feat: 实现两段历史压缩的同步处理，替换会话的活动历史记录

src/agent/context_compressor.rs

@@ -5,7 +5,6 @@ use crate::bus::{
 use crate::config::LLMProviderConfig;
 use crate::providers::{ChatCompletionRequest, LLMProvider, Message, create_provider};
 use crate::text::{char_count, take_prefix_chars};
-
 use crate::agent::{AgentError, AgentRuntimeConfig};
 
 const TOKEN_ESTIMATE_SAFETY_MULTIPLIER: f64 = 1.2;
@@ -13,6 +12,145 @@ const CJK_CHARS_PER_TOKEN: f64 = 2.0;
 const OTHER_CHARS_PER_TOKEN: f64 = 4.0;
 const JSON_OVERHEAD_PER_MESSAGE: usize = 50;
 
+/// System context marker for the heavy compression (older segment) summary.
+pub const SYSTEM_CONTEXT_HISTORY_COMPACTION_OLDER: &str = "history_compaction_older";
+/// System context marker for the light compression (newer segment) summary.
+pub const SYSTEM_CONTEXT_HISTORY_COMPACTION_NEWER: &str = "history_compaction_newer";
+
+/// Default threshold ratio: compress when estimated tokens exceed 70% of context window.
+const DEFAULT_THRESHOLD_RATIO: f64 = 0.7;
+
+/// Default budget split: older segment gets 30% of summary budget, newer gets 70%.
+const OLDER_BUDGET_RATIO: f64 = 0.3;
+const NEWER_BUDGET_RATIO: f64 = 0.7;
+
+// ============================================================================
+// HistoryUnit — atomic message units for compression
+// ============================================================================
+
+/// An indivisible unit of conversation history. Compression operates at the
+/// unit level, so tool call sequences can never be split across boundaries.
+#[derive(Debug, Clone)]
+enum HistoryUnit {
+    /// Always-preserved system messages (agent_prompt, scheduled_prompt).
+    SystemGuard(ChatMessage),
+    /// A user message.
+    UserMessage(ChatMessage),
+    /// An atomic tool call sequence: assistant(tool_calls) + all following
+    /// tool-result messages. Must be preserved or compressed as a whole.
+    ToolRound {
+        assistant: ChatMessage,
+        results: Vec<ChatMessage>,
+    },
+    /// A plain assistant text response (no tool_calls).
+    AssistantText(ChatMessage),
+}
+
+impl HistoryUnit {
+    /// Estimate tokens for this unit alone.
+    fn estimate_tokens(&self) -> usize {
+        match self {
+            HistoryUnit::SystemGuard(msg) | HistoryUnit::UserMessage(msg) | HistoryUnit::AssistantText(msg) => {
+                estimate_tokens(std::slice::from_ref(msg))
+            }
+            HistoryUnit::ToolRound { assistant, results } => {
+                let mut all = vec![assistant.clone()];
+                all.extend(results.clone());
+                estimate_tokens(&all)
+            }
+        }
+    }
+
+    /// Expand this unit back into a flat list of ChatMessages.
+    #[allow(dead_code)]
+    fn into_messages(self) -> Vec<ChatMessage> {
+        match self {
+            HistoryUnit::SystemGuard(msg)
+            | HistoryUnit::UserMessage(msg)
+            | HistoryUnit::AssistantText(msg) => vec![msg],
+            HistoryUnit::ToolRound { assistant, results } => {
+                let mut msgs = vec![assistant];
+                msgs.extend(results);
+                msgs
+            }
+        }
+    }
+
+    /// Returns true if this unit is a ToolRound.
+    #[allow(dead_code)]
+    fn is_tool_round(&self) -> bool {
+        matches!(self, HistoryUnit::ToolRound { .. })
+    }
+}
+
+// ============================================================================
+// Unit parser — one forward pass, O(n)
+// ============================================================================
+
+fn is_system_guard(msg: &ChatMessage) -> bool {
+    msg.role == "system"
+        && (msg.has_system_context(SYSTEM_CONTEXT_AGENT_PROMPT)
+            || msg.has_system_context(SYSTEM_CONTEXT_SCHEDULED_PROMPT))
+}
+
+fn is_assistant_with_tool_calls(msg: &ChatMessage) -> bool {
+    msg.role == "assistant"
+        && msg
+            .tool_calls
+            .as_ref()
+            .map_or(false, |calls| !calls.is_empty())
+}
+
+/// Parse a flat message list into atomic units. Orphaned tool results
+/// (without a preceding assistant with tool_calls) are silently dropped.
+fn parse_to_units(messages: &[ChatMessage]) -> Vec<HistoryUnit> {
+    let mut units = Vec::new();
+    let mut i = 0;
+
+    while i < messages.len() {
+        let msg = &messages[i];
+
+        if is_system_guard(msg) {
+            units.push(HistoryUnit::SystemGuard(msg.clone()));
+            i += 1;
+        } else if msg.role == "user" {
+            units.push(HistoryUnit::UserMessage(msg.clone()));
+            i += 1;
+        } else if is_assistant_with_tool_calls(msg) {
+            let assistant = msg.clone();
+            let mut results = Vec::new();
+            i += 1;
+            while i < messages.len() && messages[i].role == "tool" {
+                results.push(messages[i].clone());
+                i += 1;
+            }
+            units.push(HistoryUnit::ToolRound { assistant, results });
+        } else if msg.role == "assistant" {
+            units.push(HistoryUnit::AssistantText(msg.clone()));
+            i += 1;
+        } else if msg.role == "tool" {
+            // Orphaned tool result — drop with a warning
+            tracing::warn!(
+                tool_call_id = ?msg.tool_call_id,
+                message_index = i,
+                "Dropping orphaned tool result during unit parsing"
+            );
+            i += 1;
+        } else {
+            // system messages that aren't guards — keep as plain messages
+            units.push(HistoryUnit::AssistantText(msg.clone()));
+            i += 1;
+        }
+    }
+
+    units
+}
+
+/// Estimate total tokens from a slice of units.
+fn estimate_tokens_from_units(units: &[HistoryUnit]) -> usize {
+    units.iter().map(|u| u.estimate_tokens()).sum()
+}
+
 /// Check if a character is CJK (Chinese, Japanese, Korean)
 fn is_cjk_char(c: char) -> bool {
     matches!(c,
@@ -97,7 +235,7 @@ impl Default for ContextCompressionConfig {
 pub struct ContextCompressor {
     config: ContextCompressionConfig,
     context_window: usize,
-    /// Threshold ratio to trigger compression (50% of context window)
+    /// Threshold ratio to trigger compression (70% of context window).
     threshold_ratio: f64,
 }
 
