@koda-sl/baker-cli 0.262.0-dev.c998276fe → 0.263.0-dev.1f1c09c80

package/README.md

@@ -5796,6 +5796,7 @@ This CLI is designed for AI agent consumption. Key patterns:
 - **0.252.0**: avatar voices removed. `--voice-id` / `--voice-description` are gone from `baker avatars create|update`, and an avatar no longer holds a voice at all — every clip is voiced by the video model as it renders, directed by the profile's `--accent`. The picker they replaced offered a voice, played its preview, and then shipped a clip that did not sound like it: the re-voice was speech-to-speech, which replaces timbre and keeps pronunciation, so the accent the user picked could never survive. Three other routes were built and judged on video before removing it — TTS + lip-sync (mouth), audio-driven avatar models (picture), and Seedance 2.5 driven by audio (refuses AI faces). Reasoning and the re-open condition: `docs/adr/0005-an-avatar-is-a-face-not-a-voice.md`.
 - **0.253.0**: the pre-render gate runs `hyperframes check` — lint, runtime, layout, motion and contrast in one browser session, sampled at transition seams — instead of the static `lint` + `inspect` pair, falling back to them when the installed binary predates `check`. The browser half is signal the parser could never produce: measured on HyperFrames' own `warm-grain` example, lint/runtime/layout/motion all reported `ok: true` while the composition rendered an almost entirely blank ten-second video, the only trace being six `GSAP target #a-roll not found` warnings — a tween pointed at an element that does not exist, so nothing it animates ever appears. Baker promotes that warning to blocking, alongside `sweep_static` ("the timeline never advanced", i.e. a still image billed as a video). Warnings are grouped by code with a count (one run returned 62, 53 of them identical) and the highest-value findings carry advice on what to do rather than only what was seen. Requires `hyperframes@0.8.17` in the sandbox, bumped from 0.7.5 in the same change. Reasoning: `docs/adr/0007-the-render-gate-runs-in-a-browser.md`.
 - **0.254.0**: the nested-composition smoke test now looks at the frames instead of trusting the exit code — exiting 0 is what a blank render does too. `hyperframes snapshot --describe` sends the captured frames to Gemini and writes back what is actually in them, and the credential Studio already uses for Gemini and Omni (`GOOGLE_GENERATIVE_AI_API_KEY`) is the `GEMINI_API_KEY` it reads; the backend now ships it to the sandbox by omission, so a deployment without one keeps the plain smoke test. Verified against a real render: the pass reported "the image is completely empty" and named the single visible element without being told what to look for. Frame verdicts are condensed into the node log, one line per frame — a file the agent would have to know to open is a file it does not read.
+- **0.263.0**: shot briefs refuse the frame that contradicts its own caption, and the vision pass reads the numbers back. Observed verbatim: a phone screen reading "$215.40 SAVED THIS MONTH" under a caption reading "Save over £900 a year" — wrong currency, and a figure that contradicts the claim it sits beneath, since 215.40 a month is 2,584 a year. The model invents a number whenever a screen or document is in shot, so the frames no longer carry legible text at all: the caption track is where an ad's numbers live. Same brief also pins each object's real-world FORM after a phone rendered with a second screen on its back, and names the market's currency for anything unavoidably visible. The frame-vision question now reads out every number and word inside the picture and says whether it contradicts the caption — the defect only exists between the two, so neither alone can see it.
 - **0.262.0**: `scaffold-ad` carries the brand. The spec takes `brand.palette` (hex tokens) and `brand.logo` (repo path to the mark), and emits them as the `style.palette` and `branding` the engine already reads to colour its brand plate and CTA. The palette dresses the plate and the type and is deliberately kept OUT of the shot briefs — hex in a frame prompt makes the model paint the whole scene that colour, which reads as a filter. Also takes `market`, since an ad that does not say where it is set is an ad the model sets wherever it likes.
 - **0.261.0**: four defects that reached a finished ad, fixed in `scaffold-ad` where they originate. Frames render on **`openai/gpt-image-2`**, pinned rather than inheriting the catalogue default (Gemini Flash — right for cheap high-volume work, wrong for the opening still a five-second clip is built from). Every shot brief now states the MARKET, inferred from the voice language when unstated: an ad for Spain came back full of British houses because nothing said where it was. Every shot brief also rules out the impossible — paper and screens are opaque, objects are at real-world scale — after a bill rendered with the text from its reverse showing through, and a solar panel five metres tall. And beat durations snap to lengths the video model can actually render: free-form durations were rounded down per clip, so 11 beats summed to 27.0s of picture against a 31.4s timeline and the 4.4s difference rendered as a black tail. The read rate was re-measured off a real Spanish ad (81 words in 26.48s = 3.06 w/s); the old 2.6 over-predicted every beat by 18%.
 - **0.260.0**: burned-in captions break where the script punctuates instead of every N words. The karaoke composition sliced the transcript into fixed groups, and a word count knows nothing about the sentence it is cutting — an ad the engine rendered showed cards reading "OTRA VEZ, EN" and "TU CASA GENERA", half-thoughts ending on a preposition. A card now closes when the clause does; the word count stays as a ceiling so a long clause is still split, and a linking word ("de", "al", "y") never ends a card because it belongs to what follows. Same rule the pre-render gate already checked for, now enforced where the cards are actually built.

package/dist/{chunk-R4MBS2OI.js → chunk-ZK5S7JV2.js}

@@ -1034,12 +1034,17 @@ function failureDetail(reason) {
   }
   return { message: describeFailureReason(reason) };
 }
+function retryVerdict(retryable) {
+  if (retryable === true) return "transient \u2014 retry the run";
+  if (retryable === false) return "not retryable \u2014 retrying will not help, fix the input";
+  return "retryability unknown";
+}
 function describeCause(c) {
   switch (c.kind) {
     case "local":
       return c.cause instanceof Error ? c.cause.message : String(c.cause);
     case "remote":
-      return `[${c.code}] ${c.message}${c.provider ? ` (${c.provider})` : ""}`;
+      return `[${c.code}] ${c.message}${c.provider ? ` (${c.provider})` : ""} \u2014 ${retryVerdict(c.retryable)}`;
     case "timeout":
       return `timeout${c.provider ? ` (${c.provider})` : ""}`;
     case "network":
@@ -7443,7 +7448,7 @@ function buildSnapshotArgs(dir, frames, vision) {
   return vision ? [
     ...base,
     "--describe",
-    "Answer three things about this frame. (1) Is it empty or near-empty? Name every element actually visible. (2) Could what is shown physically happen \u2014 are object sizes, human scale and actions plausible? Say so plainly if something is the wrong size or impossible. (3) Is any on-screen text legible against what is behind it?"
+    "Answer four things about this frame. (1) Is it empty or near-empty? Name every element actually visible. (2) Could what is shown physically happen \u2014 object sizes, human scale, and whether every object has its real-world form (a phone has ONE screen, on its front). Say so plainly if something is impossible. (3) Read out every number, currency symbol and word visible INSIDE the picture \u2014 phone screens, documents, signage \u2014 and say whether any of it contradicts the caption on this frame. (4) Is the caption legible against what is behind it?"
   ] : [...base, "--describe", "false"];
 }
 function usesNestedCompositions(indexHtml) {
@@ -9267,4 +9272,4 @@ export {
   defaultRegistry,
   createEngineFromEnv
 };
-//# sourceMappingURL=chunk-R4MBS2OI.js.map
+//# sourceMappingURL=chunk-ZK5S7JV2.js.map