feat: Complete DeepSeek API adapter implementation

core/config/onboarding.ts

@@ -23,6 +23,30 @@ const GEMINI_MODEL_CONFIG = {
   apiKeyInputName: "GEMINI_API_KEY",
 };
 
+const DEEPSEEK_MODEL_CONFIG = {
+  apiKeyInputName: "DEEPSEEK_API_KEY",
+  models: [
+    {
+      slug: "deepseek/deepseek-chat",
+      model: "deepseek-chat",
+      name: "DeepSeek Chat",
+      contextLength: 131072,
+      maxTokens: 8192,
+      apiBase: "https://api.deepseek.com/",
+      roles: undefined,
+    },
+    {
+      slug: "deepseek/deepseek-reasoner",
+      model: "deepseek-reasoner",
+      name: "DeepSeek Reasoner",
+      contextLength: 131072,
+      maxTokens: 32000,
+      apiBase: "https://api.deepseek.com/",
+      roles: undefined,
+    },
+  ],
+};
+
 /**
  * We set the "best" chat + autocopmlete models by default
  * whenever a user doesn't have a config.json
@@ -97,6 +121,25 @@ export function setupProviderConfig(
         },
       }));
       break;
+    case "deepseek":
+      newModels = DEEPSEEK_MODEL_CONFIG.models.map((modelConfig) => {
+        const model: any = {
+          name: modelConfig.name,
+          provider: "deepseek",
+          model: modelConfig.model,
+          apiKey,
+          contextLength: modelConfig.contextLength,
+          defaultCompletionOptions: {
+            maxTokens: modelConfig.maxTokens,
+          },
+          roles: modelConfig.roles,
+        };
+        if (modelConfig.apiBase) {
+          model.apiBase = modelConfig.apiBase;
+        }
+        return model;
+      });
+      break;
     default:
       throw new Error(`Unknown provider: ${provider}`);
   }

core/config/yaml/models.ts

@@ -2,6 +2,7 @@ import {
   mergeConfigYamlRequestOptions,
   ModelConfig,
 } from "@continuedev/config-yaml";
+import { findLlmInfo } from "@continuedev/llm-info";
 
 import { ContinueConfig, ILLMLogger, LLMOptions } from "../..";
 import { BaseLLM } from "../../llm";
@@ -15,6 +16,82 @@ function getModelClass(
   return LLMClasses.find((llm) => llm.providerName === model.provider);
 }
 
+function applyCapabilities(options: LLMOptions, model: ModelConfig): void {
+  const { capabilities } = model;
+  // Model capabilities - need to be undefined if not found
+  // To fallback to our autodetection
+  if (capabilities?.find((c) => c === "tool_use")) {
+    options.capabilities = {
+      ...options.capabilities,
+      tools: true,
+    };
+  }
+
+  if (capabilities?.find((c) => c === "image_input")) {
+    options.capabilities = {
+      ...options.capabilities,
+      uploadImage: true,
+    };
+  }
+}
+
+function applyEmbedOptions(options: LLMOptions, model: ModelConfig): void {
+  if (model.embedOptions?.maxBatchSize) {
+    options.maxEmbeddingBatchSize = model.embedOptions.maxBatchSize;
+  }
+  if (model.embedOptions?.maxChunkSize) {
+    options.maxEmbeddingChunkSize = model.embedOptions.maxChunkSize;
+  }
+}
+
+function applyEnvOptions(
+  options: LLMOptions,
+  env: Record<string, unknown>,
+): void {
+  if (
+    "useLegacyCompletionsEndpoint" in env &&
+    typeof env.useLegacyCompletionsEndpoint === "boolean"
+  ) {
+    options.useLegacyCompletionsEndpoint = env.useLegacyCompletionsEndpoint;
+  }
+  if ("apiType" in env && typeof env.apiType === "string") {
+    options.apiType = env.apiType;
+  }
+  if ("apiVersion" in env && typeof env.apiVersion === "string") {
+    options.apiVersion = env.apiVersion;
+  }
+  if ("deployment" in env && typeof env.deployment === "string") {
+    options.deployment = env.deployment;
+  }
+  if ("deploymentId" in env && typeof env.deploymentId === "string") {
+    options.deploymentId = env.deploymentId;
+  }
+  if ("projectId" in env && typeof env.projectId === "string") {
+    options.projectId = env.projectId;
+  }
+  if ("region" in env && typeof env.region === "string") {
+    options.region = env.region;
+  }
+  if ("profile" in env && typeof env.profile === "string") {
+    options.profile = env.profile;
+  }
+  if ("accessKeyId" in env && typeof env.accessKeyId === "string") {
+    options.accessKeyId = env.accessKeyId;
+  }
+  if ("secretAccessKey" in env && typeof env.secretAccessKey === "string") {
+    options.secretAccessKey = env.secretAccessKey;
+  }
+  if ("modelArn" in env && typeof env.modelArn === "string") {
+    options.modelArn = env.modelArn;
+  }
+  if ("aiGatewaySlug" in env && typeof env.aiGatewaySlug === "string") {
+    options.aiGatewaySlug = env.aiGatewaySlug;
+  }
+  if ("accountId" in env && typeof env.accountId === "string") {
+    options.accountId = env.accountId;
+  }
+}
+
 // function getContinueProxyModelName(
 //   ownerSlug: string,
 //   packageSlug: string,
@@ -49,24 +126,40 @@ async function modelConfigToBaseLLM({
     config.requestOptions,
   );
 
