@vanillaskyai/video 0.10.11 → 0.10.13

package/CHANGELOG.md

@@ -4,6 +4,14 @@ VanillaSky follows semantic versioning. This changelog begins with the 0.1 beta.
 
 ## Unreleased
 
+## 0.10.13
+
+- Plan concise AI answers around the configured video-attempt budget, including the ending, while retaining complete authored narration and chapter recovery.
+
+## 0.10.12
+
+- Use query context to choose between equally relevant Pexels subject matches in the starter, while preserving subject, activity and equipment priority.
+
 ## 0.10.11
 
 - Keep the opening chapter visible until the first body visual and narration are ready, without mounting a second player or replaying the opening.

package/dist/server.js

@@ -1177,7 +1177,8 @@ function createVideoChatResponseInstructions(generatedVideoAvailable, openingAlr
     "Write a complete, intentful video answer as newline-delimited JSON. Match the user's form and tone; mixed intents can combine directions.",
     `First write one brief: {"type":"answer","intent":"explanation|story|comedy|imagination|practical","opening":"a short inviting spoken introduction of 6\u20139 words","subject":"literal visual subject","development":"the essential development of this answer","visualDirection":"consistent subjects, appearance and visual approach","ending":{"title":"short meaningful chapter title, at most 65 characters","narration":"the authored payoff","subject":"literal subject","action":"visible action or change","durationSec":${clipDurationSec},"continuity":"cut|continue"}}.`,
     `Then stream each developing shot on its own line: {"type":"shot","title":"short meaningful chapter title, at most 65 characters","narration":"the exact spoken beat","subject":"2\u20138 literal filmable words, at most 80 characters","action":"concrete subject, action or visible change and useful framing","durationSec":${clipDurationSec},"continuity":"cut|continue"}.`,
-    `The selected footage mode is ${mode === "pexels" ? "Pexels stock search: use literal filmable subjects; never imply stock proves a mechanism or depicts fictional events exactly" : `AI video, with at most ${generatedVideoAvailable ? maxGeneratedVideos : 0} generation attempts`}. Missing footage becomes the authored chapter title, with complete narration. Preserve the full answer rather than shortening it to fit credits. The host selects providers; do not make source choices.`,
+    `The selected footage mode is ${mode === "pexels" ? "Pexels stock search: use literal filmable subjects; never imply stock proves a mechanism or depicts fictional events exactly" : `AI video, with at most ${generatedVideoAvailable ? maxGeneratedVideos : 0} generation attempts`}. Missing footage becomes the authored chapter title, with complete narration. Never truncate already-authored narration when footage fails. The host selects providers; do not make source choices.`,
+    ...mode === "cinematic" ? [generatedVideoAvailable && maxGeneratedVideos > 0 ? `Plan at most ${maxGeneratedVideos} generated-video beats in total, including the saved ending. Use at most ${maxGeneratedVideos - 1} developing shot records; the ending uses the remaining beat. The opening chapter does not consume a generated clip. Before writing, choose a concise, complete treatment that fits this budget: combine related ideas, preserve essential facts and qualifiers, and finish the requested answer. Do not plan an extra chapter tail simply because generation attempts will run out.${maxGeneratedVideos === 1 ? " Put the complete answer in the saved ending, set development to an empty string, and emit no developing shot records." : ""}` : "No generated-video attempts are available; plan a complete chapter-led answer with a useful ending. Do not omit the answer to satisfy a zero clip budget."] : [],
     ...mode === "pexels" ? ["Stock queries must retain the essential subject, activity and distinguishing equipment in the shot's subject field, within its word limit. That field alone is the search query; action and visualDirection do not refine it. Prefer common observable actions with usable framing. Do not replace the required actor or activity with scenery, a different sport or a loosely related setting. Preserve fictional or comic narration, but do not depend on stock showing an exact invented expression or sequence; choose an illustrative action that supports the beat."] : [],
     ...mode === "pexels" ? ['Include stockSelection on every shot and the saved ending when the essential subject is known: "stockSelection":{"subject":"essential actor or object category","activity":"optional literal activity","equipment":"optional distinguishing equipment","exclude":["optional contradictory subject or activity"]}. Each phrase must be 1\u20134 words and at most 48 characters; exclude has at most 3 phrases. The essential subject is separate from the setting: do not use scenery, mood, camera framing or incidental appearance as the actor. Keep the search query broad enough to find footage; the optional hint helps select results without substituting a different actor or task. Use exclusions only for actual contradictions, not every detail absent from the story. Omit unknown fields or the whole hint rather than inventing an anchor. This is selection guidance, not verification that footage depicts the exact narration.'] : [],
     "For a very short answer whose ending alone fulfills the request, development may be empty and no developing shots are needed. Otherwise, develop the essential content before the ending.",

package/docs/reference/provider-adapters.md

@@ -96,7 +96,10 @@ stream-reconnect contract.
 
 Set `maxGeneratedVideos` on `createVideoChatHandler` to a nonnegative safe integer
 (default `5`). This is a per-response generation attempt limit, including
-failures. AI mode uses authored chapter recovery after the allowance is reached;
+failures. The default AI planner is instructed to fit a concise, complete answer
+within that many visual beats, including the ending. This is model guidance, not
+a guaranteed shot count: already-authored content is retained if the model
+exceeds the budget. AI mode uses authored chapter recovery after the allowance is reached;
 `0` skips all generated footage. Pexels mode searches stock independently and
 never consumes the generated-video allowance. Retries inside your
 provider callback can incur additional charges; bound those separately. Never

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@vanillaskyai/video",
-  "version": "0.10.11",
+  "version": "0.10.13",
   "description": "Open-source voice-and-video chat SDK for AI applications.",
   "keywords": [
     "video-chat",

package/starters/video-chat/package.json

@@ -9,7 +9,7 @@
     "preview": "vite preview"
   },
   "dependencies": {
-    "@vanillaskyai/video": "0.10.11",
+    "@vanillaskyai/video": "0.10.13",
     "react": "^19.2.8",
     "react-dom": "^19.2.8",
     "@ai-sdk/anthropic": "^3.0.0",

package/starters/video-chat/stock.ts

@@ -71,7 +71,9 @@ export async function findStockFootage(query: string, orientation: VideoOrientat
     if (selection && subject.length) {
       const covers = (phrase: string) => terms(phrase).every(word => subject.includes(word));
       if (!covers(selection.subject) || selection.exclude?.some(covers)) continue;
-      matches = 2 + Number(Boolean(selection.activity && covers(selection.activity)))
+      // Query context breaks equal hint matches without outweighing a hint.
+      const contextScore = matches / (tokens.length + 1);
+      matches = 2 + contextScore + Number(Boolean(selection.activity && covers(selection.activity)))
         + Number(Boolean(selection.equipment && covers(selection.equipment)));
     }
     // The documented Video resource can have only a numeric page URL and no