@devmotion/cli 0.7.0 → 0.8.0

package/studio/server/chunks/_server.ts-CGs8hOo6.js

@@ -0,0 +1,760 @@
+import { tool, ToolLoopAgent, convertToModelMessages } from 'ai';
+import { e as error, c as isHttpError } from './index-Dp2_pJiZ.js';
+import { env } from './env-AoZ3P0fg.js';
+import { createOpenRouter } from '@openrouter/ai-sdk-provider';
+import { z } from 'zod';
+import { P as ProjectSchema, p as projectDataSchema } from './animation-sOd5R1cE.js';
+import { a as analyzeAndSaveMediaAsset, M as MediaSegmentTypeSchema } from './media-analysis-DXUN3NUM.js';
+import { a as actionRegistry } from './registry-DcyeMfNZ.js';
+import { d as designPlaybookSection, p as presetCatalogSection, c as assetLibrarySection, e as assetProtocolSection, n as narrationWorkflowSection, v as voiceCatalogSection, U as UngroupLayersInputSchema, G as GroupLayersInputSchema, C as ConfigureProjectInputSchema, R as RemoveLayerInputSchema, a as UpsertLayerInputSchema } from './design-knowledge-D2NUeAAo.js';
+import { d as dmElements, e as describeLayerAttributes } from './attribute-protocol-DQ-1julf.js';
+import { c as canUserAccessAI, b as getModel, l as logAIUsage } from './ai-access-bDlOKhrS.js';
+import { existsSync, mkdirSync, writeFileSync } from 'fs';
+import { join } from 'path';
+import { t as trackEvent } from './posthog-kZp2zKvY.js';
+import { a as asset } from './api-Do--xZ9M.js';
+import { inArray, eq, desc } from 'drizzle-orm';
+import './shared-server-hNytdPP9.js';
+import './branding-5rLT-BvK.js';
+import './index6-BwFC1GDh.js';
+import '@aws-sdk/client-s3';
+import '@aws-sdk/s3-request-presigner';
+import 'nanoid';
+import './filename-sanitizer-yQZjRxRx.js';
+import './find-stock-image-Cwkwf5gV.js';
+import './base-qepI6vHK.js';
+import './media-v9Bq2kuR.js';
+import './presets-CU-ej2J9.js';
+import './project-layer-yZ50OnjM.js';
+import './plans-BAwNKt6H.js';
+import './subscription-BC4gCTG6.js';
+import './root-DjEsp22s.js';
+import './index9-D9BBZH26.js';
+import './async-lVJA8xJZ.js';
+import './event-DmhHnWT5.js';
+import './query-D9bdUhU3.js';
+import './shared-hgBL35K7.js';
+import 'drizzle-orm/pg-core';
+import './optional-services-BGRY7tmz.js';
+
+const AnalyzeMediaInputSchema = z.object({
+  layerId: z.string().describe(
+    "Video or audio layer ID or name to analyze. Sends media to multimodal AI for detailed scene-by-scene breakdown."
+  ),
+  context: z.string().optional().describe(
+    'Optional hint about content type (e.g., "tech tutorial about React hooks") to improve analysis accuracy'
+  )
+});
+const ComposeFromMediaInputSchema = z.object({
+  layerId: z.string().describe(
+    "Source video/audio layer ID or name. Must have been analyzed with analyze_media first."
+  ),
+  mode: z.enum(["keep-highlights", "remove-dead-space", "custom"]).describe(
+    "Editing mode: keep-highlights = only highlight segments, remove-dead-space = everything except silence/filler, custom = filter by type or specify exact time ranges"
+  ),
+  customSegments: z.array(
+    z.object({
+      sourceStart: z.number().min(0).describe("Start time in source media (seconds)"),
+      sourceEnd: z.number().min(0).describe("End time in source media (seconds)"),
+      label: z.string().optional().describe("Label for this segment (becomes layer name)")
+    })
+  ).optional().describe('For "custom" mode: exact time ranges to keep from the source'),
+  includeTypes: z.array(MediaSegmentTypeSchema).optional().describe(
+    'For "custom" mode: only keep segments of these types (e.g., ["content", "highlight"])'
+  ),
+  excludeTypes: z.array(MediaSegmentTypeSchema).optional().describe(
+    'For "custom" mode: remove segments of these types (e.g., ["dead-space", "tangent"])'
+  ),
+  gapBetweenSegments: z.number().min(0).max(5).default(0).describe("Gap between consecutive segments in the output timeline (seconds, default: 0)")
+});
+const RunActionInputSchema = z.object({
+  layerId: z.string().optional().describe(
+    "Source layer ID or name. Required for: generate-captions, dub-audio, extract-audio, generate-speech. Omit for: generate-video."
