deep-context

Run the deep-context workflow after your import completes. It collects messages per person, synthesizes dossiers, merges duplicates, and walks you through worth and LinkedIn review before building the final merged people list and search index.

Deduplication runs in tiers: a free deterministic pass merges people sharing a name plus phone or email, then a paid LLM judge settles only the pairs the deterministic pass could not resolve. A free parents step then rebuilds the canonical person layer.

Collection, dedupe, and realization are free and local. Paid stages like synthesis, the duplicate judge, and identity lookups always show a dry-run cost estimate first; small estimates auto-approve under set thresholds and larger ones require explicit approval.

Yes. The lookup, check, validate, and review commands run standalone as narrow read-only paths, so asking who a person is by name, phone, or email never triggers the full staged build.

Raw message samples stay in a gitignored local directory, and dossiers contain synthesized facts rather than verbatim messages. WhatsApp group bodies are never read, and iMessage collection requires Full Disk Access on your machine.

The review server holds a single-writer session lock and serves the Python code loaded at startup, so each staged launch stops any running server and starts fresh to avoid stale code. Review state is file-driven, so nothing is lost on restart.