diff --git a/.changeset/dedupe-provider-tool-results.md b/.changeset/dedupe-provider-tool-results.md
new file mode 100644
index 000000000..f5f9e7916
--- /dev/null
+++ b/.changeset/dedupe-provider-tool-results.md
@@ -0,0 +1,5 @@
+---
+"eve": patch
+---
+
+Fixed a turn failure when a provider-executed tool (e.g. gateway `web_search` on Opus 5) reported its result both inline in the assistant message and as a separately synthesized tool-error for the same call. The harness now keeps a single `tool_result` per tool-use id, so the next model call no longer fails with "each tool_use must have a single result".
diff --git a/packages/eve/src/harness/provider-tool-history.test.ts b/packages/eve/src/harness/provider-tool-history.test.ts
index 94a517378..e980c0aa0 100644
--- a/packages/eve/src/harness/provider-tool-history.test.ts
+++ b/packages/eve/src/harness/provider-tool-history.test.ts
@@ -1,7 +1,10 @@
 import type { ModelMessage } from "ai";
 import { describe, expect, it } from "vitest";
 
-import { normalizeProviderToolHistory } from "#harness/provider-tool-history.js";
+import {
+  dedupeToolResultsByCallId,
+  normalizeProviderToolHistory,
+} from "#harness/provider-tool-history.js";
 
 describe("normalizeProviderToolHistory", () => {
   it("preserves text ordering around provider-executed tool results", () => {
@@ -142,3 +145,82 @@ describe("normalizeProviderToolHistory", () => {
     expect(normalized.outcomeEndsResponse).toBe(true);
   });
 });
+
+describe("dedupeToolResultsByCallId", () => {
+  it("drops a second result for a provider call that already has an inline result", () => {
+    // Captured from a live Opus 5 web_search whose arguments failed to parse:
+    // the provider returns its own error result inline, and the AI SDK adds a
+    // second synthesized tool-error message for the same id.
+    const messages: ModelMessage[] = [
+      {
+        role: "assistant",
+        content: [
+          {
+            type: "tool-call",
+            toolCallId: "toolu_dup",
+            toolName: "web_search",
+            input: {},
+            providerExecuted: true,
+          },
+          {
+            type: "tool-result",
+            toolCallId: "toolu_dup",
+            toolName: "web_search",
+            output: { type: "error-json", value: { error: "invalid_input" } },
+          },
+        ],
+      },
+      {
+        role: "tool",
+        content: [
+          {
+            type: "tool-result",
+            toolCallId: "toolu_dup",
+            toolName: "web_search",
+            output: { type: "error-text", value: "AI_InvalidToolInputError" },
+          },
+        ],
+      },
+    ];
+
+    const result = dedupeToolResultsByCallId(messages);
+
+    expect(result.droppedCallIds).toEqual(["toolu_dup"]);
+    // The inline provider result survives; the empty tool message is pruned.
+    expect(result.messages).toEqual([messages[0]]);
+  });
+
+  it("keeps distinct client tool results untouched", () => {
+    const messages: ModelMessage[] = [
+      {
+        role: "assistant",
+        content: [{ type: "tool-call", toolCallId: "a", toolName: "add", input: {} }],
+      },
+      {
+        role: "tool",
+        content: [
+          {
+            type: "tool-result",
+            toolCallId: "a",
+            toolName: "add",
+            output: { type: "json", value: 1 },
+          },
+          {
+            type: "tool-result",
+            toolCallId: "b",
+            toolName: "sub",
+            output: { type: "json", value: 2 },
+          },
+        ],
+      },
+    ];
+
+    const result = dedupeToolResultsByCallId(messages);
+
+    expect(result.droppedCallIds).toEqual([]);
+    expect(result.messages).toEqual(messages);
+    // Untouched messages keep their identity (no needless copies).
+    expect(result.messages[0]).toBe(messages[0]);
+    expect(result.messages[1]).toBe(messages[1]);
+  });
+});
diff --git a/packages/eve/src/harness/provider-tool-history.ts b/packages/eve/src/harness/provider-tool-history.ts
index b95d70039..bf2648ffe 100644
--- a/packages/eve/src/harness/provider-tool-history.ts
+++ b/packages/eve/src/harness/provider-tool-history.ts
@@ -83,6 +83,63 @@ export function normalizeProviderToolHistory(input: {
   };
 }
 
