feat(llm): LlmCycle 切换 IR 消息类型，移除 Phase 0 桥接层

核心改动： - LlmCycle.messages 从 Vec<OpenaiChatMessage> 切换为 Vec<Message>，移除 Phase 0 引入的 chat_message_to_message / message_to_chat_message 转换函数 - 移除 LlmCycle.system_prompt 字段（FIX-D），调用方通过 Message::system_text() + with_messages() 管理系统提示； with_system_prompt() 标记 #[deprecated] 保留为过渡期 shim - submit_stream 简化（FIX-E）：Item = StreamEvent 不再带 Result 包装，错误用 StreamEvent::Error 传出；self.messages 不再自动 push_message，调用方在收到 MessageComplete 后手动调用 cycle.push_message() - compact.rs 适配 Message 类型 + ContentBlock 估算；microcompact 跳过 is_error:true 的 ToolResult（FIX-F），保留 LLM 错误诊断 - ConversationMemory 持久化从 OpenaiChatMessage 切换为 Message - ToolInvocation 增加 tool_call_id 字段（FIX-A）， invoke/invoke_all 签名增加 tool_call_id 参数调用方迁移： - AgentSession::submit_turn 用 Message::system_text() + with_messages() - simple_visit example 同步迁移测试 177 passed / 0 failed（compact 新增 5 个单元测试覆盖 FIX-F + 估计全变体）。
2026-07-02 22:48:00 +08:00
parent a74c24b6fe
commit 9e4f50c955
6 changed files with 364 additions and 310 deletions
@@ -1,6 +1,6 @@
 //! 上下文自动压缩 —— 当对话历史过长时自动压缩。

-use crate::llm::types::{ContentField, OpenaiChatMessage, OpenaiContentPart};
+use crate::llm::types::message::{ContentBlock, Message};

 const AUTOCOMPACT_BUFFER_TOKENS: u32 = 13_000;
 const RESERVED_OUTPUT_TOKENS: u32 = 20_000;
@@ -72,37 +72,43 @@ impl CompactState {
 }

 /// 粗略估计消息列表的 token 数（基于字符数，4 字符 ≈ 1 token）。
-pub fn estimate_message_tokens(messages: &[OpenaiChatMessage]) -> u32 {
+pub fn estimate_message_tokens(messages: &[Message]) -> u32 {
    messages
        .iter()
        .map(estimate_single_message_tokens)
        .sum()
 }

-fn estimate_single_message_tokens(msg: &OpenaiChatMessage) -> u32 {
+fn estimate_single_message_tokens(msg: &Message) -> u32 {
    let role_overhead: u32 = 4;
    let content_tokens = match msg {
-        OpenaiChatMessage::Developer { content, .. }
-        | OpenaiChatMessage::System { content, .. }
-        | OpenaiChatMessage::User { content, .. }
-        | OpenaiChatMessage::Assistant { content, .. }
-        | OpenaiChatMessage::Function { content, .. } => estimate_content_tokens(content),
-        OpenaiChatMessage::Tool { content, .. } => estimate_content_tokens(content),
+        Message::System { content }
+        | Message::User { content }
+        | Message::Assistant { content }
+        | Message::ToolResult { content, .. } => estimate_content_blocks_tokens(content),
+        Message::UserImage { .. } => 50,
    };
    role_overhead + content_tokens
 }

-fn estimate_content_tokens(content: &ContentField) -> u32 {
-    match content {
-        ContentField::String(s) => estimate_text_tokens(s),
-        ContentField::Array(parts) => parts.iter().map(estimate_part_tokens).sum(),
-    }
+fn estimate_content_blocks_tokens(blocks: &[ContentBlock]) -> u32 {
+    blocks.iter().map(estimate_block_tokens).sum()
 }

-fn estimate_part_tokens(part: &OpenaiContentPart) -> u32 {
-    match part {
-        OpenaiContentPart::Text { text } => estimate_text_tokens(text),
-        _ => 50,
+fn estimate_block_tokens(block: &ContentBlock) -> u32 {
+    match block {
+        ContentBlock::Text { text } => estimate_text_tokens(text),
+        ContentBlock::Thinking { text, .. } => estimate_text_tokens(text),
+        ContentBlock::ToolUse { input, .. } => {
+            estimate_text_tokens(&input.to_string())
+        }
+        ContentBlock::ToolResult { content, .. } => estimate_content_blocks_tokens(content),
+        // ponytail: Image / Audio / File / Extension 在 IR 中固定估算。
+        // 无文本的视觉/音频 block 用兜底估算，避免 token 计数膨胀。
+        ContentBlock::Image { .. }
+        | ContentBlock::Audio { .. }
+        | ContentBlock::File { .. }
+        | ContentBlock::Extension { .. } => 50,
    }
 }

