migrate

~/.claude/skills is itself a symlink into this repo's skills/ directory (set up by setup.js), so the migration CLI is reachable from anywhere via a home-relative path — no need to know where this repo is checked out:

node ~/.claude/skills/migrate/migrate.js

On Windows (PowerShell): node "$env:USERPROFILE\.claude\skills\migrate\migrate.js"

Run this command from the directory of the project you want to migrate (its .claude/ is what step 2 acts on) — cwd is what matters, not where migrate.js itself lives. From within this config repo's own root, npm run migrate is a shortcut for the same command.

This does the following, all idempotent (safe to re-run, no-op once current):

1. Repo links — removes any ~/.claude / ~/.codex symlinks that point into this repo but no longer correspond to an entry in CLAUDE_LINKS / CODEX_LINKS (e.g. a renamed/removed skill or config), then re-runs the normal link-creation pass for the current layout.
2. Orphaned CLI aliases — removes wrapper files this repo wrote (recognised by their marker comment) whose name is no longer in WRAPPER_NAMES (scripts/setup/install-cli-wrappers.js). This is what stops a retired provider leaving <name>.cmd behind after its <name> is gone. It sweeps the claude bin dir, so on a Codex-only host it is a no-op.

3. Gitignore hygiene — across the current repo and every nested git repo: (a) normalizes .gitignore to the depth-agnostic .claude/ template this skill owns (CLAUDE_GITIGNORE_TEMPLATE — base **/.claude/** exclusion plus the agents/skills/commands/workflows/settings/output-styles/rules/memory/docs re-includes, with MEMORY.md + _meta.json ignored last), preserving all unrelated lines; (b) untracks (via git rm --cached, keeping files on disk) anything the refreshed rules now ignore but is still committed; (c) deletes stray .gitignore files nested inside a .claude/ tree. This skill is the single writer of the template; rem only ensures ignores for the two files it generates (MEMORY.md, _meta.json).

   Overwriting .gitignore is confirmed first. When any repo's .gitignore differs from the template, an interactive TTY run asks: [O]verwrite (default) / [A]I edit (script writes nothing — you hand-merge afterward) / [S]kip. Choose AI edit / Skip for third-party projects whose .claude/ holds content outside the template's allowlist (e.g. hooks/, settings.local.json), since overwrite + untrack would stop tracking it. Pass --gitignore=overwrite|skip|ai to set it non-interactively.

   AI edit vs. Skip differ only in your follow-up — the script's file behavior is identical (it leaves .gitignore untouched, untracking only per existing rules). Skip abandons the repo. AI edit obligates you to finish the job: the run prints an AI-EDIT REQUIRED block listing the template lines and each target .gitignore. After the run, for every listed file: Read it, splice the template block in as one contiguous group, drop superseded managed lines, keep the repo's own rules, write it back with Edit, then git rm --cached any now-ignored tracked files in that repo. Do not end the turn with the merge undone — an unactioned AI edit is just a Skip.

   When running this skill for the user, dry-run first, show which repos would change, and confirm the mode before applying.

4. Project .claude/ — for every cc-market plugin that provides a migrations/migrate.mjs (e.g. rem's memory/frontmatter normalization, sharp-review's legacy finding-file consolidation), runs it against the current project.

Preview repo-link changes only, without touching anything:

node ~/.claude/skills/migrate/migrate.js --dry-run

(--dry-run covers steps 1–3 fully — it prints which links, wrappers and files would be removed/untracked, without touching them; for step 4 it lists which plugins have migrations but does not run them, since they are write-only and self-detecting.)

If a plugin's migration reports changed: true, read its summary lines — they describe exactly what was moved/rewritten.