ima2-gen 3.15.0 → 3.15.1

package/README.md

@@ -182,7 +182,7 @@ If no provider is specified, the app keeps the current GPT OAuth/default behavio
 
 Grok image generation exposes a model picker (`grok-imagine-image` / `grok-imagine-image-quality`) and a size picker (aspect ratio + 1k/2k resolution). The Settings page prefers the Grok Build weekly credits percentage and reset time from `GET /v1/billing?format=credits`; if that source is unavailable, it falls back to the legacy monthly billing window and `$used/$limit`. A **Switch Account** button starts a device-code OAuth flow (`POST /api/auth/switch`) for re-authenticating without leaving the app.
 
-Grok video generation defaults to canonical `grok-imagine-video-1.5`; `grok-imagine-video` remains available for base-model-only Ref2V, V2V edit, and extension paths, and the legacy `grok-imagine-video-1.5-preview` string is accepted as an alias. Three modes are auto-detected from reference count: text-to-video (0 refs), image-to-video (1 ref), and reference-to-video (2-7 refs, max 10s duration). 1080p is available for `grok-imagine-video-1.5` prompt-only text-to-video and single image/frame image-to-video; prompt-only 1.5 uses the internal white-canvas I2V shim before the upstream request. Video controls include duration (1-15s), resolution (480p, 720p, 1080p when supported), and aspect ratio (1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, auto).
+Grok video generation defaults to canonical `grok-imagine-video-1.5`; `grok-imagine-video` remains available for base-model-only Ref2V, V2V edit, and extension paths, and the legacy `grok-imagine-video-1.5-preview` string is accepted as an alias. Three modes are auto-detected from reference count: text-to-video (0 refs), image-to-video (1 ref), and reference-to-video (2-14 refs; up to 15s on grok-imagine-video-1.5, 10s on grok-imagine-video). 1080p is available for `grok-imagine-video-1.5` prompt-only text-to-video and single image/frame image-to-video; prompt-only 1.5 uses the internal white-canvas I2V shim before the upstream request. Video controls include duration (1-15s), resolution (480p, 720p, 1080p when supported), and aspect ratio (1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, auto).
 
 ![Settings workspace showing GPT OAuth active and API key provider available.](assets/screenshots/settings-oauth-generation.png)
 

package/bin/commands/video.js

@@ -5,7 +5,7 @@ import { out, die, color, exitCodeForError, fail } from "../lib/output.js";
 import { writeFile, mkdir, readFile, stat } from "node:fs/promises";
 import { basename, dirname } from "node:path";
 import { runVideoGenerate } from "../lib/videoMcp.js";
-import { GROK_VIDEO_MODEL_15, GROK_VIDEO_MODEL_15_PREVIEW_ALIAS, GROK_VIDEO_MODEL_BASE, validateVideoResolutionForRequest, } from "../../lib/imageModels.js";
+import { GROK_VIDEO_MODEL_15, GROK_VIDEO_MODEL_15_PREVIEW_ALIAS, GROK_VIDEO_MODEL_BASE, MAX_VIDEO_DURATION, MIN_VIDEO_DURATION, validateVideoResolutionForRequest, } from "../../lib/imageModels.js";
 import { VIDEO_CLIENT_TIMEOUT_SEC } from "../../lib/videoClientTimeouts.js";
 import { deriveVideoProviderIds } from "../../lib/providers/derive.js";
 import { listProviders } from "../../lib/mcp/providerRegistry.js";
@@ -291,7 +291,7 @@ async function videoExtendCmd(argv) {
     const args = parseArgs(argv, spec);
     rejectUnknownFlags(args);
     if (args.help) {
-        out(`  ima2 video extend <prompt> --video <url|file_id|generated-file> [--duration 6]\n\n  Extend video from its last frame.\n  Model: grok-imagine-video only. Extension: 2-10s.\n\n  Options:\n        --video <value>   Source video HTTPS URL, xAI file_id, data URL, or generated filename (required)\n        --duration <2-10> Extension duration (default: 6)\n    -o, --out <file>      Download extended video to file\n        --output <file>   Alias for --out\n        --json            Print JSON result\n        --timeout <sec>   Default: 5400\n        --server <url>    Override server URL`);
+        out(`  ima2 video extend <prompt> --video <url|file_id|generated-file> [--duration 6]\n\n  Extend video from its last frame.\n  Model: grok-imagine-video only (1.5 rejects extensions). Extension: 1-15s.\n  duration is the length of the ADDED segment, not the total output.\n\n  Options:\n        --video <value>   Source video HTTPS URL, xAI file_id, data URL, or generated filename (required)\n        --duration <1-15> Added-segment duration (default: 6)\n    -o, --out <file>      Download extended video to file\n        --output <file>   Alias for --out\n        --json            Print JSON result\n        --timeout <sec>   Default: 5400\n        --server <url>    Override server URL`);
         return;
     }
     const prompt = args.positional.join(" ");
@@ -300,8 +300,10 @@ async function videoExtendCmd(argv) {
     if (!args.video)
         die(2, "--video <url> is required");
     const duration = parseIntegerFlag(args.duration, 6, "--duration");
-    if (duration < 2 || duration > 10)
-        die(2, "--duration must be between 2 and 10");
+    // Measured 1-15, not the documented 2-10: the endpoint accepts 1 and 11 and refuses
+    // only 0 and 16. devlog/_plan/260908_xai_imagine_spec_resync/000_research.md (D11)
+    if (duration < MIN_VIDEO_DURATION || duration > MAX_VIDEO_DURATION)
+        die(2, `--duration must be between ${MIN_VIDEO_DURATION} and ${MAX_VIDEO_DURATION}`);
     parseTimeoutSeconds(args.timeout);
     const server = await resolveServer({ serverFlag: args.server });
     const res = await fetchServer(server.base, "/api/video/extend", { method: "POST", headers: { "Content-Type": "application/json" }, body: JSON.stringify({ prompt, videoUrl: args.video, duration }), signal: timeoutSignal(args.timeout) });

package/docs/API.md

@@ -520,7 +520,7 @@ Generate a video via the Grok video provider. Returns Server-Sent Events on the
 |---|---|---|
 | No images | text-to-video | 1–15s |
 | 1 image (`sourceImage` or `sourceFilename`) | image-to-video | 1–15s |
-| 2–7 images (`referenceImages` / `referenceFilenames`) | reference-to-video | 1–10s |
+| 2–14 images (`referenceImages` / `referenceFilenames`) | reference-to-video | 1–15s on grok-imagine-video-1.5, 1–10s on grok-imagine-video |
 
