cktovd

name: cktovd description: "Migrate from claudekit (ck) to the vd-cli control plane — install clean-room hooks, convert .ck.json to .vd.json, audit CK_*→VD_* env consumers, and move a repo's plans/ artifacts into feature-first .workbench folders under /.workbench/features//{plans,reports,journals,visuals,state}. Use when the user says 'cktovd', 'migrate to vd', 'migrate plans to .workbench', 'enable the .workbench umbrella', or 'switch this repo off claudekit'." license: MIT argument-hint: "[repo-path] [--machine] [--check]" metadata: author: vanducng version: "1.1.0"

cktovd — claudekit → vd-cli migration

Migrate two layers, in order:

1. Machine (once): vd-cli owns the Claude Code control plane — clean-room hooks in ~/.claude/hooks/, config in ~/.claude/.vd.json.
2. Repo (per project): opt into the feature-first .workbench umbrella via local <git-root>/.vd.json and physically move plans/ artifacts into per-feature folders.

Detect scope from the argument: --machine → machine layer only; a repo path (default: cwd git root) → repo layer (runs a machine preflight first); --check → verify only, change nothing.

Layout contract (source of truth: vd-cli hooks/lib/paths.cjs)

The umbrella dir is named tool-neutrally (.workbench, not .vd) because multiple agents share the repo. Git worktrees are a separate top-level <git-root>/.worktrees/ (not under .workbench); artifacts written from inside a worktree resolve to the main repo's .workbench/ (they survive git worktree remove), while docs/ stays branch-local.

Config precedence: DEFAULT ← global ~/.claude/.vd.json ← project <git-root>/.vd.json. There is no .ck.json read fallback (removed): vd reads .vd.json only. A lingering .ck.json without its .vd.json is no longer silently honored — it raises a migration error ("run the cktovd skill / rename to .vd.json"). So the cutover is: create the .vd.json (from .ck.json if present). Once .vd.json exists the old file is inert (ignored); it's safe to delete — the master backup is your real safety net. vd never writes .ck.json.

paths.umbrella must be a relative path with no ../empty segments that resolves inside the git root. An invalid value (absolute, traversal, empty) never errors — it's silently coerced to null, which restores the byte-identical legacy layout, so a bad value presents as "migration didn't take effect". A plausible typo (.wrok) passes sanitization and creates a wrong-named umbrella dir. Umbrella resolution anchors via git rev-parse --show-toplevel — a non-git dir never activates it; git init first.

Machine layer

1. Master backup (the cutover is destructive; the full-restore recipe depends on this):

   ts=$(date +%Y%m%dT%H%M%S); bak=~/.claude-ck-removal-backup-$ts
   mkdir -p "$bak" && cp -Rp ~/.claude/hooks "$bak"/ && cp ~/.claude/settings.json "$bak"/
   cp ~/.claude/.ck.json "$bak"/ 2>/dev/null || true
   

2. Preflight: vd --version (hooks support landed in 2.4.0; the .vd.json rename is post-2.5.0 — brew upgrade vd if older). Never patch deployed hooks in place — ck's ck update clobbers its files, and vd's are redeployed by vd install hooks.
3. Cut over:

   npm uninstall -g claudekit-cli
   vd install hooks
   grep -n '\$HOME/.claude/hooks' ~/.claude/settings.json   # all entries must show literal $HOME
   

   vd install hooks is idempotent: byte-identical files are skipped, a differing vd-owned file is backed up once as <name>.bak.<UTC-ts>.cjs, unknown files are never touched. settings.json edits are surgical (only vd's hook entries + the top-level statusLine key are patched; a one-time settings.json.bak is taken first). Registered commands keep $HOME literal — the hook runner shell-expands it; never bake a personal absolute path in. Caution: registration matches existing entries by .cjs filename substring and rewrites them to vd's canonical command — a custom wrapper referencing the same filename gets clobbered.
4. Config rename (required — prevents the migration error): [ -f ~/.claude/.ck.json ] && [ ! -f ~/.claude/.vd.json ] && cp ~/.claude/.ck.json ~/.claude/.vd.json. With the fallback gone, a global ~/.claude/.ck.json without ~/.claude/.vd.json errors on every session, so this cp is mandatory when a legacy global config exists. After .vd.json exists, rm ~/.claude/.ck.json (it's inert and backed up). Do not set paths.umbrella globally — that flips every repo at once and strands un-migrated plans/ content. Go per-repo; consider the global flip only after all active repos are migrated.
5. CK_ consumer audit* — the env rename is a hard cut: every CK_* session var has a VD_* successor (VD_REPORTS_PATH, VD_PLANS_PATH, VD_GIT_ROOT, VD_ACTIVE_PLAN, …) and anything still reading CK_* silently gets nothing. Temp files renamed ck-session-* → vd-session-*.

   grep -rn 'CK_' ~/.claude/ --include='*.cjs' --include='*.json' --include='*.sh' \
     | grep -v node_modules | grep -v '.bak'
   

   When fixing hits, work from an explicit inventory of your variable names — a blanket CK_→VD_ replace hits substring false-positives (SLACK_WEBHOOK_URL, BLOCK_LOW_AND_ABOVE, …) and third-party vars.

