AstrBot 4.10.5__py3-none-any.whl → 4.11.0__py3-none-any.whl

astrbot/api/event/filter/__init__.py

@@ -21,6 +21,9 @@ from astrbot.core.star.register import (
 from astrbot.core.star.register import register_on_llm_request as on_llm_request
 from astrbot.core.star.register import register_on_llm_response as on_llm_response
 from astrbot.core.star.register import register_on_platform_loaded as on_platform_loaded
+from astrbot.core.star.register import (
+    register_on_waiting_llm_request as on_waiting_llm_request,
+)
 from astrbot.core.star.register import register_permission_type as permission_type
 from astrbot.core.star.register import (
     register_platform_adapter_type as platform_adapter_type,
@@ -46,6 +49,7 @@ __all__ = [
     "on_llm_request",
     "on_llm_response",
     "on_platform_loaded",
+    "on_waiting_llm_request",
     "permission_type",
     "platform_adapter_type",
     "regex",

astrbot/builtin_stars/builtin_commands/commands/tts.py

@@ -14,13 +14,13 @@ class TTSCommand:
     async def tts(self, event: AstrMessageEvent):
         """开关文本转语音（会话级别）"""
         umo = event.unified_msg_origin
-        ses_tts = SessionServiceManager.is_tts_enabled_for_session(umo)
+        ses_tts = await SessionServiceManager.is_tts_enabled_for_session(umo)
         cfg = self.context.get_config(umo=umo)
         tts_enable = cfg["provider_tts_settings"]["enable"]
 
         # 切换状态
         new_status = not ses_tts
-        SessionServiceManager.set_tts_status_for_session(umo, new_status)
+        await SessionServiceManager.set_tts_status_for_session(umo, new_status)
 
         status_text = "已开启" if new_status else "已关闭"
 

astrbot/cli/__init__.py

@@ -1 +1 @@
-__version__ = "4.10.5"
+__version__ = "4.11.0"

