@koda-sl/baker-cli 0.180.0 → 0.181.1

package/dist/cli.js

@@ -25,6 +25,7 @@ import {
   isPersistedAssetRef,
   looksLikeHttpUrl,
   nearestSupportedAspectRatio,
+  nearestSupportedImageSize,
   parseRefExpr,
   platformFormats,
   requireCredentialsFromEnv,
@@ -32,11 +33,12 @@ import {
   sha256Hex,
   spineInputFlags,
   spineInputOps,
+  supportsLastFrame,
   supportsParam,
   toModelSafeImage,
   ulid,
   validateCanvasDeep
-} from "./chunk-P2T3IZRE.js";
+} from "./chunk-RWHEFQXI.js";
 import {
   csvOrJson,
   daysAgoIso,
@@ -3412,6 +3414,7 @@ var imageGenerateModelSchema = z11.enum([
   // 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-3.1-flash-lite-image",
   "google/gemini-3-pro-image-preview",
   "recraft/recraft-v4.1-pro-vector"
 ]);
@@ -19770,6 +19773,20 @@ function videoResolutionParam(videoModel, resolution) {
   if (!VIDEO_MODELS_WITH_RESOLUTION.has(videoModel)) return {};
   return { resolution: resolution ?? DEFAULT_VIDEO_RESOLUTION };
 }
+var VIDEO_MODELS_WITH_DURATION = new Set(
+  Object.entries(MODEL_REGISTRY.video_generate).filter(([, spec]) => "duration" in spec.params).map(([id]) => id)
+);
+var VIDEO_MODELS_WITH_AUDIO_TOGGLE = new Set(
+  Object.entries(MODEL_REGISTRY.video_generate).filter(([, spec]) => "generate_audio" in spec.params).map(([id]) => id)
+);
+function videoDurationParam(videoModel, duration) {
+  if (!VIDEO_MODELS_WITH_DURATION.has(videoModel)) return {};
+  return { duration };
+}
+function videoAudioParam(videoModel, generateAudio) {
+  if (!VIDEO_MODELS_WITH_AUDIO_TOGGLE.has(videoModel)) return {};
+  return { generate_audio: generateAudio };
+}
 var WORDS_PER_SECOND = 2.5;
 function wordCount(text) {
   return text.trim().split(/\s+/).filter(Boolean).length;
@@ -20374,6 +20391,7 @@ function extendPresenceByPromptMentions(slots, blueprint) {
   });
 }
 var ACTOR_SHEET_MODEL = "google/gemini-3-pro-image-preview";
+var KEYFRAME_IMAGE_SIZE = "2K";
 var SHEET_SUBJECT_TYPE = {
   person: "person",
   animal: "character",
@@ -20563,8 +20581,11 @@ function buildFrameRef(edge, url, framePrompt, present2, ctx, nodes) {
   const genParams = {
     model: ctx.imageModel,
     // gpt-image-2 derives pixel dimensions from the ratio and has no size knob,
-    // so asking for 2K there is an `unknown_param` at validate.
-    ...supportsParam("image_generate", ctx.imageModel, "image_size") ? { image_size: "2K" } : {},
+    // so asking for 2K there is an `unknown_param` at validate. Models that DO take
+    // one still differ in how far up they go — Nano Banana 2 Lite renders 1K only —
+    // so snap to the best tier the chosen model actually offers rather than pinning
+    // a literal that a cheaper model would reject outright.
+    ...supportsParam("image_generate", ctx.imageModel, "image_size") ? { image_size: nearestSupportedImageSize("image_generate", ctx.imageModel, KEYFRAME_IMAGE_SIZE) } : {},
     // Per-model image defaults (gpt-image: quality=high — OpenRouter forwards it; we do
     // NOT send input_fidelity, which gpt-image-2 forces high automatically).
     ...imageProfile?.paramDefaults ?? {},
@@ -20854,11 +20875,11 @@ function emitSceneClip(i, scene, present2, mode, nativeTurn, ambientBroll, frame
       opts.uiRouted,
       opts.videoModel
     ),
-    duration: lengths.genDur,
+    ...videoDurationParam(opts.videoModel, lengths.genDur),
     ...videoResolutionParam(opts.videoModel, opts.resolution),
-    // Native talking scene → Seedance generates the spoken audio + lip-sync; an opt-in
+    // Native talking scene → the model generates the spoken audio + lip-sync; an opt-in
     // ambient b-roll beat generates diegetic ambient only; otherwise the clip is silent.
-    generate_audio: Boolean(nativeTurn) || ambientBroll,
+    ...videoAudioParam(opts.videoModel, Boolean(nativeTurn) || ambientBroll),
     // Intent-keyed "astonishing default" overrides — merged LAST so a hero/reveal
     // beat can claim the 1080p ceiling over the aspect-derived resolution.
