wick-migrate

name	wick-migrate
description	Make a Wick agent fully portable between machines — consolidate a host's auto-memory into the Wick memory/ layer, disable the shadow portably, and relativize all paths and tooling. Guided, gated, archive-don't-delete. Use when moving an agent to a new machine, or to converge a single-machine agent onto one governed memory layer. Full reference in PORTABILITY.md.
license	MIT

Wick — Migrate (portability)

Consolidate everything into one portable, git-mergeable memory layer and disable the machine-keyed shadow. This skill is the executable form of PORTABILITY.md — read that for the full rationale and the failure modes; this is the operating checklist with the gates.

When to invoke

• Moving a Wick agent to a second machine (the memory/ folder must travel whole)
• A host keeps its own auto-memory (Claude Code, Cursor) and you want one governed layer
• Convergence: corrections or facts live only in the host layer and you want them in memory/

The load-bearing facts

• Freeze before you migrate. Auto-memory is live. Disable it (Step 1) and verify in a fresh session before consolidating — else the source mutates under you and the layers diverge invisibly. This makes it a two-session job, not one.
• Verify behavior, not form. "The setting says false" and "the path is relative" are the two ways this silently fails. Confirm by a fresh session and by resolving paths from the launch CWD.
• Archive, don't delete. Discards and prunes go to memory/_archived-auto-memory/ (the reversibility net), never to nothing.

Procedure (gates marked 🚦 — get explicit sign-off)

1. 🚦 Freeze. Find ~/.claude/projects/<path-encoded-folder>/memory/. If present, write { "autoMemoryEnabled": false } to the project's in-folder .claude/settings.json (verify the key against current docs; check settings.local.json doesn't re-enable). Then start a fresh session and confirm no auto-memory loaded. Do the rest there.
2. Inventory both layers (Wick memory/ + frozen auto-memory). Note duplicates, convention-bleed (feedback_X.md), and what's already in CLAUDE.md. Present it.
3. 🚦 Sort auto-memory: Keep / Merge / Discard / Conflict. Never auto-resolve a conflict — surface both versions, recommend, let the user decide. Present the sort.
4. 🚦 Audit Wick memory: redundancy, overlap, bloat. Never flag by mtime — judge by whether content is still true. Present findings.
5. Consolidate feedback into memory/feedback/*.md topic files — only if it isn't already in CLAUDE.md. Strip YAML frontmatter; keep tables/values; <200 lines each.
6. Write kept/merged content (read target first; append, dated). Project detail → projects/<name>/ with a status line + pointer.
7. Archive the raw auto-memory memory/ subfolder → memory/_archived-auto-memory/ with an INERT README. Not the *.jsonl transcripts.
8. 🚦 Prune now-redundant Wick files (approval required). Immutable logs stay as written.
9. Fix dangling references in memory/, CLAUDE.md, and tooling.
10. Rewrite memory/index.md — relative paths, <100 lines, no "as of " blocks, archives excluded.
11. Update CLAUDE.md memory protocol — deterministic session-start loading, feedback capture, write discipline, portability note (memory/ is the single source of truth).
12. Relativize to the launch directory — pin the frame in CLAUDE.md; run node tools/wick-path-audit.mjs. Watch the two "relative-but-resolves-wrong" traps (own-folder-name doubling; missing memory/ prefix).
13. Tooling sweep — scripts self-locate; .mcp.json uses ${CLAUDE_PROJECT_DIR}; no plaintext secrets (node tools/wick-scrub.mjs); hooks use relative cwd.
14. Verify behavior — fresh session loads nothing; sample paths resolve from launch CWD; all four scanners green; nothing lost; no conflict silently resolved. Report.

What this skill will never do

• Pass a 🚦 gate without explicit user sign-off.
• Delete the only copy of anything (archive first).
• Auto-resolve a conflict between the host layer and CLAUDE.md.
• Claim "disabled" or "portable" from text inspection — only from a fresh session and real path resolution.