The all-in-one local AI studio for your desktop: chat, image and video generation and a coding agent in one free, open source app. Windows and Linux. No Docker, no terminal, no cloud required.
TypeScript agentai-chatai-studioall-in-onecoding-agentdesktop-appfluximage-generationlinuxllmlocal-ailocal-llmopen-sourceprivacyself-hostedstable-diffusionuncensoreduncensored-aivideo-generationwindows
import { getImageModels, getVideoModels, isI2VModel, isT2VCapable } from './comfyui'
import type { ModelPickKind } from '../stores/modelPickStore'
import type { ToolArgs } from './mcp/types'
/**
* Model-Picker gate (v2.5.3). Runs inside executeImageGenerate /
* executeVideoGenerate BEFORE vramHandoffGenerate — i.e. before the VRAM
* swap — and returns the model name the generation should use, or null to
* leave the existing auto-selection untouched. The pick is OURS (LU UI),
* not the LLM's:
*
* - explicit `model` arg from the user/LLM → no picker (intent wins)
* - Create tab selection installed → silent use (the user's
* visible model choice in the product, see chooseFromUserSelection)
* - saved preference installed → silent use (Change-Model
* affordance shows in the tool call instead)
* - nothing chosen anywhere (or it got uninstalled) → ModelPickerCard
* renders in the running tool call; save icon persists the choice
* - ComfyUI unreachable / no models → null (the existing decide
* phase reports its own actionable error — UX identical to before)
*
* Model filtering mirrors vram-handoff's decide phase exactly: image =
* getImageModels; video splits by inputImage into I2V (isI2VModel) vs T2V
* (the rest). The preference keys are per-kind because the sets are
* disjoint (SVD can't T2V, Wan 1.3B can't I2V).
*/
/**
* The model the user has ALREADY chosen for this kind, in order of authority,
* or null when neither choice fits this call.
*
* The Create tab comes first. It is the model selection the user can see and
* changed last, so a Create tab set to Z-Image has to answer a chat request
* for a picture with Z-Image. Nebenbefund N1 of the D1 counter-check (Windows
* build, 2026-08-29): with an empty `model` argument the chat tool ignored
* that and built a Realistic Vision graph in two runs out of two, because the
* saved picker preference was the only choice this gate ever read.
*
* `saved` (the ModelPickerCard preference) stays as the second source, so an
* install that never opened Create behaves exactly as before. Both have to be
* in `names`, the installed and eligible set for THIS call: an uninstalled or
* wrong-capability choice falls through to the picker instead of building a
* graph around a file ComfyUI does not have.
*/
export function chooseFromUserSelection(
names: string[],
createChoice: string | null | undefined,
saved: string | null | undefined,
): string | null {
for (const choice of [createChoice, saved]) {
if (typeof choice === 'string' && choice && names.includes(choice)) return choice
}
return null
}
export async function pickModelForGeneration(
kind: 'image' | 'video',
args: ToolArgs,
): Promise<string | null> {
if (typeof args.model === 'string' && args.model) return null
// Same alias normalization as runHandoff — the pick must classify the call
// the same way the decide phase will (input_image / image → inputImage).
const inputImage = args.inputImage ?? args.input_image ?? args.image
const wantI2V = kind === 'video' && typeof inputImage === 'string' && !!inputImage
const pickKind: ModelPickKind = kind === 'image' ? 'image' : wantI2V ? 'video-i2v' : 'video-t2v'
let names: string[]
try {
const models = kind === 'image' ? await getImageModels() : await getVideoModels()
// T2V uses isT2VCapable (NOT !isI2VModel): Wan 2.2 TI2V is dual T2V/I2V,
// so it must appear in BOTH picker lists — the old negation hid it from
// the T2V picker (live find 2026-06-11, first wan22 T2V pick on the 3060).
const eligible = kind === 'image'
? models
: models.filter((m) => (wantI2V ? isI2VModel(m.name) : isT2VCapable(m.name)))
names = eligible.map((m) => m.name)
} catch {
return null
}
if (names.length === 0) return null
const { useSettingsStore } = await import('../stores/settingsStore')
const prefKey =
pickKind === 'image' ? 'preferredImageModel'
: pickKind === 'video-t2v' ? 'preferredVideoT2VModel'
: 'preferredVideoI2VModel'
const saved = useSettingsStore.getState().settings[prefKey]
// The Create tab's own model selection, read through the same store the
// Create page writes it to. For video the two sub-kinds share one Create
// selection; `names` already carries the I2V/T2V filter, so a T2V-only
// choice simply does not answer an image-to-video call.
const { useCreateStore } = await import('../stores/createStore')
const createState = useCreateStore.getState()
const createChoice = pickKind === 'image' ? createState.imageModel : createState.videoModel
const chosen = chooseFromUserSelection(names, createChoice, saved)
if (chosen) return chosen
// One model that fits is no choice: the card came on top of the tool
// approval and asked a second time for nothing (A12 box run, A13).
if (names.length === 1) return names[0]
const { useModelPickStore } = await import('../stores/modelPickStore')
const choice = await useModelPickStore.getState().request(pickKind, names, names[0])
if (!choice) return names[0]
if (choice.save) {
useSettingsStore.getState().updateSettings({ [prefKey]: choice.model })
}
return choice.model
}