@civitai/app-sdk 0.77.0 → 0.78.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md CHANGED
@@ -666,7 +666,7 @@ The starters in `civitai/civitai-app-starters` wire this into framework-specific
666
666
 
667
667
  ## Choosing a workflow step type
668
668
 
669
- The orchestrator is a workflow API: each request submits a list of typed steps. `WORKFLOW_STEP_TYPES` is the in-code catalog of every step `$type` it accepts, with a one-line description for each — `textToImage`, `imageGen`, `videoGen`, `comfy`, `customComfy`, `textToSpeech`, `aceStepAudio`, `transcription`, `imageUpscaler` among them (55 in total).
669
+ The orchestrator is a workflow API: each request submits a list of typed steps. `WORKFLOW_STEP_TYPES` is the in-code catalog of every step `$type` it accepts, with a one-line description for each — `textToImage`, `imageGen`, `videoGen`, `comfy`, `customComfy`, `textToSpeech`, `aceStepAudio`, `transcription`, `imageUpscaler` among them (58 in total).
670
670
 
671
671
  The catalog is pinned to the orchestrator's published OpenAPI spec two ways — an offline unit test against a transcribed copy of the spec's `WorkflowStepTemplate` discriminator mapping, and a CI job (`pnpm check:catalogs`) that re-fetches the live spec and diffs it. If a `$type` is listed here, the orchestrator accepts it.
672
672
 
@@ -741,15 +741,15 @@ What it exports:
741
741
 
742
742
  | Export | What |
743
743
  |---|---|
744
- | `WorkflowStepTemplates` | `$type` → template type, for 47 of the catalog's 55 step types. Keyed by the WIRE name, which the generated type names don't always match (`model3DPreview` → `Model3dPreviewStepTemplate`). |
744
+ | `WorkflowStepTemplates` | `$type` → template type, for 47 of the catalog's 58 step types. Keyed by the WIRE name, which the generated type names don't always match (`model3DPreview` → `Model3dPreviewStepTemplate`). |
745
745
  | `WorkflowStepTemplateFor<'videoGen'>` | One step's template. |
746
746
  | `WorkflowStepInputFor<'videoGen'>` | One step's `input` shape, without needing the generated `*Input` name. |
747
747
  | `AnyWorkflowStepTemplate` | Discriminated union of all 47 mapped templates — `Extract<…, { $type: 'comfy' }>` and exhaustive `switch` work. `@civitai/client`'s base `WorkflowStepTemplate` has `$type` as a bare `string`, so it narrows nothing. |
748
748
  | `TypedWorkflowTemplate` | The submit envelope with `steps` narrowed to that union. Pass it straight to `submitWorkflow` / `estimateWorkflow`. |
749
749
 
750
- > 🔴 **The map is not total over the catalog, and that is the expected state.** `WORKFLOW_STEP_TYPES` documents 55 `$type`s; this map covers 47. The 8 with no generated template in the pinned `@civitai/client` are `imageScanning`, `liveTextToSpeech`, `liveTranscription`, `merge`, `preprocessVideo`, `promptModeration`, `soniloAudioGen`, `yuE2`, and `WorkflowStepTemplateFor<…>` is a compile error for each of them.
750
+ > 🔴 **The map is not total over the catalog, and that is the expected state.** `WORKFLOW_STEP_TYPES` documents 58 `$type`s; this map covers 47. The 11 with no generated template in the pinned `@civitai/client` are `archiveExtraction`, `audioModeration`, `imageScanning`, `liveTextToSpeech`, `liveTranscription`, `merge`, `multimodalEmbedding`, `preprocessVideo`, `promptModeration`, `soniloAudioGen`, `yuE2`, and `WorkflowStepTemplateFor<…>` is a compile error for each of them.
751
751
  >
752
- > The two surfaces move independently on purpose: the catalog tracks the **live** orchestrator spec (a daily job syncs it), while these types track whatever `@civitai/client` was last published from. So the catalog runs ahead and the client catches up. The gap is never silent — `test/orchestrator/step-templates.test-d.ts` carries it as a `never` ledger plus one `@ts-expect-error` per gap `$type`, and `test/orchestrator/step-count-prose.test.ts` derives all four numbers (55, 47, 8, and the names) from `WORKFLOW_STEP_TYPES` and the map's own AST, then fails if this paragraph or its twin in `src/orchestrator/steps.ts` disagrees by one character.
752
+ > The two surfaces move independently on purpose: the catalog tracks the **live** orchestrator spec (a daily job syncs it), while these types track whatever `@civitai/client` was last published from. So the catalog runs ahead and the client catches up. The gap is never silent — `test/orchestrator/step-templates.test-d.ts` carries it as a `never` ledger plus one `@ts-expect-error` per gap `$type`, and `test/orchestrator/step-count-prose.test.ts` derives all four numbers (58, 47, 11, and the names) from `WORKFLOW_STEP_TYPES` and the map's own AST, then fails if this paragraph or its twin in `src/orchestrator/steps.ts` disagrees by one character.
753
753
 
