fix(coding-agent): harden clipboard copy

closes #3639
2026-04-24 12:52:32 +02:00
parent 9fdb12f985
commit 1e33492525
24 changed files with 561 additions and 108 deletions
--- a/packages/ai/CHANGELOG.md
+++ b/packages/ai/CHANGELOG.md
@@ -2,8 +2,13 @@

 ## [Unreleased]

+### Added
+
+- Added DeepSeek as a built-in OpenAI-compatible provider with V4 Flash and V4 Pro models and `DEEPSEEK_API_KEY` authentication.
+
 ### Fixed

+- Fixed DeepSeek V4 session replay 400 errors by adding `thinkingFormat: "deepseek"` (sends `thinking: { type }` + `reasoning_effort`), a `reasoningEffortMap`, and `requiresReasoningContentOnAssistantMessages` compat that injects empty `reasoning_content` on all replayed assistant messages when reasoning is enabled ([#3636](https://github.com/badlogic/pi-mono/issues/3636))
 - Fixed GPT-5.5 generated context window metadata to use the observed 272k limit.

 ## [0.70.0] - 2026-04-23
--- a/packages/ai/README.md
+++ b/packages/ai/README.md
@@ -50,6 +50,7 @@ Unified LLM API with automatic model discovery, provider configuration, token an
 - **OpenAI**
 - **Azure OpenAI (Responses)**
 - **OpenAI Codex** (ChatGPT Plus/Pro subscription, requires OAuth, see below)
+- **DeepSeek**
 - **Anthropic**
 - **Google**
 - **Vertex AI** (Gemini via Vertex AI)
@@ -857,7 +858,8 @@ interface OpenAICompletionsCompat {
  requiresToolResultName?: boolean;  // Whether tool results require the `name` field (default: false)
  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)
-  thinkingFormat?: 'openai' | 'zai' | 'qwen' | 'qwen-chat-template'; // Format for reasoning param: 'openai' uses reasoning_effort, 'zai' uses thinking: { type: "enabled" }, 'qwen' uses enable_thinking: boolean, 'qwen-chat-template' uses chat_template_kwargs.enable_thinking (default: openai)
+  requiresReasoningContentOnAssistantMessages?: boolean; // Whether all replayed assistant messages must include empty reasoning_content when reasoning is enabled (default: auto-detected for DeepSeek)
+  thinkingFormat?: 'openai' | 'deepseek' | 'zai' | 'qwen' | 'qwen-chat-template'; // Format for reasoning param: 'openai' uses reasoning_effort, 'deepseek' uses thinking: { type } plus reasoning_effort, 'zai' uses enable_thinking, 'qwen' uses enable_thinking, 'qwen-chat-template' uses chat_template_kwargs.enable_thinking (default: openai)
  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: {})
@@ -1020,6 +1022,7 @@ In Node.js environments, you can set environment variables to avoid passing API
 | OpenAI | `OPENAI_API_KEY` |
 | Azure OpenAI | `AZURE_OPENAI_API_KEY` + `AZURE_OPENAI_BASE_URL` or `AZURE_OPENAI_RESOURCE_NAME` (optional `AZURE_OPENAI_API_VERSION`, `AZURE_OPENAI_DEPLOYMENT_NAME_MAP` like `model=deployment,model2=deployment2`) |
 | Anthropic | `ANTHROPIC_API_KEY` or `ANTHROPIC_OAUTH_TOKEN` |
+| DeepSeek | `DEEPSEEK_API_KEY` |
 | Google | `GEMINI_API_KEY` |
 | Vertex AI | `GOOGLE_CLOUD_API_KEY` or `GOOGLE_CLOUD_PROJECT` (or `GCLOUD_PROJECT`) + `GOOGLE_CLOUD_LOCATION` + ADC |
 | Mistral | `MISTRAL_API_KEY` |
--- a/packages/ai/scripts/generate-models.ts
+++ b/packages/ai/scripts/generate-models.ts
@@ -1012,6 +1012,57 @@ async function generateModels() {
 		});
 	}

