feat(ai): add chat-template thinking compat

closes #5673

packages/ai/CHANGELOG.md

@@ -2,6 +2,10 @@
 
 ## [Unreleased]
 
+### Added
+
+- Added configurable `chat-template` thinking support for OpenAI-compatible providers that use `chat_template_kwargs`, such as DeepSeek models behind vLLM ([#5673](https://github.com/earendil-works/pi/issues/5673)).
+
 ### Fixed
 
 - Fixed GitHub Copilot OAuth model availability to use the authenticated account's model picker catalog ([#5897](https://github.com/earendil-works/pi/issues/5897)).

packages/ai/README.md

@@ -967,7 +967,8 @@ interface OpenAICompletionsCompat {
   requiresAssistantAfterToolResult?: boolean; // Whether tool results must be followed by an assistant message (default: false)
   requiresThinkingAsText?: boolean;  // Whether thinking blocks must be converted to text (default: false)
   requiresReasoningContentOnAssistantMessages?: boolean; // Whether all replayed assistant messages must include empty reasoning_content when reasoning is enabled (default: auto-detected for DeepSeek)
-  thinkingFormat?: 'openai' | 'openrouter' | 'deepseek' | 'together' | 'zai' | 'qwen' | 'qwen-chat-template' | 'string-thinking' | 'ant-ling'; // Format for reasoning param: 'openai' uses reasoning_effort, 'openrouter' uses reasoning: { effort }, 'deepseek' uses thinking: { type } plus reasoning_effort when supported, 'together' uses reasoning: { enabled } plus reasoning_effort when supported, 'zai' uses enable_thinking, 'qwen' uses enable_thinking, 'qwen-chat-template' uses chat_template_kwargs.enable_thinking, 'string-thinking' uses top-level thinking, 'ant-ling' uses reasoning: { effort } only for mapped efforts (default: openai)
+  thinkingFormat?: 'openai' | 'openrouter' | 'deepseek' | 'together' | 'zai' | 'qwen' | 'chat-template' | 'qwen-chat-template' | 'string-thinking' | 'ant-ling'; // Format for reasoning param: 'openai' uses reasoning_effort, 'openrouter' uses reasoning: { effort }, 'deepseek' uses thinking: { type } plus reasoning_effort when supported, 'together' uses reasoning: { enabled } plus reasoning_effort when supported, 'zai' uses thinking: { type }, 'qwen' uses enable_thinking, 'chat-template' uses configurable chat_template_kwargs, 'qwen-chat-template' uses chat_template_kwargs.enable_thinking and preserve_thinking, 'string-thinking' uses top-level thinking, 'ant-ling' uses reasoning: { effort } only for mapped efforts (default: openai)
+  chatTemplateKwargs?: Record<string, string | number | boolean | null | { '$var': 'thinking.enabled' | 'thinking.effort'; omitWhenOff?: boolean }>; // chat_template_kwargs values; use $var for pi-controlled thinking values
   cacheControlFormat?: 'anthropic';  // Anthropic-style cache_control on system prompt, last tool, and last user/assistant text content
   openRouterRouting?: OpenRouterRouting; // OpenRouter routing preferences (default: {})
   vercelGatewayRouting?: VercelGatewayRouting; // Vercel AI Gateway routing preferences (default: {})

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

@@ -15,6 +15,7 @@ import { calculateCost, clampThinkingLevel } from "../models.ts";
 import type {
 	AssistantMessage,
 	CacheRetention,
+	ChatTemplateKwargValue,
 	Context,
 	ImageContent,
 	Message,
@@ -91,6 +92,8 @@ type ResolvedOpenAICompletionsCompat = Omit<Required<OpenAICompletionsCompat>, "
 	cacheControlFormat?: OpenAICompletionsCompat["cacheControlFormat"];
 };
 
