diff --git a/server/services/github.js b/server/services/github.js
index bc375a268f..edd9bc597c 100644
--- a/server/services/github.js
+++ b/server/services/github.js
@@ -96,13 +96,13 @@ function ghBackoffActive(key, now = Date.now()) {
   return Boolean(entry && now < entry.retryAfter);
 }
 
-function recordGhCallOutcome(key, ok, now = Date.now()) {
+function recordGhCallOutcome(key, ok, maxMs = GH_BACKOFF_MAX_MS, now = Date.now()) {
   if (ok) {
     ghCallBackoff.delete(key);
     return;
   }
   const failures = (ghCallBackoff.get(key)?.failures || 0) + 1;
-  const delay = Math.min(GH_BACKOFF_BASE_MS * 2 ** (failures - 1), GH_BACKOFF_MAX_MS);
+  const delay = Math.min(GH_BACKOFF_BASE_MS * 2 ** (failures - 1), maxMs);
   ghCallBackoff.set(key, { failures, retryAfter: now + delay });
 }
 
@@ -120,9 +120,14 @@ export function __resetGhCallBackoff() {
  * cleared on normal exit so it never fires for a completed run. `input`, when
  * supplied, is written to stdin (used by structured `gh api --input -` calls).
  * `backoffKey`, when supplied, opts this call into the consecutive-failure
- * backoff above — see the comment there for why it's opt-in.
+ * backoff above — see the comment there for why it's opt-in. `backoffMaxMs`
+ * overrides the default cap: it MUST exceed the calling scheduler's own tick
+ * interval, or the cooldown always expires before the next tick and never
+ * actually suppresses a retry — a real attempt fires every tick regardless of
+ * `backoffKey` (caught in review on the updateChecker caller: its 30-min
+ * interval exceeds the 15-min default cap, so the backoff there was a no-op).
  */
-export function execGh(args, timeoutMs = DEFAULT_EXEC_GH_TIMEOUT_MS, { cwd = null, env = null, input = null, backoffKey = null } = {}) {
+export function execGh(args, timeoutMs = DEFAULT_EXEC_GH_TIMEOUT_MS, { cwd = null, env = null, input = null, backoffKey = null, backoffMaxMs = GH_BACKOFF_MAX_MS } = {}) {
   if (backoffKey !== null && ghBackoffActive(backoffKey)) {
     return Promise.reject(new Error(`gh command backing off for ${backoffKey} after repeated failures`));
   }
@@ -140,7 +145,7 @@ export function execGh(args, timeoutMs = DEFAULT_EXEC_GH_TIMEOUT_MS, { cwd = nul
     const settle = (ok) => {
       if (settled) return;
       settled = true;
-      if (backoffKey !== null) recordGhCallOutcome(backoffKey, ok);
+      if (backoffKey !== null) recordGhCallOutcome(backoffKey, ok, backoffMaxMs);
     };
     const timer = setTimeout(() => {
       timedOut = true;
diff --git a/server/services/github.test.js b/server/services/github.test.js
index 8911590652..e2199167f8 100644
--- a/server/services/github.test.js
+++ b/server/services/github.test.js
@@ -253,6 +253,40 @@ describe('execGh backoffKey', () => {
     await expect(p2).rejects.toThrow(/exited with code 1/); // real second attempt, not a backoff rejection
     expect(spawn).toHaveBeenCalledTimes(2);
   });
+
+  it('honors a caller-supplied backoffMaxMs instead of the default cap', async () => {
+    // Regression for a real defect: a caller whose own retry interval exceeds
+    // the default 15-min cap (e.g. updateChecker's 30-min scheduler) would see
+    // the cooldown always expire before its next attempt, making the backoff a
+    // silent no-op — a real gh call fires every tick regardless of consecutive
+    // failures. A longer backoffMaxMs must actually widen the window past what
+    // the default cap would allow.
+    //
+    // 6 consecutive failures push the uncapped exponential delay (30s * 2^5 =
+    // 960s = 16min) past the DEFAULT 15-min cap but under a 60-min custom cap —
+    // the exact boundary where the two caps disagree. Each iteration advances
+    // time past its own delay first, so the prior failure's cooldown never
+    // blocks the next attempt from actually spawning.
+    for (let i = 0; i < 6; i++) {
+      const uncappedDelay = 30_000 * 2 ** i;
+      if (i > 0) vi.advanceTimersByTime(uncappedDelay + 1000);
+      const failing = makeChild();
+      spawn.mockReturnValueOnce(failing);
+      const attempt = execGh(['api', 'releases/latest'], 5000, { backoffKey: 'k', backoffMaxMs: 60 * 60 * 1000 });
+      attempt.catch(() => {});
+      failing.emit('close', 1);
+      await expect(attempt).rejects.toThrow();
+    }
+    expect(spawn).toHaveBeenCalledTimes(6);
+
+    // Past the 15-min DEFAULT cap (which would have already expired and let a
+    // real 7th attempt through) but short of the 16-min delay the custom
+    // 60-min cap actually applies.
+    vi.advanceTimersByTime(15 * 60 * 1000 + 1000);
+    await expect(execGh(['api', 'releases/latest'], 5000, { backoffKey: 'k', backoffMaxMs: 60 * 60 * 1000 }))
+      .rejects.toThrow(/backing off/);
+    expect(spawn).toHaveBeenCalledTimes(6); // still no 7th spawn — the custom cap, not the default, governed this
+  });
 });
 
