diff --git a/src/selfhost/maintenance-admission.ts b/src/selfhost/maintenance-admission.ts
index 3d3bfefeda..f71d088843 100644
--- a/src/selfhost/maintenance-admission.ts
+++ b/src/selfhost/maintenance-admission.ts
@@ -73,18 +73,31 @@ export function isMaintenanceJobType(type: string): boolean {
 }
 
 export interface MaintenancePressureSignals {
+  /** Foreground-priority rows in pending/processing regardless of run_after -- includes work deliberately
+   *  scheduled for later (e.g. agent-regate-pr's staggered/rate-deferred per-PR backlog, index.ts:24-29's
+   *  "normal, expected, can legitimately stay nonzero for long periods"). Retained ONLY for the
+   *  gittensory_queue_live_pending observability gauge (server.ts) -- evaluateMaintenanceAdmission deliberately
+   *  does NOT gate on this (#selfhost-maintenance-admission-runnable-signal): a raw count would starve
+   *  maintenance on backlog that was never actually competing for a claim slot. Use liveRunnableNowCount for
+   *  any real pressure decision. */
   livePendingCount: number;
+  /** Age in ms of the oldest live row by created_at, regardless of run_after -- same "observability only,
+   *  not an admission signal" caveat as livePendingCount above; a deliberately future-scheduled job inflates
+   *  this without meaning anything is stuck. Use oldestLiveRunnableAgeMs for a real pressure decision. */
   oldestLivePendingAgeMs: number | null;
-  /** Foreground-priority pending jobs that are RUNNABLE right now (run_after<=now), i.e. not currently
-   *  deferred by any mechanism -- distinct from livePendingCount, which also includes deferred/processing
-   *  work. #selfhost-queue-liveness's own diagnostic: "queue large but intentionally deferred" (this count can
-   *  be 0 with livePendingCount > 0, transiently, and that is fine) vs. "queue stuck" (this count stays 0
-   *  while oldestLiveRunnableAgeMs -- once something IS runnable -- climbs, or while releaseStaleForegroundDeferrals
-   *  keeps finding stale work every sweep). */
+  /** Foreground-priority jobs that are genuinely active RIGHT NOW: either 'processing' (already claimed,
+   *  real in-flight resource use) or 'pending' AND due (run_after<=now, not currently deferred by any
+   *  mechanism) -- distinct from livePendingCount, which also includes work deliberately deferred to the
+   *  future. #selfhost-queue-liveness's own diagnostic: "queue large but intentionally deferred" (this count
+   *  can be 0 with livePendingCount > 0, transiently, and that is fine) vs. "queue stuck" (this count stays 0
+   *  while oldestLiveRunnableAgeMs -- once something IS active -- climbs, or while releaseStaleForegroundDeferrals
+   *  keeps finding stale work every sweep). This is the field evaluateMaintenanceAdmission's live_pending_high
+   *  check actually gates on. */
   liveRunnableNowCount: number;
-  /** Age in ms of the oldest RUNNABLE (run_after<=now) foreground pending job -- null when none is runnable
-   *  right now. Distinct from oldestLivePendingAgeMs, which is dominated by a job intentionally scheduled far
-   *  in the future and says nothing about how long already-due work has sat unclaimed. */
+  /** Age in ms of the oldest genuinely-active (processing, or pending AND due) foreground job -- null when
+   *  none qualifies right now. Distinct from oldestLivePendingAgeMs, which is dominated by a job intentionally
+   *  scheduled far in the future and says nothing about how long already-active work has sat unclaimed/running.
+   *  This is the field evaluateMaintenanceAdmission's live_job_age_high check actually gates on. */
   oldestLiveRunnableAgeMs: number | null;
   maintenancePendingCount: number;
   oldestMaintenancePendingAgeMs: number | null;
@@ -220,8 +233,14 @@ export function evaluateMaintenanceAdmission(
 ): MaintenanceAdmissionDecision {
   if (!config.enabled) return { admit: true, reason: "disabled" };
   if (nowMs - pendingSinceMs >= config.maxDeferAgeMs) return { admit: true, reason: "trickle_max_defer_age" };
-  if (signals.livePendingCount > config.maxLivePendingCount) return { admit: false, reason: "live_pending_high" };
-  if (signals.oldestLivePendingAgeMs !== null && signals.oldestLivePendingAgeMs > config.maxLiveJobAgeMs) {
+  // Gated on genuinely ACTIVE live work (processing, or pending AND due), not the raw pending/processing
