resolve-merge-conflicts

TL;DR

The point is to keep the agent's context small. A bundled Python script first summarizes the conflict: which files are unmerged, which index stages exist and how many hunks each file has. The agent then takes one file at a time, asking the script for just that file: nearby lines, the ours, base and theirs versions of each hunk, and a compact diff between ours and theirs. It takes one side whole with git checkout --ours or --theirs, or edits the file and removes the markers, reading more of the file only when the compact view isn't enough to decide. Conflicts that leave no markers in the file, such as add/add or modify/delete, are shown from the index stages. Options print every conflicted file at once, output JSON, or change how many context lines and how much output it shows. It's done when no unmerged paths and no conflict markers remain, targeted tests, builds or linters have run for the touched area, and the resolved files are staged.

Summarizes the skill at 8792394.