+type ResolvedChatTemplateKwargValue = string | number | boolean | null;
+
 type ChatCompletionInstructionMessageParam = ChatCompletionDeveloperMessageParam | ChatCompletionSystemMessageParam;
 
 type ChatCompletionTextPartWithCacheControl = ChatCompletionContentPartText & {
@@ -585,6 +588,11 @@ function buildParams(
 			enable_thinking: !!options?.reasoningEffort,
 			preserve_thinking: true,
 		};
+	} else if (compat.thinkingFormat === "chat-template" && model.reasoning) {
+		const chatTemplateKwargs = buildChatTemplateKwargs(model, options, compat);
+		if (chatTemplateKwargs) {
+			(params as any).chat_template_kwargs = chatTemplateKwargs;
+		}
 	} else if (compat.thinkingFormat === "deepseek" && model.reasoning) {
 		if (options?.reasoningEffort) {
 			(params as any).thinking = { type: "enabled" };
@@ -655,6 +663,44 @@ function buildParams(
 	return params;
 }
 
+function buildChatTemplateKwargs(
+	model: Model<"openai-completions">,
+	options: OpenAICompletionsOptions | undefined,
+	compat: ResolvedOpenAICompletionsCompat,
+): Record<string, ResolvedChatTemplateKwargValue> | undefined {
+	const kwargs: Record<string, ResolvedChatTemplateKwargValue> = {};
+
+	for (const [key, value] of Object.entries(compat.chatTemplateKwargs)) {
+		const resolved = resolveChatTemplateKwargValue(model, options, value);
+		if (resolved !== undefined) {
+			kwargs[key] = resolved;
+		}
+	}
+
+	return Object.keys(kwargs).length > 0 ? kwargs : undefined;
+}
+
+function resolveChatTemplateKwargValue(
+	model: Model<"openai-completions">,
+	options: OpenAICompletionsOptions | undefined,
+	value: ChatTemplateKwargValue,
+): ResolvedChatTemplateKwargValue | undefined {
+	if (typeof value !== "object" || value === null) {
+		return value;
+	}
+
+	const reasoningEffort = options?.reasoningEffort;
+	if (!reasoningEffort && value.omitWhenOff) {
+		return undefined;
+	}
+	if (value.$var === "thinking.enabled") {
+		return !!reasoningEffort;
+	}
+
+	const mappedValue = reasoningEffort ? model.thinkingLevelMap?.[reasoningEffort] : model.thinkingLevelMap?.off;
+	return mappedValue === undefined ? reasoningEffort : typeof mappedValue === "string" ? mappedValue : undefined;
+}
+
 function getCompatCacheControl(
 	compat: ResolvedOpenAICompletionsCompat,
 	cacheRetention: CacheRetention,
@@ -1167,6 +1213,7 @@ function detectCompat(model: Model<"openai-completions">): ResolvedOpenAIComplet
 							: "openai",
 		openRouterRouting: {},
 		vercelGatewayRouting: {},
+		chatTemplateKwargs: {},
 		zaiToolStream: false,
 		supportsStrictMode: !isMoonshot && !isTogether && !isCloudflareAiGateway && !isNvidia,
 		cacheControlFormat,
@@ -1205,6 +1252,7 @@ function getCompat(model: Model<"openai-completions">): ResolvedOpenAICompletion
 		thinkingFormat: model.compat.thinkingFormat ?? detected.thinkingFormat,
 		openRouterRouting: model.compat.openRouterRouting ?? {},
 		vercelGatewayRouting: model.compat.vercelGatewayRouting ?? detected.vercelGatewayRouting,
+		chatTemplateKwargs: model.compat.chatTemplateKwargs ?? detected.chatTemplateKwargs,
 		zaiToolStream: model.compat.zaiToolStream ?? detected.zaiToolStream,
 		supportsStrictMode: model.compat.supportsStrictMode ?? detected.supportsStrictMode,
 		cacheControlFormat: model.compat.cacheControlFormat ?? detected.cacheControlFormat,

packages/ai/src/types.ts

@@ -65,6 +65,15 @@ export type ImagesProvider = KnownImagesProvider | string;
 export type ThinkingLevel = "minimal" | "low" | "medium" | "high" | "xhigh";
 export type ModelThinkingLevel = "off" | ThinkingLevel;
 export type ThinkingLevelMap = Partial<Record<ModelThinkingLevel, string | null>>;
+export type ChatTemplateKwargValue =
+	| string
+	| number
+	| boolean
+	| null
+	| {
+			$var: "thinking.enabled" | "thinking.effort";
+			omitWhenOff?: boolean;
+	  };
 
