1199 lines
61 KiB
Python
1199 lines
61 KiB
Python
from __future__ import annotations
|
||
|
||
import asyncio
|
||
import json
|
||
import time
|
||
import re
|
||
import zipfile
|
||
from collections import defaultdict, Counter, deque
|
||
from datetime import datetime, timedelta
|
||
from pathlib import Path
|
||
from typing import Any, Dict, List, Optional, Tuple
|
||
|
||
from werkzeug.utils import secure_filename
|
||
|
||
from config import (
|
||
OUTPUT_FORMATS,
|
||
AUTO_FIX_TOOL_CALL,
|
||
AUTO_FIX_MAX_ATTEMPTS,
|
||
MAX_ITERATIONS_PER_TASK,
|
||
MAX_CONSECUTIVE_SAME_TOOL,
|
||
MAX_TOTAL_TOOL_CALLS,
|
||
TOOL_CALL_COOLDOWN,
|
||
MAX_UPLOAD_SIZE,
|
||
DEFAULT_CONVERSATIONS_LIMIT,
|
||
MAX_CONVERSATIONS_LIMIT,
|
||
CONVERSATIONS_DIR,
|
||
DEFAULT_RESPONSE_MAX_TOKENS,
|
||
DEFAULT_PROJECT_PATH,
|
||
LOGS_DIR,
|
||
AGENT_VERSION,
|
||
THINKING_FAST_INTERVAL,
|
||
PROJECT_MAX_STORAGE_MB,
|
||
PROJECT_MAX_STORAGE_BYTES,
|
||
UPLOAD_SCAN_LOG_SUBDIR,
|
||
)
|
||
from modules.personalization_manager import (
|
||
load_personalization_config,
|
||
save_personalization_config,
|
||
THINKING_INTERVAL_MIN,
|
||
THINKING_INTERVAL_MAX,
|
||
)
|
||
from modules.upload_security import UploadSecurityError
|
||
from modules.user_manager import UserWorkspace
|
||
from modules.usage_tracker import QUOTA_DEFAULTS
|
||
from core.web_terminal import WebTerminal
|
||
from utils.tool_result_formatter import format_tool_result_for_context
|
||
from utils.conversation_manager import ConversationManager
|
||
from config.model_profiles import get_model_context_window, get_model_profile
|
||
|
||
from .auth_helpers import api_login_required, resolve_admin_policy, get_current_user_record, get_current_username
|
||
from .context import with_terminal, get_gui_manager, get_upload_guard, build_upload_error_response, ensure_conversation_loaded, reset_system_state, get_user_resources, get_or_create_usage_tracker
|
||
from .utils_common import (
|
||
build_review_lines,
|
||
debug_log,
|
||
log_backend_chunk,
|
||
log_frontend_chunk,
|
||
log_streaming_debug_entry,
|
||
brief_log,
|
||
DEBUG_LOG_FILE,
|
||
CHUNK_BACKEND_LOG_FILE,
|
||
CHUNK_FRONTEND_LOG_FILE,
|
||
STREAMING_DEBUG_LOG_FILE,
|
||
)
|
||
from .security import rate_limited, format_tool_result_notice, compact_web_search_result, consume_socket_token, prune_socket_tokens, validate_csrf_request, requires_csrf_protection, get_csrf_token
|
||
from .monitor import cache_monitor_snapshot, get_cached_monitor_snapshot
|
||
from .extensions import socketio
|
||
from .state import (
|
||
MONITOR_FILE_TOOLS,
|
||
MONITOR_MEMORY_TOOLS,
|
||
MONITOR_SNAPSHOT_CHAR_LIMIT,
|
||
MONITOR_MEMORY_ENTRY_LIMIT,
|
||
RATE_LIMIT_BUCKETS,
|
||
FAILURE_TRACKERS,
|
||
pending_socket_tokens,
|
||
usage_trackers,
|
||
MONITOR_SNAPSHOT_CACHE,
|
||
MONITOR_SNAPSHOT_CACHE_LIMIT,
|
||
PROJECT_STORAGE_CACHE,
|
||
PROJECT_STORAGE_CACHE_TTL_SECONDS,
|
||
RECENT_UPLOAD_EVENT_LIMIT,
|
||
RECENT_UPLOAD_FEED_LIMIT,
|
||
THINKING_FAILURE_KEYWORDS,
|
||
TITLE_PROMPT_PATH,
|
||
get_last_active_ts,
|
||
user_manager,
|
||
container_manager,
|
||
custom_tool_registry,
|
||
user_terminals,
|
||
terminal_rooms,
|
||
connection_users,
|
||
stop_flags,
|
||
get_stop_flag,
|
||
set_stop_flag,
|
||
clear_stop_flag,
|
||
)
|
||
from .chat_flow_helpers import (
|
||
detect_malformed_tool_call as _detect_malformed_tool_call,
|
||
detect_tool_failure,
|
||
get_thinking_state,
|
||
mark_force_thinking as _mark_force_thinking,
|
||
mark_suppress_thinking,
|
||
apply_thinking_schedule as _apply_thinking_schedule,
|
||
update_thinking_after_call as _update_thinking_after_call,
|
||
maybe_mark_failure_from_message as _maybe_mark_failure_from_message,
|
||
generate_conversation_title_background as _generate_conversation_title_background,
|
||
)
|
||
|
||
|
||
from .chat_flow_runner_helpers import (
|
||
extract_intent_from_partial,
|
||
resolve_monitor_path,
|
||
resolve_monitor_memory,
|
||
capture_monitor_snapshot,
|
||
)
|
||
|
||
|
||
from .chat_flow_runtime import (
|
||
generate_conversation_title_background,
|
||
mark_force_thinking,
|
||
apply_thinking_schedule,
|
||
update_thinking_after_call,
|
||
maybe_mark_failure_from_message,
|
||
detect_malformed_tool_call,
|
||
)
|
||
|
||
from .chat_flow_pending_writes import finalize_pending_append, finalize_pending_modify
|
||
from .chat_flow_task_support import process_sub_agent_updates, wait_retry_delay, cancel_pending_tools
|
||
from .chat_flow_tool_loop import execute_tool_calls
|
||
|
||
async def handle_task_with_sender(terminal: WebTerminal, workspace: UserWorkspace, message, images, sender, client_sid, username: str, videos=None):
|
||
"""处理任务并发送消息 - 集成token统计版本"""
|
||
web_terminal = terminal
|
||
conversation_id = getattr(web_terminal.context_manager, "current_conversation_id", None)
|
||
videos = videos or []
|
||
raw_sender = sender
|
||
|
||
def sender(event_type, data):
|
||
"""为关键事件补充会话标识，便于前端定位报错归属。"""
|
||
if not isinstance(data, dict):
|
||
raw_sender(event_type, data)
|
||
return
|
||
payload = data
|
||
if event_type in {"error", "quota_exceeded", "task_stopped", "task_complete"}:
|
||
payload = dict(data)
|
||
current_conv = conversation_id or getattr(web_terminal.context_manager, "current_conversation_id", None)
|
||
if current_conv:
|
||
payload.setdefault("conversation_id", current_conv)
|
||
task_id = getattr(web_terminal, "task_id", None) or client_sid
|
||
if task_id:
|
||
payload.setdefault("task_id", task_id)
|
||
if client_sid:
|
||
payload.setdefault("client_sid", client_sid)
|
||
raw_sender(event_type, payload)
|
||
|
||
# 如果是思考模式，重置状态
|
||
if web_terminal.thinking_mode:
|
||
web_terminal.api_client.start_new_task(force_deep=web_terminal.deep_thinking_mode)
|
||
state = get_thinking_state(web_terminal)
|
||
state["fast_streak"] = 0
