agent-context 0.1.5

//! [`AgentContext`] Actor 及所有 [`Message`] 实现。
//!
//! 管理三区消息模型的状态，提供消息增删改查、模型对话、上下文压缩等操作。

use kameo::prelude::*;

use super::events::{
    CompressStrategy, NotifyCompressedForReply, RequestAppend, RequestClear,
    RequestCompress, RequestCompressed, RequestEstimateTokens, RequestExtend, RequestFindByRole,
    RequestImportIncremental, RequestGet, RequestImmutable, RequestIncremental, RequestInsert,
    RequestIsEmpty, RequestLen, RequestMessages, RequestPop, RequestRemove, RequestRetain,
    RequestSend, RequestSendStream, RequestSubscribeCompressed, RequestUnsubscribeCompressed,
    RequestExportIncremental, RequestExportAll, RequestUpdate,
};
use super::stream::AgentSendStream;
use super::types::ContextBackend;
use crate::error::AgentError;

type CompressSubscriberReply<M> = (Vec<M>, Vec<M>);
type CompressSubscriberRecipient<M> = ReplyRecipient<NotifyCompressedForReply<M>, CompressSubscriberReply<M>>;

use crate::message::ContextMessage;
use crate::readonly::ReadOnly;

// ---------------------------------------------------------------------------
// AgentContext Actor
// ---------------------------------------------------------------------------

/// LLM 对话上下文管理器，kameo Actor。
///
/// 管理三区 + 可选 Scratch 消息模型（immutable → compressed → incremental → scratch），提供：
/// - 消息增删改查（[`RequestAppend`]、[`RequestUpdate`]、[`RequestRemove`] 等）
/// - 对话发送（[`RequestSend`]、[`RequestSendStream`]），支持通过 [`super::CommonOpts::scratch`] 追加临时元数据
/// - 上下文压缩（[`RequestCompress`]）
/// - Token 估算（[`RequestEstimateTokens`]）
/// - 压缩订阅管理（[`RequestSubscribeCompressed`]、[`RequestUnsubscribeCompressed`]）
///
/// ## 构造
///
/// ```ignore
/// let actor_ref = AgentContext::spawn(AgentContext::new(backend, vec![]));
/// ```
#[derive(Actor)]
pub struct AgentContext<B: ContextBackend> {
    backend: B,
    immutable: ReadOnly<B::Message>,
    compressed: Vec<B::Message>,
    incremental: Vec<B::Message>,
    on_compressed: Option<CompressSubscriberRecipient<B::Message>>,
}

impl<B: ContextBackend> AgentContext<B> {
    /// 创建新的上下文管理器。
    ///
    /// - `backend`: 实现了 [`ContextBackend`] 的 LLM 后端实例
    /// - `immutable`: 初始不可变消息（系统提示词等），放入 immutable 区
    pub fn new(backend: B, immutable: Vec<B::Message>) -> Self {
        Self {
            backend,
            immutable: ReadOnly::from(immutable),
            compressed: Vec::new(),
            incremental: Vec::new(),
            on_compressed: None,
        }
    }

    fn default_summary_prompt() -> String {
        "请将以下对话历史压缩为简洁摘要，保留关键信息、决策和上下文。输出一条 system 消息。"
            .to_string()
    }


    /// 将摘要和保留消息交给压缩结果订阅者处理，若无订阅者则原样返回。
    async fn notify_compressed_subscriber(
        &self,
        summary: Vec<B::Message>,
        kept: Vec<B::Message>,
    ) -> Result<(Vec<B::Message>, Vec<B::Message>), AgentError> {
        if let Some(subscriber) = &self.on_compressed {
            subscriber
                .ask(NotifyCompressedForReply {
                    summary,
                    kept,
                })
                .send()
                .await
                .map_err(|e| AgentError::Context(e.to_string()))
        } else {
            Ok((summary, kept))
        }
    }

