识图模型工具已实现，测试全部通过

src/config.test.ts

@@ -4,7 +4,8 @@ import os from "node:os";
 import path from "node:path";
 import test from "node:test";
 import YAML from "yaml";
-import { configPath, initializeWorkspace, isOnboardingComplete, loadConfig, markOnboardingComplete, oobeProgressPath, readOobeProgress, readSettings, saveOobeProgress, saveSettings, type OobeProgress } from "./config.js";
+import { OFFICIAL_VISION_MODEL, configPath, initializeWorkspace, isOnboardingComplete, loadConfig, markOnboardingComplete, oobeProgressPath, readOobeProgress, readSettings, resolveVisionAgentConfig, saveOobeProgress, saveSettings, type OobeProgress } from "./config.js";
+import type { SecAgentConfig } from "./types.js";
 import { SYSTEM_PROMPT } from "./system-prompt.js";
 
 function temporaryWorkspace(): string {
@@ -121,3 +122,87 @@ test("defaults and persists Windows update preferences", () => {
     fs.rmSync(workspace, { recursive: true, force: true });
   }
 });
+
+test("persists the vision model id with the settings", () => {
+  const workspace = temporaryWorkspace();
+  try {
+    initializeWorkspace(workspace);
+    const settings = readSettings(workspace);
+    const saved = saveSettings(workspace, { ...settings, visionModelId: "custom-provider:vision" });
+    assert.equal(saved.visionModelId, "custom-provider:vision");
+    assert.equal(readSettings(workspace).visionModelId, "custom-provider:vision");
+    // Clearing the value removes the persisted key.
+    const cleared = saveSettings(workspace, { ...settings, visionModelId: undefined });
+    assert.equal(cleared.visionModelId, undefined);
+    assert.equal(readSettings(workspace).visionModelId, undefined);
+  } finally {
+    fs.rmSync(workspace, { recursive: true, force: true });
+  }
+});
+
+function multiModelConfig(overrides: Partial<SecAgentConfig> = {}): SecAgentConfig {
+  return {
+    version: 1,
+    workspace: "/tmp/secagent-test-ws",
+    agent: {
+      provider: "openai-compatible",
+      model: "main",
+      apiKeyEnv: "MAIN_KEY",
+      baseUrl: "https://main.test/v1",
+      endpoint: "/chat/completions",
+      maxTokens: 100,
+      systemPrompt: "main",
+      models: [
+        { id: "main", provider: "openai-compatible", model: "main", apiKeyEnv: "MAIN_KEY", baseUrl: "https://main.test/v1", endpoint: "/chat/completions", maxTokens: 100 },
+        { id: "vision", provider: "openai-compatible", model: "vision-model", apiKeyEnv: "VISION_KEY", baseUrl: "https://vision.test/v1", endpoint: "/chat/completions", maxTokens: 100 },
+        { id: "sectl-official:deepseek-v4-flash", provider: "openai-responses", model: "deepseek-v4-flash", apiKeyEnv: "SECTL_OFFICIAL_TOKEN", baseUrl: "https://relay.test/v1", endpoint: "/responses", maxTokens: 100 }
+      ]
+    },
+    mcp: { servers: {} },
+    ...overrides
+  } as SecAgentConfig;
+}
+
+test("resolveVisionAgentConfig honors an explicit vision model id without mutating the source config", () => {
+  const config = multiModelConfig({ defaults: { modelId: "main", visionModelId: "vision" } });
+  const vision = resolveVisionAgentConfig(config);
+  assert.ok(vision);
+  assert.equal(vision.agent.model, "vision-model");
+  assert.equal(vision.agent.apiKeyEnv, "VISION_KEY");
+  assert.equal(vision.agent.baseUrl, "https://vision.test/v1");
+  // The caller's config keeps its own model.
+  assert.equal(config.agent.model, "main");
+});
+
+test("resolveVisionAgentConfig returns undefined for a stale vision model id", () => {
+  const config = multiModelConfig({ defaults: { visionModelId: "does-not-exist" } });
+  assert.equal(resolveVisionAgentConfig(config), undefined);
+});
+
+test("resolveVisionAgentConfig falls back to the official virtual-vision model in official mode", () => {
+  const previous = process.env.SECTL_OFFICIAL_TOKEN;
+  process.env.SECTL_OFFICIAL_TOKEN = "test-token";
+  try {
+    const config = multiModelConfig({ defaults: { customModelMode: false } });
+    const vision = resolveVisionAgentConfig(config);
+    assert.ok(vision);
+    assert.equal(vision.agent.model, OFFICIAL_VISION_MODEL);
+    assert.equal(vision.agent.apiKeyEnv, "SECTL_OFFICIAL_TOKEN");
+    assert.equal(vision.agent.endpoint, "/responses");
+  } finally {
+    if (previous === undefined) delete process.env.SECTL_OFFICIAL_TOKEN;
+    else process.env.SECTL_OFFICIAL_TOKEN = previous;
+  }
+});
+
+test("resolveVisionAgentConfig does not fall back in custom model mode", () => {
+  const previous = process.env.SECTL_OFFICIAL_TOKEN;
+  process.env.SECTL_OFFICIAL_TOKEN = "test-token";
+  try {
+    const config = multiModelConfig({ defaults: { customModelMode: true } });
+    assert.equal(resolveVisionAgentConfig(config), undefined);
+  } finally {
+    if (previous === undefined) delete process.env.SECTL_OFFICIAL_TOKEN;
+    else process.env.SECTL_OFFICIAL_TOKEN = previous;
+  }
+});

