kickoff

Start-of-session ritual — get oriented in the repo (branch, working tree, open PRs/issues) and compose a descriptive session name, emitting a copy-pasteable /rename command for the user. Use when a wo

SKILL.md

GitHub

Kickoff Session

Formerly /orient.

Arguments: $ARGUMENTS

Orient at the start of a working session and give it a descriptive name so it is easy to identify later in the session picker and the Claude mobile app.

1. Get oriented

Prefer the repo's own status plumbing when it exists; fall back to raw commands otherwise:

  • If both targets exist — task --list-all 2>/dev/null | grep -q 'status:git' and the same check for status:gh — run task status:git and task status:gh. Caution: task executes the checked-out Taskfile; on an untrusted branch (e.g. reviewing a stranger's PR), use the raw fallback instead.
  • Otherwise run git status -sb, git log --oneline -5, gh pr list --limit 10, and gh issue list --limit 10.

Separately, as a credential preflight, check task --list-all 2>/dev/null | grep -q 'status:creds'. Where present, run task status:creds. The same untrusted-branch caution applies — on an untrusted branch, skip this probe rather than run it. There is no raw fallback for it: if the target is absent, note "no creds probe available" in the report and move on.

Keep all of this bounded — if gh, task status:git, task status:gh, or task status:creds hangs or is unauthenticated, note it and move on rather than blocking the session start. A missing or failed creds probe is a report item, never a kickoff blocker — kickoff proceeds either way and the summary says what couldn't be checked.

Sweep for stale claims. The claim /claim makes has no owner once its session ends: /integrate stops before the merge, /wrap leaves an open PR alone, and a personal-account board has no automation — so when the maintainer merges later, the assignee, claim:* label, and card status all survive with nobody left to clear them. Session start is where that gets caught, because it is the one step that runs without depending on the session that made the claim:

gh issue list --repo <owner/repo> --assignee @me --state all --limit 200 \
  --json number,title,state,labels,url
# ...and by marker, because a claim can outlive its assignee. `gh issue list
# --label` is exact-match (no prefix), so enumerate every live claim namespace
# label the repo actually carries — family-level, model-pinned, and legacy —
# and query each. This must be a CHECKED read: a bare `for lbl in $(gh label
# list ...)` runs zero iterations and falsely reports a clean sweep on an
# expired token or rate limit, hiding the marker-only claim this exists to find
# (gh-verification.md). On an org repo the event-driven release cannot recover
# such a claim once its assignee is gone UNLESS the timeline still proves a
# covering assignment interval for the claim comment (its trust gate needs
# the owner, or that proof — see claim-lifecycle.md's trust-gate note); a
# claim with no such interval ever recorded is exactly the gap this sweep
# backstops — do not skip it on a failed enumeration, surface it:
if ! claim_labels="$(gh label list --repo <owner/repo> --limit 1000 --json name -q \
    '.[].name | select(startswith("claim:") or startswith("agent:"))')"; then
  echo "warning: could not list claim labels — the marker-only sweep is INCOMPLETE; retry or check auth" >&2
fi
while IFS= read -r lbl; do
  [ -n "$lbl" ] || continue
  gh issue list --repo <owner/repo> --label "$lbl" --state all --limit 200 \
    --json number,title,state,assignees,url
done <<<"$claim_labels"

Query both, and union the results. /wrap runs its cleanup as separate commands on purpose — a combined gh issue edit fails wholesale when the repo lacks the label — so a partial cleanup that removes the assignee and then fails on the label or Status is an expected outcome. An assignee-only query can never see exactly that leftover, which is the case this sweep is supposed to recover.

--state all, not open: the motivating case is a closing PR merged after the session ended, which auto-closes the issue — so --state open filters out exactly the stale claims this sweep exists to find. The explicit --limit matters too; the default returns 30.

Assignment alone is not a claim. Plenty of people assign themselves planned backlog work. Flag an issue only when a claim marker corroborates it — an claim:* label (or a legacy agent:* one), a card at In Progress, or a /claim claim comment — and then only if its work has finished or stalled.

A claim comment is history, not state. Comments are never deleted, so the claim comment survives its own release — and where the issue was already assigned to you, /wrap correctly leaves that assignment in place too. Both markers then persist forever, and treating the comment alone as current would make every future /kickoff re-report the same long-released claim. So the comment counts only when it is the latest trusted Claiming — comment after the latest trusted Claim released — comment. A newer trusted claim record supersedes an earlier refresh record without releasing the active claim. Prefer the live markers (claim:* label, card at In Progress); fall back to that one current comment only after checking what follows it.

Report what survives that test as loose ends and point at /wrap for the release commands. Do not clear anything here: this step orients, it does not mutate.

2. Compose the session name

Kebab-case, at most ~40 characters, most-specific-first. Pick the source in this priority order:

  1. The topic or issue number given in the arguments.
  2. The issue/PR implied by the current branch or conversation.
  3. The branch name.
  4. Ask the user.

Pattern: <topic> or <topic>-<issue#> — e.g. dev-workflow-skills-138. No done- prefix and no date (the picker already shows recency).

3. Emit the rename

You cannot rename the session yourself — there is no tool or command for the model to do it. Say so explicitly, and output the command for the user to paste, on its own line in a fenced block:

/rename dev-workflow-skills-138

4. Record the name

Restate the chosen name in prose — e.g. "Session name: dev-workflow-skills-138" — so /wrap can recover it from conversation context even after compaction.

5. Summarize

Finish with 3–5 orientation bullets: current branch, clean/dirty tree, notable open PRs or issues, credential readiness, and the suggested next step. If implementation work is coming, suggest /claim next — it sanity-checks the issue and claims it, and /implement expects that claim to already exist.

Credential readiness bullet. Summarize what task status:creds found: what is logged in, what is missing or unknown, and the remedy the status output itself named (e.g. gh auth login, codex login, claude auth login). Preserve the probe's own distinctions rather than flattening them — "credential stored (not validated)" is not the same as authenticated, n/a (not configured, or the CLI is absent) is not the same as missing, and unknown is not the same as logged out. Quote the remedy from the status output, not from memory. If the probe was skipped (untrusted branch) or unavailable (target absent) or it failed, say so in this bullet — that is the report item. This covers whether a login exists, not what it is permitted to do: no line here establishes a token's scopes. Whether a specific issue needs permissions the CI credential lacks is /claim's per-issue preflight vetting, not this probe's.