Subagents can't write! Context gets consumed

[electro56435]

The main context eating culprits, in order:

1. 20 subagent task notifications — Each of the 20 review agents returned its full JSON result as text output in the notification (because  
Explore agents are read-only and couldn't write files). Each notification contained 150-400 lines of JSON embedded in the message. That's
~20 large task result blobs in the conversation.                                                                                            
2. 20 Write tool calls — Each JSON file was written inline, meaning the full content appeared in the conversation twice (once in the      
notification, once in the Write call).
3. Verbose desloppify CLI output — The scan and import commands produce long formatted output that gets included verbatim.

The root cause was the agent type mismatch: Explore agents are read-only, so they couldn't write their results to disk and fell back to
returning everything as text output. If general-purpose agents had been used instead, they could have written the files directly and only
returned a brief confirmation — keeping the notifications tiny.