@lightcone-ai/daemon 0.16.2 → 0.17.1

package/mcp-servers/official/media-tools/index.js

@@ -5,6 +5,56 @@ import { z } from 'zod';
 
 import { addTitleEffects } from './lib/render.js';
 import { SUPPORTED_PRESETS } from './lib/presets.js';
+import { runSynthesisTtsTool } from '../../../src/tools/synthesize-tts.js';
+import { runPlanVideoSegmentsTool } from '../../../src/tools/plan-video-segments.js';
+import { runComposeVideoV2Tool } from '../../../src/tools/compose-video-v2.js';
+import { runRecordUrlNarrationTool } from '../../../src/tools/record-url-narration.js';
+import { runRenderTextToImageTool } from '../../../src/tools/render-text-to-image.js';
+import { runRenderHtmlToImageTool } from '../../../src/tools/render-html-to-image.js';
+import { runTakePageScreenshotTool } from '../../../src/tools/take-page-screenshot.js';
+import { hasFreshVideoBrief } from '../../../src/video-brief-flag.js';
+import { lightconeApi, CURRENT_WORKSPACE_ID, CURRENT_AGENT_ID } from './lib/lightcone-api.js';
+
+const WORKSPACE_DIR = String(process.env.WORKSPACE_DIR ?? '');
+
+// CVMAX editor_in_chief block: in one workspace, the editor_in_chief agent
+// must not run video production tools directly (short_video_scripter owns
+// that role). Env-gated so ops can rotate workspace/agent IDs without code
+// changes. Previously lived in chat-bridge; moved here alongside the tool
+// it gates (V4 migration). submit_to_library, the other gated tool, still
+// lives in chat-bridge and keeps its own copy of this check.
+const CVMAX_WORKSPACE_ID = String(process.env.BLOCKED_EDITOR_WORKSPACE_ID ?? '');
+const CVMAX_EDITOR_IN_CHIEF_AGENT_ID = String(process.env.BLOCKED_EDITOR_AGENT_ID ?? '');
+
+function isBlockedCvmaxEditorVideoTool(toolName) {
+  return CURRENT_WORKSPACE_ID === CVMAX_WORKSPACE_ID
+    && CURRENT_AGENT_ID === CVMAX_EDITOR_IN_CHIEF_AGENT_ID
+    && CVMAX_WORKSPACE_ID
+    && CVMAX_EDITOR_IN_CHIEF_AGENT_ID
+    && toolName === 'record_url_narration';
+}
+
+function cvmaxEditorVideoToolError(toolName) {
+  return {
+    isError: true,
+    content: [{
+      type: 'text',
+      text:
+        `Error: ${toolName} blocked for editor_in_chief in CvMax. `
+        + 'In this workspace, @short_video_scripter owns video production. '
+        + 'editor_in_chief may route, review, or assist with OCR/verification, but must not run video production tools directly.',
+    }],
+  };
+}
+
+// Session-scoped flag set when plan_video_segments runs successfully.
+// compose_video_v2 refuses TTS-bearing segments (any segment with audio_path)
+// unless this is true — the agent must route audio through plan_video_segments
+// first so durations / subtitle_text are mechanically aligned. media-tools is
+// spawned per-agent, so a fresh agent session must call plan_video_segments
+// fresh. (This flag previously lived in chat-bridge module scope; moved here
+// alongside the tools it gates, see V2 migration.)
+let _planVideoSegmentsCalledThisSession = false;
 
 const PRESET_ENUM = z.enum(SUPPORTED_PRESETS);
 const POSITION_ENUM = z.enum(['top', 'center', 'bottom']);
@@ -63,6 +113,244 @@ server.tool(
   }
 );
 
