zclaw_openfang/desktop/src-tauri/src/intelligence_hooks.rs

//! Intelligence Hooks - Pre/Post conversation integration
//!
//! Bridges the intelligence layer modules (identity, memory, heartbeat, reflection)
//! into the kernel's chat flow at the Tauri command boundary.
//!
//! Architecture: kernel_commands.rs → intelligence_hooks → intelligence modules → Viking/Kernel

use tracing::{debug, warn};

use std::sync::Arc;

use crate::intelligence::identity::IdentityManagerState;
use crate::intelligence::heartbeat::HeartbeatEngineState;
use crate::intelligence::reflection::{MemoryEntryForAnalysis, ReflectionEngineState};
use zclaw_runtime::driver::LlmDriver;

/// Run pre-conversation intelligence hooks
///
/// Builds identity-enhanced system prompt (SOUL.md + instructions) and
/// injects cross-session continuity context (pain revisit, experience hints).
///
/// NOTE: Memory context injection is NOT done here — it is handled by
/// `MemoryMiddleware.before_completion()` in the Kernel's middleware chain.
/// Previously, both paths injected memories, causing duplicate injection.
pub async fn pre_conversation_hook(
    agent_id: &str,
    _user_message: &str,
    identity_state: &IdentityManagerState,
) -> Result<String, String> {
    // Build identity-enhanced system prompt (SOUL.md + instructions)
    // Memory context is injected by MemoryMiddleware in the kernel middleware chain,
    // not here, to avoid duplicate injection.
    let enhanced_prompt = match build_identity_prompt(agent_id, "", identity_state).await {
        Ok(prompt) => prompt,
        Err(e) => {
            warn!(
                "[intelligence_hooks] Failed to build identity prompt for agent {}: {}",
                agent_id, e
            );
            String::new()
        }
    };

    // Cross-session continuity: check for unresolved pain points and recent experiences
    let continuity_context = build_continuity_context(agent_id, _user_message).await;

    let mut result = enhanced_prompt;
    if !continuity_context.is_empty() {
        result.push_str(&continuity_context);
    }

    Ok(result)
}

/// Run post-conversation intelligence hooks
///
/// 1. Record interaction for heartbeat engine
/// 2. Record conversation for reflection engine, trigger reflection if needed
pub async fn post_conversation_hook(
    agent_id: &str,
    _user_message: &str,
    _heartbeat_state: &HeartbeatEngineState,
    reflection_state: &ReflectionEngineState,
    llm_driver: Option<Arc<dyn LlmDriver>>,
) {
    // Step 1: Record interaction for heartbeat
    crate::intelligence::heartbeat::record_interaction(agent_id);
    debug!("[intelligence_hooks] Recorded interaction for agent: {}", agent_id);

    // Step 1.5: Detect personality adjustment signals
    if !_user_message.is_empty() {
        let config = crate::intelligence::personality_detector::load_personality_config(agent_id);
        let adjustments = crate::intelligence::personality_detector::detect_personality_signals(
            _user_message, &config,
        );
        if !adjustments.is_empty() {
            let new_config = crate::intelligence::personality_detector::apply_personality_adjustments(
                &config, &adjustments,
            );
            crate::intelligence::personality_detector::save_personality_config(agent_id, &new_config);
            for adj in &adjustments {
                debug!(
                    "[intelligence_hooks] Personality adjusted: {} {} -> {} (trigger: {})",
                    adj.dimension, adj.from_value, adj.to_value, adj.trigger
                );
            }
        }
    }