@@ -113,6 +251,10 @@ impl ContextCompressor {
             .clamp(MIN_SUMMARY_CHARS, MAX_SUMMARY_CHARS)
     }
 
+    // =========================================================================
+    // Transcript building (shared helpers, unchanged from original)
+    // =========================================================================
+
     fn format_transcript_entry(message: &ChatMessage) -> String {
         let role = match message.role.as_str() {
             "assistant" => "Assistant",
@@ -214,6 +356,86 @@ Be concise, aim for {} characters or less.
 
 {}
 
+"#,
+            target_chars, transcript
+        )
+    }
+
+    // =========================================================================
+    // Two-segment compression: heavy (older) + light (newer) prompts
+    // =========================================================================
+
+    /// Prompt for the older segment — aggressive compression, keep only essentials.
+    fn build_heavy_summary_prompt(transcript: &str, target_chars: usize) -> String {
+        format!(
+            r#"You are a conversation compaction engine. Aggressively summarize the following OLDER conversation segment. These events happened earlier in the session and will be marked as "较早的操作记录（已压缩）".
+
+A newer portion (not shown here) follows this segment. The model can use tools to re-read files for exact data, so you do NOT need to preserve full tool outputs.
+
+=== MUST PRESERVE (keep verbatim or near-verbatim) ===
+- All file paths (absolute and relative)
+- All URLs
+- Key numeric results and conclusions (one sentence per result)
+- Error messages (exact text if important)
+- User preferences and decisions
+- The original user task or goal
+
+=== SHOULD CONDENSE ===
+- Each tool call → one line summary: "tool_name: 关键参数 → 结果概要"
+- Multiple similar operations → group: "对 ./data/ 目录下的 10 个 CSV 文件执行了统计分析"
+- Verbose tool output → drop entirely unless it contains a critical result or error
+
+=== SHOULD DROP ===
+- Full tool output (hundreds of lines of raw data)
+- Repeated attempts of the same operation
+- Greetings, filler text, markdown formatting
+
+Be concise, aim for {} characters or less. Output the summary in Chinese if the original conversation was in Chinese.
+
+---
+
+OLDER SEGMENT (events from earlier in the session):
+
+{}
+
+"#,
+            target_chars, transcript
+        )
+    }
+
+    /// Prompt for the newer segment — lighter compression, keep more detail.
+    fn build_light_summary_prompt(transcript: &str, target_chars: usize) -> String {
+        format!(
+            r#"You are a conversation compaction engine. Lightly summarize the following RECENT conversation segment. These events happened just before the current moment and will be marked as "近期操作记录（轻度压缩，保留了更多细节）".
+
+An older segment (already summarized separately) precedes this. Keep enough detail so the model can continue the task seamlessly without re-reading files for data that appears in these recent results.
+
+=== MUST PRESERVE (keep with more detail than a normal summary) ===
+- All file paths, URLs, and identifiers
+- The exact sequence of recent operations (step by step)
+- Tool parameters (especially file paths, search queries, command strings)
+- Key outputs and results (shortened but keep the substance)
+- Error messages (complete, not truncated)
+- Current task status and what should happen next
+
+=== SHOULD CONDENSE ===
+- Tool outputs → truncate to the most meaningful parts (key data, conclusions, not raw output)
+- Long text → keep the essence but not every word
+- Repeated similar outputs → note the pattern but keep a representative example
+
+=== SHOULD DROP ===
+- Completely irrelevant debug output
+- Boilerplate text with no informational value
+- Trivial operations that have no bearing on the task
+
+Be concise, aim for {} characters or less. Output the summary in Chinese if the original conversation was in Chinese.
+
+---
+
+RECENT SEGMENT (events from just before the current moment):
+
+{}
+
 "#,
             target_chars, transcript
         )
@@ -295,12 +517,339 @@ Be concise, aim for {} characters or less.
         Ok(take_prefix_chars(transcript, target_chars))
     }
 
