Workflow skills for Claude Code and Codex.
Claude Code:
/plugin marketplace add killallgit/killall-skills
/plugin install killall-skills@killallgit
Codex:
codex plugin marketplace add killallgit/killall-skills
codex plugin add killall-skills@killallgit
GitHub Copilot CLI:
copilot plugin marketplace add killallgit/killall-skills
copilot plugin install killall-skills@killallgit
Cursor, from the agent chat:
/plugin marketplace add killallgit/killall-skills
Then install killall-skills from the Marketplace tab. The Cursor CLI can add
the marketplace with cursor-agent plugin marketplace add, and the install
step stays in that tab.
killall-skills installs every skill. To install one, use its name instead,
for example git-janitor@killallgit. Pick either the bundle or individual
skills, not both.
Any host, skills only, through the skills CLI:
npx skills add killallgit/killall-skills
Claude Code runs a skill as /name; Codex runs it as $name.
task manifests regenerates the catalogs from each skill's SKILL.md
frontmatter and scripts/marketplace.config.json:
.claude-plugin/marketplace.jsonfor Claude Code, Codex, andnpx skills.cursor-plugin/marketplace.jsonfor Cursor.github/plugin/marketplace.jsonfor GitHub Copilot CLI After adding, renaming, or editing a skill description:
task manifests
CI fails when the committed file is out of date.
A skill can ship hooks in skills/<name>/hooks/hooks.json; the generator inlines
them into that skill's entry and into the killall-skills bundle.
The bundle is listed last so the skills CLI groups all skills under
Killall Skills instead of repeating each skill name as a group heading.
Before releasing, update your checkout to the latest main: the release runner
is loaded from your checkout, not the temporary worktree. Older runners do not
support the relocated scripts/marketplace.config.json.
Then run one of these from anywhere in the checkout. The release runs in
a temporary worktree off origin/main, so your checkout is never switched. It
runs the tests and generate_manifests.py --check first, and on a failed push
it removes the tag it created:
task tag:patch
task tag:minor
task tag:major
It bumps version in scripts/marketplace.config.json, regenerates the manifest,
commits, tags vX.Y.Z, and pushes the commit and tag together. The release is
built from origin/main, so uncommitted changes in your checkout are ignored
and never released.
Agent definitions and project setup live in killall-agents.
In Claude Code, /git-janitor cleans the current repository. Pass a path to
target another repository, or add audit-only to inspect without removing
anything. The plugin-qualified command is /killall-skills:git-janitor.
For a separate conversation context, mention @agent-killall-agents:git-janitor
in Claude Code with the repository path and any scope limits. This requires
the separate agents plugin. The agent uses the same skill and returns a cleanup receipt.
Its context is separate; it operates on the repository's actual worktrees.
/independent-review gets the current branch reviewed by something that did
not write it. It opens a PR if the branch has none, then waits, within a fixed
budget, for observable CI and bot-review activity without triggering reviews.
It reads actual current-head review verdicts and resolves their findings. When the repo has none, or none reviews (rate limited,
skipped, slow), it dispatches a local review agent with fresh context (never the
current session) and resolves that agent's findings instead. It reports what
was fixed and declined, and leaves the PR open. If a local reviewer fails, it
tries other viable independent routes. If none works, it reports BLOCKED
and stops the calling workflow before merge or release unless the user
explicitly waives review. Self-review and passing tests are not substitutes;
a waiver is reported as WAIVED, not reviewed. Other skills, such as ship
and lets-finish, run it too.
python3 skills/wait-for-action/scripts/wait_for_action.py [--pr N] does one thing:
wait for observable PR activity to settle. It discovers checks, statuses,
Actions workflows, and bot-review signals in a single read-only loop. It never triggers or requires reviews. Exit 0 means no observable work
is pending, not merge approval or a completed independent review. Exit 5 means
no current-head execution/review was observed; newly pushed work may not have
registered yet, even after other fast checks have completed.
Only --pr, --repo, --ttl, and --log-dir are supported. Polling is fixed
at five seconds, failures stop immediately, and API calls share the total budget.
There are no reviewer-specific modes, triggers, or workflow-only waits. Requires
a resolvable PR, authenticated gh 2.48+, and Python 3.10+.
/lets-finish (or "let's finish", "wrap up", "ship it") lands work that is ready. It runs on a small model, so fixes go to the builder agent. Words after the command skip or stop steps: /lets-finish don't review, /lets-finish stop after CI. /lets-finish and release also runs the project's release after the merge: its release or deploy skill, or the process its docs describe. It never releases unless you say so. A branch with no PR gets one,
waits for CI, and goes through independent-review first. For each PR it confirms CI
passed and that any required review is done (a CI reviewer or a human), then
squash-merges. It finds the beads, Jira tickets, and GitHub issues the PR
relates to and updates them: closed when the PR closes them, commented when it
only mentions them. It prints one [+], [-], or [x] line per step, stops
at the first failed step, and ends by asking [y/n] whether to remove the branch and
worktree. It calls the wait-for-action and independent-review skills. User-invoked only.
/end-session closes out a session. It fetches, fast-forwards local branches,
and rebases the session branch onto the default branch. Work that already
landed or was superseded is discarded with its worktree. Otherwise it asks
once whether a PR is required, opens one if so, and archives the session.
User-invoked only.
/session-eject saves the session's work before usage runs out. It tells
running subagents to wrap up and waits for their reports, then commits and
pushes every changed worktree. An open PR gets a comment on what landed and
what is left; a branch without one gets a draft PR titled
[ejected session]: <short desc>. /session-eject now skips the wait and
commits, pushes, and opens the draft PR at once. Work on the default branch
moves to an eject/ branch first. It ends by printing a handoff paragraph to
paste into a fresh agent so it can resume. User-invoked only.
/architectural-review reviews a codebase's architecture. A script ranks
hotspots (churn × size) and change-coupled files from git history; a
fresh-context agent walks them for friction, judged against a fixed design
vocabulary (module, interface, depth, seam, adapter, locality, leverage) and
the house-rules lenses from
killall-craft. It renders an HTML report of deepening
candidates with before/after Mermaid diagrams and proposed glossary terms, and
prints the bd create commands for an epic plus one bead per candidate, which
it runs only for the candidates you pick. It keeps GLOSSARY.md,
ARCHITECTURE.md, and docs/adr/ current as decisions land. Words after the
command name a direction. User-invoked only.
The trim-comments plugin hook runs on every Stop in Claude Code. It finds the
comments and docstrings in the files the session edited, in any repo, plus every
uncommitted or untracked file in the repo the session started in, including work
that predates the session. A reviewer on a different model deletes the ones the code
already says. The reviewer may edit only the flagged files, and an edit that changes
code is restored; outside Python, a # line inside a multi-line string (a YAML block
scalar, a heredoc) reads as a comment and can be removed.
It acts only in repos that opt in with a .commentrc at the root, and it needs
uv and the claude CLI. /setup-trim-comments surveys a repo and writes that file;
/trim-comments runs the trimmer by hand. See skills/trim-comments/SKILL.md for
the config keys and the exact checks. Both are user-invoked only.