+  // count (#selfhost-maintenance-admission-runnable-signal): agent-regate-pr's normal, expected, staggered/
+  // rate-deferred per-PR backlog (index.ts:24-29) sits in 'pending' for a long time by design without being
+  // due yet, so counting it here would starve maintenance on work that was never actually competing for a
+  // claim slot -- exactly the "queue large but intentionally deferred" case liveRunnableNowCount's own doc
+  // comment (above) distinguishes from a genuinely stuck queue.
+  if (signals.liveRunnableNowCount > config.maxLivePendingCount) return { admit: false, reason: "live_pending_high" };
+  if (signals.oldestLiveRunnableAgeMs !== null && signals.oldestLiveRunnableAgeMs > config.maxLiveJobAgeMs) {
     return { admit: false, reason: "live_job_age_high" };
   }
   if (signals.backlogConvergencePendingCount > config.maxBacklogConvergencePendingCount) {
diff --git a/src/selfhost/pg-queue.ts b/src/selfhost/pg-queue.ts
index a59585578f..932a422514 100644
--- a/src/selfhost/pg-queue.ts
+++ b/src/selfhost/pg-queue.ts
@@ -416,10 +416,14 @@ export function createPgQueue(
    *  distinguishes "queue large but intentionally deferred" from "queue stuck, nothing runnable" without
    *  manual SQL. Host load is an independent, optional signal. */
   async function maintenancePressureSignals(now: number): Promise<MaintenancePressureSignals> {
+    // runnable_cnt/oldest_runnable count a row as genuinely active RIGHT NOW when it's either already
+    // 'processing' (real, in-flight resource use) or 'pending' AND due (run_after<=now) -- NOT merely present
+    // in the outer pending/processing set, which also includes work deliberately deferred to the future (see
+    // maintenance-admission.ts's MaintenancePressureSignals doc comments).
     const liveRes = await pool.query(
       `SELECT COUNT(*) AS cnt, MIN(created_at) AS oldest,
-              COUNT(*) FILTER (WHERE status='pending' AND run_after<=$2) AS runnable_cnt,
-              MIN(created_at) FILTER (WHERE status='pending' AND run_after<=$2) AS oldest_runnable
+              COUNT(*) FILTER (WHERE status='processing' OR run_after<=$2) AS runnable_cnt,
+              MIN(created_at) FILTER (WHERE status='processing' OR run_after<=$2) AS oldest_runnable
          FROM ${TABLE} WHERE status IN ('pending','processing') AND priority>=$1`,
       [FOREGROUND_QUEUE_PRIORITY_FLOOR, now],
     );
diff --git a/src/selfhost/sqlite-queue.ts b/src/selfhost/sqlite-queue.ts
index 672af9665b..19eb5071fd 100644
--- a/src/selfhost/sqlite-queue.ts
+++ b/src/selfhost/sqlite-queue.ts
@@ -1379,10 +1379,14 @@ function backfillJobForegroundLanes(driver: SqliteDriver): number {
  *  distinguishes "queue large but intentionally deferred" from "queue stuck, nothing runnable" without manual
  *  SQL. Host load is an independent, optional signal (see host-pressure.ts). */
 function maintenancePressureSignals(driver: SqliteDriver, now: number): MaintenancePressureSignals {
+  // runnable_cnt/oldest_runnable count a row as genuinely active RIGHT NOW when it's either already
+  // 'processing' (real, in-flight resource use) or 'pending' AND due (run_after<=now) -- NOT merely present
+  // in the outer pending/processing set, which also includes work deliberately deferred to the future (see
+  // maintenance-admission.ts's MaintenancePressureSignals doc comments).
   const live = driver.query(
     `SELECT COUNT(*) as cnt, MIN(created_at) as oldest,
-            SUM(CASE WHEN status='pending' AND run_after<=? THEN 1 ELSE 0 END) as runnable_cnt,
-            MIN(CASE WHEN status='pending' AND run_after<=? THEN created_at ELSE NULL END) as oldest_runnable
+            SUM(CASE WHEN status='processing' OR run_after<=? THEN 1 ELSE 0 END) as runnable_cnt,
+            MIN(CASE WHEN status='processing' OR run_after<=? THEN created_at ELSE NULL END) as oldest_runnable
        FROM ${TABLE} WHERE status IN ('pending','processing') AND priority>=?`,
     [now, now, FOREGROUND_QUEUE_PRIORITY_FLOOR],
   ).rows[0] as { cnt: number; oldest: number | null; runnable_cnt: number | null; oldest_runnable: number | null };
diff --git a/test/unit/selfhost-maintenance-admission.test.ts b/test/unit/selfhost-maintenance-admission.test.ts
index 3308daf4e3..6c3eb8514c 100644
--- a/test/unit/selfhost-maintenance-admission.test.ts
+++ b/test/unit/selfhost-maintenance-admission.test.ts
@@ -71,7 +71,7 @@ describe("evaluateMaintenanceAdmission", () => {
 