+    // =========================================================================
+    // Two-segment compression
+    // =========================================================================
+
+    /// Find a safe split point in the unit list. The split ensures:
+    /// - Accumulated tokens up to `ratio` of total are placed in the older segment.
+    /// - The split never lands on a SystemGuard (they're always preserved).
+    /// - The split lands on a unit boundary (ToolRound, AssistantText, or UserMessage).
+    fn find_safe_split_point(&self, units: &[HistoryUnit], ratio: f64) -> usize {
+        let total_tokens = estimate_tokens_from_units(units);
+        let target = (total_tokens as f64 * ratio) as usize;
+        let mut accumulated = 0;
+
+        for (i, unit) in units.iter().enumerate() {
+            // Never split on a SystemGuard — they must be preserved
+            if matches!(unit, HistoryUnit::SystemGuard(_)) {
+                continue;
+            }
+            accumulated += unit.estimate_tokens();
+            if accumulated >= target {
+                return i + 1;
+            }
+        }
+
+        // Fallback: put everything in the older segment
+        units.len()
+    }
+
+    /// Summarize a transcript with a custom prompt builder function.
+    async fn summarize_with_prompt(
+        &self,
+        provider: &dyn LLMProvider,
+        transcript: &str,
+        target_chars: usize,
+        build_prompt: fn(&str, usize) -> String,
+    ) -> Result<String, AgentError> {
+        let request = ChatCompletionRequest {
+            messages: vec![
+                Message::system("You are a helpful assistant."),
+                Message::user(build_prompt(transcript, target_chars)),
+            ],
+            temperature: Some(0.3),
+            max_tokens: Some(1000),
+            tools: None,
+        };
+
+        let response = provider
+            .chat(request)
+            .await
+            .map_err(|e| AgentError::LlmError(e.to_string()))?;
+        Ok(response.content)
+    }
+
+    /// Summarize a segment of units with chunked fallback.
+    async fn summarize_units_segment(
+        &self,
+        provider: &dyn LLMProvider,
+        messages: &[ChatMessage],
+        transcript: &str,
+        target_chars: usize,
+        build_prompt: fn(&str, usize) -> String,
+    ) -> Result<String, AgentError> {
+        if messages.is_empty() {
+            return Ok(String::new());
+        }
+
+        if char_count(transcript) <= target_chars {
+            return self
+                .summarize_with_prompt(provider, transcript, target_chars, build_prompt)
+                .await;
+        }
+
+        // Use chunked summarization for oversized transcripts
+        let target = target_chars.max(1);
+        let mut layer = Self::chunk_messages_for_summary(messages, target);
+
+        if layer.is_empty() {
+            layer.push(transcript.to_string());
+        }
+
+        for _ in 0..6 {
+            if layer.len() == 1 && char_count(&layer[0]) <= target {
+                return self
+                    .summarize_with_prompt(provider, &layer[0], target, build_prompt)
+                    .await;
+            }
+
+            let per_chunk_target = (target / layer.len().max(1)).max(500).min(target);
+            let mut summaries = Vec::with_capacity(layer.len());
+            for chunk in &layer {
+                summaries.push(
+                    self.summarize_with_prompt(provider, chunk, per_chunk_target, build_prompt)
+                        .await?,
+                );
+            }
+
+            if summaries.len() == 1 {
+                let summary = summaries.pop().unwrap_or_default();
+                if char_count(&summary) <= target {
+                    return Ok(summary);
+                }
+                layer = Self::split_text_chunks(&summary, target);
+                continue;
+            }
+
+            let merged = summaries.join("\n\n");
+            if char_count(&merged) <= target {
+                return self
+                    .summarize_with_prompt(provider, &merged, target, build_prompt)
+                    .await;
+            }
+
+            layer = Self::split_text_chunks(&merged, target);
+        }
+
