bilibili-page-reader

name	bilibili-page-reader
description	Get content from Bilibili videos: official subtitles, danmaku (density/peaks/sample), comments. When no subtitles exist (no 投稿字幕), fall back to downloading audio + ASR transcription via FunASR.

Bilibili Page Reader

Core Rules

• Browser access via kimi-webbridge for page state, login-only data, Bilibili Evolved (BE) providers.
• Do not click BE download buttons. Call providers directly via evaluate.
• Always use a named session. A session-less evaluate may land on a different tab and silently return wrong data.
• Network calls outside the browser (audio download, playurl API) use Node.js — PowerShell's curl.exe and Python's requests both get blocked by Bilibili CDN TLS fingerprinting on this platform.

---

Workflow Overview

                    ┌─ BE downloadSubtitles provider ──→ 投稿字幕 (timestamps)
                    │
BiliBili video ───┼─ BE downloadDanmaku provider ────→ danmaku density/peaks/sample
                    │
                    └─ No subtitles? ──→ Audio transcription fallback
                                         1. Get audio stream URL (Node.js → playurl API)
                                         2. Download .m4s audio
                                         3. ffmpeg → .m4a
                                         4. FunASR paraformer-zh → SRT with timestamps

---

Phase 1: Subtitles (preferred — via BE)

Use the one-shot evaluate below. It returns both subtitles and danmaku in a single call.

Setup

~/.kimi-webbridge/bin/kimi-webbridge status

{"action":"navigate","args":{"url":"https://www.bilibili.com/video/BV.../","newTab":true},"session":"bilibili"}

Wait 2–3 seconds for BE to fully initialize.

One-shot evaluate

(async () => {
  const pa = window.bilibiliEvolved.pluginApis;

  // ── Identifiers ──
  const s = window.__INITIAL_STATE__ || {};
  const vd = s.videoData || {};
  const bvid = vd.bvid || s.bvid || location.pathname.match(/BV[\w]+/)?.[0];
  const aid = vd.aid || s.aid;
  const pages = vd.pages || [];
  const p = parseInt(new URLSearchParams(location.search).get('p') || '1') - 1;
  const cid = pages[p]?.cid || vd.cid || s.cid || pages[0]?.cid;
  const title = vd.title || document.title;

  // ── Register providers ──
  pa.registerData('downloadVideo.assets', []);

  // ── Poll for providers (downloadDanmaku loads async, ~1-2s) ──
  function getProviders() {
    const g = pa.();
     .(g[]) ? g.() : g;
  }
   deadline = .() + ;
   providers = ();
   (!providers.( p. === ) && .() < deadline) {
      ( (r, ));
    providers = ();
  }

  
   subResult = { : , : , :  };
   subProvider = providers.( p. === );
   (subProvider) {
     {
       subAssets =  subProvider.([{ : {} }], { : , :  });
       subRaw = subAssets[].;
       subText;
       (subRaw  ) {
         buf =  subRaw.();
        subText =  ().(buf);
      }  {
        subText = (subRaw);
      }
       subtitles = .(subText);
       subLines = subtitles.( {
         totalSec = .(s.);
         h = .(totalSec / );
         m = .((totalSec % ) / );
         sec = (totalSec % ).(, );
         (h > ) {
            + h +  + (m).(, ) +  + sec +  + s.;
        }
          + m +  + sec +  + s.;
      });
      subResult = { : subtitles., : subLines.(), :  };
    } (e) {
      subResult = { : , : , :  };
    }
  }

  
   dmk = providers.( p. === );
   dmkAssets =  dmk.(
    [{ : { : (aid), : (cid) } }],
    { : , :  }
  );
   dmkRaw = dmkAssets[].;
   dmkText;
   (dmkRaw  ) {
     buf =  dmkRaw.();
    dmkText =  ().(buf);
  }  {
    dmkText = (dmkRaw);
  }
   danmaku = .(dmkText);

  
   bucketSize = ;
   buckets = {};
   ( d  danmaku) {
     b = .(d. /  / bucketSize) * bucketSize;
    buckets[b] = (buckets[b] || ) + ;
  }
   density = .(buckets)
    .( [(t), c])
    .( a[] - b[]);

  
   peaks = density.().( b[] - a[]).(, );

  
   sampleCount = .(, danmaku.);
   step = .(, .(danmaku. / sampleCount));
   danmakuSample = [];
   ( i = ; i < danmaku. && danmakuSample. < sampleCount; i += step) {
    danmakuSample.({
      : .(danmaku[i]. / ),
      : danmaku[i].
    });
  }

   totalDuration = danmaku. > 
    ? .(...danmaku.( d.)) : ;

