refactor: refactor model capability management and improve multimodal model inspection

Author: qorzj

src/common/model-capabilities.ts

@@ -0,0 +1,16 @@
+export const DEEPSEEK_V4_MODELS = new Set(["deepseek-v4-flash", "deepseek-v4-pro"]);
+
+export const NON_MULTIMODAL_MODELS = new Set([
+  "deepseek-v4-pro",
+  "deepseek-v4-flash",
+  "deepseek-chat",
+  "deepseek-reasoner",
+]);
+
+export function defaultsToThinkingMode(model: string): boolean {
+  return DEEPSEEK_V4_MODELS.has(model);
+}
+
+export function supportsMultimodal(model: string): boolean {
+  return !NON_MULTIMODAL_MODELS.has(model.trim());
+}

src/model-capabilities.ts

@@ -3,8 +3,10 @@ import * as fs from "fs";
 import * as os from "os";
 import * as path from "path";
 import { fileURLToPath } from "url";
+import ejs from "ejs";
 import type { SessionMessage } from "./session";
 import { findGitBashPath, resolveShellPath } from "./common/shell-utils";
+import { supportsMultimodal } from "./common/model-capabilities";
 
 export const AGENT_DRIFT_GUARD_SKILL = `
 ---
@@ -248,23 +250,28 @@ const SYSTEM_PROMPT_BASE = `你是名叫Deep Code的交互式CLI工具，帮助
 重要：严禁编造任何非编程相关的 URL。对于编程链接，仅限使用：1) 用户提供的上下文；2) 你确定的官方文档主域名。在输出前，必须自查该链接是否存在于你的上下文记忆中；若不存在，请明确说明无法提供。`;
 
 type PromptToolOptions = {
+  model?: string;
   webSearchEnabled?: boolean;
 };
 
-function readToolDocs(extensionRoot: string, _options: PromptToolOptions = {}): string {
+function readToolDocs(extensionRoot: string, options: PromptToolOptions = {}): string {
   const toolsDir = path.join(extensionRoot, "templates", "tools");
   if (!fs.existsSync(toolsDir)) {
     return "";
   }
 
   const entries = fs.readdirSync(toolsDir);
   const docs = entries
-    .filter((entry) => entry.endsWith(".md"))
+    .filter((entry) => entry.endsWith(".md") || entry.endsWith(".md.ejs"))
     .sort()
     .map((entry) => {
       const fullPath = path.join(toolsDir, entry);
       try {
-        return fs.readFileSync(fullPath, "utf8").trim();
+        const template = fs.readFileSync(fullPath, "utf8");
+        const content = entry.endsWith(".ejs")
+          ? ejs.render(template, { supportsMultimodal: supportsMultimodal(options.model ?? "") })
+          : template;
+        return content.trim();
       } catch {
         return "";
       }

src/prompt.ts

@@ -8,7 +8,7 @@ import ejs from "ejs";
 import type { ChatCompletionMessageParam, ChatCompletionContentPart } from "openai/resources/chat/completions";
 import { launchNotifyScript } from "./notify";
 import { buildThinkingRequestOptions } from "./openai-thinking";
-import { DEEPSEEK_V4_MODELS } from "./model-capabilities";
+import { DEEPSEEK_V4_MODELS, supportsMultimodal } from "./common/model-capabilities";
 import { getCompactPrompt, getSystemPrompt, getTools, AGENT_DRIFT_GUARD_SKILL, type ToolDefinition } from "./prompt";
 import { ToolExecutor, type CreateOpenAIClient } from "./tools/executor";
 import { McpManager } from "./mcp/mcp-manager";
@@ -156,7 +156,7 @@ export type SkillInfo = {
 type SessionManagerOptions = {
   projectRoot: string;
   createOpenAIClient: CreateOpenAIClient;
-  getResolvedSettings: () => { webSearchTool?: string; mcpServers?: Record<string, McpServerConfig> };
+  getResolvedSettings: () => { model: string; webSearchTool?: string; mcpServers?: Record<string, McpServerConfig> };
   renderMarkdown: (text: string) => string;
   onAssistantMessage: (message: SessionMessage, shouldConnect: boolean) => void;
   onSessionEntryUpdated?: (entry: SessionEntry) => void;
@@ -175,7 +175,11 @@ export type LlmStreamProgress = {
 export class SessionManager {
   private readonly projectRoot: string;
   private readonly createOpenAIClient: CreateOpenAIClient;
-  private readonly getResolvedSettings: () => { webSearchTool?: string; mcpServers?: Record<string, McpServerConfig> };
+  private readonly getResolvedSettings: () => {
+    model: string;
+    webSearchTool?: string;
+    mcpServers?: Record<string, McpServerConfig>;
+  };
   private readonly onAssistantMessage: (message: SessionMessage, shouldConnect: boolean) => void;
   private readonly onSessionEntryUpdated?: (entry: SessionEntry) => void;
   private readonly onLlmStreamProgress?: (progress: LlmStreamProgress) => void;
@@ -1017,7 +1021,7 @@ ${skillMd}
           await this.compactSession(sessionId, sessionController.signal);
         }
 
