diff --git a/modules/personalization_manager.py b/modules/personalization_manager.py
index 6f10471..4f2a8c1 100644
--- a/modules/personalization_manager.py
+++ b/modules/personalization_manager.py
@@ -23,6 +23,19 @@ MAX_CONSIDERATION_ITEMS = 10
TONE_PRESETS = ["健谈", "幽默", "直言不讳", "鼓励性", "诗意", "企业商务", "打破常规", "同理心"]
THINKING_INTERVAL_MIN = 1
THINKING_INTERVAL_MAX = 50
+DEFAULT_SHALLOW_COMPRESS_TRIGGER_TOKENS = 80_000
+DEFAULT_SHALLOW_COMPRESS_KEEP_RECENT_TOOLS = 15
+DEFAULT_SHALLOW_COMPRESS_MAX_REPLACE_PER_ROUND = 10
+DEFAULT_SHALLOW_COMPRESS_TRIGGER_TOOL_CALLS_INTERVAL = 10
+DEFAULT_DEEP_COMPRESS_TRIGGER_TOKENS = 150_000
+MIN_COMPRESSION_TRIGGER_TOKENS = 1_000
+MAX_COMPRESSION_TRIGGER_TOKENS = 2_000_000
+MIN_SHALLOW_KEEP_RECENT_TOOLS = 0
+MAX_SHALLOW_KEEP_RECENT_TOOLS = 500
+MIN_SHALLOW_MAX_REPLACE_PER_ROUND = 1
+MAX_SHALLOW_MAX_REPLACE_PER_ROUND = 200
+MIN_SHALLOW_TRIGGER_TOOL_CALLS_INTERVAL = 1
+MAX_SHALLOW_TRIGGER_TOOL_CALLS_INTERVAL = 200
DEFAULT_PERSONALIZATION_CONFIG: Dict[str, Any] = {
"enabled": False,
@@ -44,6 +57,11 @@ DEFAULT_PERSONALIZATION_CONFIG: Dict[str, Any] = {
"image_compression": "original", # original / 1080p / 720p / 540p
"auto_shallow_compress_enabled": False,
"auto_deep_compress_enabled": False,
+ "shallow_compress_trigger_tokens": None,
+ "shallow_compress_keep_recent_tools": None,
+ "shallow_compress_max_replace_per_round": None,
+ "shallow_compress_trigger_tool_calls_interval": None,
+ "deep_compress_trigger_tokens": None,
"silent_tool_disable": False, # 禁用工具时不向模型插入提示
"enhanced_tool_display": True, # 增强工具显示
}
@@ -58,6 +76,8 @@ __all__ = [
"ensure_personalization_config",
"build_personalization_prompt",
"sanitize_personalization_payload",
+ "resolve_context_compression_settings",
+ "validate_context_compression_settings",
]
@@ -198,6 +218,71 @@ def sanitize_personalization_payload(
else:
base["auto_deep_compress_enabled"] = bool(base.get("auto_deep_compress_enabled", False))
+ if "shallow_compress_trigger_tokens" in data:
+ base["shallow_compress_trigger_tokens"] = _sanitize_optional_int(
+ data.get("shallow_compress_trigger_tokens"),
+ min_value=MIN_COMPRESSION_TRIGGER_TOKENS,
+ max_value=MAX_COMPRESSION_TRIGGER_TOKENS,
+ )
+ else:
+ base["shallow_compress_trigger_tokens"] = _sanitize_optional_int(
+ base.get("shallow_compress_trigger_tokens"),
+ min_value=MIN_COMPRESSION_TRIGGER_TOKENS,
+ max_value=MAX_COMPRESSION_TRIGGER_TOKENS,
+ )
+
+ if "shallow_compress_keep_recent_tools" in data:
+ base["shallow_compress_keep_recent_tools"] = _sanitize_optional_int(
+ data.get("shallow_compress_keep_recent_tools"),
+ min_value=MIN_SHALLOW_KEEP_RECENT_TOOLS,
+ max_value=MAX_SHALLOW_KEEP_RECENT_TOOLS,
+ )
+ else:
+ base["shallow_compress_keep_recent_tools"] = _sanitize_optional_int(
+ base.get("shallow_compress_keep_recent_tools"),
+ min_value=MIN_SHALLOW_KEEP_RECENT_TOOLS,
+ max_value=MAX_SHALLOW_KEEP_RECENT_TOOLS,
+ )
+
+ if "shallow_compress_max_replace_per_round" in data:
+ base["shallow_compress_max_replace_per_round"] = _sanitize_optional_int(
+ data.get("shallow_compress_max_replace_per_round"),
+ min_value=MIN_SHALLOW_MAX_REPLACE_PER_ROUND,
+ max_value=MAX_SHALLOW_MAX_REPLACE_PER_ROUND,
+ )
+ else:
