Asking before filing, per the repo policy on cross-repo issues.
What I'd file at max-sixty/tend: poll_pr_checks.py treats a check that is waiting on a GitHub environment approval as ordinary pending, so it polls until its own time limit. Here that check is Hosted PR preview → deploy, gated by the hosted-preview environment, and it never moves until a maintainer approves it. Today this costs about 10 minutes of runner time per review. Once tend releases max-sixty/tend#1448 (merged 2026-09-29, after 0.3.4), the poll no longer stops at 9 minutes. It waits until the session deadline, so every tend-review on a PR that touches hosted/ would stay parked for most of its job timeout.
What happened here
In the last 24 hours, 9 tend-review sessions used up the full poll on deploy and then reported it as "still pending, unverified": #814, #815, #817, #818, #819, #823, #832, #840, and #842 (the #842 session took 21 minutes). The GitHub API reports the check as status: waiting and conclusion: null, for example deploy on #842 at 924adc2d. The reviews themselves were right. They approved on the finished checks and said plainly that deploy was unverified, so the only cost is wall-clock time. #1448 turns that cost into hours.
The gap upstream
On tend main at bb5ca166, reduce_rollup counts every CheckRun whose status != "COMPLETED" as pending. _settle then keeps polling while anything is pending, and its docstring already names this case as one that only the deadline ends: "a job waiting on an environment approval — would otherwise hold the wait until the session is killed". GraphQL gives this state its own CheckStatusState value, WAITING, so the poll can recognize it on the first read.
Proposed fix
In reduce_rollup, report a CheckRun with status == "WAITING" as not settled, and don't count it as pending. That means listing it under unverified (or a new gated bucket) with a note like "waiting on an environment approval". _settle then returns as soon as everything else has finished. approval() already handles unverified by approving and naming the check, which is what the sessions here did by hand. A test would build a rollup where one check is WAITING and the rest are green, and assert that the poll returns without waiting for the deadline.
I'll file it at max-sixty/tend once a maintainer gives permission here.
Session evidence
tend-review runs whose final summaries record deploy as still pending when the poll ran out: 36448096828 (#814), 36448133617 (#815), 36451630623 (#817), 36451659386 (#818), 36462597788 (#819), 36475943580 (#823), 36498908465 (#832), 36536769467 (#840), 36572097896 (#842). The waiting runs of Hosted PR preview that are still queued go back to 2026-09-24 (gh api 'repos/diffplug/dormouse/actions/runs?status=waiting').
Asking before filing, per the repo policy on cross-repo issues.
What I'd file at
max-sixty/tend:poll_pr_checks.pytreats a check that is waiting on a GitHub environment approval as ordinary pending, so it polls until its own time limit. Here that check isHosted PR preview→deploy, gated by thehosted-previewenvironment, and it never moves until a maintainer approves it. Today this costs about 10 minutes of runner time per review. Once tend releases max-sixty/tend#1448 (merged 2026-09-29, after 0.3.4), the poll no longer stops at 9 minutes. It waits until the session deadline, so everytend-reviewon a PR that toucheshosted/would stay parked for most of its job timeout.What happened here
In the last 24 hours, 9
tend-reviewsessions used up the full poll ondeployand then reported it as "still pending, unverified": #814, #815, #817, #818, #819, #823, #832, #840, and #842 (the #842 session took 21 minutes). The GitHub API reports the check asstatus: waitingandconclusion: null, for exampledeployon #842 at924adc2d. The reviews themselves were right. They approved on the finished checks and said plainly thatdeploywas unverified, so the only cost is wall-clock time. #1448 turns that cost into hours.The gap upstream
On tend
mainatbb5ca166,reduce_rollupcounts everyCheckRunwhosestatus != "COMPLETED"as pending._settlethen keeps polling while anything is pending, and its docstring already names this case as one that only the deadline ends: "a job waiting on an environment approval — would otherwise hold the wait until the session is killed". GraphQL gives this state its ownCheckStatusStatevalue,WAITING, so the poll can recognize it on the first read.Proposed fix
In
reduce_rollup, report aCheckRunwithstatus == "WAITING"as not settled, and don't count it as pending. That means listing it underunverified(or a newgatedbucket) with a note like "waiting on an environment approval"._settlethen returns as soon as everything else has finished.approval()already handlesunverifiedby approving and naming the check, which is what the sessions here did by hand. A test would build a rollup where one check isWAITINGand the rest are green, and assert that the poll returns without waiting for the deadline.I'll file it at
max-sixty/tendonce a maintainer gives permission here.Session evidence
tend-reviewruns whose final summaries recorddeployas still pending when the poll ran out: 36448096828 (#814), 36448133617 (#815), 36451630623 (#817), 36451659386 (#818), 36462597788 (#819), 36475943580 (#823), 36498908465 (#832), 36536769467 (#840), 36572097896 (#842). Thewaitingruns ofHosted PR previewthat are still queued go back to 2026-09-24 (gh api 'repos/diffplug/dormouse/actions/runs?status=waiting').