@lightcone-ai/daemon 0.17.0 → 0.17.1

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

@@ -12,6 +12,7 @@ import { runRecordUrlNarrationTool } from '../../../src/tools/record-url-narrati
 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 ?? '');
@@ -222,6 +223,18 @@ server.tool(
         + '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 });
   }
 );
@@ -265,6 +278,25 @@ server.tool(
     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,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lightcone-ai/daemon",
-  "version": "0.17.0",
+  "version": "0.17.1",
   "type": "module",
   "main": "src/index.js",
   "bin": {

package/src/chat-bridge.js

@@ -16,6 +16,7 @@ import { runSubmitToLibraryTool } from './submit-to-library-tool.js';
 // 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 {
@@ -795,6 +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}` }] };
 });
 

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/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;
+}