   it("admits unconditionally when disabled, even under extreme pressure", () => {
     const decision = evaluateMaintenanceAdmission(
-      { ...CLEAR_SIGNALS, livePendingCount: 999 },
+      { ...CLEAR_SIGNALS, liveRunnableNowCount: 999 },
       { ...CONFIG, enabled: false },
       now - 1_000,
       now,
@@ -79,9 +79,9 @@ describe("evaluateMaintenanceAdmission", () => {
     expect(decision).toEqual({ admit: true, reason: "disabled" });
   });
 
-  it("defers when live pending count exceeds the threshold", () => {
+  it("defers when live runnable-now count exceeds the threshold", () => {
     const decision = evaluateMaintenanceAdmission(
-      { ...CLEAR_SIGNALS, livePendingCount: 6 },
+      { ...CLEAR_SIGNALS, liveRunnableNowCount: 6 },
       CONFIG,
       now - 1_000,
       now,
@@ -89,9 +89,9 @@ describe("evaluateMaintenanceAdmission", () => {
     expect(decision).toEqual({ admit: false, reason: "live_pending_high" });
   });
 
-  it("admits when live pending count is AT (not over) the threshold", () => {
+  it("admits when live runnable-now count is AT (not over) the threshold", () => {
     const decision = evaluateMaintenanceAdmission(
-      { ...CLEAR_SIGNALS, livePendingCount: 5 },
+      { ...CLEAR_SIGNALS, liveRunnableNowCount: 5 },
       CONFIG,
       now - 1_000,
       now,
@@ -99,9 +99,23 @@ describe("evaluateMaintenanceAdmission", () => {
     expect(decision.admit).toBe(true);
   });
 
-  it("defers when the oldest live job has waited past the max age", () => {
+  // Regression (#4669): a raw pending/processing count includes agent-regate-pr's normal, expected,
+  // staggered/rate-deferred per-PR backlog (index.ts:24-29), which can sit well above maxLivePendingCount for
+  // long periods without any of it actually being due -- before this fix, that alone starved the maintenance
+  // lane even with zero real live pressure.
+  it("does NOT defer on a high raw live-pending count alone when nothing is actually runnable", () => {
     const decision = evaluateMaintenanceAdmission(
-      { ...CLEAR_SIGNALS, oldestLivePendingAgeMs: 120_001 },
+      { ...CLEAR_SIGNALS, livePendingCount: 999, liveRunnableNowCount: 0 },
+      CONFIG,
+      now - 1_000,
+      now,
+    );
+    expect(decision).toEqual({ admit: true, reason: "pressure_clear" });
+  });
+
+  it("defers when the oldest RUNNABLE live job has waited past the max age", () => {
+    const decision = evaluateMaintenanceAdmission(
+      { ...CLEAR_SIGNALS, oldestLiveRunnableAgeMs: 120_001 },
       CONFIG,
       now - 1_000,
       now,
@@ -109,9 +123,9 @@ describe("evaluateMaintenanceAdmission", () => {
     expect(decision).toEqual({ admit: false, reason: "live_job_age_high" });
   });
 
