[hhff]

What

• scripts/devteam.mjs, wait: wait records when it last woke the supervisor (run/wait-since, registered in RUN_KNOWN so fleet status doesn't flag it). An open message to the supervisor, or an owner line, created after that moment is reported straight away, even though it's already in wait's first snapshot. Other items keep the old rule: only what's new since the first snapshot wakes it.
• test/kit.test.mjs:
  • A message sent before wait starts wakes it at once, and only once, when left open.
  • The fleet-park test joins the pane's wrapped lines before matching. Run inside an agent session there are nine markers to strip, and the shell wrapped the env -u line at 80 columns, mid-token.

Why

wait took its first snapshot as the baseline. A worker's message that landed between the end of the supervisor's tick and that snapshot was taken for an old one, so the supervisor slept the whole next tick, 15 minutes by default, while the worker waited. The snapshot runs metrics, which on a long log (Stacks: about 8.9k lines) takes a while, so the window was wide. The same race failed "a worker's message wakes it" three times on a loaded Mac.

Design note

"New" means since the last time wait woke the supervisor, not since its last log line. A message the supervisor deliberately leaves open would otherwise wake it on every wait, which is a busy loop. With the marker it wakes it once.

Authority

This only touches when wait returns. It writes nothing to the log, and run/wait-since is a timestamp that nothing authorizes from.

Migration

None.

Tests

Local, sequential, before merging main: Node 24.14.1 42/42, 20.19.6 42/42, 18.20.4 41/42; the failure was the pane wrap, fixed here and rerun green. Main (#93) is merged in; CI runs at the head.

🤖 Generated with Claude Code

[hhff]

Review (upstream loop, author's own PR):

• Authority: wait only changes when it returns. It writes no log line, and run/wait-since is a timestamp nothing authorizes from. No path to the owner.
• Busy-loop check: a message left open wakes the supervisor once. The new subtest asserts the second wait says "nothing new".
• run/wait-since is in RUN_KNOWN, so fleet status doesn't block a move on it.
• Test robustness: a shared serverUp waits up to 30s for each test dashboard and fails there if it never answers. The fleet-park pane check joins wrapped lines.
• Tests at the merged head: CI 18/20/22/24 green at aeee7f2. Local Node 24.14.1 and 20.19.6 43/43. Local Node 18.20.4: one failure in the dashboard review test's final review-answer POST ('fetch failed'), macOS + Node 18 only and not touched by this PR. Investigating separately.