+  const llmInfo = findLlmInfo(model.model, model.provider);
+  const contextLength =
+    model.defaultCompletionOptions?.contextLength ?? llmInfo?.contextLength;
+  const maxCompletionTokens = llmInfo?.maxCompletionTokens;
+  const defaultMaxTokens =
+    maxCompletionTokens && contextLength
+      ? Math.min(maxCompletionTokens, contextLength / 4)
+      : undefined;
+
   let options: LLMOptions = {
     ...rest,
-    contextLength: model.defaultCompletionOptions?.contextLength,
+    contextLength: contextLength,
     completionOptions: {
       ...(model.defaultCompletionOptions ?? {}),
       model: model.model,
       maxTokens:
         model.defaultCompletionOptions?.maxTokens ??
-        cls.defaultOptions?.completionOptions?.maxTokens,
+        cls.defaultOptions?.completionOptions?.maxTokens ??
+        defaultMaxTokens,
     },
     logger: llmLogger,
     uniqueId,
     title: model.name,
     template: model.promptTemplates?.chat,
     promptTemplates: model.promptTemplates,
-    baseAgentSystemMessage: model.chatOptions?.baseAgentSystemMessage,
-    basePlanSystemMessage: model.chatOptions?.basePlanSystemMessage,
-    baseChatSystemMessage: model.chatOptions?.baseSystemMessage,
+    baseAgentSystemMessage:
+      model.chatOptions?.baseAgentSystemMessage ??
+      cls.defaultOptions?.baseAgentSystemMessage,
+    basePlanSystemMessage:
+      model.chatOptions?.basePlanSystemMessage ??
+      cls.defaultOptions?.basePlanSystemMessage,
+    baseChatSystemMessage:
+      model.chatOptions?.baseSystemMessage ??
+      cls.defaultOptions?.baseChatSystemMessage,
     toolOverrides: model.chatOptions?.toolOverrides
       ? Object.entries(model.chatOptions.toolOverrides).map(([name, o]) => ({
           name,
@@ -83,76 +176,14 @@ async function modelConfigToBaseLLM({
     requestOptions: mergedRequestOptions,
   };
 
-  // Model capabilities - need to be undefined if not found
-  // To fallback to our autodetection
-  if (capabilities?.find((c) => c === "tool_use")) {
-    options.capabilities = {
-      ...options.capabilities,
-      tools: true,
-    };
-  }
+  // Apply capabilities from model config
+  applyCapabilities(options, model);
 
-  if (capabilities?.find((c) => c === "image_input")) {
-    options.capabilities = {
-      ...options.capabilities,
-      uploadImage: true,
-    };
-  }
+  applyEmbedOptions(options, model);
 
-  if (model.embedOptions?.maxBatchSize) {
-    options.maxEmbeddingBatchSize = model.embedOptions.maxBatchSize;
-  }
-  if (model.embedOptions?.maxChunkSize) {
-    options.maxEmbeddingChunkSize = model.embedOptions.maxChunkSize;
-  }
-
-  // These are params that are at model config level in JSON
-  // But we decided to move to nested `env` in YAML
-  // Since types vary and we don't want to blindly spread env for now,
-  // Each one is handled individually here
+  // Apply environment-specific options
   const env = model.env ?? {};
-  if (
-    "useLegacyCompletionsEndpoint" in env &&
-    typeof env.useLegacyCompletionsEndpoint === "boolean"
-  ) {
-    options.useLegacyCompletionsEndpoint = env.useLegacyCompletionsEndpoint;
-  }
-  if ("apiType" in env && typeof env.apiType === "string") {
-    options.apiType = env.apiType;
-  }
-  if ("apiVersion" in env && typeof env.apiVersion === "string") {
-    options.apiVersion = env.apiVersion;
-  }
-  if ("deployment" in env && typeof env.deployment === "string") {
-    options.deployment = env.deployment;
-  }
-  if ("deploymentId" in env && typeof env.deploymentId === "string") {
-    options.deploymentId = env.deploymentId;
-  }
-  if ("projectId" in env && typeof env.projectId === "string") {
-    options.projectId = env.projectId;
-  }
-  if ("region" in env && typeof env.region === "string") {
-    options.region = env.region;
-  }
-  if ("profile" in env && typeof env.profile === "string") {
-    options.profile = env.profile;
-  }
-  if ("accessKeyId" in env && typeof env.accessKeyId === "string") {
-    options.accessKeyId = env.accessKeyId;
-  }
-  if ("secretAccessKey" in env && typeof env.secretAccessKey === "string") {
-    options.secretAccessKey = env.secretAccessKey;
-  }
-  if ("modelArn" in env && typeof env.modelArn === "string") {
-    options.modelArn = env.modelArn;
-  }
-  if ("aiGatewaySlug" in env && typeof env.aiGatewaySlug === "string") {
-    options.aiGatewaySlug = env.aiGatewaySlug;
-  }
-  if ("accountId" in env && typeof env.accountId === "string") {
-    options.accountId = env.accountId;
-  }
+  applyEnvOptions(options, env);
 
