@koda-sl/baker-cli 0.180.0 → 0.181.1-dev.3fc14023b

package/README.md

@@ -1937,6 +1937,7 @@ baker images generate "flat geometric mascot, brand palette" \
 | Model | Best for | Aspect ratios | Sizes |
 |---|---|---|---|
 | `google/gemini-3.1-flash-image-preview` **(default)** | Fast iteration, reference editing, extreme aspect ratios (Nano Banana flash) | standard **+** `1:4` `4:1` `1:8` `8:1` | `0.5K`–`4K` |
+| `google/gemini-3.1-flash-lite-image` | Cheapest tier (Nano Banana 2 Lite) — half the flash price per output token. High-volume variants and any 1K-sufficient frame, portrait ad formats included | standard **+** `1:4` `4:1` `1:8` `8:1` | `1K` **only** |
 | `google/gemini-3-pro-image-preview` | Highest fidelity (Nano Banana Pro) — strong prompt adherence and reference grounding | standard set | `1K` `2K` `4K` |
 | `openai/gpt-image-2` | Photoreal + cleanest in-image text — ad/landing reproduction | standard set **minus** `4:5` `5:4` | *(none — derived from the ratio)* |
 | `openai/gpt-5.4-image-2` | **Legacy** — accepted so canvases authored before the gpt-image-2 switch still re-run. Don't pick it for new work. | standard set | `1K` `2K` `4K` (ignored upstream) |
@@ -1946,13 +1947,15 @@ Standard aspect ratios: `1:1` `2:3` `3:2` `3:4` `4:3` `4:5` `5:4` `9:16` `16:9`
 
 `openai/gpt-image-2` accepts neither `image_size` nor the `4:5` / `5:4` portrait ratios — OpenAI picks the pixel dimensions from the ratio. To ship a 4:5 Meta feed ad from it, render the hero at `3:4` and adapt with `image_aspect_adapt`; the scaffolders do this for you.
 
+`google/gemini-3.1-flash-lite-image` renders at `1K` and nothing else — `2K` / `4K` is a provider 400 (`resolution: not supported. Accepted: 1K`), not a soft clamp. It costs half of `google/gemini-3.1-flash-image-preview` per output image token ($0.00003 vs $0.00006), so it is the cheap-iteration tier for images. Keep hero stills and reference sheets that need `2K`+ on the flash or pro tier.
+
 **Flags:**
 
 | Flag | Description |
 |---|---|
 | `--model` | Model id (default `google/gemini-3.1-flash-image-preview`) |
 | `--aspect-ratio` | Output aspect ratio (default `1:1`) |
-| `--image-size` | Resolution: `1K` (default) `2K` `4K` (Gemini flash also `0.5K`) |
+| `--image-size` | Resolution: `1K` (default) `2K` `4K` (Gemini flash also `0.5K`; Gemini flash-lite is `1K` only) |
 | `--reference` | Comma-separated visual references, each either a **public image URL** (Pinterest / stock / library `imageUrl`) **or a local file path** (a sandbox image — brand logo, product shot, cropped photo, screenshot). Local files are downscaled (≤1536px) and inlined automatically — no manual upload. Applied in order; the biggest quality lever for photographed, on-brand output. Split is on `,`, so a URL containing a literal comma in its query string would be torn in two (rare for image CDNs — pass it alone if it occurs); a single `data:` URL is taken whole. |
 | `--strength` | Recraft only: vectorization strength 0–1 |
 | `--rgb-colors` | Recraft only: JSON palette `[[r,g,b],…]` |
@@ -3361,6 +3364,24 @@ Gemini flash image model — the default. Fast iteration, reference editing, ext
 | `aspect_ratio` | enum | no | EXTREME AR |
 | `image_size` | enum | no | `0.5K \| 1K \| 2K \| 4K` |
 
+###### Model: `google/gemini-3.1-flash-lite-image`
+
+"Nano Banana 2 Lite" — the cheap-iteration tier. Same Gemini image family as the flash model (same extreme ratios, same reference editing, up to 14 reference images) at **half the output-image-token price** ($0.00003 vs $0.00006), so a 1K render lands around 5 credits. What it gives up is resolution: `1K` is the only size it renders.
+
+| Name | Type | Required | Notes |
+|---|---|---|---|
+| `model` | literal | yes | `"google/gemini-3.1-flash-lite-image"` |
+| `prompt` | string | yes | non-empty |
+| `aspect_ratio` | enum | no | EXTREME AR — keeps `4:5` / `5:4`, so Meta/Instagram portrait formats come straight out of it |
+| `image_size` | enum | no | `1K` only — `2K` / `4K` is a hard provider 400 (`resolution: not supported. Accepted: 1K`), so `validate` rejects it before the call costs anything |
+
+Reach for it for high-volume variants and any frame `1K` covers. Send hero stills and the `image_reference_sheet` subject sheet — anything that needs `2K`+ — to `google/gemini-3.1-flash-image-preview` or `google/gemini-3-pro-image-preview`. Registered for `image_generate` only; `image_reference_sheet` and `image_aspect_adapt` do not accept it.
+
+```json
+{ "id": "variant_a", "type": "image_generate",
+  "params": { "model": "google/gemini-3.1-flash-lite-image", "prompt": "Portrait feed ad for a citrus energy drink …", "aspect_ratio": "4:5", "image_size": "1K" } }
+```
+
 ###### Model: `google/gemini-3-pro-image-preview`
 
