Add session condensation for structured checkpoint data

[james-in-a-box]

Context

The Entireio CLI consolidates shadow branch data and live transcripts into clean, permanent checkpoint records on each commit. The condensed record includes token usage, attribution data, modified files, and an optional AI-generated summary. This ensures checkpoint data is structured and queryable rather than raw dumps.

Current State

Our checkpoint system captures data on commit/session-end triggers, but the data is a relatively raw snapshot — full transcript, tool calls, files touched. There's no consolidation step that synthesizes this into a structured, queryable summary with computed metrics.

Proposal

Add a condensation step to checkpoint creation that produces a structured summary:

• Computed token usage and cost per session
• List of meaningful actions taken (not just raw tool calls)
• Files modified with change summaries
• Optional AI-generated session summary (intent, outcomes, learnings)
• Deduplicated and normalized data (e.g., collapsing repeated file reads)

This would make checkpoint data more useful for analysis and reporting without requiring consumers to parse raw transcripts.

Reference

See entireio/cli — session condensation in the strategy package and AI summarization logic.

Authored-by: egg