Input Router

Input Router

Use this skill as the unified input entry layer for the system.

This skill is a system-level dispatcher. It does not implement document parsing, transcription, table extraction, PPT grounding, archive unpacking, or meeting summarization logic by itself.

However, it does not stop at routing analysis. After identifying the correct input type, it must immediately invoke the correct downstream skill and continue the task until that downstream grounding workflow is completed.

What This Skill Does

This skill must:

1. inspect the input path,
2. determine the file type strictly by extension,
3. select the correct downstream grounding pipeline,
4. invoke the corresponding existing skill,
5. and continue until the selected downstream skill's workflow has completed.

What This Skill Does Not Do

This skill must not directly implement the grounding logic that belongs to downstream skills. It must not:

• transcribe audio by itself
• extract content from documents by itself
• parse tables by itself
• ground PPTX files by itself
• unpack archives by itself
• write grounded.md by bypassing the downstream skill

Those tasks belong to the selected downstream skill.

Routing Principle

Route inputs strictly by file extension or URL pattern.

URL Detection (Priority 1)

If the input is a URL instead of a local file:

URL routing workflow:

1. Detect URL pattern
2. Invoke remote-input skill to download
3. Use returned local path for downstream routing
   • If merge_failed: true, use audio_path instead of path
4. Continue with extension-based routing

Note on merge failure: When remote-input returns merge_failed: true, the actual file for downstream routing is the audio file (.webm etc.), which routes to meeting-audio-grounding instead of meeting-video-grounding.

File Extension Detection (Priority 2)

For local files, route strictly by file extension:

Do not use filename semantics, directory names, or inferred task intent to override the extension-based mapping.

Supported Routing Table

URL Patterns (Remote Inputs)

File Extensions (Local Inputs)

Required Workflow

When using this skill, you must follow this workflow:

1. Detect input type:

   • If input contains URL patterns (arxiv.org, youtube.com, youtu.be, bilibili.com, b23.tv): goto URL handling
   • Otherwise: goto file handling

2. URL handling:

   1. Match URL pattern to remote-input
   2. Invoke remote-input skill
   3. Receive returned local file path
   4. Use the local path for downstream routing

3. File handling:

   1. Inspect the input path
   2. Determine the file extension
   3. Match the extension to the routing table
   4. Select the corresponding downstream skill

4. Continue:

   1. Immediately invoke the selected downstream skill
   2. Follow the downstream skill's own workflow strictly
   3. Do not stop after only reporting the routing decision

URL Detection Patterns

arXiv URLs

• Pattern: https?://(?:www\.)?arxiv\.org/(abs|pdf)/
• Examples:
  • https://arxiv.org/abs/2301.07041
  • https://arxiv.org/pdf/2301.07041.pdf
  • http://arxiv.org/abs/2301.07041

YouTube URLs

• Pattern: https?://(?:www\.)?youtube\.com/(watch|shorts)/ or https?://youtu\.be/
• Examples:
  • https://www.youtube.com/watch?v=dQw4w9WgXcQ
  • https://youtu.be/dQw4w9WgXcQ
  • https://www.youtube.com/shorts/abc123

Bilibili URLs

• Pattern: https?://(?:www\.)?bilibili\.com/video/ or https?://b23\.tv/
• Examples:
  • https://bilibili.com/video/BV1xx411c7JZ
  • https://www.bilibili.com/video/av12345678
  • https://b23.tv/abc123

Merge Failure Handling

When downloading video from URL (YouTube or Bilibili), the downloader may produce separate video and audio files instead of a merged video file (merge failure). This happens when:

• The video has separate video and audio streams that couldn't be merged
• Common pattern: .mp4 file without audio + .webm audio file

Detection and Recovery

After remote-input downloads video content:

1. Check return values for merge_failed field:

   • merge_failed: false → Use path as normal (video with audio)
   • merge_failed: true → Use audio_path instead of path

2. Reroute based on actual file type:

   • If audio_path is used → route to meeting-audio-grounding
   • If path is used → route based on extension (video → meeting-video-grounding)

Routing After Merge Failure

When merge_failed: true, the local file is typically a pure audio file (.webm, .mp3, etc.):

Workflow Example

User: https://youtube.com/watch?v=xxx
           ↓
remote-input downloads video
           ↓
Returns: {
  "path": "video.mp4",           // video without audio
  "audio_path": "audio.webm",    // separate audio file
  "merge_failed": true
}
           ↓
Since merge_failed=true, use audio_path
           ↓
audio.webm extension → .webm
           ↓
Route to: meeting-audio-grounding

Completion Rule

The task is not complete after:

• identifying the URL pattern or extension,
• naming the downstream skill,
• or reporting the selected pipeline.

The task is complete only after:

• for URL inputs: remote-input has successfully downloaded the content
• the selected downstream skill has been invoked,
• its required workflow has been followed,
• and the expected grounding result for that input type has been produced.

Unsupported Inputs

If the input is neither a supported URL nor a supported file extension:

• report that the input is unsupported by the current router,
• do not guess a pipeline,
• do not force the file through an unrelated skill.

Dispatch Behavior

The router is an entry skill, not a stopping point.

For URL inputs:

• do not stop after saying remote-input; run it
• use the returned local path to continue routing

For audio inputs, do not stop after saying meeting-audio-grounding; run it For video inputs, do not stop after saying meeting-video-grounding; run it For document inputs, do not stop after saying document-grounding; run it For table inputs, do not stop after saying table-grounding; run it For PPT inputs, do not stop after saying pptx-grounding; run it For ZIP inputs, do not stop after saying archive-grounding; run it

Practical Meaning

From the user's perspective, this skill should behave like a unified input layer:

For URL inputs:

• the user provides a URL (arxiv paper or YouTube video),
• this skill downloads the content to data/raw_inputs/remote/,
• the task continues through the appropriate grounding pipeline.

For local file inputs:

• the user provides one input file,
• this skill selects the correct downstream grounding pipeline,
• and the task continues through that pipeline until grounding is actually completed.

The user should not need to manually re-enter another instruction just because the router has already identified the correct skill.