-        const messages = this.buildOpenAIMessages(this.listSessionMessages(sessionId), thinkingEnabled);
+        const messages = this.buildOpenAIMessages(this.listSessionMessages(sessionId), thinkingEnabled, model);
         const thinkingOptions = buildThinkingRequestOptions(thinkingEnabled, baseURL, reasoningEffort);
         const response = await this.createChatCompletionStream(
           client,
@@ -1208,8 +1212,9 @@ ${skillMd}
     this.saveSessionMessages(sessionId, sessionMessages);
   }
 
-  private getPromptToolOptions(): { webSearchEnabled: boolean } {
+  private getPromptToolOptions(): { model: string; webSearchEnabled: boolean } {
     return {
+      model: this.getResolvedSettings().model,
       webSearchEnabled: true,
     };
   }
@@ -1678,7 +1683,11 @@ ${skillMd}
     return { waitingForUser };
   }
 
-  private buildOpenAIMessages(messages: SessionMessage[], thinkingEnabled: boolean): ChatCompletionMessageParam[] {
+  private buildOpenAIMessages(
+    messages: SessionMessage[],
+    thinkingEnabled: boolean,
+    model: string
+  ): ChatCompletionMessageParam[] {
     const activeMessages = messages.filter((message) => !message.compacted);
     const toolPairings = this.pairToolMessages(activeMessages);
     const openAIMessages: ChatCompletionMessageParam[] = [];
@@ -1689,7 +1698,7 @@ ${skillMd}
         continue;
       }
 
-      openAIMessages.push(this.sessionMessageToOpenAIMessage(message, thinkingEnabled));
+      openAIMessages.push(this.sessionMessageToOpenAIMessage(message, thinkingEnabled, model));
 
       const toolCalls = this.getAssistantToolCalls(message);
       if (toolCalls.length === 0) {
@@ -1704,7 +1713,9 @@ ${skillMd}
 
         const pairedToolIndex = toolPairings.get(this.buildToolPairingKey(index, toolCallIndex));
         if (pairedToolIndex != null) {
-          openAIMessages.push(this.sessionMessageToOpenAIMessage(activeMessages[pairedToolIndex], thinkingEnabled));
+          openAIMessages.push(
+            this.sessionMessageToOpenAIMessage(activeMessages[pairedToolIndex], thinkingEnabled, model)
+          );
           continue;
         }
 
@@ -1715,7 +1726,11 @@ ${skillMd}
     return openAIMessages;
   }
 
-  private sessionMessageToOpenAIMessage(message: SessionMessage, thinkingEnabled: boolean): ChatCompletionMessageParam {
+  private sessionMessageToOpenAIMessage(
+    message: SessionMessage,
+    thinkingEnabled: boolean,
+    model: string
+  ): ChatCompletionMessageParam {
     const content = this.renderOpenAIMessageContent(message);
     const base: ChatCompletionMessageParam = {
       role: message.role,
@@ -1748,7 +1763,7 @@ ${skillMd}
       const params = Array.isArray(message.contentParams) ? message.contentParams : [message.contentParams];
       for (const param of params) {
         const part = param as ChatCompletionContentPart;
-        if (part && part.type !== "image_url") {
+        if (part && (part.type !== "image_url" || supportsMultimodal(model))) {
           contentParts.push(part);
         }
       }

src/session.ts

@@ -1,4 +1,4 @@
-import { defaultsToThinkingMode } from "./model-capabilities";
+import { defaultsToThinkingMode } from "./common/model-capabilities";
 
 export type DeepcodingEnv = Record<string, string | undefined> & {
   MODEL?: string;

src/settings.ts

@@ -17,9 +17,17 @@ test("getSystemPrompt always includes WebSearch docs", () => {
   assert.equal(prompt.includes("## WebSearch"), true);
 });
 
+test("getSystemPrompt renders Read docs for non-multimodal models", () => {
+  const prompt = getSystemPrompt("/tmp/project", { model: "deepseek-chat" });
+  assert.equal(prompt.includes("the current model is not multimodal"), true);
+  assert.equal(prompt.includes("the contents are presented visually"), false);
+});
+
 test("runtime prompt assets live under templates", () => {
   assert.equal(fs.existsSync(path.join(repoRoot, "templates", "tools", "web-search.md")), true);
+  assert.equal(fs.existsSync(path.join(repoRoot, "templates", "tools", "read.md.ejs")), true);
   assert.equal(fs.existsSync(path.join(repoRoot, "templates", "prompts", "init_command.md.ejs")), true);
+  assert.equal(fs.existsSync(path.join(repoRoot, "templates", "tools", "read.md")), false);
   assert.equal(fs.existsSync(path.join(repoRoot, "docs", "tools")), false);
   assert.equal(fs.existsSync(path.join(repoRoot, "docs", "prompts")), false);
 });