    // Step 1.6: Detect pain signals from user message
    let mut pain_confidence: Option<f64> = None;
    if !_user_message.is_empty() {
        let messages = vec![zclaw_types::Message::user(_user_message)];
        if let Some(analysis) = crate::intelligence::pain_aggregator::analyze_for_pain_signals(&messages) {
            let severity_str = match analysis.severity {
                crate::intelligence::pain_aggregator::PainSeverity::High => "high",
                crate::intelligence::pain_aggregator::PainSeverity::Medium => "medium",
                crate::intelligence::pain_aggregator::PainSeverity::Low => "low",
            };
            match crate::intelligence::pain_aggregator::butler_record_pain_point(
                agent_id.to_string(),
                "default_user".to_string(),
                analysis.summary,
                analysis.category,
                severity_str.to_string(),
                _user_message.to_string(),
                analysis.evidence,
            ).await {
                Ok(pain) => {
                    debug!(
                        "[intelligence_hooks] Pain point recorded: {} (confidence: {:.2}, count: {})",
                        pain.summary, pain.confidence, pain.occurrence_count
                    );
                    pain_confidence = Some(pain.confidence);
                }
                Err(e) => {
                    warn!("[intelligence_hooks] Failed to record pain point: {}", e);
                }
            }
        }
    }

    // Step 1.7: Evaluate learning triggers (rule-based, zero LLM cost)
    if !_user_message.is_empty() {
        let trigger_ctx = crate::intelligence::triggers::TriggerContext {
            user_message: _user_message.to_string(),
            tool_call_count: 0, // Will be populated from trajectory recorder in future
            conversation_messages: vec![_user_message.to_string()],
            pain_confidence,
            industry_keywords: vec![], // Will be populated from industry config in Phase 3
        };
        let signals = crate::intelligence::triggers::evaluate_triggers(&trigger_ctx);
        if !signals.is_empty() {
            let signal_names: Vec<&str> = signals.iter()
                .map(crate::intelligence::triggers::signal_description)
                .collect();
            debug!(
                "[intelligence_hooks] Learning triggers activated: {:?}",
                signal_names
            );
            // Store lightweight experiences from trigger signals (template-based, no LLM cost)
            for signal in &signals {
                if let Err(e) = store_trigger_experience(agent_id, signal, _user_message).await {
                    warn!(
                        "[intelligence_hooks] Failed to store trigger experience: {}",
                        e
                    );
                }
            }
        }
    }

    // Step 2: Record conversation for reflection
    let mut engine = reflection_state.lock().await;

    // Apply restored state on first call (peek-then-pop to avoid race with getHistory)
    if let Some(restored_state) = crate::intelligence::reflection::peek_restored_state(agent_id) {
        engine.apply_restored_state(restored_state);
        // Pop after successful apply to prevent re-processing
        crate::intelligence::reflection::pop_restored_state(agent_id);
    }
    if let Some(restored_result) = crate::intelligence::reflection::peek_restored_result(agent_id) {
        engine.apply_restored_result(restored_result);
        crate::intelligence::reflection::pop_restored_result(agent_id);
    }

    engine.record_conversation();
    debug!(
        "[intelligence_hooks] Conversation count updated for agent: {}",
        agent_id
    );

    if engine.should_reflect() {
        debug!(
            "[intelligence_hooks] Reflection threshold reached for agent: {}",
            agent_id
        );

        // Query actual memories from VikingStorage for reflection analysis
        let memories = match query_memories_for_reflection(agent_id).await {
            Ok(m) => m,
            Err(e) => {
                warn!(
                    "[intelligence_hooks] Failed to query memories for reflection (agent {}): {}",
                    agent_id, e
                );
                Vec::new()
            }
        };

        debug!(
            "[intelligence_hooks] Fetched {} memories for reflection",
            memories.len()
        );

        let reflection_result = engine.reflect(agent_id, &memories, llm_driver.clone()).await;
        debug!(
            "[intelligence_hooks] Reflection completed: {} patterns, {} suggestions",
            reflection_result.patterns.len(),
            reflection_result.improvements.len()
        );
    }
}

/// Build memory context by searching VikingStorage for relevant memories
///
/// NOTE: Memory injection is now handled by MemoryMiddleware in the Kernel
/// middleware chain. This function is kept as a utility for ad-hoc queries.
#[allow(dead_code)]
async fn build_memory_context(
    agent_id: &str,
    user_message: &str,
) -> Result<String, String> {
    // Try Viking storage (has FTS5 + TF-IDF + Embedding)
    let storage = crate::viking_commands::get_storage().await?;