754
754
  The generated `*StepTemplate` and `*Input` types are **not** re-exported individually. Using this subpath already requires `@civitai/client` installed, so if you want one by name, import it straight from there — `import type { TextToImageStepTemplate } from '@civitai/client'`.
755
755
 
@@ -788,7 +788,7 @@ Several of the 47 exist to serve Civitai's own pipelines rather than third-party
788
788
 
789
789
  `training` and `imageResourceTraining` are the exception: an App Block's host explicitly **allows** them on the pass-through arm (see `WorkflowBodyPassThroughStep` under Primitives above for the `maxBuzz` bound, and how a trained epoch is published). For a standalone app, the orchestrator's own authorization still decides.
790
790
 
791
- Note that `WORKFLOW_STEP_TYPES` does **not** mark most of them: of its 55 entries exactly three — `comfyNodepackSnapshot`, `qwenImageBench` and `promptModeration` — sit under its "Platform internals" heading, and the rest are ordinary documented entries (`webScrape` even carries usage notes). The reason the platform steps are typed anyway is not that the catalog flags them as internal; it's that the catalog *documents* them, so skipping them would make `WorkflowStepTemplateFor<'training'>` a compile error for a step type the SDK documents — which is exactly what is live today for each `$type` the pinned client cannot type, and is why that gap is spelled out above rather than left to be discovered.
791
+ Note that `WORKFLOW_STEP_TYPES` does **not** mark most of them: of its 58 entries exactly three — `comfyNodepackSnapshot`, `qwenImageBench` and `promptModeration` — sit under its "Platform internals" heading, and the rest are ordinary documented entries (`webScrape` even carries usage notes). The reason the platform steps are typed anyway is not that the catalog flags them as internal; it's that the catalog *documents* them, so skipping them would make `WorkflowStepTemplateFor<'training'>` a compile error for a step type the SDK documents — which is exactly what is live today for each `$type` the pinned client cannot type, and is why that gap is spelled out above rather than left to be discovered.
792
792
 
