diff --git a/convex/ai/resilience.ts b/convex/ai/resilience.ts
index dce6bebe..a6fdbf75 100644
--- a/convex/ai/resilience.ts
+++ b/convex/ai/resilience.ts
@@ -333,10 +333,16 @@ async function finalizePendingMessages(
   });
   for (const message of result.page) {
     if (message.status !== "pending") continue;
-    await ctx.runMutation(components.agent.messages.finalizeMessage, {
-      messageId: message._id,
-      result: { status: "failed", error: reason },
-    });
+    try {
+      await ctx.runMutation(components.agent.messages.finalizeMessage, {
+        messageId: message._id,
+        result: { status: "failed", error: reason },
+      });
+    } catch {
+      // The @convex-dev/agent library may have already finalized this message
+      // internally when the stream failed (before our retry logic ran). Ignore
+      // the error and continue so any remaining pending messages are processed.
+    }
   }
 }
 
diff --git a/convex/ai/resilienceFinalize.test.ts b/convex/ai/resilienceFinalize.test.ts
index d497c6c1..6f95b60e 100644
--- a/convex/ai/resilienceFinalize.test.ts
+++ b/convex/ai/resilienceFinalize.test.ts
@@ -13,6 +13,15 @@ describe("getFinalizeCodeForError", () => {
     expect(finalizeCode).not.toContain("high demand");
   });
 
+  it("maps 'model is currently overloaded' to provider_overload", () => {
+    const error = new Error("The model is currently overloaded. Please try again later.");
+
+    const finalizeCode = getFinalizeCodeForError(error);
+
+    expect(finalizeCode).toBe("provider_overload");
+    expect(finalizeCode).not.toContain("overloaded");
+  });
+
   it("uses the Error name for non-transient errors", () => {
     const error = new Error("database blew up unexpectedly");
 
@@ -27,4 +36,23 @@ describe("getFinalizeCodeForError", () => {
 
     expect(finalizeCode).toBe("unknown_error");
   });
+
+  it("returns a sanitized code (never the raw error message) for every transient kind", () => {
+    // Ensure no transient error leaks its raw message into the finalize reason.
+    const cases: [string, string][] = [
+      ["This model is currently experiencing high demand.", "provider_overload"],
+      ["The model is currently overloaded.", "provider_overload"],
+      ["Service busy — please try again later.", "provider_overload"],
+      ["Request timed out after 180s", "timeout"],
+      ["fetch failed", "network"],
+    ];
+
+    for (const [msg, expectedCode] of cases) {
+      const error = msg === "fetch failed" ? new TypeError(msg) : new Error(msg);
+      if (msg.includes("timed out")) error.name = "TimeoutError";
+      const code = getFinalizeCodeForError(error);
+      expect(code, `expected sanitized code for: ${msg}`).toBe(expectedCode);
+      expect(code, `raw message leaked for: ${msg}`).not.toContain(msg.slice(0, 10));
+    }
+  });
 });
diff --git a/src/lib/posthogBeforeSend.test.ts b/src/lib/posthogBeforeSend.test.ts
index 8d59a855..ad7048a7 100644
--- a/src/lib/posthogBeforeSend.test.ts
+++ b/src/lib/posthogBeforeSend.test.ts
@@ -100,6 +100,20 @@ describe("shouldDropPosthogEvent", () => {
     expect(dropped).toBe(true);
   });
 
+  it("drops Gemini overloaded errors (alternate 503 phrasing)", () => {
+    const event = makeEvent({
+      properties: {
+        $exception_values: [
+          { value: "The model is currently overloaded. Please try again later." },
+        ],
+      },
+    });
+
+    const dropped = shouldDropPosthogEvent(event);
+
+    expect(dropped).toBe(true);
+  });
+
   it("drops Firefox reader-mode injection errors", () => {
     const event = makeEvent({
       properties: {
diff --git a/src/lib/posthogBeforeSend.ts b/src/lib/posthogBeforeSend.ts
index f1abbddd..f319afbe 100644
--- a/src/lib/posthogBeforeSend.ts
+++ b/src/lib/posthogBeforeSend.ts
@@ -37,6 +37,8 @@ const SUPPRESSED_MESSAGE_SUBSTRINGS: readonly string[] = [
   "function call turn comes immediately after",
   "exceeded your current quota",
   "model is currently experiencing high demand",
+  // Gemini's alternative 503 overload phrasing (mirrors sentryBeforeSend.ts).
+  "model is currently overloaded",
   "RESOURCE_EXHAUSTED",
   // Gemini paid-tier billing exhaustion: surfaces as "Your prepayment credits
   // are depleted." for both BYOK and house-key users. classifyByokError maps
diff --git a/src/lib/sentryBeforeSend.test.ts b/src/lib/sentryBeforeSend.test.ts
index 73bb9270..c6da1f2d 100644
--- a/src/lib/sentryBeforeSend.test.ts
+++ b/src/lib/sentryBeforeSend.test.ts
@@ -84,6 +84,52 @@ describe("shouldDropSentryEvent", () => {
     expect(shouldDropSentryEvent(eventWithValue(payload), hintWithError(payload))).toBe(true);
   });
 