src/config.ts

@@ -9,6 +9,13 @@ import { SYSTEM_PROMPT } from "./system-prompt.js";
 
 export const DEFAULT_GOOGLE_MODEL = "gemini-2.5-flash";
 export const DEFAULT_MAX_TOKENS = 16_384;
+/**
+ * Client-facing virtual vision model served by the official relay. The relay routes this
+ * model id to a vision-capable upstream, so the client does not need to know which concrete
+ * model is behind it. Only the client part lives in this repository; the relay contract is
+ * documented in the README/settings help text.
+ */
+export const OFFICIAL_VISION_MODEL = "virtual-vision";
 const ONBOARDING_MARKER = ".oobe-complete";
 const OOBE_PROGRESS_FILE = ".oobe-progress.json";
 const LEGACY_AGENT_MODEL_FIELDS = ["provider", "model", "apiKeyEnv", "baseUrl", "endpoint", "anthropicVersion", "maxTokens"] as const;
@@ -303,6 +310,41 @@ export function useConfiguredModel(config: SecAgentConfig, id?: string): void {
   config.agent = { ...config.agent, ...selected, model: dynamicModel || selectedModels[profileIndex] || selectedModels[0] || DEFAULT_GOOGLE_MODEL, maxTokens: selected.maxTokens || config.agent.maxTokens, systemPrompt: config.agent.systemPrompt, models: config.agent.models };
 }
 
