From 7c3fddc4491822fc685bce9e53093a71e3b7bc58 Mon Sep 17 00:00:00 2001
From: clawsweeper <274271284+clawsweeper[bot]@users.noreply.github.com>
Date: Sat, 1 Aug 2026 06:19:14 -0700
Subject: [PATCH] fix(ci): release throttled live-item checks for delayed retry

---
 .github/workflows/sweep.yml | 16 ++++++++++++++++
 CHANGELOG.md                |  1 +
 test/sweep-workflow.test.ts |  6 ++++++
 3 files changed, 23 insertions(+)

diff --git a/.github/workflows/sweep.yml b/.github/workflows/sweep.yml
index 4b2ebf404f..fd5c0b16c2 100644
--- a/.github/workflows/sweep.yml
+++ b/.github/workflows/sweep.yml
@@ -758,6 +758,22 @@ jobs:
               echo "::notice::Completing $TARGET_REPO#$ITEM_NUMBER because the repository is accessible but the item is missing."
               exit 0
             fi
+            if grep -Eqi 'rate limit exceeded|secondary rate limit|HTTP 429' "$live_item_error"; then
+              cat "$live_item_error" >&2
+              rm -f "$live_item_error"
+              retry_at="$(node -e 'process.stdout.write(new Date(Date.now() + 20 * 60_000).toISOString())')"
+              echo "::notice::GitHub throttled the live-item check for $TARGET_REPO#$ITEM_NUMBER; releasing the claim for retry at $retry_at without spending failure budget."
+              {
+                echo "admission_retry=true"
+                echo "retry_at=$retry_at"
+                echo "proceed=false"
+                echo "terminal_noop=false"
+                echo "terminal_missing=false"
+                echo "guarded_open=false"
+                echo "guarded_open_action="
+              } >> "$GITHUB_OUTPUT"
+              exit 0
+            fi
             cat "$live_item_error" >&2
             rm -f "$live_item_error"
             exit 1
diff --git a/CHANGELOG.md b/CHANGELOG.md
index edd353ad1b..b60db653d8 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -18,6 +18,7 @@ checkpoint, and status-only commits are intentionally omitted.
 
 ### Changed
 
+- GitHub-throttled live-item checks now release the review claim for a delayed retry instead of failing the run and spending failure budget.
 - Event-review artifact publication completes as a superseded no-op when the reviewed branch vanished upstream (force-push or deletion) instead of failing the run.
 - Worker record requests now retry transient blank/invalid 2xx bodies from the edge within the bounded budget instead of failing hydration on the first occurrence.
 - Exact reviews of items that closed after enqueue now complete as superseded no-ops, and GitHub-throttled reservations defer as held retries — neither spends the item's review-failure budget.
diff --git a/test/sweep-workflow.test.ts b/test/sweep-workflow.test.ts
index 208b1a20e8..e86427229f 100644
--- a/test/sweep-workflow.test.ts
+++ b/test/sweep-workflow.test.ts
@@ -649,6 +649,12 @@ test("exact event review publishes directly with a queue-bounded canonical fallb
   );
   assert.match(liveItem.run ?? "", /Resolved invalid queued target branch/);
   assert.match(liveItem.run ?? "", /admission_retry=true/);
+  assert.match(
+    liveItem.run ?? "",
+    /rate limit exceeded\|secondary rate limit\|HTTP 429/,
+    "a throttled live-item check must release the claim for retry instead of failing",
+  );
+  assert.match(liveItem.run ?? "", /throttled the live-item check/);
   assert.match(liveItem.run ?? "", /decision\.targetBranch = process\.env\.TARGET_BRANCH/);
   assert.match(
     step(reviewer, "Review exact event item").if ?? "",