 /** Token budgets for each thinking level (token-based providers only) */
 export interface ThinkingBudgets {
@@ -403,7 +412,7 @@ export interface OpenAICompletionsCompat {
 	requiresThinkingAsText?: boolean;
 	/** Whether all replayed assistant messages must include an empty reasoning_content field when reasoning is enabled. Default: auto-detected from URL. */
 	requiresReasoningContentOnAssistantMessages?: boolean;
-	/** Format for reasoning/thinking parameter. "openai" uses reasoning_effort, "openrouter" uses reasoning: { effort }, "deepseek" uses thinking: { type } plus reasoning_effort when supported, "together" uses reasoning: { enabled } plus reasoning_effort when supported, "zai" uses thinking: { type }, "qwen" uses top-level enable_thinking: boolean, "qwen-chat-template" uses chat_template_kwargs.enable_thinking, "string-thinking" uses top-level thinking: string, and "ant-ling" uses reasoning: { effort } only when the mapped effort is non-null. Default: "openai". */
+	/** Format for reasoning/thinking parameter. "openai" uses reasoning_effort, "openrouter" uses reasoning: { effort }, "deepseek" uses thinking: { type } plus reasoning_effort when supported, "together" uses reasoning: { enabled } plus reasoning_effort when supported, "zai" uses thinking: { type }, "qwen" uses top-level enable_thinking: boolean, "qwen-chat-template" uses chat_template_kwargs.enable_thinking and preserve_thinking, "chat-template" uses configurable chat_template_kwargs, "string-thinking" uses top-level thinking: string, and "ant-ling" uses reasoning: { effort } only when the mapped effort is non-null. Default: "openai". */
 	thinkingFormat?:
 		| "openai"
 		| "openrouter"
@@ -411,9 +420,12 @@ export interface OpenAICompletionsCompat {
 		| "together"
 		| "zai"
 		| "qwen"
+		| "chat-template"
 		| "qwen-chat-template"
 		| "string-thinking"
 		| "ant-ling";
+	/** Kwargs to send as `chat_template_kwargs` when `thinkingFormat` is `chat-template`. Use `{ "$var": "thinking.enabled" }` or `{ "$var": "thinking.effort" }` for pi-controlled thinking values. */
+	chatTemplateKwargs?: Record<string, ChatTemplateKwargValue>;
 	/** OpenRouter-compatible routing preferences sent as the `provider` request field. */
 	openRouterRouting?: OpenRouterRouting;
 	/** Vercel AI Gateway routing preferences. Only used when baseUrl points to Vercel AI Gateway. */

packages/ai/test/openai-completions-thinking-as-text.test.ts

@@ -34,6 +34,7 @@ const compat = {
 	thinkingFormat: "openai",
 	openRouterRouting: {},
 	vercelGatewayRouting: {},
+	chatTemplateKwargs: {},
 	zaiToolStream: false,
 	supportsStrictMode: true,
 	cacheControlFormat: undefined,

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

@@ -3,7 +3,7 @@ import { beforeEach, describe, expect, it, vi } from "vitest";
 import { stream, streamSimple } from "../src/index.ts";
 import { getModel } from "../src/models.ts";
 import { convertMessages } from "../src/providers/openai-completions.ts";
-import type { AssistantMessage, Model, Tool, ToolResultMessage } from "../src/types.ts";
+import type { AssistantMessage, Model, SimpleStreamOptions, Tool, ToolResultMessage } from "../src/types.ts";
 
 const mockState = vi.hoisted(() => ({
 	lastParams: undefined as unknown,
@@ -64,6 +64,46 @@ vi.mock("openai", () => {
 	return { default: FakeOpenAI };
 });
 