 "Nano Banana Pro" — heavier, higher-quality Gemini.
@@ -3665,7 +3686,7 @@ Fail-fast: an invalid model selection (wrong count, duplicate or out-of-range pi
 
 ##### `video_generate`
 
-Generate video. Async with polling. Two curated models.
+Generate video. Curated model roster (per-model params below). Async with polling, except `google/gemini-omni-flash`, which returns its clip inline.
 
 **Inputs**
 
@@ -3688,6 +3709,36 @@ Accepted ref-image MIMEs vary by model — see per-model sections below.
 
 ---
 
+###### Model: `google/gemini-omni-flash` (default)
+
+The default `video_generate` model (`DEFAULT_VIDEO_GENERATE_MODEL`). Natively multimodal — one call returns picture **and** synchronized audio. **Not on OpenRouter** (its `GET /api/v1/videos/models` catalogue of 17 video models does not include it): routed direct to Google's **Interactions API**, `POST https://generativelanguage.googleapis.com/v1beta/interactions`, wire model `gemini-omni-flash-preview`. Its own `generateContent` endpoint refuses the model ("This model only supports Interactions API"), and it does not use Veo's `predictLongRunning` shape. The call is synchronous (~30–40s) and returns the MP4 inline as base64 — no polling.
+
+Ref-image MIMEs: `image/png`, `image/jpeg`, `image/webp`. Takes **one** conditioning image — there is no `last_frame`, and wiring one raises `VIDEO_MODEL_IGNORES_INPUT`. Tasks: `text_to_video`, `image_to_video`, `reference_to_video`, `edit`, `extend` — the only curated model besides Seedance that accepts `references`.
+
+| Name | Type | Required | Notes |
+|---|---|---|---|
+| `model` | literal | yes | `"google/gemini-omni-flash"` |
+| `prompt` | string | yes | non-empty |
+| `aspect_ratio` | enum | no | `16:9 \| 9:16` (`1:1` is rejected) |
+| `duration` | enum | no | `3, 4, 5, 6, 7, 8, 9, 10` seconds. Omitted ⇒ **10s** — the model's own default *and* its ceiling. |
+
+> Those are the **only** accepted params. `resolution`, `seed`, `negative_prompt`, `generate_audio`, `person_generation`, `enhance_prompt` and `conditioning_scale` each fail pre-dispatch with a 400 `Unknown parameter`. Output is fixed: **720p** (1280×720 or 720×1280), 24fps, h264 + native stereo AAC — audio is always on.
+
+> `duration` is the exception. Omni has **no duration field** — every spelling at every nesting level (`duration`, `duration_seconds`, `seconds`, `length`, `video_length`, `num_frames`, at top level, inside `response_format`, inside `generation_config.video_config`) comes back `Unknown parameter`. But it **honours a length asked for in the prompt**, so the backend adapter (`convex/lib/geminiOmni.ts`) translates the canonical `duration` param into a prompt directive. Measured: "exactly 3 seconds" → 3.008s, "exactly 6 seconds" → 6.016s, no length named → 10.005s.
+
+> **The 3–10s range is ours to enforce.** Outside it the model does not error — it silently snaps (2s came back 3.008s, 15s came back 10.005s) and bills the length it actually rendered. That is why `baker canvas validate` gates the enum, for free, before the render.
+
+```json
+{ "id": "clip", "type": "video_generate",
+  "inputs": { "first_frame": "$ref:start.images#0" },
+  "params": { "model": "google/gemini-omni-flash", "prompt": "Same person, talking head, …",
+              "aspect_ratio": "9:16", "duration": 6 } }
+```
+
+> **Billed per token, which resolves to a clean per-second rate** — $17.50 per 1M video output tokens ($1.50 per 1M input, noise), and 720p video is a flat **5,792 tokens per second at every length** (3.008s → 17,376 tokens; 6.016s → 34,752; 10.005s → 57,920). So **$0.10136 per second**: a 10s clip ~$1.01 (102 credits), a 4s clip ~$0.41 (41 credits), a 3s clip ~$0.30 (31 credits). Cutting a clip short is a real saving. A 5s 720p Seedance clip is $2.50.
+
+> **Move off the default** when a beat needs 1080p, a length outside 3–10s, a `1:1` frame, or a photoreal human face on camera (that still routes to Veo, unchanged).
+
 ###### Model: `bytedance/seedance-2.0`
 