+        Ok(take_prefix_chars(transcript, target))
+    }
+
+    /// Main entry point for two-segment compression.
+    ///
+    /// Splits history into older and newer segments, then summarizes each
+    /// with LLM using different prompts and budget allocations. The result
+    /// contains no tool_calls or tool-result messages — only system summaries
+    /// and the original user message — eliminating any risk of API 400 errors
+    /// from orphaned tool call sequences.
+    pub async fn compress_two_segment(
+        &self,
+        history: &[ChatMessage],
+        provider_config: &LLMProviderConfig,
+    ) -> Result<Vec<ChatMessage>, AgentError> {
+        let tokens = estimate_tokens(history);
+        if tokens <= self.threshold() {
+            tracing::info!(
+                tokens = tokens,
+                threshold = self.threshold(),
+                msg_count = history.len(),
+                "Two-segment compression not needed (under threshold)"
+            );
+            return Ok(history.to_vec());
+        }
+
+        tracing::info!(
+            tokens = tokens,
+            threshold = self.threshold(),
+            msg_count = history.len(),
+            "Starting two-segment compression"
+        );
+
+        // Step 1: Parse into atomic units
+        let units = parse_to_units(history);
+
+        // Step 2: Separate system guards + user messages from compressible units
+        let mut system_guards: Vec<ChatMessage> = Vec::new();
+        let mut user_messages: Vec<ChatMessage> = Vec::new();
+        let mut compressible: Vec<HistoryUnit> = Vec::new();
+
+        for unit in units {
+            match unit {
+                HistoryUnit::SystemGuard(msg) => system_guards.push(msg),
+                HistoryUnit::UserMessage(msg) => user_messages.push(msg),
+                other => compressible.push(other),
+            }
+        }
+
+        // Keep only the latest user message in full; older ones go into compression
+        let latest_user_msg = user_messages.pop();
+
+        // Step 3: Find split point in compressible units
+        let total_compressible = estimate_tokens_from_units(&compressible);
+        let split_point = if total_compressible > 0 {
+            self.find_safe_split_point(&compressible, 0.5)
+        } else {
+            0
+        };
+        let older_units: Vec<&HistoryUnit> = compressible[..split_point].iter().collect();
+        let newer_units: Vec<&HistoryUnit> = compressible[split_point..].iter().collect();
+
+        // Step 4: Build transcripts
+        let older_messages: Vec<ChatMessage> = older_units
+            .iter()
+            .flat_map(|u| match u {
+                HistoryUnit::ToolRound { assistant, results } => {
+                    let mut msgs = vec![assistant.clone()];
+                    msgs.extend(results.clone());
+                    msgs
+                }
+                HistoryUnit::AssistantText(msg) => vec![msg.clone()],
+                HistoryUnit::UserMessage(msg) => {
+                    // Older user messages go into the compressible transcript
+                    vec![msg.clone()]
+                }
+                _ => vec![],
+            })
+            .collect();
+
+        let newer_messages: Vec<ChatMessage> = newer_units
+            .iter()
+            .flat_map(|u| match u {
+                HistoryUnit::ToolRound { assistant, results } => {
+                    let mut msgs = vec![assistant.clone()];
+                    msgs.extend(results.clone());
+                    msgs
+                }
+                HistoryUnit::AssistantText(msg) => vec![msg.clone()],
+                _ => vec![],
+            })
+            .collect();
+
+        // Include older user messages in the older transcript
+        let older_user_msgs: Vec<ChatMessage> = user_messages
+            .iter()
+            .map(|m| {
+                let mut msg = m.clone();
+                msg.role = "user".to_string();
+                msg
+            })
+            .collect();
+        let all_older_messages: Vec<ChatMessage> = older_user_msgs
+            .iter()
+            .chain(older_messages.iter())
+            .cloned()
+            .collect();
+
+        let older_transcript = Self::build_transcript(&all_older_messages);