+// ── synthesize_tts (migrated from chat-bridge) ────────────────────────────
+// Pure TTS atomic tool: text → mp3 file. The lightcone server proxies to
+// MiniMax TTS; this tool downloads the resulting mp3 to a local tmp path so
+// downstream tools (plan_video_segments / compose_video_v2) can read it.
+//
+// Per the video-synthesis-design migration (see docs/upload-pipeline-design.md
+// and docs/scenario-content-creation/video-synthesis-design.md), this tool
+// lives in media-tools rather than chat-bridge so the video pipeline is a
+// single coherent MCP server.
+server.tool(
+  'synthesize_tts',
+  'Run MiniMax TTS on a snippet of narration text and download the resulting mp3 to a local tmp path. '
+  + 'Returns the local path and duration. Call once per video segment — do not concatenate all narration '
+  + 'into a single call (segment-level audio is required for plan_video_segments to align video durations).',
+  {
+    text: z.string().min(1).describe('Narration text for this segment. Will be synthesized as a single mp3.'),
+    voice_id: z.string().optional().describe('TTS voice preset. Omit to use workspace default.'),
+    workspace_id: z.string().optional().describe('Target workspace. Defaults to the current workspace.'),
+  },
+  async ({ text, voice_id, workspace_id }) => runSynthesisTtsTool({
+    text,
+    voice_id,
+    workspace_id,
+    currentWorkspaceId: CURRENT_WORKSPACE_ID,
+    api: lightconeApi,
+  })
+);
+
+// ── plan_video_segments (migrated from chat-bridge; TTS decoupled) ────────
+// Pure planner — takes per-segment {text, audio_path, visual_kind, ...} and
+// returns segments with audio_duration_ms / presentation.duration / dwell_ms
+// / subtitle_text filled in. Caller MUST run synthesize_tts per segment first
+// and pass the resulting audio_path here. No longer synthesizes TTS itself
+// (V2 migration; see docs/scenario-content-creation/video-synthesis-design.md).
+server.tool(
+  'plan_video_segments',
+  'Universal audio-video sync planning step. For each segment, reads the supplied audio_path via ffprobe, '
+  + 'measures audio duration, and returns a planned segments array with audio_path / audio_duration_ms / '
+  + 'subtitle_text / presentation.duration / dwell_ms filled in — ready to pass directly to both '
+  + 'record_url_narration (as the recording plan) AND compose_video_v2 (as the segment list). '
+  + 'Must be called before compose_video_v2 when any segment has audio_path.\n\n'
+  + 'Inputs per segment: {text, audio_path (required, from synthesize_tts), visual_kind, visual_path or visual_paths, '
+  + 'optionally transition / presentation.style}. Standard chain: synthesize_tts × N → plan_video_segments → '
+  + 'record_url_narration + compose_video_v2 (both use the same plan output).',
+  {
+    segments: z.array(z.object({
+      text: z.string().describe('Narration text for this segment — used as subtitle_text in the output.'),
+      audio_path: z.string().describe('Absolute path to the segment\'s mp3 (from synthesize_tts).'),
+      visual_kind: z.enum(['image', 'video', 'gif', 'carousel']).describe('Type of visual for compose_video_v2.'),
+      visual_path: z.string().optional().describe('Absolute path to a single image / video / gif file.'),
+      visual_paths: z.array(z.string()).optional().describe('For carousel: array of image paths, one per card.'),
+      transition: z.enum(['cut', 'fade', 'crossfade']).optional().describe('Transition to next segment. Default cut.'),
+      presentation: z.object({
+        style: z.enum(['static', 'scroll']).optional(),
+      }).optional().describe('Optional presentation hints (style only). duration/per_card_duration are computed.'),
+      dwell_ms: z.number().optional().describe('Optional override for record_url_narration phase duration. Default = audio_duration_ms.'),
+    })).describe('Segments to plan. audio_path is required for each.'),
+  },
+  async ({ segments }) => {
+    const result = await runPlanVideoSegmentsTool({ segments });
+    if (!result?.isError) _planVideoSegmentsCalledThisSession = true;
+    return result;
+  }
+);
+
+// ── compose_video_v2 (migrated from chat-bridge) ──────────────────────────
+// Tool-level enforcement of the standard chain: TTS-bearing segments require
+// plan_video_segments to have run earlier in this session. Without it manual
+// dwell/duration math has repeatedly produced misaligned subtitles, silent
+// tails, and re-records (Task #25/#26 trial).
+server.tool(
+  'compose_video_v2',
+  'Compose a video from a list of segments using ffmpeg. Each segment has a visual source (image / scroll / '
+  + 'carousel / video / gif), optional audio, and optional subtitle text. Subtitles are burned in when '
+  + 'subtitle_text is provided. Segments are concatenated in order; outro clips are appended after.\n\n'
+  + 'When any segment has audio_path, MUST be preceded by plan_video_segments in the same session '
+  + '(plan_video_segments fills duration/subtitle_text/audio_path mechanically; manual alignment is rejected). '
+  + 'Returns a local mp4 path + size_bytes.',
+  {
+    segments: z.array(z.object({
+      visual_path: z.string().optional().describe('Absolute path to a single image / video / gif.'),
+      visual_paths: z.array(z.string()).optional().describe('For carousel: array of image paths, one per card.'),
+      visual_kind: z.enum(['image', 'video', 'gif', 'carousel']).describe('Type of visual.'),
+      presentation: z.object({
+        style: z.enum(['static', 'scroll']).optional(),
+        duration: z.number().optional().describe('Segment duration in seconds. Required for image/scroll.'),
+        per_card_duration: z.number().optional().describe('Seconds per card for carousel.'),
+      }).optional(),
+      audio_path: z.string().nullable().optional().describe('Absolute path to audio (mp3). null/omit for silence.'),
+      subtitle_text: z.string().optional().describe('Narration text to burn as subtitle. Displayed for the full segment duration.'),
+      transition: z.enum(['cut', 'fade', 'crossfade']).optional().describe('Transition to next segment. Default cut.'),
+    })).describe('Ordered list of video segments.'),
+    outro_paths: z.array(z.string()).optional().describe('Absolute paths to outro video clips appended at end.'),
+    resolution: z.string().optional().describe('Output resolution WxH. Default "1080x1920".'),
+    output_path: z.string().optional().describe('Absolute output path. Auto-generated if omitted.'),
+  },
+  async (args) => {
+    const segments = Array.isArray(args?.segments) ? args.segments : [];
+    const hasNarration = segments.some(s => typeof s?.audio_path === 'string' && s.audio_path.trim());
+    if (hasNarration && !_planVideoSegmentsCalledThisSession) {
+      return toolError(
+        'compose_video_v2 refused: TTS-bearing segments (audio_path present) require plan_video_segments '
+        + 'to have run earlier in this session — it mechanically aligns audio_duration / video_duration / '
+        + 'subtitle_text with a safety buffer. Manual dwell/duration math has repeatedly produced misaligned '
+        + 'subtitles and silent tails that force re-recording.\n\n'
+        + 'Standard chain: synthesize_tts × N (per segment) → plan_video_segments(segments with text + audio_path + '
+        + 'visual_kind + visual_path) → compose_video_v2 (use the returned segments verbatim, only swap '
+        + 'visual_path/visual_kind for the real media). Call plan_video_segments now and pass its output here.'
+      );
+    }
+    if (hasNarration && !hasFreshVideoBrief({ workspaceId: CURRENT_WORKSPACE_ID, agentId: CURRENT_AGENT_ID })) {
+      return toolError(
+        'compose_video_v2 refused: must send a 确认稿 (production-brief) to the user via send_message before '
+        + 'compositing a narration video. The system scans send_message content for a brief — a message that '
+        + 'BOTH asks the user to confirm (确认 / 你看 / OK 吗 / 可以吗 / 同意 / 通过 / 行不行) AND describes '
+        + 'at least two of: 画面 / 时长 / 文案 / 口播 / 字幕 / 顺序 / 口吻 / 分镜 / 配音 — no such message '
+        + 'was sent in the last 6 hours for this workspace+agent.\n\n'
+        + '"已生成 TTS" / "开始合成" / progress reports do NOT count. Send a concrete confirmation draft '
+        + 'first (e.g. "我准备这么做：画面是真录屏，时长约 1 分钟，文案如下…，字幕开启，公司顺序 A→B→C，'
+        + '口吻是…—— 你 OK 吗？") and wait for the user to reply OK before calling compose_video_v2 again.'
+      );
+    }
+    return runComposeVideoV2Tool({ ...args, workspaceDir: WORKSPACE_DIR });
+  }
+);
+
+// ── record_url_narration (migrated from chat-bridge) ──────────────────────
+// Records a silent mp4 of a URL via Chromium+Xvfb+Playwright recordVideo,
+// driven by a beat-by-beat plan. Hard-block: requires plan_video_segments to
+// have run in this session — hand-written dwell_ms has drifted from TTS
+// audio in production runs (Tasks #20/#25/#26), forcing re-records.
+server.tool(
+  'record_url_narration',
+  'Record a silent video of a URL by driving Chromium on an Xvfb display and capturing it with Playwright recordVideo, driven by a video plan; ffmpeg then transcodes the recording to mp4. Outputs a silent mp4 that can be passed to compose_video_v2 as a video-kind segment with an audio_path for narration.\n\nUse this as the canonical recording step for URL-narration videos. Falls back: if the page needs interactions outside the visual_action vocabulary (clicks, waits, OCR loops), use Monitor (Bash) with custom Playwright instead.\n\nMUST be preceded by plan_video_segments in the same session — feed plan_video_segments\'s `segments` array as `plan.sections` so dwell_ms aligns mechanically with TTS audio_duration_ms (hand-written dwell_ms has drifted and forced re-records in production).\n\nRuntime requirements: this tool only works on a Linux daemon machine with Xvfb + Chromium + ffmpeg installed (ffmpeg is used to transcode the recording to mp4; no x11grab device support needed). macOS / Windows daemons will fail at startup.',
+  {
+    url: z.string().describe('Page URL to record'),
+    plan: z.record(z.any()).describe(
+      'A video plan: an object with `phases` (or `sections`), each a "visual beat" with '
+      + '`action` (scroll_to_dwell / linear_scroll_during / scroll_back / hold / ...), a '
+      + 'target (`target_y` or `focus_region:[y1,y2]`) for scroll-type actions, and '
+      + '`dwell_ms` (how long to hold that beat — should match the segment\'s TTS duration).\n\n'
+      + 'Standard chain: pass plan_video_segments\'s `segments` array directly as `plan.sections` — '
+      + 'each segment\'s `dwell_ms` is already set to its `audio_duration_ms`.\n\n'
+      + 'For RECRUITMENT URLs (mp.weixin.qq.com / 校招 / 实习 / 岗位 content), each section MUST '
+      + 'also declare `target_y_content_label` — a short Chinese label describing what content '
+      + 'sits at that pixel y position on the page (e.g. "标题区" / "岗位信息卡片" / "公司介绍" / '
+      + '"届别说明"). Labels matching forbidden regions ("二维码" / "扫码" / "投递入口" / "投递方式" / '
+      + '"联系方式" / "微信号" / "QR" / "阅读原文" / "外链") will cause the tool to refuse the '
+      + 'recording — recruitment content must NOT dwell on these areas (see fragments.md '
+      + 'frag.short.recruitment_url_mode_policy). Pick a different target_y in the 标题/岗位 '
+      + 'information area and rewrite that section.'
+    ),
+    output_path: z.string().optional().describe('Workspace-relative output mp4 path. Default tmp/wx3_video/recorded-{ts}.mp4'),
+    events_path: z.string().optional().describe('Workspace-relative events.json path. Default ${output_path}.events.json'),
+    viewport: z.object({
+      width: z.number().optional(),
+      height: z.number().optional(),
+    }).optional().describe('Default 1080x1920 (mobile portrait). Override only if the plan requires a different shape.'),
+    fps: z.number().optional().describe('Default 30. Do not change unless needed.'),
+    settle_ms: z.number().optional().describe('Default 4000. Settle wait after navigation before recording starts.'),
+  },
+  async (args) => {
+    if (isBlockedCvmaxEditorVideoTool('record_url_narration')) {
+      return cvmaxEditorVideoToolError('record_url_narration');
+    }
+    // record_url_narration is part of the narration-video pipeline (paired
+    // with synthesize_tts + plan_video_segments + compose_video_v2), so it
+    // requires the same 确认稿 gate as compose_video_v2 — catching the skip
+    // earlier saves TTS + recording time.
+    if (!hasFreshVideoBrief({ workspaceId: CURRENT_WORKSPACE_ID, agentId: CURRENT_AGENT_ID })) {
+      return {
+        isError: true,
+        content: [{ type: 'text', text:
+          'Error: record_url_narration refused: must send a 确认稿 (production-brief) to the user via '
+          + 'send_message before starting a narration recording. The system scans send_message content for '
+          + 'a brief — a message that BOTH asks the user to confirm (确认 / 你看 / OK 吗 / 可以吗 / 同意 / '
+          + '通过 / 行不行) AND describes at least two of: 画面 / 时长 / 文案 / 口播 / 字幕 / 顺序 / 口吻 / '
+          + '分镜 / 配音 — no such message was sent in the last 6 hours for this workspace+agent.\n\n'
+          + '"已生成 TTS" / "开始合成" / progress reports do NOT count. Send a concrete confirmation draft '
+          + 'first (e.g. "我准备这么做：画面是真录屏，时长约 1 分钟，文案如下…，字幕开启，公司顺序 A→B→C，'
+          + '口吻是…—— 你 OK 吗？") and wait for the user to reply OK before calling record_url_narration.'
+        }],
+      };
+    }
+    return runRecordUrlNarrationTool({
+      args,
+      currentWorkspaceId: CURRENT_WORKSPACE_ID,
+      workspaceDir: WORKSPACE_DIR,
+      planVideoSegmentsCalled: _planVideoSegmentsCalledThisSession,
+    });
+  }
+);
+
+// ── render_text_to_image (migrated from chat-bridge) ──────────────────────
+server.tool(
+  'render_text_to_image',
+  'Render text content into image(s) for video synthesis. style=scroll produces a single tall image (for a scrolling video segment); style=carousel produces one image per card (for a slide-show segment). Returns local file paths.',
+  {
+    content: z.union([z.string(), z.array(z.string())]).describe('Text content. For carousel, pass an array of strings — one per card. For scroll, pass a single string (or array joined with line breaks).'),
+    style: z.enum(['scroll', 'carousel']).describe('scroll: one tall image; carousel: one image per card.'),
+    theme: z.enum(['dark', 'light']).optional().describe('Color theme. Default dark.'),
+    width: z.number().optional().describe('Image width in pixels. Default 1080.'),
+    card_height: z.number().optional().describe('Card height in pixels (carousel) or viewport height (scroll baseline). Default 1920.'),
+    font_size: z.number().optional().describe('Base font size in pixels. Default 48.'),
+  },
+  async (args) => runRenderTextToImageTool(args)
+);
+
+// ── render_html_to_image (migrated from chat-bridge) ──────────────────────
+server.tool(
+  'render_html_to_image',
+  'Render a raw HTML string to a PNG image by navigating to it as a local file:// page. Unlike evaluate_script+document.write on about:blank, this preserves file:// origin so <img src="file:///..."> references load correctly. Returns the output image path.',
+  {
+    html: z.string().describe('Full HTML document to render (including <!doctype>, <html>, <head>, <body>).'),
+    output_path: z.string().optional().describe('Absolute path to save the PNG. Auto-generated in /tmp if omitted.'),
+    viewport_width: z.number().optional().describe('Viewport width in pixels. Default 1080.'),
+    viewport_height: z.number().optional().describe('Viewport height in pixels. Default 1920.'),
+    wait_until: z.enum(['load', 'networkidle', 'domcontentloaded']).optional().describe('Navigation wait condition. Default load.'),
+  },
+  async (args) => runRenderHtmlToImageTool(args)
+);
+
+// ── take_page_screenshot (migrated from chat-bridge) ──────────────────────
+server.tool(
+  'take_page_screenshot',
+  'Open a URL with a headless browser and capture a screenshot. crop=above_fold captures only the visible viewport (ideal for thumbnail-style frames); crop=full_page captures the entire page height.',
+  {
+    url: z.string().describe('Page URL to screenshot.'),
+    crop: z.enum(['above_fold', 'full_page']).optional().describe('Capture mode. Default above_fold.'),
+    viewport: z.object({
+      width: z.number().optional(),
+      height: z.number().optional(),
+    }).optional().describe('Viewport size. Default 390×844 (mobile).'),
+    wait_for: z.enum(['load', 'networkidle', 'domcontentloaded']).optional().describe('Page load event to wait for before screenshotting. Default networkidle.'),
+  },
+  async (args) => runTakePageScreenshotTool(args)
+);
+
 const transport = new StdioServerTransport();
 await server.connect(transport);
 console.error('[official-media-tools] MCP Server started');

