[vandazia]

This PR implements "Native Multimodal Autonomous Sensory" capabilities, allowing the agent to natively fetch, understand, and process multimodal context blocks (e.g., images) directly via tool executions, without relying on external text-only transcriptions.

Key Enhancements

1. Standardized Content Blocks Exchange:

   • Upgraded tool return types in nanobot/agent/tools/base.py and registry.py from str to Any to support returning lists of multimodal blocks.
   • Tools can now return standard [{"type": "image_url", ...}, {"type": "text", ...}] structures.

2. Sensory Output in Tools:

   • ReadFileTool (filesystem.py): Detects image MIME types. If an image is read, it encodes the raw bytes into base64 and returns a native image block instead of failing or reading as UTF-8.
   • WebFetchTool (web.py): Prioritizes a streaming GET request to detect the content-type. If it detects an image, it captures it directly as a native image block, bypassing textual translation services (like Jina) that would otherwise strip the raw visual data.

3. Sensory Injection & History Persistence:

   • ContextBuilder.add_tool_result now blindly accepts structured blocks and injects them natively into the LLM context.
   • Updated AgentLoop._save_turn to safely persist tool results. Base64 payloads in image_url blocks are automatically stripped and replaced with [image: <path>] placeholders before saving to .jsonl history, preventing context bloat while maintaining full multimodal perception for the live inference round.

4. Prompt Awareness:

   • Explicitly updated the guidelines in ContextBuilder._get_identity and SubagentManager._build_subagent_prompt so the agent is consciously aware of its native multimodal perception capabilities and will proactively read visual resources when requested.

This transforms the agent into a "general sensory entity" capable of direct visual reasoning over local files and web images!