|
||
state["force_next"] = False
|
||
state["suppress_next"] = False
|
||
|
||
# 添加到对话历史
|
||
history_len_before = len(getattr(web_terminal.context_manager, "conversation_history", []) or [])
|
||
is_first_user_message = history_len_before == 0
|
||
web_terminal.context_manager.add_conversation("user", message, images=images, videos=videos)
|
||
|
||
if is_first_user_message and getattr(web_terminal, "context_manager", None):
|
||
try:
|
||
personal_config = load_personalization_config(workspace.data_dir)
|
||
except Exception:
|
||
personal_config = {}
|
||
auto_title_enabled = personal_config.get("auto_generate_title", True)
|
||
if auto_title_enabled:
|
||
conv_id = getattr(web_terminal.context_manager, "current_conversation_id", None)
|
||
socketio.start_background_task(
|
||
generate_conversation_title_background,
|
||
web_terminal,
|
||
conv_id,
|
||
message,
|
||
username
|
||
)
|
||
|
||
# === 移除：不在这里计算输入token，改为在每次API调用前计算 ===
|
||
|
||
# 构建上下文和消息（用于API调用）
|
||
context = web_terminal.build_context()
|
||
messages = web_terminal.build_messages(context, message)
|
||
tools = web_terminal.define_tools()
|
||
try:
|
||
profile = get_model_profile(getattr(web_terminal, "model_key", None) or "kimi-k2.5")
|
||
web_terminal.apply_model_profile(profile)
|
||
except Exception as exc:
|
||
debug_log(f"更新模型配置失败: {exc}")
|
||
|
||
# === 上下文预算与安全校验（避免超出模型上下文） ===
|
||
max_context_tokens = get_model_context_window(getattr(web_terminal, "model_key", None) or "kimi-k2.5")
|
||
current_tokens = web_terminal.context_manager.get_current_context_tokens(conversation_id)
|
||
# 提前同步给底层客户端，动态收缩 max_tokens
|
||
web_terminal.api_client.update_context_budget(current_tokens, max_context_tokens)
|
||
if max_context_tokens:
|
||
if current_tokens >= max_context_tokens:
|
||
err_msg = (
|
||
f"当前对话上下文已达 {current_tokens} tokens，超过模型上限 "
|
||
f"{max_context_tokens}，请先使用压缩功能或清理对话后再试。"
|
||
)
|
||
debug_log(err_msg)
|
||
web_terminal.context_manager.add_conversation("system", err_msg)
|
||
sender('error', {
|
||
'message': err_msg,
|
||
'status_code': 400,
|
||
'error_type': 'context_overflow'
|
||
})
|
||
return
|
||
usage_percent = (current_tokens / max_context_tokens) * 100
|
||
warned = web_terminal.context_manager.conversation_metadata.get("context_warning_sent", False)
|
||
if usage_percent >= 70 and not warned:
|
||
warn_msg = (
|
||
f"当前对话上下文约占 {usage_percent:.1f}%（{current_tokens}/{max_context_tokens}），"
|
||
"建议使用压缩功能。"
|
||
)
|
||
web_terminal.context_manager.conversation_metadata["context_warning_sent"] = True
|
||
web_terminal.context_manager.auto_save_conversation(force=True)
|
||
sender('context_warning', {
|
||
'title': '上下文过长',
|
||
'message': warn_msg,
|
||
'type': 'warning',
|
||
'conversation_id': conversation_id
|
||
})
|
||
|
||
# 开始新的AI消息
|
||
sender('ai_message_start', {})
|
||
|
||
# 增量保存相关变量
|
||
accumulated_response = "" # 累积的响应内容
|
||
is_first_iteration = True # 是否是第一次迭代
|
||
|
||
# 统计和限制变量
|
||
total_iterations = 0
|
||
total_tool_calls = 0
|
||
consecutive_same_tool = defaultdict(int)
|
||
last_tool_name = ""
|
||
auto_fix_attempts = 0
|
||
last_tool_call_time = 0
|
||
detected_tool_intent: Dict[str, str] = {}
|
||
|
||
# 设置最大迭代次数（API 可覆盖）；None 表示不限制
|
||
max_iterations_override = getattr(web_terminal, "max_iterations_override", None)
|
||
max_iterations = max_iterations_override if max_iterations_override is not None else MAX_ITERATIONS_PER_TASK
|
||
max_api_retries = 4
|
||
retry_delay_seconds = 10
|
||
|
||
pending_append = None # {"path": str, "tool_call_id": str, "buffer": str, ...}
|
||
append_probe_buffer = ""
|
||
pending_modify = None # {"path": str, "tool_call_id": str, "buffer": str, ...}
|
||
modify_probe_buffer = ""
|
||
|
||
iteration = 0
|
||
while max_iterations is None or iteration < max_iterations:
|
||
current_iteration = iteration + 1
|
||
iteration += 1
|
||
total_iterations += 1
|
||
iteration_limit_label = max_iterations if max_iterations is not None else "∞"
|
||
debug_log(f"\n--- 迭代 {current_iteration}/{iteration_limit_label} 开始 ---")
|
||
|
||
# 检查是否超过总工具调用限制
|
||
if MAX_TOTAL_TOOL_CALLS is not None and total_tool_calls >= MAX_TOTAL_TOOL_CALLS:
|
||
debug_log(f"已达到最大工具调用次数限制 ({MAX_TOTAL_TOOL_CALLS})")
|
||
sender('system_message', {
|
||
'content': f'⚠️ 已达到最大工具调用次数限制 ({MAX_TOTAL_TOOL_CALLS})，任务结束。'
|
||
})
|
||
mark_force_thinking(web_terminal, reason="tool_limit")
|
||
break
|
||
|
||
apply_thinking_schedule(web_terminal)
|
||
|
||
full_response = ""
|
||
tool_calls = []
|
||
current_thinking = ""
|
||
detected_tools = {}
|
||
last_usage_payload = None
|
||
|
||
# 状态标志
|
||
in_thinking = False
|
||
thinking_started = False
|
||
thinking_ended = False
|
||
text_started = False
|
||
text_has_content = False
|
||
text_streaming = False
|
||
text_chunk_index = 0
|
||
last_text_chunk_time: Optional[float] = None
|
||
|
||
# 计数器
|
||
chunk_count = 0
|
||
reasoning_chunks = 0
|
||
content_chunks = 0
|
||
tool_chunks = 0
|
||
append_break_triggered = False
|
||
append_result = {"handled": False}
|
||
modify_break_triggered = False
|
||
modify_result = {"handled": False}
|
||
last_finish_reason = None
|
||
|
||
thinking_expected = web_terminal.api_client.get_current_thinking_mode()
|
||
debug_log(f"思考模式: {thinking_expected}")
|
||
quota_allowed = True
|
||
quota_info = {}
|
||
if hasattr(web_terminal, "record_model_call"):
|
||
quota_allowed, quota_info = web_terminal.record_model_call(bool(thinking_expected))
|
||
if not quota_allowed:
|
||
quota_type = 'thinking' if thinking_expected else 'fast'
|
||
socketio.emit('quota_notice', {
|
||
'type': quota_type,
|
||
'reset_at': quota_info.get('reset_at'),
|
||
'limit': quota_info.get('limit'),
|
||
'count': quota_info.get('count')
|
||
}, room=f"user_{getattr(web_terminal, 'username', '')}")
|
||
