fix(ai): preserve Z.AI thinking content

closes #6083

packages/ai/CHANGELOG.md

@@ -15,6 +15,7 @@
 - Fixed `streamSimple()` to send a context-aware max-token cap so providers that count input and output against one context window do not reject long requests ([#5595](https://github.com/earendil-works/pi/issues/5595)).
 - Fixed OpenAI Responses streams to preserve reasoning replay state when output items finish out of order ([#6009](https://github.com/earendil-works/pi/issues/6009)).
 - Fixed retry classification for provider errors that explicitly tell callers to retry the request ([#6019](https://github.com/earendil-works/pi/issues/6019)).
+- Fixed Z.AI preserved thinking requests to send `thinking.clear_thinking: false` when thinking is enabled, allowing replayed `reasoning_content` to participate in provider caching ([#6083](https://github.com/earendil-works/pi/issues/6083)).
 
 ## [0.80.2] - 2026-06-23
 

packages/ai/src/api/openai-completions.ts

@@ -593,10 +593,10 @@ function buildParams(
 
 	if (compat.thinkingFormat === "zai" && model.reasoning) {
 		const zaiParams = params as Omit<typeof params, "reasoning_effort"> & {
-			thinking?: { type: "enabled" | "disabled" };
+			thinking?: { type: "enabled" | "disabled"; clear_thinking?: boolean };
 			reasoning_effort?: string;
 		};
-		zaiParams.thinking = { type: options?.reasoningEffort ? "enabled" : "disabled" };
+		zaiParams.thinking = options?.reasoningEffort ? { type: "enabled", clear_thinking: false } : { type: "disabled" };
 		if (options?.reasoningEffort && compat.supportsReasoningEffort) {
 			const mappedEffort = model.thinkingLevelMap?.[options.reasoningEffort];
 			const effort = mappedEffort === undefined ? options.reasoningEffort : mappedEffort;

packages/ai/test/openai-completions-tool-choice.test.ts

@@ -343,11 +343,71 @@ describe("openai-completions tool_choice", () => {
 			).result();
 
 			const params = (payload ?? mockState.lastParams) as { thinking?: unknown; reasoning_effort?: string };
-			expect(params.thinking).toEqual({ type: "enabled" });
+			expect(params.thinking).toEqual({ type: "enabled", clear_thinking: false });
 			expect(params.reasoning_effort).toBe(testCase.effort);
 		}
 	});
 
+	it("preserves z.ai thinking when replaying reasoning_content", async () => {
+		const model = getModel("zai", "glm-5.2")!;
+		const assistantMessage: AssistantMessage = {
+			role: "assistant",
+			api: "openai-completions",
+			provider: "zai",
+			model: "glm-5.2",
+			content: [
+				{ type: "thinking", thinking: "prior reasoning", thinkingSignature: "reasoning_content" },
+				{ type: "toolCall", id: "call_1", name: "read", arguments: { path: "README.md" } },
+			],
+			usage: {
+				input: 0,
+				output: 0,
+				cacheRead: 0,
+				cacheWrite: 0,
+				totalTokens: 0,
+				cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+			},
+			stopReason: "toolUse",
+			timestamp: Date.now(),
+		};
+		const toolResult: ToolResultMessage = {
+			role: "toolResult",
+			toolCallId: "call_1",
+			toolName: "read",
+			content: [{ type: "text", text: "contents" }],
+			isError: false,
+			timestamp: Date.now(),
+		};
+		let payload: unknown;
+
+		await streamSimple(
+			model,
+			{
+				messages: [
+					{ role: "user", content: "Read README.md", timestamp: Date.now() },
+					assistantMessage,
+					toolResult,
+					{ role: "user", content: "Continue", timestamp: Date.now() },
+				],
+			},
+			{
+				apiKey: "test",
+				reasoning: "high",
+				onPayload: (params: unknown) => {
+					payload = params;
+				},
+			},
+		).result();
+
+		const params = (payload ?? mockState.lastParams) as {
+			messages?: Array<Record<string, unknown>>;
+			thinking?: unknown;
+		};
+		const replayedAssistant = params.messages?.find((message) => message.role === "assistant");
+		expect(replayedAssistant).toMatchObject({ reasoning_content: "prior reasoning" });
+		expect(params.thinking).toEqual({ type: "enabled", clear_thinking: false });
+	});
+
 	it("omits z.ai GLM-5.2 reasoning_effort when thinking is off", async () => {
 		const model = getModel("zai", "glm-5.2")!;
 		let payload: unknown;