+        let newer_transcript = Self::build_transcript(&newer_messages);
+
+        // Step 5: Budget allocation
+        let total_budget = self.config.summary_max_chars;
+        let older_budget = (total_budget as f64 * OLDER_BUDGET_RATIO) as usize;
+        let newer_budget = (total_budget as f64 * NEWER_BUDGET_RATIO) as usize;
+
+        // Step 6: Create provider and run both summaries (can be parallel)
+        let runtime_config = AgentRuntimeConfig::from(provider_config.clone());
+        let provider = create_provider(runtime_config.provider)
+            .map_err(|e| AgentError::ProviderCreation(e.to_string()))?;
+
+        let (older_result, newer_result) = if older_units.is_empty() && newer_units.is_empty() {
+            (Ok(String::new()), Ok(String::new()))
+        } else if older_units.is_empty() {
+            let result = self
+                .summarize_units_segment(
+                    provider.as_ref(),
+                    &newer_messages,
+                    &newer_transcript,
+                    newer_budget,
+                    Self::build_light_summary_prompt,
+                )
+                .await;
+            (Ok(String::new()), result)
+        } else if newer_units.is_empty() {
+            let result = self
+                .summarize_units_segment(
+                    provider.as_ref(),
+                    &all_older_messages,
+                    &older_transcript,
+                    older_budget,
+                    Self::build_heavy_summary_prompt,
+                )
+                .await;
+            (result, Ok(String::new()))
+        } else {
+            let older_fut = self.summarize_units_segment(
+                provider.as_ref(),
+                &all_older_messages,
+                &older_transcript,
+                older_budget,
+                Self::build_heavy_summary_prompt,
+            );
+            let newer_fut = self.summarize_units_segment(
+                provider.as_ref(),
+                &newer_messages,
+                &newer_transcript,
+                newer_budget,
+                Self::build_light_summary_prompt,
+            );
+            tokio::join!(older_fut, newer_fut)
+        };
+
+        let older_summary = older_result?;
+        let newer_summary = newer_result?;
+
+        // Step 7: Assemble compressed history
+        let mut compressed: Vec<ChatMessage> = Vec::with_capacity(4);
+
+        // System guards first
+        compressed.extend(system_guards);
+
+        // Latest user message
+        if let Some(user_msg) = latest_user_msg {
+            compressed.push(user_msg);
+        }
+
+        // Heavy compression summary (older)
+        if !older_summary.is_empty() {
+            compressed.push(ChatMessage::system_with_context(
+                format!("## 较早的操作记录（已压缩）\n\n{}", older_summary),
+                Some(SYSTEM_CONTEXT_HISTORY_COMPACTION_OLDER.to_string()),
+            ));
+        }
+
+        // Light compression summary (newer)
+        if !newer_summary.is_empty() {
+            compressed.push(ChatMessage::system_with_context(
+                format!(
+                    "## 近期操作记录（轻度压缩，保留了更多细节）\n\n{}\n\n---\n以上为近期操作记录。如需精确数据，可使用工具重新读取相关文件。",
+                    newer_summary
+                ),
+                Some(SYSTEM_CONTEXT_HISTORY_COMPACTION_NEWER.to_string()),
+            ));
+        }
+
+        tracing::info!(
+            original_tokens = tokens,
+            final_tokens = estimate_tokens(&compressed),
+            final_msg_count = compressed.len(),
+            older_units = older_units.len(),
+            newer_units = newer_units.len(),
+            "Two-segment compression completed"
+        );
+
+        Ok(compressed)
+    }
+
+    // =========================================================================
+    // Legacy methods (kept for backward compatibility during migration)
+    // =========================================================================
+
     /// Create a new compressor with the given context window size.
     pub fn new(context_window: usize) -> Self {
         Self {
             config: ContextCompressionConfig::default(),
             context_window,
-            threshold_ratio: 0.5,
+            threshold_ratio: DEFAULT_THRESHOLD_RATIO,
         }
     }
 