+  ),
+  actionId: z.enum(Object.keys(actionRegistry)).describe(
+    "Action to run: " + Object.values(actionRegistry).map((a) => `"${a.id}" — ${a.description}`).join(" | ")
+  ),
+  params: z.record(z.string(), z.unknown()).optional().describe(
+    'Action-specific parameters. generate-captions: { language? (e.g., "en"), prompt? (context), position? ("bottom"|"top"|"middle"), mode? ("highlight"|"single-word"|"typewriter"|"block"), fontFamily?, fontSize?, fontWeight?, textColor?, bgColor?, textEffect? }. dub-audio: { targetLang ("es"|"fr"|"de"|"it"|"pt"|"ru"|"ja"|"ko"|"zh"|"ar"|"hi"|"nl"|"pl"|"tr"), sourceLang? }. extract-audio: no params needed. generate-video: { prompt (required, describe the video), duration? (4–15s, default 5), resolution? ("480p"|"720p"), generateAudio? (bool, default true) }. generate-image: { prompt (required, describe the image), aspect? ("square"|"landscape"|"portrait"), transparentBackground? (bool, for logos/cutouts) }. find-stock-image: { query (required, search terms), orientation? ("square"|"horizontal"|"vertical"), imageType? ("photo"|"illustration"|"3d") }. generate-speech: { text (required), voiceId (required — exact case-sensitive value from the TTS Voice Catalog), modelId? (eleven_v3|eleven_v3_turbo|eleven_multilingual_v2|eleven_turbo_v2_5|eleven_flash_v2_5), stability? (0–1), similarityBoost? (0–1) }.'
+  )
+});
+const animationTools = {
+  upsert_layer: tool({
+    description: `Create or modify a layer. All sizes/positions in % of project width (not pixels).
+- CREATE: omit layerId, provide layer {type, props}. Omitted props use schema defaults.
+- EDIT: pass layerId (from a previous upsert_layer/run_action response, or the layer name). Everything is a patch — provide only what changes. layer {type, props} only when changing props (type must match the target layer). Use animation.preset/keyframes to animate existing layers, including media layers added by run_action.
+
+Example — create a text title with fade-in:
+{"name":"Title","layer":{"type":"text","props":{"content":"Hello","fontSize":4,"color":"#ffffff"}},"transform":{"position":{"x":0,"y":-20}},"animation":{"preset":{"id":"fade-in","startTime":0,"duration":0.5}}}
+
+Example — retime an existing layer:
+{"layerId":"abc123","enterTime":2,"exitTime":6,"animation":{"preset":{"id":"fade-out","startTime":5.6,"duration":0.4}}}`,
+    inputSchema: UpsertLayerInputSchema
+  }),
+  remove_layer: tool({
+    description: "Delete a layer by ID (from upsert_layer) or name.",
+    inputSchema: RemoveLayerInputSchema
+  }),
+  configure_project: tool({
+    description: "Set project name, dimensions (px), duration (s), or background (hex string or gradient object).",
+    inputSchema: ConfigureProjectInputSchema
+  }),
+  group_layers: tool({
+    description: "Group multiple layers into a folder. Transform/rotate/scale the group to affect all children together.",
+    inputSchema: GroupLayersInputSchema
+  }),
+  ungroup_layers: tool({
+    description: "Dissolve a group. Children become top-level; group transforms are baked into each child.",
+    inputSchema: UngroupLayersInputSchema
+  }),
+  run_action: tool({
+    description: `Run an AI-powered media action:
+- generate-captions: Transcribe audio/video to synced captions layer. Needs layerId. Params: { language?, prompt?, position? "bottom"|"top"|"middle", mode? "highlight"|"single-word"|"typewriter"|"block", fontFamily?, fontSize?, fontWeight?, textColor?, bgColor?, textEffect?, textEffectColor?, textEffectSize?, transitionEffects?, transitionDuration? }
+- dub-audio: AI voice-clone dubbing into another language. Needs layerId. Params: { targetLang (required, e.g. "es"), sourceLang? }
+- extract-audio: Rip audio track from video. Needs layerId. No params.
+- generate-video: Seedance 2.0 text-to-video. No layerId. Params: { prompt (required), duration? 4–15s, resolution? "480p"|"720p", generateAudio? }
+- generate-image: AI text-to-image, saved as asset + image layer. No layerId. Params: { prompt (required), aspect? "square"|"landscape"|"portrait", transparentBackground? (logos/cutouts) }
+- find-stock-image: Search real stock photos (Lummi), import best match as asset + image layer with attribution. No layerId. Params: { query (required), orientation?, imageType? }. The response includes layersAdded[].props.src/fileKey for reuse.
+- generate-speech: ElevenLabs TTS synthesis. Needs layerId of an audio layer (create one first if needed). Params: { text (required), voiceId (required, exact case-sensitive value from the TTS Voice Catalog), modelId?, stability? 0–1, similarityBoost? 0–1 }
+
+Slow/credit-consuming actions (generate-video, generate-image, dub-audio, generate-speech) pause for user approval before running — the user sees the action and its inputs and can approve or deny.`,
+    inputSchema: RunActionInputSchema,
+    // Human-in-the-loop: expensive external API actions pause for user approval.
+    // Flag lives on the action definition — single source for chat and any
+    // future approval-aware surface.
+    needsApproval: (input) => actionRegistry[input.actionId]?.requiresApproval ?? false
+  }),
+  compose_from_media: tool({
+    description: `Split an analyzed video/audio into sequential segments. Requires analyze_media first.