 1080p is accepted for `grok-imagine-video-1.5` prompt-only text-to-video and image-to-video with one image/frame source, including `continueFromVideo` after the server extracts the parent video's last frame. Prompt-only 1.5 text-to-video uses the internal white-canvas image-to-video shim before the upstream request. 1.5 does not add Ref2V, V2V edit, or extension support.
 

package/docs/API.zh-CN.md

@@ -386,7 +386,7 @@ Grok节点SSE回复不包括回复API `partial`图像事件是因为xAI图片API
 |---|---|---|
 |没有图片|文本转视频| 1–15s |
 |1 张图片（`sourceImage`或者`sourceFilename`) |图像到视频| 1–15s |
-|2–7 张图像 (`referenceImages` / `referenceFilenames`) |参考视频| 1–10s |
+|2–14 张图像 (`referenceImages` / `referenceFilenames`) |参考视频| 1.5: 1–15s / base: 1–10s |
 
 1080p 可接受`grok-imagine-video-1.5`仅提示文本到视频和图像到视频，具有一个图像/帧源，包括`continueFromVideo`服务器提取父视频的最后一帧后。仅提示 1.5 文本到视频在上游请求之前使用内部白色画布图像到视频 shim。 1.5 不添加 Ref2V、V2V 编辑或扩展支持。
 

package/docs/API.zh-TW.md

@@ -386,7 +386,7 @@ Grok節點SSE回覆不包括回覆API `partial`影像事件是因為xAI圖片API
 |---|---|---|
 |沒有圖片|文字轉視頻| 1–15s |
 |1 張圖片（`sourceImage`或者`sourceFilename`) |影像到視頻| 1–15s |
-|2–7 張圖像 (`referenceImages` / `referenceFilenames`) |參考影片| 1–10s |
+|2–14 張圖像 (`referenceImages` / `referenceFilenames`) |參考影片| 1.5: 1–15s / base: 1–10s |
 
 1080p 可接受`grok-imagine-video-1.5`僅提示文字到視頻和圖像到視頻，具有一個圖像/幀源，包括`continueFromVideo`伺服器提取父影片的最後一幀後。僅提示 1.5 文字到影片在上游請求之前使用內部白色畫布圖像到影片 shim。 1.5 不新增 Ref2V、V2V 編輯或擴充支援。
 

package/docs/CLI.md

@@ -273,7 +273,7 @@ Video mode is auto-detected from `--ref` count:
 |---|---|
 | 0 | text-to-video |
 | 1 | image-to-video |
-| 2–7 | reference-to-video (max 10s duration) |
+| 2–14 | reference-to-video (15s on grok-imagine-video-1.5, 10s on grok-imagine-video) |
 
 `grok-imagine-video-1.5` supports 1080p for prompt-only text-to-video and single image/frame image-to-video. Prompt-only 1.5 text-to-video is submitted through the internal white-canvas image-to-video shim because upstream 1.5 rejects raw T2V. The old `grok-imagine-video-1.5-preview` name is accepted as an alias and normalized before the upstream request. 1.5 does not support `reference_images` reference-to-video, V2V edit, or video extension. For 2+ refs, use `grok-imagine-video`; if ima2 auto-retries a 1.5 Ref2V request to the base model, read `video.effectiveModel` and `video.modelFallback` from CLI `--json`, or `effectiveModel` and `modelFallback` from SSE.
 

package/docs/CLI.zh-CN.md

@@ -201,7 +201,7 @@ and rain build, end on a still close-up after the line finishes`。
 |---|---|
 | 0 |文本转视频|
 | 1 |图像到视频|
-| 2–7 |参考视频（最长 10 秒持续时间）|
+| 2–14 |参考视频（grok-imagine-video-1.5 最长 15 秒，grok-imagine-video 最长 10 秒）|
 
 `grok-imagine-video-1.5`支持 1080p 仅提示文本到视频和单图像/帧图像到视频。仅提示 1.5 文本到视频通过内部白色画布图像到视频 shim 提交，因为上游 1.5 拒绝原始 T2V。旧的`grok-imagine-video-1.5-preview`name 在上游请求之前被接受为别名并进行规范化。 1.5 不支持`reference_images`视频参考、V2V 编辑或视频扩展。对于 2 个以上参考，请使用`grok-imagine-video`;如果ima2自动重试向基本模型发出 1.5 Ref2V 请求，读取`video.effectiveModel`和`video.modelFallback`从CLI `--json`， 或者`effectiveModel`和`modelFallback`从SSE.
 

package/docs/CLI.zh-TW.md

@@ -201,7 +201,7 @@ and rain build, end on a still close-up after the line finishes`。
 |---|---|
 | 0 |文字轉視頻|
 | 1 |影像到視頻|
-| 2–7 |參考影片（最長 10 秒持續時間）|
+| 2–14 |參考影片（grok-imagine-video-1.5 最長 15 秒，grok-imagine-video 最長 10 秒）|
 
 `grok-imagine-video-1.5`支援 1080p 僅提示文字到影片和單圖像/幀圖像到影片。僅提示 1.5 文字到影片透過內部白色畫布圖像到影片 shim 提交，因為上游 1.5 拒絕原始 T2V。舊的`grok-imagine-video-1.5-preview`name 在上游請求之前被接受為別名並進行規範化。 1.5 不支持`reference_images`影片參考、V2V 編輯或影片擴充。對於 2 個以上參考，請使用`grok-imagine-video`;如果ima2自動重試向基本模型發出 1.5 Ref2V 請求，讀取`video.effectiveModel`和`video.modelFallback`從CLI `--json`， 或者`effectiveModel`和`modelFallback`從SSE.
 

package/docs/README.ja.md

@@ -120,7 +120,7 @@ API key が env/config に存在する場合、生成エンドポイントで `p
 
 Settings に **API key provider available** と表示される場合、API key が検出され、生成・編集・multimode・node request に使用できるという意味です。
 