  
   .({
    : , bvid, : (aid), : (cid), : p + , title,
    : subResult,
    : {
      : danmaku.,
      : .(totalDuration / ),
      density,
      : peaks.( ({ : pk[], : pk[] })),
      : danmakuSample
    }
  });
})()

Response shape

{
  "sub": {
    "count": 243,
    "text": "[0:00] 大家好\n[0:01] 这个视频...",
    "source": "投稿字幕"
  },
  "dmk": {
    "count": 247,
    "density": [[0,11], [30,2], ...],
    "peakMoments": [{"timeSec": 1170, "count": 16}, ...],
    "sample": [{    ...
  

If sub.count === 0, no 投稿字幕 was available. Proceed to Phase 2 below.

---

Phase 2: Audio Transcription Fallback (when no subtitles exist)

Use when Phase 1 returns sub.count === 0. This replaces the old "AI subtitle API" fallback which is unreliable — the /x/player/v2 API often returns stale/empty subtitle data.

Why this approach

The Bilibili CDN uses TLS fingerprinting that blocks curl and Python requests. Reliable paths:

Method	Works?	Notes
curl.exe with browser headers	❌ Exit code 35	SSL blocked
Python requests / urllib	❌ Blocked	Same reason
Node.js https.get()	✅	Use User-Agent + Referer headers
Browser fetch (in-page)	✅ but evaluate timeout <1s	Only for fast API calls
Node.js to download audio	✅	Backup URL is most stable

Step 1: Get audio stream URL via Node.js

node -e "
const https = require('https');
const url = 'https://api.bilibili.com/x/player/playurl' +
  '?bvid=BVxxxxxxxxx&cid=xxxxxxxxx&qn=0&fnval=4048&fourk=1&platform=web';
https.get(url, {
  headers: {
    'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36',
    'Referer': 'https://www.bilibili.com/'
  }
}, (res) => {
  let data = '';
  res.on('data', (chunk) => { data += chunk; });
  res.on('end', () => {
    const d = JSON.parse(data);
    if (d.code === 0 && d.data && d.data.dash && d.data.dash.audio) {
      const audio = d.data.dash.audio;
      const best = audio.reduce((a,b) => a.bandwidth > b.bandwidth ? a : b);
      console.log(best.baseUrl + '|' + (best.backupUrl ? best.backupUrl[0] || '' : ''));
    } else {
      console.log('ERROR: no dash audio data');
    }
  });
}).on('error', (e) => { console.log('ERROR: ' + e.message); });
"

The fnval=4048 flag requests DASH format with separate audio/video streams.

Step 2: Download audio (use backup URL — most reliable)

node -e "
const https = require('https');
const fs = require('fs');
const url = 'BACKUP_URL_FROM_STEP_1';
const file = fs.createWriteStream('output.m4s');
https.get(url, {
  headers: {
    'User-Agent': 'Mozilla/5.0 ...',
    'Referer': 'https://www.bilibili.com/'
  },
  timeout: 60000
}, (res) => {
  if (res.statusCode !== 200) { console.log('HTTP ' + res.statusCode); return; }
  res.pipe(file);
  res.on('end', () => { console.log('OK size=' + file.bytesWritten); });
}).on('error', (e) => { console.log('ERR: ' + e.message); });
"

Step 3: Convert to standard audio (stream copy, no re-encode)

ffmpeg -i output.m4s -c copy output.m4a -y

Step 4: ASR transcription via FunASR

set MODELSCOPE_CACHE=.\cache\modelscope
set MODELSCOPE_CREDENTIAL_PATH=.\cache\modelscope_cred
python audio2srt.py output.m4a --srt --model paraformer-zh --punc-model ct-punc --spk-model cam++

Step 5: Cleanup — remove intermediate audio files

del output.m4s output.m4a

The only artifacts worth keeping are the .srt subtitle file. The raw .m4s, the converted .m4a, and any .json debug output are all intermediate and should be cleaned up after the SRT is confirmed valid.

What the script does:

• Loads paraformer-zh (Chinese ASR, 220M params) + fsmn-vad + ct-punc + cam++
• Outputs SRT with per-sentence timestamps via res[0]["sentence_info"]
• CPU perf: ~15x realtime (21 min audio ≈ 90 sec)
• Handles ModelScope cache/credential path setup automatically

If cam++ is too slow to download, omit --spk-model and group res[0]["timestamp"] (per-character ms array) by punctuation boundaries manually.

