Paperclip stranded-work watch

One job: tell Lacy when fleet work has gone quiet and the platform did not say so.

scripts/paperclip-stranded-watch.ts, run by launchd every 15 minutes, reads the Paperclip Postgres read-only and Slack-DMs one line per newly parked issue.

Why it exists (LAC-4568)

Paperclip parks an issue in blocked when a run dies before the agent gets control. That park is deliberate: the platform files a stranded_assigned_issue recovery action with ownerType: "board" and a next action that reads "Board operator: inspect the evidence, then retry the original owner." It sends no message. blocked wakes no assignee. So the platform hands Lacy a job and never tells him.

On 2026-10-16 the laptop slept at 1% battery. The 08:00Z Doug store watch slot (LAC-4564) died mid-setup with setup_failed: Legacy controller lease lost, was parked at 09:29Z, and was never retried. Doug's stores went unwatched for 9.5h. When this check was written the board held 128 such issues: 101 escalated to a board owner who was never told, 27 with no wake path at all, the oldest 132 days.

What it reports

An issue counts as parked when it is blocked and has no unresolved first-class blocker. Three outcomes:

Class	Meaning	Alerted
escalated	an active recovery action owns it, nobody was told	yes
orphaned	no recovery action, no declared wait, nothing will wake it	yes
declared wait	`unblock_descriptor` names the owner and the thing	no

A declared wait is a person parking work on purpose. Alerting on it would bury the hundred cases where nobody chose anything.

A park younger than 30 minutes is ignored: platform revalidation clears a transient one in seconds, and an agent mid-PATCH holds the state for a moment.

Reading the alert

First run counts the backlog instead of listing it, and says the listing starts now. Every run after that names up to 5 new parks, one line each:

LAC-4564 (Founding Engineer, 4.5h, stranded_assigned_issue): Doug store watch (every 4h, read-only)

One DM per park. The dedupe key is the issue plus the moment it parked, so a park that lasts a week produces one message, and the same issue parking again later produces another.

Run it by hand

export PAPERCLIP_DB_URL='postgres://paperclip:paperclip@127.0.0.1:54329/paperclip'
bun scripts/paperclip-stranded-watch.ts                  # print, do not DM
bun scripts/paperclip-stranded-watch.ts --notify          # print and DM
bun scripts/paperclip-stranded-watch.ts --state /tmp/s.json   # scratch state

Exit codes: 0 nothing parked, 1 something is, 2 could not look (missing PAPERCLIP_DB_URL, database unreachable). 2 never consumes an alert.

Needs STRANDED_WATCH_SLACK_TOKEN and STRANDED_WATCH_SLACK_CHANNEL to DM. Silence from a healthy board is the designed output, same contract as the store watch and its freshness check.

Schedule

~/Library/LaunchAgents/com.buildandserve.paperclip-stranded-watch.plist, StartInterval 900, log at ~/.paperclip/logs/paperclip-stranded-watch.log, state at ~/.paperclip/stranded-watch-state.json.

launchctl bootout gui/$UID/com.buildandserve.paperclip-stranded-watch 2>/dev/null
launchctl bootstrap gui/$UID ~/Library/LaunchAgents/com.buildandserve.paperclip-stranded-watch.plist
launchctl kickstart -p gui/$UID/com.buildandserve.paperclip-stranded-watch

Scheduled by launchd, not by a Paperclip routine, for the same reason as the freshness check: Paperclip is the thing that fails. It was down for two hours the morning this was written.

Delete it when

The platform retries a run that died before the agent got control, or tells the owner of a recovery action that it filed one. Then this file and its plist go, and the removal gets noted on LAC-4568. Tracked with LAC-4513 and LAC-4514; all three wait on the LAC-4482 decision about shipping platform patches.

Known gaps

Read-only on purpose. It never unparks anything. Unparking 100 issues at once would stampede the fleet, and the platform re-parks configuration_incomplete seconds after a write.
It watches the board, not the machine. The laptop sleeping at 1% battery is what started this, and nothing here notices that. The fleet power watch is the machine half: it warns before the battery runs out, so this one has less to report afterwards.
The GHOST company (6002a0d1-…, split 2026-09-24) is skipped; its 25 blocked issues are a frozen copy.