@lightcone-ai/daemon 0.22.1 → 0.23.1

package/src/tools/record-url-narration.js

@@ -1,17 +1,16 @@
-// record_url_narration — atomic recording tool.
+// V6 record_url_narration daemon tool wrapper.
 //
-// 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.
+// Drives Chromium on Xvfb + Playwright recordVideo to capture a silent mp4
+// per section, then ffmpeg-transcodes + slices. The resulting silent mp4s
+// feed compose_video_v2 as video-kind segments 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.
+// V6 contract (see docs/scenario-content-creation/video-synthesis-design.md):
+//   - page_understanding (from analyze_page) is required — drives safe-region
+//     check and preheat consistency
+//   - plan.sections each carry operations[] of atom calls; V5 fields are rejected
+//   - mp.weixin.qq.com-only keyword blacklist is gone — unsafe_regions from
+//     page_understanding is the universal safety mechanism
+//   - plan_video_segments must run earlier in this session (standard chain)
 
 import { mkdirSync } from 'fs';
 import path from 'path';
@@ -56,89 +55,9 @@ function deriveDurationMs(recorderOutput) {
   return lastTms > 0 ? lastTms : null;
 }
 
-function planSegments(plan) {
+function planSections(plan) {
   if (!isPlainObject(plan)) return null;
-  for (const key of ['phases', 'sections', 'segments']) {
-    if (Array.isArray(plan[key]) && plan[key].length > 0) return plan[key];
-  }
-  return null;
-}
-
-function derivePhaseCount({ plan, recorderOutput }) {
-  const explicit = normalizeNumberOrNull(recorderOutput?.phases);
-  if (explicit != null) return explicit;
-
-  const segments = planSegments(plan);
-  return segments ? segments.length : null;
-}
-
-function assertPipelineCompliance(plan) {
-  if (!isPlainObject(plan)) return;
-  if (!planSegments(plan)) {
-    throw new Error(
-      'record_url_narration: `plan` must contain a non-empty `phases` (or `sections` / `segments`) array — '
-      + 'either hand-written or from plan_video_segments. Each entry should carry a visual action and a duration.'
-    );
-  }
-}
-
-// Forbidden region keywords for recruitment content. If a section's
-// target_y_content_label matches, we refuse to record — the resulting video
-// would show 投递入口 / 二维码 / contact info, which violates the recruitment
-// content policy (see fragments.md frag.short.recruitment_url_mode_policy).
-//
-// Origin: in production runs the agent's plan repeatedly declared a target_y
-// without checking what content lived at that pixel position, and ended up
-// dwelling on QR codes / 投递 entries / 联系方式. The prompt-level rule
-// requiring `target_y_content_label` has been ignored often enough that we
-// enforce it at the tool layer instead.
-const FORBIDDEN_REGION_PATTERNS = [
-  /二维码/, /扫码/, /扫一扫/,
-  /投递入口/, /投递方式/, /投递通道/, /投递渠道/, /报名入口/, /报名方式/,
-  /联系方式/, /联系人/, /微信号/, /\bWeChat\b/i, /\bQQ群\b/,
-  /阅读原文/, /外链/, /\bQR\b/i,
-];
-
-function isRecruitmentLikeUrl(url) {
-  if (typeof url !== 'string') return false;
-  return /mp\.weixin\.qq\.com/.test(url);
-}
-
-function describeForbiddenMatch(label) {
-  for (const pattern of FORBIDDEN_REGION_PATTERNS) {
-    if (pattern.test(label)) return pattern.source;
-  }
-  return null;
-}
-
-function checkSafeRegionLabels({ url, plan }) {
-  if (!isRecruitmentLikeUrl(url)) return null;
-  const segments = planSegments(plan);
-  if (!segments) return null;
-  for (let i = 0; i < segments.length; i += 1) {
-    const seg = segments[i] ?? {};
-    const label = normalizeText(seg.target_y_content_label ?? seg.targetYContentLabel ?? '');
-    if (!label) {
-      return (
-        `record_url_narration: section[${i}] is missing required field `
-        + `\`target_y_content_label\`. For recruitment URLs (mp.weixin.qq.com / `
-        + `校招 / 实习等) you MUST label what content lives at target_y so the `
-        + `tool can verify it is not 二维码/投递入口/联系方式. Look at the page `
-        + `screenshot, find what is at target_y=${seg.target_y ?? '<unset>'}, `
-        + `and add a short label like "标题区" / "岗位信息卡片" / "公司介绍".`
-      );
-    }
-    const match = describeForbiddenMatch(label);
-    if (match) {
-      return (
-        `record_url_narration: section[${i}] target_y=${seg.target_y ?? '?'} `
-        + `is labeled "${label}", which matches a forbidden region pattern `
-        + `/${match}/. Recruitment content must NOT dwell on 投递入口 / 二维码 / `
-        + `联系方式 areas. Pick a different target_y inside the 标题区 / 岗位 `
-        + `信息卡片 / 公司介绍 area and rewrite this section.`
-      );
-    }
-  }
+  if (Array.isArray(plan.sections) && plan.sections.length > 0) return plan.sections;
   return null;
 }
 