     ...clipParamRecipe(profile, clipIntentOf(scene, i))
@@ -20867,7 +20888,13 @@ function emitSceneClip(i, scene, present2, mode, nativeTurn, ambientBroll, frame
   nodes.push({
     id: `s${i}${tag}_clip`,
     type: "video_generate",
-    inputs: { first_frame: frames.first, ...frames.last ? { last_frame: frames.last } : {} },
+    // An end frame can exist purely as the NEXT scene's splice frame, so having one
+    // is not permission to wire it — a model that takes a single conditioning image
+    // silently drops it.
+    inputs: {
+      first_frame: frames.first,
+      ...frames.last && supportsLastFrame(opts.videoModel) ? { last_frame: frames.last } : {}
+    },
     params: clipParams
   });
   const base = `$ref:s${i}${tag}_clip.video`;
@@ -21010,7 +21037,7 @@ function buildCompositeScene(layout, regions, comp, scene, i, present2, mode, na
     const startPrompt = region.frame_prompt ?? scene.start_frame_prompt;
     const endPrompt = region.frame_prompt ?? scene.end_frame_prompt;
     const first = buildFrameRef("start", void 0, startPrompt, regionSlots, ctx, nodes);
-    const last = buildFrameRef("end", void 0, endPrompt, regionSlots, ctx, nodes);
+    const last = supportsLastFrame(opts.videoModel) ? buildFrameRef("end", void 0, endPrompt, regionSlots, ctx, nodes) : void 0;
     const regionNative = isPresenter ? nativeTurn : void 0;
     const regionScene = {
       ...scene,
@@ -21578,14 +21605,16 @@ function emitPhraseClip(phrase, voiceNode, env, nodes, out) {
   );
   const lastShown = phrase.shownScenes[phrase.shownScenes.length - 1] ?? anchor;
   const lastScene = env.blueprint.scenes[lastShown] ?? anchorScene;
-  const last = buildFrameRef(
+  const modelTakesLastFrame = supportsLastFrame(env.opts.videoModel);
+  const nextSplicesFromEnd = Boolean(env.blueprint.scenes[lastShown + 1]?.continues_previous);
+  const last = modelTakesLastFrame || nextSplicesFromEnd ? buildFrameRef(
     "end",
     lastScene.end_frame_asset?.url,
     lastScene.end_frame_prompt,
     slotsForFrame(env.slots, lastShown, "end"),
     ctx,
     nodes
-  );
+  ) : void 0;
   const clipStart = phrase.shownScenes.reduce(
     (m, s) => Math.min(m, env.blueprint.scenes[s]?.start_s ?? phrase.start_s),
     phrase.start_s
@@ -21608,19 +21637,19 @@ function emitPhraseClip(phrase, voiceNode, env, nodes, out) {
       env.uiRouted.has(anchor),
       env.opts.videoModel
     ),
-    duration: genDur,
+    ...videoDurationParam(env.opts.videoModel, genDur),
     ...videoResolutionParam(env.opts.videoModel, env.opts.resolution),
-    generate_audio: true,
+    ...videoAudioParam(env.opts.videoModel, true),
     ...clipParamRecipe(clipProfile, clipIntentOf(anchorScene, anchor))
   };
   clipParams.aspect_ratio = env.genAr;
   nodes.push({
     id: `s${anchor}_clip`,
     type: "video_generate",
-    inputs: { first_frame: first, last_frame: last },
+    inputs: { first_frame: first, ...last && modelTakesLastFrame ? { last_frame: last } : {} },
     params: clipParams
   });
-  out.phraseEnd = { ref: last, scene: lastShown };
+  out.phraseEnd = last ? { ref: last, scene: lastShown } : void 0;
   const clipRef = `$ref:s${anchor}_clip.video`;
   const speechOffset = Math.max(0, phrase.start_s - clipStart);
   const extractLen = Math.min(Math.max(0.5, phrase.end_s - phrase.start_s), Math.max(0.5, genDur - speechOffset));
@@ -21828,14 +21857,15 @@ function brollKeyframes(scene, i, env, ctx, lengths, prevEndFrame, nodes) {
     nodes
   );
   const nextContinues = Boolean(env.blueprint.scenes[i + 1]?.continues_previous);
-  const last = lengths.trimTarget <= 4 && !nextContinues ? void 0 : buildFrameRef(
+  const guidesMotion = supportsLastFrame(env.opts.videoModel) && lengths.trimTarget > 4;
+  const last = guidesMotion || nextContinues ? buildFrameRef(
     "end",
     scene.end_frame_asset?.url,
     scene.end_frame_prompt,
     slotsForFrame(env.slots, i, "end"),
     ctx,
     nodes
-  );
+  ) : void 0;
   return { first, last, sharesPrevFrame };
 }
 function emitBrollScene(scene, i, isLast, env, nodes, out, prevEndFrame) {
@@ -22685,7 +22715,7 @@ var VIDEO_GUIDE = [
   "1. Edit each frame's prompt IN PLACE. Every `s<i>_start` keyframe node has its OWN self-contained `params.prompt` (the FRAME DESCRIPTION) \u2014 editing one changes only that frame. Rewrite the cast, product, claims, palette into the ad you want. The frame is RECAST to the el_* reference images you drop (the source ad's people are never reused), so describe pose/action/framing here and let the references carry identity.",
