@koda-sl/baker-cli 0.282.0-dev.37d65c47b → 0.282.0-dev.555a170a5

package/dist/cli.js

@@ -19,6 +19,7 @@ import {
   VIDEO_GENERATE_MODELS,
   ValidationError,
   YtDlpError,
+  bestSupportedQuality,
   canonicalUrl,
   clipParamRecipe,
   clipProfileFor,
@@ -59,7 +60,7 @@ import {
   ulid,
   validateCanvasDeep,
   ytDlpBlockSignal
-} from "./chunk-6KTKYFT4.js";
+} from "./chunk-LOUE7GTU.js";
 import {
   csvOrJson,
   daysAgoIso,
@@ -6195,8 +6196,9 @@ var imagesPinterestRequestSchema = z18.object({
 });
 var imagesPinterestResponseSchema = providerHitsResponseSchema();
 var imageGenerateModelSchema = z18.enum([
+  "openai/gpt-image-2.5-sunburst",
+  // Legacy — see the registry entries; kept so pre-switch canvases still run.
   "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-3.1-flash-lite-image",
@@ -6555,7 +6557,7 @@ var studioGenerateRequestSchema = z21.object({
   useBrand: z21.boolean().optional(),
   // Per-model knobs; the backend refuses one the chosen model does not advertise.
   imageSize: z21.string().optional(),
-  quality: z21.enum(["auto", "low", "medium", "high"]).optional(),
+  quality: z21.enum(["auto", "low", "medium", "high", "xhigh", "max"]).optional(),
   strength: z21.coerce.number().min(0).max(1).optional(),
   rgbColors: z21.array(rgbTriple).optional(),
   backgroundRgbColor: rgbTriple.optional()
@@ -29842,9 +29844,17 @@ function thinProfileHint(profile, handle) {
   if (missing.length === 0) return null;
   return `MISSING ${missing.map((field) => `--${field}`).join(" ")}. A bare subject description holds the face still and nothing else, so clips of @${handle} come out stiff and generic \u2014 measured, filling these in nearly doubled the movement in the same brief. --accent is the one nobody guesses and the only one that fails silently: it decides the accent the avatar SPEAKS with in every clip, so name the region the user is writing to you from ("Spanish from Spain, Madrid accent") \u2014 "Spanish" on its own renders as no accent at all. The user almost certainly told you how this person talks and carries themselves; put it in: \`baker avatars update ${handle} ${missing.map((field) => `--${field} "\u2026"`).join(" ")}\`.`;
 }
+function parseFaceChoice(raw) {
+  if (raw === void 0) return "ask";
+  const value = raw.trim().toLowerCase();
+  if (value === "ask" || value === "auto") return value;
+  return null;
+}
 function creationHints(created, profile) {
   const thin = thinProfileHint(profile, created.handle);
-  const casting = created.likeness === "synthetic" ? [
+  const casting = created.likeness === "synthetic" && created.faceChoice === "auto" ? [
+    `No one was asked about this face \u2014 \`--face auto\` cast it and it is kept. Do not tell the user they chose it, and do not offer them a pick you are not going to run. If this should have been theirs, say so and create it again with \`--face ask\`.`
+  ] : created.likeness === "synthetic" ? [
     `The user chooses this face, and it only happens if you ask: call \`request_avatar_casting { avatarHandle: "${created.handle}" }\` now. It shows them real portraits and blocks until they pick \u2014 never stand in a question of your own, which can only offer them words. The sheet building right now is fused from a face nobody has chosen.`
   ] : [];
   const hints2 = [
@@ -30051,7 +30061,12 @@ registerSchema({
       required: false
     },
     "voice-id": { type: "string", description: "Voice id to speak with in video", required: false },
-    "voice-description": { type: "string", description: "How the voice should sound", required: false }
+    "voice-description": { type: "string", description: "How the voice should sound", required: false },
+    face: {
+      type: "string",
+      description: "Who picks the face, for an invented avatar: 'ask' (default) blocks on the user through `request_avatar_casting`, 'auto' keeps the one this build casts and asks nobody. Use 'auto' only when there is no one to answer \u2014 a Schedule, a batch, an explicit \"do it yourself\" \u2014 because a person who is there almost always wants this choice.",
+      required: false
+    }
   }
 });
 var createCommand2 = defineCommand96({
@@ -30086,7 +30101,12 @@ var createCommand2 = defineCommand96({
       default: false
     },
     "voice-id": { type: "string", description: "Voice id", required: false },
-    "voice-description": { type: "string", description: "How the voice should sound", required: false }
+    "voice-description": { type: "string", description: "How the voice should sound", required: false },
+    face: {
+      type: "string",
+      description: "Who picks the face of an invented avatar: 'ask' (default, blocks on the user) or 'auto' (asks nobody)",
+      required: false
+    }
   },
   run: async ({ args }) => {