If no sentence_info in the output: the spk_model triggers sentence segmentation. Without it, you only get raw text + per-character timestamps.

The audio2srt.py script

Located next to this SKILL.md file. Copy it into the current workspace or call it by its skill-directory path. It handles:

• Model loading with ModelScope credential path override (sandbox-safe)
• Multiple output formats: SRT, VTT, JSON
• Parsing sentence_info, timestamp, and raw text fallback
• Cleaning SenseVoice special tags (<|zh|>, <|HAPPY|>, etc.)
• Segment deduplication and time-sorting

After transcription completes, delete *.m4s and *.m4a — only the .srt is the final output.

---

Danmaku Fallback: Protobuf API (when BE's downloadDanmaku is absent)

(async () => {
  const be = window.bilibiliEvolved;
  const proto = await be.runtimeLibrary.protobufLibrary;

  const schema = {
    nested: {
      DmSegMobileReply: {
        fields: { elems: { rule: 'repeated', type: 'DanmakuElem', id: 1 } },
      },
      DanmakuElem: {
        fields: {
          id: { type: 'int64', id: 1 },
          progress: { type: 'int32', id: 2 },
          mode: { type: 'int32', id: 3 },
          fontsize: { type: 'int32', id: 4 },
          color: { type: 'uint32', id: 5 },
          midHash: { type: 'string', id: 6 },
          content: { type: 'string', id:  },
          : { : , :  },
          : { : , :  },
          : { : , :  },
          : { : , :  },
          : { : , :  },
        },
      },
    },
  };
   root = proto..(schema);
    = root.();

   raw =   ( {
    be..({
      : ,
      : ,
      : { :  },
      : ,
      :  (r.),
      :  (r.)
    });
  });
   decoded = .( (raw));
   .(decoded. || []);
})()

If segment 1 has data and segment 2 is empty, treat segment 1 as complete. Apply the same density/peak/sample logic from Phase 1.

---

Comments

https://api.bilibili.com/x/v2/reply/main?jsonp=jsonp&type=1&oid={aid}&mode={mode}&ps=20&next=0

Modes: 2 = newest, 3 = popular, 0 = popular (server-dependent). Use DOM to confirm render order.

---

Data Structures

投稿字幕 (from BE downloadSubtitles)

[
  { "from": 0.3, "to": 5.3, "sid": 1, "location": 2, "content": "文字内容", "music": 0 },
  ...
]

Danmaku

{ "progress": 40000, "mode": 1, "fontsize": 25, "color": 16777215,
  "content": "弹幕内容", ... }

progress is in milliseconds.

DASH audio stream (from playurl API)

• fnval=4048 → DASH response with dash.audio[]
• Audio IDs: 30216 (~65kbps), 30232 (~74kbps), 30280 (~128kbps AAC)
• Codec: mp4a.40.2 (AAC-LC, 48kHz, stereo)
• Format: .m4s = fragmented MP4, valid for ffmpeg stream copy

ASR output (sentence_info)

[
  { "start": 110, "end": 2190, "text": "话说大家有没有见过那种反驳型人格", "spk": 0 },
  { "start": 2330, "end": 3750, "text": "就是不管听到你说什么，", "spk": 0 },
  ...
]

start/end in milliseconds. spk is the speaker ID (from cam++ diarization).

---

Gotchas

Issue	What to do
downloadDanmaku not in providers	Polling loop handles this. If absent after 5s, fall back to Protobuf API.
Session-less evaluate	Always use "session":"bilibili" or similar.
registerData() without []	Always pass []. Provider callbacks crash without data.
Blob.text() hangs	Use arrayBuffer() + TextDecoder('utf-8').decode() instead.
CDN blocks curl.exe	Use Node.js https.get() with browser UA + Referer.
playurl API returns no dash	Don't use platform=html5. Use fnval=4048&platform=web.
Audio URL returns 403	URLs expire after ~5 min (deadline param). Fetch + download immediately in one pass.
ModelScope cache permission denied	Set MODELSCOPE_CACHE + MODELSCOPE_CREDENTIAL_PATH to writable workspace dirs. The audio2srt.py script handles this.
No sentence_info in output	Add --spk-model cam++ to trigger sentence segmentation.
Nested array from getData	Always flatten: Array.isArray(g[0]) ? g.flat() : g.
curl.exe downloading	Use node -e with https.get() instead.