sender('quota_exceeded', {
|
||
'type': quota_type,
|
||
'reset_at': quota_info.get('reset_at')
|
||
})
|
||
sender('error', {
|
||
'message': "配额已达到上限，暂时无法继续调用模型。",
|
||
'quota': quota_info
|
||
})
|
||
return
|
||
|
||
tool_call_limit_label = MAX_TOTAL_TOOL_CALLS if MAX_TOTAL_TOOL_CALLS is not None else "∞"
|
||
print(f"[API] 第{current_iteration}次调用 (总工具调用: {total_tool_calls}/{tool_call_limit_label})")
|
||
|
||
api_error = None
|
||
for api_attempt in range(max_api_retries + 1):
|
||
api_error = None
|
||
if api_attempt > 0:
|
||
full_response = ""
|
||
tool_calls = []
|
||
current_thinking = ""
|
||
detected_tools = {}
|
||
last_usage_payload = None
|
||
in_thinking = False
|
||
thinking_started = False
|
||
thinking_ended = False
|
||
text_started = False
|
||
text_has_content = False
|
||
text_streaming = False
|
||
text_chunk_index = 0
|
||
last_text_chunk_time = None
|
||
chunk_count = 0
|
||
reasoning_chunks = 0
|
||
content_chunks = 0
|
||
tool_chunks = 0
|
||
append_break_triggered = False
|
||
append_result = {"handled": False}
|
||
modify_break_triggered = False
|
||
modify_result = {"handled": False}
|
||
last_finish_reason = None
|
||
|
||
# 收集流式响应
|
||
async for chunk in web_terminal.api_client.chat(messages, tools, stream=True):
|
||
chunk_count += 1
|
||
|
||
# 检查停止标志
|
||
client_stop_info = get_stop_flag(client_sid, username)
|
||
if client_stop_info:
|
||
stop_requested = client_stop_info.get('stop', False) if isinstance(client_stop_info, dict) else client_stop_info
|
||
if stop_requested:
|
||
debug_log(f"检测到停止请求，中断流处理")
|
||
if pending_append:
|
||
append_result, pending_append, append_probe_buffer = await finalize_pending_append(pending_append=pending_append, append_probe_buffer=append_probe_buffer, response_text=full_response, stream_completed=False, finish_reason="user_stop", web_terminal=web_terminal, sender=sender, debug_log=debug_log)
|
||
if pending_modify:
|
||
modify_result, pending_modify, modify_probe_buffer = await finalize_pending_modify(pending_modify=pending_modify, modify_probe_buffer=modify_probe_buffer, response_text=full_response, stream_completed=False, finish_reason="user_stop", web_terminal=web_terminal, sender=sender, debug_log=debug_log)
|
||
cancel_pending_tools(tool_calls_list=tool_calls, sender=sender, messages=messages)
|
||
sender('task_stopped', {
|
||
'message': '命令执行被用户取消',
|
||
'reason': 'user_stop'
|
||
})
|
||
clear_stop_flag(client_sid, username)
|
||
return
|
||
|
||
if isinstance(chunk, dict) and chunk.get("error"):
|
||
api_error = chunk.get("error")
|
||
break
|
||
|
||
# 先尝试记录 usage（有些平台会在最后一个 chunk 里携带 usage 但 choices 为空）
|
||
usage_info = chunk.get("usage")
|
||
if usage_info:
|
||
last_usage_payload = usage_info
|
||
|
||
if "choices" not in chunk:
|
||
debug_log(f"Chunk {chunk_count}: 无choices字段")
|
||
continue
|
||
if not chunk.get("choices"):
|
||
debug_log(f"Chunk {chunk_count}: choices为空列表")
|
||
continue
|
||
choice = chunk["choices"][0]
|
||
if not usage_info and isinstance(choice, dict) and choice.get("usage"):
|
||
# 兼容部分供应商将 usage 放在 choice 内的格式（例如部分 Kimi/Qwen 返回）
|
||
last_usage_payload = choice.get("usage")
|
||
delta = choice.get("delta", {})
|
||
finish_reason = choice.get("finish_reason")
|
||
if finish_reason:
|
||
last_finish_reason = finish_reason
|
||
|
||
# 处理思考内容（兼容 reasoning_content / reasoning_details）
|
||
reasoning_content = ""
|
||
if "reasoning_content" in delta:
|
||
reasoning_content = delta.get("reasoning_content") or ""
|
||
elif "reasoning_details" in delta:
|
||
details = delta.get("reasoning_details")
|
||
if isinstance(details, list):
|
||
parts = []
|
||
for item in details:
|
||
if isinstance(item, dict):
|
||
text = item.get("text")
|
||
if text:
|
||
parts.append(text)
|
||
if parts:
|
||
reasoning_content = "".join(parts)
|
||
if reasoning_content:
|
||
reasoning_chunks += 1
|
||
debug_log(f" 思考内容 #{reasoning_chunks}: {len(reasoning_content)} 字符")
|
||
|
||
if not thinking_started:
|
||
in_thinking = True
|
||
thinking_started = True
|
||
sender('thinking_start', {})
|
||
await asyncio.sleep(0.05)
|
||
|
||
current_thinking += reasoning_content
|
||
sender('thinking_chunk', {'content': reasoning_content})
|
||
|
||
# 处理正常内容
|
||
if "content" in delta:
|
||
content = delta["content"]
|
||
if content:
|
||
content_chunks += 1
|
||
debug_log(f" 正式内容 #{content_chunks}: {repr(content[:100] if content else 'None')}")
|
||
|
||
if in_thinking and not thinking_ended:
|
||
in_thinking = False
|
||
thinking_ended = True
|
||
sender('thinking_end', {'full_content': current_thinking})
|
||
await asyncio.sleep(0.1)
|
||
|
||
|
||
expecting_modify = bool(pending_modify) or bool(getattr(web_terminal, "pending_modify_request", None))
|
||
expecting_append = bool(pending_append) or bool(getattr(web_terminal, "pending_append_request", None))
|
||
|
||
if pending_modify:
|
||
if not pending_modify.get("start_seen"):
|
||
probe_buffer = pending_modify.get("probe_buffer", "") + content
|
||
if len(probe_buffer) > 10000:
|
||
probe_buffer = probe_buffer[-10000:]
|
||
marker = pending_modify.get("start_marker")
|
||
marker_index = probe_buffer.find(marker)
|
||
if marker_index == -1:
|
||
pending_modify["probe_buffer"] = probe_buffer
|
||
continue
|
||
after_marker = marker_index + len(marker)
|
||
remainder = probe_buffer[after_marker:]
|
||
pending_modify["buffer"] = remainder
|
||
pending_modify["raw_buffer"] = marker + remainder
|
||
pending_modify["start_seen"] = True
|
||
pending_modify["detected_blocks"] = set()
|
||
pending_modify["probe_buffer"] = ""
|
||
if pending_modify.get("display_id"):
|
||
sender('update_action', {
|
||
'id': pending_modify["display_id"],
|
||
'status': 'running',
|
||
'preparing_id': pending_modify.get("tool_call_id"),
|
||
'message': f"正在修改 {pending_modify['path']}..."