 Production-quality ad-creative model. Routed via **Replicate** (`bytedance/seedance-2.0`). NOTE: ByteDance's upstream "real person" likeness filter still blocks photorealistic human reference frames on **any** reseller — the escape is a synthetic/AI presenter face or routing real faces to Veo, not the provider.
@@ -4356,7 +4407,7 @@ baker canvas run ./reference-ad.video.canvas.json
 | `--deconstruct-model <id>` | `~google/gemini-pro-latest` | Override the `video_deconstruct` model. |
 | `--select-model <id>` | `~google/gemini-flash-latest` | Override the element-selection `text_generate` model. |
 | `--image-model <id>` | `openai/gpt-image-2` | Override the per-frame `image_generate` model (defaults to the strongest, matching `scaffold-static-ad`). |
-| `--video-model <id>` | scored router | Override the `video_generate` model **and skip the router**. Curated roster: `bytedance/seedance-2.0` (workhorse), `google/veo-3.1` (cine ceiling + real-face), `google/veo-3.1-fast` (cheap Veo), `kwaivgi/kling-v3.0-pro` (motion-transfer). |
+| `--video-model <id>` | scored router | Override the `video_generate` model **and skip the router**. Curated roster: `google/gemini-omni-flash` (**the default video model** — native audio + reference images, ~$0.101 per second of clip, 720p/`16:9`\|`9:16`, `duration` 3–10s and 10s when unset), `bytedance/seedance-2.0` (workhorse), `google/veo-3.1` (cine ceiling + real-face), `google/veo-3.1-fast` (cheap Veo), `kwaivgi/kling-v3.0-pro` (motion-transfer). |
 | `--real-face` | auto-detected | Force the Veo route for a photoreal human presenter (dodges the ByteDance real-person filter). **Usually unnecessary** — the scaffold auto-detects a photoreal on-camera person/animal cast from the deconstruct and routes it to Veo on its own; set this only to force it when the cast isn't detected. Covers AI-generated photoreal faces too, not just real ones. |
 | `--motion-transfer` | off | Motion driven from a reference / hyper-dynamic → the router picks Kling. |
 | `--identity` | off | Same character/product across clips → the router picks the Seedance workhorse. |
@@ -5022,6 +5073,7 @@ This CLI is designed for AI agent consumption. Key patterns:
 - **0.179.0**: `baker ads google keywords update --final-url` / `--clear-final-url` — changing where one keyword sends its clicks no longer costs the keyword its history. The write contract has always accepted `finalUrls` on a keyword update (`baker capabilities google-ads` reported it as a field of `google.keyword.update`), but no CLI flag reached it, so an agent that followed capabilities got "update needs at least one field" and the only route left was remove + re-add — a new criterion to Google, learning from zero, for a change that was only ever about the destination. `--clear-final-url` stages the empty set Google reads as "inherit the ad's landing page", which is the undo for the misuse `--final-url` already warns about. Same nudge as `add` on the way in.
 
   A keyword's final URL now also shows on its row when the keyword is being **removed**. 0.178.0 read it from the stage-time snapshot, but the snapshot never asked Google for it, so the removal half of that change silently did nothing — the card said which keyword was going away but not which landing page it had been sending its clicks to. The batch hint says outright that `--final-url` has no per-keyword form rather than implying destinations could differ within one call, and the tool doc now lists each `keywords` subcommand's own flags instead of one shared set (`--ad-group-ref`, `--text` and `--match-type` are `add`-only; `remove` takes a target and nothing else).
+- **0.180.1**: `baker tag-manager --help` said the container version Baker cuts at publish is **NOT** published, and told the agent to have the user publish it in Tag Manager. The executor publishes it — a clean run takes the client's container live — so the help understated an irreversible change to a production site. Corrected, and the stale `versionPendingMessage` helper (no callers, same wrong claim) removed from `@baker/api/tag-manager`. `draft list` already reported both outcomes correctly and is unchanged.
 