+
+Modes:
+- keep-highlights: only highlight segments
+- remove-dead-space: everything except silence/filler
+- custom: filter by segment type (includeTypes/excludeTypes) or exact time ranges (customSegments)
+
+Source layer is replaced by N sequential layers, each pointing to the same file with different contentOffset and timing. Add gapBetweenSegments for breathing room.`,
+    inputSchema: ComposeFromMediaInputSchema
+  })
+};
+function compactLayerPropsReference() {
+  const lines = [];
+  for (const entry of dmElements) {
+    const { attributes } = describeLayerAttributes(entry.layerType, entry.contentProp);
+    const parts = attributes.map((a) => {
+      if (a.options?.length) return `${a.prop}[${a.options.join("|")}]`;
+      if (a.type === "number" && (a.min !== void 0 || a.max !== void 0)) {
+        return `${a.prop}(num ${a.min ?? ""}…${a.max ?? ""})`;
+      }
+      return a.prop;
+    });
+    const content = entry.contentProp ? `${entry.contentProp}(text), ` : "";
+    lines.push(`- ${entry.layerType}: ${content}${parts.join(", ")}`);
+  }
+  return lines.join("\n");
+}
+const fps = 30;
+const name = "Untitled Project";
+const width = 1280;
+const height = 720;
+const schemaVersion = 2;
+const layers = [{ "id": "random-id-1", "name": "Text", "type": "text", "props": { "color": "#ffffff", "width": 31.25, "content": "Creating", "fontSize": 3.75, "autoWidth": true, "textAlign": "center", "fontFamily": "Poppins", "fontWeight": "normal" }, "style": { "opacity": 1 }, "locked": false, "visible": true, "keyframes": [{ "id": "random-id-2", "time": 0, "value": -5.27, "property": "position.x", "interpolation": { "family": "continuous", "strategy": "ease-in-out" } }, { "id": "random-id-3", "time": 0, "value": 0, "property": "position.y", "interpolation": { "family": "continuous", "strategy": "ease-in-out" } }, { "id": "random-id-4", "time": 0, "value": 0, "property": "opacity", "interpolation": { "family": "continuous", "strategy": "ease-in-out" } }, { "id": "random-id-5", "time": 0, "value": 0.7, "property": "scale.x", "interpolation": { "family": "continuous", "strategy": "ease-in-out" } }, { "id": "random-id-5", "time": 0, "value": 0.7, "property": "scale.y", "interpolation": { "family": "continuous", "strategy": "ease-in-out" } }, { "id": "random-id-6", "time": 0, "value": "Creating", "property": "props.content", "interpolation": { "family": "text", "strategy": "char-reveal" } }, { "id": "random-id-6", "time": 0.625, "value": "Creating", "property": "props.content", "interpolation": { "family": "text", "strategy": "word-reveal", "separator": " " } }, { "id": "random-id-7", "time": 0.6278125, "value": 1, "property": "scale.y", "interpolation": { "family": "continuous", "strategy": "ease-in-out" } }, { "id": "random-id-8", "time": 0.6278125, "value": 1, "property": "scale.x", "interpolation": { "family": "continuous", "strategy": "ease-in-out" } }, { "id": "random-id-9", "time": 0.6278125, "value": -5.27, "property": "position.x", "interpolation": { "family": "continuous", "strategy": "ease-in-out" } }, { "id": "random-id-10", "time": 1.0278125, "value": 1, "property": "opacity", "interpolation": { "family": "continuous", "strategy": "ease-in-out" } }, { "id": "random-id-11", "time": 1.1840000000000008, "value": -24.27, "property": "position.x", "interpolation": { "family": "continuous", "strategy": "ease-in-out" } }, { "id": "random-id-12", "time": 1.5478125, "value": "Creating Motion design videos", "property": "props.content", "interpolation": { "family": "text", "strategy": "word-reveal", "separator": " " } }, { "id": "random-id-13", "time": 1.7278125, "value": 1, "property": "opacity", "interpolation": { "family": "continuous", "strategy": "ease-in" } }, { "id": "random-id-14", "time": 1.7278125, "value": 0, "property": "position.y", "interpolation": { "family": "continuous", "strategy": "ease-in-out" } }, { "id": "random-id-15", "time": 2.4078125000000017, "value": 0, "property": "opacity", "interpolation": { "family": "continuous", "strategy": "linear" } }, { "id": "random-id-16", "time": 2.4078125000000017, "value": 3.28, "property": "position.y", "interpolation": { "family": "continuous", "strategy": "ease-in-out" } }], "transform": { "scale": { "x": 0.7, "y": 0.7 }, "anchor": "center-left", "position": { "x": 0, "y": 0, "z": 0 }, "rotation": { "x": 0, "y": 0, "z": 0 } } }, { "id": "random-id-17", "name": "Text", "type": "text", "props": { "color": "#ffffff", "width": 31.25, "content": "is", "fontSize": 3.75, "autoWidth": true, "textAlign": "center", "fontFamily": "Poppins", "fontWeight": "normal" }, "style": { "opacity": 1 }, "locked": false, "visible": true, "exitTime": 5.31, "enterTime": 2.4, "keyframes": [{ "id": "random-id-18", "time": 2.4378125, "value": 0, "property": "opacity", "interpolation": { "family": "continuous", "strategy": "ease-out" } }, { "id": "random-id-19", "time": 2.4478125, "value": "is", "property": "props.content", "interpolation": { "family": "text", "strategy": "char-reveal" } }, { "id": "random-id-20", "time": 2.4478125, "value": 0, "property": "opacity", "interpolation": { "family": "continuous", "strategy": "ease-out" } }, { "id": "random-id-21", "time": 