You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
dev-lead retry sweep ignores dev-lead:hands-off: it posts retry markers and dispatch guards and dispatches passes on held PRs #2089
scripts/dev-lead-retry.sh decides whether to resume a PR with pr_resume_suppressed (scripts/lib/pr-automation-budget.sh). That check only knows about needs-human-review, through pr_has_escalation_label. It never consults the shared hold set in scripts/lib/hold-gate.sh (needs-human-review, dev-lead:needs-human, dev-lead:hands-off). #1595 introduced that set so callers would stop drifting apart on which labels count as a hold.
So for a PR held with dev-lead:hands-off or dev-lead:needs-human, the sweep still:
posts dev-lead-bot-comment-retry markers,
posts dev-lead-dispatch-guard markers,
dispatches a dev-lead pass.
The dispatched pass is then skipped by dev-lead-intent.sh (hands-off-label). Nothing is pushed, so this is not a safety breach. But every sweep:
adds marker comments to a PR a human parked;
spends a retry attempt on a run that can't act, so the attempt count may already be used up when the hold is lifted;
starts a workflow run that only skips.
Evidence
Seen on 2026-10-09 on two PRs held with dev-lead:hands-off:
In pr_resume_suppressed, suppress when any hold-gate label is present, not only the escalation label. The labels JSON is already passed in, so this needs no extra API call. Something like:
local held
held=$(jq -r '.[]'<<<"$labels_json"2>/dev/null | hold_gate_first_match "$(cat)")&& {
echo" [suppress] PR #${pr} in ${repo} carries ${held} — human hold; not resuming (#1595)">&2return 0
}
The library would need to source lib/hold-gate.sh.
Acceptance criteria
With dev-lead:hands-off or dev-lead:needs-human on a PR, the retry sweep posts no retry marker or dispatch guard and dispatches nothing. A bats test covers each of the three retry paths: rate-limit/partial, bot-comment and bot-thread.
Thanks so much for the detailed report and the clear evidence from #2085 and #1902 — it makes the problem easy to follow. To make sure the fix covers everything, could you confirm whether any other label-gated paths in the retry sweep (beyond the three listed: rate-limit/partial, bot-comment and bot-thread) also skip the hold-gate check, or whether pr_resume_suppressed is the only place it's missing?
Activity
github-actions commentedon Oct 9, 2026
Thanks so much for the detailed report and the clear evidence from #2085 and #1902 — it makes the problem easy to follow. To make sure the fix covers everything, could you confirm whether any other label-gated paths in the retry sweep (beyond the three listed: rate-limit/partial, bot-comment and bot-thread) also skip the hold-gate check, or whether
pr_resume_suppressedis the only place it's missing?fix(holds/pr-review): honour every hold label; pin the escalation hol…