+ base["shallow_compress_max_replace_per_round"] = _sanitize_optional_int(
+ base.get("shallow_compress_max_replace_per_round"),
+ min_value=MIN_SHALLOW_MAX_REPLACE_PER_ROUND,
+ max_value=MAX_SHALLOW_MAX_REPLACE_PER_ROUND,
+ )
+
+ if "deep_compress_trigger_tokens" in data:
+ base["deep_compress_trigger_tokens"] = _sanitize_optional_int(
+ data.get("deep_compress_trigger_tokens"),
+ min_value=MIN_COMPRESSION_TRIGGER_TOKENS,
+ max_value=MAX_COMPRESSION_TRIGGER_TOKENS,
+ )
+ else:
+ base["deep_compress_trigger_tokens"] = _sanitize_optional_int(
+ base.get("deep_compress_trigger_tokens"),
+ min_value=MIN_COMPRESSION_TRIGGER_TOKENS,
+ max_value=MAX_COMPRESSION_TRIGGER_TOKENS,
+ )
+
+ if "shallow_compress_trigger_tool_calls_interval" in data:
+ base["shallow_compress_trigger_tool_calls_interval"] = _sanitize_optional_int(
+ data.get("shallow_compress_trigger_tool_calls_interval"),
+ min_value=MIN_SHALLOW_TRIGGER_TOOL_CALLS_INTERVAL,
+ max_value=MAX_SHALLOW_TRIGGER_TOOL_CALLS_INTERVAL,
+ )
+ else:
+ base["shallow_compress_trigger_tool_calls_interval"] = _sanitize_optional_int(
+ base.get("shallow_compress_trigger_tool_calls_interval"),
+ min_value=MIN_SHALLOW_TRIGGER_TOOL_CALLS_INTERVAL,
+ max_value=MAX_SHALLOW_TRIGGER_TOOL_CALLS_INTERVAL,
+ )
+
# 静默禁用工具提示
if "silent_tool_disable" in data:
base["silent_tool_disable"] = bool(data.get("silent_tool_disable"))
@@ -242,6 +327,7 @@ def save_personalization_config(base_dir: PathLike, payload: Dict[str, Any]) ->
"""Persist sanitized personalization config and return it."""
existing = load_personalization_config(base_dir)
config = sanitize_personalization_payload(payload, fallback=existing)
+ validate_context_compression_settings(config)
path = _to_path(base_dir)
_ensure_parent(path)
with open(path, "w", encoding="utf-8") as f:
@@ -249,6 +335,78 @@ def save_personalization_config(base_dir: PathLike, payload: Dict[str, Any]) ->
return config
+def _sanitize_optional_int(value: Any, *, min_value: int, max_value: int) -> Optional[int]:
+ if value is None or value == "":
+ return None
+ if isinstance(value, bool):
+ return None
+ try:
+ parsed = int(value)
+ except (TypeError, ValueError):
+ return None
+ if parsed < min_value:
+ return min_value
+ if parsed > max_value:
+ return max_value
+ return parsed
+
+
+def resolve_context_compression_settings(config: Optional[Dict[str, Any]]) -> Dict[str, int]:
+ data = config or {}
+ shallow_trigger = _sanitize_optional_int(
+ data.get("shallow_compress_trigger_tokens"),
+ min_value=MIN_COMPRESSION_TRIGGER_TOKENS,
+ max_value=MAX_COMPRESSION_TRIGGER_TOKENS,
+ ) or DEFAULT_SHALLOW_COMPRESS_TRIGGER_TOKENS
+ shallow_keep_recent = _sanitize_optional_int(
+ data.get("shallow_compress_keep_recent_tools"),
+ min_value=MIN_SHALLOW_KEEP_RECENT_TOOLS,
+ max_value=MAX_SHALLOW_KEEP_RECENT_TOOLS,
+ )
+ if shallow_keep_recent is None:
+ shallow_keep_recent = DEFAULT_SHALLOW_COMPRESS_KEEP_RECENT_TOOLS
+ shallow_max_replace = _sanitize_optional_int(
+ data.get("shallow_compress_max_replace_per_round"),
+ min_value=MIN_SHALLOW_MAX_REPLACE_PER_ROUND,
+ max_value=MAX_SHALLOW_MAX_REPLACE_PER_ROUND,
+ ) or DEFAULT_SHALLOW_COMPRESS_MAX_REPLACE_PER_ROUND
+ shallow_trigger_interval = _sanitize_optional_int(
