64/588 chunks extracted so far (~1949 activities) but in a fresh
session we should switch the subagent model from Opus to Sonnet —
the task is structured JSON extraction with a fixed schema, no
complex reasoning needed, and Sonnet's 200K context easily fits the
~25k-token prompt and ~20k-token output per chunk. Document captures
the exact resume procedure: pending-chunk discovery, the Agent call
template with model:"sonnet", and the finalization steps
(validate -> build_database -> needs_review bulk merge).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