   "1b. STORYBOARD FIRST \u2014 align the look on the cheap stills before clips bill. Each scene's keyframe IS your storyboard; `metadata.video.motion_board` lays out each scene's frame, time window, spoken line, and the graphics scheduled in it. Lock the keyframes + check each graphic lands on its spoken beat, THEN run the clips (images are cheap, videos aren't; the cache re-bills only what you change). See references/video-flow.md.",
   "2. Drop ONE real source image at each `el_*` ingest `[TODO]` path. Each recurring element (person/product/logo) is reused across every frame it appears in, so the same identity stays consistent. `baker canvas run` REFUSES to start until every `[TODO]` slot holds a real source \u2014 so this is mandatory, not optional.",
-  "3. Talking heads are NATIVE: a scene with one on-camera speaker is voiced by Seedance itself (the line is in the clip's prompt + `generate_audio`), so lips and voice are generated together \u2014 no separate tts, no post-hoc lip-sync. Edit the line in the scene's `s<i>_clip` prompt to re-author the words TRUE for your brand. Off-camera narration scenes still use a sequenced `tts` per turn.",
+  "3. Talking heads are NATIVE: a scene with one on-camera speaker is voiced by the video model itself (the line goes in the clip's prompt; models with a `generate_audio` toggle get it set, and the default model always renders audio), so lips and voice are generated together \u2014 no separate tts, no post-hoc lip-sync. Edit the line in the scene's `s<i>_clip` prompt to re-author the words TRUE for your brand. Off-camera narration scenes still use a sequenced `tts` per turn.",
   "4. Voice consistency: every native talking clip's audio is re-voiced to ONE brand voice via `audio_voice_convert` (timing preserved \u2192 lips stay matched). Confirm the `voice_select` casting (one per speaker) \u2014 its `voice_id` is the brand voice; set its gender/language so the voice matches the creator.",
   "5. Overlays are REAL HTML you paint. Open `video-overlay-composition/index.html`: the reference's overlays are seeded inside `#overlay-root` as plain elements (text + a `.pos-*` class + `data-start`/`data-dur`). Restyle the CSS freely \u2014 build lower-thirds, a ticker, whatever the look needs \u2014 and replace a logo placeholder with a real `<img>` you source (`baker images icon/sticker/gif/logo`) and drop in that dir. The runtime only shows/hides by timestamp; it makes no styling decisions. Drop `brand-bold.otf` / `brand-regular.otf` there for on-brand type.",
   "6. `baker canvas validate` (proves native-audio + timing for free) then `baker canvas run` (generates many billed image/video/audio assets \u2014 not free).",
@@ -22762,9 +22792,9 @@ function buildVideoTodo(report, overlayCount, floatingCount, opts, blueprint) {
       voice_description: d.voice_description,
       line: d.line
     })),
-    talking_head_note: "SHOT-NATIVE: a presenter shot is ONE Seedance clip (its line quoted in s<anchor>_clip's prompt + generate_audio) so lips+voice are generated together \u2014 no tts, no veed-lipsync. Two adjacent presenter shots at a hard cut are SEPARATE clips; a cutaway phrase (the presenter on camera, cut to b-roll, back on camera) stays one clip whose on-camera windows are cut inside the spine's per-input chains (no extra nodes). Edit the line in the s<anchor>_clip prompt to re-author it. A pure-voiceover phrase (speaker never shown) is one ElevenLabs tts read instead.",
+    talking_head_note: "SHOT-NATIVE: a presenter shot is ONE video clip (its line quoted in s<anchor>_clip's prompt, plus generate_audio on the models that expose it) so lips+voice are generated together \u2014 no tts, no veed-lipsync. Two adjacent presenter shots at a hard cut are SEPARATE clips; a cutaway phrase (the presenter on camera, cut to b-roll, back on camera) stays one clip whose on-camera windows are cut inside the spine's per-input chains (no extra nodes). Edit the line in the s<anchor>_clip prompt to re-author it. A pure-voiceover phrase (speaker never shown) is one ElevenLabs tts read instead.",
     voice_note: "ONE voice per person: a single voice_select is reused across all that person's shots (on-camera AND off \u2014 the deconstruct's `voiceover` label folds into the sole presenter). Every presenter clip's native audio is extracted and re-voiced to that brand voice through a SINGLE merged audio_voice_convert per speaker (<voice>_conv, eleven_multilingual_sts_v2, timing preserved so lips stay matched) \u2014 so timbre stays consistent across the separate shot clips. Set voice_select.voice_id's gender/language to match the creator.",