   const llm = new cls(options);
   return llm;

core/edit/recursiveStream.ts

@@ -39,7 +39,6 @@ export async function* recursiveStream(
   const injectApplyToken = type === "apply" && shouldInjectApplyToken(llm);
   if (typeof prompt === "string") {
     const finalPrompt = injectApplyToken ? prompt + APPLY_UNIQUE_TOKEN : prompt;
-
     const generator = llm.streamComplete(finalPrompt, abortController.signal, {
       raw: true,
       prediction: undefined,
@@ -84,9 +83,11 @@ export async function* recursiveStream(
     });
 
     for await (const chunk of generator) {
-      yield chunk;
-      const rendered = renderChatMessage(chunk);
-      buffer += rendered;
+      if (chunk.role === "assistant") {
+        yield chunk;
+        const rendered = renderChatMessage(chunk);
+        buffer += rendered;
+      }
       totalTokens += countTokens(chunk.content);
 
       if (totalTokens >= safeTokens) {

core/llm/autodetect.ts

@@ -248,6 +248,7 @@ const PARALLEL_PROVIDERS: string[] = [
   "vertexai",
   "function-network",
   "scaleway",
+  "deepseek",
 ];
 
 function llmCanGenerateInParallel(provider: string, model: string): boolean {
@@ -273,6 +274,9 @@ function isProviderHandlesTemplatingOrNoTemplateTypeRequired(
     modelName.includes("kimi") ||
     modelName.includes("mercury") ||
     modelName.includes("glm") ||
+    modelName.includes("deepseek-chat") ||
+    modelName.includes("deepseek-reasoner") ||
+    modelName.includes("deepseek-fim-beta") ||
     /^o\d/.test(modelName)
   );
 }
@@ -512,6 +516,10 @@ function autodetectPromptTemplates(
     editTemplate = gptEditPrompt;
   } else if (model.includes("codestral")) {
     editTemplate = osModelsEditPrompt;
+  } else if (
+    ["deepseek-chat", "deepseek-reasoner", "deepseek-fim-beta"].includes(model)
+  ) {
+    editTemplate = osModelsEditPrompt;
   }
 
   if (editTemplate !== null) {

core/llm/countTokens.test.ts

@@ -366,3 +366,57 @@ describe("extractToolSequence", () => {
     expect(messages).toHaveLength(1); // User message should remain
   });
 });
+
+describe("compileChatMessages with prefill scenarios", () => {
+  test("should handle prefill scenario (last message is assistant)", () => {
+    const messages: ChatMessage[] = [
+      {
+        role: "user",
+        content: "Please edit this code",
+      },
+      {
+        role: "assistant",
+        content: "Sure! Here's the edited code:",
+      },
+    ];
+
+    // This should not throw an error about missing user/tool message
+    expect(() => {
+      compileChatMessages({
+        modelName: "gpt-4",
+        msgs: messages,
+        knownContextLength: 1000,
+        maxTokens: 100,
+        supportsImages: false,
+      });
+    }).not.toThrow();
+  });
+
+  test("should handle prefill scenario with system message", () => {
+    const messages: ChatMessage[] = [
+      {
+        role: "system",
+        content: "You are a helpful assistant",
+      },
+      {
+        role: "user",
+        content: "Please edit this code",
+      },
+      {
+        role: "assistant",
+        content: "Sure! Here's the edited code:",
+      },
+    ];
+
+    // This should not throw an error about missing user/tool message
+    expect(() => {
+      compileChatMessages({
+        modelName: "gpt-4",
+        msgs: messages,
+        knownContextLength: 1000,
+        maxTokens: 100,
+        supportsImages: false,
+      });
+    }).not.toThrow();
+  });
+});