End

/end — wrap up, or refuse to

The skill is a gate, then a closeout, then (maybe) an exit. It exists so "I'm done" is a checked claim, not a feeling. Three rules:

1. Refuse to end with unsaved work. Uncommitted files you touched, unpushed commits, dirty worktrees, or a missing journal entry → you are not done.
2. Closeout is short and stands alone. The last message must make sense to someone who reads only it.
3. Exit only when it's cheap to come back. Light session → exit the harness. Substantial session → stay alive so the human can review before the context is gone. (--exit / --stay override.)

Arguments

$ARGUMENTS — optional flags:

Step 1 — Run the gate (always)

# SKILL_DIR: $CLAUDE_SKILL_DIR in Claude Code; otherwise the directory of this
# SKILL.md (e.g. gptme-contrib/skills/end or ~/.codex/skills/end).
python3 "$SKILL_DIR/scripts/end-check.py"

It reports, per check, ✅ ok / ℹ️ info / ⚠️ warn / ❌ block, and ends with one of:

"This session" = files/commits newer than the harness process start (from /proc), or commits carrying this session's id. Other sessions' dirt in a shared worktree is listed as info, not a blocker. Pass --since 2h if the process-start heuristic is wrong (e.g. a resumed session).

Shared worktree with parallel sessions (Bob's brain repo): siblings share your time window, so declare your footprint — the repo paths you touched plus the absolute paths of any linked worktrees you created. Only dirt/commits under those paths and only declared worktrees can block; the rest is info:

python3 "$SKILL_DIR/scripts/end-check.py" \
  --paths journal/2026-08-22/my-session.md scripts/foo.py /tmp/worktrees/my-feature

A parallel session's fresh worktree can still show as a blocker; if git -C <wt> log -1 proves it isn't yours, say so in the closeout and treat it as info — don't touch it.

Also do the judgment checks the script can't:

• Did the user ask something you never answered? → not done.
• Did you promise "I'll …" anywhere in the last few messages? → do it or retract it.
• Is a subagent or background job still running that produces a deliverable? → wait or kill it explicitly.
• Is a task/claim still marked active for this session? → release it (Bob: uv run coordination work-list --claimed, then work-complete; gptodo edit <id> --set state …).

Step 2 — If BLOCKED: say so, then fix

Open with the refusal, verbatim shape:

Not ending — N blocker(s): <summary list from the check>

Then fix what is fixable without the human:

git commit <explicit paths> -m "..."          # brain repo: git-safe-commit --scope-only <paths> -m "..."
git push                                      # brain repo: git-safe-push-master
# journal entry for the work (Bob: journal/YYYY-MM-DD/<session-file>.md, append-only)

Re-run the check. If a blocker needs the human (a decision, a credential, a conflict you shouldn't resolve alone), stay alive and ask — one question, with the default you'd pick. Never "wrap up anyway".

Step 3 — Closeout message (≤ 12 lines)

## Session closeout
**Shipped**: <commits/PRs with links — or "nothing shipped">
**Verified**: <how you know it works — tests run, symptom re-checked — or "not verified: …">
**Not done / deferred**: <explicitly, so nobody assumes it happened>
**Next**: <the one action that unblocks continuation>
**Closing**: exiting now | staying alive for review (say `/end --exit` to close)

Bob-specific: the closeout goes into the journal file too if it isn't there already (append-only), and the journal commit must be pushed before Step 4 — re-run the check once more after writing it.

Step 4 — Exit decision

To exit, the closeout message must already be sent; then, as the final tool call:

python3 "$SKILL_DIR/scripts/end-exit.py"        # SIGTERM to the harness after 3s
python3 "$SKILL_DIR/scripts/end-exit.py" --dry-run   # show the target first if unsure

It finds the interactive harness process (Claude Code via CLAUDE_PID; gptme and Codex via the /proc parent chain) and schedules the signal from a detached helper so your tool call returns first. Non-interactive runs (claude -p, gptme -n, codex exec) are left alone — they end with the turn. gptme caveat: SIGTERM skips SESSION_END hooks; when a human is present prefer telling them to type /exit.

Rationalizations this skill exists to block

Per-runtime install

Run once per machine (idempotent):

bash <this-dir>/install.sh          # symlinks into ~/.claude/skills/end and ~/.codex/skills/end
bash <this-dir>/install.sh --project   # also .claude/skills/end in the current repo

• Claude Code: ~/.claude/skills/end (or project .claude/skills/end) → /end.
• Codex: ~/.codex/skills/end → skills are slash-command packages; /end or $end.
• gptme: skills in configured lesson dirs are already indexed; /skill:end lands with gptme PR "invoke skills as slash commands". Until then, say "wrap up the session" — the when_to_use matcher loads this skill.

Related

• skills/autonomous-session-workflow (Bob) — Phase 5 "Persist" is what this gate enforces.
• lessons/workflow/merged-is-not-live (Bob) — why the closeout has a Verified line.
• scripts/forward-drive-probe.py (Bob) — the harness-side sibling: the same invariants checked on every completion event fleet-wide, dispatching a full session when a gap is found. Design: knowledge/technical-designs/forward-drive-completion-reactor.md.
• scripts/closed-loop-check.py (Bob) — post-hoc session-altitude check; the report-only ancestor of end-check.py.