+const localOpenAICompletionsModel = {
+	api: "openai-completions",
+	provider: "local-vllm",
+	baseUrl: "http://localhost:8000/v1",
+	reasoning: true,
+	input: ["text"],
+	cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+	contextWindow: 128000,
+	maxTokens: 8192,
+} satisfies Omit<Model<"openai-completions">, "id" | "name" | "compat">;
+
+type CapturedParams = {
+	chat_template_kwargs?: Record<string, unknown>;
+	thinking?: unknown;
+	reasoning_effort?: string;
+};
+
+async function captureSimpleParams(
+	model: Model<"openai-completions">,
+	reasoning?: SimpleStreamOptions["reasoning"],
+): Promise<CapturedParams> {
+	let payload: unknown;
+
+	await streamSimple(
+		model,
+		{
+			messages: [{ role: "user", content: "Hi", timestamp: Date.now() }],
+		},
+		{
+			apiKey: "test",
+			reasoning,
+			onPayload: (params: unknown) => {
+				payload = params;
+			},
+		},
+	).result();
+
+	return (payload ?? mockState.lastParams) as CapturedParams;
+}
+
 describe("openai-completions tool_choice", () => {
 	beforeEach(() => {
 		mockState.lastParams = undefined;
@@ -1142,6 +1182,7 @@ describe("openai-completions tool_choice", () => {
 				thinkingFormat: "openai",
 				openRouterRouting: {},
 				vercelGatewayRouting: {},
+				chatTemplateKwargs: {},
 				zaiToolStream: false,
 				supportsStrictMode: true,
 				sendSessionAffinityHeaders: false,
@@ -1449,6 +1490,77 @@ describe("openai-completions tool_choice", () => {
 		expect(params.reasoning_effort).toBeUndefined();
 	});
 
+	it("uses configurable chat template boolean thinking kwargs", async () => {
+		const model = {
+			...localOpenAICompletionsModel,
+			id: "deepseek-ai/DeepSeek-V3.1",
+			name: "DeepSeek V3.1 via vLLM",
+			compat: {
+				thinkingFormat: "chat-template",
+				supportsReasoningEffort: false,
+				chatTemplateKwargs: { thinking: { $var: "thinking.enabled" } },
+			},
+		} satisfies Model<"openai-completions">;
+
+		for (const testCase of [
+			{ reasoning: "high" as const, expected: true },
+			{ reasoning: undefined, expected: false },
+		]) {
+			const params = await captureSimpleParams(model, testCase.reasoning);
+
+			expect(params.chat_template_kwargs).toEqual({ thinking: testCase.expected });
+			expect(params.thinking).toBeUndefined();
+			expect(params.reasoning_effort).toBeUndefined();
+		}
+	});
+
+	it("uses qwen chat template thinking kwargs", async () => {
+		const model = {
+			...localOpenAICompletionsModel,
+			id: "Qwen/Qwen3-Coder",
+			name: "Qwen3 Coder via vLLM",
+			compat: {
+				thinkingFormat: "qwen-chat-template",
+				supportsReasoningEffort: false,
+			},
+		} satisfies Model<"openai-completions">;
+
+		for (const testCase of [
+			{ reasoning: "high" as const, expected: true },
+			{ reasoning: undefined, expected: false },
+		]) {
+			const params = await captureSimpleParams(model, testCase.reasoning);
+
+			expect(params.chat_template_kwargs).toEqual({
+				enable_thinking: testCase.expected,
+				preserve_thinking: true,
+			});
+			expect(params.reasoning_effort).toBeUndefined();
+		}
+	});
+
+	it("uses configurable chat template effort kwargs with static kwargs", async () => {
+		const model = {
+			...localOpenAICompletionsModel,
+			id: "unsloth/gpt-oss-120b-GGUF",
+			name: "GPT OSS via vLLM",
+			thinkingLevelMap: { xhigh: "max" },
+			compat: {
+				thinkingFormat: "chat-template",
+				supportsReasoningEffort: false,
+				chatTemplateKwargs: {
+					preserve_thinking: true,
+					reasoning_effort: { $var: "thinking.effort", omitWhenOff: true },
+				},
+			},
+		} satisfies Model<"openai-completions">;
+
+		const params = await captureSimpleParams(model, "xhigh");
+
+		expect(params.chat_template_kwargs).toEqual({ preserve_thinking: true, reasoning_effort: "max" });
+		expect(params.reasoning_effort).toBeUndefined();
+	});
+
 	it("uses Ant Ling compatibility metadata", async () => {
 		const model = getModel("ant-ling", "Ring-2.6-1T")!;
 		let payload: unknown;