-Grok 動画の既定値は正式名 `grok-imagine-video-1.5` です。`grok-imagine-video` は Ref2V、V2V edit、extension の互換パスで引き続き使用し、従来の `grok-imagine-video-1.5-preview` も互換 alias として受け付けます。参照数に応じて T2V(0)、I2V(1)、Ref2V(2-7、最大10秒)が自動選択され、1080p は `grok-imagine-video-1.5` の prompt-only T2V と単一画像/フレーム I2V で有効です。duration(1-15s)、resolution(480p/720p/対応時 1080p)、aspect ratio を設定できます。
+Grok 動画の既定値は正式名 `grok-imagine-video-1.5` です。`grok-imagine-video` は Ref2V、V2V edit、extension の互換パスで引き続き使用し、従来の `grok-imagine-video-1.5-preview` も互換 alias として受け付けます。参照数に応じて T2V(0)、I2V(1)、Ref2V(2-14、grok-imagine-video-1.5 は最大15秒 / grok-imagine-video は最大10秒)が自動選択され、1080p は `grok-imagine-video-1.5` の prompt-only T2V と単一画像/フレーム I2V で有効です。duration(1-15s)、resolution(480p/720p/対応時 1080p)、aspect ratio を設定できます。
 
 設定画面の QuotaCard に Grok billing `$used/$limit` バーと **Switch Account** ボタン（`POST /api/auth/switch`）が表示されます。
 

package/docs/README.ko.md

@@ -117,7 +117,7 @@ Ctrl+C로 DB, 소켓, 자식 프로세스를 정리할 수 있습니다. Windows
 - **Video 생성**: 텍스트, 이미지, 또는 여러 레퍼런스에서 짧은 영상을 만듭니다. SSE로 기획→제출→진행률→완료를 실시간 표시합니다. 생성된 영상에서 First/Mid/Last 프레임 복사 버튼으로 키프레임을 추출할 수 있습니다.
 - **Storyboard mode**: 컴포저에서 스토리보드 모드를 켜면 연속 프레임의 인물·장면 연속성을 유지합니다. 이미지와 영상 생성 모두 지원합니다.
 - **Local gallery**: 생성물을 내 컴퓨터에 저장하고 세션별 히스토리로 봅니다. 기본적으로 현재 세션만 보이며 All Images 토글로 전체 히스토리를 볼 수 있습니다. 각 이미지의 생성 시간·reasoning effort가 메타데이터에 기록됩니다.
-- **Reference images**: 레퍼런스를 드래그, 붙여넣기, 파일 선택으로 추가합니다. 이미지 최대 5장, 영상 최대 7장. 큰 이미지는 업로드 전에 자동 압축됩니다.
+- **Reference images**: 레퍼런스를 드래그, 붙여넣기, 파일 선택으로 추가합니다. 이미지 최대 5장, 영상 최대 14장. 큰 이미지는 업로드 전에 자동 압축됩니다.
 - **Prompt library imports**: 로컬 prompt pack, GitHub folder, curated GPT-image hint를 내장 prompt library로 가져옵니다.
 - **Mobile shell**: 작은 화면에서는 app bar, compose sheet, compact settings toggle로 조작합니다.
 - **Observable jobs**: 진행 중인 작업과 최근 완료된 작업을 request ID로 추적합니다.
@@ -139,7 +139,7 @@ Ctrl+C로 DB, 소켓, 자식 프로세스를 정리할 수 있습니다. Windows
 
 Grok은 Classic, Node, Agent 흐름을 지원합니다. Classic 레퍼런스, Node 부모 이미지, Agent 현재 이미지가 있으면 최종 Grok 호출은 xAI image edit 경로로 전환되어 image-to-image 맥락을 유지합니다. 기본 이미지 모델은 `grok-imagine-image-quality`입니다.
 
-Grok video 기본값은 정식 `grok-imagine-video-1.5`입니다. `grok-imagine-video`는 Ref2V, V2V edit, extension 호환 경로에서 계속 쓰며, 기존 `grok-imagine-video-1.5-preview` 문자열도 호환 alias로 받습니다. 레퍼런스 수에 따라 T2V(0), I2V(1), Ref2V(2-7, 최대 10초)가 자동 선택되며, 1080p는 `grok-imagine-video-1.5` 프롬프트 전용 T2V와 단일 이미지/프레임 I2V에서 활성화됩니다. 프롬프트 전용 1.5 T2V는 upstream 요청 전에 내부 흰 캔버스 I2V shim을 사용합니다. duration(1-15s), resolution(480p/720p/지원 시 1080p), aspect ratio 컨트롤을 제공합니다.
+Grok video 기본값은 정식 `grok-imagine-video-1.5`입니다. `grok-imagine-video`는 Ref2V, V2V edit, extension 호환 경로에서 계속 쓰며, 기존 `grok-imagine-video-1.5-preview` 문자열도 호환 alias로 받습니다. 레퍼런스 수에 따라 T2V(0), I2V(1), Ref2V(2-14, grok-imagine-video-1.5는 15초 / grok-imagine-video는 10초)가 자동 선택되며, 1080p는 `grok-imagine-video-1.5` 프롬프트 전용 T2V와 단일 이미지/프레임 I2V에서 활성화됩니다. 프롬프트 전용 1.5 T2V는 upstream 요청 전에 내부 흰 캔버스 I2V shim을 사용합니다. duration(1-15s), resolution(480p/720p/지원 시 1080p), aspect ratio 컨트롤을 제공합니다.
 
 설정 화면의 QuotaCard에 Grok billing `$used/$limit` 바와 **Switch Account** 버튼(`POST /api/auth/switch`)이 표시됩니다.
 

package/docs/README.zh-CN.md

@@ -176,7 +176,7 @@ ima2 skill install --tmp            # install to temp dir (fallback)
 
 Grok图像生成公开了模型选择器（`grok-imagine-image` / `grok-imagine-image-quality`）和尺寸选择器（长宽比 + 1k/2k 分辨率）。设置页面更喜欢Grok建立每周积分百分比并重置时间`GET /v1/billing?format=credits`;如果该来源不可用，则会退回到传统的每月计费窗口，并且`$used/$limit`. A **切换账户**按钮启动设备代码OAuth流动 （`POST /api/auth/switch`）无需离开应用程序即可重新进行身份验证。
 
