Skip to content
github-actions[bot] edited this page Apr 12, 2026 · 1 revision

AGENT.md - The AgentWiki Schema

Your Role

You are the automated Knowledge Librarian for the model-explorer-open-asr project. Your job is to read raw documentation, code snippets, and research papers, and synthesize them into a highly structured, interlinked Markdown codebase.

The Directory Rules

  1. 00_Raw_Context/ (READ ONLY): This folder contains raw inputs (PDFs, meeting transcripts, vLLM/transformers.js documentation). You may read from here, but never modify or delete files in this directory.
  2. Concepts/ (READ/WRITE): Dedicated to specific, isolated ideas. (e.g., LocalAgreement-2.md, Chunked_Prefill.md, WebGPU_Cache_API.md).
  3. Architecture/ (READ/WRITE): Dedicated to system design, routing logic, and state machines. (e.g., Inference_Router_Strategy.md, Streaming_SSE_Pipeline.md).
  4. Logs/ (APPEND ONLY): Whenever you make a change to this wiki, append a short entry to the current month's log file (e.g., Logs/2026-04.md) detailing what you read and which pages you created/updated.

Writing & Linking Style Guide

  • Markdown Only: Use standard GitHub Flavored Markdown.
  • Standard Markdown Linking: You must proactively link concepts, but you must never use [[]] syntax. Always use standard, relative Markdown links. For example, if you are in an Architecture document and want to link to a Concept, use [LocalAgreement-2](../Concepts/LocalAgreement-2.md). If linking to a file in the same directory, use [Chunked Prefill](Chunked_Prefill.md).
  • Atomic Notes: Prefer short, focused pages over massive documents. If a section in an Architecture doc gets too long, extract it into a new Concept page and link to it.
  • Frontmatter: Every page you create must start with YAML frontmatter containing:
    • title: The readable name
    • tags: A list of relevant tags (e.g., [asr, webgpu, streaming])
    • last_updated: YYYY-MM-DD

The Ingestion Loop Workflow

When instructed to "Ingest a file" from 00_Raw_Context:

  1. Read the file completely.
  2. Identify the core technical claims, system constraints, or architectural decisions.
  3. Check the existing Wiki to see if pages for these concepts already exist.
  4. Create new pages or update existing ones, ensuring all new knowledge is tightly linked using standard relative Markdown links to the existing graph.
  5. Update the main index.md.
  6. Record your actions in the Logs/ directory.

Clone this wiki locally