package/mcp-servers/official/media-tools/lib/lightcone-api.js

@@ -0,0 +1,41 @@
+// Minimal HTTP helper for media-tools to call lightcone server's internal API.
+// Wraps fetch with the right URL prefix + auth headers + JSON encoding.
+//
+// Unlike daemon/src/chat-bridge.js's `api`, this helper does NOT route through
+// the governance/cache layer — media-tools is a separate stdio MCP server and
+// governance integration is chat-bridge-specific. If a tool here needs
+// governance-mediated execution, route it through chat-bridge's thin-proxy
+// instead (see weixin-tools for the pattern).
+
+const SERVER_URL = String(process.env.SERVER_URL ?? '').replace(/\/+$/, '');
+const MACHINE_API_KEY = String(process.env.MACHINE_API_KEY ?? '');
+const AGENT_ID = String(process.env.AGENT_ID ?? '');
+
+if (!SERVER_URL) throw new Error('media-tools: SERVER_URL env var is required');
+if (!MACHINE_API_KEY) throw new Error('media-tools: MACHINE_API_KEY env var is required');
+if (!AGENT_ID) throw new Error('media-tools: AGENT_ID env var is required');
+
+export async function lightconeApi(method, apiPath, body) {
+  const url = `${SERVER_URL}/internal/agent/${encodeURIComponent(AGENT_ID)}${apiPath}`;
+  const res = await fetch(url, {
+    method,
+    headers: {
+      'Content-Type': 'application/json',
+      'Authorization': `Bearer ${MACHINE_API_KEY}`,
+    },
+    body: body != null ? JSON.stringify(body) : undefined,
+  });
+  if (!res.ok) {
+    let text = '';
+    try { text = await res.text(); } catch { /* ignore */ }
+    const err = new Error(`lightcone ${method} ${apiPath} → ${res.status}: ${text.slice(0, 400)}`);
+    err.status = res.status;
+    err.body = text;
+    throw err;
+  }
+  return res.json();
+}
+
+// Exposed so tools can construct workspace-aware fallback identifiers.
+export const CURRENT_AGENT_ID = AGENT_ID;
+export const CURRENT_WORKSPACE_ID = String(process.env.WORKSPACE_ID ?? '');

package/mcp-servers/official/media-tools/manifest.json

@@ -5,7 +5,14 @@
   "runtime": "node",
   "entrypoint": "index.js",
   "tool_declarations": [
-    { "name": "add_title_effects", "classification": "cacheable" }
+    { "name": "add_title_effects", "classification": "cacheable" },
+    { "name": "synthesize_tts", "classification": "mandatory" },
+    { "name": "plan_video_segments", "classification": "mandatory" },
+    { "name": "compose_video_v2", "classification": "mandatory" },
+    { "name": "record_url_narration", "classification": "mandatory" },
+    { "name": "render_text_to_image", "classification": "cacheable" },
+    { "name": "render_html_to_image", "classification": "cacheable" },
+    { "name": "take_page_screenshot", "classification": "cacheable" }
   ],
   "smoke_test": {
     "tool": "add_title_effects",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lightcone-ai/daemon",
-  "version": "0.16.2",
+  "version": "0.17.1",
   "type": "module",
   "main": "src/index.js",
   "bin": {

package/src/chat-bridge.js

@@ -2,23 +2,21 @@
 import { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js';
 import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';
 import { z } from 'zod';
-import { createReadStream, existsSync, mkdirSync, readFileSync, writeFileSync, statSync } from 'fs';
+import { existsSync, mkdirSync, readFileSync, writeFileSync, statSync } from 'fs';
 import { createHash, randomUUID } from 'crypto';
 import path, { extname } from 'path';
 import os from 'os';
-import { recordUrlNarration } from './_vendor/video/recorder/index.js';
 import { writeLocalFileToWorkspace, resolveWorkspaceFileUploadPlan } from './workspace-file-upload.js';
 import { UploadJobManager } from './upload-job-manager.js';
 import { createUploadServerApi } from './upload-server-api.js';
-import { runRecordUrlNarrationTool } from './record-url-narration-tool.js';
+// record_url_narration moved to media-tools MCP server (V4 migration);
+// recorder import / handler are now consumed there, not from chat-bridge.
 import { runSubmitToLibraryTool } from './submit-to-library-tool.js';
-import { runRenderTextToImageTool } from './tools/render-text-to-image.js';
-import { runRenderHtmlToImageTool } from './tools/render-html-to-image.js';
-import { runSynthesisTtsTool } from './tools/synthesize-tts.js';
-import { runPlanVideoSegmentsTool } from './tools/plan-video-segments.js';
-import { runComposeVideoV2Tool } from './tools/compose-video-v2.js';
-import { runTakePageScreenshotTool } from './tools/take-page-screenshot.js';
+// render_text_to_image, render_html_to_image, take_page_screenshot moved to
+// media-tools MCP server (V5 migration). Handlers still live in
+// daemon/src/tools/ as shared modules and are imported there.
 import { runGetLibraryFileTool } from './tools/get-library-file.js';
+import { markVideoBriefSent, looksLikeVideoBrief } from './video-brief-flag.js';
 import { isLeaseInvalidated, clearInvalidatedLease } from './governance-state.js';
 import { classifyLeaseWindow } from './lease-window.js';
 import {
@@ -72,8 +70,9 @@ let currentWorkspaceId = WORKSPACE_ID;
 // Remove entirely once the new atomic tool framework is stable and the legacy pipeline retires.
 const CVMAX_WORKSPACE_ID = process.env.BLOCKED_EDITOR_WORKSPACE_ID ?? '';
 const CVMAX_EDITOR_IN_CHIEF_AGENT_ID = process.env.BLOCKED_EDITOR_AGENT_ID ?? '';
+// record_url_narration moved to media-tools and carries its own copy of this
+// block. submit_to_library stays here.
 const CVMAX_EDITOR_BLOCKED_VIDEO_TOOLS = new Set([
-  'record_url_narration',
   'submit_to_library',
 ]);
 