astrbot/core/agent/context/compressor.py

@@ -0,0 +1,243 @@
+from typing import TYPE_CHECKING, Protocol, runtime_checkable
+
+from ..message import Message
+
+if TYPE_CHECKING:
+    from astrbot import logger
+else:
+    try:
+        from astrbot import logger
+    except ImportError:
+        import logging
+
+        logger = logging.getLogger("astrbot")
+
+if TYPE_CHECKING:
+    from astrbot.core.provider.provider import Provider
+
+from ..context.truncator import ContextTruncator
+
+
+@runtime_checkable
+class ContextCompressor(Protocol):
+    """
+    Protocol for context compressors.
+    Provides an interface for compressing message lists.
+    """
+
+    def should_compress(
+        self, messages: list[Message], current_tokens: int, max_tokens: int
+    ) -> bool:
+        """Check if compression is needed.
+
+        Args:
+            messages: The message list to evaluate.
+            current_tokens: The current token count.
+            max_tokens: The maximum allowed tokens for the model.
+
+        Returns:
+            True if compression is needed, False otherwise.
+        """
+        ...
+
+    async def __call__(self, messages: list[Message]) -> list[Message]:
+        """Compress the message list.
+
+        Args:
+            messages: The original message list.
+
+        Returns:
+            The compressed message list.
+        """
+        ...
+
+
+class TruncateByTurnsCompressor:
+    """Truncate by turns compressor implementation.
+    Truncates the message list by removing older turns.
+    """
+
+    def __init__(self, truncate_turns: int = 1, compression_threshold: float = 0.82):
+        """Initialize the truncate by turns compressor.
+
+        Args:
+            truncate_turns: The number of turns to remove when truncating (default: 1).
+            compression_threshold: The compression trigger threshold (default: 0.82).
+        """
+        self.truncate_turns = truncate_turns
+        self.compression_threshold = compression_threshold
+
+    def should_compress(
+        self, messages: list[Message], current_tokens: int, max_tokens: int
+    ) -> bool:
+        """Check if compression is needed.
+
+        Args:
+            messages: The message list to evaluate.
+            current_tokens: The current token count.
+            max_tokens: The maximum allowed tokens.
+
+        Returns:
+            True if compression is needed, False otherwise.
+        """
+        if max_tokens <= 0 or current_tokens <= 0:
+            return False
+        usage_rate = current_tokens / max_tokens
+        return usage_rate > self.compression_threshold
+
+    async def __call__(self, messages: list[Message]) -> list[Message]:
+        truncator = ContextTruncator()
+        truncated_messages = truncator.truncate_by_dropping_oldest_turns(
+            messages,
+            drop_turns=self.truncate_turns,
+        )
+        return truncated_messages
+
+
+def split_history(
+    messages: list[Message], keep_recent: int
+) -> tuple[list[Message], list[Message], list[Message]]:
+    """Split the message list into system messages, messages to summarize, and recent messages.
+
+    Ensures that the split point is between complete user-assistant pairs to maintain conversation flow.
+
+    Args:
+        messages: The original message list.
+        keep_recent: The number of latest messages to keep.
+
+    Returns:
+        tuple: (system_messages, messages_to_summarize, recent_messages)
+    """
+    # keep the system messages
+    first_non_system = 0
+    for i, msg in enumerate(messages):
+        if msg.role != "system":
+            first_non_system = i
+            break
+
+    system_messages = messages[:first_non_system]
+    non_system_messages = messages[first_non_system:]
+
+    if len(non_system_messages) <= keep_recent:
+        return system_messages, [], non_system_messages
+
+    # Find the split point, ensuring recent_messages starts with a user message
+    # This maintains complete conversation turns
+    split_index = len(non_system_messages) - keep_recent
+
+    # Search backward from split_index to find the first user message
+    # This ensures recent_messages starts with a user message (complete turn)
+    while split_index > 0 and non_system_messages[split_index].role != "user":
+        # TODO: +=1 or -=1 ? calculate by tokens
+        split_index -= 1
+
+    # If we couldn't find a user message, keep all messages as recent
+    if split_index == 0:
+        return system_messages, [], non_system_messages
+
+    messages_to_summarize = non_system_messages[:split_index]
+    recent_messages = non_system_messages[split_index:]
+
+    return system_messages, messages_to_summarize, recent_messages
+
+
+class LLMSummaryCompressor:
+    """LLM-based summary compressor.
+    Uses LLM to summarize the old conversation history, keeping the latest messages.
+    """
+
+    def __init__(
+        self,
+        provider: "Provider",
+        keep_recent: int = 4,
+        instruction_text: str | None = None,
+        compression_threshold: float = 0.82,
+    ):
+        """Initialize the LLM summary compressor.
+
+        Args:
+            provider: The LLM provider instance.
+            keep_recent: The number of latest messages to keep (default: 4).
+            instruction_text: Custom instruction for summary generation.
+            compression_threshold: The compression trigger threshold (default: 0.82).
+        """
+        self.provider = provider
+        self.keep_recent = keep_recent
+        self.compression_threshold = compression_threshold
+
+        self.instruction_text = instruction_text or (
+            "Based on our full conversation history, produce a concise summary of key takeaways and/or project progress.\n"
+            "1. Systematically cover all core topics discussed and the final conclusion/outcome for each; clearly highlight the latest primary focus.\n"
+            "2. If any tools were used, summarize tool usage (total call count) and extract the most valuable insights from tool outputs.\n"
+            "3. If there was an initial user goal, state it first and describe the current progress/status.\n"
+            "4. Write the summary in the user's language.\n"
+        )
+
+    def should_compress(
+        self, messages: list[Message], current_tokens: int, max_tokens: int
+    ) -> bool:
+        """Check if compression is needed.
+
+        Args:
+            messages: The message list to evaluate.
+            current_tokens: The current token count.
+            max_tokens: The maximum allowed tokens.
+
+        Returns:
+            True if compression is needed, False otherwise.
+        """
+        if max_tokens <= 0 or current_tokens <= 0:
+            return False
+        usage_rate = current_tokens / max_tokens
+        return usage_rate > self.compression_threshold
+
+    async def __call__(self, messages: list[Message]) -> list[Message]:
+        """Use LLM to generate a summary of the conversation history.
+
+        Process:
+        1. Divide messages: keep the system message and the latest N messages.
+        2. Send the old messages + the instruction message to the LLM.
+        3. Reconstruct the message list: [system message, summary message, latest messages].
+        """
+        if len(messages) <= self.keep_recent + 1:
+            return messages
+
+        system_messages, messages_to_summarize, recent_messages = split_history(
+            messages, self.keep_recent
+        )
+
+        if not messages_to_summarize:
+            return messages
+
+        # build payload
+        instruction_message = Message(role="user", content=self.instruction_text)
+        llm_payload = messages_to_summarize + [instruction_message]
+
+        # generate summary
+        try:
+            response = await self.provider.text_chat(contexts=llm_payload)
+            summary_content = response.completion_text
+        except Exception as e:
+            logger.error(f"Failed to generate summary: {e}")
+            return messages
+
+        # build result
+        result = []
+        result.extend(system_messages)
+
+        result.append(
+            Message(
+                role="user",
+                content=f"Our previous history conversation summary: {summary_content}",
+            )
+        )
+        result.append(
+            Message(
+                role="assistant",
+                content="Acknowledged the summary of our previous conversation history.",
+            )
+        )
+
+        result.extend(recent_messages)
+
+        return result

