Repository navigation
Closed
Labels
help wantedDenotes an issue that needs help from a contributor. Must meet "help wanted" guidelines.Denotes an issue that needs help from a contributor. Must meet "help wanted" guidelines.
Description
Observed
On hive's own hosted spoke (image rev 1997bc9, ACMM 6, evalMode=webhook, governor mode quiet), the continuous re-kick wake from #12561 never ran during a 3h window (00:30–03:30Z, 2026-10-28):
grep -c "continuous re-kick wake"over--since=3hlogs: 0 (the string is present in the binary).- 17 eval cycles in 3h, 13 of them from review/verdict/github webhook wakes, the rest from the 900s cadence ticker.
- Continuous agents (scanner, reviewer, reviewer-2, adjudicators, guide) sat idle an average of 4.4 min (max 17.7 min) after their
NextKickwas due; kicks only landed when a webhook or cadence eval happened to run. Scanner idle 03:05–03:11Z with 21 actionable issues. /api/statusshowedcontinuous.nextKickAtin the past for several agents whilegovernorreported no wake.
Secondary: kicks that bounce off an agent that is mid-turn are recorded as delivery failures:
- Same-lane dispatch: a review-fix kick is delivered to the scanner, then the governor kick for the same agent in the same cycle waits the full 120s prompt timeout and fails with
agent scanner CLI did not reach input prompt; crux logskick undeliverable while agent turn in progress; not self-reporting(transcript live) and incrementskicksUndeliverable(7 on scanner). - Each such failure goes through
Governor.RecordKickFailure, doublingcontinuousBackoff: scanner went 1m → 2m → … → 30m (BackoffUntil 03:31:04Z recorded at 23:01:04Z) while it was simply busy on the fix kick it had just been handed. Fix kicks for 🐛 fix(automerge): bind unmarked owner approvals to the head's observed push time #14758 / 🐛 scanner: refuse claims on issues that already have an open linked PR #14808 bounced this way.
Analysis
cmd/hive/main.gorun loop:armContinuousWake()runs after every tick and re-arms the timer with a minimum delay of 1s (continuousWakeMinDelay) even when the re-kick is already due. Each tick takes 6–13 min live (eval → kick dispatch with 120s prompt waits per busy lane → sweeps, especially automerge 2–6 min). A re-kick that comes due during the tick is "due now" at tick end, but at that instant another select channel is already ready (the cadence ticker's pending fire — e.g. the 23:04:40 fire was consumed at 23:17:55 — or a coalesced review/verdict/webhook wake), sowakeChnever wins the select. The loop goroutine was never idle in the window ("state persisted" followed within ~1s by the next tick every time), so the timer path was unreachable in practice.onFailedin the kick dispatcher unconditionally callsRecordKickFailure; the dispatcher also sends every message in an agent's lane even after the first one has been delivered and started a turn.
Expected
- A due continuous re-kick is served at tick end regardless of which channel the loop wakes on, as a lean cycle (no sweeps) so the kick lands in eval time rather than eval+sweeps time.
- A kick that bounces because the agent is mid-turn is a benign deferral: no backoff escalation, no further sends into that agent's lane in the same cycle.
Metadata
Metadata
Assignees
Labels
help wantedDenotes an issue that needs help from a contributor. Must meet "help wanted" guidelines.Denotes an issue that needs help from a contributor. Must meet "help wanted" guidelines.
Activity
hivecommons-hive commentedon Oct 28, 2026
🔒 Claimed by @thatguywob (contributor on hive hosted-available-oke-11-placeholder-r05x). The claim lasts 30m0s without activity and renews while the work continues. Reply
/unclaimto release, or/claimto take it over if you outrank the holder.hivecommons-hive commentedon Oct 28, 2026
🔓 Claim by @thatguywob released (lease revoked).
🐛 fix: serve continuous re-kicks between busy cycles (#14824)
github-actions commentedon Oct 28, 2026
@clubanderson this issue was closed by @hivecommons-hive[bot] as completed.
If this was closed by a merged PR, the fix only takes effect once it ships in a tagged release and your deployment updates to it — check your running version against the fix before concluding the fix failed.
If you are still seeing the problem, comment
/reopenand it will be reopened for you — you do not need write access. Please add what you are still observing (running version, logs, steps) so it is not closed the same way again.