| name | pm-forgotten-pr |
| description | One-shot triage of your open PRs with no activity above a configurable age threshold — classifies each as close or merge, renders a Forgotten PRs block, and dispatches confirmed merges as phase-c-merger subagents. Invoked inline by /pm Step 1D; can also be run standalone. Triggers on "pm-forgotten-pr", "forgotten PRs", "idle PRs", "stale PRs". |
| triggers | ["forgotten PRs","idle PRs","stale PRs"] |
| argument-hint | [days] (optional — default 3; PRs idle longer than this are surfaced) |
One-shot startup triage of open PRs that have gone idle. Does not enter a monitoring loop — continuous PR-fleet monitoring remains /pr-monitor-and-manage's job.
Resolve the required detector and merge-order helper before Step 1:
resolve_script() {
local name="$1" candidate
for candidate in \
"$HOME/.claude/skills-worktree/.claude/scripts/$name" \
"$HOME/.claude/scripts/$name" \
".claude/scripts/$name"; do
if [[ -x "$candidate" ]]; then echo "$candidate"; return 0; fi
done
return 1
}
FORGOTTEN_PR_TRIAGE_SH=$(resolve_script forgotten-pr-triage.sh || true)
MERGE_SEQUENCE_SH=$(resolve_script merge-sequence.sh || true)
[[ -n "$FORGOTTEN_PR_TRIAGE_SH" ]] || { echo "ERROR: forgotten-pr-triage.sh not found (checked all three paths) — forgotten PR triage unavailable" >&2; exit 1; }
When invoked by /pm Step 1D inline, $GH_USER and $FORGOTTEN_PR_DAYS are already set from the calling context. When run standalone, parse $ARGUMENTS as the days threshold (e.g. /pm-forgotten-pr 7 → 7-day threshold); non-numeric or absent values fall back to 3 safely.
Step 1: Detection
if [[ -z "${FORGOTTEN_PR_DAYS:-}" ]] && [[ -n "${ARGUMENTS:-}" ]]; then
_ARG=$(printf '%s' "$ARGUMENTS" | awk '{print $1}')
[[ "$_ARG" =~ ^[1-9][0-9]*$ ]] && FORGOTTEN_PR_DAYS="$_ARG"
fi
DAYS="${FORGOTTEN_PR_DAYS:-3}"
"$FORGOTTEN_PR_TRIAGE_SH" --json --days "$DAYS" ${GH_USER:+--author "$GH_USER"}
forgotten-pr-triage.sh (read-only — it never closes, merges, or deletes) enumerates your open PRs, keeps those whose last activity (updatedAt, the documented age basis) is strictly more than the threshold ago, and classifies each close or merge using exactly two close signals (first match wins; otherwise merge):
- linked issue closed — the PR's
Closes/Fixes #N issue is already CLOSED (rationale linked issue #N closed).
- superseded / already in main — the PR contributes zero net-new commits to
main (every commit already landed, including via another merged PR, by git cherry patch-id equivalence; rationale superseded / already in main).
CI/conflict state is not a close signal — a red or conflicted PR is surfaced as a merge candidate and will fail the gate visibly downstream. Each JSON record carries number, title, url, headRefName, age_days, recommendation, and rationale, so the block below renders and the action paths act without extra gh calls. Exit 0 (including the empty set) is success; only exit 2/3 (usage / environment) is a real error to surface.
Step 2: Render
## Forgotten PRs (>N days)
- **PR #123 — {title}** — {age_days}d idle → **close** (superseded / already in main)
- **PR #130 — {title}** — {age_days}d idle → **close** (linked issue #99 closed)
- **PR #141 — {title}** — {age_days}d idle → **merge**
N is the active --days threshold (default 3). If the forgotten set is empty, print a single line — ## Forgotten PRs — none older than N days — and skip the action paths entirely.
Step 3: Close flow (confirmation-gated)
For every close-classified PR, present the PR and its rationale and ask for explicit confirmation before touching anything (recommend → confirm → apply, mirroring /pm-clean). Declining leaves the PR open.
