Capture assistant thinking for action tracing #52

Background

When diagnosing what happened during a session, it would be useful to have a record of the assistant's reasoning — not the step-by-step tool narration ("I need to open this file…"), but the substantive conclusions and decisions made during thinking.

Proposal

Optionally capture assistant thinking output and store it alongside turns, filtered or summarized to exclude mechanical tool-use narration. The goal is a lightweight audit trail for tracing "what did the assistant decide and why" after the fact.

Notes

• Raw thinking is likely too noisy to store verbatim
• A summary or filtered version (conclusions only) would be more useful
• Should be opt-in / configurable
• Low priority — can be deferred until there's a concrete need