+	const deepseekCompat: OpenAICompletionsCompat = {
+		requiresReasoningContentOnAssistantMessages: true,
+		thinkingFormat: "deepseek",
+		reasoningEffortMap: {
+			minimal: "high",
+			low: "high",
+			medium: "high",
+			high: "high",
+			xhigh: "max",
+		},
+	};
+	const deepseekV4Models: Model<"openai-completions">[] = [
+		{
+			id: "deepseek-v4-flash",
+			name: "DeepSeek V4 Flash",
+			api: "openai-completions",
+			baseUrl: "https://api.deepseek.com",
+			provider: "deepseek",
+			reasoning: true,
+			input: ["text"],
+			cost: {
+				input: 0.14,
+				output: 0.28,
+				cacheRead: 0.028,
+				cacheWrite: 0,
+			},
+			contextWindow: 1000000,
+			maxTokens: 384000,
+			compat: deepseekCompat,
+		},
+		{
+			id: "deepseek-v4-pro",
+			name: "DeepSeek V4 Pro",
+			api: "openai-completions",
+			baseUrl: "https://api.deepseek.com",
+			provider: "deepseek",
+			reasoning: true,
+			input: ["text"],
+			cost: {
+				input: 1.74,
+				output: 3.48,
+				cacheRead: 0.145,
+				cacheWrite: 0,
+			},
+			contextWindow: 1000000,
+			maxTokens: 384000,
+			compat: deepseekCompat,
+		},
+	];
+	allModels.push(...deepseekV4Models);
+
 	const minimaxDirectSupportedIds = new Set(["MiniMax-M2.7", "MiniMax-M2.7-highspeed"]);

