Clean Stop

Clean stop

Purpose

Before a machine may disappear, whether that is end of day, a spot instance about to reclaim, a container about to be torn down, or just walking away, in-flight work is easy to strand: a commit that never pushed, a branch with no PR, a plan that lives only in a local file, a worktree nobody will find again. This skill sweeps everything touched this session to a durable, linked resting state and reports whether it is genuinely safe to shut down.

Where /session-flow:keep-going recovers a session AFTER an interruption, clean-stop is its go/stop mirror. It makes the interruption safe BEFOREHAND, so there is nothing to recover. It supersedes a local /session-flow:handoff when the machine itself may go away: a handoff file is a machine-local save-point, and a save-point that dies with the disk is no save-point. clean-stop leaves the resumable state where it survives the machine, in the PR and issue bodies on the remote.

Steps

1. Inspect first, never assume. Enumerate every repository and worktree this session touched and read its real state: uncommitted changes, stashes, unpushed commits, branches without a PR, which PRs/issues already carry the resumable context, and any ignored, machine-local state that is not reproducible, a generated handoff, local config, hand-made credentials, which git status hides but which dies with the disk just the same. Act only on the gaps. Running clean-stop when a branch is already pushed and its PR already describes the work must NOT re-push, open a duplicate PR, or re-file issues. The sweep converges on what is missing, not on redoing what is done.
2. Durability sweep. Make the gaps durable. Push unpushed commits; for coherently committable work, including a clean, coherent stash, commit it with a real message and push. What is NOT coherently committable, ambiguous work-in-progress, a half-edit, a stash that will not cleanly apply, is not force-committed and not silently dropped: surface it in the verdict as an explicit "not durable, do X before shutdown" item. A bad commit and a lost stash are both worse than a named dangling item. Ignored, machine-local state that must not go to the remote, secrets, credentials, deliberately-ignored files, is never pushed; when it is not reproducible, surface it as a "preserve off the machine before shutdown" item so it is not silently lost with the disk. The conversation itself is another machine-local, non-durable item: transcripts live only in this machine's ~/.claude tree and are retention-swept besides. When the built-in export command resolves in your session and the session is worth keeping, first verify the resolved memory root's self-ignore guard (a .gitignore containing *, created and announced when absent), then offer the one-line export /export <memory_dir>/exports/<YYYYMMDDTHHMMSSZ>-<topic>.txt. Offer only, never run: built-ins are user-invoked, and nothing verifies the export happened. That destination is itself machine-local, so when the machine may go away, also surface copying the export off the machine (or a durable destination of the user's choice) as a "preserve off the machine" item.
3. Linkage + breadcrumbs (redact before any remote write). Before a PR or issue body is created or updated, sweep everything outbound: remaining tasks, dependencies, resume context, any pasted terminal output or diffs, for secrets, API keys, tokens, credentials, connection strings, and PII, and replace each hit with a shape marker (<REDACTED: API key>), never the value. This scrub is a hard gate: the breadcrumbs land on a remote shared with others and pass no human review first, so nothing outbound is written before it runs. Then link the work: every pushed non-default branch gets a PR. A branch that is the repo's default (main / master / trunk, or whatever git symbolic-ref refs/remotes/origin/HEAD resolves to) carries no PR and needs none once pushed. Every follow-up gets filed as an issue LINKED to that PR; dependencies, remaining tasks, and the context needed to continue live in the PR and issue bodies, never only in a local file. The acceptance bar is a cold agent: someone who finds the PR or issue with no memory of this session should have everything needed to offer to continue. Every remote artifact is created unattended, so the issue-linkage decision is made up front, never left for an interactive prompt: use the closing keyword when the branch or context names an issue, otherwise an explicit no-linkage reason. For an orphan branch, a hotfix, refactor, or drift sweep with no issue in its name, drive a direct gh pr create supplying that no-linkage reason, the guaranteed non-interactive path. Route through an installed pull-request or work-item capability only when it can run unattended with that decision passed in; if it would stop to ask, use the direct git / gh path carrying the same decision instead.
4. Local hygiene. Prune only what is provably safe: branches whose work is fully merged, worktrees with no uncommitted or stashed state and no irreplaceable ignored files, background work that has genuinely finished. A plain git status hides ignored files, but git worktree remove deletes them, so before removing a worktree, inspect its ignored content explicitly (git status --ignored) and preserve or surface anything not reproducible: a local .env, credentials, generated handoff data, or files a bootstrap copied in. Anything still holding unmerged commits, dirty state, irreplaceable ignored files, or a running job is NOT pruned. It becomes a named item in the verdict.
5. Free-and-clear verdict. Close with one of two honest outcomes: "clean to shut down, everything durable and linked", or a named list of the dangling items that still need a hand, each with the one action that would clear it. Never report clean while step 2 or 4 left something surfaced.

Autonomy policy. Make durable freely, gate destructive cleanup

• Make durable without asking. Pushing a commit, opening a PR for a pushed branch, and filing a linked follow-up issue are loss-prevention with low blast radius and are the whole point of the invocation: do them once the step 3 redaction gate has scrubbed anything outbound. Idempotency comes from step 1: act on gaps, never re-fire what already landed.
• GATE destructive hygiene. Deleting a branch, removing a worktree, or any force operation is irreversible. Do it only when step 1 PROVED it safe (fully merged, no uncommitted or stashed state, and, for a worktree, no irreplaceable ignored files a plain status hides). When safety cannot be proven, do not prune; name the item in the verdict instead. Losing unmerged work, or the only copy of an ignored .env, to a tidy-up is the exact failure this skill exists to prevent.

Nothing-dirty case

If the sweep finds everything already durable and linked, nothing uncommitted, nothing unpushed, every non-default branch that carries session work has a PR (work committed straight to the default branch is durable once pushed and needs none), nothing strandable left only on local disk, including non-reproducible ignored state a plain status hides, say so and give the free-and-clear verdict directly. A clean session is a valid, common outcome; do not manufacture work to look thorough.

What this skill does NOT do

• Does not leave resumable state in a local file. Context goes into PR and issue bodies precisely because a local file dies with the machine, that is what separates clean-stop from /session-flow:handoff and why it supersedes it when the machine may go away.
• Does not force-commit ambiguous work. Work that is not coherently committable is surfaced as a dangling item, never wrapped in a junk commit to make the tree look clean.
• Does not prune unmerged or dirty state. Destructive cleanup is gated on provable safety; when in doubt it names the item rather than deleting it.
• Does not duplicate capability mechanics. It routes PR, issue, and worktree work to whatever capabilities are installed and falls back to direct git / gh; it does not reimplement them.
• Does not report clean while something dangles. The verdict is honest, either genuinely free-and-clear or a named list.

Gotchas

• "Pushed" is not "linked". A branch can be safely on the remote and still strand its context if no PR or issue carries the why and the next step. The durability sweep and the breadcrumb bar are two separate checks; passing the first does not pass the second.
• Idempotency is load-bearing. clean-stop is often run more than once as a session winds down; every run must inspect real state first so the second run is a no-op on what the first already made durable, not a source of duplicate PRs or issues.
• A stash is invisible to a casual glance and cannot be pushed. Inspect for stashes explicitly; each one is either converted to a durable commit or named in the verdict, never left to die with the machine.