    /// 检查上下文是否已满，如果已满且 `auto_compress` 则自动压缩。
    async fn compress_if_full(&mut self, opts: &B::Opts) -> Result<(), AgentError> {
        let common = opts.as_ref();
        let mut all: Vec<B::Message> = self
            .immutable
            .iter()
            .chain(self.compressed.iter())
            .chain(self.incremental.iter())
            .cloned()
            .collect();
        if let Some(ref scratch) = common.scratch {
            all.push(self.backend.system_message(scratch.clone()));
        }
        let tokens = self
            .backend
            .estimate_tokens(&all)
            .await
            .unwrap_or(usize::MAX);
        if tokens < common.context_window {
            return Ok(());
        }
        if !common.auto_compress {
            return Err(AgentError::Context("上下文已满且未启用自动压缩".into()));
        }
        let total = self.incremental.len();
        let keep = total / 2;
        if total <= keep {
            return Ok(());
        }
        let split = total - keep;
        let to_summarize: Vec<B::Message> = self.incremental.drain(..split).collect();
        if to_summarize.is_empty() {
            return Ok(());
        }
        let mut summary_messages =
            vec![self.backend.system_message(Self::default_summary_prompt())];
        summary_messages.append(&mut self.compressed);
        summary_messages.extend(to_summarize);
        let response = self.backend.send(&summary_messages, opts).await?;
        let raw_msgs = self
            .backend
            .extract_messages(std::slice::from_ref(&response))?;
        let request_msgs = self.backend.to_request_messages(raw_msgs)?;
        let summary: Vec<B::Message> = request_msgs
            .into_iter()
            .map(|msg| self.backend.to_system_message(msg))
            .collect();
        let kept: Vec<B::Message> = self.incremental.drain(..).collect();
        let (final_summary, final_kept) =
            self.notify_compressed_subscriber(summary, kept).await?;
        self.compressed = final_summary;
        self.incremental = final_kept;
        Ok(())
    }
}

// ---------------------------------------------------------------------------
// 变更消息
// ---------------------------------------------------------------------------

impl<B: ContextBackend> Message<RequestAppend<B::Message>> for AgentContext<B> {
    type Reply = ();

    async fn handle(
        &mut self,
        msg: RequestAppend<B::Message>,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        self.incremental.push(msg.message);
    }
}

impl<B: ContextBackend> Message<RequestExtend<B::Message>> for AgentContext<B> {
    type Reply = ();

    async fn handle(
        &mut self,
        msg: RequestExtend<B::Message>,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        for m in msg.messages {
            self.incremental.push(m);
        }
    }
}

impl<B: ContextBackend> Message<RequestUpdate<B::Message>> for AgentContext<B> {
    type Reply = Result<(), AgentError>;

    async fn handle(
        &mut self,
        msg: RequestUpdate<B::Message>,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        if msg.index >= self.incremental.len() {
            return Err(AgentError::Context("索引越界".into()));
        }
        self.incremental[msg.index] = msg.message;
        Ok(())
    }
}

impl<B: ContextBackend> Message<RequestInsert<B::Message>> for AgentContext<B> {
    type Reply = Result<(), AgentError>;

    async fn handle(
        &mut self,
        msg: RequestInsert<B::Message>,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        if msg.index > self.incremental.len() {
            return Err(AgentError::Context("索引越界".into()));
        }
        self.incremental.insert(msg.index, msg.message);
        Ok(())
    }
}

impl<B: ContextBackend> Message<RequestRemove> for AgentContext<B> {
    type Reply = Result<(), AgentError>;

    async fn handle(
        &mut self,
        msg: RequestRemove,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        if msg.index >= self.incremental.len() {
            return Err(AgentError::Context("索引越界".into()));
        }
        self.incremental.remove(msg.index);
        Ok(())
    }
}

impl<B: ContextBackend> Message<RequestPop> for AgentContext<B> {
    type Reply = Option<B::Message>;

    async fn handle(
        &mut self,
        _msg: RequestPop,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        self.incremental.pop()
    }
}

impl<B: ContextBackend> Message<RequestRetain> for AgentContext<B> {
    type Reply = ();

    async fn handle(
        &mut self,
        msg: RequestRetain,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        let mut removed = Vec::new();
        let role = msg.role;
        self.incremental.retain(|m| {
            if m.role() == role {
                true
            } else {
                removed.push(m.clone());
                false
            }
        });
    }
}

impl<B: ContextBackend> Message<RequestClear> for AgentContext<B> {
    type Reply = ();

    async fn handle(
        &mut self,
        _msg: RequestClear,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        self.incremental.clear();
    }
}