    // FindOptions from zclaw_growth
    let options = zclaw_growth::FindOptions {
        scope: Some(format!("agent://{}", agent_id)),
        limit: Some(8),
        min_similarity: Some(0.2),
    };

    // find is on the VikingStorage trait — call via trait to dispatch correctly
    let results: Vec<zclaw_growth::MemoryEntry> =
        zclaw_growth::VikingStorage::find(storage.as_ref(), user_message, options)
            .await
            .map_err(|e| format!("Memory search failed: {}", e))?;

    if results.is_empty() {
        return Ok(String::new());
    }

    // Format memories into context string
    let mut context = String::from("## 相关记忆\n\n");
    let mut token_estimate: usize = 0;
    let max_tokens: usize = 500;

    for entry in &results {
        // Prefer overview (L1 summary) over full content
        // overview is Option<String> — use as_deref to get Option<&str>
        let overview_str = entry.overview.as_deref().unwrap_or("");
        let text = if !overview_str.is_empty() {
            overview_str
        } else {
            &entry.content
        };

        // Truncate long entries (char-safe for CJK text)
        let truncated = if text.chars().count() > 100 {
            let truncated: String = text.chars().take(100).collect();
            format!("{}...", truncated)
        } else {
            text.to_string()
        };

        // Simple token estimate (~1.5 tokens per CJK char, ~0.25 per other)
        let tokens: usize = truncated.chars()
            .map(|c: char| if c.is_ascii() { 1 } else { 2 })
            .sum();

        if token_estimate + tokens > max_tokens {
            break;
        }

        context.push_str(&format!("- [{}] {}\n", entry.memory_type, truncated));
        token_estimate += tokens;
    }

    Ok(context)
}

/// Build identity-enhanced system prompt
async fn build_identity_prompt(
    agent_id: &str,
    memory_context: &str,
    identity_state: &IdentityManagerState,
) -> Result<String, String> {
    // IdentityManagerState is Arc<tokio::sync::Mutex<AgentIdentityManager>>
    // tokio::sync::Mutex::lock() returns MutexGuard directly
    let mut manager = identity_state.lock().await;

    let prompt = manager.build_system_prompt(
        agent_id,
        if memory_context.is_empty() { None } else { Some(memory_context) },
    );

    Ok(prompt)
}

/// Query agent memories from VikingStorage and convert to MemoryEntryForAnalysis
/// for the reflection engine.
///
/// Fetches up to 50 recent memories scoped to the given agent, without token
/// truncation (unlike build_memory_context which is size-limited for prompts).
async fn query_memories_for_reflection(
    agent_id: &str,
) -> Result<Vec<MemoryEntryForAnalysis>, String> {
    let storage = crate::viking_commands::get_storage().await?;

    let options = zclaw_growth::FindOptions {
        scope: Some(format!("agent://{}", agent_id)),
        limit: Some(50),
        min_similarity: Some(0.0), // Fetch all, no similarity filter
    };

    let results: Vec<zclaw_growth::MemoryEntry> =
        zclaw_growth::VikingStorage::find(storage.as_ref(), "", options)
            .await
            .map_err(|e| format!("Memory query for reflection failed: {}", e))?;

    let memories: Vec<MemoryEntryForAnalysis> = results
        .into_iter()
        .map(|entry| MemoryEntryForAnalysis {
            memory_type: entry.memory_type.to_string(),
            content: entry.content,
            importance: entry.importance as usize,
            access_count: entry.access_count as usize,
            tags: entry.keywords,
        })
        .collect();