793
793
  ## Public vs. confidential clients
794
794
 
@@ -196,6 +196,14 @@ export declare const WORKFLOW_STEP_TYPES: {
196
196
  readonly imageScanning: "Unified image moderation scan (NSFW level, AI/anime recognition, tagging, age detection)";
197
197
  /** Shieldstral text/prompt safety moderation (`mode: 'prompt' | 'text'`). */
198
198
  readonly shieldstralModeration: "Text / prompt safety moderation (Shieldstral)";
199
+ /**
200
+ * Transcribe the audio at `mediaUrl` and classify the lyrics for sexualized
201
+ * minors. `transcriber` is `auto`, `aceStep` or `asr`; `auto` sends speech in
202
+ * a supported language to ASR and everything else to ACE-Step. The output
203
+ * carries the `lyrics`, a `score`, `sexualizedMinors` and `blocked`, plus
204
+ * `hasAudio` and a `skippedReason`.
205
+ */
206
+ readonly audioModeration: "Audio moderation — transcribe lyrics and flag sexualized minors";
199
207
  /** ClamAV scan a model file for malware. */
200
208
  readonly modelClamScan: "Antivirus scan a model file";
201
209
  /** Pickle-scan a model file for unsafe pickles. */
@@ -206,6 +214,13 @@ export declare const WORKFLOW_STEP_TYPES: {
206
214
  readonly chatCompletion: "LLM chat completion";
207
215
  /** Generate a richer prompt from a short seed prompt. */
208
216
  readonly promptEnhancement: "Prompt expansion via LLM";
217
+ /**
218
+ * Embed 1–128 `inputs` (each a unique `id` carrying `text`, an `imageUrl` or
219
+ * a `videoUrl` of at most 60 s / 100 MiB, audio ignored) into normalized
220
+ * vectors with `wemm-2b`. `dimensions` is 64–2048 in powers of two, default
221
+ * 256.
222
+ */
223
+ readonly multimodalEmbedding: "Text / image / video embeddings (wemm-2b)";
209
224
  /**
210
225
  * Fetch one page and return its content. `url` and `formats` are both
211
226
  * REQUIRED; `formats` selects which representations come back — the spec
@@ -226,6 +241,13 @@ export declare const WORKFLOW_STEP_TYPES: {
226
241
  readonly echo: "Echo step — round-trip the input for testing";
227
242
  /** Package multiple blobs into a zip archive. */
228
243
  readonly blobArchive: "Zip multiple blobs into an archive";
244
+ /**
245
+ * Unpack a `.zip`, `.tar` or `.tar.gz` at `sourceUrl`. `mediaTypes` limits
246
+ * which kinds of entry are extracted (default all) and `maxEntries` caps the
247
+ * count (1–2000, default 2000); anything filtered or over the cap comes back
248
+ * in `rejected` rather than failing the step. `entries` are in archive order.
249
+ */
250
+ readonly archiveExtraction: "Extract the entries of a zip / tar / tar.gz archive";
229
251
  /** Combine fields from earlier steps' outputs into a single result. */
230
252
  readonly merge: "Merge — combine fields from earlier steps' outputs into one result";
231
253
  /** Snapshot the installed ComfyUI custom-node packs on a worker. */
@@ -272,6 +294,14 @@ export declare const IMAGE_GEN_ENGINES: {
272
294
  * workers) — picking this engine pins the first-party vendor path.
273
295
  */
274
296
  readonly krea: "Krea (Krea 2, first-party API)";
297
+ /**
298
+ * Krea 3: text-to-image, or an edit of up to six `images`. `resolution` is
299
+ * `1k` or `2k` (default `2k`), `thinkingMode` is `instant` / `medium` /
300
+ * `ultra`, and `quantity` is capped at 4. Omit `aspectRatio` to let Krea
301
+ * choose. Note the spec marks `images` required (no minimum length) while
302
+ * its own description says to omit it for text-to-image.
303
+ */
304
+ readonly krea3: "Krea 3 (generation and editing, 1K / 2K)";
275
305
  /** Self-hosted SDCpp (stable-diffusion.cpp) gen. */
276
306
  readonly sdcpp: "SDCpp (self-hosted diffusion)";
277
307
  /** fal.ai routed gen. */
@@ -204,6 +204,14 @@ export const WORKFLOW_STEP_TYPES = {
204
204
  imageScanning: 'Unified image moderation scan (NSFW level, AI/anime recognition, tagging, age detection)',
205
205
  /** Shieldstral text/prompt safety moderation (`mode: 'prompt' | 'text'`). */
206
206
  shieldstralModeration: 'Text / prompt safety moderation (Shieldstral)',
207
+ /**
208
+ * Transcribe the audio at `mediaUrl` and classify the lyrics for sexualized
