cx-duplicate-detectionlisted

# Finding duplicate conversations Duplicates inflate volume, split context across tickets so agents answer without the full history, and make a customer feel unheard. They also distort every metric you compute — two tickets for one problem doubles the contact rate and halves the apparent first-contact resolution rate. This produces a **reviewable merge plan**, not a change. Applying it is a separate skill, deliberately. ## Why detection and merging are separate skills Merging is irreversible on every major helpdesk. So the decision and the action are split: ``` canonical export ──▶ cx-duplicate-detection ──▶ merge-plan.jsonl ──▶ <platform>-apply-merges (read-only) (read-only) (reviewable) (writes, --apply) ``` Three things this buys you, and the reason every mutation in this catalog works this way: - **An agent can safely propose a bulk merge** without being able to perform one. - **The plan is a diff a human can review** before anything happens, entry by entry. - **The applier re-validates against live state**, so a plan that has gone stale cannot cause a wrong merge. ## What counts as a duplicate Not "similar topic". A duplicate is **the same customer raising the same problem more than once**, which happens for mundane reasons: - The form was submitted twice, or an email retried. - The customer got no reply and tried another channel. - The customer replied to a closed ticket and the platform opened a new one. - An integration