 ## Publishing
 

package/dist/{chunk-P2T3IZRE.js → chunk-UZ37VVP4.js}

@@ -7,9 +7,9 @@ import {
   __toESM
 } from "./chunk-YL3HDEIJ.js";
 
-// ../../node_modules/.pnpm/safe-stable-stringify@2.5.0/node_modules/safe-stable-stringify/index.js
+// ../../.pnpm-store/v10/links/@/safe-stable-stringify/2.5.0/810146e81bae4e3a061fe487864f2fde80c4b03b886877dc0f1fffbc6480b67e/node_modules/safe-stable-stringify/index.js
 var require_safe_stable_stringify = __commonJS({
-  "../../node_modules/.pnpm/safe-stable-stringify@2.5.0/node_modules/safe-stable-stringify/index.js"(exports, module) {
+  "../../.pnpm-store/v10/links/@/safe-stable-stringify/2.5.0/810146e81bae4e3a061fe487864f2fde80c4b03b886877dc0f1fffbc6480b67e/node_modules/safe-stable-stringify/index.js"(exports, module) {
     "use strict";
     var { hasOwnProperty } = Object.prototype;
     var stringify = configure2();
@@ -1075,7 +1075,7 @@ function resolveAdaptFormats(params) {
   return params.formats ?? [];
 }
 