2.8478125, "value": 1, "property": "opacity", "interpolation": { "family": "continuous", "strategy": "linear" } }, { "id": "random-id-22", "time": 2.9378125, "value": 1, "property": "opacity", "interpolation": { "family": "continuous", "strategy": "linear" } }, { "id": "random-id-23", "time": 4.7178125, "value": 1, "property": "opacity", "interpolation": { "family": "continuous", "strategy": "ease-in" } }, { "id": "random-id-24", "time": 4.8478125, "value": "is slow & boring", "property": "props.content", "interpolation": { "family": "text", "strategy": "char-reveal" } }, { "id": "random-id-25", "time": 5.187812500000001, "value": 0, "property": "opacity", "interpolation": { "family": "continuous", "strategy": "linear" } }], "transform": { "scale": { "x": 1, "y": 1 }, "anchor": "center", "position": { "x": 0, "y": 0, "z": 0 }, "rotation": { "x": 0, "y": 0, "z": 0 } } }];
+const duration = 10;
+const background = { "type": "linear", "angle": 110, "stops": [{ "color": "#c4f9ff", "position": 11 }, { "color": "#eec5f6", "position": 42 }, { "color": "#edb5fe", "position": 72 }, { "color": "#7bb7fd", "position": 100 }] };
+const fontFamily = "Raleway";
+const goodExampleRaw = {
+  fps,
+  name,
+  width,
+  height,
+  schemaVersion,
+  layers,
+  duration,
+  background,
+  fontFamily
+};
+const exampleProject = projectDataSchema.parse(goodExampleRaw);
+function formatTimestamp(seconds) {
+  const mins = Math.floor(seconds / 60);
+  const secs = Math.round(seconds % 60);
+  return `${mins}:${secs.toString().padStart(2, "0")}`;
+}
+function buildMediaContextSection(mediaContexts) {
+  if (mediaContexts.length === 0) return "";
+  const lines = ["## Media in Project\n"];
+  for (const media of mediaContexts) {
+    lines.push(`### "${media.layerName}" (${media.layerType})`);
+    if (!media.annotation) {
+      lines.push("Not yet analyzed. Call **analyze_media** first.\n");
+      continue;
+    }
+    const ann = media.annotation;
+    lines.push(
+      `Duration: ${formatTimestamp(ann.duration)} | Model: ${ann.modelUsed.split("/").pop()}`
+    );
+    lines.push(`Summary: ${ann.summary}`);
+    if (ann.topics.length > 0) {
+      lines.push(`Topics: ${ann.topics.join(", ")}`);
+    }
+    const typeCounts = {};
+    for (const seg of ann.segments) {
+      typeCounts[seg.type] = (typeCounts[seg.type] || 0) + 1;
+    }
+    const segSummary = Object.entries(typeCounts).map(([type, count]) => `${count} ${type}`).join(", ");
+    lines.push(`Segments: ${ann.segments.length} total (${segSummary})`);
+    const highlights = ann.segments.filter((s) => s.type === "highlight");
+    if (highlights.length > 0) {
+      const hlTimes = highlights.map((h) => `${formatTimestamp(h.start)}-${formatTimestamp(h.end)}: ${h.label}`).join("; ");
+      lines.push(`Highlights: ${hlTimes}`);
+    }
+    if (ann.suggestedCuts.length > 0) {
+      const cuts = ann.suggestedCuts.map((c) => `${formatTimestamp(c.start)}-${formatTimestamp(c.end)} (${c.reason})`).join("; ");
+      lines.push(`Suggested cuts: ${cuts}`);
+    }
+    lines.push("Segments detail:");
+    for (const seg of ann.segments) {
+      lines.push(
+        `  ${formatTimestamp(seg.start)}-${formatTimestamp(seg.end)} [${seg.type}] "${seg.label}" (energy:${seg.energy.toFixed(1)}, relevance:${seg.relevance.toFixed(1)})`
+      );
+    }
+    lines.push("");
+  }
+  return lines.join("\n");
+}
+function buildSystemPrompt(project, mediaContexts = [], assetLibrary = [], voices = []) {
+  const halfHeightPct = Math.round(project.height / 2 / project.width * 100);
+  const thirdHeightPct = Math.round(project.height / 3 / project.width * 100);
+  const mediaContextSection = buildMediaContextSection(mediaContexts);
+  return `You are a motion-graphics designer creating polished video animations.
+
+## Workflow
+
+- **Full video request**: Plan the scene (duration, layers, layout, animations), then execute all tool calls.
+- **Spot request** (add/edit one element): Skip planning — apply best practices and fulfill the ask directly.
+
+Steps for full requests: present brief plan → execute tools → confirm.
+
+## Coordinate System (CRITICAL)
+
+Canvas: ${project.width}x${project.height}px | Duration: ${project.duration}s | FPS: ${project.fps}
+
+ALL positions, sizes, font sizes, radii, and stroke widths are **% of project width** (NOT pixels).
+- (0,0) = canvas center (with default positionOrigin: center)
+- X range: -50 (left edge) to +50 (right edge)
+- Y range: -${halfHeightPct} (top) to +${halfHeightPct} (bottom)
+- width:50 = half the canvas width
+- fontSize:4 = text ~4% of canvas width (e.g., ~29px on 720px canvas)
+- All percentage values scale proportionally regardless of canvas size
+
+transform.positionOrigin: shifts the (0,0) reference point on the canvas.
+Default "center" makes (0,0) the canvas center. Use "top-left" to make (0,0) the top-left corner.
+
+## Tools
+
+- **upsert_layer**: Create a layer (omit layerId; layer.type+props required) or patch-edit an existing one (pass layerId; provide only what changes, layer.type must match). Supports transform/style/timing and animation.preset/keyframes — use it to animate layers created by run_action too.
