publish

name	publish
description	Ship a Batterie shard or CLI change end-to-end in one verb — bump plugin.json, commit, push, trigger the assemble CI, watch it green, then pull this machine current. Use when you've edited a batterie source repo (instructions.md, a skill, a hook, CLI code) and want it live. Triggers on 'publish this', 'ship this shard', 'release this', 'cut a version', '/batterie:publish'.
allowed-tools	["Bash","Read"]

Publish a Batterie shard

The dance — bump version → commit → push → assemble → pull — collapsed into one command. The engine is scripts/publish.py; this skill is the safe wrapper: it previews with --dry-run, you confirm, then it runs for real.

Publishing is a capability, not a host list. Any machine meeting three preconditions can publish: a clone of the source repo being published (plus this repo for the central bump), gh authenticated with repo + workflow scopes (gh auth status), and uv for the PEP 723 engine script. Check those rather than asking which machine you're on — host lists rot. If the clone is missing there's nothing to push from — use /batterie:update to pull instead, or clone first.

This repo (publish target)

!python3 - <<'PYEOF' import json, subprocess from pathlib import Path cwd = Path.cwd() pj = cwd / ".claude-plugin" / "plugin.json" suite_pj = Path.home() / "repos/spm1001/batterie-de-savoir/.claude-plugin/plugin.json" if not pj.exists(): print(f"cwd {cwd} has no .claude-plugin/plugin.json — cd into the source repo you want to publish.") else: d = json.load(open(pj)) suite_v = json.load(open(suite_pj))["version"] if suite_pj.exists() else "?" is_suite = cwd.resolve() == suite_pj.parent.parent.resolve() print(f"publishing: {d['name']} ({cwd.name})") print(f"bump target: SUITE version {suite_v} (batterie-de-savoir) — every plugin ships this one number") print(f"mode: {'single-repo (cwd IS the suite repo)' if is_suite else '2-repo push (content here + suite bump in batterie-de-savoir)'}") st = subprocess.run(["git", "status", "--short"], capture_output=True, text=True) pending = st.stdout.rstrip() print("\nPending in this content repo (tracked edits ship in the content commit; untracked files are REFUSED unless you pass --all — bds-fifuko):") print(pending or " (clean tree)") PYEOF

Your task

1. Pick the bump level and the changelog line

Bump level — default patch. Use minor for a new capability, major for a breaking change. Map the user's words ("ship the docs fix" → patch; "release the new skill" → minor) and state which you picked. Under single-version (bds-suwoho) the bump always applies to the suite version (batterie-de-savoir's plugin.json) — the one number every plugin ships — whichever repo you're publishing from.

Changelog line — the suite ships ONE canonical CHANGELOG (bds-mawitu), and publish.py prepends an entry to it every release, so a shipped plugin's changelog can never predate its version. Draft a one-line human narrative of what this release does (e.g. "Carrying mise: fetch handles Shared Drives") and pass it as --changelog "…". If you don't, the commit message is used verbatim — fine for a mechanical release, but a written line reads far better to the family Claudes who orient off these files. Confirm the line with the user alongside the bump level.

2. Dry-run first — always

Resolve the engine from the local source tree and preview:

uv run --script ~/repos/spm1001/batterie-de-savoir/scripts/publish.py --patch --changelog "one-line narrative" --dry-run

(run from the cwd of the repo being published — it operates on cwd). Swap --patch for the chosen level. The dry-run shows the changelog: [x.y.z] … line it will prepend.

3. Review the plan with the human

Show the dry-run output and call out the staging list explicitly. In the content repo publish.py stages tracked modifications (git add -u), so every tracked pending file there ships in the content commit — if a tracked edit looks unrelated to this release, stop and ask. Untracked files are a hard stop (bds-fifuko): the engine refuses rather than sweep them, so a stray scratch/WIP file can't leak into a pushed, marketplace-triggering commit. If an untracked file genuinely belongs in the release (e.g. a brand-new skill file), re-run with --all — deliberately, after the human has seen it. (The suite-version bump is a separate, targeted plugin.json+CHANGELOG commit in batterie-de-savoir, so it can't sweep that repo's WIP.) Confirm: suite bump (old → new), commit message, wait+pull on, and — when publishing a non-batterie repo — that it's a 2-repo push (content here, suite bump in batterie-de-savoir).

4. Run it for real

On confirmation, drop --dry-run (keep the same --changelog). If the release legitimately includes new/untracked files the dry-run flagged, also pass --all (the engine refuses untracked files otherwise):

uv run --script ~/repos/spm1001/batterie-de-savoir/scripts/publish.py --patch --changelog "one-line narrative"

It will: push the content change (cwd repo), bump the suite version + prepend the CHANGELOG entry in batterie-de-savoir (one combined commit if you're publishing batterie-de-savoir itself; a targeted plugin.json+CHANGELOG commit otherwise), push, trigger assemble.yml, and watch the run to green (~1–2 min). The assembler stamps every plugin to the new suite number and regenerates each plugin's shipped CHANGELOG stub (which just points back at the canonical suite changelog — so no shipped changelog can ever look stale). A red run is almost always a suite-level version-ratchet quarantine — content drifted but the suite version wasn't bumped; the message names the failing run. Then it pulls this machine current for the published plugin (marketplace update → plugin update → CLI reinstall from source with --no-cache if this repo ships a CLI).

Escape hatches if asked: --no-wait (trigger, don't watch — also skips the pull, since an unwatched publish can't confirm what shipped), --no-pull (push only), -m "msg" (commit message).

5. Report + restart

Report the new version and that it went green. If the pull ran, remind the user:

Restart Claude Code (/exit then claude) to activate hook/skill changes —
SessionStart hooks only fire on a full restart. (Any CLI reinstall is already live.)

Notes

• Why no local assemble step: assemble.yml CI owns assembling every repo into the marketplace (daily 07:00 UTC + the dispatch this verb triggers). Running assemble.sh locally would fight it. publish.py triggers and watches the CI; it never assembles.
• One plugin, not all: the suite bump re-versions every plugin in the marketplace, but this pulls only the plugin you just shipped. The others show "update available" until /batterie:update — the "bring everything current" sibling.
• Lazy CLI convergence (jomiwa): when you publish a non-batterie source repo, its own plugin.json is also stamped to the new suite version inside the content commit — so that CLI's --version reads "the suite release that last changed it" and ticks toward the suite number as it changes. Only the repo being published is stamped (never its siblings), so there's no per-release multi-repo dance — CLIs converge lazily, not on every bump.