|
||
})
|
||
else:
|
||
pending_modify["buffer"] += content
|
||
pending_modify["raw_buffer"] += content
|
||
|
||
if pending_modify.get("start_seen"):
|
||
block_text = pending_modify["buffer"]
|
||
for match in re.finditer(r"\[replace:(\d+)\]", block_text):
|
||
try:
|
||
block_index = int(match.group(1))
|
||
except ValueError:
|
||
continue
|
||
detected_blocks = pending_modify.setdefault("detected_blocks", set())
|
||
if block_index not in detected_blocks:
|
||
detected_blocks.add(block_index)
|
||
if pending_modify.get("display_id"):
|
||
sender('update_action', {
|
||
'id': pending_modify["display_id"],
|
||
'status': 'running',
|
||
'preparing_id': pending_modify.get("tool_call_id"),
|
||
'message': f"正在对 {pending_modify['path']} 进行第 {block_index} 处修改..."
|
||
})
|
||
|
||
if pending_modify.get("start_seen"):
|
||
end_pos = pending_modify["buffer"].find(pending_modify["end_marker"])
|
||
if end_pos != -1:
|
||
pending_modify["end_index"] = end_pos
|
||
modify_break_triggered = True
|
||
debug_log("检测到<<<END_MODIFY>>>，即将终止流式输出并应用修改")
|
||
break
|
||
continue
|
||
elif expecting_modify:
|
||
modify_probe_buffer += content
|
||
if len(modify_probe_buffer) > 10000:
|
||
modify_probe_buffer = modify_probe_buffer[-10000:]
|
||
|
||
marker_match = re.search(r"<<<MODIFY:\s*([\s\S]*?)>>>", modify_probe_buffer)
|
||
if marker_match:
|
||
detected_raw_path = marker_match.group(1)
|
||
detected_path = detected_raw_path.strip()
|
||
marker_full = marker_match.group(0)
|
||
after_marker_index = modify_probe_buffer.find(marker_full) + len(marker_full)
|
||
remainder = modify_probe_buffer[after_marker_index:]
|
||
modify_probe_buffer = ""
|
||
|
||
if not detected_path:
|
||
debug_log("检测到 MODIFY 起始标记但路径为空，忽略。")
|
||
continue
|
||
|
||
pending_modify = {
|
||
"path": detected_path,
|
||
"tool_call_id": None,
|
||
"buffer": remainder,
|
||
"raw_buffer": marker_full + remainder,
|
||
"start_marker": marker_full,
|
||
"end_marker": "<<<END_MODIFY>>>",
|
||
"start_seen": True,
|
||
"end_index": None,
|
||
"display_id": None,
|
||
"detected_blocks": set()
|
||
}
|
||
if hasattr(web_terminal, "pending_modify_request"):
|
||
web_terminal.pending_modify_request = {"path": detected_path}
|
||
debug_log(f"直接检测到modify起始标记，构建修改缓冲: {detected_path}")
|
||
|
||
end_pos = pending_modify["buffer"].find(pending_modify["end_marker"])
|
||
if end_pos != -1:
|
||
pending_modify["end_index"] = end_pos
|
||
modify_break_triggered = True
|
||
debug_log("检测到<<<END_MODIFY>>>，即将终止流式输出并应用修改")
|
||
break
|
||
continue
|
||
|
||
if pending_append:
|
||
pending_append["buffer"] += content
|
||
|
||
if pending_append.get("content_start") is None:
|
||
marker_index = pending_append["buffer"].find(pending_append["start_marker"])
|
||
if marker_index != -1:
|
||
pending_append["content_start"] = marker_index + len(pending_append["start_marker"])
|
||
debug_log(f"检测到追加起始标识: {pending_append['start_marker']}")
|
||
|
||
if pending_append.get("content_start") is not None:
|
||
end_index = pending_append["buffer"].find(
|
||
pending_append["end_marker"],
|
||
pending_append["content_start"]
|
||
)
|
||
if end_index != -1:
|
||
pending_append["end_index"] = end_index
|
||
append_break_triggered = True
|
||
debug_log("检测到<<<END_APPEND>>>，即将终止流式输出并写入文件")
|
||
break
|
||
|
||
# 继续累积追加内容
|
||
continue
|
||
elif expecting_append:
|
||
append_probe_buffer += content
|
||
# 限制缓冲区大小防止过长
|
||
if len(append_probe_buffer) > 10000:
|
||
append_probe_buffer = append_probe_buffer[-10000:]
|
||
|
||
marker_match = re.search(r"<<<APPEND:\s*([\s\S]*?)>>>", append_probe_buffer)
|
||
if marker_match:
|
||
detected_raw_path = marker_match.group(1)
|
||
detected_path = detected_raw_path.strip()
|
||
if not detected_path:
|
||
append_probe_buffer = append_probe_buffer[marker_match.end():]
|
||
continue
|
||
marker_full = marker_match.group(0)
|
||
after_marker_index = append_probe_buffer.find(marker_full) + len(marker_full)
|
||
remainder = append_probe_buffer[after_marker_index:]
|
||
append_probe_buffer = ""
|
||
pending_append = {
|
||
"path": detected_path,
|
||
"tool_call_id": None,
|
||
"buffer": remainder,
|
||
"start_marker": marker_full,
|
||
"end_marker": "<<<END_APPEND>>>",
|
||
"content_start": 0,
|
||
"end_index": None,
|
||
"display_id": None
|
||
}
|
||
if hasattr(web_terminal, "pending_append_request"):
|
||
web_terminal.pending_append_request = {"path": detected_path}
|
||
debug_log(f"直接检测到append起始标记，构建追加缓冲: {detected_path}")
|
||
# 检查是否立即包含结束标记
|
||
if pending_append["buffer"]:
|
||
end_index = pending_append["buffer"].find(pending_append["end_marker"], pending_append["content_start"])
|
||
if end_index != -1:
|
||
pending_append["end_index"] = end_index
|
||
append_break_triggered = True
|
||
debug_log("检测到<<<END_APPEND>>>，即将终止流式输出并写入文件")
|
||
break
|
||
continue
|
||
|
||
if not text_started:
|
||
text_started = True
|
||
text_streaming = True
|
||
sender('text_start', {})
|
||
brief_log("模型输出了内容")
|
||
await asyncio.sleep(0.05)
|
||
|
||
if not pending_append:
|
||
full_response += content
|
||
accumulated_response += content
|
||
text_has_content = True
|
||
emit_time = time.time()
|
||
elapsed = 0.0 if last_text_chunk_time is None else emit_time - last_text_chunk_time
|
||
last_text_chunk_time = emit_time
|
||
text_chunk_index += 1
|
||
log_backend_chunk(
|
||
conversation_id,
|
||
current_iteration,
|
||
text_chunk_index,
|
||
elapsed,
|
||
len(content),
|
||
content[:32]
|
||
)
|
||
sender('text_chunk', {
|
||