packages/ai/test/openai-completions-tool-result-images.test.ts

@@ -32,6 +32,7 @@ const compat: Required<OpenAICompletionsCompat> = {
 	thinkingFormat: "openai",
 	openRouterRouting: {},
 	vercelGatewayRouting: {},
+	chatTemplateKwargs: {},
 	zaiToolStream: false,
 	supportsStrictMode: true,
 	cacheControlFormat: "anthropic",

packages/coding-agent/CHANGELOG.md

@@ -2,6 +2,10 @@
 
 ## [Unreleased]
 
+### Added
+
+- Added inherited configurable `chat-template` thinking support for OpenAI-compatible providers that use `chat_template_kwargs`, such as DeepSeek models behind vLLM ([#5673](https://github.com/earendil-works/pi/issues/5673)).
+
 ### Fixed
 
 - Fixed fuzzy `edit` matches to preserve untouched line blocks instead of rewriting the whole file through normalized content ([#5899](https://github.com/earendil-works/pi/issues/5899)).

packages/coding-agent/docs/custom-provider.md

@@ -229,7 +229,7 @@ models: [{
 }]
 ```
 
-Use `openrouter` for OpenRouter-style `reasoning: { effort }` controls. Use `together` for Together-style `reasoning: { enabled }` controls; with `supportsReasoningEffort`, it also sends `reasoning_effort`. Use `qwen-chat-template` instead for local Qwen-compatible servers that read `chat_template_kwargs.enable_thinking`.
+Use `openrouter` for OpenRouter-style `reasoning: { effort }` controls. Use `together` for Together-style `reasoning: { enabled }` controls; with `supportsReasoningEffort`, it also sends `reasoning_effort`. Use `qwen-chat-template` for local Qwen-compatible servers that read `chat_template_kwargs.enable_thinking` and need `preserve_thinking`.
 Use `cacheControlFormat: "anthropic"` for OpenAI-compatible providers that expose Anthropic-style prompt caching via `cache_control` on the system prompt, last tool definition, and last user/assistant text content.
 
 For Anthropic-compatible providers using `api: "anthropic-messages"`, set `compat.forceAdaptiveThinking: true` on models or providers whose upstream model requires adaptive thinking (`thinking.type: "adaptive"` plus `output_config.effort`). Built-in adaptive Claude models set this automatically. Set `compat.allowEmptySignature: true` only for providers that emit empty thinking signatures and expect `signature: ""` on replay.
@@ -718,7 +718,8 @@ interface ProviderModelConfig {
     requiresAssistantAfterToolResult?: boolean;
     requiresThinkingAsText?: boolean;
     requiresReasoningContentOnAssistantMessages?: boolean;
-    thinkingFormat?: "openai" | "openrouter" | "deepseek" | "together" | "zai" | "qwen" | "qwen-chat-template";
+    thinkingFormat?: "openai" | "openrouter" | "deepseek" | "together" | "zai" | "qwen" | "chat-template" | "qwen-chat-template" | "string-thinking" | "ant-ling";
+    chatTemplateKwargs?: Record<string, string | number | boolean | null | { "$var": "thinking.enabled" | "thinking.effort"; omitWhenOff?: boolean }>;
     cacheControlFormat?: "anthropic";
 
     // anthropic-messages
@@ -732,5 +733,5 @@ interface ProviderModelConfig {
 }
 ```
 