-Grok视频生成默认为规范`grok-imagine-video-1.5`; `grok-imagine-video`仍然可用于仅限基本模型的 Ref2V、V2V 编辑和扩展路径，以及旧版本`grok-imagine-video-1.5-preview`字符串被接受作为别名。根据引用计数自动检测三种模式：文本到视频（0 引用）、图像到视频（1 引用）和引用到视频（2-7 引用，最长 10 秒持续时间）。 1080p 可用于`grok-imagine-video-1.5`仅提示文本到视频和单图像/帧图像到视频；仅提示 1.5 在上游请求之前使用内部白色画布 I2V 填充程序。视频控制包括持续时间（1-15秒）、分辨率（480p、720p、1080p（如果支持））和宽高比（1:1、16:9、9:16、4:3、3:4、3:2、2:3、自动）。
+Grok视频生成默认为规范`grok-imagine-video-1.5`; `grok-imagine-video`仍然可用于仅限基本模型的 Ref2V、V2V 编辑和扩展路径，以及旧版本`grok-imagine-video-1.5-preview`字符串被接受作为别名。根据引用计数自动检测三种模式：文本到视频（0 引用）、图像到视频（1 引用）和引用到视频（2-14 引用；grok-imagine-video-1.5 最长 15 秒，grok-imagine-video 最长 10 秒）。 1080p 可用于`grok-imagine-video-1.5`仅提示文本到视频和单图像/帧图像到视频；仅提示 1.5 在上游请求之前使用内部白色画布 I2V 填充程序。视频控制包括持续时间（1-15秒）、分辨率（480p、720p、1080p（如果支持））和宽高比（1:1、16:9、9:16、4:3、3:4、3:2、2:3、自动）。
 
 ![设置工作区显示GPT OAuth活跃和API可用的密钥提供者。](../assets/screenshots/settings-oauth-generation.png)
 

package/docs/README.zh-TW.md

@@ -170,7 +170,7 @@ Web UI 使用單一 `GET /api/events` 伺服器發送事件（SSE）連線，接
 
 Grok圖像生成公開了模型選擇器（`grok-imagine-image` / `grok-imagine-image-quality`）和尺寸選擇器（長寬比 + 1k/2k 解析度）。設定頁面更喜歡Grok建立每週積分百分比並重置時間`GET /v1/billing?format=credits`;如果該來源不可用，則會退回到傳統的每月計費窗口，並且`$used/$limit`. A **切換帳戶**按鈕啟動設備代碼OAuth流動 （`POST /api/auth/switch`）無需離開應用程式即可重新進行身份驗證。
 
-Grok影片產生預設為規範`grok-imagine-video-1.5`; `grok-imagine-video`仍可用於僅限基本型號的 Ref2V、V2V 編輯和擴展路徑，以及舊版本`grok-imagine-video-1.5-preview`字串被接受作為別名。根據引用計數自動偵測三種模式：文字到影片（0 引用）、圖像到影片（1 引用）和引用到影片（2-7 引用，最長 10 秒持續時間）。 1080p 可用於`grok-imagine-video-1.5`僅提示文字到影片和單圖像/幀圖像到影片；僅提示 1.5 在上游請求之前使用內部白色畫布 I2V 填充程式。視訊控制包括持續時間（1-15秒）、解析度（480p、720p、1080p（如果支援））和寬高比（1:1、16:9、9:16、4:3、3:4、3:2、2:3、自動）。
+Grok影片產生預設為規範`grok-imagine-video-1.5`; `grok-imagine-video`仍可用於僅限基本型號的 Ref2V、V2V 編輯和擴展路徑，以及舊版本`grok-imagine-video-1.5-preview`字串被接受作為別名。根據引用計數自動偵測三種模式：文字到影片（0 引用）、圖像到影片（1 引用）和引用到影片（2-14 引用；grok-imagine-video-1.5 最長 15 秒，grok-imagine-video 最長 10 秒）。 1080p 可用於`grok-imagine-video-1.5`僅提示文字到影片和單圖像/幀圖像到影片；僅提示 1.5 在上游請求之前使用內部白色畫布 I2V 填充程式。視訊控制包括持續時間（1-15秒）、解析度（480p、720p、1080p（如果支援））和寬高比（1:1、16:9、9:16、4:3、3:4、3:2、2:3、自動）。
 
 ![設定工作區顯示GPT OAuth活躍和API可用的密鑰提供者。](../assets/screenshots/settings-oauth-generation.png)
 

package/docs/grok-video-i2v-research.md

@@ -125,7 +125,7 @@ Status values to handle:
 | default resolution | `480p` |
 | video edit input max | 8.7 seconds — **owner-measured, not documented by xAI.** Re-verified as still-believed 2026-08-20; no public source states this number, so do not drop it just because doc sweeps come back empty |
 | video extension input | 2-15 seconds |
-| extension duration | 2-10 seconds |
+| extension duration | 2-10 seconds (superseded: re-measured as 1-15 on 2026-09-08, see devlog/_plan/260908_xai_imagine_spec_resync) |
 | generated URL | temporary; download promptly |
 
 ### Pricing Observed / Documented

package/docs/grok-video-i2v-research.zh-CN.md

@@ -125,7 +125,7 @@
 |默认分辨率| `480p` |
 |视频编辑最大输入|8.7秒|
 |视频扩展输入|2-15秒|
-|延长期限|2-10秒|
+|延长期限|2-10秒（已更新：2026-09-08 实测为 1-15 秒）|
 |生成的URL |暂时的;及时下载|
 
 ### 观察/记录的定价

package/docs/grok-video-i2v-research.zh-TW.md

@@ -125,7 +125,7 @@
 |預設解析度| `480p` |
 |影片編輯最大輸入|8.7秒|
 |視訊擴充輸入|2-15秒|
-|延長期限|2-10秒|
+|延長期限|2-10秒（已更新：2026-09-08 實測為 1-15 秒）|
 |產生的URL |暫時的;及時下載|
 
 ### 觀察/記錄的定價

package/docs/migration/runtime-test-inventory.md

@@ -4,7 +4,7 @@ Generated by `npm run test:inventory` (script: `scripts/classify-tests.mjs`).
 
 _Tests considered "runtime-importing" if they import from `../lib/`, `../routes/`, `../bin/`, `../server`, or `../config`._
 
-Total: 488 (runtime: 225, contract: 263)
+Total: 492 (runtime: 227, contract: 265)
 
 ## Runtime-importing tests
 - `tests/agent-image-reference-contract.test.ts`
@@ -232,6 +232,8 @@ Total: 488 (runtime: 225, contract: 263)
 - `tests/videoExtendedRoute.test.ts`
 - `tests/videoLineage.test.ts`
 - `tests/videoRoute.test.ts`
+- `tests/voice-picker-contract.test.ts`
+- `tests/xai-video-model-alias-contract.test.ts`
 