- On confirmation:
gh pr close <N> --comment "<rationale>".
- Separately — a second, independent gate, never bundled with the close confirmation — offer to delete the PR's head branch. When confirmed, delete
<headRefName> honoring stale-cleanup.sh's branch-deletion safety rules (never a protected name — main/master/develop; never a branch checked out in a worktree), skipping with a one-line note if any check fails. Do not reimplement those safety checks — treat stale-cleanup.sh as their source of truth; its out-of-band sweep (via /pm-update) also reaps the branch later once it ages past the stale threshold.
Step 4: Merge flow (confirmation-gated, delegated to /wrap)
For the merge-classified PRs, present them and require an explicit "yes" before acting on any of them. This is a scoped exception to the global auto-merge default in CLAUDE.md "PR MERGE AUTHORIZATION": these PRs are triage-discovered — enumerated by forgotten-pr-triage.sh, not created or monitored by PM's own A→B→C pipeline — and have not earned the standing auto-merge authorization that inline pipeline PRs carry. Issue #733's authorship caution reinforces this gate: acting on a PR whose CI and review history PM never tracked warrants an upfront consent step. This confirmation is this step's own triage design (matching the sibling close gate in Step 3), not a paraphrase of the merge-authorization rule. Declining leaves everything unmerged. The "yes" is the authorization; do not add a second per-PR merge prompt.
On confirmation, dispatch one subagent per approved PR that executes the /wrap #N workflow (the arbitrary-PR form) so the existing merge gate + AC verification + squash-merge run unchanged — do not reimplement merge logic. Spawn per .claude/agents/README.md and subagent-orchestration.md: subagent_type: "phase-c-merger", mode: "bypassPermissions", explicit model: "sonnet", the verbatim SAFETY block, and the handoff path. Respect the 3–4 concurrent-pipeline ceiling from subagent-orchestration.md — launch up to the cap and queue the rest, starting a queued merge as each running one finishes. This is a one-shot hand-off: once the merge subagents are dispatched, this step is done — it does not poll them (polling is monitoring, which belongs to /pr-monitor-and-manage).
Sequence the approved set by file overlap before dispatching (issue #756). Two approved PRs touching the same file merge in whichever order they finish, and the second inherits a conflict the first created. Run the planner once over the approved numbers and dispatch in its order:
MERGE_SEQUENCE_AVAILABLE=1
MERGE_SEQUENCE_RC=0
if [[ -z "$MERGE_SEQUENCE_SH" ]]; then
MERGE_SEQUENCE_AVAILABLE=0
SEQ=""
echo "DEGRADED: merge-sequence.sh not found (checked all three paths) — approved merge ordering unavailable; holding all merge candidates" >&2
else
SEQ=$("$MERGE_SEQUENCE_SH" --prs "$APPROVED_CSV" --skip-missing) || MERGE_SEQUENCE_RC=$?
fi
When MERGE_SEQUENCE_AVAILABLE=0, do not dispatch unsequenced merges; report each approved PR as held for the next /pm run or /pr-monitor-and-manage. Otherwise branch on MERGE_SEQUENCE_RC: 0/1 → dispatch merge and batch PRs and hold back anything the plan marks hold, saying so in one line naming the shared file; any other value → the resolved planner failed, so log one line and dispatch in the original order rather than treating empty output as a valid plan. A held PR is not dropped: it stays a merge candidate for the next /pm run or for /pr-monitor-and-manage, which owns the across-tick hold state. Model: .claude/reference/merge-sequencing.md.
Never run two PRs from the same group concurrently. The 3–4 parallel phase-c-merger dispatch above applies across groups, not within one. Members of a group share a file by construction, so merging two of them in parallel recreates exactly the conflict the plan just avoided — and batch members are always same-group. Dispatch at most one PR per SEQ group at a time (plan[N].group; null = no overlap, always parallel-safe), starting the next member of a group only after the previous one merges or blocks. PRs in different groups still fill the remaining slots in parallel.
