diff --git a/src/api/transform/__tests__/sanitize-messages.spec.ts b/src/api/transform/__tests__/sanitize-messages.spec.ts
new file mode 100644
index 0000000000..e5744ed0b6
--- /dev/null
+++ b/src/api/transform/__tests__/sanitize-messages.spec.ts
@@ -0,0 +1,172 @@
+import { sanitizeMessagesForProvider } from "../sanitize-messages"
+import type { RooMessage } from "../../../core/task-persistence/rooMessage"
+
+describe("sanitizeMessagesForProvider", () => {
+	it("should preserve role and content on user messages", () => {
+		const messages: RooMessage[] = [{ role: "user", content: [{ type: "text", text: "Hello" }] }]
+
+		const result = sanitizeMessagesForProvider(messages)
+
+		expect(result).toHaveLength(1)
+		expect(result[0]).toEqual({
+			role: "user",
+			content: [{ type: "text", text: "Hello" }],
+		})
+	})
+
+	it("should preserve role, content, and providerOptions on assistant messages", () => {
+		const messages: RooMessage[] = [
+			{
+				role: "assistant",
+				content: [{ type: "text", text: "Hi" }],
+				providerOptions: { openrouter: { reasoning_details: [{ type: "reasoning.text", text: "thinking" }] } },
+			} as any,
+		]
+
+		const result = sanitizeMessagesForProvider(messages)
+
+		expect(result).toHaveLength(1)
+		expect(result[0]).toEqual({
+			role: "assistant",
+			content: [{ type: "text", text: "Hi" }],
+			providerOptions: { openrouter: { reasoning_details: [{ type: "reasoning.text", text: "thinking" }] } },
+		})
+	})
+
+	it("should strip reasoning_details from messages", () => {
+		const messages = [
+			{
+				role: "assistant",
+				content: [{ type: "text", text: "Response" }],
+				reasoning_details: [{ type: "reasoning.encrypted", data: "encrypted_data" }],
+			},
+		] as any as RooMessage[]
+
+		const result = sanitizeMessagesForProvider(messages)
+
+		expect(result).toHaveLength(1)
+		expect(result[0]).not.toHaveProperty("reasoning_details")
+		expect(result[0]).toEqual({
+			role: "assistant",
+			content: [{ type: "text", text: "Response" }],
+		})
+	})
+
+	it("should strip reasoning_content from messages", () => {
+		const messages = [
+			{
+				role: "assistant",
+				content: [{ type: "text", text: "Response" }],
+				reasoning_content: "some reasoning content",
+			},
+		] as any as RooMessage[]
+
+		const result = sanitizeMessagesForProvider(messages)
+
+		expect(result).toHaveLength(1)
+		expect(result[0]).not.toHaveProperty("reasoning_content")
+	})
+
+	it("should strip metadata fields (ts, condenseId, etc.)", () => {
+		const messages = [
+			{
+				role: "user",
+				content: "Hello",
+				ts: 1234567890,
+				condenseId: "cond-1",
+				condenseParent: "cond-0",
+				truncationId: "trunc-1",
+				truncationParent: "trunc-0",
+				isTruncationMarker: true,
+				isSummary: true,
+			},
+		] as any as RooMessage[]
+
+		const result = sanitizeMessagesForProvider(messages)
+
+		expect(result).toHaveLength(1)
+		expect(result[0]).toEqual({
+			role: "user",
+			content: "Hello",
+		})
+		expect(result[0]).not.toHaveProperty("ts")
+		expect(result[0]).not.toHaveProperty("condenseId")
+		expect(result[0]).not.toHaveProperty("condenseParent")
+		expect(result[0]).not.toHaveProperty("truncationId")
+		expect(result[0]).not.toHaveProperty("truncationParent")
+		expect(result[0]).not.toHaveProperty("isTruncationMarker")
+		expect(result[0]).not.toHaveProperty("isSummary")
+	})
+
+	it("should strip any unknown extra fields", () => {
+		const messages = [
