Run one task across every wisdom repo (the *-wisdom siblings of wisdom-tweaker) so it doesn't have to be repeated by hand in each one. Use when the user types /l-each, or asks to apply a change, fix, dependency bump, or any dev task "to each / across all / for every wisdom repo". The argument is either a slash command run verbatim per repo (e.g. /l-each /dev-bump-zudo-deps) or a natural-language dev request that defaults to /x -m -a per repo (e.g. /l-each footer link to zudo-doc is 404, fix it).
Run one task across every wisdom repo (the *-wisdom siblings of wisdom-tweaker) so it doesn't have to be repeated by hand in each one. Use when the user types /l-each, or asks to apply a change, fix, dependency bump, or any dev task "to each / across all / for every wisdom repo". The argument is either a slash command run verbatim per repo (e.g. /l-each /dev-bump-zudo-deps) or a natural-language dev request that defaults to /x -m -a per repo (e.g. /l-each footer link to zudo-doc is 404, fix it).
user-invocable
true
disable-model-invocation
true
argument-hint
<slash-command | natural-language dev request>
/l-each — run a task across every wisdom repo
wisdom-tweaker is the control repo for a family of documentation sites that live as siblings:
../<name>-wisdom/. Each is an independent git repo. /l-each applies the SAME task to all of
them, after first making sure each one is safe to touch.
The task to run is: $ARGUMENTS
Routing: what "the task" means
Trim $ARGUMENTS, then:
Starts with / → it is a slash command. Run it verbatim in each repo (e.g.
). That command governs its own commit/push behavior.
/dev-bump-zudo-deps
Anything else → it is a natural-language dev request. Run /x -m -a <request> in each
repo. /x -m -a plans, implements, reviews, merges to main, and cleans up autonomously — the
full hands-off chain, once per repo.
If the user named specific repos in the invocation ("…for css and test only"), operate on just
those; otherwise operate on every discovered repo.
Phase 1 — Discover the repos
.claude/skills/l-each/scripts/discover-repos.sh
Prints one absolute repo path per line (every *-wisdom git repo next to wisdom-tweaker). New
repos are picked up automatically just by being cloned as a sibling. Work from this list.
Phase 2 — Preparation safety gate (before ANY task runs)
Each task — especially /x -m -a, which merges to main on its own — must never run on top of
uncommitted work. Assess every repo first, read-only:
It prints VERDICT: CLEAN (exit 0) or VERDICT: DIRTY (exit 3) plus details, and never mutates
the repo.
Meaningful vs. noise — DIRTY means real uncommitted work: modified or staged tracked files, or
untracked source files (e.g. a half-finished .mdx). Build/editor artifacts like .astro/ are
filtered out as noise, even in a repo that forgot to gitignore them. When unsure, the script errs
toward meaningful — better to ask than to bulldoze.
Then:
Any DIRTY repo → STOP. Do not run the task anywhere yet. Show the user a consolidated report
of every problem repo with its details (branch + what is uncommitted), because a DIRTY repo often
means a doc edit that was never committed — exactly the thing they would want to know. Ask whether
to (a) proceed on the CLEAN repos only, or (b) abort so they can resolve it first. Wait for the
answer.
CLEAN repos → if not already on main, switch them:
git -C <repo-path> checkout main
Checking out main is what makes /x branch from and merge back into main. Do not pull or fetch
unless asked. If prep-check reported UNPUSHED_COMMITS, mention it — it is not a blocker.
Phase 3 — Run the task in each ready repo
Dispatch one subagent per ready repo (via the Agent tool), all in a single batch so they run
concurrently. Each subagent prompt must:
Pin the work to that repo: every shell command runs with the repo as its working directory (start
with cd <repo-path> &&, or use git -C <repo-path>); all file paths stay under it.
Run the task — the verbatim slash command, or /x -m -a <request> — via the Skill tool.
Report back tersely: what changed, branch / PR / merge / CI outcome, or the failure reason.
/l-each itself does not commit, push, or merge — the task (/x -m -a or the passed command) owns
that. Caveat: a target repo's own project skills are not auto-loaded in the subagent (it is a
sibling, not the session root); /x and the finalize flow rely on global skills, so this is fine in
practice.
If the user asked to watch merges one at a time, run the subagents sequentially instead of batched.
Phase 4 — Report
Summarize per repo in one table: outcome (done / failed / skipped-dirty), what changed, and
PR / merge / CI status. Make skipped DIRTY repos and any failures impossible to miss.
