# Plan — overnight run (consolidate → /upgrade-all → morning report)

**Created:** 2026-06-02 ~02:10 UTC. **Runs via:** `cc-ci-plan/overnight-run.sh` in tmux `cc-ci-overnight`.

## Goal
By morning, have a full picture of CI + all recipes: let things settle overnight, then run the weekly
**`/upgrade-all`** as a full test run and leave a report — so we can see whether it makes it through,
how many recipe PRs it produced, and the overall state.

## Sequence (the runner enforces it; polls in pure bash so it doesn't burn the shared claude budget)
1. **Gate A — Assistant done.** Wait for the PR-consolidation task to finish (marker
   `/srv/cc-ci/.cc-ci-logs/pr-consolidation.done`, which the assistant writes as its last step). So
   `/upgrade-all` starts from a clean one-open-PR-per-recipe state. (4h timeout → proceed anyway.)
2. **Gate B — usage reset.** Don't launch `/upgrade-all` until after the claude usage-limit reset
   (~03:30 UTC) so it has budget. (The loops + assistant are at ~90% now and may stall until reset;
   the watchdog resumes the loops on lift.)
3. **Gate C — Swarm free.** Wait (≤1h) for the Builder/Adversary to be idle so `/upgrade-all` doesn't
   contend with the regression canaries on the shared Swarm; proceed anyway after the cap.
4. **Run `/upgrade-all`** (`launch-upgrader.py fresh` → DEFAULT mode: surveys every enrolled recipe,
   opens/updates a recipe PR per upgradeable one verified by `!testme`, **NEVER merges**). Wait for it
   to finish (it self-terminates + writes `…/upgrades/upgrade-all-<date>.md`; ≤8h).
5. **Morning report** → `/srv/cc-ci/.cc-ci-logs/overnight-report-<date>.md`: the `/upgrade-all`
   summary, open-PR-per-recipe survey, and `launch.py status`. Then ping the orchestrator to send the
   operator a proactive PushNotification + journal it.

## Guardrails / notes
- `/upgrade-all` DEFAULT **never merges** anything and never weakens tests.
- Shared claude budget: a full `/upgrade-all` is heavy; if it re-exhausts the limit it may stall
  mid-run — the report will show how far it got (a partial result is still a result, per the ask).
- Neither the upgrader nor the assistant is watchdog-managed; if either limit-stalls, it stays stalled
  until nudged. The runner has bounded timeouts so it never blocks forever.
- This is a one-off overnight run (not the recurring Sunday-02:00 timer, which stays as-is).