@bolloon/bolloon-agent 0.2.13 → 0.2.15

package/dist/cli-entry.js

@@ -23,7 +23,7 @@ const YELLOW = '\x1b[33m';
 const GREEN = '\x1b[32m';
 const MAGENTA = '\x1b[35m';
 // 版本信息 — 与 package.json:version 同步, 否则 banner 会显示过时版本误导用户
-const VERSION = '0.2.13';
+const VERSION = '0.2.15';
 function log(msg, color = RESET) {
     console.log(`${color}${msg}${RESET}`);
 }

package/dist/llm/config-store.js

@@ -38,7 +38,7 @@ export const DEFAULT_PROVIDER_CONFIGS = {
         enabled: false,
         apiKey: '',
         baseUrl: 'https://generativelanguage.googleapis.com/v1beta',
-        model: 'gemini-3.5-pro',
+        model: 'gemini-2.5-pro',
         temperature: 0.7,
         maxTokens: 4096,
         requiresApiKey: true
@@ -121,7 +121,12 @@ export const PROVIDER_INFO = {
     openai: { name: 'OpenAI', description: 'GPT-4, GPT-3.5 等模型', requiresApiKey: true, models: ['gpt-4.1', 'gpt-4o', 'gpt-4o-mini', 'gpt-4-turbo', 'gpt-3.5-turbo'] },
     anthropic: { name: 'Anthropic', description: 'Claude 3.5+ 系列模型', requiresApiKey: true, models: ['claude-sonnet-4-5-20250929', 'claude-3-5-sonnet-20241022', 'claude-3-5-haiku-20241022', 'claude-3-opus-20240229'] },
     openrouter: { name: 'OpenRouter', description: '聚合多个 AI 供应商', requiresApiKey: true, models: ['anthropic/claude-sonnet-4.5', 'anthropic/claude-3.5-sonnet'] },
-    gemini: { name: 'Google Gemini', description: 'Gemini 系列模型', requiresApiKey: true, models: ['gemini-2.5-pro', 'gemini-2.0-flash', 'gemini-1.5-pro', 'gemini-1.5-flash'] },
+    gemini: { name: 'Google Gemini', description: 'Gemini 系列模型', requiresApiKey: true, models: [
+            'gemini-3.5-flash', // 2026-06 官方推荐 stable 旗舰 (https://ai.google.dev/gemini-api/docs/models)
+            'gemini-2.5-pro', // 高级推理, 仍为 GA
+            'gemini-3.1-flash-lite', // 成本敏感场景
+            'gemini-flash-latest', // 滚动 alias → 当前 stable Flash
+        ] },
     ollama: { name: 'Ollama', description: '本地 LLM 运行框架', requiresApiKey: false },
     minimax: {
         name: 'MiniMax',
@@ -209,6 +214,18 @@ function getDefaultConfig() {
 class LLMConfigStore {
     config = null;
     initialized = false;
+    // v0.2.15: single-flight lock around read-modify-write of `~/.bolloon/llm-config.json`.
+    // Prevents concurrent save() calls from clobbering each other when the user
+    // configures two providers back-to-back (e.g. saving gemini, then anthropic, in
+    // quick succession). One operation at a time, in call order.
+    writeChain = Promise.resolve();
+    async withWriteLock(fn) {
+        // Chain the new op after the previous one; swallow the previous op's
+        // rejection so a single failed save does not poison subsequent writes.
+        const next = this.writeChain.then(fn, fn);
+        this.writeChain = next.then(() => undefined, () => undefined);
+        return next;
+    }
     async initialize() {
         if (this.initialized)
             return;
@@ -270,19 +287,23 @@ class LLMConfigStore {
         if (providerConfig.requiresApiKey && !providerConfig.apiKey) {
             throw new Error(`${provider} requires an API key but none is configured`);
         }
-        this.config.activeProvider = provider;
-        await this.save();
+        await this.withWriteLock(async () => {
+            this.config.activeProvider = provider;
+            await this.save();
+        });
     }
     async updateProvider(provider, updates) {
         await this.initialize();
         if (!this.config?.providers[provider]) {
             throw new Error(`Unknown provider: ${provider}`);
         }
-        this.config.providers[provider] = {
-            ...this.config.providers[provider],
-            ...updates
-        };
-        await this.save();
+        await this.withWriteLock(async () => {
+            this.config.providers[provider] = {
+                ...this.config.providers[provider],
+                ...updates
+            };
+            await this.save();
+        });
     }
     async testProvider(provider) {
         await this.initialize();
@@ -295,10 +316,16 @@ class LLMConfigStore {
         }
         const startTime = Date.now();
         try {
-            const response = await fetch(`${config.baseUrl}/models`, {
-                method: 'GET',
-                headers: this.buildHeaders(provider, config)
-            });
+            // v0.2.15: per-provider test endpoint. The old unified `GET baseUrl/models`
+            // is wrong on at least two providers:
+            //   - Anthropic has no GET /v1/models endpoint -> always 404.
+            //   - Google's GET /v1beta/models returns the public model catalog
+            //     without auth -> always 200, even with a wrong/expired key, so
