picturereader 2.0.1 → 3.0.0

package/src/vlm.js

@@ -10,16 +10,153 @@
  */
 
 import { spawn } from 'node:child_process';
-import { stat } from 'node:fs/promises';
+import { stat, readFile } from 'node:fs/promises';
+import { join } from 'node:path';
+import { homedir } from 'node:os';
+import { getRuntimeConfig } from './runtime.js';
 
 // ---------------------------------------------------------------------------
-// Configuration (all via environment variables, defaults are empty/disabled)
+// Configuration (environment variables first, then DSH settings.yaml fallback)
 // ---------------------------------------------------------------------------
 
+/**
+ * Try to read DSH settings.yaml and extract tool-vision config.
+ * @returns {Promise<{baseURL: string, model: string, apiKey: string}|null>}
+ */
+async function readDshToolVisionConfig() {
+  try {
+    const settingsPath = join(homedir(), '.dsh', 'settings.yaml');
+    const content = await readFile(settingsPath, 'utf-8');
+    // Simple YAML parsing for tool-vision section
+    const lines = content.split('\n');
+    let inToolVision = false;
+    let baseURL = '';
+    let model = '';
+    let apiKey = '';
+    let indent = -1;
+
+    for (const line of lines) {
+      // Detect start of tool-vision section
+      if (line.match(/^tool-vision:\s*$/)) {
+        inToolVision = true;
+        indent = -1;
+        continue;
+      }
+
+      if (inToolVision) {
+        // Check if we've left the section (non-empty line with less or equal indent)
+        const match = line.match(/^(\s*)\S/);
+        if (match) {
+          const currentIndent = match[1].length;
+          if (indent === -1) {
+            indent = currentIndent;
+          } else if (currentIndent <= 0) {
+            // New top-level key, exit tool-vision section
+            break;
+          }
+        }
+
+        // Parse key-value pairs
+        const kvMatch = line.match(/^\s+(\w+):\s*(.+)$/);
+        if (kvMatch) {
+          const [, key, value] = kvMatch;
+          const cleanValue = value.replace(/^["']|["']$/g, '').trim();
+          if (key === 'baseURL') baseURL = cleanValue;
+          else if (key === 'model') model = cleanValue;
+          else if (key === 'apiKey') apiKey = cleanValue;
+        }
+      }
+    }
+
+    if (baseURL) {
+      return { baseURL, model, apiKey };
+    }
+  } catch {
+    // Settings file not found or parse error, ignore
+  }
+  return null;
+}
+
+/**
+ * Read DSH settings.yaml and extract the `picturereader` namespace's VLM
+ * fields (vlm_base / vlm_model / vlm_key). These come from the plugin's own
+ * settings card and take priority over the legacy tool-vision namespace,
+ * so the user's configured endpoint is honoured even when the plugin packs
+ * a newer schema.
+ * @returns {Promise<{baseURL: string, model: string, apiKey: string}|null>}
+ */
+async function readDshPicturereaderConfig() {
+  try {
+    const settingsPath = join(homedir(), '.dsh', 'settings.yaml');
+    const content = await readFile(settingsPath, 'utf-8');
+    const lines = content.split('\n');
+    let inNs = false;
+    let baseURL = '';
+    let model = '';
+    let apiKey = '';
+    let indent = -1;
+    const KEY_MAP = { vlm_base: 'baseURL', vlm_model: 'model', vlm_key: 'apiKey' };
+
+    for (const line of lines) {
+      if (line.match(/^picturereader:\s*$/)) {
+        inNs = true;
+        indent = -1;
+        continue;
+      }
+      if (inNs) {
+        const match = line.match(/^(\s*)\S/);
+        if (match) {
+          const currentIndent = match[1].length;
+          if (indent === -1) {
+            indent = currentIndent;
+          } else if (currentIndent <= 0) {
+            break;
+          }
+        }
+        const kvMatch = line.match(/^\s+(\w+):\s*(.+)$/);
+        if (kvMatch) {
+          const [, key, value] = kvMatch;
+          const target = KEY_MAP[key];
+          const cleanValue = value.replace(/^["']|["']$/g, '').trim();
+          if (target === 'baseURL') baseURL = cleanValue;
+          else if (target === 'model') model = cleanValue;
+          else if (target === 'apiKey') apiKey = cleanValue;
+        }
+      }
+    }
+    if (baseURL) return { baseURL, model, apiKey };
+  } catch {
+    // ignore parse errors
+  }
+  return null;
+}
+
+// Read config: priority env → picturereader namespace → legacy tool-vision namespace
+const dshConfig = await readDshToolVisionConfig();
+const dshPictConfig = await readDshPicturereaderConfig();
+
+// ---------------------------------------------------------------------------
+// GLM-4V-Flash: free built-in vision model from Zhipu AI
+// https://open.bigmodel.cn — register to get your API key
+// ---------------------------------------------------------------------------
+
+/** GLM-4V-Flash default endpoint (OpenAI-compatible). */
+const GLM4V_BASE = 'https://open.bigmodel.cn/api/paas/v4';
+/** GLM-4V-Flash model name. */
+const GLM4V_MODEL = 'glm-4v-flash';
+
+/**
+ * Resolve the GLM-4V-Flash API key.
+ * Priority: env SEE_API_KEY > env GLM_API_KEY > DSH settings.yaml apiKey > empty.
+ */
+function resolveGlm4vKey() {
+  return process.env.SEE_API_KEY ?? process.env.GLM_API_KEY ?? dshConfig?.apiKey ?? '';
+}
+
 /** OpenAI-compatible VLM endpoint (empty = VLM disabled). */
-export const DEFAULT_BASE = process.env.SEE_BASE ?? '';
+export const DEFAULT_BASE = process.env.SEE_BASE ?? dshPictConfig?.baseURL ?? dshConfig?.baseURL ?? GLM4V_BASE;
 /** VLM model name. */
-export const DEFAULT_MODEL = process.env.SEE_MODEL ?? '';
+export const DEFAULT_MODEL = process.env.SEE_MODEL ?? dshPictConfig?.model ?? dshConfig?.model ?? GLM4V_MODEL;
 /** Local llama-server executable path. */
 export const DEFAULT_SERVER_EXE = process.env.SEE_SERVER_EXE ?? '';
 /** Local model GGUF path. */
@@ -32,18 +169,53 @@ export const DEFAULT_PORT = Number(process.env.SEE_SERVER_PORT ?? 8080);
 export const DEFAULT_NGL = process.env.SEE_SERVER_NGL ?? '20';
 /** Context size for local server. */
 export const DEFAULT_CTX = Number(process.env.SEE_SERVER_CTX ?? 16384);
-/** API key for remote endpoints. */
-export const DEFAULT_API_KEY = process.env.SEE_API_KEY ?? '';
+/** API key for remote endpoints (env > picturereader ns > legacy tool-vision ns > GLM). */
+export const DEFAULT_API_KEY =
+  process.env.SEE_API_KEY ?? dshPictConfig?.apiKey ?? dshConfig?.apiKey ?? process.env.GLM_API_KEY ?? '';
 