-// ../../node_modules/.pnpm/safe-stable-stringify@2.5.0/node_modules/safe-stable-stringify/esm/wrapper.js
+// ../../.pnpm-store/v10/links/@/safe-stable-stringify/2.5.0/810146e81bae4e3a061fe487864f2fde80c4b03b886877dc0f1fffbc6480b67e/node_modules/safe-stable-stringify/esm/wrapper.js
 var import__ = __toESM(require_safe_stable_stringify(), 1);
 var configure = import__.default.configure;
 var wrapper_default = import__.default;
@@ -1122,9 +1122,14 @@ var OPENROUTER_IMAGE_AR_EXTREME = [...OPENROUTER_IMAGE_AR, "1:4", "4:1", "1:8",
 var OPENROUTER_IMAGE_AR_GPT = ["1:1", "2:3", "3:2", "3:4", "4:3", "9:16", "16:9", "21:9"];
 var OPENROUTER_IMAGE_SIZE = ["1K", "2K", "4K"];
 var OPENROUTER_IMAGE_SIZE_EXTENDED = ["0.5K", ...OPENROUTER_IMAGE_SIZE];
+var GEMINI_LITE_IMAGE_SIZE = ["1K"];
 var OPENROUTER_IMAGE_QUALITY = ["auto", "low", "medium", "high"];
 var SEEDANCE_DURATIONS = [4, 5, 6, 8, 10, 12, 15];
 var KLING_DURATIONS = [3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15];
+var GEMINI_OMNI_DURATIONS = [3, 4, 5, 6, 7, 8, 9, 10];
+var GEMINI_OMNI_DEFAULT_SECONDS = 10;
+var GEMINI_OMNI_ASPECT_RATIOS = ["16:9", "9:16"];
+var GEMINI_OMNI_IMAGE_MIMES = ["image/png", "image/jpeg", "image/webp"];
 var ELEVENLABS_OUTPUT_FORMATS = [
   "mp3_22050_32",
   "mp3_44100_32",
@@ -1144,16 +1149,30 @@ var IMAGE_GENERATE_MODELS = [
   "openai/gpt-image-2",
   "openai/gpt-5.4-image-2",
   "google/gemini-3.1-flash-image-preview",
+  "google/gemini-3.1-flash-lite-image",
   "google/gemini-3-pro-image-preview",
   "recraft/recraft-v4.1-pro-vector"
 ];
 var VIDEO_GENERATE_MODELS = [
   "bytedance/seedance-2.0",
+  "google/gemini-omni-flash",
   "google/veo-3.1",
   "google/veo-3.1-fast",
   "kwaivgi/kling-v3.0-pro"
 ];
-var DEFAULT_VIDEO_GENERATE_MODEL = "bytedance/seedance-2.0";
+var DEFAULT_VIDEO_GENERATE_MODEL = "google/gemini-omni-flash";
+function isGeminiOmniModel(model) {
+  return model.startsWith("google/gemini-omni");
+}
+var LAST_FRAME_TO_VIDEO_MODELS = [
+  "bytedance/seedance-2.0",
+  "google/veo-3.1",
+  "google/veo-3.1-fast",
+  "kwaivgi/kling-v3.0-pro"
+];
+function supportsLastFrame(model) {
+  return LAST_FRAME_TO_VIDEO_MODELS.includes(model);
+}
 var MODEL_REGISTRY = {
   text_generate: {
     "~google/gemini-flash-latest": {
@@ -1286,6 +1305,31 @@ var MODEL_REGISTRY = {
         image_size: { kind: "string", enum: OPENROUTER_IMAGE_SIZE_EXTENDED }
       }
     },
+    "google/gemini-3.1-flash-lite-image": {
+      // Nano Banana 2 Lite. Same Gemini image family as the flash tier — same
+      // extreme aspect-ratio set, same reference editing — at HALF the output-token
+      // price ($0.00003 vs $0.00006 per output image token), which makes it the
+      // cheap-iteration tier for image work the way `veo-3.1-fast` is for video.
+      //
+      // The one thing it gives up is resolution: it renders 1K and nothing else.
+      // That is a hard provider 400, not a soft clamp, so it is gated here rather
+      // than discovered at dispatch. A 2K/4K hero frame belongs on the flash or pro
+      // tier. Unlike gpt-image it keeps 4:5 and 5:4, so Meta/Instagram portrait
+      // formats still come straight out of it.
+      //
+      // Reached through the normal OpenRouter image gateway (Google Vertex and
+      // Google AI Studio both serve it) — no direct-provider exception needed.
+      // Source: GET /api/v1/images/models/google/gemini-3.1-flash-lite-image/endpoints.
+      label: "Google Gemini 3.1 Flash Lite Image",
+      inputs: [],
+      optional_inputs: [{ kind: "image", mimes: OPENROUTER_IMAGE_MIMES }],
+      required: ["prompt"],
+      params: {
+        prompt: { kind: "string" },
+        aspect_ratio: { kind: "string", enum: OPENROUTER_IMAGE_AR_EXTREME },
+        image_size: { kind: "string", enum: GEMINI_LITE_IMAGE_SIZE }
+      }
+    },
     "google/gemini-3-pro-image-preview": {
       label: "Google Gemini 3 Pro Image (Nano Banana Pro)",
       inputs: [],
@@ -1294,8 +1338,10 @@ var MODEL_REGISTRY = {
