feat: 回退的妥协的tool_use呼叫，合并到系统提示词

Mag1cFall · Mag1cFall · commit e64caa17c8f8 · 2025-12-29T13:07:50.000+08:00
diff --git a/src/api/request_processor.py b/src/api/request_processor.py
@@ -17,6 +17,33 @@
 from .abort_detector import AbortSignalDetector, AbortSignalHandler
 from browser.page_controller import PageController
 
+def _merge_tools_to_system_prompt(system_prompt: str, tools: Optional[list], logger, req_id: str) -> str:
+    if not tools:
+        return system_prompt
+    function_declarations = []
+    for tool in tools:
+        if not isinstance(tool, dict):
+            continue
+        if tool.get('google_search_retrieval') is not None:
+            continue
+        if tool.get('function', {}).get('name') == 'googleSearch':
+            continue
+        if 'function' in tool:
+            func = tool['function']
+            declaration = {
+                'name': func.get('name', ''),
+                'description': func.get('description', ''),
+            }
+            if 'parameters' in func:
+                declaration['parameters'] = func['parameters']
+            function_declarations.append(declaration)
+    if not function_declarations:
+        return system_prompt
+    tools_json = json.dumps(function_declarations, indent=2, ensure_ascii=False)
+    logger.info(f"[{req_id}] 🔧 合并 {len(function_declarations)} 个函数到系统提示词")
+    tools_section = f"<tools>\n{tools_json}\n</tools>\n\n"
+    return tools_section + system_prompt
+
 async def _initialize_request_context(req_id: str, request: ChatCompletionRequest) -> dict:
     from server import logger, page_instance, is_page_ready, parsed_model_list, current_ai_studio_model_id, model_switching_lock, page_params_cache, params_cache_lock
     request_manager.register_request(req_id, {'model': request.model, 'stream': request.stream, 'message_count': len(request.messages)})
@@ -670,7 +697,10 @@ async def _process_request_refactored(req_id: str, request: ChatCompletionReques
 
         await _handle_parameter_cache(req_id, context)
         await page_controller.adjust_parameters(request.model_dump(exclude_none=True), context['page_params_cache'], context['params_cache_lock'], context['model_id_to_use'], context['parsed_model_list'], check_client_disconnected)
-        await page_controller.set_system_instructions(system_prompt, check_client_disconnected)
+        
+        # 合并tools到system prompt
+        final_system_prompt = _merge_tools_to_system_prompt(system_prompt, request.tools, context['logger'], req_id)
+        await page_controller.set_system_instructions(final_system_prompt, check_client_disconnected)
         check_client_disconnected('提交提示前最终检查')
         await page_controller.submit_prompt(prepared_prompt, image_list, check_client_disconnected)
         response_result = await _handle_response_processing(req_id, request, page, context, result_future, submit_button_locator, check_client_disconnected, disconnect_check_task)
diff --git a/src/browser/page_controller.py b/src/browser/page_controller.py
@@ -5,6 +5,7 @@
 import re
 import os
 from playwright.async_api import Page as AsyncPage, expect as expect_async, TimeoutError, Locator
+import json as json_module
 from config import TEMPERATURE_INPUT_SELECTOR, MAX_OUTPUT_TOKENS_SELECTOR, STOP_SEQUENCE_INPUT_SELECTOR, MAT_CHIP_REMOVE_BUTTON_SELECTOR, TOP_P_INPUT_SELECTOR, SUBMIT_BUTTON_SELECTOR, SUBMIT_BUTTON_SELECTORS, OVERLAY_SELECTOR, PROMPT_TEXTAREA_SELECTOR, PROMPT_TEXTAREA_SELECTORS, RESPONSE_CONTAINER_SELECTOR, RESPONSE_TEXT_SELECTOR, EDIT_MESSAGE_BUTTON_SELECTOR, USE_URL_CONTEXT_SELECTOR, UPLOAD_BUTTON_SELECTOR, UPLOAD_BUTTON_SELECTORS, INSERT_BUTTON_SELECTOR, INSERT_BUTTON_SELECTORS, HIDDEN_FILE_INPUT_SELECTORS, THINKING_MODE_TOGGLE_SELECTOR, SET_THINKING_BUDGET_TOGGLE_SELECTOR, THINKING_BUDGET_INPUT_SELECTOR, GROUNDING_WITH_GOOGLE_SEARCH_TOGGLE_SELECTOR, ZERO_STATE_SELECTOR, SYSTEM_INSTRUCTIONS_BUTTON_SELECTOR, SYSTEM_INSTRUCTIONS_TEXTAREA_SELECTOR, SKIP_PREFERENCE_VOTE_BUTTON_SELECTOR, CLICK_TIMEOUT_MS, WAIT_FOR_ELEMENT_TIMEOUT_MS, CLEAR_CHAT_VERIFY_TIMEOUT_MS, DEFAULT_TEMPERATURE, DEFAULT_MAX_OUTPUT_TOKENS, DEFAULT_STOP_SEQUENCES, DEFAULT_TOP_P, ENABLE_URL_CONTEXT, ENABLE_THINKING_BUDGET, DEFAULT_THINKING_BUDGET, ENABLE_GOOGLE_SEARCH, THINKING_LEVEL_SELECT_SELECTOR, THINKING_LEVEL_OPTIONS, DEFAULT_THINKING_LEVEL, ADVANCED_SETTINGS_EXPANDER_SELECTOR
 from config.timeouts import (
     MAX_RETRIES, SLEEP_RETRY, SLEEP_SHORT, SLEEP_MEDIUM, SLEEP_LONG, SLEEP_TICK,
@@ -71,6 +72,7 @@ async def handle_tools_panel():
                 await self._open_url_content(check_client_disconnected)
             else:
                 self.logger.info(f'[{self.req_id}] URL Context 功能已禁用，跳过调整。')
+            # NOTE: Function Calling 改为合并到 system prompt，不再通过 UI 填写
 