-`openrouter` sends `reasoning: { effort }`. `deepseek` sends `thinking: { type: "enabled" | "disabled" }` and `reasoning_effort` when enabled. `together` sends `reasoning: { enabled }` and also `reasoning_effort` when `supportsReasoningEffort` is enabled. `qwen` is for DashScope-style top-level `enable_thinking`. Use `qwen-chat-template` for local Qwen-compatible servers that read `chat_template_kwargs.enable_thinking`.
+`openrouter` sends `reasoning: { effort }`. `deepseek` sends `thinking: { type: "enabled" | "disabled" }` and `reasoning_effort` when enabled. `together` sends `reasoning: { enabled }` and also `reasoning_effort` when `supportsReasoningEffort` is enabled. `qwen` is for DashScope-style top-level `enable_thinking`. Use `qwen-chat-template` for local Qwen-compatible servers that read `chat_template_kwargs.enable_thinking` and need `preserve_thinking`. Use `chat-template` for configurable `chat_template_kwargs`, for example DeepSeek V3.x behind vLLM with `chatTemplateKwargs: { "thinking": { "$var": "thinking.enabled" } }`.
 `cacheControlFormat: "anthropic"` applies Anthropic-style `cache_control` markers to the system prompt, last tool definition, and last user/assistant text content.

packages/coding-agent/docs/models.md

@@ -399,14 +399,15 @@ For providers with partial OpenAI compatibility, use the `compat` field.
 | `requiresAssistantAfterToolResult` | Insert an assistant message before a user message after tool results |
 | `requiresThinkingAsText` | Convert thinking blocks to plain text |
 | `requiresReasoningContentOnAssistantMessages` | Include empty `reasoning_content` on all replayed assistant messages when reasoning is enabled |
-| `thinkingFormat` | Use `reasoning_effort`, `openrouter`, `deepseek`, `together`, `zai`, `qwen`, or `qwen-chat-template` thinking parameters |
+| `thinkingFormat` | Use `reasoning_effort`, `openrouter`, `deepseek`, `together`, `zai`, `qwen`, `chat-template`, or `qwen-chat-template` thinking parameters |
+| `chatTemplateKwargs` | `chat_template_kwargs` values for `thinkingFormat: "chat-template"`; use `{ "$var": "thinking.enabled" }` or `{ "$var": "thinking.effort" }` for pi-controlled thinking values |
 | `cacheControlFormat` | Use Anthropic-style `cache_control` markers on the system prompt, last tool definition, and last user/assistant text content. Currently only `anthropic` is supported. |
 | `supportsStrictMode` | Include the `strict` field in tool definitions |
 | `supportsLongCacheRetention` | Whether the provider accepts long cache retention when cache retention is `long`: `prompt_cache_retention: "24h"` for OpenAI prompt caching, or `cache_control.ttl: "1h"` when `cacheControlFormat` is `anthropic`. Default: `true`. |
 | `openRouterRouting` | OpenRouter provider routing preferences. This object is sent as-is in the `provider` field of the [OpenRouter API request](https://openrouter.ai/docs/guides/routing/provider-selection). |
 | `vercelGatewayRouting` | Vercel AI Gateway routing config for provider selection (`only`, `order`) |
 
-`openrouter` uses `reasoning: { effort }`. `together` uses `reasoning: { enabled }` and also `reasoning_effort` when `supportsReasoningEffort` is enabled. `qwen` uses top-level `enable_thinking`. Use `qwen-chat-template` for local Qwen-compatible servers that require `chat_template_kwargs.enable_thinking`.
+`openrouter` uses `reasoning: { effort }`. `together` uses `reasoning: { enabled }` and also `reasoning_effort` when `supportsReasoningEffort` is enabled. `qwen` uses top-level `enable_thinking`. Use `qwen-chat-template` for local Qwen-compatible servers that require `chat_template_kwargs.enable_thinking` and `preserve_thinking`. Use `chat-template` for vLLM/Hugging Face chat templates that need configurable `chat_template_kwargs`, such as `chatTemplateKwargs: { "thinking": { "$var": "thinking.enabled" } }` for DeepSeek V3.x templates.
 