@@ -162,7 +161,6 @@ const DEFAULT_TOOL_CLASSIFICATION = {
   update_goal_field: 'mandatory',
   supersede_goal_field: 'mandatory',
   request_credential_auth: 'mandatory',
-  record_url_narration: 'mandatory',
   submit_to_library: 'mandatory',
   register_data_source: 'mandatory',
   bind_workspace_scenario: 'mandatory',
@@ -491,37 +489,6 @@ async function directApi(method, apiPath, body) {
   return res.json();
 }
 
-async function directApiVideoUpload(apiPath, {
-  localPath,
-  filename,
-  contentType = 'video/mp4',
-}) {
-  const url = `${SERVER_URL}/internal/agent/${AGENT_ID}${apiPath}`;
-  const headers = {
-    'Authorization': `Bearer ${MACHINE_API_KEY}`,
-    'Content-Type': contentType,
-  };
-  if (filename) headers['X-File-Name'] = filename;
-
-  let res;
-  try {
-    res = await fetch(url, {
-      method: 'POST',
-      headers,
-      body: createReadStream(localPath),
-      duplex: 'half',
-    });
-  } catch (error) {
-    throw buildDirectApiTransportError({ method: 'POST', apiPath, error });
-  }
-
-  if (!res.ok) {
-    const text = await res.text();
-    throw buildDirectApiHttpError({ method: 'POST', apiPath, status: res.status, text });
-  }
-  return res.json();
-}
-
 async function callGovernance(payload, { retry = true } = {}) {
   const attempts = retry ? 2 : 1;
   let lastError = null;
@@ -609,63 +576,6 @@ async function governanceRoundTrip({ method, apiPath, body, toolName, classifica
   return directApi(method, apiPath, nextBody);
 }
 
-async function runMandatoryLocalTool({ toolName, toolInput = {}, executor }) {
-  const classification = TOOL_CLASSIFICATION[toolName] ?? 'mandatory';
-  const traceId = randomUUID();
-  enqueueBundleEvent('tool_call_started', {
-    trace_id: traceId,
-    tool_name: toolName,
-    tool_classification: classification,
-    method: 'LOCAL',
-    api_path: '/local-tool',
-  });
-
-  try {
-    await ensureGovernanceContext();
-    const governancePayload = {
-      spawn_bundle_id: governanceContext.spawnBundleId,
-      policy_version: governanceContext.policyVersion,
-      tool_name: toolName,
-      tool_input: toolInput,
-      tool_classification: classification,
-      agent_id: AGENT_ID,
-      idempotency_key: randomUUID(),
-      lease_id: governanceContext.lease?.lease_id ?? null,
-    };
-    const governance = await callGovernance(governancePayload, { retry: true });
-    if (governance.policy_lease) applyPolicyLease(governance.policy_lease);
-    if (governance.verdict === 'reject' || governance.verdict === 'defer_human') {
-      throw governanceError(governanceReasonCode(governance.reason));
-    }
-
-    const checkedInput = (governance.verdict === 'modify' && governance.modified_input && typeof governance.modified_input === 'object')
-      ? { ...toolInput, ...governance.modified_input }
-      : toolInput;
-    const result = await executor(checkedInput);
-    if (CACHE_INVALIDATION_TOOLS.has(toolName)) {
-      governanceContext.cache.clear();
-    }
-
-    enqueueBundleEvent('tool_call_succeeded', {
-      trace_id: traceId,
-      tool_name: toolName,
-      tool_classification: classification,
-      source: 'governance_roundtrip',
-    });
-    return result;
-  } catch (error) {
-    if (shouldEmitToolCallFailed(error)) {
-      enqueueBundleEvent('tool_call_failed', {
-        trace_id: traceId,
-        tool_name: toolName,
-        tool_classification: classification,
-        reason: toolCallFailedReason(error),
-      });
-    }
-    throw error;
-  }
-}
-
 function renewCacheInBackground({ method, apiPath, body, toolName, cacheKey }) {
   if (governanceContext.renewalInFlight.has(cacheKey)) return;
   governanceContext.renewalInFlight.add(cacheKey);
@@ -886,6 +796,15 @@ server.tool('send_message', 'Send a message to a workspace, DM, or thread', {
   content: z.string().describe('Message content'),
 }, async ({ target, content }) => {
   const data = await api('POST', '/send', { target, content });
+  // Heuristic: if this looks like a video-production 确认稿 (asks permission +
+  // describes plan), mark a cross-process flag so media-tools' compose_video_v2
+  // / record_url_narration can verify a brief was actually sent before running.
+  // See daemon/src/video-brief-flag.js for the detection rules.
+  if (looksLikeVideoBrief(content) && AGENT_ID && currentWorkspaceId) {
+    try {
+      markVideoBriefSent({ workspaceId: currentWorkspaceId, agentId: AGENT_ID, content });
+    } catch { /* best-effort; failure to mark is non-fatal */ }
+  }
   return { content: [{ type: 'text', text: `Sent. messageId=${data.messageId} threadTarget=${data.threadTarget}` }] };
 });
 
@@ -1375,141 +1294,12 @@ server.tool('request_credential_auth',
   }
 );
 