         tasks = [
             self._adjust_temperature(temp_to_set, page_params_cache, params_cache_lock, check_client_disconnected),
@@ -87,16 +89,20 @@ async def handle_tools_panel():
     async def set_system_instructions(self, system_prompt: str, check_client_disconnected: Callable):
         if not system_prompt:
             return
-        self.logger.info(f'[{self.req_id}] 正在设置系统指令...')
+        self.logger.info(f'[{self.req_id}] 正在设置系统指令 (长度: {len(system_prompt)} chars)...')
         await self._check_disconnect(check_client_disconnected, 'Start System Instructions')
         try:
             sys_prompt_button = self.page.locator(SYSTEM_INSTRUCTIONS_BUTTON_SELECTOR)
             sys_prompt_textarea = self.page.locator(SYSTEM_INSTRUCTIONS_TEXTAREA_SELECTOR)
             await self._click_and_verify(sys_prompt_button, sys_prompt_textarea, 'System Instructions Button', 'System Instructions Textarea')
-            await expect_async(sys_prompt_textarea).to_be_editable(timeout=5000)
-            await sys_prompt_textarea.fill(system_prompt, timeout=5000)
-            await expect_async(sys_prompt_textarea).to_have_value(system_prompt, timeout=5000)
-            self.logger.info(f'[{self.req_id}] 系统指令已成功填充并验证。')
+            await expect_async(sys_prompt_textarea).to_be_editable(timeout=TIMEOUT_ELEMENT_VISIBLE)
+            await sys_prompt_textarea.fill(system_prompt)
+            await asyncio.sleep(DELAY_AFTER_FILL)
+            filled_value = await sys_prompt_textarea.input_value(timeout=TIMEOUT_INPUT_VALUE)
+            if len(filled_value) >= len(system_prompt) * 0.9:
+                self.logger.info(f'[{self.req_id}] ✅ 系统指令已填充 (验证长度: {len(filled_value)} chars)')
+            else:
+                self.logger.warning(f'[{self.req_id}] ⚠️ 系统指令填充可能不完整 (期望: {len(system_prompt)}, 实际: {len(filled_value)})')
             for close_attempt in range(1, 4):
                 try:
                     await self.page.keyboard.press("Escape")
@@ -108,10 +114,14 @@ async def set_system_instructions(self, system_prompt: str, check_client_disconn
                 except Exception:
                     pass
         except Exception as e:
-            self.logger.error(f'[{self.req_id}] 设置系统指令时出错: {e}')
+            err_msg = str(e)
+            if len(err_msg) > 200:
+                err_msg = err_msg[:200] + '...[truncated]'
+            self.logger.error(f'[{self.req_id}] 设置系统指令时出错: {err_msg}')
             if isinstance(e, ClientDisconnectedError):
                 raise
 
