feat(ai/providers): dropped auto tool choice for reasoning models

- Add `shouldDropAutoToolChoiceForReasoning` helper to handle compatibility constraints for models where tool choice disables reasoning.
- Update OpenAI completions and responses parameter building to drop automatic tool choices when reasoning options are active.

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

@@ -106,6 +106,7 @@ import {
 	resolveOpenAICompletionsOutputClamp,
 	resolveOpenAIOutputTokenParam,
 	resolveOpenAIRequestSetup,
+	shouldDropAutoToolChoiceForReasoning,
 	shouldRetryWithoutStrictTools,
 } from "./openai-shared";
 import { transformMessages } from "./transform-messages";
@@ -1692,6 +1693,10 @@ function buildParams(
 		delete params.tool_choice;
 	}
 
+	if (shouldDropAutoToolChoiceForReasoning(model, initialCompat, params.tool_choice, options)) {
+		delete params.tool_choice;
+	}
+
 	const finalPolicy = resolveOpenAICompatPolicy(model, {
 		endpoint: "chat-completions",
 		reasoning: options?.reasoning,

packages/ai/src/providers/openai-responses.ts

@@ -99,6 +99,7 @@ import {
 	resolveOpenAIOutputTokenParam,
 	resolveOpenAIRequestSetup,
 	resolveOpenAIResponsesOutputClamp,
+	shouldDropAutoToolChoiceForReasoning,
 	shouldRetryWithoutStrictTools,
 } from "./openai-shared";
 
@@ -1276,6 +1277,10 @@ export function buildParams(
 		}
 	}
 
+	if (shouldDropAutoToolChoiceForReasoning(model, model.compat, params.tool_choice, options)) {
+		delete params.tool_choice;
+	}
+
 	const reasoningPolicy = resolveOpenAICompatPolicy(model, {
 		endpoint: "responses",
 		reasoning: options?.reasoning,

packages/ai/src/providers/openai-shared.ts

@@ -844,6 +844,29 @@ function isImplicitDisableWhenNotRequested(disableMode: OpenAIReasoningDisableMo
 	);
 }
 
+/**
+ * Whether a redundant `tool_choice: "auto"` should be dropped to keep
+ * reasoning alive. Hosts with `disableReasoningOnToolChoice` (DeepSeek family
+ * on e.g. Fireworks) silently turn reasoning off whenever any `tool_choice`
+ * is present. "auto" is already the provider default, so omitting it is
+ * wire-neutral for tool selection; forced and "none" choices are semantic and
+ * still win over reasoning (#1207).
+ */
+export function shouldDropAutoToolChoiceForReasoning(
+	model: Pick<Model, "reasoning">,
+	compat: { disableReasoningOnToolChoice: boolean },
+	toolChoice: unknown,
+	options: { reasoning?: string; disableReasoning?: boolean } | undefined,
+): boolean {
+	return (
+		toolChoice === "auto" &&
+		compat.disableReasoningOnToolChoice &&
+		Boolean(model.reasoning) &&
+		options?.reasoning !== undefined &&
+		!options.disableReasoning
+	);
+}
+
 export function resolveOpenAICompatPolicy<TApi extends Api>(
 	model: Model<TApi>,
 	options: ResolveOpenAICompatPolicyOptions,