 ## Contract-only tests
 - `tests/a11y-modal-contract.test.ts`
@@ -332,6 +334,7 @@ Total: 488 (runtime: 225, contract: 263)
 - `tests/core-selection-transport.test.mjs`
 - `tests/current-image-actions-readiness-contract.test.js`
 - `tests/direct-mode-visual-contract.test.js`
+- `tests/dropped-media-sorting-contract.test.ts`
 - `tests/duration-slider-contract.test.js`
 - `tests/e2e-app-environment.test.ts`
 - `tests/e2e-runtime-build.test.mjs`
@@ -489,6 +492,7 @@ Total: 488 (runtime: 225, contract: 263)
 - `tests/vectorize-panel-contract.test.ts`
 - `tests/video-continuity-ui-contract.test.js`
 - `tests/video-defaults-persistence-contract.test.js`
+- `tests/video-edit-client-contract.test.ts`
 - `tests/video-extend-ui-contract.test.js`
 - `tests/video-frame-local-file.test.js`
 - `tests/video-gallery-refresh-contract.test.ts`

package/lib/capabilities.js

@@ -5,7 +5,7 @@ import { buildCatalog, catalogSummary } from "./contracts/catalog.js";
 import { loadAllBundledSnapshots } from "./mcp/snapshotStore.js";
 import { KEY_TO_ENV, WRITABLE_CONFIG_KEYS } from "./configKeys.js";
 import { DEFAULT_IMAGE_QUALITY, VALID_IMAGE_QUALITIES } from "./oauthNormalize.js";
-import { MAX_REF2V_REFERENCES, MAX_REFERENCE_AUDIOS, MAX_VIDEO_DURATION, MIN_VIDEO_DURATION } from "./imageModels.js";
+import { MAX_REF2V_DURATION_15, MAX_REF2V_DURATION_BASE, MAX_REF2V_REFERENCES, MAX_REFERENCE_AUDIOS, MAX_VIDEO_DURATION, MIN_VIDEO_DURATION, } from "./imageModels.js";
 import { deriveProviderIds, getProviderSurfaceSupport } from "./providers/derive.js";
 import { PROVIDER_SURFACES } from "./providers/surfaceSupport.js";
 const VALID_MODES = ["auto", "direct"];
@@ -86,18 +86,21 @@ export function buildIma2Capabilities({ appConfig = runtimeConfigDefault, packag
             },
             videoModels: {
                 supported: ["grok-imagine-video", "grok-imagine-video-1.5"],
-                aliases: { "grok-imagine-video-1.5-preview": "grok-imagine-video-1.5" },
+                aliases: {
+                    "grok-imagine-video-1.5-preview": "grok-imagine-video-1.5",
+                    "grok-imagine-video-1.5-2026-05-30": "grok-imagine-video-1.5",
+                },
                 resolutions: ["480p", "720p", "1080p"],
                 resolutionNotes: { "1080p": "grok-imagine-video-1.5 text-to-video canvas shim or image-to-video; reference-to-video unsupported" },
                 aspectRatios: ["1:1", "16:9", "9:16", "4:3", "3:4", "3:2", "2:3", "auto"],
                 durationRange: [1, 15],
-                maxReferences: 7,
+                maxReferences: MAX_REF2V_REFERENCES,
                 // Flat numbers above describe the widest case, which is not what any single
                 // request is allowed to do. A client that reads only those draws controls the
                 // server will reject. These per-mode entries come from the same constants the
                 // request path enforces, so the advertisement cannot drift from the behavior.
-                // Verified against api.x.ai on 2026-08-20:
-                // devlog/_plan/260820_grok15_multi_reference_video/000_research.md
+                // Verified against api.x.ai on 2026-09-08:
+                // devlog/_plan/260908_xai_imagine_spec_resync/000_research.md
                 modes: {
                     "text-to-video": {
                         maxReferences: 0,
@@ -113,9 +116,15 @@ export function buildIma2Capabilities({ appConfig = runtimeConfigDefault, packag
                     },
                     "reference-to-video": {
                         maxReferences: MAX_REF2V_REFERENCES,
-                        durationRange: [MIN_VIDEO_DURATION, MAX_VIDEO_DURATION],
+                        // The widest case, which only grok-imagine-video-1.5 can actually reach.
+                        // durationRangeByModel below is what a single request must obey.
+                        durationRange: [MIN_VIDEO_DURATION, MAX_REF2V_DURATION_15],
+                        durationRangeByModel: {
+                            "grok-imagine-video-1.5": [MIN_VIDEO_DURATION, MAX_REF2V_DURATION_15],
+                            "grok-imagine-video": [MIN_VIDEO_DURATION, MAX_REF2V_DURATION_BASE],
+                        },
                         resolutions: ["480p", "720p"],
-                        notes: "References guide the subject without locking the first frame. 1080p is rejected upstream.",
+                        notes: `References guide the subject without locking the first frame. 1080p is rejected upstream. Up to ${MAX_REF2V_DURATION_15}s on grok-imagine-video-1.5, ${MAX_REF2V_DURATION_BASE}s on grok-imagine-video. A preset voice alone also selects this mode.`,
                     },
                 },
                 referenceAudio: {
@@ -130,6 +139,23 @@ export function buildIma2Capabilities({ appConfig = runtimeConfigDefault, packag
                         "atlas",
                     ],
                     presetsAreAuthoritative: false,
+                    // Named by the 400 itself: "Custom voice ids created via the /v1/custom-voices
+                    // API are also accepted." The endpoint answers 200 with a 30-voice cap that is
+                    // documented nowhere else.
+                    customVoiceApi: "/v1/custom-voices",
+                },
+                // Which model takes a video in, and which takes audio in. The two split on input
+                // modality, so there is no path that edits a video on 1.5 — it answers 400.
+                videoInputModes: {
+                    edit: {
+                        models: ["grok-imagine-video"],
+                        notes: "grok-imagine-video-1.5 answers 400 'Video editing is not supported for this model.'",
+                    },
+                    extend: {
+                        models: ["grok-imagine-video"],
+                        durationRange: [MIN_VIDEO_DURATION, MAX_VIDEO_DURATION],
+                        notes: "duration is the added segment, not the total. Measured 1-15s, wider than the documented 2-10.",
+                    },
                 },
             },
             reasoningEfforts: toArray(appConfig.imageModels.validReasoningEfforts),