 	for (const candidate of allModels) {
--- a/packages/ai/src/env-api-keys.ts
+++ b/packages/ai/src/env-api-keys.ts
@@ -68,6 +68,7 @@ function getApiKeyEnvVars(provider: string): readonly string[] | undefined {
 	const envMap: Record<string, string> = {
 		openai: "OPENAI_API_KEY",
 		"azure-openai-responses": "AZURE_OPENAI_API_KEY",
+		deepseek: "DEEPSEEK_API_KEY",
 		google: "GEMINI_API_KEY",
 		"google-vertex": "GOOGLE_CLOUD_API_KEY",
 		groq: "GROQ_API_KEY",
--- a/packages/ai/src/models.generated.ts
+++ b/packages/ai/src/models.generated.ts
@@ -328,6 +328,40 @@ export const MODELS = {
 			contextWindow: 1000000,
 			maxTokens: 64000,
 		} satisfies Model<"bedrock-converse-stream">,
+		"au.anthropic.claude-opus-4-6-v1": {
+			id: "au.anthropic.claude-opus-4-6-v1",
+			name: "AU Anthropic Claude Opus 4.6",
+			api: "bedrock-converse-stream",
+			provider: "amazon-bedrock",
+			baseUrl: "https://bedrock-runtime.us-east-1.amazonaws.com",
+			reasoning: true,
+			input: ["text", "image"],
+			cost: {
+				input: 16.5,
+				output: 82.5,
+				cacheRead: 0.5,
+				cacheWrite: 6.25,
+			},
+			contextWindow: 1000000,
+			maxTokens: 128000,
+		} satisfies Model<"bedrock-converse-stream">,
+		"au.anthropic.claude-sonnet-4-6": {
+			id: "au.anthropic.claude-sonnet-4-6",
+			name: "AU Anthropic Claude Sonnet 4.6",
+			api: "bedrock-converse-stream",
+			provider: "amazon-bedrock",
+			baseUrl: "https://bedrock-runtime.us-east-1.amazonaws.com",
+			reasoning: true,
+			input: ["text", "image"],
+			cost: {
+				input: 3.3,
+				output: 16.5,
+				cacheRead: 0.33,
+				cacheWrite: 4.125,
+			},
+			contextWindow: 1000000,
+			maxTokens: 128000,
+		} satisfies Model<"bedrock-converse-stream">,
 		"deepseek.r1-v1:0": {
 			id: "deepseek.r1-v1:0",
 			name: "DeepSeek-R1",
@@ -2715,6 +2749,44 @@ export const MODELS = {
 			maxTokens: 40000,
 		} satisfies Model<"openai-completions">,
 	},
+	"deepseek": {
+		"deepseek-v4-flash": {
+			id: "deepseek-v4-flash",
+			name: "DeepSeek V4 Flash",
+			api: "openai-completions",
+			provider: "deepseek",
+			baseUrl: "https://api.deepseek.com",
+			compat: {"requiresReasoningContentOnAssistantMessages":true,"thinkingFormat":"deepseek","reasoningEffortMap":{"minimal":"high","low":"high","medium":"high","high":"high","xhigh":"max"}},
+			reasoning: true,
+			input: ["text"],
+			cost: {
+				input: 0.14,
+				output: 0.28,
+				cacheRead: 0.028,
+				cacheWrite: 0,
+			},
+			contextWindow: 1000000,
+			maxTokens: 384000,
+		} satisfies Model<"openai-completions">,
+		"deepseek-v4-pro": {
+			id: "deepseek-v4-pro",
+			name: "DeepSeek V4 Pro",
+			api: "openai-completions",
+			provider: "deepseek",
+			baseUrl: "https://api.deepseek.com",
+			compat: {"requiresReasoningContentOnAssistantMessages":true,"thinkingFormat":"deepseek","reasoningEffortMap":{"minimal":"high","low":"high","medium":"high","high":"high","xhigh":"max"}},
+			reasoning: true,
+			input: ["text"],
+			cost: {
+				input: 1.74,
+				output: 3.48,
+				cacheRead: 0.145,
+				cacheWrite: 0,
+			},
+			contextWindow: 1000000,
+			maxTokens: 384000,
+		} satisfies Model<"openai-completions">,
+	},
 	"fireworks": {
 		"accounts/fireworks/models/deepseek-v3p1": {
 			id: "accounts/fireworks/models/deepseek-v3p1",
@@ -8162,6 +8234,40 @@ export const MODELS = {
 			contextWindow: 163840,
 			maxTokens: 65536,
 		} satisfies Model<"openai-completions">,
+		"deepseek/deepseek-v4-flash": {
+			id: "deepseek/deepseek-v4-flash",
+			name: "DeepSeek: DeepSeek V4 Flash",
+			api: "openai-completions",
