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.

The script is in the plugin, which is where this skill itself came from. The harness states this skill's own base directory when it loads it — Base directory for this skill: <plugin>/skills/init — so the plugin root is two directories above that, and the script is <plugin>/scripts/init.sh. Write that absolute path:

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

Until 0.26.0 this step carried a 33-line hand copy of the shim's six-way search for the plugin, on the grounds that init runs before the repository has anything that could do the finding. Measured 2026-09-22: the harness hands the path over in-band at skill load, so the root is two directories up from a string that is already on screen — no cache glob, no sort -V, no Cursor SHA ordering to get right. The copy had been wrong once before it was removed (four branches of six, for as long as it had existed; found 2026-09-09).

If the base directory is genuinely not stated — a harness that does not print it — do not guess a cache path: say so and stop. $CLAUDE_PLUGIN_ROOT and $CURSOR_PLUGIN_ROOT are worth trying first, and a checkout the human can name is worth asking for; both beat a search whose failure mode is running the wrong copy of the plugin.

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.