"""Carry trusted VM screenshots through the kernel's native vision pipeline.""" import base64 import json def computer_result(name, value): if not name.startswith('computer_') or not isinstance(value, dict) or value.get('isError'): return json.dumps(value, ensure_ascii=False) content = value.get('content') images = [part for part in content if isinstance(part, dict) and part.get('type') == 'image'] if isinstance(content, list) else [] if not images: return json.dumps(value, ensure_ascii=False) # The shared Box bridge produces exactly one PNG per capture. Invalid media # must be an explicit error, never a huge base64 text result/spillover file. try: if len(images) != 1 or images[0].get('mimeType') != 'image/png': raise ValueError('unsupported screenshot envelope') data = images[0].get('data') if not isinstance(data, str) or len(data) > 32 * 1024 * 1024: raise ValueError('invalid screenshot size') raw = base64.b64decode(data, validate=True) if not raw.startswith(b'\x89PNG\r\n\x1a\n'): raise ValueError('invalid screenshot bytes') from agent.image_eviction_policy import image_dimensions_from_bytes from tools.computer_use.backend import CaptureResult from tools.computer_use.tool import _capture_response dimensions = image_dimensions_from_bytes(raw) if not dimensions or not all(0 < n <= 16384 for n in dimensions): raise ValueError('invalid screenshot dimensions') except (ValueError, TypeError): return json.dumps({'isError': True, 'error': 'VM screenshot payload is invalid; no other computer was used.'}) note = '\n'.join(part['text'] for part in content if isinstance(part, dict) and part.get('type') == 'text' and isinstance(part.get('text'), str)) # Reuse Hermes native/aux-vision routing, provider limits and private image # cache. This only renders the returned pixels; it never captures the Host. return _capture_response(CaptureResult(mode='vision', width=dimensions[0], height=dimensions[1], png_b64=data, png_bytes_len=len(raw), image_mime_type='image/png', note=note))