       params: {
         prompt: { kind: "string" },
         aspect_ratio: { kind: "string", enum: OPENROUTER_IMAGE_AR },
-        image_size: { kind: "string", enum: OPENROUTER_IMAGE_SIZE },
-        quality: { kind: "string", enum: OPENROUTER_IMAGE_QUALITY }
+        image_size: { kind: "string", enum: OPENROUTER_IMAGE_SIZE }
+        // No `quality`: OpenRouter's per-endpoint schema lists it for gpt-image and
+        // for NO Gemini image model, pro included. It is accepted-then-dropped, so
+        // declaring it invites the scaffolder to emit a param that does nothing.
       }
     },
     "recraft/recraft-v4.1-pro-vector": {
@@ -1407,6 +1453,41 @@ var MODEL_REGISTRY = {
         generate_audio: { kind: "boolean" }
       }
     },
+    "google/gemini-omni-flash": {
+      // Google Gemini Omni Flash — the DEFAULT video model. Natively multimodal:
+      // it reasons over text + images and renders video WITH synchronized audio in
+      // one call, and it is the only curated model that takes reference images
+      // through its own first-party API (task `reference_to_video`).
+      //
+      // NOT on OpenRouter. `GET /api/v1/videos/models` lists 17 video models and
+      // Omni is not among them, so this model is the sanctioned direct-provider
+      // exception (see `convex/lib/geminiOmni.ts`): Google's Interactions API,
+      // `POST /v1beta/interactions`. Its own `generateContent` endpoint refuses it
+      // ("This model only supports Interactions API"), so the Veo `predictLongRunning`
+      // shape does not apply either.
+      //
+      // The param set below is DELIBERATELY tiny and was read off the live API, not
+      // off the docs — every knob the other video models take is rejected with a
+      // pre-dispatch 400 (`Unknown parameter '<name>'`): no `resolution`, no `seed`,
+      // no `negative_prompt`, no `generate_audio`, no `person_generation` /
+      // `enhance_prompt` / `conditioning_scale`. Resolution is fixed at 720p/24fps
+      // and audio is always on. Widening this list does not widen the model; it just
+      // moves a free validation error into a billed one.
+      //
+      // `duration` is the exception, and the reason it looks odd: Omni has no
+      // duration FIELD, but it honours a length named in the prompt (3-10s), so the
+      // adapter turns this param into a prompt directive. It is a real, billable
+      // knob — cost scales exactly with the seconds rendered.
+      label: "Google Gemini Omni Flash",
+      inputs: [],
+      optional_inputs: [{ kind: "image", mimes: GEMINI_OMNI_IMAGE_MIMES }],
+      required: ["prompt"],
+      params: {
+        prompt: { kind: "string" },
+        aspect_ratio: { kind: "string", enum: GEMINI_OMNI_ASPECT_RATIOS },
+        duration: { kind: "number", enum: GEMINI_OMNI_DURATIONS }
+      }
+    },
     "google/veo-3.1": {
       // Photoreal CINE CEILING + the real-face fallback (Veo generates adult
       // humans from a keyframe, dodging ByteDance's real-person filter). Same
@@ -1661,6 +1742,21 @@ function nearestSupportedAspectRatio(kind, model, target) {
   }
   return best;
 }
+function supportedImageSizes(kind, model) {
+  const schema = MODEL_REGISTRY[kind]?.[model]?.params.image_size;
+  return schema?.kind === "string" ? schema.enum : void 0;
+}
+var IMAGE_SIZE_ORDER = ["0.5K", "1K", "2K", "4K"];
+function nearestSupportedImageSize(kind, model, target) {
+  const supported = supportedImageSizes(kind, model);
+  if (!supported || supported.includes(target)) return target;
+  const wanted = IMAGE_SIZE_ORDER.indexOf(target);
+  if (wanted < 0) return target;
+  const ranked = supported.filter((s) => IMAGE_SIZE_ORDER.includes(s)).sort((a, b) => IMAGE_SIZE_ORDER.indexOf(a) - IMAGE_SIZE_ORDER.indexOf(b));
+  if (ranked.length === 0) return target;
+  const belowOrEqual = ranked.filter((s) => IMAGE_SIZE_ORDER.indexOf(s) <= wanted);
+  return belowOrEqual.length > 0 ? belowOrEqual[belowOrEqual.length - 1] : ranked[0];
+}
 
