mirror of
https://github.com/Significant-Gravitas/AutoGPT.git
synced 2026-02-06 12:55:05 -05:00
dcb3550910
When CoPilot executes blocks that produce binary outputs (images, PDFs), the data is now automatically detected and saved to the user's workspace, replacing raw base64 data with workspace:// references. Uses content-based detection (not field-name based) because: - Code execution blocks return base64 in stdout_logs, not structured fields - The png/jpeg/pdf fields only populate from Jupyter display mechanisms - Other blocks use various field names (image, result, output, etc.) Detection strategy: 1. Data URI detection with mimetype whitelist 2. Raw base64 with magic number validation (PNG, JPEG, PDF, GIF, WebP) 3. Size threshold > 1KB to filter tokens/hashes Features: - Scans ALL string values recursively, regardless of field name - Content deduplication within single block execution via SHA-256 hash - Graceful degradation (original value preserved on save failure) - ~97% token reduction observed (17k -> 500 tokens for PDF generation) Closes SECRT-1887