+
     async def _control_thinking_mode_toggle(self, should_be_checked: bool, check_client_disconnected: Callable) -> bool:
         toggle_selector = THINKING_MODE_TOGGLE_SELECTOR
         action = '啟用' if should_be_checked else '停用'
@@ -385,6 +395,10 @@ async def _adjust_google_search(self, request_params: Dict[str, Any], check_clie
         self.logger.error(f"[{self.req_id}] ❌ Google Search 設定失敗，已重試 {max_retries} 次")
 
 
+    # NOTE: _extract_function_declarations 和 _adjust_function_calling 已移除
+    # Function Calling 改为将 tools 定义合并到 system prompt 中
+
+
     async def _ensure_advanced_settings_expanded(self, check_client_disconnected: Callable):
         max_retries = MAX_RETRIES
         expander_locator = self.page.locator(ADVANCED_SETTINGS_EXPANDER_SELECTOR)
diff --git a/src/config/constants.py b/src/config/constants.py
@@ -1,24 +1,48 @@
+# 常量配置
 import os
 import json
 from dotenv import load_dotenv
 load_dotenv()
+
+# 模型相关
 MODEL_NAME = os.environ.get('MODEL_NAME', 'AI-Studio_Proxy_API')
 CHAT_COMPLETION_ID_PREFIX = os.environ.get('CHAT_COMPLETION_ID_PREFIX', 'chatcmpl-')
 DEFAULT_FALLBACK_MODEL_ID = os.environ.get('DEFAULT_FALLBACK_MODEL_ID', 'no model list')
+
+# 默认参数
 DEFAULT_TEMPERATURE = float(os.environ.get('DEFAULT_TEMPERATURE', '1.0'))
 DEFAULT_MAX_OUTPUT_TOKENS = int(os.environ.get('DEFAULT_MAX_OUTPUT_TOKENS', '65536'))
 DEFAULT_TOP_P = float(os.environ.get('DEFAULT_TOP_P', '0.95'))
+
+# 功能开关
 ENABLE_URL_CONTEXT = os.environ.get('ENABLE_URL_CONTEXT', 'false').lower() in ('true', '1', 'yes')
 ENABLE_THINKING_BUDGET = os.environ.get('ENABLE_THINKING_BUDGET', 'false').lower() in ('true', '1', 'yes')
 DEFAULT_THINKING_BUDGET = int(os.environ.get('DEFAULT_THINKING_BUDGET', '8192'))
 ENABLE_GOOGLE_SEARCH = os.environ.get('ENABLE_GOOGLE_SEARCH', 'false').lower() in ('true', '1', 'yes')
+
+# 停止序列
 try:
     DEFAULT_STOP_SEQUENCES = json.loads(os.environ.get('DEFAULT_STOP_SEQUENCES', '["用户:"]'))
 except (json.JSONDecodeError, TypeError):
     DEFAULT_STOP_SEQUENCES = ['用户:']
+
+# URL和端点
 AI_STUDIO_URL_PATTERN = os.environ.get('AI_STUDIO_URL_PATTERN', 'aistudio.google.com/')
 MODELS_ENDPOINT_URL_CONTAINS = os.environ.get('MODELS_ENDPOINT_URL_CONTAINS', 'MakerSuiteService/ListModels')
+
+# 消息标记
 USER_INPUT_START_MARKER_SERVER = os.environ.get('USER_INPUT_START_MARKER_SERVER', '__USER_INPUT_START__')
 USER_INPUT_END_MARKER_SERVER = os.environ.get('USER_INPUT_END_MARKER_SERVER', '__USER_INPUT_END__')
+
+# 模型过滤
 EXCLUDED_MODELS_FILENAME = os.environ.get('EXCLUDED_MODELS_FILENAME', 'excluded_models.txt')
-STREAM_TIMEOUT_LOG_STATE = {'consecutive_timeouts': 0, 'last_error_log_time': 0.0, 'suppress_until_time': 0.0, 'max_initial_errors': int(os.environ.get('STREAM_MAX_INITIAL_ERRORS', '3')), 'warning_interval_after_suppress': float(os.environ.get('STREAM_WARNING_INTERVAL_AFTER_SUPPRESS', '60.0')), 'suppress_duration_after_initial_burst': float(os.environ.get('STREAM_SUPPRESS_DURATION_AFTER_INITIAL_BURST', '400.0'))}
+
+# 流式超时日志状态
+STREAM_TIMEOUT_LOG_STATE = {
+    'consecutive_timeouts': 0,
+    'last_error_log_time': 0.0,
+    'suppress_until_time': 0.0,
+    'max_initial_errors': int(os.environ.get('STREAM_MAX_INITIAL_ERRORS', '3')),
+    'warning_interval_after_suppress': float(os.environ.get('STREAM_WARNING_INTERVAL_AFTER_SUPPRESS', '60.0')),
+    'suppress_duration_after_initial_burst': float(os.environ.get('STREAM_SUPPRESS_DURATION_AFTER_INITIAL_BURST', '400.0'))
+}
diff --git a/src/config/selectors.py b/src/config/selectors.py
@@ -1,3 +1,4 @@
+# 输入区域
 PROMPT_TEXTAREA_SELECTORS = [
     'ms-prompt-box textarea[aria-label="Enter a prompt"]',
     'ms-prompt-box textarea[placeholder="Start typing a prompt"]',
@@ -9,13 +10,15 @@
 INPUT_SELECTOR = PROMPT_TEXTAREA_SELECTOR
 INPUT_SELECTOR2 = PROMPT_TEXTAREA_SELECTOR
 