+- **remove_layer**: Delete layer by ID or name
+- **configure_project**: Set name, dimensions, duration, background
+- **group_layers** / **ungroup_layers**: Manage layer groups
+- **analyze_media**: Analyze video/audio via multimodal AI → transcript, segments, highlights, dead space, editing suggestions. Results persist on asset.
+- **compose_from_media**: Build edited cut from analyzed media. Modes: keep-highlights (keep only highlights), remove-dead-space (keep everything except silence/filler), custom (filter by segment type or exact time ranges).
+- **run_action**: AI-powered operations: generate-captions (transcribe audio→captions layer), dub-audio (AI voice cloning translation), extract-audio (rip audio from video), generate-video (Seedance 2.0 text→video), generate-image (AI text→image asset+layer), find-stock-image (real stock photo→asset+layer), generate-speech (ElevenLabs TTS).
+
+### Generated / Stock Media Workflow
+1. To get missing media, call **run_action** with actionId "generate-image", "generate-video", or "find-stock-image".
+2. The action creates the media layer for you. Use the returned layersAdded[].id for all subsequent edits.
+3. The returned layersAdded[].props contains the real media source (for images: props.src, props.fileKey, props.fileName, attribution). Treat those props as authoritative.
+4. Some actions update an existing layer instead of adding a new one. Use returned layersUpdated[].changes as authoritative for src/fileKey/contentDuration/timing after the action completes.
+5. Do not remove and recreate generated/stock media layers just to animate or resize them. Use **upsert_layer** with the returned layer ID and props/timing/transform/style/animation.
+6. Only create a new media layer manually when you already have a real src from the Asset Library, user attachment, user-provided URL, or a tool result.
+
+### Fade / Crossfade Rules
+- enterTime/exitTime only control hard visibility. They do NOT create fades.
+- For a fade-in, add opacity keyframes 0 → 1 over the fade window.
+- For a fade-out, add opacity keyframes 1 → 0 before exitTime.
+- For a crossfade, overlap the outgoing/incoming layers AND add opacity keyframes on both layers during the overlap.
+- Do not claim a layer fades or crossfades unless you created opacity keyframes or used a fade preset.
+
+### Media Editing Workflow
+1. Call **analyze_media** on a video/audio layer to get full analysis
+2. Review results and propose an editing strategy
+3. Call **compose_from_media** to split source into sequential segments
+4. Fine-tune segments with upsert_layer/remove_layer
+
+${designPlaybookSection()}
+
+### Backgrounds — capabilities
+- Gradients: linear (angle + color stops), radial, conic — multiple stops with transparency supported
+- Set project-wide background via configure_project (backgroundColor accepts hex string or gradient object)
+- For animated/ambient backgrounds, create shape layers with their own background property
+- Interpolation by property type: position/scale/opacity/rotation → continuous with easing;
+  colors → continuous (RGB); text content (props.content) → text (char/word-reveal);
+  booleans/enums → discrete (step-end/step-start/step-mid)
+
+${presetCatalogSection()}
+
+## Interpolation Reference
+
+- **continuous**: Smooth numeric/color transitions. Strategies: linear, ease-in, ease-out, ease-in-out, ease-in-quad, ease-out-quad, ease-in-out-quad, ease-in-back, ease-out-back, ease-in-bounce, ease-out-bounce, ease-in-elastic, ease-out-elastic
+- **discrete**: Instant jumps. Strategies: step-end (hold old value until keyframe), step-start (jump immediately), step-mid (jump at midpoint)
+- **quantized**: Smooth animation snapped to integers (strategy: integer) or grid (strategy: snap-grid + increment)
+- **text**: Character/word reveal (char-reveal, word-reveal). ONLY for props.content on text/terminal layers.
+
+## HTML Keyframe Syntax
+
+When editing HTML scenes directly, use \`data-dm-keyframe-<property>\` with
+\`time:value@strategy\`. The \`@strategy\` suffix maps to interpolation family:
+- \`@ease-out\`, \`@ease-in-out\`, etc. → continuous
+- \`@step-start\`, \`@step-end\`, \`@step-mid\` → discrete
+- \`@integer\`, \`@snap-grid\` → quantized (\`snap-grid\` uses \`data-dm-grid-increment\`)
+- \`@char-reveal\`, \`@word-reveal\` → text interpolation for \`props.content\`
+
+Examples:
+\`data-dm-keyframe-opacity="0:0@ease-out; 0.5:1@ease-out"\`
+\`data-dm-keyframe-props-content="0:@char-reveal; 1.4:Welcome@char-reveal"\`
+\`data-dm-word-separator=" "\` with \`@word-reveal\`
+\`data-dm-grid-increment="0.25"\` with \`@snap-grid\`
+
+For polished \`dm-text\` character/word reveals, add fragment transition attrs:
+\`transition-effects="fade,slide-up,blur"\`, \`transition-duration="350"\`,
+\`transition-easing="ease-in-quad"\`, \`slide-distance="2.7"\`,
+\`scale-from="0.5"\`, \`blur-amount="0.78"\`.
+
+## Layer Props (per type)
+
+Use these exact prop names with upsert_layer. Derived from the layer
+schemas — they are authoritative. \`[a|b]\` = allowed enum values; \`(num lo…hi)\`
+= numeric range. \`(text)\` props are also the element's child text in HTML.