-    native_timing: "Clips separate at COMPLETE BREAKS between shots, but the VOICE stays continuous where it should: a voiceover narration is ONE read across the b-roll it plays over, and a cutaway leaves the presenter's read continuous under the insert. Each clip is generated long enough for its estimated speech. `metadata.video.talking_scenes` carries each shot's scene_s vs est_speech_s. CAVEAT: a b-roll cutaway INSIDE a phrase lands at an approximate (proportional) time \u2014 Seedance exposes no word timing \u2014 so if a cutaway is off its beat, nudge the scene boundary (it's a starting point). NOTE: a native clip's voice is extracted for the SPOKEN window only (`s<i>_voextract`'s `-t`) \u2014 when the scene runs longer than the line, the picture tail is deliberately silent (extracting the full clip would put post-line breathing/babble on the voice bus); trim the scene or extend the line if the dead air reads wrong.",
+    native_timing: "Clips separate at COMPLETE BREAKS between shots, but the VOICE stays continuous where it should: a voiceover narration is ONE read across the b-roll it plays over, and a cutaway leaves the presenter's read continuous under the insert. Each clip is generated long enough for its estimated speech. `metadata.video.talking_scenes` carries each shot's scene_s vs est_speech_s. CAVEAT: a b-roll cutaway INSIDE a phrase lands at an approximate (proportional) time \u2014 no video model exposes word timing \u2014 so if a cutaway is off its beat, nudge the scene boundary (it's a starting point). NOTE: a native clip's voice is extracted for the SPOKEN window only (`s<i>_voextract`'s `-t`) \u2014 when the scene runs longer than the line, the picture tail is deliberately silent (extracting the full clip would put post-line breathing/babble on the voice bus); trim the scene or extend the line if the dead air reads wrong.",
     craft: {
       note: "Production-craft principles that raise every clip's realism. Full rationale: references/video-craft.md (production craft); references/script-craft.md + meta-ads-playbook for the hook/message layer.",
       principles: [
@@ -24967,6 +24997,7 @@ var SEEDANCE = "bytedance/seedance-2.0";
 var VEO = "google/veo-3.1";
 var VEO_FAST = "google/veo-3.1-fast";
 var KLING = "kwaivgi/kling-v3.0-pro";
+var GEMINI_OMNI = "google/gemini-omni-flash";
 function routeVideoModel(input) {
   const budget = input.budget ?? "standard";
   const signals = [];
@@ -24979,8 +25010,8 @@ function routeVideoModel(input) {
   if (input.needsIdentity && !input.hasRealFace) {
     signals.push({ model: SEEDANCE, weight: 300, because: "identity/product consistency \u2192 Seedance workhorse" });
   }
-  signals.push({ model: SEEDANCE, weight: 100, because: "default identity/product workhorse" });
-  const scores = { [SEEDANCE]: 0, [VEO]: 0, [VEO_FAST]: 0, [KLING]: 0 };
+  signals.push({ model: GEMINI_OMNI, weight: 100, because: "default \u2014 native audio in one call" });
+  const scores = { [SEEDANCE]: 0, [VEO]: 0, [VEO_FAST]: 0, [KLING]: 0, [GEMINI_OMNI]: 0 };
   let top = signals[0];
   for (const s of signals) {
     scores[s.model] = (scores[s.model] ?? 0) + s.weight;
@@ -35116,8 +35147,11 @@ Then stage changes (nothing is sent to Tag Manager until publish):
   baker tag-manager tag create --json '{"name":"GA4 Quote","type":"gaawe","firingTriggerId":["gtm_temp_trigger_..."]}'
   baker tag-manager draft list                 \u2014 review everything staged
 
-On publish, staged changes are written into a new Tag Manager version that is NOT published.
-Tell the user to publish that version in Tag Manager to make it live.
+On publish, staged changes are written into the container and closed as a new Tag Manager
+version, and a clean run PUBLISHES that version \u2014 so the changes go live on the user's site.
+Say they go live when the user publishes the chat; never that they are live already, and never
+that someone still has to approve them inside Tag Manager. (If the publish call itself fails the
+version stands unpublished, and \`draft list\` says so and what is left to do.)
 
 This owns the CONTENTS of the container. Installing the container snippet on the site is a
 different job \u2014 that is \`baker tags\` with the googleTagManager tag.