package/lib/grokVideoAdapter.js

@@ -5,7 +5,7 @@ import { detectImageMimeFromB64 } from "./refs.js";
 import { aspectToCanvas, generateWhiteCanvasB64 } from "./grokVideoCanvas.js";
 import { downloadVideo } from "./grokVideoDownload.js";
 import { buildGrokVideoPlannerSystemPrompt, composeFallbackVideoPrompt, formatDurationPacingGuidance } from "./grokVideoPlannerPrompt.js";
-import { GROK_VIDEO_MODEL_15, GROK_VIDEO_MODEL_15_PREVIEW_ALIAS, GROK_VIDEO_MODEL_BASE, MAX_REF2V_REFERENCES, MAX_REFERENCE_AUDIOS, validateVideoResolutionForRequest, } from "./imageModels.js";
+import { GROK_VIDEO_MODEL_15, GROK_VIDEO_MODEL_BASE, MAX_REF2V_REFERENCES, MAX_REFERENCE_AUDIOS, canonicalGrokVideoModel, validateVideoDurationForRequest, validateVideoResolutionForRequest, } from "./imageModels.js";
 import { formatVideoContinuityForPlanner } from "./videoContinuity.js";
 import { DEFAULT_GROK_PLANNER_MODEL } from "../config.js";
 import { videoConfig, videoEndpoint, withTimeoutSignal, } from "./grokVideoShared.js";
@@ -41,9 +41,7 @@ export function isDegradablePlannerFailure(err) {
         return { reason: "failed", message };
     return null;
 }
-function canonicalVideoModel(model) {
-    return model === GROK_VIDEO_MODEL_15_PREVIEW_ALIAS ? GROK_VIDEO_MODEL_15 : model;
-}
+const canonicalVideoModel = canonicalGrokVideoModel;
 function sourceImageUrl(image, mime) {
     if (image.startsWith("data:") || image.startsWith("http"))
         return image;
@@ -314,6 +312,13 @@ export function buildVideoGenerationPayload(plan, opts) {
     if (!("ok" in resolutionCheck)) {
         throw grokError(resolutionCheck.error, resolutionCheck.status, resolutionCheck.code);
     }
+    // Answered here as well as at the route because the agent path builds payloads without
+    // going through routes/video.ts. Two checks of the same rule are two defenses, not a
+    // duplicated rule: both read the same function.
+    const durationCheck = validateVideoDurationForRequest(model, plan.duration, plan.mode);
+    if (!("ok" in durationCheck)) {
+        throw grokError(durationCheck.error, durationCheck.status, durationCheck.code);
+    }
     const payload = { model, prompt: plan.prompt, duration: plan.duration, resolution: plan.resolution };
     if (plan.aspectRatio && plan.aspectRatio !== "auto")
         payload.aspect_ratio = plan.aspectRatio;

package/lib/imageModels.js

@@ -171,10 +171,15 @@ export function normalizeComfyWorkflowModel(rawModel) {
 export const GROK_VIDEO_MODEL_BASE = "grok-imagine-video";
 export const GROK_VIDEO_MODEL_15 = "grok-imagine-video-1.5";
 export const GROK_VIDEO_MODEL_15_PREVIEW_ALIAS = "grok-imagine-video-1.5-preview";
+// GET /v1/video-generation-models lists this alongside -preview. Without it a legal
+// model id is rejected locally with INVALID_GROK_VIDEO_MODEL before the request ever
+// reaches xAI. Verified 2026-09-08: devlog/_plan/260908_xai_imagine_spec_resync.
+export const GROK_VIDEO_MODEL_15_DATED_ALIAS = "grok-imagine-video-1.5-2026-05-30";
 export const GROK_FALLBACK_VIDEO_MODEL = GROK_VIDEO_MODEL_15;
 export const VALID_GROK_VIDEO_MODELS = new Set([
     ...deriveModels("grok", "video"),
     GROK_VIDEO_MODEL_15_PREVIEW_ALIAS,
+    GROK_VIDEO_MODEL_15_DATED_ALIAS,
 ]);
 export const VALID_VIDEO_RESOLUTIONS = new Set(["480p", "720p", "1080p"]);
 export const VALID_VIDEO_ASPECT_RATIOS = new Set([
@@ -189,10 +194,25 @@ export const VALID_VIDEO_ASPECT_RATIOS = new Set([
 ]);
 export const MIN_VIDEO_DURATION = 1;
 export const MAX_VIDEO_DURATION = 15;
-// reference-to-video (xAI): up to 7 reference images (8 -> 400), 1-15s, 720p max.
-// Verified against api.x.ai on 2026-08-20:
-// devlog/_plan/260820_grok15_multi_reference_video/000_research.md
-export const MAX_REF2V_REFERENCES = 7;
+// reference-to-video (xAI): up to 14 reference images, 720p max.
+//
+// The cap doubled from 7. 15 returns 400 "Too many reference images: 15. Maximum
+// allowed is 14." and 30 names the same number, on BOTH video models, so it is a
+// model-independent limit. @imagine announced it on 2026-09-02; the API docs still
+// say 7, which is why the measurement is the source here and not the page.
+// Re-verified against api.x.ai on 2026-09-08:
+// devlog/_plan/260908_xai_imagine_spec_resync/000_research.md
+export const MAX_REF2V_REFERENCES = 14;
+// reference-to-video carries its own duration ceiling, and it differs by model. The
+// same 16s request answers "maximum allowed for reference-to-video, which is 15s" on
+// grok-imagine-video-1.5 and "... is 10s" on grok-imagine-video.
+//
+// This is NOT the ceiling removed in 260820. That one was invented locally and applied
+// to every model against evidence. This one is the upstream's own per-model rule, and
+// omitting it turns a legal-looking base-model request into an upstream 400 the user
+// pays a planning round for.
+export const MAX_REF2V_DURATION_15 = 15;
+export const MAX_REF2V_DURATION_BASE = 10;
 // reference_audios: preset voices, grok-imagine-video-1.5 only. 4 -> 400.
 export const MAX_REFERENCE_AUDIOS = 3;
 /**
@@ -216,6 +236,55 @@ export function deriveVideoMode(refCount) {
 export function isGrokVideoModel(value) {
     return typeof value === "string" && VALID_GROK_VIDEO_MODELS.has(value);
 }
+/**
+ * Folds every alias onto the id xAI bills and validates against.
+ *
+ * Three call sites used to inline `model === PREVIEW_ALIAS ? GROK_VIDEO_MODEL_15 : model`,
+ * which is exactly one alias' worth of knowledge duplicated three times. Adding the dated
+ * alias would have had to find all three; one of them would have been missed.
+ *
+ * Passes non-Grok ids through unchanged: comfy hands a workflow id down this same
+ * `model` field (routes/video.ts), and rewriting it would be worse than leaving it alone.
+ */
+export function canonicalGrokVideoModel(model) {
+    return model === GROK_VIDEO_MODEL_15_PREVIEW_ALIAS || model === GROK_VIDEO_MODEL_15_DATED_ALIAS
+        ? GROK_VIDEO_MODEL_15
+        : model;
+}
+/**
+ * The reference-to-video duration ceiling for a model, or null when the rule does not
+ * apply to it.
+ *
+ * Null is not "unknown" and not "no limit": it means xAI's per-model r2v rule has no
+ * jurisdiction here. A comfy workflow id arrives through the same `model` field, and
+ * answering the base model's 10s for it would clamp a lane whose ceiling xAI does not
+ * own. Callers must handle null explicitly rather than let it fall through to a number.
+ */
+export function maxRef2vDuration(model) {
+    if (!isGrokVideoModel(model))
+        return null;
+    return canonicalGrokVideoModel(model) === GROK_VIDEO_MODEL_15
+        ? MAX_REF2V_DURATION_15
+        : MAX_REF2V_DURATION_BASE;
+}
+/**
+ * Validates a duration against the ceiling its mode and model actually impose.
+ *
+ * Returns ok for every non-reference-to-video request: normalizeVideoDuration already
+ * owns the shared 1-15 bound, and re-checking it here would give one rule two owners.
+ */
+export function validateVideoDurationForRequest(model, duration, mode) {
+    if (mode !== "reference-to-video")
+        return { ok: true };
+    const ceiling = maxRef2vDuration(model);
+    if (ceiling === null || duration <= ceiling)
+        return { ok: true };
+    return {
+        error: `reference-to-video on ${model} allows at most ${ceiling} seconds`,
+        code: "INVALID_VIDEO_DURATION",
+        status: 400,
+    };
+}
 export function normalizeGrokVideoModel(rawModel) {
     if (typeof rawModel !== "string" || rawModel.length === 0) {
         return { model: GROK_FALLBACK_VIDEO_MODEL };
@@ -227,7 +296,7 @@ export function normalizeGrokVideoModel(rawModel) {
             status: 400,
         };
     }
-    return { model: rawModel === GROK_VIDEO_MODEL_15_PREVIEW_ALIAS ? GROK_VIDEO_MODEL_15 : rawModel };
+    return { model: canonicalGrokVideoModel(rawModel) };
 }
 export function normalizeVideoResolution(raw) {
     if (raw === undefined || raw === null || raw === "")
@@ -242,13 +311,17 @@ export function normalizeVideoResolution(raw) {
     return { resolution: raw };
 }
 export function usesGrokVideo15TextCanvasShim(model, mode) {
-    const canonicalModel = model === GROK_VIDEO_MODEL_15_PREVIEW_ALIAS ? GROK_VIDEO_MODEL_15 : model;
-    return canonicalModel === GROK_VIDEO_MODEL_15 && mode === "text-to-video";
+    return canonicalGrokVideoModel(model) === GROK_VIDEO_MODEL_15 && mode === "text-to-video";
 }
 export function validateVideoResolutionForRequest(model, resolution, mode, options = {}) {
     if (resolution !== "1080p")
         return { ok: true };
-    const canonicalModel = model === GROK_VIDEO_MODEL_15_PREVIEW_ALIAS ? GROK_VIDEO_MODEL_15 : model;
+    // A non-Grok model reaches this function too: comfy passes its workflow id through the
+    // same field. xAI's 1080p rule says nothing about a locally-run workflow, so refusing
+    // one here would be this codebase inventing a limit rather than enforcing a real one.
+    if (!isGrokVideoModel(model))
+        return { ok: true };
+    const canonicalModel = canonicalGrokVideoModel(model);
     if (canonicalModel === GROK_VIDEO_MODEL_15 && mode === "image-to-video") {
         return { ok: true };
     }