'content': content,
|
||
'index': text_chunk_index,
|
||
'elapsed': elapsed
|
||
})
|
||
|
||
# 收集工具调用 - 实时发送准备状态
|
||
if "tool_calls" in delta:
|
||
tool_chunks += 1
|
||
for tc in delta["tool_calls"]:
|
||
found = False
|
||
for existing in tool_calls:
|
||
if existing.get("index") == tc.get("index"):
|
||
if "function" in tc and "arguments" in tc["function"]:
|
||
arg_chunk = tc["function"]["arguments"]
|
||
existing_fn = existing.get("function", {})
|
||
existing_args = existing_fn.get("arguments", "")
|
||
existing_fn["arguments"] = (existing_args or "") + arg_chunk
|
||
existing["function"] = existing_fn
|
||
|
||
combined_args = existing_fn.get("arguments", "")
|
||
tool_id = existing.get("id") or tc.get("id")
|
||
tool_name = (
|
||
existing_fn.get("name")
|
||
or tc.get("function", {}).get("name", "")
|
||
)
|
||
intent_value = extract_intent_from_partial(combined_args)
|
||
if (
|
||
intent_value
|
||
and tool_id
|
||
and detected_tool_intent.get(tool_id) != intent_value
|
||
):
|
||
detected_tool_intent[tool_id] = intent_value
|
||
brief_log(f"[intent] 增量提取 {tool_name}: {intent_value}")
|
||
sender('tool_intent', {
|
||
'id': tool_id,
|
||
'name': tool_name,
|
||
'intent': intent_value,
|
||
'conversation_id': conversation_id
|
||
})
|
||
debug_log(f" 发送工具意图: {tool_name} -> {intent_value}")
|
||
await asyncio.sleep(0.01)
|
||
found = True
|
||
break
|
||
|
||
if not found and tc.get("id"):
|
||
tool_id = tc["id"]
|
||
tool_name = tc.get("function", {}).get("name", "")
|
||
arguments_str = tc.get("function", {}).get("arguments", "") or ""
|
||
|
||
# 新工具检测到，立即发送准备事件
|
||
if tool_id not in detected_tools and tool_name:
|
||
detected_tools[tool_id] = tool_name
|
||
|
||
# 尝试提前提取 intent
|
||
intent_value = None
|
||
if arguments_str:
|
||
intent_value = extract_intent_from_partial(arguments_str)
|
||
if intent_value:
|
||
detected_tool_intent[tool_id] = intent_value
|
||
brief_log(f"[intent] 预提取 {tool_name}: {intent_value}")
|
||
|
||
# 立即发送工具准备中事件
|
||
brief_log(f"[tool] 准备调用 {tool_name} (id={tool_id}) intent={intent_value or '-'}")
|
||
sender('tool_preparing', {
|
||
'id': tool_id,
|
||
'name': tool_name,
|
||
'message': f'准备调用 {tool_name}...',
|
||
'intent': intent_value,
|
||
'conversation_id': conversation_id
|
||
})
|
||
debug_log(f" 发送工具准备事件: {tool_name}")
|
||
await asyncio.sleep(0.1)
|
||
|
||
tool_calls.append({
|
||
"id": tool_id,
|
||
"index": tc.get("index"),
|
||
"type": "function",
|
||
"function": {
|
||
"name": tool_name,
|
||
"arguments": arguments_str
|
||
}
|
||
})
|
||
# 尝试从增量参数中抽取 intent，并单独推送
|
||
if tool_id and arguments_str:
|
||
intent_value = extract_intent_from_partial(arguments_str)
|
||
if intent_value and detected_tool_intent.get(tool_id) != intent_value:
|
||
detected_tool_intent[tool_id] = intent_value
|
||
sender('tool_intent', {
|
||
'id': tool_id,
|
||
'name': tool_name,
|
||
'intent': intent_value,
|
||
'conversation_id': conversation_id
|
||
})
|
||
debug_log(f" 发送工具意图: {tool_name} -> {intent_value}")
|
||
await asyncio.sleep(0.01)
|
||
debug_log(f" 新工具: {tool_name}")
|
||
|
||
# 检查是否被停止
|
||
client_stop_info = get_stop_flag(client_sid, username)
|
||
if client_stop_info:
|
||
stop_requested = client_stop_info.get('stop', False) if isinstance(client_stop_info, dict) else client_stop_info
|
||
if stop_requested:
|
||
debug_log("任务在流处理完成后检测到停止状态")
|
||
sender('task_stopped', {
|
||
'message': '命令执行被用户取消',
|
||
'reason': 'user_stop'
|
||
})
|
||
cancel_pending_tools(tool_calls_list=tool_calls, sender=sender, messages=messages)
|
||
clear_stop_flag(client_sid, username)
|
||
return
|
||
|
||
# === API响应完成后只计算输出token ===
|
||
if last_usage_payload:
|
||
try:
|
||
web_terminal.context_manager.apply_usage_statistics(last_usage_payload)
|
||
debug_log(
|
||
f"Usage统计: prompt={last_usage_payload.get('prompt_tokens', 0)}, "
|
||
f"completion={last_usage_payload.get('completion_tokens', 0)}, "
|
||
f"total={last_usage_payload.get('total_tokens', 0)}"
|
||
)
|
||
except Exception as e:
|
||
debug_log(f"Usage统计更新失败: {e}")
|
||
else:
|
||
debug_log("未获取到usage字段，跳过token统计更新")
|
||
|
||
|
||
if api_error:
|
||
try:
|
||
debug_log(f"API错误原始数据: {json.dumps(api_error, ensure_ascii=False)}")
|
||
except Exception:
|
||
debug_log(f"API错误原始数据(不可序列化): {repr(api_error)}")
|
||
error_message = ""
|
||
error_status = None
|
||
error_type = None
|
||
error_code = None
|
||
error_text = ""
|
||
request_dump = None
|
||
error_base_url = None
|
||
error_model_id = None
|
||
if isinstance(api_error, dict):
|
||
error_status = api_error.get("status_code")
|
||
error_type = api_error.get("error_type") or api_error.get("type")
|
||
error_code = api_error.get("error_code") or api_error.get("code")
|
||
error_text = api_error.get("error_text") or ""
|
||
error_message = (
|
||
api_error.get("error_message")
|
||
or api_error.get("message")
|
||
or error_text
|
||
or ""
|
||
)
|
||
request_dump = api_error.get("request_dump")
|
||
error_base_url = api_error.get("base_url")
|
||
error_model_id = api_error.get("model_id")
|
||
elif isinstance(api_error, str):
|
||
error_message = api_error
|
||
if not error_message:
|
||
if error_status:
|
||
error_message = f"API 请求失败（HTTP {error_status}）"
|
||
else:
|
||
error_message = "API 请求失败"
|
||
# 若命中阿里云配额错误，立即写入状态并切换到官方 API
|
||
try:
|
||
from utils.aliyun_fallback import compute_disabled_until, set_disabled_until
|
||
disabled_until, reason = compute_disabled_until(error_message)
|
||
if disabled_until and reason:
|
