diff --git a/migrations/0181_alert_dedup_claims.sql b/migrations/0181_alert_dedup_claims.sql
new file mode 100644
index 0000000000..78f888687f
--- /dev/null
+++ b/migrations/0181_alert_dedup_claims.sql
@@ -0,0 +1,16 @@
+-- #8901: src/review/alerts.ts's runAnomalyAlerts (the anomaly-alert port) writes hourly dedup CLAIMS with the
+-- shape (id, project, target_id, notification_key, status) and ON CONFLICT(project, target_id, notification_key)
+-- DO NOTHING. It was pointing those raw INSERTs at `notification_deliveries` (migration 0031), whose real schema
+-- has none of those columns and no such unique constraint — so the moment runAnomalyAlerts is wired to a cron
+-- path, its first INSERT throws. This gives the port its own correctly-shaped table with the real unique index
+-- it needs, independent of whether it is ever wired.
+CREATE TABLE IF NOT EXISTS alert_dedup_claims (
+  id                TEXT PRIMARY KEY,   -- opaque claim id (newId("hc")/newId("anm"))
+  project           TEXT NOT NULL,      -- agent slug
+  target_id         TEXT NOT NULL,      -- '__healthcheck__' or '__anomaly__'
+  notification_key  TEXT NOT NULL,      -- hashed per-hour / per-condition-set key
+  status            TEXT NOT NULL DEFAULT 'sent',
+  -- The claim mechanism relies on this exact unique constraint: the ON CONFLICT(...) DO NOTHING lets exactly one
+  -- writer per (project, target_id, notification_key) win the hourly claim; the losers see zero changed rows.
+  UNIQUE (project, target_id, notification_key)
+);
diff --git a/src/review/alerts.ts b/src/review/alerts.ts
index 255a53ea19..1cee22214a 100644
--- a/src/review/alerts.ts
+++ b/src/review/alerts.ts
@@ -230,7 +230,7 @@ export async function runAnomalyAlerts(env: Env, config: AlertAgentConfig, deps:
   // computes + maybe alerts; the other 59 short-circuit here before touching D1.
   const hourBucket = new Date().toISOString().slice(0, 13); // YYYY-MM-DDTHH
   const checkClaim = await storage(env).prepare(
-    `INSERT INTO notification_deliveries (id, project, target_id, notification_key, status)
+    `INSERT INTO alert_dedup_claims (id, project, target_id, notification_key, status)
      VALUES (?, ?, '__healthcheck__', ?, 'sent')
      ON CONFLICT(project, target_id, notification_key) DO NOTHING`,
   )
@@ -246,7 +246,7 @@ export async function runAnomalyAlerts(env: Env, config: AlertAgentConfig, deps:
   // Throttle: claim a per-(condition-set, hour) key so a repeated condition alerts at most hourly.
   const key = await sha256Hex(`anomaly:${anomalies.join("|")}:${hourBucket}`);
   const claim = await storage(env).prepare(
-    `INSERT INTO notification_deliveries (id, project, target_id, notification_key, status)
+    `INSERT INTO alert_dedup_claims (id, project, target_id, notification_key, status)
      VALUES (?, ?, '__anomaly__', ?, 'sent')
      ON CONFLICT(project, target_id, notification_key) DO NOTHING`,
   )
diff --git a/test/unit/alerts.test.ts b/test/unit/alerts.test.ts
index 5cb034c7f8..a1a9aa6369 100644
--- a/test/unit/alerts.test.ts
+++ b/test/unit/alerts.test.ts
@@ -7,6 +7,7 @@ import {
   detectAnomalies,
   runAnomalyAlerts,
 } from "../../src/review/alerts";
+import { createTestEnv } from "../helpers/d1";
 
 const healthy: AgentHealth = {
   byStatus: {},
@@ -229,6 +230,32 @@ describe("runAnomalyAlerts — send path", () => {
     expect(body.embeds[0]?.description).toMatch(/calibration drift|DEAD-LETTERED|permanently failed/);
   });
 
+  it("REGRESSION (#8901): the dedup-claim inserts succeed against the REAL migrated schema (alert_dedup_claims)", async () => {
+    // The other send-path tests use a mocked claim store (claimEnv), which is exactly why the
+    // notification_deliveries collision stayed latent. Against a REAL migrated D1, the old code's
+    // `INSERT INTO notification_deliveries (id, project, target_id, notification_key, status)` threw
+    // "no column named project"; alert_dedup_claims (migration 0181) has those columns + the unique index.
+    const fetchSpy = vi.fn(async () => new Response(null, { status: 204 }));
+    vi.stubGlobal("fetch", fetchSpy);
+    const env = createTestEnv();
+    const config = { slug: "anomaly-agent", features: { discordNotify: true }, secrets: {}, discordWebhookUrl: WEBHOOK } as AlertAgentConfig;
+    const deps: AnomalyAlertDeps = { computeAgentHealth: async () => anomalousHealth, computeCalibration: async () => driftCal };
+
+    await runAnomalyAlerts(env, config, deps);
+    expect(fetchSpy).toHaveBeenCalledTimes(1); // reached the send path => both claim inserts succeeded
+
+    // Both claims landed in the new table, and its unique constraint powers the throttle: a second run in the
+    // same hour loses the healthcheck claim (ON CONFLICT DO NOTHING => changes=0) and short-circuits, so no
+    // second Discord POST fires.
+    const rows = await env.DB.prepare("SELECT target_id FROM alert_dedup_claims WHERE project = ? ORDER BY target_id")
+      .bind("anomaly-agent")
+      .all<{ target_id: string }>();
+    expect((rows.results ?? []).map((r) => r.target_id)).toEqual(["__anomaly__", "__healthcheck__"]);
+
+    await runAnomalyAlerts(env, config, deps);
+    expect(fetchSpy).toHaveBeenCalledTimes(1); // still once: this hour's healthcheck claim was already taken
+  });
+
   it("resolves the webhook from a named secret (config.secrets.discordWebhook → env[name])", async () => {
     const fetchSpy = vi.fn(async () => new Response(null, { status: 204 }));
     vi.stubGlobal("fetch", fetchSpy);