 // The merge-follow-up reaper turns "the forge says this PR is OPEN" into a
diff --git a/server/services/updateChecker.js b/server/services/updateChecker.js
index c4732e19df..f598b33030 100644
--- a/server/services/updateChecker.js
+++ b/server/services/updateChecker.js
@@ -213,11 +213,15 @@ export async function checkForUpdate({ manual = false } = {}) {
     // consecutive-failure backoff — the 30-min scheduler retries a failure
     // on its very next tick with no cooldown of its own, which piled up
     // alongside branch-reconcile's identical gap during a real `gh` blip
-    // (both logged the same incident).
+    // (both logged the same incident). `backoffMaxMs` must exceed
+    // CHECK_INTERVAL_MS (30 min) — execGh's default 15-min cap always expires
+    // before this scheduler's own next tick, so it would never actually skip
+    // a scheduled attempt; a comfortable margin above the interval is what
+    // lets repeated failures widen the gap between real attempts.
     const raw = await execGh(
       ['api', `repos/${UPSTREAM_OWNER}/${UPSTREAM_REPO}/releases/latest`],
       undefined,
-      manual ? {} : { backoffKey: 'update-check' }
+      manual ? {} : { backoffKey: 'update-check', backoffMaxMs: CHECK_INTERVAL_MS * 3 }
     );
     let data;
     try { data = JSON.parse(raw); } catch { throw new Error(`Failed to parse GitHub release response: ${raw.slice(0, 200)}`); }
diff --git a/server/services/updateChecker.test.js b/server/services/updateChecker.test.js
index 19284ead1d..31bcfc6d6f 100644
--- a/server/services/updateChecker.test.js
+++ b/server/services/updateChecker.test.js
@@ -304,11 +304,14 @@ describe('checkForUpdate', () => {
     await checkForUpdate();
     // The scheduler retries a failed check on its next 30-min tick with no
     // cooldown of its own — the actual backoff mechanics live in execGh
-    // itself (github.js, mocked wholesale here); see github.test.js.
+    // itself (github.js, mocked wholesale here); see github.test.js. The cap
+    // must exceed the 30-min scheduler interval (90 min = 3x here) or the
+    // cooldown always expires before the next tick and never actually
+    // suppresses a retry — a real gh call fires every tick regardless.
     expect(execGh).toHaveBeenCalledWith(
       ['api', 'repos/atomantic/PortOS/releases/latest'],
       undefined,
-      { backoffKey: 'update-check' }
+      { backoffKey: 'update-check', backoffMaxMs: 90 * 60 * 1000 }
     );
   });