+			provider: "openrouter",
+			baseUrl: "https://openrouter.ai/api/v1",
+			reasoning: true,
+			input: ["text"],
+			cost: {
+				input: 0.14,
+				output: 0.28,
+				cacheRead: 0.028,
+				cacheWrite: 0,
+			},
+			contextWindow: 1048576,
+			maxTokens: 384000,
+		} satisfies Model<"openai-completions">,
+		"deepseek/deepseek-v4-pro": {
+			id: "deepseek/deepseek-v4-pro",
+			name: "DeepSeek: DeepSeek V4 Pro",
+			api: "openai-completions",
+			provider: "openrouter",
+			baseUrl: "https://openrouter.ai/api/v1",
+			reasoning: true,
+			input: ["text"],
+			cost: {
+				input: 1.74,
+				output: 3.48,
+				cacheRead: 0.145,
+				cacheWrite: 0,
+			},
+			contextWindow: 1048576,
+			maxTokens: 384000,
+		} satisfies Model<"openai-completions">,
 		"essentialai/rnj-1-instruct": {
 			id: "essentialai/rnj-1-instruct",
 			name: "EssentialAI: Rnj 1 Instruct",
@@ -12380,6 +12486,40 @@ export const MODELS = {
 			contextWindow: 128000,
 			maxTokens: 64000,
 		} satisfies Model<"anthropic-messages">,
+		"deepseek/deepseek-v4-flash": {
+			id: "deepseek/deepseek-v4-flash",
+			name: "DeepSeek V4 Flash",
+			api: "anthropic-messages",
+			provider: "vercel-ai-gateway",
+			baseUrl: "https://ai-gateway.vercel.sh",
+			reasoning: true,
+			input: ["text"],
+			cost: {
+				input: 0.14,
+				output: 0.28,
+				cacheRead: 0.014,
+				cacheWrite: 0,
+			},
+			contextWindow: 1000000,
+			maxTokens: 384000,
+		} satisfies Model<"anthropic-messages">,
+		"deepseek/deepseek-v4-pro": {
+			id: "deepseek/deepseek-v4-pro",
+			name: "DeepSeek V4 Pro",
+			api: "anthropic-messages",
+			provider: "vercel-ai-gateway",
+			baseUrl: "https://ai-gateway.vercel.sh",
+			reasoning: true,
+			input: ["text"],
+			cost: {
+				input: 1.74,
+				output: 3.48,
+				cacheRead: 0.145,
+				cacheWrite: 0,
+			},
+			contextWindow: 1000000,
+			maxTokens: 384000,
+		} satisfies Model<"anthropic-messages">,
 		"google/gemini-2.0-flash": {
 			id: "google/gemini-2.0-flash",
 			name: "Gemini 2.0 Flash",
--- a/packages/ai/src/providers/openai-completions.ts
+++ b/packages/ai/src/providers/openai-completions.ts
@@ -526,6 +526,11 @@ function buildParams(
 			enable_thinking: !!options?.reasoningEffort,
 			preserve_thinking: true,
 		};
+	} else if (compat.thinkingFormat === "deepseek" && model.reasoning) {
+		(params as any).thinking = { type: options?.reasoningEffort ? "enabled" : "disabled" };
+		if (options?.reasoningEffort) {
+			(params as any).reasoning_effort = mapReasoningEffort(options.reasoningEffort, compat.reasoningEffortMap);
+		}
 	} else if (compat.thinkingFormat === "openrouter" && model.reasoning) {
 		// OpenRouter normalizes reasoning across providers via a nested reasoning object.
 		const openRouterParams = params as typeof params & { reasoning?: { effort?: string } };
@@ -831,6 +836,13 @@ export function convertMessages(
 					(assistantMsg as any).reasoning_details = reasoningDetails;
 				}
 			}
+			if (
+				compat.requiresReasoningContentOnAssistantMessages &&
+				model.reasoning &&
+				(assistantMsg as { reasoning_content?: string }).reasoning_content === undefined
+			) {
+				(assistantMsg as { reasoning_content?: string }).reasoning_content = "";
+			}
 			// Skip assistant messages that have no content and no tool calls.
 			// Some providers require "either content or tool_calls, but not none".
 			// Other providers also don't accept empty assistant messages.
@@ -1021,10 +1033,18 @@ function detectCompat(model: Model<"openai-completions">): ResolvedOpenAIComplet