@@ -156,6 +75,24 @@ export function validateRecordUrlNarrationArgs(args = {}) {
     throw error;
   }
 
+  if (!planSections(args.plan)) {
+    const error = new Error(
+      'plan.sections is required (non-empty array). Each section: { id?, text?, audio_path?, dwell_ms?, operations: [{atom, duration_ms, ...}] }.',
+    );
+    error.code = 'PLAN_SECTIONS_REQUIRED';
+    throw error;
+  }
+
+  if (!isPlainObject(args.page_understanding)) {
+    const error = new Error(
+      'page_understanding is required — call analyze_page(url) first and pass its output here. '
+      + 'V6 uses page_understanding.unsafe_regions[] to validate scroll_to.y / cursor_focus.y, and '
+      + 'page_understanding.preheat_strategy to align the record browser with the analyze browser.',
+    );
+    error.code = 'PAGE_UNDERSTANDING_REQUIRED';
+    throw error;
+  }
+
   return {
     ...(args ?? {}),
     url: normalizedUrl,
@@ -219,49 +156,22 @@ export async function runRecordUrlNarrationTool({
     return toolError(`Error: ${error.message}`);
   }
 
-  try {
-    assertPipelineCompliance(validatedInput.plan);
-  } catch (error) {
-    return toolError(`Error: ${error.message}`);
-  }
-
-  // Safe-region check for recruitment URLs — refuse plans that dwell on
-  // forbidden regions (二维码 / 投递入口 / 联系方式) before we even start
-  // Chromium. The agent must label each target_y with the content that lives
-  // there, and the labels are pattern-matched against a forbidden list.
-  const safeRegionError = checkSafeRegionLabels({
-    url: validatedInput.url,
-    plan: validatedInput.plan,
-  });
-  if (safeRegionError) {
-    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.
+  // ran in this session. plan_video_segments is what aligns operations[]
+  // duration sums to the per-section TTS audio duration; skipping it lets
+  // audio/visual drift accumulate across sections.
   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.'
+      + 'session so per-section operations.duration_ms is reconciled with TTS audio_duration_ms.\n\n'
+      + 'V6 standard chain: analyze_page(url) → synthesize_tts × N (per section) → '
+      + 'plan_video_segments(segments with text + audio_path + visual_kind + operations) → '
+      + 'record_url_narration(url, page_understanding, plan=…, output_paths=[…]) + '
+      + 'compose_video_v2(segments=…, variants=[…]). Call plan_video_segments now, then retry.',
     );
   }
 
   try {
-    // output_paths is REQUIRED. The legacy "default output_path master file"
-    // mode is gone — agents kept defaulting to one-call-per-section because
-    // that was the lowest-friction path. Now every recording is sliced, even
-    // single-section ones (which are just a 1-element output_paths array).
     let resolvedOutputPaths;
     try {
       resolvedOutputPaths = resolveOutputPaths(validatedInput.output_paths, { workspaceDir });
@@ -272,22 +182,18 @@ export async function runRecordUrlNarrationTool({
       return toolError(
         'Error: output_paths is required — one workspace-relative mp4 path per plan.sections entry. '
         + 'Single-section recording is a 1-element array. Multi-section recording records once '
-        + 'continuously (one browser session, one scrollTop) and slices the result at section '
-        + 'boundaries. See frag.short.video_synthesis_tools.',
+        + 'continuously (one browser session) and slices the result at section boundaries.',
       );
     }
-    const planSectionCount = (planSegments(validatedInput.plan) ?? []).length;
-    if (resolvedOutputPaths.length !== planSectionCount) {
+    const sectionCount = (planSections(validatedInput.plan) ?? []).length;
+    if (resolvedOutputPaths.length !== sectionCount) {
       return toolError(
         `Error: output_paths length (${resolvedOutputPaths.length}) must match `
-        + `plan.sections length (${planSectionCount}). Each section produces exactly one mp4 — `
+        + `plan.sections length (${sectionCount}). Each section produces exactly one mp4 — `
         + `don't pad or truncate.`,
       );
     }
 