// ---------------------------------------------------------------------------
// 查询消息
// ---------------------------------------------------------------------------

impl<B: ContextBackend> Message<RequestLen> for AgentContext<B> {
    type Reply = usize;

    async fn handle(
        &mut self,
        _msg: RequestLen,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        self.immutable.len() + self.compressed.len() + self.incremental.len()
    }
}

impl<B: ContextBackend> Message<RequestIsEmpty> for AgentContext<B> {
    type Reply = bool;

    async fn handle(
        &mut self,
        _msg: RequestIsEmpty,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        self.immutable.is_empty() && self.compressed.is_empty() && self.incremental.is_empty()
    }
}

impl<B: ContextBackend> Message<RequestGet> for AgentContext<B> {
    type Reply = Option<B::Message>;

    async fn handle(
        &mut self,
        msg: RequestGet,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        let idx = msg.0;
        let imm_len = self.immutable.len();
        let comp_len = self.compressed.len();
        if idx < imm_len {
            Some(self.immutable[idx].clone())
        } else if idx < imm_len + comp_len {
            Some(self.compressed[idx - imm_len].clone())
        } else if idx < imm_len + comp_len + self.incremental.len() {
            Some(self.incremental[idx - imm_len - comp_len].clone())
        } else {
            None
        }
    }
}

impl<B: ContextBackend> Message<RequestMessages> for AgentContext<B> {
    type Reply = Vec<B::Message>;

    async fn handle(
        &mut self,
        _msg: RequestMessages,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        self.immutable
            .iter()
            .chain(self.compressed.iter())
            .chain(self.incremental.iter())
            .cloned()
            .collect()
    }
}

impl<B: ContextBackend> Message<RequestImmutable> for AgentContext<B> {
    type Reply = Vec<B::Message>;

    async fn handle(
        &mut self,
        _msg: RequestImmutable,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        self.immutable.to_vec()
    }
}

impl<B: ContextBackend> Message<RequestCompressed> for AgentContext<B> {
    type Reply = Vec<B::Message>;

    async fn handle(
        &mut self,
        _msg: RequestCompressed,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        self.compressed.clone()
    }
}

impl<B: ContextBackend> Message<RequestIncremental> for AgentContext<B> {
    type Reply = Vec<B::Message>;

    async fn handle(
        &mut self,
        _msg: RequestIncremental,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        self.incremental.clone()
    }
}

impl<B: ContextBackend> Message<RequestFindByRole> for AgentContext<B> {
    type Reply = Vec<B::Message>;

    async fn handle(
        &mut self,
        msg: RequestFindByRole,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        self.immutable
            .iter()
            .chain(self.compressed.iter())
            .chain(self.incremental.iter())
            .filter(|m| m.role() == msg.0)
            .cloned()
            .collect()
    }
}

// ---------------------------------------------------------------------------
// 对话消息
// ---------------------------------------------------------------------------

impl<B: ContextBackend> Message<RequestSend<B::Opts>> for AgentContext<B> {
    type Reply = Result<B::Response, AgentError>;

    async fn handle(
        &mut self,
        msg: RequestSend<B::Opts>,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        self.compress_if_full(&msg.opts).await?;
        let scratch = msg.opts.as_ref().scratch.clone();
        let mut all_messages: Vec<B::Message> = self
            .immutable
            .iter()
            .chain(self.compressed.iter())
            .chain(self.incremental.iter())
            .cloned()
            .collect();
        if let Some(content) = scratch {
            all_messages.push(self.backend.system_message(content));
        }
        let response = self.backend.send(&all_messages, &msg.opts).await?;
        let raw_msgs = self
            .backend
            .extract_messages(std::slice::from_ref(&response))?;
        let request_msgs = self.backend.to_request_messages(raw_msgs)?;
        for msg in &request_msgs {
            self.incremental.push(msg.clone());
        }
        Ok(response)
    }
}

impl<B: ContextBackend + Clone> Message<RequestSendStream<B::Opts>> for AgentContext<B> {
    type Reply = Result<AgentSendStream<B>, AgentError>;

    async fn handle(
        &mut self,
        msg: RequestSendStream<B::Opts>,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        self.compress_if_full(&msg.opts).await?;
        let scratch = msg.opts.as_ref().scratch.clone();
        let mut all_messages: Vec<B::Message> = self
            .immutable
            .iter()
            .chain(self.compressed.iter())
            .chain(self.incremental.iter())
            .cloned()
            .collect();
        if let Some(content) = scratch {
            all_messages.push(self.backend.system_message(content));
        }
        let stream = self.backend.send_stream(all_messages, msg.opts);
        Ok(AgentSendStream::new(stream))
    }
}