@@ -323,11 +872,11 @@ Be concise, aim for {} characters or less.
         Self {
             config,
             context_window,
-            threshold_ratio: 0.5,
+            threshold_ratio: DEFAULT_THRESHOLD_RATIO,
         }
     }
 
-    /// Get the compression threshold in tokens.
+    /// Get the compression threshold in tokens (70% of context window).
     fn threshold(&self) -> usize {
         (self.context_window as f64 * self.threshold_ratio) as usize
     }
@@ -707,7 +1256,7 @@ mod tests {
     #[test]
     fn test_threshold() {
         let compressor = ContextCompressor::new(128_000);
-        assert_eq!(compressor.threshold(), 64_000);
+        assert_eq!(compressor.threshold(), 89_600); // 70% of 128_000
     }
 
     #[test]

src/gateway/compaction.rs

@@ -6,97 +6,60 @@ use crate::agent::AgentError;
 
 use super::session::Session;
 
+/// Run two-segment history compression synchronously.
+///
+/// Unlike the previous background approach (tokio::spawn), this holds the
+/// session lock during the LLM calls (2–5 seconds).  Since the agent loop
+/// has already finished by this point there is no response-time impact, and
+/// the synchronous guarantee means the next execution always starts with
+/// freshly compacted history.
 pub(crate) async fn schedule_background_history_compaction(
     session: Arc<Mutex<Session>>,
     chat_id: impl Into<String>,
 ) -> Result<(), AgentError> {
     let chat_id = chat_id.into();
 
-    let snapshot = {
+    let mut session_guard = session.lock().await;
-        let mut session_guard = session.lock().await;
+    session_guard.ensure_persistent_session(&chat_id)?;
-        let session_record = session_guard.ensure_persistent_session(&chat_id)?;
+    session_guard.ensure_chat_loaded(&chat_id)?;
-        session_guard.ensure_chat_loaded(&chat_id)?;
 