-    // The master / events JSON paths are agent-optional debug artifacts.
-    // Default master to a tmp path next to the first output; events default
-    // to <master>.events.json. Agent can override either if they care.
     const { resolvedOutputPath: masterPath, resolvedEventsPath } = resolveRecordUrlNarrationPaths({
       workspaceDir,
       outputPath: validatedInput.output_path,
@@ -300,6 +206,7 @@ export async function runRecordUrlNarrationTool({
     const recorderOutput = await recordUrlNarrationFn({
       url: validatedInput.url,
       plan: validatedInput.plan,
+      page_understanding: validatedInput.page_understanding,
       output_path: masterPath,
       events_path: resolvedEventsPath,
       output_paths: resolvedOutputPaths,

package/src/_vendor/video/recorder/phase-duration.js

@@ -1,18 +0,0 @@
-function normalizeInteger(value, fallback = null) {
-  const parsed = Number.parseInt(String(value ?? ''), 10);
-  if (!Number.isFinite(parsed)) return fallback;
-  return parsed;
-}
-
-export function resolveDurationMs(phase, fallback = 0) {
-  const parsed = normalizeInteger(phase?.duration_ms, null);
-  if (parsed !== null && parsed >= 0) return parsed;
-
-  const dwellMs = normalizeInteger(phase?.dwell_ms, null);
-  if (dwellMs !== null && dwellMs >= 0) return dwellMs;
-
-  const secs = Number(phase?.duration_s);
-  if (Number.isFinite(secs) && secs >= 0) return Math.round(secs * 1000);
-
-  return fallback;
-}

package/src/_vendor/video/recorder/plan-estimator.js

@@ -1,43 +0,0 @@
-import { resolveDurationMs } from './phase-duration.js';
-import { normalizePlanPhases } from './plan-executor.js';
-
-export function estimatePlanDurationMs(plan = {}) {
-  let phases = [];
-  try {
-    phases = normalizePlanPhases(plan);
-  } catch {
-    phases = [];
-  }
-
-  return phases.reduce((total, phase) => {
-    const action = String(phase?.action ?? phase?.visual_action?.type ?? '').trim().toLowerCase();
-    const durationMs = resolveDurationMs(phase, Number.NaN);
-    const dwellMs = Number(phase?.dwell_ms);
-    const transitionMs = Number(phase?.transition_ms ?? phase?.visual_action?.transition_ms);
-    const effectiveHoldMs = Number.isFinite(dwellMs) && dwellMs > 0
-      ? dwellMs
-      : durationMs;
-
-    if (action === 'hold' && Number.isFinite(effectiveHoldMs) && effectiveHoldMs > 0) {
-      return total + effectiveHoldMs;
-    }
-    if (action === 'linear_scroll_during') {
-      if (Number.isFinite(effectiveHoldMs) && effectiveHoldMs > 0) return total + effectiveHoldMs;
-      return total + 1200;
-    }
-    if (action === 'scroll_to_dwell' || action === 'cursor_focus' || action === 'scroll_back') {
-      let next = total;
-      if (Number.isFinite(transitionMs) && transitionMs > 0) next += transitionMs;
-      if (Number.isFinite(effectiveHoldMs) && effectiveHoldMs > 0) next += effectiveHoldMs;
-      if (next === total) next += 1200;
-      return next;
-    }
-    if (Number.isFinite(transitionMs) && transitionMs > 0) {
-      return total + transitionMs;
-    }
-    if (Number.isFinite(durationMs) && durationMs > 0) {
-      return total + durationMs;
-    }
-    return total + 800;
-  }, 0);
-}