Triage findings

You have a findings.json (from extract-findings). Filing every row blindly creates a noisy, unprioritized backlog that nobody can action. This skill is the human-in-the-loop gate: show the shape of the findings, agree on a scope and order, emit a smaller set, then hand off to classify-work-items. The win is an auditable wave structure — wave 1 is the high-value, confirmed work; later waves pick up the rest after review.

This is a reasoning/filtering step — there is no script to run. You read JSON, summarize, and write a smaller JSON. See references/data-contracts.md for the exact findings.json shape (the key, severity, status, kind, section fields).

1. Read and profile the findings

Read the findings.json the user points at (default name findings.json in the working folder). Then show counts before asking anything — people choose well when they can see the distribution. Break it down across the dimensions that matter for triage: