القائمة
File-Augmented Retrieval — generate persistent .meta sidecar files for PDFs, images, spreadsheets, videos and more, making every binary file readable to AI coding agents.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
| name | far |
| description | File-Augmented Retrieval — generate persistent .meta sidecar files for PDFs, images, spreadsheets, videos and more, making every binary file readable to AI coding agents. |
| allowed-tools | Read, Write, Edit, Bash, Glob |
| user-invocable | true |
"Making Every File Readable to Coding Agents via Persistent .meta Sidecars"
FAR is a file augmentation protocol that generates persistent .meta sidecar files for binary documents (PDF, DOCX, XLSX, PPTX, Images, etc.). This allows AI coding agents (like OpenClaw, Cursor, GitHub Copilot) to "read" non-text files directly from the file system without requiring external RAG infrastructure.
Current Version: 1.0.0 Author: Kelly Peilin Chan
AI agents operating in a repository can read code (.py, .js, .md), but they are blind to 30-40% of critical context stored in binary formats:
requirements.docxarchitecture.pngbudget.xlsxagreement.pdfWhen an agent encounters budget.xlsx, it sees opaque bytes. It cannot reason about the content.
FAR solves this by generating a human-readable and machine-readable .meta file next to every binary file.
Example:
project/
├── budget.xlsx (Binary, opaque)
└── budget.xlsx.meta (Markdown, readable)
The agent simply reads budget.xlsx.meta to understand the spreadsheet. No vector database, no API calls, no runtime overhead.
FAR extracts text and structure from a wide range of formats:
.pdf): Full text extraction with layout preservation. OCR fallback for scanned PDFs. Embedded image extraction (pdfimages + OCR/Vision)..docx, .doc): Text extraction..xlsx): Sheet data converted to Markdown tables..pptx): Slide text extraction..png, .jpg, .jpeg, .gif, .bmp, .webp): OCR (Tesseract) + Apple Vision (on-device, macOS) + AI Vision (GPT-4o) if configured..mp4, .mov, .mp3, .wav, .m4a, .flac): Technical metadata (FFprobe) + Apple Vision video-frame analysis (macOS) + AI Transcription (Whisper) if configured..csv): Data rendered as Markdown tables (up to 100 rows)..ipynb): Markdown cells, code cells, and outputs..epub): Full text extracted from all chapters in spine order..tar, .tar.gz, .tgz, .bz2, .xz): File listing with sizes..eml, .msg): Headers, body text, and attachment list..rtf): Plain text extraction via control word stripping..db, .sqlite, .sqlite3): Table schemas + latest 20 rows per table as Markdown tables..parquet): Schema and row count via pyarrow (optional). [Metadata only].fig, .sketch, .xd): File size and page/canvas count. [Metadata only].py, .js, .ts, .go, .rs, .md, .json, ...): Direct content mirroring.FAR is designed for speed. It uses a two-layer caching mechanism:
For directory summaries, FAR also avoids noisy rewrites: if .dir.meta summary content is unchanged, it keeps the file untouched (no extracted_at churn).
.dir.meta)FAR generates a .dir.meta file in every directory, providing a high-level summary of all files within. This allows agents to "browse" a folder and understand its contents without reading every single file.
FAR automatically handles Git LFS pointer files. It will attempt to pull the real content before processing, ensuring you don't get empty metadata for LFS-tracked files.
The skill is pre-installed in the OpenClaw workspace:
~/.openclaw/workspace/.agents/skills/far/far_gen.py
A symlink is available as far:
far [directory_or_file]
Scan current directory (Recursive):
far
Scan specific directory:
far ~/Documents/projects/files
Process single file:
far report.pdf
Force regeneration (Ignore cache):
far . --force
To enable AI features (Audio Transcription, Image Description), create a .env file in the skill directory or your home folder ~/.far.env.
Copy .env.example as a template:
OPENAI_API_KEY=sk-your-key-here
# Optional
OPENAI_BASE_URL=https://api.openai.com/v1
# Optional (macOS default: enabled)
FAR_USE_APPLE_VISION=1
# Optional (macOS default: enabled)
FAR_USE_MACOS_METADATA=1
# Optional tuning
FAR_APPLE_VISION_TIMEOUT=25
FAR_APPLE_VISION_MAX_FRAMES=6
If API keys are missing, FAR gracefully falls back to local tools (Tesseract, FFprobe), and on macOS it also uses Apple Vision + Spotlight metadata on-device by default.
Create a .farignore file in your project root to exclude files or directories from scanning.
# .farignore
node_modules
.git
secrets/
*.tmp
Add one rule to AGENTS.md or your system prompt:
## File Reading with FAR
When navigating a directory, first read `.dir.meta` for a high-level overview
before opening individual files.
When you encounter a binary file you cannot read directly
(.png, .pdf, .xlsx, .mp4, .fig, .db, etc.), check for a .meta file
beside it — it contains the extracted content as Markdown.
Navigation strategy:
1. Read <dir>/.dir.meta to understand directory structure
2. Read <file>.meta to access binary file content
3. Fall back to reading the source file only if no .meta exists
Each .meta file follows a strict format with a YAML frontmatter header and Markdown body.
--far_version: 1
source:
sha256: 5a1cc2b8d...
mime: application/pdf
size: 129509
mtime: 1708845210.5
extract:
pipeline: far_gen_v14
extracted_at: 2026-02-27T10:00:00Z
deterministic: true
layout:
pages: 24
---
# filename.pdf
[Extracted Content Here...]
Cache invalidation: a .meta is stale if sha256 or pipeline version has changed.
Roadmap:
.dir.meta).farignore