||
set_disabled_until(getattr(web_terminal, "model_key", None) or "kimi-k2.5", disabled_until, reason)
|
||
profile = get_model_profile(getattr(web_terminal, "model_key", None) or "kimi-k2.5")
|
||
web_terminal.apply_model_profile(profile)
|
||
except Exception as exc:
|
||
debug_log(f"处理阿里云配额回退失败: {exc}")
|
||
can_retry = (
|
||
api_attempt < max_api_retries
|
||
and not full_response
|
||
and not tool_calls
|
||
and not current_thinking
|
||
and not pending_append
|
||
and not pending_modify
|
||
)
|
||
sender('error', {
|
||
'message': error_message,
|
||
'status_code': error_status,
|
||
'error_type': error_type,
|
||
'error_code': error_code,
|
||
'error_text': error_text,
|
||
'request_dump': request_dump,
|
||
'base_url': error_base_url,
|
||
'model_id': error_model_id,
|
||
'retry': bool(can_retry),
|
||
'retry_in': retry_delay_seconds if can_retry else None,
|
||
'attempt': api_attempt + 1,
|
||
'max_attempts': max_api_retries + 1
|
||
})
|
||
if can_retry:
|
||
try:
|
||
profile = get_model_profile(getattr(web_terminal, "model_key", None) or "kimi-k2.5")
|
||
web_terminal.apply_model_profile(profile)
|
||
except Exception as exc:
|
||
debug_log(f"重试前更新模型配置失败: {exc}")
|
||
cancelled = await wait_retry_delay(delay_seconds=retry_delay_seconds, client_sid=client_sid, username=username, sender=sender, get_stop_flag=get_stop_flag, clear_stop_flag=clear_stop_flag)
|
||
if cancelled:
|
||
return
|
||
continue
|
||
cancel_pending_tools(tool_calls_list=tool_calls, sender=sender, messages=messages)
|
||
return
|
||
break
|
||
|
||
# 流结束后的处理
|
||
debug_log(f"\n流结束统计:")
|
||
debug_log(f" 总chunks: {chunk_count}")
|
||
debug_log(f" 思考chunks: {reasoning_chunks}")
|
||
debug_log(f" 内容chunks: {content_chunks}")
|
||
debug_log(f" 工具chunks: {tool_chunks}")
|
||
debug_log(f" 收集到的思考: {len(current_thinking)} 字符")
|
||
debug_log(f" 收集到的正文: {len(full_response)} 字符")
|
||
debug_log(f" 收集到的工具: {len(tool_calls)} 个")
|
||
|
||
if not append_result["handled"] and pending_append:
|
||
append_result, pending_append, append_probe_buffer = await finalize_pending_append(pending_append=pending_append, append_probe_buffer=append_probe_buffer, response_text=full_response, stream_completed=True, finish_reason=last_finish_reason, web_terminal=web_terminal, sender=sender, debug_log=debug_log)
|
||
if not modify_result["handled"] and pending_modify:
|
||
modify_result, pending_modify, modify_probe_buffer = await finalize_pending_modify(pending_modify=pending_modify, modify_probe_buffer=modify_probe_buffer, response_text=full_response, stream_completed=True, finish_reason=last_finish_reason, web_terminal=web_terminal, sender=sender, debug_log=debug_log)
|
||
|
||
# 结束未完成的流
|
||
if in_thinking and not thinking_ended:
|
||
sender('thinking_end', {'full_content': current_thinking})
|
||
await asyncio.sleep(0.1)
|
||
|
||
|
||
# 确保text_end事件被发送
|
||
if text_started and text_has_content and not append_result["handled"] and not modify_result["handled"]:
|
||
debug_log(f"发送text_end事件，完整内容长度: {len(full_response)}")
|
||
sender('text_end', {'full_content': full_response})
|
||
await asyncio.sleep(0.1)
|
||
text_streaming = False
|
||
|
||
if full_response.strip():
|
||
debug_log(f"流式文本内容长度: {len(full_response)} 字符")
|
||
|
||
if append_result["handled"]:
|
||
append_metadata = append_result.get("assistant_metadata")
|
||
append_content_text = append_result.get("assistant_content")
|
||
if append_content_text:
|
||
web_terminal.context_manager.add_conversation(
|
||
"assistant",
|
||
append_content_text,
|
||
metadata=append_metadata
|
||
)
|
||
debug_log("💾 增量保存：追加正文快照")
|
||
|
||
payload_info = append_metadata.get("append_payload") if append_metadata else {}
|
||
sender('append_payload', {
|
||
'path': payload_info.get("path") or append_result.get("path"),
|
||
'forced': payload_info.get("forced", False),
|
||
'lines': payload_info.get("lines"),
|
||
'bytes': payload_info.get("bytes"),
|
||
'tool_call_id': payload_info.get("tool_call_id") or append_result.get("tool_call_id"),
|
||
'success': payload_info.get("success", append_result.get("success", False)),
|
||
'conversation_id': conversation_id
|
||
})
|
||
|
||
if append_result["tool_content"]:
|
||
tool_call_id = append_result.get("tool_call_id") or f"append_{int(time.time() * 1000)}"
|
||
system_notice = format_tool_result_notice("append_to_file", tool_call_id, append_result["tool_content"])
|
||
web_terminal.context_manager.add_conversation("system", system_notice)
|
||
append_result["tool_call_id"] = tool_call_id
|
||
debug_log("💾 增量保存：append_to_file 工具结果（system 通知）")
|
||
|
||
finish_reason = append_result.get("finish_reason")
|
||
path_for_prompt = append_result.get("path")
|
||
need_follow_prompt = (
|
||
finish_reason == "length" or
|
||
append_result.get("forced") or
|
||
not append_result.get("success")
|
||
)
|
||
|
||
if need_follow_prompt and path_for_prompt:
|
||
prompt_lines = [
|
||
f"append_to_file 在处理 {path_for_prompt} 时未完成，需要重新发起写入。"
|
||
]
|
||
if finish_reason == "length":
|
||
prompt_lines.append(
|
||
"上一次输出达到系统单次输出上限，已写入的内容已保存。"
|
||
)
|
||
if append_result.get("forced"):
|
||
prompt_lines.append(
|
||
"收到的内容缺少 <<<END_APPEND>>> 标记，系统依据流式结束位置落盘。"
|
||
)
|
||
if not append_result.get("success"):
|
||
prompt_lines.append("系统未能识别有效的追加标记。")
|
||
prompt_lines.append(
|
||
"请再次调用 append_to_file 工具获取新的写入窗口，并在工具调用的输出中遵循以下格式："
|
||
)
|
||
prompt_lines.append(f"<<<APPEND:{path_for_prompt}>>>")
|
||
prompt_lines.append("...填写剩余正文，如内容已完成可留空...")