+# 提交按钮
 SUBMIT_BUTTON_SELECTORS = [
     'button[aria-label="Run"].run-button',
     'ms-run-button button[aria-label="Run"]',
     'ms-prompt-box ms-run-button button',
 ]
 SUBMIT_BUTTON_SELECTOR = SUBMIT_BUTTON_SELECTORS[0]
 
+# 文件上传
 INSERT_BUTTON_SELECTORS = [
     'button[data-test-id="add-media-button"]',
     'button[aria-label="Insert images, videos, audio, or files"]',
@@ -40,39 +43,65 @@
 HIDDEN_FILE_INPUT_SELECTOR = HIDDEN_FILE_INPUT_SELECTORS[0]
 
 UPLOADED_MEDIA_ITEM_SELECTOR = 'ms-prompt-box .multi-media-row ms-media-chip'
+
+# 响应区域
 SKIP_PREFERENCE_VOTE_BUTTON_SELECTOR = 'button[data-test-id="skip-button"][aria-label="Skip preference vote"]'
 RESPONSE_CONTAINER_SELECTOR = 'ms-chat-turn .chat-turn-container.model'
 RESPONSE_TEXT_SELECTOR = 'ms-cmark-node.cmark-node'
 
+# 加载状态
 LOADING_SPINNER_SELECTORS = [
     'button[aria-label="Run"].run-button svg .stoppable-spinner',
     'ms-run-button button svg .stoppable-spinner',
     'ms-prompt-box ms-run-button button svg .stoppable-spinner',
 ]
 LOADING_SPINNER_SELECTOR = LOADING_SPINNER_SELECTORS[0]
 
