loose-endslisted

# Loose Ends You start a branch, get most of the way, and write a file so the thinking is not lost. Then you move on. Weeks later there are forty of those files: some superseded, some duplicated in two branches, some describing a repo that no longer exists — and none of them read since the day they were written. This skill finds all of them and makes you judge each one. Four verdicts, one recommended per item, and the answers are recorded so a file judged once never comes back. If it's a person driving rather than an agent, they don't need any of the steps below — `bin/loose-ends` runs the whole pipeline itself and just asks plainly when it needs a decision. The procedure that follows is for an agent, which wants the stages kept separate: read the scan summary before deciding whether a review session is even worth opening, and check a big fold's extraction against the source before recommending it. ## Overview Four stages. Everything before the last one is read-only. ``` scan.py → candidates.json → review.py or build_page.py → verdicts.json → apply.py (git) (facts) (your judgment) (decisions) (does it) ``` `scan.py` has no model in the loop — it is deterministic, so two runs on an unchanged repo produce the same file. Your job is the parts it cannot do: reading what a note actually says, and drafting what survives it. ## Where it looks Five hiding places, because the fourth and fifth never show up in `ls`: 1. the working