@@ -115,11 +121,7 @@ fn estimate_text_tokens(text: &str) -> u32 {
 }

 /// 判断是否需要触发自动压缩。
-pub fn should_compact(
-    messages: &[OpenaiChatMessage],
-    config: &CompactConfig,
-    state: &CompactState,
-) -> bool {
+pub fn should_compact(messages: &[Message], config: &CompactConfig, state: &CompactState) -> bool {
    if state.consecutive_failures >= MAX_CONSECUTIVE_FAILURES {
        return false;
    }
@@ -133,7 +135,10 @@ pub fn should_compact(
 /// 保留最近的 `keep_recent` 条消息不变。
 ///
 /// 返回释放的估算 token 数。
-pub fn microcompact(messages: &mut [OpenaiChatMessage], keep_recent: usize) -> u32 {
+///
+/// **审查 FIX-F**：仅压缩 `is_error: false` 的 `ToolResult` —— 错误结果包含对 LLM
+/// 理解失败原因至关重要的诊断信息，压缩后 LLM 无法理解。
+pub fn microcompact(messages: &mut [Message], keep_recent: usize) -> u32 {
    if messages.len() <= keep_recent {
        return 0;
    }
@@ -141,19 +146,119 @@ pub fn microcompact(messages: &mut [OpenaiChatMessage], keep_recent: usize) -> u
    let prune_start = messages.len() - keep_recent;
    let mut freed_tokens: u32 = 0;

+    // 第一遍：计算可释放 token（仅非错误 ToolResult）
    for msg in &messages[..prune_start] {
-        if matches!(msg, OpenaiChatMessage::Tool { .. }) {
+        if matches!(msg, Message::ToolResult { is_error: false, .. }) {
            freed_tokens += estimate_single_message_tokens(msg);
        }
    }

+    // 第二遍：替换内容（仅非错误 ToolResult）
    for msg in &mut messages[..prune_start] {
-        if let OpenaiChatMessage::Tool { content, .. } = msg {
-            *content = ContentField::Array(vec![OpenaiContentPart::Text {
+        if let Message::ToolResult { content, is_error: false, .. } = msg {
+            *content = vec![ContentBlock::Text {
                text: "[pruned]".to_string(),
-            }]);
+            }];
        }
    }

    freed_tokens
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn user_msg(s: &str) -> Message {
+        Message::user_text(s)
+    }
+
+    #[test]
+    fn estimate_message_tokens_handles_all_variants() {
+        let messages = vec![
+            Message::System {
+                content: vec![ContentBlock::Text {
+                    text: "sys".into(),
+                }],
+            },
+            Message::user_text("hi"),
+            Message::assistant("ans"),
+            Message::user_image("b64", "image/png", crate::llm::types::shared::ImageDetail::Auto),
+            Message::tool_result("call_1", "tool res", false),
+        ];
+        let tokens = estimate_message_tokens(&messages);
+        // 至少 5 条消息 × 4 role overhead = 20 + 文本/估算
+        assert!(tokens > 20);
+    }
+
+    #[test]
+    fn microcompact_replaces_old_tool_result_with_pruned() {
+        let mut messages = vec![
+            user_msg("hi"),
+            Message::tool_result("call_1", "raw result".repeat(50), false),
+            user_msg("again"),
+            user_msg("keep recent 1"),
+            user_msg("keep recent 2"),
+        ];
+        let before_len = messages.len();
+        let freed = microcompact(&mut messages, 2);
+        assert!(freed > 0);
+        assert_eq!(messages.len(), before_len); // 只改内容，不删消息
+        // 索引 1 是被压缩的 ToolResult
+        if let Message::ToolResult { content, is_error, .. } = &messages[1] {
+            assert_eq!(content.len(), 1);
+            assert!(matches!(&content[0], ContentBlock::Text { text } if text == "[pruned]"));
+            assert!(!is_error);
+        } else {
+            panic!("expected ToolResult at index 1");
+        }
+    }
+
+    /// FIX-F 验证：错误结果不被压缩，诊断信息完整保留。
+    #[test]
+    fn microcompact_preserves_error_tool_results() {
+        let mut messages = vec![
+            user_msg("hi"),
+            Message::tool_result("call_1", "important error info: backend down", true),
+            user_msg("keep recent 1"),
+            user_msg("keep recent 2"),
+        ];
+        let before_len = messages.len();
+        let freed = microcompact(&mut messages, 2);
+        assert_eq!(freed, 0); // 错误 ToolResult 不计入
+        assert_eq!(messages.len(), before_len);
+        // 错误信息保留完整
+        if let Message::ToolResult { content, is_error, .. } = &messages[1] {
+            assert!(is_error);
+            assert!(matches!(&content[0], ContentBlock::Text { text } if text.contains("backend down")));
+        } else {
+            panic!("expected ToolResult at index 1");
+        }
+    }
+
+    #[test]
+    fn microcompact_keeps_recent_messages() {
+        let mut messages = vec![
+            Message::tool_result("c1", "old", false),
+            user_msg("m1"),
+            user_msg("m2"),
+            user_msg("recent"),
+        ];
+        let freed = microcompact(&mut messages, 1);
+        // keep_recent=1 → 仅最近 1 条不动，其余 ToolResult 压缩
+        // 索引 0 (ToolResult) 被压缩，索引 1-3 保留
+        assert!(freed > 0);
+        if let Message::ToolResult { content, .. } = &messages[0] {
+            assert!(matches!(&content[0], ContentBlock::Text { text } if text == "[pruned]"));
+        }
+        assert!(matches!(messages[3], Message::User { .. }));
+    }
+
+    #[test]
+    fn should_compact_respects_threshold() {
+        let cfg = CompactConfig::default();
+        let state = CompactState::new();
+        let empty: Vec<Message> = vec![];
+        assert!(!should_compact(&empty, &cfg, &state));
+    }
+}