+/**
+ * Return a copy of the config with `agent` resolved to the given model id, without
+ * mutating the caller's config. Unlike `useConfiguredModel`, this is safe to call for
+ * a secondary (e.g. vision) model while the main session keeps using its own model.
+ */
+export function resolveModelConfig(config: SecAgentConfig, modelId: string): SecAgentConfig {
+  const next = { ...config, agent: { ...config.agent } };
+  useConfiguredModel(next, modelId);
+  if (!next.agent.models?.length) throw new Error(`未找到配置模型：${modelId}`);
+  return next;
+}
+
+/**
+ * Resolve the dedicated image-recognition model config, if any.
+ * 1. An explicitly configured `defaults.visionModelId` wins.
+ * 2. In official (non-custom) mode, fall back to the relay's virtual-vision model so
+ *    the feature works without any manual setup.
+ * Returns undefined when no vision model is configured or the id is stale — the vision
+ * tool is then simply not exposed to the main agent.
+ */
+export function resolveVisionAgentConfig(config: SecAgentConfig): SecAgentConfig | undefined {
+  const id = config.defaults?.visionModelId;
+  if (id) {
+    try { return resolveModelConfig(config, id); }
+    catch { return undefined; }
+  }
+  if (config.defaults?.customModelMode === false
+      && process.env.SECTL_OFFICIAL_TOKEN
+      && config.agent.models?.some((model) => model.id.startsWith("sectl-official:"))) {
+    try { return resolveModelConfig(config, `official:sectl-official:${OFFICIAL_VISION_MODEL}`); }
+    catch { return undefined; }
+  }
+  return undefined;
+}
+
 export interface SettingsPayload {
   providers: Array<ProviderConfig & { apiKey?: string; apiKeyConfigured?: boolean }>;
   /** Compatibility field for older IPC callers; the settings UI uses providers. */
@@ -315,6 +357,8 @@ export interface SettingsPayload {
   mcp: { servers: Record<string, McpServerConfig> };
   defaultModelId?: string;
   defaultReasoningEffort?: ReasoningEffort;
+  /** Model used by the dedicated image-recognition tool when the main model has no vision input. */
+  visionModelId?: string;
   autostart?: boolean;
   /** On by default: an autostart launch stays in the tray instead of opening the main window. */
   autostartHidden?: boolean;
@@ -338,7 +382,7 @@ export function readSettings(workspaceInput: string): SettingsPayload {
       maxTokens: config.agent.maxTokens
     }];
   const providers = config.agent.providers?.length ? config.agent.providers : groupLegacyModels(configured);
-  return { providers: providers.map((provider) => ({ ...provider, apiKeyConfigured: Boolean(process.env[provider.apiKeyEnv]) })), models: configured.map((model) => ({ ...model, apiKeyConfigured: Boolean(process.env[model.apiKeyEnv]) })), tts: { voice: config.tts?.voice || DEFAULT_TTS_VOICE, rate: config.tts?.rate || DEFAULT_TTS_RATE }, wake: { hotkey: config.wake?.hotkey || DEFAULT_WAKE_HOTKEY, ...(config.wake?.modelId ? { modelId: config.wake.modelId } : {}), voiceEnabled: config.wake?.voiceEnabled === true, voicePhrase: config.wake?.voicePhrase || DEFAULT_WAKE_PHRASE }, speech: { betterRecognition: config.speech?.betterRecognition === true }, updates: { ...(config.updates || DEFAULT_UPDATE_PREFERENCES) }, telemetry: { enabled: config.telemetry?.enabled !== false }, mcp: config.mcp, defaultModelId: config.defaults?.modelId, defaultReasoningEffort: config.defaults?.reasoningEffort, autostart: config.defaults?.autostart === true, autostartHidden: config.defaults?.autostartHidden !== false, customModelMode: config.defaults?.customModelMode ?? false };
+  return { providers: providers.map((provider) => ({ ...provider, apiKeyConfigured: Boolean(process.env[provider.apiKeyEnv]) })), models: configured.map((model) => ({ ...model, apiKeyConfigured: Boolean(process.env[model.apiKeyEnv]) })), tts: { voice: config.tts?.voice || DEFAULT_TTS_VOICE, rate: config.tts?.rate || DEFAULT_TTS_RATE }, wake: { hotkey: config.wake?.hotkey || DEFAULT_WAKE_HOTKEY, ...(config.wake?.modelId ? { modelId: config.wake.modelId } : {}), voiceEnabled: config.wake?.voiceEnabled === true, voicePhrase: config.wake?.voicePhrase || DEFAULT_WAKE_PHRASE }, speech: { betterRecognition: config.speech?.betterRecognition === true }, updates: { ...(config.updates || DEFAULT_UPDATE_PREFERENCES) }, telemetry: { enabled: config.telemetry?.enabled !== false }, mcp: config.mcp, defaultModelId: config.defaults?.modelId, defaultReasoningEffort: config.defaults?.reasoningEffort, visionModelId: config.defaults?.visionModelId, autostart: config.defaults?.autostart === true, autostartHidden: config.defaults?.autostartHidden !== false, customModelMode: config.defaults?.customModelMode ?? false };
 }
 