-        let history = session_guard.get_or_create_history(&chat_id).clone();
+    let history = session_guard.get_or_create_history(&chat_id).clone();
-        let compressor = session_guard.compressor().clone();
+    let compressor = session_guard.compressor().clone();
-        if !compressor.should_compress(&history) {
-            return Ok(());
-        }
 
-        if !session_guard.try_start_background_compaction(&chat_id) {
+    if !compressor.should_compress(&history) {
-            return Ok(());
+        return Ok(());
-        }
+    }
 
-        (
+    let store = session_guard.store();
-            session_guard.store(),
+    let session_id = session_guard.persistent_session_id(&chat_id);
-            session_guard.persistent_session_id(&chat_id),
+    let provider_config = session_guard.provider_config().clone();
-            session_record.message_count,
-            history,
-            compressor,
-            session_guard.provider_config().clone(),
-        )
-    };
 
-    let (
+    tracing::info!(
-        store,
+        chat_id = %chat_id,
-        session_id,
+        msg_count = history.len(),
-        snapshot_end_seq,
+        "Starting synchronous two-segment compression"
-        history,
+    );
-        compressor,
-        provider_config,
-    ) = snapshot;
-    let session_for_task = session.clone();
-    let chat_id_for_task = chat_id.clone();
 
-    tokio::spawn(async move {
+    // Synchronous compression — holds lock during LLM calls.
-        tracing::info!(chat_id = %chat_id_for_task, snapshot_end_seq, "Starting background history compaction");
+    // compress_two_segment guarantees the result contains no tool_calls,
+    // so there is no risk of orphaned tool call sequences.
+    let compressed = compressor
+        .compress_two_segment(&history, &provider_config)
+        .await?;
 
-        let compaction_result = compressor
+    // Replace the entire history with the compressed result.
-            .build_compaction_plan(&history, &provider_config)
+    // Since we hold the lock, no concurrent modifications can occur.
-            .await;
+    store
-        let mut committed = false;
+        .replace_active_history(&session_id, &compressed)
+        .map_err(|e| AgentError::Other(format!("replace_active_history error: {}", e)))?;
 
-        match compaction_result {
+    tracing::info!(
-            Ok(Some(plan)) => match store.compact_active_history(
+        chat_id = %chat_id,
-                &session_id,
+        compressed_msg_count = compressed.len(),
-                snapshot_end_seq,
+        "Two-segment compression committed"
-                &plan.preserved_system_messages,
+    );
-                &plan.summary_message,
-                &plan.preserved_messages,
-            ) {
-                Ok(true) => {
-                    committed = true;
-                    tracing::info!(
-                        chat_id = %chat_id_for_task,
-                        snapshot_end_seq,
-                        compressed_turns = plan.compressed_turns,
-                        preserved_turns = plan.preserved_turns,
-                        "Background history compaction committed"
-                    );
-                }
-                Ok(false) => {
-                    tracing::info!(chat_id = %chat_id_for_task, snapshot_end_seq, "Background history compaction skipped due to stale snapshot");
-                }
-                Err(error) => {
-                    tracing::warn!(chat_id = %chat_id_for_task, error = %error, "Background history compaction commit failed");
-                }
-            },
-            Ok(None) => {
-                tracing::debug!(chat_id = %chat_id_for_task, "Background history compaction not needed after snapshot analysis");
-            }
-            Err(error) => {
-                tracing::warn!(chat_id = %chat_id_for_task, error = %error, "Background history compaction build failed");
-            }
-        }
 
-        let mut session_guard = session_for_task.lock().await;
+    session_guard.reload_chat_history(&chat_id)?;
-        if committed {
-            if let Err(error) = session_guard.reload_chat_history(&chat_id_for_task) {
-                tracing::warn!(chat_id = %chat_id_for_task, error = %error, "Failed to reload history after background compaction");
-            }
-        }
-        session_guard.finish_background_compaction(&chat_id_for_task);
-    });
 