+ data.get("shallow_compress_trigger_tool_calls_interval"),
+ min_value=MIN_SHALLOW_TRIGGER_TOOL_CALLS_INTERVAL,
+ max_value=MAX_SHALLOW_TRIGGER_TOOL_CALLS_INTERVAL,
+ ) or DEFAULT_SHALLOW_COMPRESS_TRIGGER_TOOL_CALLS_INTERVAL
+ deep_trigger = _sanitize_optional_int(
+ data.get("deep_compress_trigger_tokens"),
+ min_value=MIN_COMPRESSION_TRIGGER_TOKENS,
+ max_value=MAX_COMPRESSION_TRIGGER_TOKENS,
+ ) or DEFAULT_DEEP_COMPRESS_TRIGGER_TOKENS
+ if deep_trigger <= shallow_trigger:
+ deep_trigger = shallow_trigger + 1
+ return {
+ "shallow_trigger_tokens": shallow_trigger,
+ "shallow_keep_recent_tools": shallow_keep_recent,
+ "shallow_max_replace_per_round": shallow_max_replace,
+ "shallow_trigger_tool_calls_interval": shallow_trigger_interval,
+ "deep_trigger_tokens": deep_trigger,
+ }
+
+
+def validate_context_compression_settings(config: Optional[Dict[str, Any]]) -> None:
+ data = config or {}
+ shallow_trigger = _sanitize_optional_int(
+ data.get("shallow_compress_trigger_tokens"),
+ min_value=MIN_COMPRESSION_TRIGGER_TOKENS,
+ max_value=MAX_COMPRESSION_TRIGGER_TOKENS,
+ ) or DEFAULT_SHALLOW_COMPRESS_TRIGGER_TOKENS
+ deep_trigger = _sanitize_optional_int(
+ data.get("deep_compress_trigger_tokens"),
+ min_value=MIN_COMPRESSION_TRIGGER_TOKENS,
+ max_value=MAX_COMPRESSION_TRIGGER_TOKENS,
+ ) or DEFAULT_DEEP_COMPRESS_TRIGGER_TOKENS
+ if deep_trigger <= shallow_trigger:
+ raise ValueError("深压缩触发上下文必须大于浅压缩触发上下文")
+
+
def build_personalization_prompt(
config: Optional[Dict[str, Any]],
include_header: bool = True
diff --git a/server/chat_flow_task_main.py b/server/chat_flow_task_main.py
index a8c44cf..6542b46 100644
--- a/server/chat_flow_task_main.py
+++ b/server/chat_flow_task_main.py
@@ -38,6 +38,7 @@ from modules.personalization_manager import (
save_personalization_config,
THINKING_INTERVAL_MIN,
THINKING_INTERVAL_MAX,
+ resolve_context_compression_settings,
)
from modules.skill_hint_manager import SkillHintManager
from modules.upload_security import UploadSecurityError
@@ -566,11 +567,12 @@ async def handle_task_with_sender(terminal: WebTerminal, workspace: UserWorkspac
personal_config = load_personalization_config(workspace.data_dir)
except Exception:
personal_config = {}
+ compression_settings = resolve_context_compression_settings(personal_config)
auto_deep_enabled = bool(personal_config.get("auto_deep_compress_enabled", False))
current_tokens_for_deep = web_terminal.context_manager.get_current_context_tokens(conversation_id)
if (
auto_deep_enabled
- and current_tokens_for_deep > 150_000
+ and current_tokens_for_deep > compression_settings["deep_trigger_tokens"]
and not web_terminal.context_manager.is_compression_in_progress()
):
web_terminal.context_manager._set_meta_flag("is_ultra_long_conversation", True)
diff --git a/server/chat_flow_tool_loop.py b/server/chat_flow_tool_loop.py
index 1248475..5d30e1f 100644
--- a/server/chat_flow_tool_loop.py
+++ b/server/chat_flow_tool_loop.py
@@ -12,8 +12,9 @@ from .security import compact_web_search_result
from .chat_flow_helpers import detect_tool_failure
from .chat_flow_runner_helpers import resolve_monitor_path, resolve_monitor_memory, capture_monitor_snapshot
