Email Thread Reducer

Email threads accumulate quoted history exponentially, so naive summarization re-ingests the same content five times. Dedup-by-normalized-content first, then summarize, gives you linear input length and a cleaner output.

1. Parse the thread, normalize each message (strip quoted history, signature blocks, tracking footers). 2. De-duplicate identical or near-identical content using hash + fuzzy match. 3. Order messages chronologically with speaker and timestamp. 4. Feed the deduped sequence to the LLM with a strict output schema (Context / Decisions / Open Questions / Action Items). 5. Post-check: every Action Item must name an owner and a date — otherwise mark 'unassigned' explicitly rather than dropping.