209
+ * minors. `transcriber` is `auto`, `aceStep` or `asr`; `auto` sends speech in
210
+ * a supported language to ASR and everything else to ACE-Step. The output
211
+ * carries the `lyrics`, a `score`, `sexualizedMinors` and `blocked`, plus
212
+ * `hasAudio` and a `skippedReason`.
213
+ */
214
+ audioModeration: 'Audio moderation — transcribe lyrics and flag sexualized minors',
207
215
  /** ClamAV scan a model file for malware. */
208
216
  modelClamScan: 'Antivirus scan a model file',
209
217
  /** Pickle-scan a model file for unsafe pickles. */
@@ -215,6 +223,13 @@ export const WORKFLOW_STEP_TYPES = {
215
223
  chatCompletion: 'LLM chat completion',
216
224
  /** Generate a richer prompt from a short seed prompt. */
217
225
  promptEnhancement: 'Prompt expansion via LLM',
226
+ /**
227
+ * Embed 1–128 `inputs` (each a unique `id` carrying `text`, an `imageUrl` or
228
+ * a `videoUrl` of at most 60 s / 100 MiB, audio ignored) into normalized
229
+ * vectors with `wemm-2b`. `dimensions` is 64–2048 in powers of two, default
230
+ * 256.
231
+ */
232
+ multimodalEmbedding: 'Text / image / video embeddings (wemm-2b)',
218
233
  // ----- Web ---------------------------------------------------------------
219
234
  /**
220
235
  * Fetch one page and return its content. `url` and `formats` are both
@@ -237,6 +252,13 @@ export const WORKFLOW_STEP_TYPES = {
237
252
  echo: 'Echo step — round-trip the input for testing',
238
253
  /** Package multiple blobs into a zip archive. */
239
254
  blobArchive: 'Zip multiple blobs into an archive',
255
+ /**
256
+ * Unpack a `.zip`, `.tar` or `.tar.gz` at `sourceUrl`. `mediaTypes` limits
257
+ * which kinds of entry are extracted (default all) and `maxEntries` caps the
258
+ * count (1–2000, default 2000); anything filtered or over the cap comes back
259
+ * in `rejected` rather than failing the step. `entries` are in archive order.
260
+ */
261
+ archiveExtraction: 'Extract the entries of a zip / tar / tar.gz archive',
240
262
  /** Combine fields from earlier steps' outputs into a single result. */
241
263
  merge: "Merge — combine fields from earlier steps' outputs into one result",
242
264
  // ----- Platform internals ------------------------------------------------
@@ -286,6 +308,14 @@ export const IMAGE_GEN_ENGINES = {
286
308
  * workers) — picking this engine pins the first-party vendor path.
287
309
  */
288
310
  krea: 'Krea (Krea 2, first-party API)',
311
+ /**
312
+ * Krea 3: text-to-image, or an edit of up to six `images`. `resolution` is
313
+ * `1k` or `2k` (default `2k`), `thinkingMode` is `instant` / `medium` /
314
+ * `ultra`, and `quantity` is capped at 4. Omit `aspectRatio` to let Krea
315
+ * choose. Note the spec marks `images` required (no minimum length) while
316
+ * its own description says to omit it for text-to-image.
317
+ */
318
+ krea3: 'Krea 3 (generation and editing, 1K / 2K)',
289
319
  /** Self-hosted SDCpp (stable-diffusion.cpp) gen. */
290
320
  sdcpp: 'SDCpp (self-hosted diffusion)',
291
321
  /** fal.ai routed gen. */
@@ -3,7 +3,7 @@
3
3
  * own generated client (`@civitai/client`).
4
4
  *
5
5
  * The sibling `@civitai/app-sdk/orchestrator` module gives you the *catalog*
6
- * (`WORKFLOW_STEP_TYPES` — 55 `$type` names and what each one does) and the
6
+ * (`WORKFLOW_STEP_TYPES` — 58 `$type` names and what each one does) and the
7
7
  * fetch helpers (`submitWorkflow`, `estimateWorkflow`, …), but its body
8
8
  * builders take `input: unknown`. This module is the missing half: the actual
9
9
  * per-step input shapes, tracked against the orchestrator's OpenAPI spec by
@@ -78,7 +78,7 @@
78
78
  * orchestrator's own authorization still decides.
79
79
  *
80
80
  * ⚠️ `WORKFLOW_STEP_TYPES` does NOT mark most of them. Counted at this commit:
81
- * of its 55 entries, exactly THREE sit under its "Platform internals" heading —
81
+ * of its 58 entries, exactly THREE sit under its "Platform internals" heading —
82
82
  * `comfyNodepackSnapshot`, `qwenImageBench` and `promptModeration`. `training`,
83
83
