Skip to content

Node-based terminal manager for AI coding agents — tmux-backed terminals and parallel agent sessions as draggable nodes on an infinite pan/zoom canvas. macOS, Linux, and a browser Server Edition.

ci(release): give verify-feed one shared retry budget, not a per-file count

review5529cd3e4d1f
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 9e29c8200..731893a96 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -351,6 +351,19 @@ jobs:
# window (mirror sync / CDN delay); ADVISORY files only warn. latest.yml is advisory until the
# mirror is confirmed to serve it — move it to STRICT then. It is checked only when release-win
# itself succeeded (that job is best-effort and may finish later or need a rerun).
+ #
+ # The budget is ONE deadline for the whole job, not a retry count per file, and that is the
+ # v0.4.4 lesson. Measured on run 39079366383: publish promoted at 00:25:02, this job started at
+ # 00:25:03 — one second later — and the mirror did not serve 0.4.4 until somewhere inside
+ # 00:29:43…00:30:14, i.e. ~5 minutes behind the promote, for every file AT ONCE. Under a per-file
+ # count the first STRICT file absorbs that entire wait alone, and it ran out ~30 s short: the job
+ # failed on latest-mac.yml and then passed latest-linux.yml and latest.yml seconds later, which
+ # reads as a mac-specific cache problem and is nothing of the kind. A shared deadline charges the
+ # one lag once — the first file waits, the rest are free — and it bounds a genuinely dead mirror
+ # to BUDGET_MINUTES instead of BUDGET_MINUTES × files. If STRICT spends the whole budget then
+ # ADVISORY gets exactly one attempt, deliberately: the mirror is demonstrably behind and a second
+ # full wait buys a warning nobody is waiting for. Do NOT "fix" a red run here by reordering the
+ # file list — the lag is global, so whichever file is checked first is the one that fails.
verify-feed:
runs-on: ubuntu-latest
needs: [publish, release-win]
@@ -362,21 +375,28 @@ jobs:
STRICT: latest-mac.yml latest-linux.yml
ADVISORY: latest.yml
WIN_RESULT: ${{ needs.release-win.result }}
+ # ~2.4× the 5 min lag measured on v0.4.4, and a cap on the JOB, not on each file.
+ BUDGET_MINUTES: 12
+ INTERVAL: 30
run: |
want="${GITHUB_REF_NAME#v}"
+ # One deadline for every file, taken once up front — see the comment above the job.
+ deadline=$(( $(date +%s) + ${BUDGET_MINUTES:-12} * 60 ))
fail=0
check() {
- f="$1"; level="$2"; n=1; got=""
- while [ "$n" -le 10 ]; do
+ f="$1"; level="$2"; got=""
+ while :; do
got=$(curl -fsSL --max-time 20 "$FEED/$f" 2>/dev/null \
| sed -n 's/^version:[[:space:]]*//p' | head -n 1 | tr -d "\"' \r")
if [ "$got" = "$want" ]; then
echo "$FEED/$f serves $got"
return 0
fi
- echo "attempt $n: $FEED/$f serves '${got:-<missing>}', want $want; retrying in 30s…"
- n=$((n + 1))
- sleep 30
+ # Every file gets at least one attempt, however little budget is left.
+ left=$(( deadline - $(date +%s) ))
+ if [ "$left" -le 0 ]; then break; fi
+ echo "$FEED/$f serves '${got:-<missing>}', want $want; retrying in ${INTERVAL}s (${left}s of budget left)…"
+ sleep "$INTERVAL"
done
echo "::$level::$FEED/$f serves version '${got:-<missing>}', expected $want"
if [ "$level" = error ]; then fail=1; fi