+
+${compactLayerPropsReference()}
+
+## Transform Properties
+
+position: { x, y, z } — in % of project width
+scale: { x, y } — multiplier (1 = normal)
+rotation: { x, y, z } — radians
+anchor: pivot point for transforms (top-left, top-center, top-right, center-left, center, center-right, bottom-left, bottom-center, bottom-right)
+positionOrigin: canvas reference for (0,0) — same values as anchor. Default "center" = (0,0) at canvas center.
+
+## Layout Guidelines
+
+Three-zone distribution on a ${project.width}x${project.height} canvas:
+- Title zone: y ≈ -${thirdHeightPct} (upper third)
+- Content zone: y ≈ 0 (center)
+- Footer zone: y ≈ +${thirdHeightPct} (lower third)
+
+${assetLibrarySection(assetLibrary)}
+
+${assetProtocolSection("chat")}
+
+${narrationWorkflowSection("chat")}
+
+${voiceCatalogSection(voices)}
+
+## Rules
+
+1. Plan first, then execute (skip planning for single-element requests)
+2. Always use rich, intentional backgrounds
+3. Every property value must be a deliberate choice
+4. Plain text only — no markdown formatting in responses
+5. Suggest analyze_media if project has unanalyzed video/audio layers
+6. Use layer IDs returned by upsert_layer for subsequent edit/remove calls
+7. Follow the Asset Rules — never fabricate media URLs; ask for or generate missing assets
+
+## Reference Example
+
+Below is a valid project showing correct structure and percentage-based values.
+Note: the example shows fontSize as % of project width, positions as %, etc.
+Use it to understand the schema shape, not as a source of literal values.
+
+${JSON.stringify(exampleProject)}
+
+${mediaContextSection}`;
+}
+const GenerateRequestSchema = z.object({
+  project: ProjectSchema,
+  modelId: z.string().describe("Model ID to use for generation"),
+  messages: z.array(z.custom())
+});
+const MAX_IMAGE_SIZE = 200 * 1024;
+function getOpenRouterClient(userId, projectId) {
+  if (env.DESKTOP_MODE) {
+    throw new Error("Not supported in desktop mode");
+  }
+  const apiKey = env.OPENROUTER_API_KEY;
+  if (!apiKey) {
+    throw new Error("OPENROUTER_API_KEY is not configured");
+  }
+  return createOpenRouter({
+    apiKey,
+    ...env.HELINE_AUTH ? {
+      baseURL: "https://openrouter.helicone.ai/api/v1",
+      headers: {
+        "Helicone-Auth": `Bearer ${env.HELINE_AUTH}`,
+        "Helicone-User-Id": userId,
+        "Helicone-Property-Project": projectId ?? ""
+      }
+    } : {}
+  });
+}
+function logAIInteraction(data) {
+  try {
+    const logsDir = join(process.cwd(), "logs", "ai-chat");
+    if (!existsSync(logsDir)) {
+      mkdirSync(logsDir, { recursive: true });
+    }
+    const filename = `${data.timestamp.replace(/[:.]/g, "-")}.json`;
+    const filepath = join(logsDir, filename);
+    const logData = {
+      timestamp: data.timestamp,
+      model: {
+        id: data.modelId,
+        name: data.modelName
+      },
+      systemPrompt: {
+        length: data.systemPromptLength
+      },
+      messages: data.userMessages.map((msg, idx) => {
+        const content = JSON.stringify(msg);
+        return {
+          index: idx,
+          role: msg.role,
+          contentPreview: content.substring(0, 200) + (content.length > 200 ? "..." : ""),
+          fullMessage: msg
+        };
+      }),
+      toolCalls: data.toolCalls?.length ? {
+        count: data.toolCalls.length,
+        calls: data.toolCalls
+      } : void 0,
+      response: data.responseMessages?.length ? {
+        messageCount: data.responseMessages.length,
+        messages: data.responseMessages
+      } : void 0,
+      usage: data.usage
+    };
+    writeFileSync(filepath, JSON.stringify(logData, null, 2), "utf-8");
+    console.log(`
+${"=".repeat(80)}`);
+    console.log(`[AI Chat] ${data.timestamp}`);
+    console.log(`${"=".repeat(80)}`);
+    console.log(`Model: ${data.modelName} (${data.modelId})`);
+    console.log(`System Prompt: ${data.systemPromptLength} chars`);
+    console.log(`User Messages: ${data.userMessages.length}`);
+    if (data.toolCalls?.length) {
+      console.log(`Tool Calls: ${data.toolCalls.length}`);
+    }
+    if (data.usage) {
+      console.log(
+        `Tokens: ${data.usage.totalTokens || "N/A"} (input: ${data.usage.inputTokens || "N/A"}, output: ${data.usage.outputTokens || "N/A"})`
+      );
+    }
+    console.log(`Log file: ${filepath}`);
+    console.log(`${"=".repeat(80)}
+`);
+  } catch (err) {
+    console.error("[AI] Failed to log interaction:", err);
+  }
+}
+function validateImageSizes(messages) {
+  for (const message of messages) {
+    if (!message.parts) continue;
+    for (const part of message.parts) {
+      if ("image" in part && typeof part.image === "string") {
+        const base64Data = part.image.split(",")[1];
+        if (base64Data) {
+          const sizeInBytes = Math.floor(base64Data.length * 3 / 4);
+          if (sizeInBytes > MAX_IMAGE_SIZE) {
+            error(
+              413,
+              `Image size (${Math.round(sizeInBytes / 1024)}KB) exceeds maximum allowed size (${MAX_IMAGE_SIZE / 1024}KB)`
+            );
+          }
+        }
+      }
+    }
+  }
+}
+function filterImagesFromOldMessages(messages) {
+  let lastImageMessageIndex = -1;
+  for (let i = messages.length - 1; i >= 0; i--) {
+    if (messages[i].role === "user" && messages[i].parts?.some((p) => p.type === "file")) {