 // ../canvas-contract/src/validateParams.ts
 function isManagedNodeKind(nodeType) {
@@ -1827,8 +1923,14 @@ var USD_PER_SECOND = {
     }
   }
 };
+var GEMINI_OMNI_USD_PER_1M_VIDEO_TOKENS = 17.5;
+var GEMINI_OMNI_VIDEO_TOKENS_PER_SECOND = 5792;
+var GEMINI_OMNI_USD_PER_SECOND = GEMINI_OMNI_VIDEO_TOKENS_PER_SECOND * GEMINI_OMNI_USD_PER_1M_VIDEO_TOKENS / 1e6;
 function estimateVideoCostUsd({ model, duration, resolution, generateAudio }) {
   const seconds = duration ?? DEFAULT_VIDEO_DURATION_S;
+  if (isGeminiOmniModel(model)) {
+    return (duration ?? GEMINI_OMNI_DEFAULT_SECONDS) * GEMINI_OMNI_USD_PER_SECOND;
+  }
   if (isSeedanceModel(model)) return seedanceUsdPerSecond(resolution) * seconds;
   const published = USD_PER_SECOND[model];
   if (published) {
@@ -2688,15 +2790,41 @@ var KLING_PROFILE = {
   durationSet: KLING_DURATIONS2,
   paramDefaults: { negative_prompt: KLING_NEGATIVE_PROMPT, cfg_scale: KLING_CFG_SCALE }
 };
+var OMNI_PROFILE = {
+  id: "omni",
+  dialogueDirective: SPEAKS_PROSE,
+  // Omni renders speech and ambience natively in the same pass, so the prompt
+  // carries the audio direction; there is no separate audio toggle to set.
+  extraDirectives: ["No subtitles, no captions, no on-screen text of any kind."],
+  // Omni exposes NO negative_prompt param — the API rejects the field outright —
+  // so the anti-artifact intent has to be carried positively, Seedance-style,
+  // rather than as an "Avoid:" list Veo would have taken as a param.
+  stabilityDirectives: [
+    "hands and fingers crisp and anatomically correct",
+    "one consistent face and identity across every frame",
+    "smooth, temporally coherent motion that holds steady frame to frame"
+  ],
+  keyframeInstruction: "Preserve the composition and colors of the first frame; change only the motion described.",
+  // Omni reasons over a 1M-token context and rewards a fuller brief than the
+  // diffusion-style models, which dilute past their sweet spot.
+  wordBudget: 200,
+  // 3-10s. Omni has no duration FIELD, but honours a length asked for in the
+  // prompt, so the scaffolder plans real clip lengths here like any other model.
+  durationSet: GEMINI_OMNI_DURATIONS,
+  // Nothing to set. Every knob the other profiles use (negative_prompt,
+  // person_generation, cfg_scale) is a pre-dispatch 400 on this model.
+  paramDefaults: {}
+};
 function clipProfileFor(modelId) {
   if (/^bytedance\/seedance/.test(modelId)) return SEEDANCE_PROFILE;
+  if (/^google\/gemini-omni/.test(modelId)) return OMNI_PROFILE;
   if (/^google\/veo/.test(modelId)) return VEO_PROFILE;
   if (/^kwaivgi\/kling|^kling\//.test(modelId)) return KLING_PROFILE;
   return void 0;
 }
 function clipParamRecipe(profile, intent) {
   const out = {};
-  if (intent === "hero") {
+  if (intent === "hero" && profile.id !== "omni") {
     out.resolution = "1080p";
   }
   if (intent === "hook" && profile.id === "kling") {
@@ -2721,7 +2849,12 @@ var GEMINI_IMAGE_PROFILE = {
   id: "gemini",
   constraintPlacement: "inline",
   photorealCue: true,
-  paramDefaults: { quality: "high" }
+  // NO `quality`. OpenRouter accepts the field on a Gemini image call and returns
+  // 200 — but its per-endpoint schema does not list it and the cost is identical
+  // with and without, so Google drops it: an inert knob. Worse, the registry
+  // (correctly) does not declare it for Gemini, so emitting it fails `baker canvas
+  // validate` with `unknown_param` before the call is ever made.
+  paramDefaults: {}
 };
 var RECRAFT_IMAGE_PROFILE = {
   id: "recraft",
@@ -2855,6 +2988,7 @@ var STAGE_CODES = {
   BRANDMARK_IN_PROMPT: "VIDEO_BRANDMARK_IN_PROMPT",
   SPEECH_EXCEEDS_EXTRACT: "VIDEO_SPEECH_EXCEEDS_EXTRACT",
   PROMPT_PROFILE_MISSING: "VIDEO_PROMPT_PROFILE_MISSING",
+  MODEL_IGNORES_INPUT: "VIDEO_MODEL_IGNORES_INPUT",
   PROMPT_DECISION_MISSING: "VIDEO_PROMPT_DECISION_MISSING",
   HOOK_LAYER_MISSING: "VIDEO_HOOK_LAYER_MISSING",
   IMAGE_PROFILE_MISSING: "IMAGE_PROMPT_PROFILE_MISSING",
@@ -3281,9 +3415,14 @@ function talkingSceneSatisfied(ctx, entry, scene) {
   const nodes = ctx.canvas.nodes;
   if (typeof entry === "object" && "voice_convert_node" in entry) {
     const nativeClipRe = new RegExp(`^s${scene}(_r\\d+)?_clip$`);
-    const clipNativeAudio = nodes.some(
-      (n) => nativeClipRe.test(n.id) && n.type === "video_generate" && n.params?.generate_audio === true
-    );
+    const clipNativeAudio = nodes.some((n) => {
+      if (!nativeClipRe.test(n.id) || n.type !== "video_generate") return false;
+      const params = n.params;
+      if (typeof params?.model === "string" && !supportsParam("video_generate", params.model, "generate_audio")) {
+        return true;
+      }
+      return params?.generate_audio === true;
+    });
     const converted = nodes.some((n) => n.id === entry.voice_convert_node && n.type === "audio_voice_convert");
     return clipNativeAudio && (converted || nativeAudioReachesMix(ctx, scene));
   }
@@ -3344,6 +3483,7 @@ function checkVideoInvariants(ctx) {
   checkReferenceCompleteness(ctx, meta);
   checkClipSpanFitsModel(ctx, meta);
   checkTimelineContract(ctx, meta);
+  checkUnusableLastFrame(ctx);
   checkNativeSegOverlap(ctx);
   checkSpineNormalized(ctx, meta);
   checkOddDimensions(ctx);
@@ -3708,6 +3848,23 @@ function videoModelMaxDuration(model) {
   const nums = durations.filter((d) => typeof d === "number");
   return nums.length > 0 ? Math.max(...nums) : void 0;
 }
+function checkUnusableLastFrame(ctx) {
+  for (const n of ctx.canvas.nodes) {
+    if (n.type !== "video_generate") continue;
+    const model = n.params?.model;
+    if (typeof model !== "string" || supportsLastFrame(model)) continue;
+    const inputs = n.inputs;
+    if (!inputs?.last_frame) continue;
+    ctx.issues.push({
+      path: `nodes[${ctx.idToIndex.get(n.id) ?? -1}].inputs.last_frame`,
+      code: STAGE_CODES.MODEL_IGNORES_INPUT,
+      severity: "warning",
+      node_id: n.id,
+      node_type: "video_generate",
+      message: `"${n.id}" wires a last_frame but runs on ${model}, which takes only ONE conditioning image \u2014 the end keyframe is dropped and the image generated for it is wasted. Drop the last_frame, or move this clip to Veo/Seedance, which honour both frames`
+    });
+  }
+}
 function checkClipSpanFitsModel(ctx, meta) {
   for (const { node, span_s } of meta.clip_spans ?? []) {
     const target = ctx.canvas.nodes.find((n) => n.id === node && n.type === "video_generate");
@@ -7096,16 +7253,9 @@ var dialogueNode = delegated({
 