  * `webScrape`, `xGuardModeration`, `modelPickleScan` and the `model*` / `media*` steps are
84
84
  * ordinary documented entries under ordinary headings, and `webScrape` carries
@@ -97,10 +97,11 @@
97
97
  * (#315) without anything going red. The measured state, derived rather than
98
98
  * typed:
99
99
  *
100
- * `WORKFLOW_STEP_TYPES` documents 55 `$type`s; this map covers 47. The 8 with
101
- * no generated template in the pinned `@civitai/client` are `imageScanning`,
102
- * `liveTextToSpeech`, `liveTranscription`, `merge`, `preprocessVideo`,
103
- * `promptModeration`, `soniloAudioGen`, `yuE2`, and
100
+ * `WORKFLOW_STEP_TYPES` documents 58 `$type`s; this map covers 47. The 11 with
101
+ * no generated template in the pinned `@civitai/client` are `archiveExtraction`,
102
+ * `audioModeration`, `imageScanning`, `liveTextToSpeech`, `liveTranscription`,
103
+ * `merge`, `multimodalEmbedding`, `preprocessVideo`, `promptModeration`,
104
+ * `soniloAudioGen`, `yuE2`, and
104
105
  * `WorkflowStepTemplateFor<…>` is a compile error for each of them.
105
106
  *
106
107
  * That gap is EXPECTED and is not a defect in either surface. The catalog
@@ -309,7 +310,7 @@ interface StepTemplateMap {
309
310
  xGuardModeration: XGuardModerationStepTemplate;
310
311
  }
311
312
  /**
312
- * `$type` → its step-template type, for 47 of the catalog's 55 step types.
313
+ * `$type` → its step-template type, for 47 of the catalog's 58 step types.
313
314
  *
314
315
  * Keyed by the WIRE name rather than the generated type name, because the wire
315
316
  * name is what you actually have in hand and the generator does not always
@@ -3,7 +3,7 @@
3
3
  * own generated client (`@civitai/client`).
4
4
  *
5
5
  * The sibling `@civitai/app-sdk/orchestrator` module gives you the *catalog*
6
- * (`WORKFLOW_STEP_TYPES` — 55 `$type` names and what each one does) and the
6
+ * (`WORKFLOW_STEP_TYPES` — 58 `$type` names and what each one does) and the
7
7
  * fetch helpers (`submitWorkflow`, `estimateWorkflow`, …), but its body
8
8
  * builders take `input: unknown`. This module is the missing half: the actual
9
9
  * per-step input shapes, tracked against the orchestrator's OpenAPI spec by
@@ -78,7 +78,7 @@
78
78
  * orchestrator's own authorization still decides.
79
79
  *
80
80
  * ⚠️ `WORKFLOW_STEP_TYPES` does NOT mark most of them. Counted at this commit:
81
- * of its 55 entries, exactly THREE sit under its "Platform internals" heading —
81
+ * of its 58 entries, exactly THREE sit under its "Platform internals" heading —
82
82
  * `comfyNodepackSnapshot`, `qwenImageBench` and `promptModeration`. `training`,
83
83
  * `webScrape`, `xGuardModeration`, `modelPickleScan` and the `model*` / `media*` steps are
84
84
  * ordinary documented entries under ordinary headings, and `webScrape` carries
@@ -97,10 +97,11 @@
97
97
  * (#315) without anything going red. The measured state, derived rather than
98
98
  * typed:
99
99
  *
100
- * `WORKFLOW_STEP_TYPES` documents 55 `$type`s; this map covers 47. The 8 with
101
- * no generated template in the pinned `@civitai/client` are `imageScanning`,
102
- * `liveTextToSpeech`, `liveTranscription`, `merge`, `preprocessVideo`,
103
- * `promptModeration`, `soniloAudioGen`, `yuE2`, and
100
+ * `WORKFLOW_STEP_TYPES` documents 58 `$type`s; this map covers 47. The 11 with
101
+ * no generated template in the pinned `@civitai/client` are `archiveExtraction`,
102
+ * `audioModeration`, `imageScanning`, `liveTextToSpeech`, `liveTranscription`,
103
+ * `merge`, `multimodalEmbedding`, `preprocessVideo`, `promptModeration`,
104
+ * `soniloAudioGen`, `yuE2`, and
104
105
  * `WorkflowStepTemplateFor<…>` is a compile error for each of them.
105
106
  *
106
107
  * That gap is EXPECTED and is not a defect in either surface. The catalog
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@civitai/app-sdk",
3
- "version": "0.77.0",
3
+ "version": "0.78.0",
4
4
  "description": "OAuth + PKCE, encrypted-cookie sessions, scopes, and orchestrator helpers for building third-party Civitai apps.",
5
5
  "license": "MIT",
6
6
  "type": "module",