-  it("admits when there is no live job at all (null oldest age)", () => {
+  it("admits when there is no runnable live job at all (null oldest runnable age)", () => {
     const decision = evaluateMaintenanceAdmission(
-      { ...CLEAR_SIGNALS, oldestLivePendingAgeMs: null },
+      { ...CLEAR_SIGNALS, oldestLiveRunnableAgeMs: null },
       CONFIG,
       now - 1_000,
       now,
@@ -119,6 +133,19 @@ describe("evaluateMaintenanceAdmission", () => {
     expect(decision.admit).toBe(true);
   });
 
+  // Regression (#4669): oldestLivePendingAgeMs is created_at-based, not run_after-based, so it is dominated by
+  // a job intentionally scheduled far in the future -- before this fix, a deliberately staggered/deferred job
+  // could trip this the same way a genuinely stuck job would.
+  it("does NOT defer on a high raw oldest-live-pending age alone when nothing runnable is stale", () => {
+    const decision = evaluateMaintenanceAdmission(
+      { ...CLEAR_SIGNALS, oldestLivePendingAgeMs: 999_999_999, oldestLiveRunnableAgeMs: null },
+      CONFIG,
+      now - 1_000,
+      now,
+    );
+    expect(decision).toEqual({ admit: true, reason: "pressure_clear" });
+  });
+
   it("defers when the maintenance lane itself is already backed up", () => {
     const decision = evaluateMaintenanceAdmission(
       { ...CLEAR_SIGNALS, maintenancePendingCount: 16 },
@@ -226,9 +253,9 @@ describe("evaluateMaintenanceAdmission", () => {
     expect(decision).toEqual({ admit: true, reason: "pressure_clear" });
   });
 
-  it("admits when the oldest live job's age is AT (not over) the threshold", () => {
+  it("admits when the oldest runnable live job's age is AT (not over) the threshold", () => {
     const decision = evaluateMaintenanceAdmission(
-      { ...CLEAR_SIGNALS, oldestLivePendingAgeMs: 120_000 },
+      { ...CLEAR_SIGNALS, oldestLiveRunnableAgeMs: 120_000 },
       CONFIG,
       now - 1_000,
       now,
@@ -238,7 +265,7 @@ describe("evaluateMaintenanceAdmission", () => {
 
   it("force-admits via trickle once pending since exceeds the max defer age, even under pressure", () => {
     const decision = evaluateMaintenanceAdmission(
-      { ...CLEAR_SIGNALS, livePendingCount: 999, hostLoadAvg1PerCore: 99 },
+      { ...CLEAR_SIGNALS, liveRunnableNowCount: 999, hostLoadAvg1PerCore: 99 },
       CONFIG,
       now - CONFIG.maxDeferAgeMs,
       now,
@@ -248,7 +275,7 @@ describe("evaluateMaintenanceAdmission", () => {
 
   it("does not trickle-admit a job that hasn't waited long enough yet", () => {
     const decision = evaluateMaintenanceAdmission(
-      { ...CLEAR_SIGNALS, livePendingCount: 999 },
+      { ...CLEAR_SIGNALS, liveRunnableNowCount: 999 },
       CONFIG,
       now - (CONFIG.maxDeferAgeMs - 1),
       now,
@@ -258,7 +285,7 @@ describe("evaluateMaintenanceAdmission", () => {
 
   it("checks live pressure before maintenance-lane pressure (priority order)", () => {
     const decision = evaluateMaintenanceAdmission(
-      { ...CLEAR_SIGNALS, livePendingCount: 6, maintenancePendingCount: 16 },
+      { ...CLEAR_SIGNALS, liveRunnableNowCount: 6, maintenancePendingCount: 16 },
       CONFIG,
       now - 1_000,
       now,
@@ -266,9 +293,9 @@ describe("evaluateMaintenanceAdmission", () => {
     expect(decision.reason).toBe("live_pending_high");
   });
 