+      lastImageMessageIndex = i;
+      break;
+    }
+  }
+  return messages.map((msg, idx) => {
+    if (msg.role !== "user" || idx === lastImageMessageIndex) {
+      return msg;
+    }
+    const hasImage = msg.parts?.some((p) => p.type === "file");
+    if (!hasImage) return msg;
+    return {
+      ...msg,
+      parts: msg.parts?.map(
+        (part) => part.type === "file" ? { type: "text", text: "(image removed to reduce context size)" } : part
+      )
+    };
+  });
+}
+async function getMediaContexts(project, db) {
+  const mediaLayers = project.layers.filter((l) => l.type === "video" || l.type === "audio");
+  if (mediaLayers.length === 0) return [];
+  const fileKeyMap = /* @__PURE__ */ new Map();
+  for (const layer of mediaLayers) {
+    const fileKey = layer.props.fileKey;
+    if (fileKey && !fileKeyMap.has(fileKey)) {
+      fileKeyMap.set(fileKey, { layerName: layer.name, layerType: layer.type });
+    }
+  }
+  if (fileKeyMap.size === 0) return [];
+  const fileKeys = [...fileKeyMap.keys()];
+  const assets = await db.select({
+    storageKey: asset.storageKey,
+    annotation: asset.annotation
+  }).from(asset).where(inArray(asset.storageKey, fileKeys));
+  const annotationMap = /* @__PURE__ */ new Map();
+  for (const a of assets) {
+    annotationMap.set(a.storageKey, a.annotation);
+  }
+  return fileKeys.map((fileKey) => ({
+    layerName: fileKeyMap.get(fileKey).layerName,
+    layerType: fileKeyMap.get(fileKey).layerType,
+    fileKey,
+    annotation: annotationMap.get(fileKey) ?? null
+  }));
+}
+async function getAssetLibrary(userId, db, limit = 30) {
+  const assets = await db.select({
+    name: asset.originalName,
+    mediaType: asset.mediaType,
+    url: asset.url,
+    storageKey: asset.storageKey,
+    duration: asset.duration,
+    annotation: asset.annotation
+  }).from(asset).where(eq(asset.userId, userId)).orderBy(desc(asset.createdAt)).limit(limit);
+  return assets.map((a) => ({
+    name: a.name,
+    mediaType: a.mediaType,
+    url: a.url,
+    fileKey: a.storageKey,
+    duration: a.duration,
+    summary: a.annotation?.summary ?? null
+  }));
+}
+async function getVoiceCatalog() {
+  const apiKey = env.ELEVENLABS_API_KEY;
+  if (!apiKey) return [];
+  try {
+    const { getTtsVoices } = await import('./tts-voices-B9J2ZzGy.js');
+    return await getTtsVoices(apiKey);
+  } catch (err) {
+    console.warn("[AI] Failed to load TTS voices for prompt:", err);
+    return [];
+  }
+}
+function formatDuration(seconds) {
+  const mins = Math.floor(seconds / 60);
+  const secs = Math.round(seconds % 60);
+  return mins > 0 ? `${mins}m${secs.toString().padStart(2, "0")}s` : `${secs}s`;
+}
+const POST = async ({ request, locals }) => {
+  try {
+    let enableCacheControl = function(messages2) {
+      return messages2.map((message) => {
+        if (typeof message.content !== "string") {
+          return message;
+        }
+        return {
+          ...message,
+          providerOptions: {
+            ...message.providerOptions,
+            openrouter: {
+              ...message.providerOptions?.openrouter,
+              cacheControl: {
+                type: "ephemeral"
+                /* ttl: '1h' */
+              }
+            }
+          }
+        };
+      });
+    };
+    if (!locals.user) {
+      error(401, "Authentication required");
+    }
+    const user = locals.user;
+    const accessCheck = await canUserAccessAI(user.id, locals.db);
+    if (!accessCheck.allowed) {
+      error(403, accessCheck.reason || "AI access not enabled");
+    }
+    const body = await request.json();
+    const { project, modelId, messages } = GenerateRequestSchema.parse(body);
+    validateImageSizes(messages);
+    const filteredMessages = filterImagesFromOldMessages(messages);
+    const openrouter = getOpenRouterClient(user.email, project.id);
+    const [mediaContexts, assetLibrary, voices] = await Promise.all([
+      getMediaContexts(project, locals.db),
+      getAssetLibrary(user.id, locals.db),
+      getVoiceCatalog()
+    ]);
+    const systemPrompt = buildSystemPrompt(project, mediaContexts, assetLibrary, voices);
+    const model = getModel(modelId);
+    console.log(`[AI] Using model: ${model.name} (${model.id})`);
+    console.log(`[AI] System prompt length: ${systemPrompt.length} chars`);
+    const serverTools = {
+      analyze_media: tool({
+        description: "Analyze a video/audio layer with multimodal AI. Extracts transcript, scene segments, highlights, dead space, and editing suggestions. Results persist on the asset. Call before compose_from_media.",
+        inputSchema: AnalyzeMediaInputSchema,
+        execute: async ({ layerId, context }) => {
+          try {
+            const layer = project.layers.find(
+              (l) => l.id === layerId || l.name.toLowerCase() === layerId.toLowerCase()
+            );
+            if (!layer) {
+              const available = project.layers.map((l) => `"${l.name}" (id: ${l.id})`).join(", ");
+              return {
+                success: false,
+                message: `Layer "${layerId}" not found. Available: ${available}`,
+                error: "Layer not found"
+              };
+            }
+            if (layer.type !== "video" && layer.type !== "audio") {
