fix(ai): explicitly disable Anthropic thinking when off closes #2022

packages/ai/CHANGELOG.md

@@ -10,6 +10,7 @@
 ### Fixed
 
 - Fixed OpenAI Responses replay for foreign tool-call item IDs by hashing foreign `function_call.id` values into bounded `fc_<hash>` IDs instead of preserving backend-specific normalized shapes that OpenAI Codex rejects.
+- Fixed Anthropic thinking disable handling to send `thinking: { type: "disabled" }` for reasoning-capable models when thinking is explicitly off, and added payload and env-gated end-to-end coverage for the Anthropic provider ([#2022](https://github.com/badlogic/pi-mono/issues/2022))
 
 ## [0.61.1] - 2026-03-20
 

packages/ai/src/providers/anthropic.ts

@@ -654,20 +654,25 @@ function buildParams(
 		params.tools = convertTools(context.tools, isOAuthToken);
 	}
 
-	// Configure thinking mode: adaptive (Opus 4.6 and Sonnet 4.6) or budget-based (older models)
-	if (options?.thinkingEnabled && model.reasoning) {
-		if (supportsAdaptiveThinking(model.id)) {
-			// Adaptive thinking: Claude decides when and how much to think
-			params.thinking = { type: "adaptive" };
-			if (options.effort) {
-				params.output_config = { effort: options.effort };
+	// Configure thinking mode: adaptive (Opus 4.6 and Sonnet 4.6),
+	// budget-based (older models), or explicitly disabled.
+	if (model.reasoning) {
+		if (options?.thinkingEnabled) {
+			if (supportsAdaptiveThinking(model.id)) {
+				// Adaptive thinking: Claude decides when and how much to think
+				params.thinking = { type: "adaptive" };
+				if (options.effort) {
+					params.output_config = { effort: options.effort };
+				}
+			} else {
+				// Budget-based thinking for older models
+				params.thinking = {
+					type: "enabled",
+					budget_tokens: options.thinkingBudgetTokens || 1024,
+				};
 			}
-		} else {
-			// Budget-based thinking for older models
-			params.thinking = {
-				type: "enabled",
-				budget_tokens: options.thinkingBudgetTokens || 1024,
-			};
+		} else if (options?.thinkingEnabled === false) {
+			params.thinking = { type: "disabled" };
 		}
 	}
 

packages/ai/test/anthropic-thinking-disable.test.ts

@@ -0,0 +1,127 @@
+import { describe, expect, it } from "vitest";
+import { getModel } from "../src/models.js";
+import { streamSimple } from "../src/stream.js";
+import type { Context, Model } from "../src/types.js";
+
+interface AnthropicThinkingPayload {
+	thinking?: { type: string; budget_tokens?: number };
+	output_config?: { effort?: string };
+}
+
+function makePayloadCaptureContext(): Context {
+	return {
+		messages: [{ role: "user", content: "Hello", timestamp: Date.now() }],
+	};
+}
+
+async function capturePayload(model: Model<"anthropic-messages">): Promise<AnthropicThinkingPayload> {
+	let capturedPayload: AnthropicThinkingPayload | undefined;
+	const payloadCaptureModel: Model<"anthropic-messages"> = {
+		...model,
+		baseUrl: "http://127.0.0.1:9",
+	};
+
+	const s = streamSimple(payloadCaptureModel, makePayloadCaptureContext(), {
+		apiKey: "fake-key",
+		onPayload: (payload) => {
+			capturedPayload = payload as AnthropicThinkingPayload;
+			return payload;
+		},
+	});
+
+	await s.result();
+
+	if (!capturedPayload) {
+		throw new Error("Expected payload to be captured before request failure");
+	}
+
+	return capturedPayload;
+}
+
+interface RunResult {
+	thinkingEventCount: number;
+	thinkingCharCount: number;
+	text: string;
+	contentTypes: string[];
+}
+
+function makeE2EContext(): Context {
+	return {
+		systemPrompt: "You are a precise assistant. Follow the requested output format exactly.",
+		messages: [
+			{
+				role: "user",
+				content:
+					"Before replying, carefully solve 36863 * 5279 internally. Then reply with the word pong repeated exactly 40 times, separated by single spaces. Do not add any other text.",
+				timestamp: Date.now(),
+			},
+		],
+	};
+}
+
+function countPongs(text: string): number {
+	return text.match(/\bpong\b/gi)?.length ?? 0;
+}
+
+async function runWithoutReasoning(model: Model<"anthropic-messages">): Promise<RunResult> {
+	const s = streamSimple(model, makeE2EContext(), {
+		temperature: 0,
+		maxTokens: 160,
+	});
+
+	let thinkingEventCount = 0;
+	let thinkingCharCount = 0;
+
+	for await (const event of s) {
+		if (event.type === "thinking_start" || event.type === "thinking_end") {
+			thinkingEventCount += 1;
+		}
+		if (event.type === "thinking_delta") {
+			thinkingEventCount += 1;
+			thinkingCharCount += event.delta.length;
+		}
+	}
+
+	const response = await s.result();
+	expect(response.stopReason, response.errorMessage).toBe("stop");
+
+	const text = response.content
+		.filter((block) => block.type === "text")
+		.map((block) => block.text)
+		.join("")
+		.trim();
+
+	return {
+		thinkingEventCount,
+		thinkingCharCount,
+		text,
+		contentTypes: response.content.map((block) => block.type),
+	};
+}
+
+describe("Anthropic thinking disable payload", () => {
+	it("sends thinking.type=disabled for budget-based reasoning models when thinking is off", async () => {
+		const payload = await capturePayload(getModel("anthropic", "claude-sonnet-4-5"));
+
+		expect(payload.thinking).toEqual({ type: "disabled" });
+		expect(payload.output_config).toBeUndefined();
+	});
+
+	it("sends thinking.type=disabled for adaptive reasoning models when thinking is off", async () => {
+		const payload = await capturePayload(getModel("anthropic", "claude-opus-4-6"));
+
+		expect(payload.thinking).toEqual({ type: "disabled" });
+		expect(payload.output_config).toBeUndefined();
+	});
+});
+
+describe.skipIf(!process.env.ANTHROPIC_API_KEY)("Anthropic thinking disable E2E", () => {
+	it("disables thinking for Claude reasoning models", { retry: 2, timeout: 30000 }, async () => {
+		const result = await runWithoutReasoning(getModel("anthropic", "claude-sonnet-4-5"));
+
+		expect(result.thinkingEventCount).toBe(0);
+		expect(result.thinkingCharCount).toBe(0);
+		expect(result.contentTypes).not.toContain("thinking");
+		expect(countPongs(result.text)).toBeGreaterThanOrEqual(35);
+	});
+});