-  it("checks the oldest-live-job age before maintenance-lane pressure (priority order)", () => {
+  it("checks the oldest-runnable-live-job age before maintenance-lane pressure (priority order)", () => {
     const decision = evaluateMaintenanceAdmission(
-      { ...CLEAR_SIGNALS, oldestLivePendingAgeMs: 120_001, maintenancePendingCount: 16 },
+      { ...CLEAR_SIGNALS, oldestLiveRunnableAgeMs: 120_001, maintenancePendingCount: 16 },
       CONFIG,
       now - 1_000,
       now,
@@ -305,7 +332,7 @@ describe("evaluateMaintenanceAdmission", () => {
     );
     expect(beforeMaintenance.reason).toBe("backlog_convergence_high");
     const afterLive = evaluateMaintenanceAdmission(
-      { ...CLEAR_SIGNALS, backlogConvergencePendingCount: 11, livePendingCount: 6 },
+      { ...CLEAR_SIGNALS, backlogConvergencePendingCount: 11, liveRunnableNowCount: 6 },
       CONFIG,
       now - 1_000,
       now,
diff --git a/test/unit/selfhost-pg-queue.test.ts b/test/unit/selfhost-pg-queue.test.ts
index 5aafbf0532..90d405064a 100644
--- a/test/unit/selfhost-pg-queue.test.ts
+++ b/test/unit/selfhost-pg-queue.test.ts
@@ -3217,7 +3217,7 @@ describe("createPgQueue (durable #977)", () => {
 
     it("defers a maintenance job when live queue pressure is high", async () => {
       const m = makePool();
-      m.setPressureSignals({ live: { cnt: 6, oldest: now } }); // default threshold is 5
+      m.setPressureSignals({ live: { cnt: 6, oldest: now, runnableCnt: 6 } }); // default threshold is 5
       m.enqueueResult({ rows: [], rowCount: 0 }); // empty foreground claim
       m.enqueueResult({ rows: [maintenanceRow], rowCount: 1 }); // background claim
       const started: string[] = [];
@@ -3236,7 +3236,7 @@ describe("createPgQueue (durable #977)", () => {
 
     it("logs a deferred maintenance admission at info level, not warn (#selfhost-backpressure-noise)", async () => {
       const m = makePool();
-      m.setPressureSignals({ live: { cnt: 6, oldest: now } }); // default threshold is 5
+      m.setPressureSignals({ live: { cnt: 6, oldest: now, runnableCnt: 6 } }); // default threshold is 5
       m.enqueueResult({ rows: [], rowCount: 0 }); // empty foreground claim
       m.enqueueResult({ rows: [maintenanceRow], rowCount: 1 }); // background claim
       const logged = vi.spyOn(console, "log").mockImplementation(() => undefined);
diff --git a/test/unit/selfhost-sqlite-queue.test.ts b/test/unit/selfhost-sqlite-queue.test.ts
index 9f2d27a2ba..4e3d52a49a 100644
--- a/test/unit/selfhost-sqlite-queue.test.ts
+++ b/test/unit/selfhost-sqlite-queue.test.ts
@@ -3379,7 +3379,10 @@ describe("createSqliteQueue (durable #980)", () => {
     });
 
     // Directly occupies rows the way currently-pending/processing live work would, without needing worker-slot
-    // choreography -- the admission check reads real table state, not the consumer callback.
+    // choreography -- the admission check reads real table state, not the consumer callback. Defaults to
+    // 'processing' (stably occupied for the whole test, unlike a genuinely due 'pending' row which this
+    // queue's own pump would claim and complete during drain()) -- liveRunnableNowCount counts 'processing'
+    // rows as genuine current pressure alongside due-and-unclaimed 'pending' ones (see maintenance-admission.ts).
     function seedLiveRows(driver: ReturnType<typeof makeDriver>, count: number, status: "pending" | "processing" = "processing"): void {
       const now = Date.now();
       for (let i = 0; i < count; i += 1) {