+              return {
+                success: false,
+                message: `Layer "${layer.name}" is a ${layer.type} layer, not video/audio`,
+                error: "Not a media layer"
+              };
+            }
+            const fileKey = layer.props.fileKey;
+            if (!fileKey) {
+              return {
+                success: false,
+                message: `Layer "${layer.name}" has no media file uploaded`,
+                error: "No fileKey"
+              };
+            }
+            console.log(`[AI] Analyzing media: ${fileKey}`);
+            const result2 = await analyzeAndSaveMediaAsset(fileKey, context, user, locals.db);
+            if (!result2.success) {
+              return { success: false, message: result2.error, error: result2.error };
+            }
+            const annotation = result2.annotation;
+            const highlights = annotation.segments.filter((s) => s.type === "highlight");
+            const deadSpace = annotation.segments.filter((s) => s.type === "dead-space");
+            console.log(
+              `[AI] Media analysis complete: ${annotation.segments.length} segments, ${annotation.suggestedCuts.length} suggested cuts`
+            );
+            return {
+              success: true,
+              message: `Analyzed ${formatDuration(annotation.duration)} of ${annotation.modelUsed.split("/").pop()}. Found ${annotation.segments.length} segments (${highlights.length} highlights, ${deadSpace.length} dead space). ${annotation.suggestedCuts.length} cuts suggested.`,
+              summary: annotation.summary,
+              segmentCount: annotation.segments.length,
+              highlightCount: highlights.length,
+              deadSpaceCount: deadSpace.length,
+              suggestedCutsCount: annotation.suggestedCuts.length,
+              duration: annotation.duration,
+              topics: annotation.topics
+            };
+          } catch (err) {
+            console.error("[AI] Media analysis error:", err);
+            return {
+              success: false,
+              message: "Failed to analyze media",
+              error: err instanceof Error ? err.message : "Unknown error"
+            };
+          }
+        }
+      })
+    };
+    const allTools = { ...animationTools, ...serverTools };
+    const agent = new ToolLoopAgent({
+      model: openrouter(model.id, {
+        reasoning: {
+          effort: model.effort
+        }
+      }),
+      // Explicit output budget: without it some providers default to a few
+      // thousand tokens and long steps (reasoning + several tool calls) get
+      // truncated mid-tool-input, stranding the call in input-streaming.
+      ...model.maxOutputTokens ? { maxOutputTokens: model.maxOutputTokens } : {},
+      providerOptions: {
+        openrouter: {
+          only: model.providerOnly
+        }
+      },
+      instructions: {
+        role: "system",
+        content: systemPrompt,
+        providerOptions: {
+          openrouter: {
+            cacheControl: {
+              type: "ephemeral"
+            },
+            sort: "price"
+          }
+        }
+      },
+      tools: allTools,
+      async onFinish(event) {
+        logAIInteraction({
+          timestamp: (/* @__PURE__ */ new Date()).toISOString(),
+          modelId: model.id,
+          modelName: model.name,
+          systemPromptLength: systemPrompt.length,
+          userMessages: messages,
+          toolCalls: event.toolCalls,
+          responseMessages: event.response.messages,
+          requestData: event.request,
+          usage: event.usage
+        });
+        if (event.usage) {
+          await logAIUsage(
+            {
+              userId: user.id,
+              modelId: model.id,
+              promptTokens: event.usage.inputTokens || 0,
+              completionTokens: event.usage.outputTokens || 0,
+              metadata: {
+                projectName: project.name,
+                messageCount: messages.length,
+                toolCallCount: event.toolCalls?.length || 0
+              }
+            },
+            locals.db
+          );
+        }
+        await trackEvent(user.id, "ai_chat_message", {
+          modelId: model.id,
+          modelName: model.name,
+          projectId: project.id,
+          messageCount: messages.length,
+          toolCallCount: event.toolCalls?.length || 0,
+          inputTokens: event.usage?.inputTokens || 0,
+          outputTokens: event.usage?.outputTokens || 0,
+          plan: user.subscriptionTier
+        });
+      }
+    });
+    const result = await agent.stream({
+      messages: [
+        ...enableCacheControl([...await convertToModelMessages(filteredMessages)]),
+        {
+          role: "system",
+          content: JSON.stringify(project)
+        }
+      ]
+    });
+    return result.toUIMessageStreamResponse();
+  } catch (err) {
+    if (isHttpError(err)) {
+      throw err;
+    }
+    console.error("[AI] Error generating animation:", err);
+    if (err instanceof z.ZodError) {
+      error(400, `Validation error: ${err.message}`);
+    }
+    error(500, err instanceof Error ? err.message : "Failed to generate animation");
+  }
+};
+
+export { POST };
+//# sourceMappingURL=_server.ts-CGs8hOo6.js.map