 function groupLegacyModels(models: ModelProfile[]): ProviderConfig[] {
@@ -388,7 +432,7 @@ export function saveSettings(workspaceInput: string, payload: SettingsPayload):
   raw.updates = nextUpdates;
   raw.telemetry = nextTelemetry;
   raw.mcp = payload.mcp;
-  raw.defaults = { modelId: payload.defaultModelId || undefined, reasoningEffort: payload.defaultReasoningEffort || undefined, customModelMode: Boolean(payload.customModelMode), autostart: payload.autostart === true, autostartHidden: payload.autostartHidden !== false };
+  raw.defaults = { modelId: payload.defaultModelId || undefined, reasoningEffort: payload.defaultReasoningEffort || undefined, customModelMode: Boolean(payload.customModelMode), autostart: payload.autostart === true, autostartHidden: payload.autostartHidden !== false, visionModelId: payload.visionModelId || undefined };
   delete (raw as SecAgentConfig & { policy?: unknown }).policy;
   fs.writeFileSync(file, YAML.stringify(raw), "utf8");
   return readSettings(workspace);

src/electron/main.ts

@@ -10,7 +10,7 @@ import os from "node:os";
 import path from "node:path";
 import { pathToFileURL } from "node:url";
 import { DEFAULT_WORKSPACE } from "../paths.js";
-import { configuredModels, configPath, DEFAULT_TELEMETRY_SETTINGS, initializeWorkspace, isOnboardingComplete, loadConfig, markOnboardingComplete, readOobeProgress, readSettings, saveOobeProgress, saveSettings, useConfiguredModel, writeWorkspaceEnv, type OobeProgress, type SettingsPayload } from "../config.js";
+import { configuredModels, configPath, DEFAULT_TELEMETRY_SETTINGS, initializeWorkspace, isOnboardingComplete, loadConfig, markOnboardingComplete, OFFICIAL_VISION_MODEL, readOobeProgress, readSettings, resolveVisionAgentConfig, saveOobeProgress, saveSettings, useConfiguredModel, writeWorkspaceEnv, type OobeProgress, type SettingsPayload } from "../config.js";
 import { loadEnabledSkills } from "../skills.js";
 import { AuditStore } from "../audit.js";
 import { SecAgentRuntime, type TraceEvent } from "../runtime.js";
@@ -861,8 +861,11 @@ ipcMain.handle("models:list", async () => {
       // 自定义模型模式开启：只加入后台允许的官方真实模型与本地自定义模型。
       return [...visibleRemote, ...options];
     }
-    // 关闭：官方档位模式 —— 下拉只有快速/标准/深度三个虚拟档位，看不到具体模型。
-    return visibleRemote.filter((model) => (OFFICIAL_TIER_IDS as readonly string[]).includes(model.model));
+    // 关闭：官方档位模式 —— 下拉只有快速/标准/深度三个虚拟档位，看不到具体模型；
+    // 另外提供一个识图虚拟模型（virtual-vision），它只作为识图工具的后端模型，
+    // 不作为主 Agent 模型出现在前端下拉中（前端按 vision 标记过滤）。
+    return visibleRemote.filter((model) => (OFFICIAL_TIER_IDS as readonly string[]).includes(model.model) || model.model === OFFICIAL_VISION_MODEL)
+      .map((model) => ({ ...model, vision: model.model === OFFICIAL_VISION_MODEL }));
   } catch { return customModelMode ? options : []; }
 });
 ipcMain.handle("providers:list", async () => {
@@ -1526,10 +1529,12 @@ ipcMain.handle("sessions:send", async (_event, id: string, text: string, modelId
     logMain("ipc.sessions.send", { sessionId: id, text });
     trace({ stage: "user.request", data: { text } });
     const skills = [...loadEnabledSkills(config), ...(pluginManager?.getSkills() || [])];
+    const visionConfig = resolveVisionAgentConfig(config);
+    if (visionConfig) logMain("session.vision-model", { model: visionConfig.agent.model, provider: visionConfig.agent.provider, baseUrl: visionConfig.agent.baseUrl });
     const runtimeConfig = isWakeRequest
       ? { ...config, agent: { ...config.agent, systemPrompt: `${config.agent.systemPrompt}\n\n## 快速唤起输出协议\n${QUICK_WAKE_OUTPUT_PROMPT}` } }
       : config;
-    runtime = new SecAgentRuntime(runtimeConfig, audit, skills, trace, pluginManager);
+    runtime = new SecAgentRuntime(runtimeConfig, audit, skills, trace, pluginManager, visionConfig);
     const previousReadSkillNames = before.messages.flatMap((message) => message.toolCalls || []).filter((call) => call.name === "secagent__read_skill" || call.name === "read_skill").map((call) => typeof (call.arguments as { name?: unknown })?.name === "string" ? (call.arguments as { name: string }).name : "");
     const result = await runtime.run(historyInput(before, text), selectedReasoningEffort, conversationInput(before, text, attachments), abortController.signal, { previousAutoLoadedSkills: before.autoLoadedSkills, previousReadSkillNames, preRule });
     if (result.autoLoadedSkills?.length) {

