polish

The passes run in a fresh subagent, never in the session that loads this skill. That session is usually the work's author, reading the diff with the change's history in mind, so a comment narrating the edit, or two blocks it remembers as different, reads fine to it and not to a reviewer. A subagent has only the diff and the plan, as the reviewer will.

So launch one general-purpose subagent, in the foreground (run_in_background: false) since it edits, commits and pushes the working tree, and touch nothing until it returns. Its prompt carries:

• @.claude/skills/polish/passes.md, to load and follow — the whole procedure, which this session does not read;
• the argument, verbatim;
• the plan under docs/plans/, when the branch has one, as the record of intent.

Nothing else: no account of what was built or why.

It reports, and the session relays. Answer the intent questions this session's context settles and act on them, committed under the type the edit would carry anyway: a polish: subject would put the edit below the next run's floor (passes.md § "The floor"), unreviewed. The rest go to the operator with /dry's calls.

Putting back what a run changed, on the operator's ask, is itself a polish: commit — polish: restore <what>. That run already read what it restores, so the commit becoming the next run's floor keeps the cut from being made again. It goes straight on top of the run: anything committed in between would sink below the floor with it, unreviewed.

Where it runs

• @.claude/skills/go/SKILL.md Step 3 — after implementing, before the PR.
• @.claude/skills/finalize/SKILL.md, ahead of its numbered steps — the backstop for work that reached a PR without passing through /go. That skill owns why it goes first.
• On the operator's ask, at any point in a session — the entry the other two exist to make unnecessary and routinely don't. A task asked for and done directly ends with the passes unrun unless someone names them.
• @.claude/skills/update-muthur/SKILL.md Step 8, over a port.