diff --git a/src/settings/agent-actions.ts b/src/settings/agent-actions.ts
index 919068e204..13acf95566 100644
--- a/src/settings/agent-actions.ts
+++ b/src/settings/agent-actions.ts
@@ -145,13 +145,12 @@ export function planAgentMaintenanceActions(input: AgentActionPlanInput): Planne
   // the merge; it can't complete for this commit. A new commit makes the live head differ from mergeBlockedSha.
   const mergeTerminallyBlocked = input.pr.mergeBlockedSha != null && input.pr.headSha != null && input.pr.mergeBlockedSha === input.pr.headSha;
   const canMerge = reviewGood && !guardrailHit && acting("merge") && mergeableClean && approvalsSatisfied && !mergeTerminallyBlocked;
-  // A GOOD PR on a guarded path → held for the owner's manual safety review (NOT auto-approved or auto-merged).
-  // But a CONTRIBUTOR PR that is NOT review-good (gate blockers / red / unverified CI) OR conflicts is CLOSED
-  // one-shot REGARDLESS of the guardrail. Spec: "guarded + would-merge → hold; otherwise → closure." The
-  // guardrail exists to stop auto-MERGING/APPROVING crucial-path changes without owner review (see canMerge /
-  // approve) — it must NOT keep a rejected PR open: closing rejects bad changes and merges nothing, so it is
-  // always safe. Owner/automation PRs are still never closed (isContributor gates that). (#close-bad-guarded)
-  const willClose = isContributor && acting("close") && (!reviewGood || isConflict);
+  // A guarded/CRUCIAL path (CI, the review engine, visual) → ALWAYS held for the owner, never auto-actioned —
+  // not auto-approved, not auto-merged, AND not auto-closed. Operator decision (#hold-crucial-on-reject): a
+  // hallucinated reject on a crucial PR must NOT auto-close a good change (the #1528 near-miss); the owner
+  // verifies and closes/merges. The BULK (non-guarded) contributor PRs still auto-close one-shot on a bad
+  // verdict / conflict — only the small crucial set is held. Owner/automation PRs are never closed regardless.
+  const willClose = !guardrailHit && isContributor && acting("close") && (!reviewGood || isConflict);
   const ciReason = ciFailed
     ? `CI is failing${failingCheckNames.length ? ` (${failingCheckNames.join(", ")})` : ""}`
     : ciUnverified
diff --git a/test/unit/agent-actions.test.ts b/test/unit/agent-actions.test.ts
index b3ab4e8678..f4074bc317 100644
--- a/test/unit/agent-actions.test.ts
+++ b/test/unit/agent-actions.test.ts
@@ -130,11 +130,12 @@ describe("planAgentMaintenanceActions (#778)", () => {
       expect(plan).not.toContain("merge");
     });
 
-    it("DOES auto-close a failing contributor PR on a guarded path (the guard blocks auto-merge, NOT rejection)", () => {
-      // Spec: guarded + would-merge → hold; otherwise → closure. Closing a bad PR merges nothing, so the
-      // hard-guardrail (which exists to stop auto-MERGING crucial paths) must not keep a rejected PR open.
+    it("does NOT auto-close a failing contributor PR on a guarded path — holds it for the owner (#hold-crucial-on-reject)", () => {
+      // Operator decision: a crucial/guarded PR is NEVER auto-closed, even on a reject — a hallucinated reject
+      // must not auto-close a good crucial PR (the #1528 near-miss). The owner verifies + closes/merges. The
+      // BULK (non-guarded) still auto-closes on reject; only the small crucial set is held.
       const plan = classes(planAgentMaintenanceActions(input({ conclusion: "failure", autonomy: { close: "auto" }, blockerTitles: ["x"], ...guarded, pr: { labels: [], slopRisk: 95 } })));
-      expect(plan).toContain("close");
+      expect(plan).not.toContain("close");
     });
 
     it("does NOT approve or auto-merge a passing PR on a guarded path", () => {