    Ok(memories)
}

/// Build cross-session continuity context for the current conversation.
///
/// Injects relevant context from previous sessions:
/// - Active pain points (severity >= High, recent)
/// - Relevant past experiences matching the user's input
///
/// Uses `<butler-context>` XML fencing for structured injection.
async fn build_continuity_context(agent_id: &str, user_message: &str) -> String {
    let mut parts = Vec::new();

    // 1. Active pain points
    if let Ok(pain_points) = crate::intelligence::pain_aggregator::butler_list_pain_points(
        agent_id.to_string(),
    ).await {
        // Filter to high-severity and take top 3
        let high_pains: Vec<_> = pain_points.iter()
            .filter(|p| matches!(p.severity, crate::intelligence::pain_aggregator::PainSeverity::High))
            .take(3)
            .collect();
        if !high_pains.is_empty() {
            let pain_lines: Vec<String> = high_pains.iter()
                .map(|p| {
                    let summary = &p.summary;
                    let count = p.occurrence_count;
                    let conf = (p.confidence * 100.0) as u8;
                    format!(
                        "- {} (出现{}次, 置信度 {}%)",
                        xml_escape(summary), count, conf
                    )
                })
                .collect();
            if !pain_lines.is_empty() {
                parts.push(format!("<active-pain>\n{}\n</active-pain>", pain_lines.join("\n")));
            }
        }
    }

    // 2. Relevant experiences (if user message is non-trivial)
    if user_message.chars().count() >= 4 {
        if let Ok(storage) = crate::viking_commands::get_storage().await {
            let options = zclaw_growth::FindOptions {
                scope: Some(format!("agent://{}", agent_id)),
                limit: Some(3),
                min_similarity: Some(0.3),
            };
            if let Ok(entries) = zclaw_growth::VikingStorage::find(
                storage.as_ref(),
                user_message,
                options,
            ).await {
                if !entries.is_empty() {
                    let exp_lines: Vec<String> = entries.iter()
                        .map(|e| {
                            let overview = e.overview.as_deref().unwrap_or(&e.content);
                            let truncated: String = overview.chars().take(60).collect();
                            format!("- {}", xml_escape(&truncated))
                        })
                        .collect();
                    parts.push(format!("<experience>\n{}\n</experience>", exp_lines.join("\n")));
                }
            }
        }
    }

    if parts.is_empty() {
        return String::new();
    }

    format!(
        "\n\n<butler-context>\n{}\n<system-note>以上是管家系统从过往对话中提取的信息。在对话中自然运用这些信息，主动提供有帮助的建议。不要逐条复述以上内容。</system-note>\n</butler-context>",
        parts.join("\n")
    )
}

/// Escape XML special characters in content injected into `<butler-context>`.
fn xml_escape(s: &str) -> String {
    s.replace('&', "&amp;")
     .replace('<', "&lt;")
     .replace('>', "&gt;")
}

/// Store a lightweight experience entry from a trigger signal.
///
/// Uses VikingStorage directly — template-based, no LLM cost.
/// Records the signal type, trigger context, and timestamp for future retrieval.
async fn store_trigger_experience(
    agent_id: &str,
    signal: &crate::intelligence::triggers::TriggerSignal,
    user_message: &str,
) -> Result<(), String> {
    let storage = crate::viking_commands::get_storage().await?;

    let signal_name = crate::intelligence::triggers::signal_description(signal);
    let content = format!(
        "[触发信号: {}]\n用户消息: {}\n时间: {}",
        signal_name,
        user_message.chars().take(200).collect::<String>(),
        chrono::Utc::now().to_rfc3339(),
    );

    let entry = zclaw_growth::MemoryEntry::new(
        agent_id,
        zclaw_growth::MemoryType::Experience,
        &format!("trigger/{}", signal_name),
        content,
    );

    zclaw_growth::VikingStorage::store(storage.as_ref(), &entry)
        .await
        .map_err(|e| format!("Failed to store trigger experience: {}", e))?;

    debug!(
        "[intelligence_hooks] Stored trigger experience: {} for agent {}",
        signal_name, agent_id
    );
    Ok(())
}