Init

Init

A thin wrapper. All the work is in scripts/init.sh; this skill exists to ask the two questions the script cannot answer for itself, then run it.

1. Ask

Ask the human, in one short message, for:

• the memory directory — where this repository's durable memory lives. Suggest .agent-memory if they have no reason to want something else.
• the memory language — the language notes get written in. This is independent of the language the agent replies in (see agent-memory); suggest en if they have no preference.

Do not guess either value silently — a wrong memory directory is annoying to move later, and this only needs one short question.

2. Run the script

Before running it — this is the writing half, it lays files into --repo . — confirm that directory is the repository the human meant to set up. Several projects can be open in the same harness at once (Cursor especially), and the shell a skill runs in is not necessarily the one the human was talking about. One line naming it is enough.

Neither harness is trusted to hand a skill the plugin root, so do not rely on a variable alone. Claude Code sets CLAUDE_PLUGIN_ROOT and Cursor sets CURSOR_PLUGIN_ROOT (measured 2026-08-26), but each only in some contexts, and a variable a harness did not set is indistinguishable here from one it did. Locate the plugin all six ways .floppy/run locates it once installed — the two harness variables, then AI_FLOPPY_HOME for development, then the Claude Code cache, then Cursor's local symlink, then Cursor's cache — and fail loudly, naming the install command, if none resolves. All six, not the first four: on 2026-09-09 this block carried the Claude cache and stopped, so a Cursor user whose CURSOR_PLUGIN_ROOT the harness had not set was told "plugin not found" for a plugin .floppy/run would have resolved through cursor_local — which on the owner's own machine is the branch that answers, because the Cursor cache directory there exists and is empty. This block duplicates the has_scripts chain in shim/run rather than reading it from there, because .floppy/run does not exist yet in this repository — creating it is the first thing the script below does — and that search normally lives in the one file this plugin copies into a consumer, which has to stay self-contained. tests/test-init-bootstrap.sh extracts this block and runs it against each branch, so the two drift apart loudly rather than silently. Run:

has_scripts() { [[ -n "${1:-}" ]] && ls "$1"/scripts/*.sh >/dev/null 2>&1; }

# A cache directory that exists but holds no scripts/*.sh is not a candidate,
# and each cache is read the way its own names allow: Claude Code's last
# segment is a version, so `sort -V`; Cursor's is a commit SHA, where any
# lexicographic sort orders by hash value rather than recency, so `ls -dt`.
claude_cache="$(ls -d "$HOME"/.claude/plugins/cache/*/floppy/*/ 2>/dev/null | sort -V | tail -n1)"
cursor_local="$HOME/.cursor/plugins/local/floppy"
cursor_cache="$(ls -dt "$HOME"/.cursor/plugins/cache/*/floppy/*/ 2>/dev/null | head -n1)"

if   has_scripts "${CLAUDE_PLUGIN_ROOT:-}"; then
  floppy_root="$CLAUDE_PLUGIN_ROOT"
elif has_scripts "${CURSOR_PLUGIN_ROOT:-}"; then
  floppy_root="$CURSOR_PLUGIN_ROOT"
elif has_scripts "${AI_FLOPPY_HOME:-}"; then
  floppy_root="$AI_FLOPPY_HOME"
elif has_scripts "$claude_cache"; then
  floppy_root="$claude_cache"
elif has_scripts "$cursor_local"; then
  floppy_root="$cursor_local"
else
  floppy_root="$cursor_cache"
fi

if ! has_scripts "$floppy_root"; then
  echo "x floppy plugin not found (a cache directory with no scripts/*.sh counts as not found)." >&2
  echo "  Install it: /plugin marketplace add spscream/ai-floppy && /plugin install floppy" >&2
  echo "  If it is installed, reinstall: `plugin update` is a no-op while the version is unchanged." >&2
  exit 1
fi

bash "$floppy_root/scripts/init.sh" \
  --repo . --memory-dir <their answer> --language <their answer>

This can only run this way — through the script directly, not through .floppy/run — because .floppy/run does not exist yet in this repository; creating it is the first thing the script does.

3. Report

Read the script's own output back in short form: what was created, and what was already there and left untouched (the script says so per file). Running this again later, on the same repository, changes nothing — that's by design, not a limitation worth apologizing for.

Point out explicitly that quota.lock was not created — see agent-memory for why, and mention it will need a real measurement once there is a corpus worth measuring, not before.