+            //     the user saw "connected" while chat requests still failed.
+            // The per-provider branch below uses an endpoint that actually
+            // gates on the key.
+            const { url, init } = this.buildTestRequest(provider, config);
+            const response = await fetch(url, init);
             const latency = Date.now() - startTime;
             if (response.ok) {
                 return { success: true, latency };
@@ -307,9 +334,13 @@ class LLMConfigStore {
                 const errorText = await response.text().catch(() => 'Unknown error');
                 const hint = response.status === 401
                     ? '（API Key 无效或不匹配该供应商 — 请检查是否复制完整、有无多余空格）'
-                    : response.status === 404
-                        ? '（端点不存在 — 请检查 baseUrl）'
-                        : '';
+                    : response.status === 403
+                        ? '（API Key 没有调用此端点的权限 — 请检查 key scope 或供应商 endpoint）'
+                        : response.status === 404
+                            ? '（端点不存在 — 请检查 baseUrl）'
+                            : response.status === 429
+                                ? '（供应商限流中 — 稍候再试）'
+                                : '';
                 return { success: false, error: `HTTP ${response.status}: ${errorText.substring(0, 500)}${hint ? ' ' + hint : ''}`, latency };
             }
         }
@@ -317,6 +348,59 @@ class LLMConfigStore {
             return { success: false, error: error.message || 'Connection failed', latency: Date.now() - startTime };
         }
     }
+    /**
+     * Build the lightest "is the key + baseUrl healthy?" probe for the given
+     * provider. Each branch targets an endpoint that *actually* validates the
+     * credentials (as opposed to a public catalog or non-existent route).
+     */
+    buildTestRequest(provider, config) {
+        switch (provider) {
+            case 'anthropic':
+                // No GET /v1/models. Use a minimal /messages ping that fails fast
+                // on bad keys (401) and rate-limits (429) without burning quota.
+                return {
+                    url: `${config.baseUrl}/messages`,
+                    init: {
+                        method: 'POST',
+                        headers: this.buildHeaders(provider, config),
+                        body: JSON.stringify({
+                            model: config.model || 'claude-sonnet-4-5-20250929',
+                            max_tokens: 1,
+                            messages: [{ role: 'user', content: 'ping' }],
+                        }),
+                    },
+                };
+            case 'gemini':
+                // listModels with the key in the query string returns 400 for
+                // an invalid key, 200 for a valid one. This is the only "light"
+                // Gemini endpoint that gates on auth (generateContent would
+                // burn quota on a real prompt).
+                return {
+                    url: `${config.baseUrl}/models?key=${encodeURIComponent(config.apiKey)}`,
+                    init: { method: 'GET' },
+                };
+            case 'ollama':
+                return { url: `${config.baseUrl}/api/tags`, init: { method: 'GET' } };
+            case 'openai':
+            case 'openrouter':
+            case 'deepseek':
+            case 'kimi':
+            case 'glm':
+            case 'qwen':
+            case 'mimo':
+            case 'minimax':
+            case 'local':
+                return {
+                    url: `${config.baseUrl}/models`,
+                    init: { method: 'GET', headers: this.buildHeaders(provider, config) },
+                };
+            default:
+                return {
+                    url: `${config.baseUrl}/models`,
+                    init: { method: 'GET', headers: this.buildHeaders(provider, config) },
+                };
+        }
+    }
     buildHeaders(provider, config) {
         const headers = { 'Content-Type': 'application/json' };
         switch (provider) {

package/dist/llm/pi-ai.js

@@ -255,6 +255,8 @@ export class PiAIModel {
             anthropic: this.config.model || 'claude-sonnet-4-5-20250929',
             ollama: this.config.model || 'llama3.2',
             openrouter: this.config.model || 'anthropic/claude-sonnet-4.5',
+            // Pinned to 2.5-pro: the only `-pro` model that is GA per Google docs.
+            // The 3.x line ships as `-flash` only — there is no `gemini-3.x-pro`.
             gemini: this.config.model || 'gemini-2.5-pro',
             minimax: this.config.model || process.env.MINIMAX_MODEL || 'MiniMax-M3',
             deepseek: this.config.model || process.env.DEEPSEEK_MODEL || 'deepseek-chat',