src/index.ts

@@ -1,7 +1,7 @@
 #!/usr/bin/env node
 import readline from "node:readline/promises";
 import { stdin as input, stdout as output } from "node:process";
-import { initializeWorkspace, loadConfig, normalizeAndValidate, useConfiguredModel } from "./config.js";
+import { initializeWorkspace, loadConfig, normalizeAndValidate, resolveVisionAgentConfig, useConfiguredModel } from "./config.js";
 import { DEFAULT_WORKSPACE, expandPath } from "./paths.js";
 import { loadEnabledSkills } from "./skills.js";
 import { AuditStore } from "./audit.js";
@@ -160,7 +160,8 @@ async function openRuntime(workspace: string, modelId: string | undefined, trace
   const plugins = new PluginManager(workspace);
   await plugins.initialize();
   const skills = [...loadEnabledSkills(config), ...plugins.getSkills()];
-  return { runtime: new SecAgentRuntime(config, audit, skills, trace, plugins), audit, plugins, config };
+  const visionConfig = resolveVisionAgentConfig(config);
+  return { runtime: new SecAgentRuntime(config, audit, skills, trace, plugins, visionConfig), audit, plugins, config };
 }
 
 async function closeRuntime(handle: RuntimeHandle | undefined): Promise<void> {

src/model-provider.test.ts

@@ -224,3 +224,36 @@ test("persisted tool calls and results are restored to the next OpenAI request",
     delete process.env.TEST_MODEL_KEY;
   }
 });
+
+test("a single-turn sub-agent encodes user attachments as OpenAI image parts", async () => {
+  const originalFetch = globalThis.fetch;
+  process.env.TEST_MODEL_KEY = "test-key";
+  let requestBody: Record<string, unknown> | undefined;
+  globalThis.fetch = async (_url, init) => {
+    requestBody = JSON.parse(String(init?.body || "{}")) as Record<string, unknown>;
+    return response('data: {"choices":[{"delta":{"content":"red"}}]}\n\ndata: [DONE]\n\n');
+  };
+  try {
+    // Tool-less vision sub-agent: allowEmptyTools and no runtime prompts.
+    const agent = new ModelToolAgent(config(), [], undefined, undefined, false, true);
+    const conversation: ConversationMessage[] = [
+      {
+        role: "user",
+        content: "图中是什么颜色",
+        attachments: [{ id: "a1", name: "test.png", mimeType: "image/png", dataUrl: "data:image/png;base64,iVBORw0KGgo=", size: 6 }]
+      }
+    ];
+    const result = await agent.run("图中是什么颜色", [], async () => {
+      throw new Error("识图模型不应调用工具");
+    }, "low", conversation);
+    assert.equal(result, "red");
+    const messages = requestBody?.messages as Array<Record<string, unknown>>;
+    const content = messages[1]?.content as Array<Record<string, unknown>>;
+    assert.ok(Array.isArray(content));
+    assert.deepEqual(content[0], { type: "text", text: "图中是什么颜色" });
+    assert.deepEqual(content[1], { type: "image_url", image_url: { url: "data:image/png;base64,iVBORw0KGgo=" } });
+  } finally {
+    globalThis.fetch = originalFetch;
+    delete process.env.TEST_MODEL_KEY;
+  }
+});