     Ok(())
 }

src/gateway/session.rs

@@ -526,10 +526,12 @@ impl Session {
         &self.compressor
     }
 
+    #[allow(dead_code)]
     pub(crate) fn try_start_background_compaction(&mut self, chat_id: &str) -> bool {
         self.history.try_start_background_compaction(chat_id)
     }
 
+    #[allow(dead_code)]
     pub(crate) fn finish_background_compaction(&mut self, chat_id: &str) {
         self.history.finish_background_compaction(chat_id);
     }

src/gateway/session_history.rs

@@ -235,10 +235,12 @@ impl SessionHistory {
         Ok(())
     }
 
+    #[allow(dead_code)]
     pub(crate) fn try_start_background_compaction(&mut self, chat_id: &str) -> bool {
         self.compression_in_flight.insert(chat_id.to_string())
     }
 
+    #[allow(dead_code)]
     pub(crate) fn finish_background_compaction(&mut self, chat_id: &str) {
         self.compression_in_flight.remove(chat_id);
     }

src/storage/mod.rs

@@ -814,6 +814,59 @@ impl SessionStore {
         Ok(true)
     }
 
+    /// Replace the entire active history for a session.
+    ///
+    /// This is a simpler alternative to `compact_active_history` for when the
+    /// compressor has already produced a complete, validated message list
+    /// (e.g. two-segment compression).  It replaces all existing messages
+    /// with the new list in a single transaction.
+    pub fn replace_active_history(
+        &self,
+        session_id: &str,
+        messages: &[ChatMessage],
+    ) -> Result<(), StorageError> {
+        let conn = self.pool.get()?;
+        let tx = conn.unchecked_transaction()?;
+        let now = current_timestamp();
+
+        // Delete all existing messages for this session
+        tx.execute(
+            "DELETE FROM messages WHERE session_id = ?1",
+            params![session_id],
+        )?;
+
+        // Insert new messages with sequential seq numbers
+        let mut active_user_turn_count = 0_i64;
+        for (i, message) in messages.iter().enumerate() {
+            let seq = (i + 1) as i64;
+            if message.role == "user" {
+                active_user_turn_count += 1;
+            }
+            insert_message_with_seq(&tx, session_id, seq, message)?;
+        }
+
+        tx.execute(
+            "
+            UPDATE sessions
+            SET message_count = ?2,
+                user_turn_count = ?3,
+                updated_at = ?4,
+                last_active_at = ?4,
+                archived_at = NULL
+            WHERE id = ?1 AND deleted_at IS NULL
+            ",
+            params![
+                session_id,
+                messages.len() as i64,
+                active_user_turn_count,
+                now,
+            ],
+        )?;
+
+        tx.commit()?;
+        Ok(())
+    }
+
     pub fn mark_agent_prompt_reinjected(&self, session_id: &str) -> Result<(), StorageError> {
         let now = current_timestamp();
         let conn = self.pool.get()?;

src/storage/ports.rs

@@ -54,6 +54,15 @@ pub trait ConversationRepository: Send + Sync + 'static {
         summary_message: &ChatMessage,
         preserved_messages: &[ChatMessage],
     ) -> Result<bool, StorageError>;
+
+    /// Replace the entire active history for a session with new messages.
+    /// Used when the compressor has already produced a complete, validated
+    /// message list (e.g. two-segment compression).
+    fn replace_active_history(
+        &self,
+        session_id: &str,
+        messages: &[ChatMessage],
+    ) -> Result<(), StorageError>;
 }
 
 pub trait PromptInjectionRepository: Send + Sync + 'static {
@@ -235,6 +244,14 @@ impl ConversationRepository for super::SessionStore {
             preserved_messages,
         )
     }
+
+    fn replace_active_history(
+        &self,
+        session_id: &str,
+        messages: &[ChatMessage],
+    ) -> Result<(), StorageError> {
+        super::SessionStore::replace_active_history(self, session_id, messages)
+    }
 }
 
 impl PromptInjectionRepository for super::SessionStore {