+/**
+ * Enforces the provider invariant that each `tool_use` id carries at most one
+ * `tool_result`.
+ *
+ * Provider SDKs can emit two results for one provider-executed call whose
+ * arguments failed to parse: the provider's own outcome inline in the
+ * assistant message, plus a separately synthesized tool-error message. Both
+ * reference the same id, and replaying them fails the next model call
+ * ("each tool_use must have a single result"). This keeps the first result
+ * for each id — the provider's authoritative inline outcome, which appears
+ * before any synthesized follow-up — and drops later duplicates, pruning any
+ * message left empty.
+ */
+export function dedupeToolResultsByCallId(messages: readonly ModelMessage[]): {
+  readonly messages: ModelMessage[];
+  readonly droppedCallIds: readonly string[];
+} {
+  const seen = new Set<string>();
+  const droppedCallIds: string[] = [];
+  const deduped: ModelMessage[] = [];
+
+  for (const message of messages) {
+    if (!Array.isArray(message.content)) {
+      deduped.push(message);
+      continue;
+    }
+
+    let dropped = false;
+    const content = (message.content as { type: string; toolCallId?: string }[]).filter((part) => {
+      if (
+        (part.type !== "tool-result" && part.type !== "tool-error") ||
+        part.toolCallId === undefined
+      ) {
+        return true;
+      }
+      if (seen.has(part.toolCallId)) {
+        droppedCallIds.push(part.toolCallId);
+        dropped = true;
+        return false;
+      }
+      seen.add(part.toolCallId);
+      return true;
+    });
+
+    if (!dropped) {
+      deduped.push(message);
+      continue;
+    }
+    if (content.length === 0) {
+      continue;
+    }
+    deduped.push({ ...message, content } as ModelMessage);
+  }
+
+  return { messages: deduped, droppedCallIds };
+}
+
 function findUnmarkedProviderToolCalls(input: {
   readonly messages: readonly ModelMessage[];
   readonly providerExecutedOutcomeIds: ReadonlySet<string>;
diff --git a/packages/eve/src/harness/tool-loop.ts b/packages/eve/src/harness/tool-loop.ts
index a9f205249..946227829 100644
--- a/packages/eve/src/harness/tool-loop.ts
+++ b/packages/eve/src/harness/tool-loop.ts
@@ -117,7 +117,10 @@ import {
 } from "#harness/stale-input-responses.js";
 import { getInstrumentationConfig } from "#harness/instrumentation-config.js";
 import { normalizeUserContent, resolveAssistantStepText } from "#harness/messages.js";
-import { normalizeProviderToolHistory } from "#harness/provider-tool-history.js";
+import {
+  dedupeToolResultsByCallId,
+  normalizeProviderToolHistory,
+} from "#harness/provider-tool-history.js";
 import {
   type AuthorizationSignal,
   isAuthorizationSignal,
@@ -1834,7 +1837,13 @@ async function handleStepResult(input: {
     messages: rawResponseMessages,
     providerExecutedOutcomeIds,
   });
-  const responseMessages = normalizedProviderHistory.messages;
+  const deduped = dedupeToolResultsByCallId(normalizedProviderHistory.messages);
+  if (deduped.droppedCallIds.length > 0) {
+    log.warn("dropped duplicate tool results before persisting history", {
+      callIds: deduped.droppedCallIds,
+    });
+  }
+  const responseMessages = deduped.messages;
 
   const baseSession: HarnessSession = {
     ...session,