+			{
+				role: "assistant",
+				content: [{ type: "text", text: "Hi" }],
+				some_future_field: "should be stripped",
+				another_unknown: 42,
+			},
+		] as any as RooMessage[]
+
+		const result = sanitizeMessagesForProvider(messages)
+
+		expect(result).toHaveLength(1)
+		expect(result[0]).not.toHaveProperty("some_future_field")
+		expect(result[0]).not.toHaveProperty("another_unknown")
+	})
+
+	it("should not include providerOptions key when undefined", () => {
+		const messages: RooMessage[] = [{ role: "user", content: "Hello" }]
+
+		const result = sanitizeMessagesForProvider(messages)
+
+		expect(result).toHaveLength(1)
+		expect(Object.keys(result[0])).toEqual(["role", "content"])
+	})
+
+	it("should handle mixed message types correctly", () => {
+		const messages = [
+			{
+				role: "user",
+				content: [{ type: "text", text: "Hello" }],
+				ts: 100,
+			},
+			{
+				role: "assistant",
+				content: [{ type: "text", text: "Hi" }],
+				reasoning_details: [{ type: "thinking", thinking: "some reasoning" }],
+				reasoning_content: "some reasoning content",
+				ts: 200,
+			},
+			{
+				role: "tool",
+				content: [{ type: "tool-result", toolCallId: "call_1", toolName: "test", result: "ok" }],
+				ts: 300,
+			},
+			{
+				role: "user",
+				content: [{ type: "text", text: "Follow up" }],
+				ts: 400,
+			},
+		] as any as RooMessage[]
+
+		const result = sanitizeMessagesForProvider(messages)
+
+		expect(result).toHaveLength(4)
+
+		for (const msg of result) {
+			expect(msg).not.toHaveProperty("reasoning_details")
+			expect(msg).not.toHaveProperty("reasoning_content")
+			expect(msg).not.toHaveProperty("ts")
+		}
+
+		expect(result[0]).toEqual({
+			role: "user",
+			content: [{ type: "text", text: "Hello" }],
+		})
+		expect(result[1]).toEqual({
+			role: "assistant",
+			content: [{ type: "text", text: "Hi" }],
+		})
+	})
+})
diff --git a/src/api/transform/sanitize-messages.ts b/src/api/transform/sanitize-messages.ts
new file mode 100644
index 0000000000..27fe8155f4
--- /dev/null
+++ b/src/api/transform/sanitize-messages.ts
@@ -0,0 +1,32 @@
+import type { ModelMessage } from "ai"
+import type { RooMessage, RooRoleMessage } from "../../core/task-persistence/rooMessage"
+import { isRooReasoningMessage } from "../../core/task-persistence/rooMessage"
+
+/**
+ * Sanitize RooMessage[] for provider APIs by allowlisting only the fields
+ * that the AI SDK expects on each message.
+ *
+ * Legacy fields like `reasoning_details`, `reasoning_content`, `ts`, `condenseId`,
+ * etc. survive JSON deserialization round-trips and cause providers to reject
+ * requests with "Extra inputs are not permitted" (Anthropic 400) or similar errors.
+ *
+ * This uses an allowlist approach: only `role`, `content`, and `providerOptions`
+ * are forwarded, ensuring any future extraneous fields are also stripped.
+ *
+ * RooReasoningMessage items (standalone encrypted reasoning with no `role`) are
+ * filtered out since they have no AI SDK equivalent.
+ */
+export function sanitizeMessagesForProvider(messages: RooMessage[]): ModelMessage[] {
+	return messages
+		.filter((msg): msg is RooRoleMessage => !isRooReasoningMessage(msg))
+		.map((msg) => {
+			const clean: Record<string, unknown> = {
+				role: msg.role,
+				content: msg.content,
+			}
+			if (msg.providerOptions !== undefined) {
+				clean.providerOptions = msg.providerOptions
+			}
+			return clean as ModelMessage
+		})
+}