// ---------------------------------------------------------------------------
// 压缩消息
// ---------------------------------------------------------------------------

impl<B: ContextBackend> Message<RequestCompress<B::Opts>> for AgentContext<B> {
    type Reply = Result<(), AgentError>;

    async fn handle(
        &mut self,
        msg: RequestCompress<B::Opts>,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        match msg.strategy {
            CompressStrategy::Summarize { keep, prompt } => {
                let total = self.incremental.len();
                if total > keep {
                    let split = total - keep;
                    let to_summarize: Vec<B::Message> = self.incremental.drain(..split).collect();
                    if !to_summarize.is_empty() {
                        let summary_prompt = prompt.unwrap_or_else(Self::default_summary_prompt);
                        let mut summary_messages =
                            vec![self.backend.system_message(summary_prompt)];
                        summary_messages.append(&mut self.compressed);
                        summary_messages.extend(to_summarize);
                        let response = self.backend.send(&summary_messages, &msg.opts).await?;
                        let raw_msgs = self
                            .backend
                            .extract_messages(std::slice::from_ref(&response))?;
                        let request_msgs = self.backend.to_request_messages(raw_msgs)?;
                        let summary: Vec<B::Message> = request_msgs
                            .into_iter()
                            .map(|msg| self.backend.to_system_message(msg))
                            .collect();
                        let kept: Vec<B::Message> = self.incremental.drain(..).collect();
                        let (final_summary, final_kept) =
                            self.notify_compressed_subscriber(summary, kept).await?;
                        self.compressed = final_summary;
                        self.incremental = final_kept;
                    }
                }
                Ok(())
            }
        }
    }
}


impl<B: ContextBackend> Message<RequestSubscribeCompressed<B::Message>> for AgentContext<B> {
    type Reply = ();

    async fn handle(
        &mut self,
        msg: RequestSubscribeCompressed<B::Message>,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        self.on_compressed = Some(msg.recipient);
    }
}

impl<B: ContextBackend> Message<RequestUnsubscribeCompressed> for AgentContext<B> {
    type Reply = ();

    async fn handle(
        &mut self,
        _msg: RequestUnsubscribeCompressed,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        self.on_compressed = None;
    }
}

// ---------------------------------------------------------------------------
// 工具消息
// ---------------------------------------------------------------------------

impl<B: ContextBackend> Message<RequestEstimateTokens> for AgentContext<B> {
    type Reply = usize;

    async fn handle(
        &mut self,
        _msg: RequestEstimateTokens,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        let all: Vec<B::Message> = self
            .immutable
            .iter()
            .chain(self.compressed.iter())
            .chain(self.incremental.iter())
            .cloned()
            .collect();
        self.backend.estimate_tokens(&all).await.unwrap_or(0)
    }
}

impl<B: ContextBackend> Message<RequestExportAll> for AgentContext<B> {
    type Reply = Result<String, AgentError>;

    async fn handle(
        &mut self,
        _msg: RequestExportAll,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        let lines: Vec<String> = self
            .immutable
            .iter()
            .chain(self.compressed.iter())
            .chain(self.incremental.iter())
            .map(|m| self.backend.message_to_json(m))
            .collect::<Result<_, _>>()?;
        Ok(lines.join("\n"))
    }
}

impl<B: ContextBackend> Message<RequestExportIncremental> for AgentContext<B> {
    type Reply = Result<String, AgentError>;

    async fn handle(
        &mut self,
        _msg: RequestExportIncremental,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        let lines: Vec<String> = self
            .incremental
            .iter()
            .map(|m| self.backend.message_to_json(m))
            .collect::<Result<_, _>>()?;
        Ok(lines.join("\n"))
    }
}

impl<B: ContextBackend> Message<RequestImportIncremental> for AgentContext<B> {
    type Reply = Result<(), AgentError>;

    async fn handle(
        &mut self,
        msg: RequestImportIncremental,
        _ctx: &mut Context<Self, Self::Reply>,
    ) -> Self::Reply {
        let mut messages = Vec::new();
        for line in msg.json.lines() {
            let line = line.trim();
            if line.is_empty() {
                continue;
            }
            messages.push(self.backend.message_from_json(line)?);
        }
        self.incremental.clear();
        self.incremental = messages.clone();
        Ok(())
    }
}