 	const isGrok = provider === "xai" || baseUrl.includes("api.x.ai");
 	const isGroq = provider === "groq" || baseUrl.includes("groq.com");
+	const isDeepSeek = provider === "deepseek" || baseUrl.includes("deepseek.com");
 	const cacheControlFormat = provider === "openrouter" && model.id.startsWith("anthropic/") ? "anthropic" : undefined;

-	const reasoningEffortMap =
-		isGroq && model.id === "qwen/qwen3-32b"
+	const reasoningEffortMap = isDeepSeek
+		? {
+				minimal: "high",
+				low: "high",
+				medium: "high",
+				high: "high",
+				xhigh: "max",
+			}
+		: isGroq && model.id === "qwen/qwen3-32b"
 			? {
 					minimal: "default",
 					low: "default",
@@ -1043,11 +1063,14 @@ function detectCompat(model: Model<"openai-completions">): ResolvedOpenAIComplet
 		requiresToolResultName: false,
 		requiresAssistantAfterToolResult: false,
 		requiresThinkingAsText: false,
-		thinkingFormat: isZai
-			? "zai"
-			: provider === "openrouter" || baseUrl.includes("openrouter.ai")
-				? "openrouter"
-				: "openai",
+		requiresReasoningContentOnAssistantMessages: isDeepSeek,
+		thinkingFormat: isDeepSeek
+			? "deepseek"
+			: isZai
+				? "zai"
+				: provider === "openrouter" || baseUrl.includes("openrouter.ai")
+					? "openrouter"
+					: "openai",
 		openRouterRouting: {},
 		vercelGatewayRouting: {},
 		zaiToolStream: false,
@@ -1077,6 +1100,9 @@ function getCompat(model: Model<"openai-completions">): ResolvedOpenAICompletion
 		requiresAssistantAfterToolResult:
 			model.compat.requiresAssistantAfterToolResult ?? detected.requiresAssistantAfterToolResult,
 		requiresThinkingAsText: model.compat.requiresThinkingAsText ?? detected.requiresThinkingAsText,
+		requiresReasoningContentOnAssistantMessages:
+			model.compat.requiresReasoningContentOnAssistantMessages ??
+			detected.requiresReasoningContentOnAssistantMessages,
 		thinkingFormat: model.compat.thinkingFormat ?? detected.thinkingFormat,
 		openRouterRouting: model.compat.openRouterRouting ?? {},
 		vercelGatewayRouting: model.compat.vercelGatewayRouting ?? detected.vercelGatewayRouting,
--- a/packages/ai/src/types.ts
+++ b/packages/ai/src/types.ts
@@ -26,6 +26,7 @@ export type KnownProvider =
 	| "openai"
 	| "azure-openai-responses"
 	| "openai-codex"
+	| "deepseek"
 	| "github-copilot"
 	| "xai"
 	| "groq"
@@ -282,8 +283,10 @@ export interface OpenAICompletionsCompat {
 	requiresAssistantAfterToolResult?: boolean;
 	/** Whether thinking blocks must be converted to text blocks with <thinking> delimiters. Default: auto-detected from URL. */
 	requiresThinkingAsText?: boolean;
-	/** Format for reasoning/thinking parameter. "openai" uses reasoning_effort, "openrouter" uses reasoning: { effort }, "zai" uses top-level enable_thinking: boolean, "qwen" uses top-level enable_thinking: boolean, and "qwen-chat-template" uses chat_template_kwargs.enable_thinking. Default: "openai". */
-	thinkingFormat?: "openai" | "openrouter" | "zai" | "qwen" | "qwen-chat-template";
+	/** 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, "zai" uses top-level enable_thinking: boolean, "qwen" uses top-level enable_thinking: boolean, and "qwen-chat-template" uses chat_template_kwargs.enable_thinking. Default: "openai". */
+	thinkingFormat?: "openai" | "openrouter" | "deepseek" | "zai" | "qwen" | "qwen-chat-template";
 	/** OpenRouter-specific routing preferences. Only used when baseUrl points to OpenRouter. */
 	openRouterRouting?: OpenRouterRouting;
 	/** Vercel AI Gateway routing preferences. Only used when baseUrl points to Vercel AI Gateway. */
--- a/packages/ai/test/openai-completions-thinking-as-text.test.ts
+++ b/packages/ai/test/openai-completions-thinking-as-text.test.ts
@@ -31,6 +31,7 @@ const compat = {
 	requiresToolResultName: false,
 	requiresAssistantAfterToolResult: false,
 	requiresThinkingAsText: true,
+	requiresReasoningContentOnAssistantMessages: false,
 	thinkingFormat: "openai",
 	openRouterRouting: {},
 	vercelGatewayRouting: {},
--- a/packages/ai/test/openai-completions-tool-result-images.test.ts
+++ b/packages/ai/test/openai-completions-tool-result-images.test.ts
@@ -29,6 +29,7 @@ const compat: Required<OpenAICompletionsCompat> = {
 	requiresToolResultName: false,
 	requiresAssistantAfterToolResult: false,
 	requiresThinkingAsText: false,
+	requiresReasoningContentOnAssistantMessages: false,
 	thinkingFormat: "openai",
 	openRouterRouting: {},
 	vercelGatewayRouting: {},
--- a/packages/ai/test/stream.test.ts
+++ b/packages/ai/test/stream.test.ts
@@ -447,6 +447,33 @@ describe("Generate E2E Tests", () => {
 		});
 	});

+	describe.skipIf(!process.env.DEEPSEEK_API_KEY)(
+		"DeepSeek Provider (deepseek-v4-flash via OpenAI Completions)",
+		() => {
+			const llm = getModel("deepseek", "deepseek-v4-flash");
+
+			it("should complete basic text generation", { retry: 3 }, async () => {
+				await basicTextGeneration(llm);
+			});
+
+			it("should handle tool calling", { retry: 3 }, async () => {
+				await handleToolCall(llm);
+			});
+
+			it("should handle streaming", { retry: 3 }, async () => {
+				await handleStreaming(llm);
+			});
+
+			it("should handle thinking mode", { retry: 3 }, async () => {
+				await handleThinking(llm, { reasoningEffort: "high" });
+			});
+
+			it("should handle multi-turn with thinking and tools", { retry: 3 }, async () => {
+				await multiTurn(llm, { reasoningEffort: "high" });
+			});
+		},
+	);
+
 	describe.skipIf(!process.env.OPENAI_API_KEY)("OpenAI Responses Provider (gpt-5.4)", () => {
 		const llm = getModel("openai", "gpt-5.4");