Repo layer

1. Opt in at the git root. Default to local-only repo config unless the team explicitly wants to share it:

   cd "$(git rev-parse --show-toplevel)"
   printf '{\n  "paths": {\n    "umbrella": ".workbench",\n    "layout": "feature-first"\n  }\n}\n' > .vd.json
   grep -qxF '.vd.json' .gitignore 2>/dev/null || printf '\n.vd.json\n' >> .gitignore
   

2. Move artifacts into one feature folder per ticket/topic. Prefer vd:workbench new <slug> --ticket <ticket> to create feature.json, then move legacy files into that feature's type folders:

   node ~/.claude/skills/workbench/scripts/workbench.cjs new my-feature --ticket PROJ-123
   feature=.workbench/features/proj-123-my-feature
   mkdir -p "$feature"/{plans,reports,journals,visuals,state}
   [ -d plans/reports ]  && find plans/reports  -mindepth 1 -maxdepth 1 -exec mv {} "$feature/reports/"  \;
   [ -d plans/journals ] && find plans/journals -mindepth 1 -maxdepth 1 -exec mv {} "$feature/journals/" \;
   [ -d plans/visuals ]  && find plans/visuals  -mindepth 1 -maxdepth 1 -exec mv {} "$feature/visuals/"  \;
   [ -d plans/goals ]    && find plans/goals    -mindepth 1 -maxdepth 1 -exec mv {} "$feature/state/"    \;
   find plans -mindepth 1 -maxdepth 1 -type d -exec mv {} "$feature/plans/" \; 2>/dev/null
   rmdir plans
   

   Missing sources are fine — absent subdirs appear lazily on first use. If the current session is writing into a plan dir, move that dir last (or from a copy). Follow-ups that bite: rewrite embedded plans/goals paths inside moved state files; git ls-files .workbench/ | xargs -r git rm -r --cached for anything previously tracked under plans/ (a mv into an ignored dir doesn't untrack it).
3. Gitignore: replace the plans entry with .workbench. Don't keep both — if a non-migrated tool recreates plans/, it should show up untracked as a signal. .vd.json stays local-only via .gitignore unless the team explicitly chooses a shared config.
4. Stale references: grep the repo (CLAUDE.md, AGENTS.md, docs/, .github/, scripts) for hardcoded plans/ paths and point them at the hook-injected paths instead.

Already on the legacy .work umbrella? A repo migrated before the .workbench rename just needs the dir + config flipped: git mv .work .workbench (or mv if untracked), set paths.umbrella to .workbench in .vd.json, and swap the .gitignore entry .work/ → .workbench/. The hook re-resolves to .workbench/ on the next prompt; legacy .work/ is still recognized read-only by the resume probes during the transition.

Verify (also the whole of --check)

Run from inside the target repo — loadConfig() resolves the project .vd.json from process.cwd(); the payload cwd only anchors path strings, so piping from elsewhere false-negatives into legacy output.

cd "$(git rev-parse --show-toplevel)"
git check-ignore -v .workbench/features/example/reports/x && echo "OK .workbench ignored"
git check-ignore -v docs/x; [ $? -eq 1 ] && echo "OK docs NOT ignored"
echo "{\"cwd\":\"$PWD\",\"session_id\":\"check\"}" | node ~/.claude/hooks/dev-rules-reminder.cjs

Pass = the injected ## Paths block shows six paths (Reports/Plans/Docs/Visuals/Journals/State) with everything except Docs under .workbench/. Three paths = umbrella did not activate (invalid umbrella value, missing .vd.json, or you ran from outside the repo). A malformed .vd.json doesn't crash — it silently falls back to the 3-path legacy injection, so assert on output, never trust the config edit. No paths at all = a legacy .ck.json lingers without a .vd.json: the loader now raises a migration error and the fail-open hook injects nothing — create the .vd.json (step 4 / repo step 1). New paths take effect on the next prompt; the current session's earlier injection still shows old paths — don't chase that as a bug.

Rollback

• Repo: delete <git-root>/.vd.json (or set umbrella null) and reverse the gitignore swap → legacy layout resumes byte-identically; mv the .workbench trees back if needed.
• Machine: vd hooks rollback (restore newest *.bak.* + settings.json.bak) or vd hooks uninstall (remove + unregister vd-managed files only; third-party hooks untouched). Both support --dry-run. Caution: the uninstall allowlist can lag the deployed asset set — check --dry-run output and remove orphans by hand.
• Full restore: copy the master backup back over ~/.claude/hooks + settings.json, then npm i -g claudekit-cli.

Gotchas

• Reports under feature-first live in the feature root, not inside a plan dir: features/<feature>/reports, alongside features/<feature>/plans. Anything string-concatenating plans/reports/... breaks; always use the hook-injected Reports: path.
• task-completed-handler.cjs / teammate-idle-handler.cjs live in ~/.claude/hooks but are deliberately not in settings.json (team-runtime invoked) — an "unregistered hook" finding there is not a bug; don't register or delete them.
• Subagent prompts must pass the work-context repo's .workbench/ paths, not the cwd's, when editing another project.
• Monorepos: the umbrella anchors to the git root, not cwd — one .vd.json and one .workbench/ per repo, even when working from a subdirectory.
• Dotfiles repos at $HOME: stray-home protection anchors child projects to themselves by default; set paths.allowHomeRoot: true only when $HOME is intentionally the artifact root.
• A corrupted ancestor .git/config (e.g. a stray repo at $HOME) silently disables git-root resolution for loose dirs beneath it — umbrella just stays off, no error.