|
||
prompt_lines.append("<<<END_APPEND>>>")
|
||
prompt_lines.append("不要在普通回复中粘贴上述标记，必须通过 append_to_file 工具发送。")
|
||
follow_prompt = "\n".join(prompt_lines)
|
||
messages.append({
|
||
"role": "system",
|
||
"content": follow_prompt
|
||
})
|
||
web_terminal.context_manager.add_conversation("system", follow_prompt)
|
||
debug_log("已注入追加任务提示")
|
||
|
||
if append_result["handled"] and append_result.get("forced") and append_result.get("success"):
|
||
mark_force_thinking(web_terminal, reason="append_forced_finish")
|
||
if append_result["handled"] and not append_result.get("success"):
|
||
sender('system_message', {
|
||
'content': f'⚠️ 追加写入失败：{append_result.get("error")}'
|
||
})
|
||
maybe_mark_failure_from_message(web_terminal, f'⚠️ 追加写入失败：{append_result.get("error")}')
|
||
mark_force_thinking(web_terminal, reason="append_failed")
|
||
|
||
if modify_result["handled"]:
|
||
modify_metadata = modify_result.get("assistant_metadata")
|
||
modify_content_text = modify_result.get("assistant_content")
|
||
if modify_content_text:
|
||
web_terminal.context_manager.add_conversation(
|
||
"assistant",
|
||
modify_content_text,
|
||
metadata=modify_metadata
|
||
)
|
||
debug_log("💾 增量保存：修改正文快照")
|
||
|
||
payload_info = modify_metadata.get("modify_payload") if modify_metadata else {}
|
||
sender('modify_payload', {
|
||
'path': payload_info.get("path") or modify_result.get("path"),
|
||
'total': payload_info.get("total_blocks") or modify_result.get("total_blocks"),
|
||
'completed': payload_info.get("completed") or modify_result.get("completed_blocks"),
|
||
'failed': payload_info.get("failed") or modify_result.get("failed_blocks"),
|
||
'forced': payload_info.get("forced", modify_result.get("forced", False)),
|
||
'success': modify_result.get("success", False),
|
||
'conversation_id': conversation_id
|
||
})
|
||
|
||
if modify_result["tool_content"]:
|
||
tool_call_id = modify_result.get("tool_call_id") or f"modify_{int(time.time() * 1000)}"
|
||
system_notice = format_tool_result_notice("modify_file", tool_call_id, modify_result["tool_content"])
|
||
web_terminal.context_manager.add_conversation("system", system_notice)
|
||
modify_result["tool_call_id"] = tool_call_id
|
||
debug_log("💾 增量保存：modify_file 工具结果（system 通知）")
|
||
|
||
path_for_prompt = modify_result.get("path")
|
||
failed_blocks = modify_result.get("failed_blocks") or []
|
||
need_follow_prompt = modify_result.get("forced") or bool(failed_blocks)
|
||
|
||
if need_follow_prompt and path_for_prompt:
|
||
prompt_lines = [
|
||
f"modify_file 在处理 {path_for_prompt} 时未完成，需要重新发起补丁。"
|
||
]
|
||
if modify_result.get("forced"):
|
||
prompt_lines.append(
|
||
"刚才的内容缺少 <<<END_MODIFY>>> 标记，系统仅应用了已识别的部分。"
|
||
)
|
||
if failed_blocks:
|
||
failed_text = "、".join(str(idx) for idx in failed_blocks)
|
||
prompt_lines.append(f"以下补丁未成功：第 {failed_text} 处。")
|
||
prompt_lines.append(
|
||
"请再次调用 modify_file 工具，并在新的工具调用中按以下模板提供完整补丁："
|
||
)
|
||
prompt_lines.append(f"<<<MODIFY:{path_for_prompt}>>>")
|
||
prompt_lines.append("[replace:序号]")
|
||
prompt_lines.append("<<OLD>>")
|
||
prompt_lines.append("...原文（必须逐字匹配，包含全部缩进、空格和换行）...")
|
||
prompt_lines.append("<<END>>")
|
||
prompt_lines.append("<<NEW>>")
|
||
prompt_lines.append("...新内容，可留空表示清空，注意保持结构完整...")
|
||
prompt_lines.append("<<END>>")
|
||
prompt_lines.append("[/replace]")
|
||
prompt_lines.append("<<<END_MODIFY>>>")
|
||
prompt_lines.append("请勿在普通回复中直接粘贴补丁，必须通过 modify_file 工具发送。")
|
||
follow_prompt = "\n".join(prompt_lines)
|
||
messages.append({
|
||
"role": "system",
|
||
"content": follow_prompt
|
||
})
|
||
web_terminal.context_manager.add_conversation("system", follow_prompt)
|
||
debug_log("已注入修改任务提示")
|
||
|
||
if modify_result["handled"] and modify_result.get("failed_blocks"):
|
||
mark_force_thinking(web_terminal, reason="modify_partial_failure")
|
||
if modify_result["handled"] and modify_result.get("forced") and modify_result.get("success"):
|
||
mark_force_thinking(web_terminal, reason="modify_forced_finish")
|
||
if modify_result["handled"] and not modify_result.get("success"):
|
||
error_message = modify_result.get("summary_message") or modify_result.get("error") or "修改操作未成功，请根据提示重新执行。"
|
||
sender('system_message', {
|
||
'content': f'⚠️ 修改操作存在未完成的内容：{error_message}'
|
||
})
|
||
maybe_mark_failure_from_message(web_terminal, f'⚠️ 修改操作存在未完成的内容：{error_message}')
|
||
mark_force_thinking(web_terminal, reason="modify_failed")
|
||
|
||
if web_terminal.api_client.last_call_used_thinking and current_thinking:
|
||
web_terminal.api_client.current_task_thinking = current_thinking or ""
|
||
if web_terminal.api_client.current_task_first_call:
|
||
web_terminal.api_client.current_task_first_call = False
|
||
update_thinking_after_call(web_terminal)
|
||
|
||
# 检测是否有格式错误的工具调用
|
||
if not tool_calls and full_response and AUTO_FIX_TOOL_CALL and not append_result["handled"] and not modify_result["handled"]:
|
||
if detect_malformed_tool_call(full_response):
|
||
auto_fix_attempts += 1
|
||
|
||
if auto_fix_attempts <= AUTO_FIX_MAX_ATTEMPTS:
|
||
debug_log(f"检测到格式错误的工具调用，尝试自动修复 (尝试 {auto_fix_attempts}/{AUTO_FIX_MAX_ATTEMPTS})")
|
||
|
||
fix_message = "你使用了错误的格式输出工具调用。请使用正确的工具调用格式而不是直接输出JSON。根据当前进度继续执行任务。"
|
||
|
||
sender('system_message', {
|
||
'content': f'⚠️ 自动修复: {fix_message}'
|
||
})
|
||
maybe_mark_failure_from_message(web_terminal, f'⚠️ 自动修复: {fix_message}')
|
||
|
||