package/lib/providers/registry.js

@@ -68,9 +68,13 @@ export const REGISTRY = [
             { id: "grok-imagine-image", kind: "image", supports: EDIT },
             { id: "grok-imagine-image-quality", kind: "image", supports: EDIT },
             { id: "grok-imagine-video", kind: "video", supports: EDIT },
-            { id: "grok-imagine-video-1.5", aliases: ["grok-imagine-video-1.5-preview"], kind: "video", supports: EDIT },
+            { id: "grok-imagine-video-1.5", aliases: ["grok-imagine-video-1.5-preview", "grok-imagine-video-1.5-2026-05-30"], kind: "video", supports: EDIT },
         ],
-        referenceLimits: { image: 3, edit: 3, video: 7 },
+        // image/edit 5: measured on grok-imagine-image-2.0 through the oauth proxy on
+        // 2026-09-08 (6 -> 400 "This model supports at most 5 input image(s)"), and xAI's
+        // release notes say editing "now accepts up to 5 source images per request (was 3)".
+        // video 14: see MAX_REF2V_REFERENCES in lib/imageModels.ts.
+        referenceLimits: { image: 5, edit: 5, video: 14 },
         elementTaxonomy: "grok",
         limits: { timeoutMs: 300_000 },
         errorPrefix: "GROK_",
@@ -92,9 +96,12 @@ export const REGISTRY = [
             { id: "grok-imagine-image", kind: "image", supports: EDIT },
             { id: "grok-imagine-image-quality", kind: "image", supports: EDIT },
             { id: "grok-imagine-video", kind: "video", supports: EDIT },
-            { id: "grok-imagine-video-1.5", aliases: ["grok-imagine-video-1.5-preview"], kind: "video", supports: EDIT },
+            { id: "grok-imagine-video-1.5", aliases: ["grok-imagine-video-1.5-preview", "grok-imagine-video-1.5-2026-05-30"], kind: "video", supports: EDIT },
         ],
-        referenceLimits: { image: 3, edit: 3, video: 7 },
+        // Same upstream as the grok lane, reached with a direct key instead of the proxy, so
+        // the caps are taken to match. The 5 was measured on the proxy path only; if an edit
+        // fails at 4-5 images with a direct key, this assumption is the place to look.
+        referenceLimits: { image: 5, edit: 5, video: 14 },
         elementTaxonomy: "grok",
         limits: { timeoutMs: 300_000 },
         errorPrefix: "GROK_",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ima2-gen",
-  "version": "3.15.0",
+  "version": "3.15.1",
   "packageManager": "npm@11.18.0",
   "description": "Local-first visual generation runtime and studio for people and coding agents, with reproducible image and video workflows across multiple providers.",
   "type": "module",
@@ -127,5 +127,5 @@
     "typescript": "^5.9.3",
     "yaml": "2.9.0"
   },
-  "gitHead": "b96a11ed8a46c5782e673b5eeb61129d9f3aee9d"
+  "gitHead": "11900764a59e74146c9008ed296a23d3fc0ac06a"
 }