package/dist/web/api-config.html

@@ -421,7 +421,15 @@
       const result = document.getElementById('testResult');
 
       btn.disabled = true;
-      btn.innerHTML = '<span class="spinner"></span> 测试中...';
+      // v0.2.15: defer label mutation to the next frame. Synchronously
+      // replacing the button's innerHTML in its own click handler cancels
+      // the click event in several Chromium-derived engines — the user
+      // observed the first click "doing nothing" until they closed and
+      // reopened the modal. Wrapping the mutation in requestAnimationFrame
+      // lets the click event finish dispatching first.
+      requestAnimationFrame(() => {
+        btn.textContent = '⚡ 测试中...';
+      });
       result.style.display = 'none';
 
       const endpoint = currentProviderType === 'llm' ? '/api/llm-test' :
@@ -447,7 +455,9 @@
       }
 
       btn.disabled = false;
-      btn.innerHTML = '⚡ 测试连接';
+      requestAnimationFrame(() => {
+        btn.textContent = '⚡ 测试连接';
+      });
     }
 
     // ==================== 保存 ====================

package/dist/web/routes-llm-config.js

@@ -43,20 +43,30 @@ export function registerLlmConfigRoutes(app) {
                 config.apiKey = currentConfig.apiKey;
             }
             await llmConfigStore.updateProvider(provider, config);
-            // 如果是活跃供应商，重新初始化 Pi SDK
-            const currentActive = await llmConfigStore.getActiveProvider();
-            if (provider === currentActive) {
-                const newConfig = await llmConfigStore.getActiveProviderConfig();
-                if (newConfig) {
-                    initMinimax({
-                        provider,
-                        apiKey: newConfig.apiKey || undefined,
-                        baseUrl: newConfig.baseUrl || undefined,
-                        model: newConfig.model || undefined
-                    });
-                }
+            // v0.2.15: 当用户保存一个 LLM 配置为 enabled + 有 key 时，自动把它切为
+            // activeProvider 并 rebind runtime singleton。修原来的两个真问题：
+            //   1) frontend Save 从不调 /api/llm-provider，于是 activeProvider 一直
+            //      卡在 process 启动时的第一个值，新配置的 provider 永远不接管 chat
+            //   2) 即使用户手动 active，updateProvider 不替换 modelInstance（只有
+            //      当前 active 命中才 rebind），所以保存非 active 那个 provider 之后
+            //      runtime 还是指向旧 provider + 旧 key
+            // 现在：用户每保存一个 enabled 的 LLM 配置，bolloon 立即把这个 provider
+            // 设成 active + 重新 init MinLLM/Pi SDK，让"配置 + 立刻能跑"成立。
+            // 用户想保持旧 active，可以显式调 /api/llm-provider 改回去。
+            const newConfig = await llmConfigStore.getProvider(provider);
+            const shouldAutoActivate = newConfig?.enabled === true &&
+                // 如果 provider 不需要 key (如 ollama) 或者已经给了真 key，才激活
+                (newConfig.apiKey || !newConfig.requiresApiKey);
+            if (shouldAutoActivate) {
+                await llmConfigStore.setActiveProvider(provider);
+                initMinimax({
+                    provider: provider,
+                    apiKey: newConfig.apiKey || undefined,
+                    baseUrl: newConfig.baseUrl || undefined,
+                    model: newConfig.model || undefined
+                });
             }
-            res.json({ ok: true });
+            res.json({ ok: true, autoActivated: shouldAutoActivate });
         }
         catch (err) {
             res.status(500).json({ error: err.message });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bolloon/bolloon-agent",
-  "version": "0.2.13",
+  "version": "0.2.15",
   "type": "module",
   "description": "P2P AI Document Agent - 全局安装后执行 `bolloon` 启动产品",
   "main": "dist/cli-entry.js",
@@ -48,7 +48,7 @@
     "src/constraint-runtime"
   ],
   "dependencies": {
-    "@bolloon/bolloon-agent": "^0.2.13",
+    "@bolloon/bolloon-agent": "^0.2.15",
     "@bolloon/constraint-runtime": "0.1.0",
     "@capacitor/core": "^8.4.1",
     "@capacitor/ios": "^8.4.1",