messages.append({
|
||
"role": "user",
|
||
"content": fix_message
|
||
})
|
||
|
||
await asyncio.sleep(1)
|
||
continue
|
||
else:
|
||
debug_log(f"自动修复尝试已达上限 ({AUTO_FIX_MAX_ATTEMPTS})")
|
||
sender('system_message', {
|
||
'content': f'⌘ 工具调用格式错误，自动修复失败。请手动检查并重试。'
|
||
})
|
||
maybe_mark_failure_from_message(web_terminal, '⌘ 工具调用格式错误，自动修复失败。请手动检查并重试。')
|
||
break
|
||
|
||
# 构建助手消息（用于API继续对话）
|
||
assistant_content_parts = []
|
||
|
||
if full_response:
|
||
assistant_content_parts.append(full_response)
|
||
elif append_result["handled"] and append_result["assistant_content"]:
|
||
assistant_content_parts.append(append_result["assistant_content"])
|
||
elif modify_result["handled"] and modify_result.get("assistant_content"):
|
||
assistant_content_parts.append(modify_result["assistant_content"])
|
||
|
||
assistant_content = "\n".join(assistant_content_parts) if assistant_content_parts else ""
|
||
|
||
# 添加到消息历史（用于API继续对话，不保存到文件）
|
||
assistant_message = {
|
||
"role": "assistant",
|
||
"content": assistant_content,
|
||
"tool_calls": tool_calls
|
||
}
|
||
if current_thinking:
|
||
assistant_message["reasoning_content"] = current_thinking
|
||
|
||
messages.append(assistant_message)
|
||
if assistant_content or current_thinking or tool_calls:
|
||
web_terminal.context_manager.add_conversation(
|
||
"assistant",
|
||
assistant_content,
|
||
tool_calls=tool_calls if tool_calls else None,
|
||
reasoning_content=current_thinking or None
|
||
)
|
||
|
||
# 为下一轮迭代重置流状态标志，但保留 full_response 供上面保存使用
|
||
text_streaming = False
|
||
text_started = False
|
||
text_has_content = False
|
||
full_response = ""
|
||
|
||
if append_result["handled"] and append_result.get("tool_content"):
|
||
tool_call_id = append_result.get("tool_call_id") or f"append_{int(time.time() * 1000)}"
|
||
system_notice = format_tool_result_notice("append_to_file", tool_call_id, append_result["tool_content"])
|
||
messages.append({
|
||
"role": "system",
|
||
"content": system_notice
|
||
})
|
||
append_result["tool_call_id"] = tool_call_id
|
||
debug_log("已将 append_to_file 工具结果以 system 形式追加到对话上下文")
|
||
if modify_result["handled"] and modify_result.get("tool_content"):
|
||
tool_call_id = modify_result.get("tool_call_id") or f"modify_{int(time.time() * 1000)}"
|
||
system_notice = format_tool_result_notice("modify_file", tool_call_id, modify_result["tool_content"])
|
||
messages.append({
|
||
"role": "system",
|
||
"content": system_notice
|
||
})
|
||
modify_result["tool_call_id"] = tool_call_id
|
||
debug_log("已将 modify_file 工具结果以 system 形式追加到对话上下文")
|
||
|
||
force_continue = append_result["handled"] or modify_result["handled"]
|
||
if force_continue:
|
||
if append_result["handled"]:
|
||
debug_log("append_to_file 已处理，继续下一轮以让模型返回确认回复")
|
||
elif modify_result["handled"]:
|
||
debug_log("modify_file 已处理，继续下一轮以让模型返回确认回复")
|
||
else:
|
||
debug_log("补丁处理完成，继续下一轮以获取模型回复")
|
||
continue
|
||
|
||
if not tool_calls:
|
||
debug_log("没有工具调用，结束迭代")
|
||
break
|
||
|
||
# 检查连续相同工具调用
|
||
for tc in tool_calls:
|
||
tool_name = tc["function"]["name"]
|
||
|
||
if tool_name == last_tool_name:
|
||
consecutive_same_tool[tool_name] += 1
|
||
|
||
if (
|
||
MAX_CONSECUTIVE_SAME_TOOL is not None
|
||
and consecutive_same_tool[tool_name] >= MAX_CONSECUTIVE_SAME_TOOL
|
||
):
|
||
debug_log(f"警告: 连续调用相同工具 {tool_name} 已达 {MAX_CONSECUTIVE_SAME_TOOL} 次")
|
||
sender('system_message', {
|
||
'content': f'⚠️ 检测到重复调用 {tool_name} 工具 {MAX_CONSECUTIVE_SAME_TOOL} 次，可能存在循环。'
|
||
})
|
||
maybe_mark_failure_from_message(web_terminal, f'⚠️ 检测到重复调用 {tool_name} 工具 {MAX_CONSECUTIVE_SAME_TOOL} 次，可能存在循环。')
|
||
|
||
if consecutive_same_tool[tool_name] >= MAX_CONSECUTIVE_SAME_TOOL + 2:
|
||
debug_log(f"终止: 工具 {tool_name} 调用次数过多")
|
||
sender('system_message', {
|
||
'content': f'⌘ 工具 {tool_name} 重复调用过多，任务终止。'
|
||
})
|
||
maybe_mark_failure_from_message(web_terminal, f'⌘ 工具 {tool_name} 重复调用过多，任务终止。')
|
||
break
|
||
else:
|
||
consecutive_same_tool.clear()
|
||
consecutive_same_tool[tool_name] = 1
|
||
|
||
last_tool_name = tool_name
|
||
|
||
# 更新统计
|
||
total_tool_calls += len(tool_calls)
|
||
|
||
# 执行每个工具
|
||
tool_loop_result = await execute_tool_calls(
|
||
web_terminal=web_terminal,
|
||
tool_calls=tool_calls,
|
||
sender=sender,
|
||
messages=messages,
|
||
client_sid=client_sid,
|
||
username=username,
|
||
iteration=iteration,
|
||
conversation_id=conversation_id,
|
||
last_tool_call_time=last_tool_call_time,
|
||
process_sub_agent_updates=process_sub_agent_updates,
|
||
maybe_mark_failure_from_message=maybe_mark_failure_from_message,
|
||
mark_force_thinking=mark_force_thinking,
|
||
get_stop_flag=get_stop_flag,
|
||
clear_stop_flag=clear_stop_flag,
|
||
)
|
||
last_tool_call_time = tool_loop_result.get("last_tool_call_time", last_tool_call_time)
|
||
if tool_loop_result.get("stopped"):
|
||
return
|
||
|
||
# 标记不再是第一次迭代
|
||
is_first_iteration = False
|
||
|
||
|
||
# 最终统计
|
||
debug_log(f"\n{'='*40}")
|
||
debug_log(f"任务完成统计:")
|
||
debug_log(f" 总迭代次数: {total_iterations}")
|
||
debug_log(f" 总工具调用: {total_tool_calls}")
|
||
debug_log(f" 自动修复尝试: {auto_fix_attempts}")
|
||
debug_log(f" 累积响应: {len(accumulated_response)} 字符")
|
||
debug_log(f"{'='*40}\n")
|
||
|
||
# 发送完成事件
|
||
sender('task_complete', {
|
||
'total_iterations': total_iterations,
|
||
'total_tool_calls': total_tool_calls,
|
||
'auto_fix_attempts': auto_fix_attempts
|
||
})
|