astrbot/core/agent/context/config.py

@@ -0,0 +1,35 @@
+from dataclasses import dataclass
+from typing import TYPE_CHECKING
+
+from .compressor import ContextCompressor
+from .token_counter import TokenCounter
+
+if TYPE_CHECKING:
+    from astrbot.core.provider.provider import Provider
+
+
+@dataclass
+class ContextConfig:
+    """Context configuration class."""
+
+    max_context_tokens: int = 0
+    """Maximum number of context tokens. <= 0 means no limit."""
+    enforce_max_turns: int = -1  # -1 means no limit
+    """Maximum number of conversation turns to keep. -1 means no limit. Executed before compression."""
+    truncate_turns: int = 1
+    """Number of conversation turns to discard at once when truncation is triggered.
+    Two processes will use this value:
+
+    1. Enforce max turns truncation.
+    2. Truncation by turns compression strategy.
+    """
+    llm_compress_instruction: str | None = None
+    """Instruction prompt for LLM-based compression."""
+    llm_compress_keep_recent: int = 0
+    """Number of recent messages to keep during LLM-based compression."""
+    llm_compress_provider: "Provider | None" = None
+    """LLM provider used for compression tasks. If None, truncation strategy is used."""
+    custom_token_counter: TokenCounter | None = None
+    """Custom token counting method. If None, the default method is used."""
+    custom_compressor: ContextCompressor | None = None
+    """Custom context compression method. If None, the default method is used."""