+  it("drops Gemini overloaded errors (alternate 503 phrasing)", () => {
+    const payload = "The model is currently overloaded. Please try again later.";
+    expect(shouldDropSentryEvent(eventWithValue(payload), hintWithError(payload))).toBe(true);
+  });
+
+  it("drops high-demand error when buried in AI SDK wrapper via Error.cause", () => {
+    // The Vercel AI SDK wraps provider errors: outer.message is generic while
+    // the provider-specific text lives on Error.cause.
+    const inner = new Error(
+      "This model is currently experiencing high demand. Spikes in demand are usually temporary. Please try again later.",
+    );
+    const outer = new Error("Error reading UI message stream");
+    (outer as Error & { cause?: unknown }).cause = inner;
+
+    const genericEvent = eventWithValue("Error reading UI message stream");
+    expect(shouldDropSentryEvent(genericEvent, { originalException: outer })).toBe(true);
+  });
+
+  it("drops high-demand error buried two levels deep in cause chain", () => {
+    const root = new Error("This model is currently experiencing high demand.");
+    const mid = new Error("Failed to process stream chunk");
+    (mid as Error & { cause?: unknown }).cause = root;
+    const outer = new Error("Unhandled stream error");
+    (outer as Error & { cause?: unknown }).cause = mid;
+
+    const genericEvent = eventWithValue("Unhandled stream error");
+    expect(shouldDropSentryEvent(genericEvent, { originalException: outer })).toBe(true);
+  });
+
+  it("drops error when originalException is a non-Error object with a message property", () => {
+    // Some SDK error types don't extend Error but carry a .message field.
+    const fakeError = { message: "This model is currently experiencing high demand." };
+    const genericEvent = eventWithValue("Unknown error");
+    expect(shouldDropSentryEvent(genericEvent, { originalException: fakeError as Error })).toBe(
+      true,
+    );
+  });
+
+  it("keeps real errors that happen to have an innocent cause chain", () => {
+    const cause = new Error("Some benign internal detail");
+    const outer = new Error("Database connection failed");
+    (outer as Error & { cause?: unknown }).cause = cause;
+    const event = eventWithValue("Database connection failed");
+    expect(shouldDropSentryEvent(event, { originalException: outer })).toBe(false);
+  });
+
   it("drops Gemini RESOURCE_EXHAUSTED errors", () => {
     const payload = "Error: 429 RESOURCE_EXHAUSTED";
     expect(shouldDropSentryEvent(eventWithValue(payload), hintWithError(payload))).toBe(true);
diff --git a/src/lib/sentryBeforeSend.ts b/src/lib/sentryBeforeSend.ts
index 60dfe399..54e8c0f4 100644
--- a/src/lib/sentryBeforeSend.ts
+++ b/src/lib/sentryBeforeSend.ts
@@ -36,6 +36,10 @@ const SUPPRESSED_MESSAGE_SUBSTRINGS: readonly string[] = [
   // which is what Sentry captures here.
   "function call turn comes immediately after",
   "model is currently experiencing high demand",
+  // Gemini's alternative 503 overload phrasing. Both variants map to the same
+  // provider_overload transient kind on the backend; the client stream consumer
+  // surfaces whichever string was in the streaming error chunk.
+  "model is currently overloaded",
   "RESOURCE_EXHAUSTED",
   // Gemini free-tier quota errors. The leading "You exceeded" prefix is
   // Gemini's exact phrasing (capitalized "You exceeded ..."), so it covers
@@ -50,17 +54,35 @@ const SUPPRESSED_MESSAGE_SUBSTRINGS: readonly string[] = [
   "credits are depleted",
 ];
 
-// Collect every candidate message from the hint and event so that quota errors
-// nested in event.exception.values are not missed when hint.originalException
-// carries a generic wrapper message.
+// Collect every candidate message from the hint and event so that provider
+// errors nested inside AI SDK wrapper errors are not missed.
+//
+// The Vercel AI SDK sometimes wraps provider errors: the outer Error has a
+// generic message (e.g. "Error reading stream") while the provider-specific
+// text (e.g. "This model is currently experiencing high demand") lives on
+// Error.cause. We walk the full cause chain so the suppression list matches
+// regardless of how many wrappers the SDK adds.
 function errorMessages(event: ErrorEvent, hint: EventHint): string[] {
   const messages: string[] = [];
 
   const hintError = hint.originalException;
-  if (hintError instanceof Error && typeof hintError.message === "string") {
-    messages.push(hintError.message);
+  if (hintError instanceof Error) {
+    if (typeof hintError.message === "string") {
+      messages.push(hintError.message);
+    }
+    // Walk the cause chain — provider errors are often buried here.
+    let cause: unknown = hintError.cause;
+    while (cause instanceof Error) {
+      if (typeof cause.message === "string") messages.push(cause.message);
+      cause = cause.cause;
+    }
   } else if (typeof hintError === "string") {
     messages.push(hintError);
+  } else if (hintError !== null && typeof hintError === "object") {
+    // Some SDK error classes don't extend Error but carry a message property
+    // (duck-typed Error interface). Extract it defensively.
+    const msg = (hintError as Record<string, unknown>).message;
+    if (typeof msg === "string") messages.push(msg);
   }
 
   const values = event.exception?.values;