 let serverStartPromise = null;
 let serverChild = null;
 
 /**
- * Check if VLM is configured (has a base URL).
- * @returns {boolean} true when VLM endpoint is configured.
+ * Effective API key: runtime explicit key → runtime env var → static default.
+ * @param {object} rt - runtime config (may be undefined).
+ * @returns {string}
+ */
+function effectiveApiKey(rt) {
+  if (rt?.vlm?.apiKey) return rt.vlm.apiKey;
+  const envName = rt?.vlm?.apiKeyEnv;
+  if (envName) {
+    const fromEnv = process.env[envName];
+    if (fromEnv) return fromEnv;
+  }
+  return DEFAULT_API_KEY;
+}
+
+/**
+ * Effective endpoint base URL: runtime explicit → static default.
+ * @param {object} rt
+ * @returns {string}
+ */
+function effectiveBase(rt) {
+  return (rt?.vlm?.baseUrl || '').trim() || DEFAULT_BASE;
+}
+
+/**
+ * Check if VLM is configured (has a base URL and API key for cloud endpoints).
+ * 隐私模式（privacy）为硬门禁：即使配置了外部 API 也返回 false，绝不外呼。
+ * @returns {boolean} true when VLM endpoint is configured and ready to use.
  */
 export function isVlmConfigured() {
-  return DEFAULT_BASE.length > 0;
+  const rt = getRuntimeConfig();
+  if (rt?.mode === 'privacy') return false;
+  // 选配（vlm_enabled）未勾选：即使配置了端点/Key 也视为未启用外部 VLM。
+  if (rt?.vlm?.enabled === false) return false;
+  const base = effectiveBase(rt);
+  if (base.length === 0) return false;
+  // Cloud endpoints require an API key
+  const isLocal = isManagedEndpoint(base, DEFAULT_PORT);
+  if (!isLocal && effectiveApiKey(rt).length === 0) return false;
+  return true;
 }
 
 /**
@@ -224,7 +396,17 @@ export async function sendVisionRequest(config, images, prompt) {
     headers.authorization = `Bearer ${config.apiKey}`;
   }
 
-  const res = await fetch(`${config.baseURL}/chat/completions`, {
+  const base = String(config.baseURL || '').trim().replace(/\/+$/, '');
+  let endpoint;
+  if (/\/v\d+\/chat\/completions$/.test(base) || /\/chat\/completions$/.test(base)) {
+    endpoint = base;
+  } else if (/\/v\d+$/.test(base)) {
+    endpoint = `${base}/chat/completions`;
+  } else {
+    // OpenAI 兼容端点统一用 /v1/chat/completions（LM Studio / llama-server / 云端网关等）
+    endpoint = `${base}/v1/chat/completions`;
+  }
+  const res = await fetch(endpoint, {
     method: 'POST',
     headers,
     body: JSON.stringify(body),
@@ -250,10 +432,11 @@ export async function sendVisionRequest(config, images, prompt) {
  * @returns {object} VLM configuration.
  */
 export function defaultVlmConfig(overrides = {}) {
+  const rt = getRuntimeConfig();
   return {
-    baseURL: DEFAULT_BASE,
-    apiKey: DEFAULT_API_KEY,
-    model: DEFAULT_MODEL,
+    baseURL: effectiveBase(rt),
+    apiKey: effectiveApiKey(rt),
+    model: (rt?.vlm?.model || '').trim() || DEFAULT_MODEL,
     serverExe: DEFAULT_SERVER_EXE,
     serverModel: DEFAULT_SERVER_MODEL,
     serverMmproj: DEFAULT_SERVER_MMPROJ,
@@ -262,8 +445,8 @@ export function defaultVlmConfig(overrides = {}) {
     ctxSize: DEFAULT_CTX,
     autoStart: true,
     healthTimeoutMs: 120_000,
-    requestTimeoutMs: 300_000,
-    maxTokens: 8192,
+    requestTimeoutMs: rt?.vlm?.requestTimeoutMs ?? 300_000,
+    maxTokens: rt?.vlm?.maxTokens ?? 8192,
     ...overrides,
   };
 }