• Keyboard-first three-panel terminal UI over every *.cflow.ts in a folder:
  workflow tree with drafts nested under each workflow, vars panel, preview.
• Drafts: named var-value snapshots per workflow, always-in-a-draft, debounced
  autosave, full CRUD; the last active draft per workflow is remembered.
• Preview panel: live latent previews while a run is in flight, then the final
  output. p cycles native → ansi → off; on iTerm2 / WezTerm / VS Code
  terminals "native" paints the REAL image, elsewhere truecolor half-blocks.
• r during a run queues another prompt on the server; progress, previews and
  outputs follow every queued run, not just the first.
• Multiline prompt editor with readline word ops, line-wise motion, //
  comment toggling, - negative lines, and the injected lora keywords shown
  as non-editable chrome.
• Host panel: node/lora/embedding counts, live server queue length, re-codegen
  the SDK, restart ComfyUI, clear the pending queue, interrupt the current run.

Imports, exports, ecosystem

• toApiJson() / toWorkflowJson() (autolayouted litegraph JSON, drag it into
  the ComfyUI editor) and the reverse host.importApiJson() /
  host.importWorkflowJson().
• ComfyUI-Manager registry mirror: install custom nodes and models, plus
  generated Known* unions over the whole ecosystem.
• ComfyUI-Lora-Manager support: lora preview images in the TUI when the
  extension is present, quietly skipped when it is not.

💥 Breaking

• The great renaming, pre-1.0: ComfyPrompt → ComfyExecution,
  sendPrompt/sendPromptAndWaitUntilDone → start()/run(),
  prompt.finished → execution.done, json_forPrompt()/json_workflow() →
  toApiJson()/toWorkflowJson(), CONNECT()/DISCONNECT() →
  connect()/disconnect(), createEmptyWorkflow() → workflow(),
  MediaImageL → MediaImage, nameInCushy → nodeKey, src/livegraph/ →
  src/graph/.
• ComfyUIObjectInfoParsed and ComfySchema merged into one ComfySchema.

Under the hood

• Dual ESM/CJS build via tsdown (TypeScript 7 / tsgo emits the dts), oxlint +
  oxfmt, bun test.
• The library typechecks with NO generated sdk on disk, which is what a fresh
  clone and CI see.