astrbot/core/agent/context/manager.py

@@ -0,0 +1,120 @@
+from astrbot import logger
+
+from ..message import Message
+from .compressor import LLMSummaryCompressor, TruncateByTurnsCompressor
+from .config import ContextConfig
+from .token_counter import EstimateTokenCounter
+from .truncator import ContextTruncator
+
+
+class ContextManager:
+    """Context compression manager."""
+
+    def __init__(
+        self,
+        config: ContextConfig,
+    ):
+        """Initialize the context manager.
+
+        There are two strategies to handle context limit reached:
+        1. Truncate by turns: remove older messages by turns.
+        2. LLM-based compression: use LLM to summarize old messages.
+
+        Args:
+            config: The context configuration.
+        """
+        self.config = config
+
+        self.token_counter = config.custom_token_counter or EstimateTokenCounter()
+        self.truncator = ContextTruncator()
+
+        if config.custom_compressor:
+            self.compressor = config.custom_compressor
+        elif config.llm_compress_provider:
+            self.compressor = LLMSummaryCompressor(
+                provider=config.llm_compress_provider,
+                keep_recent=config.llm_compress_keep_recent,
+                instruction_text=config.llm_compress_instruction,
+            )
+        else:
+            self.compressor = TruncateByTurnsCompressor(
+                truncate_turns=config.truncate_turns
+            )
+
+    async def process(
+        self, messages: list[Message], trusted_token_usage: int = 0
+    ) -> list[Message]:
+        """Process the messages.
+
+        Args:
+            messages: The original message list.
+
+        Returns:
+            The processed message list.
+        """
+        try:
+            result = messages
+
+            # 1. 基于轮次的截断 (Enforce max turns)
+            if self.config.enforce_max_turns != -1:
+                result = self.truncator.truncate_by_turns(
+                    result,
+                    keep_most_recent_turns=self.config.enforce_max_turns,
+                    drop_turns=self.config.truncate_turns,
+                )
+
+            # 2. 基于 token 的压缩
+            if self.config.max_context_tokens > 0:
+                total_tokens = self.token_counter.count_tokens(
+                    result, trusted_token_usage
+                )
+
+                if self.compressor.should_compress(
+                    result, total_tokens, self.config.max_context_tokens
+                ):
+                    result = await self._run_compression(result, total_tokens)
+
+            return result
+        except Exception as e:
+            logger.error(f"Error during context processing: {e}", exc_info=True)
+            return messages
+
+    async def _run_compression(
+        self, messages: list[Message], prev_tokens: int
+    ) -> list[Message]:
+        """
+        Compress/truncate the messages.
+
+        Args:
+            messages: The original message list.
+            prev_tokens: The token count before compression.
+
+        Returns:
+            The compressed/truncated message list.
+        """
+        logger.debug("Compress triggered, starting compression...")
+
+        messages = await self.compressor(messages)
+
+        # double check
+        tokens_after_summary = self.token_counter.count_tokens(messages)
+
+        # calculate compress rate
+        compress_rate = (tokens_after_summary / self.config.max_context_tokens) * 100
+        logger.info(
+            f"Compress completed."
+            f" {prev_tokens} -> {tokens_after_summary} tokens,"
+            f" compression rate: {compress_rate:.2f}%.",
+        )
+
+        # last check
+        if self.compressor.should_compress(
+            messages, tokens_after_summary, self.config.max_context_tokens
+        ):
+            logger.info(
+                "Context still exceeds max tokens after compression, applying halving truncation..."
+            )
+            # still need compress, truncate by half
+            messages = self.truncator.truncate_by_halving(messages)
+
+        return messages

astrbot/core/agent/context/token_counter.py

@@ -0,0 +1,64 @@
+import json
+from typing import Protocol, runtime_checkable
+
+from ..message import Message, TextPart
+
+
+@runtime_checkable
+class TokenCounter(Protocol):
+    """
+    Protocol for token counters.
+    Provides an interface for counting tokens in message lists.
+    """
+
+    def count_tokens(
+        self, messages: list[Message], trusted_token_usage: int = 0
+    ) -> int:
+        """Count the total tokens in the message list.
+
+        Args:
+            messages: The message list.
+            trusted_token_usage: The total token usage that LLM API returned.
+                For some cases, this value is more accurate.
+                But some API does not return it, so the value defaults to 0.
+
+        Returns:
+            The total token count.
+        """
+        ...
+
+
+class EstimateTokenCounter:
+    """Estimate token counter implementation.
+    Provides a simple estimation of token count based on character types.
+    """
+
+    def count_tokens(
+        self, messages: list[Message], trusted_token_usage: int = 0
+    ) -> int:
+        if trusted_token_usage > 0:
+            return trusted_token_usage
+
+        total = 0
+        for msg in messages:
+            content = msg.content
+            if isinstance(content, str):
+                total += self._estimate_tokens(content)
+            elif isinstance(content, list):
+                # 处理多模态内容
+                for part in content:
+                    if isinstance(part, TextPart):
+                        total += self._estimate_tokens(part.text)
+
+            # 处理 Tool Calls
+            if msg.tool_calls:
+                for tc in msg.tool_calls:
+                    tc_str = json.dumps(tc if isinstance(tc, dict) else tc.model_dump())
+                    total += self._estimate_tokens(tc_str)
+
+        return total
+
+    def _estimate_tokens(self, text: str) -> int:
+        chinese_count = len([c for c in text if "\u4e00" <= c <= "\u9fff"])
+        other_count = len(text) - chinese_count
+        return int(chinese_count * 0.6 + other_count * 0.3)