package/routes/video.js

@@ -16,7 +16,7 @@ import { getVideoSeriesChain } from "../lib/videoSeriesChain.js";
 import { ACTIVE_VIDEO_PROMPT_GUIDANCE, appendVideoContinuityEntry, lineageFromVideoMetadata, normalizeVideoContinuityLineage, readVideoSidecar, requireActiveVideoPrompt, safeGeneratedVideoFilename, } from "../lib/videoContinuity.js";
 import { extractGeneratedVideoFrameB64 } from "../lib/videoFrameExtract.js";
 import { errorEnvelopeFields } from "../lib/errors/envelope.js";
-import { normalizeGrokVideoModel, normalizeVideoResolution, normalizeVideoAspectRatio, normalizeVideoDuration, deriveVideoMode, MAX_REF2V_REFERENCES, MAX_REFERENCE_AUDIOS, validateVideoResolutionForRequest, } from "../lib/imageModels.js";
+import { normalizeGrokVideoModel, normalizeVideoResolution, normalizeVideoAspectRatio, normalizeVideoDuration, deriveVideoMode, MAX_REF2V_REFERENCES, MAX_REFERENCE_AUDIOS, validateVideoResolutionForRequest, validateVideoDurationForRequest, } from "../lib/imageModels.js";
 import { errInfo } from "../lib/errInfo.js";
 import { requireRuntimeContext } from "../lib/runtimeContext.js";
 import { generateVideoThumbnail } from "../lib/videoThumb.js";
@@ -338,18 +338,36 @@ export function registerVideoRoutes(app, ctxRaw) {
             // devlog/_plan/260820_grok15_multi_reference_video/030_single_ref_mode_choice.md
             const composerRefCount = resolved.filter((r) => r.source === "composer").length;
             const requestedMode = typeof req.body?.mode === "string" ? req.body.mode : null;
+            // Shape only. Which voice ids exist is xAI's to answer, and its 400 names every
+            // valid voice — a list we would only get wrong, and which cannot include the
+            // caller's custom voices anyway. Parsed here rather than at the call site because
+            // the mode derivation below needs to know whether any voice was attached.
+            const referenceAudios = toArray(req.body?.referenceAudios)
+                .map((voice) => (typeof voice === "string" ? voice.trim() : ""))
+                .filter((voice) => voice.length > 0);
+            if (referenceAudios.length > MAX_REFERENCE_AUDIOS) {
+                return fail(400, "GROK_VIDEO_AUDIO_TOO_MANY", `at most ${MAX_REFERENCE_AUDIOS} reference voices`);
+            }
             const derivedMode = composerRefCount > 0
                 ? "reference-to-video"
-                : deriveVideoMode(resolved.length);
+                // A voice alone selects reference-to-video upstream: "at least one reference of
+                // either kind selects the reference-to-video mode" (xAI OpenAPI,
+                // GenerateVideoRequest.reference_audios). Calling it text-to-video here would let
+                // a 15s base-model request through while xAI applies the 10s r2v ceiling, and the
+                // rejection would name a mode this request never asked for.
+                : referenceAudios.length > 0
+                    ? "reference-to-video"
+                    : deriveVideoMode(resolved.length);
             const mode = incomingProviderUrl
                 ? "image-to-video"
                 : (requestedMode === "reference-to-video" || requestedMode === "image-to-video" || requestedMode === "text-to-video")
                     ? requestedMode
                     : derivedMode;
-            // An explicit reference-to-video with nothing to reference would ship an empty
-            // reference_images array and fail upstream with a less useful message.
-            if (mode === "reference-to-video" && resolved.length === 0) {
-                return fail(400, "GROK_VIDEO_INVALID_MODE", "reference-to-video requires at least 1 reference image");
+            // An explicit reference-to-video with nothing to reference at all would ship empty
+            // arrays and fail upstream with a less useful message. A voice counts as a
+            // reference, so audio-only is legitimate and must not be rejected here.
+            if (mode === "reference-to-video" && resolved.length === 0 && referenceAudios.length === 0) {
+                return fail(400, "GROK_VIDEO_INVALID_MODE", "reference-to-video requires at least 1 reference image or voice");
             }
             const duration = durationCheck.duration;
             const resolutionModeCheck = validateVideoResolutionForRequest(modelCheck.model, resolutionCheck.resolution, mode, {
@@ -357,6 +375,12 @@ export function registerVideoRoutes(app, ctxRaw) {
             });
             if (isNormalizeError(resolutionModeCheck))
                 return fail(resolutionModeCheck.status, resolutionModeCheck.code, resolutionModeCheck.error);
+            // The r2v ceiling is per model, and the start call is deliberately never retried
+            // (a retry could bill a second video), so a 400 we can predict must be answered
+            // before the job is admitted rather than asynchronously after planning.
+            const durationModeCheck = validateVideoDurationForRequest(modelCheck.model, duration, mode);
+            if (isNormalizeError(durationModeCheck))
+                return fail(durationModeCheck.status, durationModeCheck.code, durationModeCheck.error);
             const referenceImages = mode === "reference-to-video" ? resolved.map((r) => r.b64) : undefined;
             const sourceB64 = incomingProviderUrl || (mode === "image-to-video" ? resolved[0]?.b64 : undefined);
             const sourceFilename = resolved[0]?.filename ?? null;
@@ -483,15 +507,6 @@ export function registerVideoRoutes(app, ctxRaw) {
                 + (backgroundPreset ? ` ${backgroundPromptSuffix(backgroundPreset, "video")}` : "");
             const plannerModel = typeof req.body?.plannerModel === "string" ? req.body.plannerModel.trim() : undefined;
             const directApiKey = provider === "grok-api" ? ctx.xaiApiKey : undefined;
-            // Only the shape is checked here. Which voice ids exist is xAI's to answer, and its
-            // 400 names every valid voice — a list we would only get wrong, and which cannot
-            // include the caller's custom voices anyway.
-            const referenceAudios = toArray(req.body?.referenceAudios)
-                .map((voice) => (typeof voice === "string" ? voice.trim() : ""))
-                .filter((voice) => voice.length > 0);
-            if (referenceAudios.length > MAX_REFERENCE_AUDIOS) {
-                return fail(400, "GROK_VIDEO_AUDIO_TOO_MANY", `at most ${MAX_REFERENCE_AUDIOS} reference voices`);
-            }
             const result = await generateVideoViaGrok(effectivePrompt, ctx, {
                 model: modelCheck.model,
                 mode,