Morning

Triage the night

The loop leaves its evidence in three places and remembers nothing else. Read all three before saying anything. Announce: "Using nightshift:morning to read last night's run."

1. Gather

cfg=loop/config; . $cfg 2>/dev/null
journal=~/.local/state/nightshift/$(basename "$PWD")/journal.md

• Journal: everything after the last start: line. Each STOP: line names why the night ended; task N: lines carry cost and round outcomes.
• Pull requests: gh pr list --label land --state open --json number,title,isDraft,labels,url and the same for --label land:blocked. A draft with the blocked label is a task that failed its repair round; its body has the generator's report, the skeptic's verdict and the verifier's tail. Merged PRs since the start line: gh pr list --label land --state merged --search "merged:>=<date>".
• Run directory: ~/.local/state/nightshift/<repo>/<date>-t<N>/ holds brief.md, gen-<round>.md, skeptic-<round>.md, check-<round>.log and the JSON envelopes. Read the log tail and the skeptic before judging.
• The switch: gh variable get LANDING_STATE. The loop flips nothing; if it is not run, a human froze it or never armed it.

2. Report, one block per stop

Headline first, then the consequence, then the fix. No raw log pastes; quote the one line that matters.

• Landed. "Task 3 landed as #41, 2 rounds, $5.10." Nothing to do; say it.
• Blocked (land:blocked draft). Say which of the three refused it: verifier red (quote the ERROR <step> line), skeptic REFUTED (quote its reason), or the generator's BLOCKED: line. The consequence is that every later task waits, because the loop stops at the first blocked task.
• Waiting (open non-draft PR). CI is still running, or a check never registered (wait mode with a wrong EXPECTED_CHECKS name); the merge script's output is in the journal.
• Closed without merging. A human decided; the loop will not retry until the PR is reopened or the task is removed from the plan.
• Produced nothing. The generator made no commits; the brief was probably not self-contained. The fix is a plan edit, not a retry.
• Frozen / deadline / MAX. The night ended on purpose; say which.
• No start: line today. launchd did not fire: launchctl list | grep nightshift, and the stdout/stderr paths in loop/launchd.plist.

3. Offer the fixes, do not apply them

Each blocked task gets one of:

1. Fix on the branch. Check it out, make the change, run scripts/check, push, gh pr ready <n> and remove the label: gh pr edit <n> --remove-label land:blocked. The next run resumes it from "open PR".
2. Close the PR. The loop treats a closed unmerged PR as a human decision and stops on that task until the plan changes — unless the closed PR carries the land:retry label (gh pr edit <n> --add-label land:retry), which says "run it again from scratch" and the next night starts the task over.
3. Edit the plan. Wrong or under-specified task: a plan change is a pull request (nightshift:plan's landing step).
4. Refreeze. gh variable set LANDING_STATE --body frozen when the stops suggest the plan needs a human pass before another night.

Ask which, one question, then do that one. Never gh pr merge a task PR by hand — that is what ./loop/merge-pr.sh <n> is for, and the hook denies the bare form for a reason: the same checks judge a PR whether a human or the loop asks.