packages/ai/test/openai-completions-compat.test.ts

@@ -1792,7 +1792,9 @@ describe("kimi model detection via detectCompat", () => {
 	// Dropping reasoning_effort does not turn off the gateway's default thinking
 	// mode, so the compat descriptor itself must mark forced tool choice
 	// unsupported (no per-model override) and buildParams must downgrade the
-	// selector to "auto" while keeping the tool advertised.
+	// selector while keeping the tool advertised. The downgraded "auto" is then
+	// dropped as redundant so reasoning survives (#1207) — omission and "auto"
+	// are wire-equivalent for tool selection.
 	it("scopes the DeepSeek forced tool_choice downgrade to OpenCode gateways", async () => {
 		const todoTool: Tool = {
 			name: "todo",
@@ -1839,7 +1841,7 @@ describe("kimi model detection via detectCompat", () => {
 		const openCode = buildModel(deepseekSpec);
 		expect(openCode.compat.supportsForcedToolChoice).toBe(false);
 		const openCodePayload = await captureToolChoice(openCode);
-		expect(openCodePayload.tool_choice).toBe("auto");
+		expect(openCodePayload.tool_choice).toBeUndefined();
 		expect(
 			Array.isArray(openCodePayload.tools) &&
 				openCodePayload.tools.some(tool => getNestedObject(tool, "function")?.name === "todo"),
@@ -1853,7 +1855,7 @@ describe("kimi model detection via detectCompat", () => {
 		} satisfies ModelSpec<"openai-completions">);
 		expect(customOpenCode.compat.supportsForcedToolChoice).toBe(false);
 		const customPayload = await captureToolChoice(customOpenCode);
-		expect(customPayload.tool_choice).toBe("auto");
+		expect(customPayload.tool_choice).toBeUndefined();
 
 		const nvidia = buildModel({
 			...deepseekSpec,

packages/ai/test/openai-completions-reasoning-disable-dialects.test.ts

@@ -121,9 +121,11 @@ describe("Chat Completions reasoning-disable conflict policy (per dialect)", ()
 	});
 
 	it("applies the same per-dialect disable on the non-forced disableReasoningOnToolChoice path", async () => {
-		// Non-forced branch (any tool_choice) shares the dialect-aware helper:
-		// an OpenRouter reasoning model with disableReasoningOnToolChoice must emit
+		// Non-forced branch shares the dialect-aware helper: an OpenRouter
+		// reasoning model with disableReasoningOnToolChoice must emit
 		// reasoning={enabled:false} rather than leaving the effort object in place.
+		// A semantic "none" selector exercises the path; a bare "auto" no longer
+		// does — it is dropped as redundant so reasoning survives (#1207).
 		const model = reasoningDialectModel("openrouter", {
 			disableReasoningOnForcedToolChoice: false,
 			disableReasoningOnToolChoice: true,
@@ -134,11 +136,11 @@ describe("Chat Completions reasoning-disable conflict policy (per dialect)", ()
 			fetch: createMockFetch(),
 			signal: createAbortedSignal(),
 			reasoning: "high",
-			toolChoice: "auto",
+			toolChoice: "none",
 			onPayload: payload => resolve(payload),
 		});
 		const payload = (await promise) as Record<string, unknown>;
-		expect(payload.tool_choice).toBe("auto");
+		expect(payload.tool_choice).toBe("none");
 		expect(payload.reasoning).toEqual({ enabled: false });
 	});
 });

packages/coding-agent/src/config/model-registry.ts

@@ -108,7 +108,7 @@ export {
 
 import { ModelsConfigFile, type ProviderValidationModel, validateProviderConfiguration } from "./models-config";
 import type { ModelOverride, ModelsConfig, ProviderAuthMode } from "./models-config-schema";
-import { settings, type Settings } from "./settings";
+import { type Settings, settings } from "./settings";
 
 // DeviceCheck attestation (`x-oai-attestation`) for ChatGPT-OAuth Codex
 // requests; the pi-ai provider resolves it just-in-time per request.

packages/coding-agent/src/sdk.ts

@@ -1251,9 +1251,13 @@ async function createAgentSessionScoped(options: CreateAgentSessionOptions): Pro
 	// / session would silently miss credential_disabled events.
 	const modelRegistry =
 		options.modelRegistry ??
-		new ModelRegistry(options.authStorage ?? (await logger.time("discoverModels", discoverAuthStorage, agentDir)), undefined, {
-			settings,
-		});
+		new ModelRegistry(
+			options.authStorage ?? (await logger.time("discoverModels", discoverAuthStorage, agentDir)),
+			undefined,
+			{
+				settings,
+			},
+		);
 	// Track whether we internally created the authStorage so we can close it
 	// if construction fails before the session takes ownership.
 	const ownsAuthStorage = !options.authStorage && !options.modelRegistry;