+# 对话框/遮罩层
 OVERLAY_SELECTOR = '.mat-mdc-dialog-inner-container'
 ZERO_STATE_SELECTOR = 'ms-zero-state'
 ERROR_TOAST_SELECTOR = 'div.toast.warning, div.toast.error'
+
+# 消息编辑
 EDIT_MESSAGE_BUTTON_SELECTOR = 'button[aria-label="Edit"].toggle-edit-button:has(span:text-is("edit"))'
 MESSAGE_TEXTAREA_SELECTOR = 'ms-chat-turn:last-child ms-text-chunk ms-autosize-textarea'
 FINISH_EDIT_BUTTON_SELECTOR = 'button[aria-label="Stop editing"].toggle-edit-button'
 MORE_OPTIONS_BUTTON_SELECTOR = 'button[aria-label="Open options"]'
 COPY_MARKDOWN_BUTTON_SELECTOR = 'button[role="menuitem"]:has-text("Copy markdown")'
 COPY_MARKDOWN_BUTTON_SELECTOR_ALT = 'div[role="menu"] button:has-text("Copy Markdown")'
+
+# 高级设置
 ADVANCED_SETTINGS_EXPANDER_SELECTOR = 'button[aria-label="Expand or collapse advanced settings"]'
 MAX_OUTPUT_TOKENS_SELECTOR = 'input[aria-label="Maximum output tokens"]'
 STOP_SEQUENCE_INPUT_SELECTOR = 'input[aria-label="Add stop token"]'
 MAT_CHIP_REMOVE_BUTTON_SELECTOR = 'mat-chip-set mat-chip-row button[aria-label*="Remove"]'
 TOP_P_INPUT_SELECTOR = '//div[contains(@class, "settings-item-column") and .//h3[normalize-space()="Top P"]]//input[@type="number"]'
 TEMPERATURE_INPUT_SELECTOR = '//div[contains(@class, "settings-item-column") and .//h3[normalize-space()="Temperature"]]//input[@type="number"]'
+
+# 工具面板
 USE_URL_CONTEXT_SELECTOR = 'button[aria-label="Browse the url context"]'
+GROUNDING_WITH_GOOGLE_SEARCH_TOGGLE_SELECTOR = 'div[data-test-id="searchAsAToolTooltip"] mat-slide-toggle button'
+
+# 思考模式
 THINKING_MODE_TOGGLE_SELECTOR = 'mat-slide-toggle[data-test-toggle="enable-thinking"]'
 SET_THINKING_BUDGET_TOGGLE_SELECTOR = 'mat-slide-toggle[data-test-toggle="manual-budget"]'
 THINKING_BUDGET_INPUT_SELECTOR = 'ms-slider[data-test-slider] input[type="number"]'
-GROUNDING_WITH_GOOGLE_SEARCH_TOGGLE_SELECTOR = 'div[data-test-id="searchAsAToolTooltip"] mat-slide-toggle button'
+THINKING_LEVEL_SELECT_SELECTOR = 'mat-select[aria-label="Thinking Level"], mat-select[aria-label="Thinking level"]'
+THINKING_LEVEL_OPTIONS = {
+    'minimal': 'mat-option:has-text("Minimal")',
+    'low': 'mat-option:has-text("Low")',
+    'medium': 'mat-option:has-text("Medium")',
+    'high': 'mat-option:has-text("High")',
+}
+THINKING_LEVEL_OPTION_HIGH_SELECTOR = THINKING_LEVEL_OPTIONS['high']
+THINKING_LEVEL_OPTION_LOW_SELECTOR = THINKING_LEVEL_OPTIONS['low']
+DEFAULT_THINKING_LEVEL = "high"
+
+# 系统指令
 SYSTEM_INSTRUCTIONS_BUTTON_SELECTOR = 'button[aria-label="System instructions"]'
 SYSTEM_INSTRUCTIONS_TEXTAREA_SELECTOR = 'textarea[aria-label="System instructions"]'
+
+# 模型选择器
 MODEL_SELECTOR_CARD_TITLE = '.model-selector-card .title'
 MODEL_SELECTOR_CARD_NAME = '[data-test-id="model-name"]'
 MODEL_SELECTOR_CARD_SUBTITLE = '.model-selector-card .subtitle'
@@ -83,18 +112,15 @@
 MODEL_OPTION_CONTENT_SPAN = '.model-option-content span'
 MODEL_SELECTORS_LIST = [MODEL_SELECTOR_CARD_TITLE, MODEL_SELECTOR_CARD_NAME, MODEL_SELECTOR_CARD_SUBTITLE, MODEL_SELECTOR_LEGACY_PRIMARY, MODEL_SELECTOR_LEGACY_FALLBACK, MODEL_SELECTOR_LEGACY_GENERIC, '.model-selector span', MODEL_SELECTOR_BUTTON_SPAN, MODEL_OPTION_CONTENT_SPAN]
 