-// ── render_text_to_image ───────────────────────────────────────────────────────
-server.tool('render_text_to_image',
-  'Render text content into image(s) for video synthesis. style=scroll produces a single tall image (for a scrolling video segment); style=carousel produces one image per card (for a slide-show segment). Returns local file paths.',
-  {
-    content: z.union([z.string(), z.array(z.string())]).describe('Text content. For carousel, pass an array of strings — one per card. For scroll, pass a single string (or array joined with line breaks).'),
-    style: z.enum(['scroll', 'carousel']).describe('scroll: one tall image; carousel: one image per card.'),
-    theme: z.enum(['dark', 'light']).optional().describe('Color theme. Default dark.'),
-    width: z.number().optional().describe('Image width in pixels. Default 1080.'),
-    card_height: z.number().optional().describe('Card height in pixels (carousel) or viewport height (scroll baseline). Default 1920.'),
-    font_size: z.number().optional().describe('Base font size in pixels. Default 48.'),
-  },
-  async (args) => runRenderTextToImageTool(args)
-);
-
-// ── render_html_to_image ───────────────────────────────────────────────────────
-server.tool('render_html_to_image',
-  'Render a raw HTML string to a PNG image by navigating to it as a local file:// page. Unlike evaluate_script+document.write on about:blank, this preserves file:// origin so <img src="file:///..."> references load correctly. Returns the output image path.',
-  {
-    html: z.string().describe('Full HTML document to render (including <!doctype>, <html>, <head>, <body>).'),
-    output_path: z.string().optional().describe('Absolute path to save the PNG. Auto-generated in /tmp if omitted.'),
-    viewport_width: z.number().optional().describe('Viewport width in pixels. Default 1080.'),
-    viewport_height: z.number().optional().describe('Viewport height in pixels. Default 1920.'),
-    wait_until: z.enum(['load', 'networkidle', 'domcontentloaded']).optional().describe('Navigation wait condition. Default load.'),
-  },
-  async (args) => runRenderHtmlToImageTool(args)
-);
-
-// ── synthesize_tts ─────────────────────────────────────────────────────────────
-server.tool('synthesize_tts',
-  'Convert text to speech using the workspace MiniMax TTS credential. Returns a local mp3 file path and duration. Use this to generate narration audio for individual video segments.',
-  {
-    text: z.string().describe('Text to synthesize. Keep under 500 characters per call for reliable results.'),
-    voice_id: z.string().optional().describe('MiniMax voice ID. Omit to use the workspace default voice.'),
-    workspace_id: z.string().optional().describe('Target workspace. Defaults to current workspace context.'),
-  },
-  async (args) => runSynthesisTtsTool({ ...args, currentWorkspaceId, api })
-);
-
-// ── plan_video_segments ────────────────────────────────────────────────────────
-// Session-scoped flag set when plan_video_segments runs. compose_video_v2
-// refuses TTS-bearing segments unless this is true — the agent must route
-// audio through plan_video_segments first so durations / subtitle_text /
-// audio_path are mechanically aligned. This is a per-chat-bridge-process
-// flag, so a fresh codex session must call plan_video_segments fresh.
-let _planVideoSegmentsCalledThisSession = false;
-
-server.tool('plan_video_segments',
-  'Universal audio-video sync planning step. For each segment, call TTS to get the real audio duration, then compute the visual duration with a safety buffer. Returns a planned segments array ready to pass directly to compose_video_v2 (with audio_path, presentation.duration/per_card_duration, and subtitle_text pre-filled). Always call this before compose_video_v2 when you have narration text.',
-  {
-    segments: z.array(z.object({
-      text: z.string().describe('Narration text for this segment. TTS will be generated from this.'),
-      visual_kind: z.enum(['image', 'video', 'gif', 'carousel']).describe('Type of visual.'),
-      visual_path: z.string().optional().describe('Absolute path to a single image, video, or gif file.'),
-      visual_paths: z.array(z.string()).optional().describe('For carousel: array of image paths, one per card.'),
-      transition: z.enum(['cut', 'fade', 'crossfade']).optional().describe('Transition to next segment. Default cut.'),
-      presentation: z.object({
-        style: z.enum(['static', 'scroll']).optional(),
-      }).optional().describe('Partial presentation hints (style only). duration/per_card_duration are computed from TTS.'),
-    })).describe('Segments to plan. Each must have narration text and visual info.'),
-    voice_id: z.string().optional().describe('TTS voice ID. Omit to use workspace default.'),
-    workspace_id: z.string().optional().describe('Target workspace. Defaults to current workspace context.'),
-  },
-  async (args) => {
-    const result = await runPlanVideoSegmentsTool({ ...args, currentWorkspaceId, api });
-    if (!result?.isError) _planVideoSegmentsCalledThisSession = true;
-    return result;
-  }
-);
-
-// ── compose_video_v2 ───────────────────────────────────────────────────────────
-server.tool('compose_video_v2',
-  'Compose a video from a list of segments using ffmpeg. Each segment has a visual source (image/scroll/carousel/video/gif), optional audio, and optional subtitle text. Subtitles are burned into the video by default when subtitle_text is provided. Segments are concatenated in order; outro clips are appended at the end. Returns a local mp4 path.\n\nTypical flow: plan_video_segments → compose_video_v2 (segments output fed directly in).',
-  {
-    segments: z.array(z.object({
-      visual_path: z.string().optional().describe('Absolute path to a single image, video, or gif file.'),
-      visual_paths: z.array(z.string()).optional().describe('For carousel: array of image paths, one per card.'),
-      visual_kind: z.enum(['image', 'video', 'gif', 'carousel']).describe('Type of visual.'),
-      presentation: z.object({
-        style: z.enum(['static', 'scroll']).optional().describe('For image: static (default) or scroll (pan upward).'),
-        duration: z.number().optional().describe('Segment duration in seconds. Required for image/scroll.'),
-        per_card_duration: z.number().optional().describe('Seconds per card for carousel.'),
-      }).optional(),
-      audio_path: z.string().nullable().optional().describe('Absolute path to audio (mp3). null or omit for silence.'),
-      subtitle_text: z.string().optional().describe('Narration text to burn as subtitle for this segment. Displayed for the full segment duration.'),
-      transition: z.enum(['cut', 'fade', 'crossfade']).optional().describe('Transition to next segment. Default cut.'),
-    })).describe('Ordered list of video segments.'),
-    outro_paths: z.array(z.string()).optional().describe('Absolute paths to outro video clips appended after all segments.'),
-    resolution: z.string().optional().describe('Output resolution WxH. Default "1080x1920".'),
-    output_path: z.string().optional().describe('Absolute output path for the mp4. Auto-generated if omitted.'),
-  },
-  async (args) => {
-    // Tool-level enforcement of the synthesize_tts → plan_video_segments →
-    // compose_video_v2 standard chain. If any segment has audio_path (i.e.
-    // narration is involved) and the agent never invoked plan_video_segments
-    // in this session, refuse the compose — manual dwell/duration math is
-    // unreliable (last syllable cut, silent tails, subtitle drift). Observed
-    // twice in row: agent skipped plan_video_segments, manually estimated
-    // dwell_ms wrong, ended up with too-long records and silent tails it then
-    // re-recorded to fix — wasting record_url_narration runs that
-    // plan_video_segments would have prevented.
-    const segments = Array.isArray(args?.segments) ? args.segments : [];
-    const hasNarration = segments.some(s => typeof s?.audio_path === 'string' && s.audio_path.trim());
-    if (hasNarration && !_planVideoSegmentsCalledThisSession) {
-      return {
-        isError: true,
-        content: [{
-          type: 'text',
-          text: 'compose_video_v2 refused: TTS-bearing segments (audio_path present) require plan_video_segments '
-            + 'to have run earlier in this session — it mechanically aligns audio_duration / video_duration / '
-            + 'subtitle_text with a safety buffer. Manual dwell/duration math has repeatedly produced misaligned '
-            + 'subtitles and silent tails that force re-recording.\n\n'
-            + 'Standard chain: synthesize_tts(per segment) → plan_video_segments(with text+visual_kind+visual_path) '
-            + '→ compose_video_v2(use the returned segments verbatim, only swap visual_path/visual_kind for real '
-            + 'media). Call plan_video_segments now and pass its output here.',
-        }],
-      };
-    }
-    return runComposeVideoV2Tool({ ...args, workspaceDir: WORKSPACE_DIR });
-  }
-);
-
-// ── take_page_screenshot ───────────────────────────────────────────────────────
-server.tool('take_page_screenshot',
-  'Open a URL with a headless browser and capture a screenshot. crop=above_fold captures only the visible viewport (ideal for thumbnail-style frames); crop=full_page captures the entire page height.',
-  {
-    url: z.string().describe('Page URL to screenshot.'),
-    crop: z.enum(['above_fold', 'full_page']).optional().describe('Capture mode. Default above_fold.'),
-    viewport: z.object({
-      width: z.number().optional(),
-      height: z.number().optional(),
-    }).optional().describe('Viewport size. Default 390×844 (mobile).'),
-    wait_for: z.enum(['load', 'networkidle', 'domcontentloaded']).optional().describe('Page load event to wait for before screenshotting. Default networkidle.'),
-  },
-  async (args) => runTakePageScreenshotTool(args)
-);
+// render_text_to_image, render_html_to_image, take_page_screenshot, synthesize_tts,
+// plan_video_segments, compose_video_v2, record_url_narration — all moved to
+// media-tools MCP server (V1–V5 migration; see
+// docs/scenario-content-creation/video-synthesis-design.md). The whole video
+// pipeline now lives in one stdio server so session-scoped flags (plan-was-
+// called) can gate downstream tools.
 
 // ── get_library_file ───────────────────────────────────────────────────────────
 server.tool('get_library_file',
@@ -1521,48 +1311,10 @@ server.tool('get_library_file',
   async (args) => runGetLibraryFileTool({ ...args, currentWorkspaceId, api, SERVER_URL, MACHINE_API_KEY, workspaceDir: WORKSPACE_DIR })
 );
 