from utils.tool_result_formatter import format_tool_result_for_context
+from utils.context_manager import AUTO_SHALLOW_PLACEHOLDER
from config import TOOL_CALL_COOLDOWN
-from modules.personalization_manager import load_personalization_config
+from modules.personalization_manager import load_personalization_config, resolve_context_compression_settings
from .deep_compression import run_deep_compression
@@ -457,12 +458,21 @@ async def execute_tool_calls(*, web_terminal, tool_calls, sender, messages, clie
personal_config = load_personalization_config(workspace_data_dir) if workspace_data_dir else {}
except Exception:
personal_config = {}
+ compression_settings = resolve_context_compression_settings(personal_config)
auto_shallow_enabled = bool(personal_config.get("auto_shallow_compress_enabled", False))
auto_deep_enabled = bool(personal_config.get("auto_deep_compress_enabled", False))
compressed_count = 0
try:
compressed_count = int(
- web_terminal.context_manager.on_tool_call_finished(function_name, enable_shallow=auto_shallow_enabled) or 0
+ web_terminal.context_manager.on_tool_call_finished(
+ function_name,
+ enable_shallow=auto_shallow_enabled,
+ shallow_trigger_tokens=compression_settings["shallow_trigger_tokens"],
+ deep_trigger_tokens=compression_settings["deep_trigger_tokens"],
+ shallow_batch_size=compression_settings["shallow_max_replace_per_round"],
+ shallow_keep_recent_tools=compression_settings["shallow_keep_recent_tools"],
+ shallow_trigger_tool_calls_interval=compression_settings["shallow_trigger_tool_calls_interval"],
+ ) or 0
)
except Exception as exc:
debug_log(f"[ContextCompression] 工具后浅压缩检测失败: {exc}")
@@ -470,8 +480,34 @@ async def execute_tool_calls(*, web_terminal, tool_calls, sender, messages, clie
sender('shallow_compression', {
"conversation_id": conversation_id,
"compressed_count": compressed_count,
- "keep_recent_tools": 15,
+ "keep_recent_tools": compression_settings["shallow_keep_recent_tools"],
})
+ # 关键修复：同一轮工具循环里，API 继续调用使用的是本地 messages（不是每次重建上下文），
+ # 因此需要把已打标的旧 tool 结果同步替换为占位符，避免日志看起来“弹窗触发但请求未替换”。
+ if auto_shallow_enabled and messages:
+ try:
+ marked_keys = set()
+ for item in (web_terminal.context_manager.conversation_history or []):
+ if not isinstance(item, dict) or item.get("role") != "tool":
+ continue
+ item_meta = item.get("metadata") or {}
+ if not item_meta.get("auto_shallow_compacted"):
+ continue
+ marked_keys.add((str(item.get("tool_call_id") or ""), str(item.get("name") or "")))
+
+ if marked_keys:
+ replaced_in_loop = 0
+ for msg in messages:
+ if not isinstance(msg, dict) or msg.get("role") != "tool":
+ continue
+ key = (str(msg.get("tool_call_id") or ""), str(msg.get("name") or ""))
+ if key in marked_keys and msg.get("content") != AUTO_SHALLOW_PLACEHOLDER:
+ msg["content"] = AUTO_SHALLOW_PLACEHOLDER
+ replaced_in_loop += 1
+ if replaced_in_loop > 0:
+ debug_log(f"[ContextCompression] 同步替换本轮 messages 中已压缩 tool 结果: {replaced_in_loop}")
+ except Exception as exc:
+ debug_log(f"[ContextCompression] 同步替换本轮 messages 失败: {exc}")
debug_log(f"💾 增量保存：工具结果 {function_name}")
if function_name != "wait_sub_agent":
system_message = result_data.get("system_message") if isinstance(result_data, dict) else None