-THINKING_LEVEL_SELECT_SELECTOR = 'mat-select[aria-label="Thinking Level"], mat-select[aria-label="Thinking level"]'
-THINKING_LEVEL_OPTIONS = {
-    'minimal': 'mat-option:has-text("Minimal")',
-    'low': 'mat-option:has-text("Low")',
-    'medium': 'mat-option:has-text("Medium")',
-    'high': 'mat-option:has-text("High")',
-}
-THINKING_LEVEL_OPTION_HIGH_SELECTOR = THINKING_LEVEL_OPTIONS['high']
-THINKING_LEVEL_OPTION_LOW_SELECTOR = THINKING_LEVEL_OPTIONS['low']
-DEFAULT_THINKING_LEVEL = "high"
-
+# 速率限制
 RATE_LIMIT_CALLOUT_SELECTOR = 'ms-callout.error-callout .message, ms-callout.warning-callout .message'
 RATE_LIMIT_KEYWORDS = ["exceeded quota", "out of free generations"]
 
-
+# Function Calling (尚未使用，改为合并到system prompt)
+FUNCTION_CALLING_TOGGLE_SELECTOR = 'div[data-test-id="functionCallingTooltip"] mat-slide-toggle'
+EDIT_FUNCTION_DECLARATIONS_BUTTON_SELECTOR = 'button.edit-function-declarations-button'
+FUNCTION_DECLARATIONS_DIALOG_SELECTOR = 'ms-edit-function-declarations-dialog'
+FUNCTION_DECLARATIONS_DIALOG_CLOSE_BUTTON_SELECTOR = 'ms-edit-function-declarations-dialog h2 button[aria-label="close"]'
+FUNCTION_DECLARATIONS_CODE_EDITOR_TAB_SELECTOR = 'ms-edit-function-declarations-dialog button[role="tab"]:has-text("Code Editor")'
+FUNCTION_DECLARATIONS_TEXTAREA_SELECTOR = 'ms-edit-function-declarations-dialog ms-text-editor textarea'
+FUNCTION_DECLARATIONS_SAVE_BUTTON_SELECTOR = 'ms-edit-function-declarations-dialog mat-dialog-actions button:has-text("Save")'
diff --git a/src/config/settings.py b/src/config/settings.py
@@ -1,11 +1,18 @@
+# 应用设置
 import os
 from dotenv import load_dotenv
 load_dotenv()
+
+# 调试开关
 DEBUG_LOGS_ENABLED = os.environ.get('DEBUG_LOGS_ENABLED', 'false').lower() in ('true', '1', 'yes')
 TRACE_LOGS_ENABLED = os.environ.get('TRACE_LOGS_ENABLED', 'false').lower() in ('true', '1', 'yes')
+
+# 认证相关
 AUTO_SAVE_AUTH = os.environ.get('AUTO_SAVE_AUTH', '').lower() in ('1', 'true', 'yes')
 AUTH_SAVE_TIMEOUT = int(os.environ.get('AUTH_SAVE_TIMEOUT', '30'))
 AUTO_CONFIRM_LOGIN = os.environ.get('AUTO_CONFIRM_LOGIN', 'true').lower() in ('1', 'true', 'yes')
+
+# 目录路径
 PROJECT_ROOT = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
 DATA_DIR = os.path.join(PROJECT_ROOT, 'data')
 AUTH_PROFILES_DIR = os.path.join(DATA_DIR, 'auth_profiles')
@@ -14,6 +21,7 @@
 LOG_DIR = os.path.join(PROJECT_ROOT, 'logs')
 APP_LOG_FILE_PATH = os.path.join(LOG_DIR, 'app.log')
 
+# 工具函数
 def get_environment_variable(key: str, default: str='') -> str:
     return os.environ.get(key, default)
 
@@ -29,6 +37,8 @@ def get_int_env(key: str, default: int=0) -> int:
         return int(os.environ.get(key, str(default)))
     except (ValueError, TypeError):
         return default
+
+# 代理和脚本注入
 NO_PROXY_ENV = os.environ.get('NO_PROXY')
 ENABLE_SCRIPT_INJECTION = get_boolean_env('ENABLE_SCRIPT_INJECTION', True)
 USERSCRIPT_PATH = get_environment_variable('USERSCRIPT_PATH', 'browser/more_models.js')
diff --git a/src/config/timeouts.py b/src/config/timeouts.py