-// ── record_url_narration ────────────────────────────────────────────────────────
-server.tool('record_url_narration',
-  'Record a silent video of a URL by driving Chromium on an Xvfb display and capturing it with Playwright recordVideo, driven by a video plan; ffmpeg then transcodes the recording to mp4. Outputs a silent mp4 that can be passed to compose_video_v2 as a video-kind segment with an audio_path for narration.\n\nUse this as the canonical recording step for URL-narration videos. Falls back: if the page needs interactions outside the visual_action vocabulary (clicks, waits, OCR loops), use Monitor (Bash) with custom Playwright instead.\n\nRuntime requirements: this tool only works on a Linux daemon machine with Xvfb + Chromium + ffmpeg installed (ffmpeg is used to transcode the recording to mp4; no x11grab device support needed). macOS / Windows daemons will fail at startup.',
-  {
-    url: z.string().describe('Page URL to record'),
-    plan: z.record(z.any()).describe(
-      'A video plan: an object with `phases` (or `sections`), each a "visual beat" with '
-      + '`action` (scroll_to_dwell / linear_scroll_during / scroll_back / hold / ...), a '
-      + 'target (`target_y` or `focus_region:[y1,y2]`) for scroll-type actions, and '
-      + '`dwell_ms` (how long to hold that beat — should match the segment\'s TTS duration).\n\n'
-      + 'For RECRUITMENT URLs (mp.weixin.qq.com / 校招 / 实习 / 岗位 content), each section MUST '
-      + 'also declare `target_y_content_label` — a short Chinese label describing what content '
-      + 'sits at that pixel y position on the page (e.g. "标题区" / "岗位信息卡片" / "公司介绍" / '
-      + '"届别说明"). Look at the take_page_screenshot output, find the y-pixel, and label it. '
-      + 'Labels matching forbidden regions ("二维码" / "扫码" / "投递入口" / "投递方式" / "联系方式" / '
-      + '"微信号" / "QR" / "阅读原文" / "外链") will cause the tool to refuse the recording — '
-      + 'recruitment content must NOT dwell on these areas (see fragments.md '
-      + 'frag.short.recruitment_url_mode_policy). Pick a different target_y in the 标题/岗位 '
-      + 'information area and rewrite that section.'
-    ),
-    output_path: z.string().optional().describe('Workspace-relative output mp4 path. Default tmp/wx3_video/recorded-{ts}.mp4'),
-    events_path: z.string().optional().describe('Workspace-relative events.json path. Default ${output_path}.events.json'),
-    viewport: z.object({
-      width: z.number().optional(),
-      height: z.number().optional(),
-    }).optional().describe('Default 1080x1920 (mobile portrait). Override only if the plan requires a different shape.'),
-    fps: z.number().optional().describe('Default 30. Do not change unless needed.'),
-    settle_ms: z.number().optional().describe('Default 4000. Settle wait after navigation before recording starts.'),
-  },
-  async (args) => {
-    if (isBlockedCvmaxEditorVideoTool('record_url_narration')) {
-      return cvmaxEditorVideoToolError('record_url_narration');
-    }
-    return runRecordUrlNarrationTool({
-      args,
-      currentWorkspaceId,
-      workspaceDir: WORKSPACE_DIR,
-      runMandatoryLocalToolFn: runMandatoryLocalTool,
-      recordUrlNarrationFn: recordUrlNarration,
-    });
-  }
-);
+// record_url_narration moved to media-tools MCP server (V4 migration). The
+// session-scoped plan_video_segments check now lives there alongside the
+// other video pipeline tools (synthesize_tts / plan_video_segments /
+// compose_video_v2). The CVMAX editor_in_chief block also moved with it.
 
 // ── submit_to_library ──────────────────────────────────────────────────────────
 server.tool('submit_to_library',

package/src/mcp-config.js

@@ -89,6 +89,11 @@ const SERVER_BACKED_MCP_SERVERS = new Set([
   'audience-research',
   'hook-pattern-library',
   'weixin-tools',
+  // media-tools (V1–V5 chat-bridge → media-tools migration): synthesize_tts
+  // hits /tts/voiceover and the CvMax editor_in_chief gate + video-brief
+  // checks read CURRENT_AGENT_ID / CURRENT_WORKSPACE_ID. lib/lightcone-api.js
+  // throws at module load without the SERVER_URL/MACHINE_API_KEY/AGENT_ID triple.
+  'media-tools',
 ]);
 
 function baseEnvForServer(serverKey, { serverUrl, authToken, agentId, workspaceId, workspaceDir }) {

package/src/tools/plan-video-segments.js

@@ -1,7 +1,22 @@
-import { mkdirSync, writeFileSync } from 'fs';
-import { randomUUID } from 'crypto';
-import path from 'path';
-import os from 'os';
+// plan_video_segments — pure audio/video alignment planner.
+//
+// Takes per-segment {text, audio_path, visual_kind, ...} and returns unified
+// plan segments with:
+//   - audio_duration_ms (read via ffprobe from the provided audio_path)
+//   - subtitle_text (= text)
+//   - presentation.duration / per_card_duration (audio_duration + buffer)
+//   - dwell_ms (= audio_duration; lets the same segment drive record_url_narration)
+//
+// Previously this tool ALSO synthesized TTS internally — which duplicated
+// the work when callers had already run synthesize_tts, and caused the
+// "wrong standard chain" confusion in fragments.md. TTS is now decoupled:
+// callers must run synthesize_tts per segment first and pass the resulting
+// audio_path here. See docs/scenario-content-creation/video-synthesis-design.md.
+//
+// Lives in daemon/src/tools/ as a reusable handler module; the MCP-tool
+// registration lives in daemon/mcp-servers/official/media-tools/index.js.
+
+import { spawn } from 'node:child_process';
 
 function toolText(text) {
   return { content: [{ type: 'text', text }] };
@@ -11,125 +26,99 @@ function toolError(text) {
   return { isError: true, content: [{ type: 'text', text }] };
 }
 
-function inferAudioExt(url) {
-  const clean = String(url ?? '').split('?')[0];
-  const ext = path.extname(clean).toLowerCase();
-  return ['.mp3', '.wav', '.flac', '.aac', '.ogg'].includes(ext) ? ext : '.mp3';
-}
-
-async function synthesizeSegmentTts(text, { workspace_id, voice_id, api }) {
-  const payload = { workspace_id, text, speed: 1, format: 'mp3' };
-  if (voice_id) payload.voice_preset = String(voice_id).trim();
-
-  const data = await api('POST', '/tts/voiceover', payload);
-  const remoteAudioUrl = String(data.audio_url ?? '').trim();
-  if (!remoteAudioUrl) throw new Error('TTS API did not return audio_url');
-
-  const downloadRes = await fetch(remoteAudioUrl);
-  if (!downloadRes.ok) throw new Error(`Failed to download audio (${downloadRes.status})`);
-
-  const fileBuffer = Buffer.from(await downloadRes.arrayBuffer());
-  const outDir = path.join(os.tmpdir(), 'lightcone-tts');
-  mkdirSync(outDir, { recursive: true });
-  const ext = inferAudioExt(remoteAudioUrl);
-  const outPath = path.join(outDir, `tts-${Date.now()}-${randomUUID().slice(0, 8)}${ext}`);
-  writeFileSync(outPath, fileBuffer);
-
-  const durationMs = Number(data.duration_ms ?? 0);
-  return { audio_path: outPath, audio_duration_ms: durationMs };
+async function probeAudioDurationMs(audioPath) {
+  return new Promise((resolve, reject) => {
+    const proc = spawn('ffprobe', [
+      '-v', 'error',
+      '-show_entries', 'format=duration',
+      '-of', 'csv=p=0',
+      audioPath,
+    ], { stdio: ['ignore', 'pipe', 'pipe'] });
+    const out = [];
+    const err = [];
+    proc.stdout.on('data', c => out.push(c));
+    proc.stderr.on('data', c => err.push(c));
+    proc.on('close', code => {
+      if (code !== 0) {
+        return reject(new Error(`ffprobe exited ${code}: ${Buffer.concat(err).toString().slice(-300)}`));
+      }
+      const seconds = parseFloat(Buffer.concat(out).toString().trim());
+      if (!Number.isFinite(seconds)) {
+        return reject(new Error(`ffprobe returned non-numeric duration: ${Buffer.concat(out).toString().slice(0, 200)}`));
+      }
+      resolve(Math.round(seconds * 1000));
+    });
+    proc.on('error', err2 => reject(new Error(`ffprobe spawn failed: ${err2.message}`)));
+  });
 }
 
-// Compute segment duration from audio duration: audio + 0.5s buffer, rounded up to nearest 0.5s.
+// Plan visual duration from audio duration: audio + buffer, rounded up to the
+// nearest 0.5s. scroll-style images get a longer buffer because eyes need
+// extra time to follow the motion.
 function planDurationSec(audioDurationMs, bufferSec = 0.5) {
   const raw = audioDurationMs / 1000 + bufferSec;
-  return Math.ceil(raw * 2) / 2; // round up to nearest 0.5s
+  return Math.ceil(raw * 2) / 2;
 }
 
-// Run fn over items with a bounded number of concurrent workers (FIFO drain).
-async function mapWithConcurrency(items, limit, fn) {
-  const queue = items.map((item, index) => ({ item, index }));
-  const workers = Array.from({ length: Math.max(1, Math.min(limit, queue.length)) }, async () => {
-    while (queue.length > 0) {
-      const next = queue.shift();
-      await fn(next.item, next.index);
-    }
-  });
-  await Promise.all(workers);
-}
-
-const TTS_CONCURRENCY = 5;
-
-export async function runPlanVideoSegmentsTool({ segments, workspace_id, voice_id, currentWorkspaceId, api }) {
+export async function runPlanVideoSegmentsTool({ segments } = {}) {
   if (!Array.isArray(segments) || segments.length === 0) {
     return toolError('segments must be a non-empty array.');
   }
 