@@ -483,7 +519,7 @@ async def execute_tool_calls(*, web_terminal, tool_calls, sender, messages, clie
current_context_tokens = web_terminal.context_manager.get_current_context_tokens(conversation_id)
if (
auto_deep_enabled
- and current_context_tokens > 150_000
+ and current_context_tokens > compression_settings["deep_trigger_tokens"]
and not web_terminal.context_manager.is_compression_in_progress()
):
web_terminal.context_manager._set_meta_flag("is_ultra_long_conversation", True)
diff --git a/static/src/App.vue b/static/src/App.vue
index fb5a925..bbf9d98 100644
--- a/static/src/App.vue
+++ b/static/src/App.vue
@@ -250,7 +250,7 @@
:tool-settings="toolSettings"
:tool-settings-loading="toolSettingsLoading"
:settings-open="settingsOpen"
- :compressing="compressing"
+ :compressing="compressing || compressionInProgress"
:current-conversation-id="currentConversationId"
:icon-style="iconStyle"
:tool-category-icon="toolCategoryIcon"
diff --git a/static/src/app/methods/taskPolling.ts b/static/src/app/methods/taskPolling.ts
index bdb22f3..a1f7d0f 100644
--- a/static/src/app/methods/taskPolling.ts
+++ b/static/src/app/methods/taskPolling.ts
@@ -51,7 +51,12 @@ export const taskPollingMethods = {
// 检查事件的 conversation_id 是否匹配当前对话
// 如果不匹配，忽略该事件（避免切换对话后旧任务的事件显示到新对话中）
- if (eventType !== 'shallow_compression' && eventData.conversation_id && this.currentConversationId) {
+ const crossConversationAllowed = new Set([
+ 'shallow_compression',
+ 'conversation_resolved',
+ 'compression_finished'
+ ]);
+ if (!crossConversationAllowed.has(eventType) && eventData.conversation_id && this.currentConversationId) {
if (eventData.conversation_id !== this.currentConversationId) {
console.log(`[DEBUG_AWAITING] 忽略不匹配的事件`, {
eventType,
@@ -935,9 +940,19 @@ export const taskPollingMethods = {
if (!data || typeof data !== 'object') {
return;
}
+ const wasInProgress = !!this.compressionInProgress;
this.compressionInProgress = !!data.in_progress;
this.compressionMode = data.mode || '';
this.compressionStage = data.stage || '';
+ if (this.compressionInProgress && !wasInProgress) {
+ const modeLabel = this.compressionMode === 'manual' ? '手动' : '自动';
+ this.uiPushToast({
+ title: '压缩中',
+ message: `对话正在${modeLabel}压缩，请稍候…`,
+ type: 'info',
+ duration: 2600
+ });
+ }
if (!this.compressionInProgress) {
this.compressionError = data.error || '';
}
@@ -952,6 +967,16 @@ export const taskPollingMethods = {
if (newId && newId !== this.currentConversationId) {
await this.loadConversation(newId, { force: true });
}
+ this.conversationsOffset = 0;
+ if (typeof this.loadConversationsList === 'function') {
+ await this.loadConversationsList();
+ }
+ this.uiPushToast({
+ title: '压缩完成',
+ message: '已自动切换到压缩后的新对话',
+ type: 'success',
+ duration: 2400
+ });
},
handleShallowCompression(data: any, eventIdx?: number) {
diff --git a/static/src/components/personalization/PersonalizationDrawer.vue b/static/src/components/personalization/PersonalizationDrawer.vue
index 2d72903..59a5f3d 100644
--- a/static/src/components/personalization/PersonalizationDrawer.vue
+++ b/static/src/components/personalization/PersonalizationDrawer.vue
@@ -522,6 +522,76 @@
自动深层压缩
+
+
+
+
+
+
+
+
+
+
+
+ 自定义项留空则使用默认值；深压缩触发上下文必须大于浅压缩触发上下文。
+