astrbot/core/agent/context/truncator.py

@@ -0,0 +1,141 @@
+from ..message import Message
+
+
+class ContextTruncator:
+    """Context truncator."""
+
+    def fix_messages(self, messages: list[Message]) -> list[Message]:
+        fixed_messages = []
+        for message in messages:
+            if message.role == "tool":
+                # tool block 前面必须要有 user 和 assistant block
+                if len(fixed_messages) < 2:
+                    # 这种情况可能是上下文被截断导致的
+                    # 我们直接将之前的上下文都清空
+                    fixed_messages = []
+                else:
+                    fixed_messages.append(message)
+            else:
+                fixed_messages.append(message)
+        return fixed_messages
+
+    def truncate_by_turns(
+        self,
+        messages: list[Message],
+        keep_most_recent_turns: int,
+        drop_turns: int = 1,
+    ) -> list[Message]:
+        """截断上下文列表，确保不超过最大长度。
+        一个 turn 包含一个 user 消息和一个 assistant 消息。
+        这个方法会保证截断后的上下文列表符合 OpenAI 的上下文格式。
+
+        Args:
+            messages: 上下文列表
+            keep_most_recent_turns: 保留最近的对话轮数
+            drop_turns: 一次性丢弃的对话轮数
+
+        Returns:
+            截断后的上下文列表
+        """
+        if keep_most_recent_turns == -1:
+            return messages
+
+        first_non_system = 0
+        for i, msg in enumerate(messages):
+            if msg.role != "system":
+                first_non_system = i
+                break
+
+        system_messages = messages[:first_non_system]
+        non_system_messages = messages[first_non_system:]
+
+        if len(non_system_messages) // 2 <= keep_most_recent_turns:
+            return messages
+
+        num_to_keep = keep_most_recent_turns - drop_turns + 1
+        if num_to_keep <= 0:
+            truncated_contexts = []
+        else:
+            truncated_contexts = non_system_messages[-num_to_keep * 2 :]
+
+        # 找到第一个 role 为 user 的索引，确保上下文格式正确
+        index = next(
+            (i for i, item in enumerate(truncated_contexts) if item.role == "user"),
+            None,
+        )
+        if index is not None and index > 0:
+            truncated_contexts = truncated_contexts[index:]
+
+        result = system_messages + truncated_contexts
+
+        return self.fix_messages(result)
+
+    def truncate_by_dropping_oldest_turns(
+        self,
+        messages: list[Message],
+        drop_turns: int = 1,
+    ) -> list[Message]:
+        """丢弃最旧的 N 个对话轮次。"""
+        if drop_turns <= 0:
+            return messages
+
+        first_non_system = 0
+        for i, msg in enumerate(messages):
+            if msg.role != "system":
+                first_non_system = i
+                break
+
+        system_messages = messages[:first_non_system]
+        non_system_messages = messages[first_non_system:]
+
+        if len(non_system_messages) // 2 <= drop_turns:
+            truncated_non_system = []
+        else:
+            truncated_non_system = non_system_messages[drop_turns * 2 :]
+
+        index = next(
+            (i for i, item in enumerate(truncated_non_system) if item.role == "user"),
+            None,
+        )
+        if index is not None:
+            truncated_non_system = truncated_non_system[index:]
+        elif truncated_non_system:
+            truncated_non_system = []
+
+        result = system_messages + truncated_non_system
+
+        return self.fix_messages(result)
+
+    def truncate_by_halving(
+        self,
+        messages: list[Message],
+    ) -> list[Message]:
+        """对半砍策略，删除 50% 的消息"""
+        if len(messages) <= 2:
+            return messages
+
+        first_non_system = 0
+        for i, msg in enumerate(messages):
+            if msg.role != "system":
+                first_non_system = i
+                break
+
+        system_messages = messages[:first_non_system]
+        non_system_messages = messages[first_non_system:]
+
+        messages_to_delete = len(non_system_messages) // 2
+        if messages_to_delete == 0:
+            return messages
+
+        truncated_non_system = non_system_messages[messages_to_delete:]
+
+        index = next(
+            (i for i, item in enumerate(truncated_non_system) if item.role == "user"),
+            None,
+        )
+        if index is not None:
+            truncated_non_system = truncated_non_system[index:]
+
+        result = system_messages + truncated_non_system
+
+        return self.fix_messages(result)