-  const targetWorkspaceId = String(workspace_id ?? currentWorkspaceId ?? '').trim();
-  if (!targetWorkspaceId) {
-    return toolError('workspace_id is required (no current workspace context).');
+  // Up-front validation — fail fast before any work.
+  for (let i = 0; i < segments.length; i += 1) {
+    const seg = segments[i] ?? {};
+    if (typeof seg.audio_path !== 'string' || !seg.audio_path.trim()) {
+      return toolError(
+        `segments[${i}]: audio_path is required. plan_video_segments no longer synthesizes TTS — call synthesize_tts(text) `
+        + 'first and pass the returned path as audio_path. Standard chain: synthesize_tts × N → plan_video_segments → '
+        + 'record_url_narration + compose_video_v2 (share the same plan).'
+      );
+    }
+    const kind = String(seg.visual_kind ?? '');
+    if (!kind) {
+      return toolError(`segments[${i}]: visual_kind is required (image / video / gif / carousel).`);
+    }
   }
 
   const planned = [];
-  const errors = [];
+  const warnings = [];
+
+  for (let i = 0; i < segments.length; i += 1) {
+    const seg = segments[i];
+    const text = String(seg.text ?? '').trim();
+    const kind = String(seg.visual_kind);
 
-  // Synthesize TTS for every text-bearing segment up front, in parallel (bounded),
-  // so an N-segment plan no longer pays N sequential round-trips to the TTS API.
-  const audioResults = new Array(segments.length).fill(null);
-  const ttsJobs = segments
-    .map((seg, i) => ({ i, text: String(seg.text ?? '').trim() }))
-    .filter(job => job.text);
-  await mapWithConcurrency(ttsJobs, TTS_CONCURRENCY, async ({ i, text }) => {
+    let audioDurationMs;
     try {
-      audioResults[i] = await synthesizeSegmentTts(text, { workspace_id: targetWorkspaceId, voice_id, api });
+      audioDurationMs = await probeAudioDurationMs(seg.audio_path);
     } catch (err) {
-      errors.push(`segments[${i}]: TTS failed — ${err.message}`);
-      audioResults[i] = { audio_path: null, audio_duration_ms: 3000 }; // fallback estimate
+      warnings.push(`segments[${i}]: audio probe failed (${err.message}); falling back to 3000ms`);
+      audioDurationMs = 3000;
     }
-  });
-  errors.sort((a, b) => {
-    const na = Number((a.match(/segments\[(\d+)\]/) ?? [])[1] ?? 0);
-    const nb = Number((b.match(/segments\[(\d+)\]/) ?? [])[1] ?? 0);
-    return na - nb;
-  });
 
-  for (let i = 0; i < segments.length; i++) {
-    const seg = segments[i];
-    const text = String(seg.text ?? '').trim();
-    const kind = String(seg.visual_kind ?? 'image');
-    const audioResult = audioResults[i];
-
-    const audioDurationMs = audioResult?.audio_duration_ms ?? 0;
     let presentation;
-
     if (kind === 'carousel') {
       const numCards = Array.isArray(seg.visual_paths) ? seg.visual_paths.length : 1;
-      const totalDuration = audioDurationMs > 0 ? planDurationSec(audioDurationMs) : numCards * 4;
+      const totalDuration = planDurationSec(audioDurationMs);
       const perCard = Math.max(2, Math.ceil((totalDuration / numCards) * 2) / 2);
       presentation = { per_card_duration: perCard };
     } else {
-      // image, scroll, video, gif
-      const duration = audioDurationMs > 0 ? planDurationSec(audioDurationMs, kind === 'scroll' ? 1.0 : 0.5) : 4;
+      // image / scroll / video / gif
+      const duration = planDurationSec(audioDurationMs, kind === 'scroll' ? 1.0 : 0.5);
       presentation = { duration, ...(kind === 'scroll' ? { style: 'scroll' } : {}) };
     }
 
-    // dwell_ms lets the same segment double as a record_url_narration plan phase
-    // (the recorder reads dwell_ms / duration_ms for how long to hold each beat).
-    // Prefer the real measured audio length; fall back to the planned visual duration.
-    const dwellMs = audioDurationMs > 0
-      ? audioDurationMs
-      : Math.round((presentation.duration ?? presentation.per_card_duration ?? 4) * 1000);
-
-    const planned_seg = {
+    planned.push({
       ...seg,
-      ...(audioResult?.audio_path ? { audio_path: audioResult.audio_path } : {}),
+      audio_path: seg.audio_path,
+      audio_duration_ms: audioDurationMs,
       ...(text ? { subtitle_text: text } : {}),
       presentation: { ...presentation, ...(seg.presentation ?? {}) },
-      dwell_ms: seg.dwell_ms ?? dwellMs,
-    };
-    if (audioResult?.audio_duration_ms) {
-      planned_seg.audio_duration_ms = audioResult.audio_duration_ms;
-    }
-    planned.push(planned_seg);
+      // dwell_ms doubles as record_url_narration's per-phase hold duration so
+      // recording naturally tracks the narration audio.
+      dwell_ms: seg.dwell_ms ?? audioDurationMs,
+    });
   }
 
   const result = {
@@ -141,7 +130,7 @@ export async function runPlanVideoSegmentsTool({ segments, workspace_id, voice_i
         : (s.presentation?.duration ?? 4);
       return sum + Math.round(d * 1000);
     }, 0),
-    ...(errors.length > 0 ? { warnings: errors } : {}),
+    ...(warnings.length > 0 ? { warnings } : {}),
   };
 
   return toolText(JSON.stringify(result, null, 2));

package/src/{record-url-narration-tool.js → tools/record-url-narration.js}

@@ -1,6 +1,23 @@
+// record_url_narration — atomic recording tool.
+//
+// Drives Chromium on Xvfb + Playwright recordVideo to capture a silent mp4 of
+// a URL following a beat-by-beat visual plan, then ffmpeg-transcodes it. The
+// resulting silent mp4 feeds into compose_video_v2 as a video-kind segment
+// alongside narration audio.
+//
+// Lives in daemon/src/tools/ as a reusable handler module; the MCP-tool
+// registration lives in daemon/mcp-servers/official/media-tools/index.js.
+// Migrated out of chat-bridge.js (V4) — no longer wrapped by
+// runMandatoryLocalTool / governance round-trip. media-tools is a separate
+// stdio MCP server and governance integration is chat-bridge-specific;
+// matches the precedent set by synthesize_tts / plan_video_segments /
+// compose_video_v2 in V1/V2/V3.
+
 import { mkdirSync } from 'fs';
 import path from 'path';
 
+import { recordUrlNarration as defaultRecordUrlNarrationFn } from '../_vendor/video/recorder/index.js';
+
 function toolText(text) {
   return { content: [{ type: 'text', text }] };
 }
@@ -55,10 +72,6 @@ function derivePhaseCount({ plan, recorderOutput }) {
   return segments ? segments.length : null;
 }
 
-// record_url_narration is an atomic tool, not the tail of a fixed pipeline.
-// The plan may be hand-written by the scripter or produced by plan_video_segments;
-// it just needs a non-empty list of segments with per-segment visual action + duration
-// so the recording stays in sync with the narration audio.
 function assertPipelineCompliance(plan) {
   if (!isPlainObject(plan)) return;
   if (!planSegments(plan)) {
@@ -86,10 +99,6 @@ const FORBIDDEN_REGION_PATTERNS = [
 ];
 
 function isRecruitmentLikeUrl(url) {
-  // Conservative URL-based heuristic: mp.weixin.qq.com pages forwarding 招聘 /
-  // 校招 / 实习 / job content. Until we have content classification, treat
-  // mp.weixin.qq.com URLs as recruitment-class for safety — the cost of a
-  // mis-flag is "agent must add a label", not "recording fails permanently".
   if (typeof url !== 'string') return false;
   return /mp\.weixin\.qq\.com/.test(url);
 }
@@ -101,11 +110,6 @@ function describeForbiddenMatch(label) {
   return null;
 }
 
-/**
- * For recruitment-class URLs, every plan section must declare what content
- * sits at its target_y, and the label must NOT match the forbidden-region
- * patterns. Returns null on pass, error message string on fail.
- */
 function checkSafeRegionLabels({ url, plan }) {
   if (!isRecruitmentLikeUrl(url)) return null;
   const segments = planSegments(plan);
@@ -181,16 +185,13 @@ export async function runRecordUrlNarrationTool({
   args = {},
   currentWorkspaceId = '',
   workspaceDir = process.cwd(),
-  runMandatoryLocalToolFn,
-  recordUrlNarrationFn,
+  planVideoSegmentsCalled = false,
+  recordUrlNarrationFn = defaultRecordUrlNarrationFn,
   nowMs = () => Date.now(),
 } = {}) {
   if (!currentWorkspaceId) {
     return toolError('No workspace context.');
   }
-  if (typeof runMandatoryLocalToolFn !== 'function') {
-    return toolError('Error: record_url_narration runMandatoryLocalToolFn is required.');
-  }
   if (typeof recordUrlNarrationFn !== 'function') {
     return toolError('Error: record_url_narration executor is required.');
   }
@@ -220,51 +221,53 @@ export async function runRecordUrlNarrationTool({
     return toolError(`Error: ${safeRegionError}`);
   }
 