     const name = args.name?.trim();
@@ -30110,6 +30130,16 @@ var createCommand2 = defineCommand96({
         `--handle "${handle}" is not usable. Use lowercase letters, numbers and dashes, starting with a letter \u2014 "maria-founder", not "Maria Founder". Leave it off to have one derived from the name.`
       );
     }
+    const faceChoice = parseFaceChoice(args.face);
+    if (faceChoice === null) {
+      failValidation2(
+        `--face "${String(args.face)}" is not a mode. Use --face ask to let the user pick the face, or --face auto to keep the one this build casts.`,
+        {
+          action: "pick_a_mode",
+          explanation: "Drop the flag entirely if a person is in the conversation \u2014 that is --face ask, and it is what they almost always want. Pass --face auto only when nobody is there to answer: a Schedule, a batch run, or the user telling you to do it without them."
+        }
+      );
+    }
     const sourceImageIds = parseImageIds(args["source-image"]);
     const likenessConfirmed = Boolean(args["likeness-confirmed"]);
     if (sourceImageIds && !likenessConfirmed) {
@@ -30132,7 +30162,7 @@ var createCommand2 = defineCommand96({
         ...chatIdFromEnv() ? { chatId: chatIdFromEnv() } : {}
       };
       const data = await writeAvatars("/api/avatars", body);
-      writeJson({ ok: true, data, hints: creationHints(data, body.profile) });
+      writeJson({ ok: true, data, hints: creationHints({ ...data, faceChoice }, body.profile) });
     } catch (err) {
       failAvatarApi(err, handle);
     }
@@ -32867,6 +32897,7 @@ function extendPresenceByPromptMentions(slots, blueprint) {
 }
 var ACTOR_SHEET_MODEL = "google/gemini-3-pro-image-preview";
 var KEYFRAME_IMAGE_SIZE = "2K";
+var FRAME_QUALITY = "xhigh";
 var SHEET_SUBJECT_TYPE = {
   person: "person",
   animal: "character",
@@ -33057,6 +33088,7 @@ function buildFrameRef(edge, url, framePrompt, present2, ctx, nodes) {
   const presentOrdered = orderSlotsByIdentity(present2);
   const reference = [...presentOrdered.map((s) => s.ref), ...originalRef ? [originalRef] : []];
   const imageProfile = imageProfileFor(ctx.imageModel);
+  const quality = bestSupportedQuality("image_generate", ctx.imageModel, FRAME_QUALITY);
   const genParams = {
     model: ctx.imageModel,
     // gpt-image-2 derives pixel dimensions from the ratio and has no size knob,
@@ -33065,9 +33097,14 @@ function buildFrameRef(edge, url, framePrompt, present2, ctx, nodes) {
     // 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).
+    // Per-model image defaults (we do NOT send input_fidelity, which gpt-image forces
+    // high automatically).
     ...imageProfile?.paramDefaults ?? {},
+    // The frame IS the clip's first moment, and the video model animates whatever it is
+    // handed — a soft frame is a soft clip for its whole length, and no amount of motion
+    // direction recovers it. Worth the step the older models cannot reach; they fall back
+    // to `high` on their own rather than failing validate.
+    ...quality ? { quality } : {},
     prompt: buildFramePrompt(
       edge,
       ctx.sceneIndex,
@@ -37432,6 +37469,12 @@ function scaffoldStaticAd(input, elementsInput, opts) {
       ...supportsParam("image_generate", opts.genModel, "image_size") ? { image_size: "2K" } : {},
       // Per-model image defaults (gpt-image: quality=high).
       ...imageProfileFor(opts.genModel)?.paramDefaults ?? {},
+      // This one render IS the ad — it gets published at full size and read by a person,
+      // not consumed by another model. One image, so the best step the model has.
+      ...(() => {
+        const quality = bestSupportedQuality("image_generate", opts.genModel, "xhigh");
+        return quality ? { quality } : {};
+      })(),
       prompt
     }
   });
@@ -38808,7 +38851,7 @@ function resolveShippedCanvasDir(name, startDir, exists = existsSync5, maxDepth
 }
 
 // src/commands/canvas/scaffold-ad.ts
-var AD_IMAGE_MODEL = "openai/gpt-image-2";
+var AD_IMAGE_MODEL = "openai/gpt-image-2.5-sunburst";
 var SHIPPED_COMPOSITION_DIR = resolveShippedCanvasDir("video-overlay-composition", import.meta.dirname);
 var SHIPPED_CAPTIONS_DIR = resolveShippedCanvasDir("tiktok-captions-composition", import.meta.dirname);
 registerSchema({
@@ -39534,7 +39577,7 @@ function resolveModels2(args) {
     // Default to the strongest image model (matches the static-ad scaffold); the
     // frame generators need the most faithful text/identity reproduction. Override
     // with --image-model for a cheaper/faster pass.
-    imageModel: pick("image-model", "image_generate", "openai/gpt-image-2")
+    imageModel: pick("image-model", "image_generate", "openai/gpt-image-2.5-sunburst")
   };
 }
 function hasPhotorealCast(elements) {