What problem does it solve? Claude Code session transcripts are the only record of decisions, uncommitted work, and unresolved incidents, but they routinely reach 130-150 MB and cannot be opened directly, and embedded boilerplate (CLAUDE.md, MEMORY.md) makes naive keyword searches produce thousands of false positives. ## Core Features & Use Cases - Streaming transcript analysis: jq and grep recipes for extracting session titles, date ranges, user requests, shipped-work markers, and end-of-session outcomes from files up to 150 MB without loading them into context. - Boilerplate calibration and classification: verify keywords against CLAUDE.md and MEMORY.md, calibrate hit-count thresholds, and tag sessions as core, partial, or incidental to eliminate false positives. - Fork detection and fan-out strategy: detect forked session lineages to avoid double-counting, sweep nested subagent transcripts, and split work across agents by directory and filename range. - Use Case: When asked "what have we done on project X?", sweep all project directories, calibrate against boilerplate, classify the high-scoring sessions, and produce a structured JSON history with date ranges, workstreams, and key outputs. ## Quick Start Reconstruct everything we have done on this project by mining my Claude Code session transcripts in ~/.claude/projects, calibrating out boilerplate hits first.