+  // Standard-chain hard block: refuse recordings unless plan_video_segments
+  // ran in this session. Discovered repeatedly in Tasks #20/#25/#26 that
+  // agents hand-write dwell_ms by guessing, producing recordings whose phase
+  // boundaries drift from the TTS audio they will eventually be paired with —
+  // forcing a full re-record. plan_video_segments fills dwell_ms mechanically
+  // from ffprobe audio duration, eliminating the drift.
+  if (!planVideoSegmentsCalled) {
+    return toolError(
+      'Error: record_url_narration refused: plan_video_segments must run earlier in this '
+      + 'session so dwell_ms / phase durations are mechanically aligned with the segment\'s '
+      + 'TTS audio (audio_duration_ms). Hand-written dwell_ms has repeatedly drifted from '
+      + 'the actual TTS duration in production runs, forcing full re-records.\n\n'
+      + 'Standard chain: synthesize_tts × N (per segment) → plan_video_segments(segments with '
+      + 'text + audio_path + visual_kind=video + visual_path) → record_url_narration (feed '
+      + 'plan_video_segments output as plan.sections — each section\'s dwell_ms is already '
+      + 'set to audio_duration_ms) + compose_video_v2 (same plan output). Call plan_video_segments '
+      + 'now, then pass its `segments` array as `plan.sections` here.'
+    );
+  }
+
   try {
-    const result = await runMandatoryLocalToolFn({
-      toolName: 'record_url_narration',
-      toolInput: validatedInput,
-      executor: async (checkedInput = {}) => {
-        const mergedInput = {
-          ...validatedInput,
-          ...checkedInput,
-        };
-        const finalInput = validateRecordUrlNarrationArgs(mergedInput);
-        const { resolvedOutputPath, resolvedEventsPath } = resolveRecordUrlNarrationPaths({
-          workspaceDir,
-          outputPath: finalInput.output_path,
-          eventsPath: finalInput.events_path,
-          nowMs,
-        });
-
-        mkdirSync(path.dirname(resolvedOutputPath), { recursive: true });
-        mkdirSync(path.dirname(resolvedEventsPath), { recursive: true });
-
-        const recorderOutput = await recordUrlNarrationFn({
-          url: finalInput.url,
-          plan: finalInput.plan,
-          output_path: resolvedOutputPath,
-          events_path: resolvedEventsPath,
-          viewport: finalInput.viewport,
-          fps: finalInput.fps,
-          settle_ms: finalInput.settle_ms,
-        });
-
-        return {
-          videoPath: resolvedOutputPath,
-          eventsPath: resolvedEventsPath,
-          durationMs: deriveDurationMs(recorderOutput),
-          phases: derivePhaseCount({ plan: finalInput.plan, recorderOutput }),
-        };
-      },
+    const { resolvedOutputPath, resolvedEventsPath } = resolveRecordUrlNarrationPaths({
+      workspaceDir,
+      outputPath: validatedInput.output_path,
+      eventsPath: validatedInput.events_path,
+      nowMs,
+    });
+
+    mkdirSync(path.dirname(resolvedOutputPath), { recursive: true });
+    mkdirSync(path.dirname(resolvedEventsPath), { recursive: true });
+
+    const recorderOutput = await recordUrlNarrationFn({
+      url: validatedInput.url,
+      plan: validatedInput.plan,
+      output_path: resolvedOutputPath,
+      events_path: resolvedEventsPath,
+      viewport: validatedInput.viewport,
+      fps: validatedInput.fps,
+      settle_ms: validatedInput.settle_ms,
     });
 
     return toolText(
       `Recorded URL narration.\n`
-      + `video_path=${result.videoPath}\n`
-      + `events_path=${result.eventsPath}\n`
-      + `duration_ms=${result.durationMs ?? 'unknown'}\n`
-      + `phases=${result.phases ?? 'n/a'}`
+      + `video_path=${resolvedOutputPath}\n`
+      + `events_path=${resolvedEventsPath}\n`
+      + `duration_ms=${deriveDurationMs(recorderOutput) ?? 'unknown'}\n`
+      + `phases=${derivePhaseCount({ plan: validatedInput.plan, recorderOutput }) ?? 'n/a'}`
     );
   } catch (error) {
     return toolError(`Error: ${error.message}`);

package/src/video-brief-flag.js

@@ -0,0 +1,78 @@
+// Cross-process flag for "the agent has sent a video-production 确认稿 (brief)
+// to the user in this workspace+agent context recently". chat-bridge's
+// send_message tool writes the flag when the outgoing message heuristically
+// looks like a confirmation brief; media-tools' compose_video_v2 and
+// record_url_narration read the flag and refuse to proceed without it.
+//
+// Why a file flag instead of in-process state: send_message lives in
+// chat-bridge (one stdio MCP server), compose_video_v2/record_url_narration
+// live in media-tools (a different stdio MCP server, same machine). Both are
+// spawned by the same codex CLI session per agent, so they share env (notably
+// AGENT_ID / WORKSPACE_ID) but not memory. A flag file under ~/.lightcone is
+// the simplest cross-process medium and survives short-lived MCP restarts.
+//
+// The heuristic is intentionally specific (asks-permission marker AND
+// 2+ plan-describing markers) so casual progress reports like "已生成 TTS"
+// or "画面已就绪" do NOT satisfy it. A motivated agent could game the
+// detection by stuffing keywords into any send_message, but the default
+// codex behavior (which silently skipped the soft prompt rule) is what we
+// need to interrupt — and gaming is observable in chat history.
+
+import { mkdirSync, statSync, utimesSync, writeFileSync, existsSync } from 'node:fs';
+import path from 'node:path';
+import os from 'node:os';
+
+const TTL_MS = 6 * 60 * 60 * 1000; // 6 hours
+const FILE_NAME = 'video-brief-sent.flag';
+
+function flagDir(workspaceId, agentId) {
+  return path.join(os.homedir(), '.lightcone', 'sessions', workspaceId, agentId);
+}
+
+function flagPath(workspaceId, agentId) {
+  return path.join(flagDir(workspaceId, agentId), FILE_NAME);
+}
+
+export function markVideoBriefSent({ workspaceId, agentId, content }) {
+  if (!workspaceId || !agentId) return;
+  const dir = flagDir(workspaceId, agentId);
+  const p = flagPath(workspaceId, agentId);
+  mkdirSync(dir, { recursive: true });
+  writeFileSync(p, String(content ?? '').slice(0, 4096));
+  const now = new Date();
+  utimesSync(p, now, now);
+}
+
+export function hasFreshVideoBrief({ workspaceId, agentId, ttlMs = TTL_MS } = {}) {
+  if (!workspaceId || !agentId) return false;
+  const p = flagPath(workspaceId, agentId);
+  if (!existsSync(p)) return false;
+  try {
+    const st = statSync(p);
+    return (Date.now() - st.mtimeMs) <= ttlMs;
+  } catch { return false; }
+}
+
+// Permission-asking markers — the message must ask the user to decide / OK.
+// 确认 alone is too broad (matches "无需确认" / "已确认硬约束" / "确认收到"); require
+// a specific permission-ask shape: 请确认 / 确认稿 / 你确认 / 确认[吗?？] / 等确认.
+const PERMISSION_MARKERS = [
+  /请.*确认/, /你.*确认/, /确认\s*[吗?？]/, /等.*确认/, /确认稿/,
+  /你看/, /OK\s*吗/i, /可以吗/, /同意吗/, /通过吗/, /行不行/, /如何\?|如何？/,
+];
+
+// Plan-describing markers — must cover at least 2 different aspects of the brief.
+const PLAN_MARKERS = [
+  /画面/, /时长/, /文案/, /口播/, /字幕/, /顺序/, /口吻/, /分镜/, /配音/,
+];
+
+export function looksLikeVideoBrief(content) {
+  if (typeof content !== 'string') return false;
+  // Min length 20 — Chinese is character-dense, a plausible brief like
+  // "请确认：画面/时长/字幕已定。同意吗？" is only ~20 chars but still a valid brief.
+  if (content.length < 20) return false;
+  const hasPermissionAsk = PERMISSION_MARKERS.some(rx => rx.test(content));
+  if (!hasPermissionAsk) return false;
+  const distinctPlanHits = PLAN_MARKERS.filter(rx => rx.test(content)).length;
+  return distinctPlanHits >= 2;
+}