 // src/engine/nodes/remote/image.ts
 import { z as z17 } from "zod";
-var IMAGE_GENERATE_MODELS2 = [
-  "openai/gpt-image-2",
-  // Legacy — see the registry entry; kept so pre-switch canvases still run.
-  "openai/gpt-5.4-image-2",
-  "google/gemini-3.1-flash-image-preview",
-  "google/gemini-3-pro-image-preview",
-  "recraft/recraft-v4.1-pro-vector"
-];
+var imageModelEnum = z17.enum(IMAGE_GENERATE_MODELS);
 var ImageGenerateParams = z17.object({
-  model: z17.enum(IMAGE_GENERATE_MODELS2),
+  model: imageModelEnum,
   prompt: z17.string().min(1),
   aspect_ratio: z17.enum(["1:1", "16:9", "9:16", "4:3", "3:4", "3:2", "2:3", "4:5", "5:4", "21:9", "1:4", "4:1", "1:8", "8:1"]).optional(),
   image_size: z17.enum(["0.5K", "1K", "2K", "4K"]).optional(),
@@ -7514,8 +7664,8 @@ var videoGenerateNode = delegated({
   id: "video_generate",
   version: "2.0.0",
   category: "video",
-  summary: "Generate video for ad creatives. Curated roster: `bytedance/seedance-2.0` (identity/product workhorse, NO photoreal human faces), `google/veo-3.1` (photoreal cine ceiling + the human-face model), `google/veo-3.1-fast` (cheap Veo iteration), `kwaivgi/kling-3.0` (motion-transfer/dynamic). Async with polling.",
-  when_to_use: "Use `bytedance/seedance-2.0` for identity/product output WITHOUT a photoreal human face on camera. Any clip whose frame shows a photoreal human face \u2014 a UGC creator, presenter, or testimonial speaker, REAL OR AI-GENERATED \u2014 must use `google/veo-3.1`: ByteDance's filter rejects photoreal human faces on Seedance (content_policy_blocked/E005), and a synthetic AI face does NOT clear it. Use `google/veo-3.1-fast` while iterating to keep cost low; `kwaivgi/kling-3.0` for motion-transfer/hyper-dynamic beats. `baker canvas scaffold-video` auto-routes a photoreal cast to Veo; a Seedance clip blocked at runtime also auto-recovers on Veo. Each model gates its own durations/resolutions/aspect ratios in the registry \u2014 see the README per-model section.",
+  summary: "Generate video for ad creatives. Curated roster: `google/gemini-omni-flash` (DEFAULT \u2014 native audio + reference images in one call, 720p, 3-10s), `bytedance/seedance-2.0` (identity/product workhorse, NO photoreal human faces), `google/veo-3.1` (photoreal cine ceiling + the human-face model), `google/veo-3.1-fast` (cheap Veo iteration), `kwaivgi/kling-3.0` (motion-transfer/dynamic). Async with polling.",
+  when_to_use: "`google/gemini-omni-flash` is the DEFAULT: it renders picture and synchronized audio in one call, and is the only model here that takes reference images alongside a prompt. SET `duration` (3-10s) to the length the beat actually needs \u2014 it bills ~$0.101 PER SECOND, so leaving it out renders the 10s default and charges ~$1.01 for a beat that may need 3s (~$0.30). Its resolution is fixed at 720p and it accepts only `16:9` or `9:16`; it takes NO `resolution`, `seed`, `negative_prompt` or `generate_audio` param \u2014 sending one fails validation. Keep the default for ordinary beats; move off it when a beat needs 1080p, a length outside 3-10s, or a 1:1 frame. Use `bytedance/seedance-2.0` for identity/product output WITHOUT a photoreal human face on camera. Any clip whose frame shows a photoreal human face \u2014 a UGC creator, presenter, or testimonial speaker, REAL OR AI-GENERATED \u2014 must use `google/veo-3.1`: ByteDance's filter rejects photoreal human faces on Seedance (content_policy_blocked/E005), and a synthetic AI face does NOT clear it. Use `google/veo-3.1-fast` while iterating to keep cost low; `kwaivgi/kling-3.0` for motion-transfer/hyper-dynamic beats. `baker canvas scaffold-video` auto-routes a photoreal cast to Veo; a Seedance clip blocked at runtime also auto-recovers on Veo. Each model gates its own durations/resolutions/aspect ratios in the registry \u2014 see the README per-model section.",
   inputs: z28.object({
     first_frame: ImageRef.optional(),
     last_frame: ImageRef.optional(),
@@ -7999,9 +8149,11 @@ export {
   ELEVENLABS_MAX_MUSIC_LENGTH_MS,
   IMAGE_GENERATE_MODELS,
   DEFAULT_VIDEO_GENERATE_MODEL,
+  supportsLastFrame,
   MODEL_REGISTRY,
   supportsParam,
   nearestSupportedAspectRatio,
+  nearestSupportedImageSize,
   resolveConcurrency,
   ulid,
   isPersistedAssetRef,
@@ -8028,4 +8180,4 @@ export {
   defaultRegistry,
   createEngineFromEnv
 };
-//# sourceMappingURL=chunk-P2T3IZRE.js.map
+//# sourceMappingURL=chunk-UZ37VVP4.js.map