 `cacheControlFormat: "anthropic"` is for OpenAI-compatible providers that expose Anthropic-style prompt caching through `cache_control` markers on text content and tool definitions.
 

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

@@ -96,6 +96,13 @@ const ThinkingLevelMapSchema = Type.Object({
 	xhigh: Type.Optional(ThinkingLevelMapValueSchema),
 });
 
+const ChatTemplateKwargScalarSchema = Type.Union([Type.String(), Type.Number(), Type.Boolean(), Type.Null()]);
+const ChatTemplateKwargVariableSchema = Type.Object({
+	$var: Type.Union([Type.Literal("thinking.enabled"), Type.Literal("thinking.effort")]),
+	omitWhenOff: Type.Optional(Type.Boolean()),
+});
+const ChatTemplateKwargSchema = Type.Union([ChatTemplateKwargScalarSchema, ChatTemplateKwargVariableSchema]);
+
 const OpenAICompletionsCompatSchema = Type.Object({
 	supportsStore: Type.Optional(Type.Boolean()),
 	supportsDeveloperRole: Type.Optional(Type.Boolean()),
@@ -114,9 +121,13 @@ const OpenAICompletionsCompatSchema = Type.Object({
 			Type.Literal("deepseek"),
 			Type.Literal("zai"),
 			Type.Literal("qwen"),
+			Type.Literal("chat-template"),
 			Type.Literal("qwen-chat-template"),
+			Type.Literal("string-thinking"),
+			Type.Literal("ant-ling"),
 		]),
 	),
+	chatTemplateKwargs: Type.Optional(Type.Record(Type.String(), ChatTemplateKwargSchema)),
 	cacheControlFormat: Type.Optional(Type.Literal("anthropic")),
 	openRouterRouting: Type.Optional(OpenRouterRoutingSchema),
 	vercelGatewayRouting: Type.Optional(VercelGatewayRoutingSchema),
@@ -289,6 +300,13 @@ function mergeCompat(
 		};
 	}
 
+	if (baseCompletions?.chatTemplateKwargs || overrideCompletions.chatTemplateKwargs) {
+		mergedCompletions.chatTemplateKwargs = {
+			...baseCompletions?.chatTemplateKwargs,
+			...overrideCompletions.chatTemplateKwargs,
+		};
+	}
+
 	return merged as Model<Api>["compat"];
 }
 

packages/coding-agent/test/model-registry.test.ts

@@ -434,6 +434,43 @@ describe("ModelRegistry", () => {
 			expect(compat?.cacheControlFormat).toBe("anthropic");
 		});
 
+		test("compat schema accepts chat template thinking configuration", () => {
+			writeRawModelsJson({
+				demo: {
+					baseUrl: "https://example.com/v1",
+					apiKey: "DEMO_KEY",
+					api: "openai-completions",
+					models: [
+						{
+							id: "demo-model",
+							reasoning: true,
+							input: ["text"],
+							cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+							contextWindow: 1000,
+							maxTokens: 100,
+							compat: {
+								thinkingFormat: "chat-template",
+								chatTemplateKwargs: {
+									preserve_thinking: true,
+									thinking: { $var: "thinking.enabled" },
+								},
+							},
+						},
+					],
+				},
+			});
+
+			const registry = ModelRegistry.create(authStorage, modelsJsonPath);
+			const compat = registry.find("demo", "demo-model")?.compat as OpenAICompletionsCompat | undefined;
+
+			expect(registry.getError()).toBeUndefined();
+			expect(compat?.thinkingFormat).toBe("chat-template");
+			expect(compat?.chatTemplateKwargs).toEqual({
+				preserve_thinking: true,
+				thinking: { $var: "thinking.enabled" },
+			});
+		});
+
 		test("compat schema accepts Anthropic eager tool input streaming flag", () => {
 			writeRawModelsJson({
 				demo: {