consciousness/src/agent/mod.rs

// agent.rs — Core agent loop
//
// The simplest possible implementation of the agent pattern:
// send messages + tool definitions to the model, if it responds
// with tool calls then dispatch them and loop, if it responds
// with text then display it and wait for the next prompt.
//
// Uses streaming by default so text tokens appear as they're
// generated. Tool calls are accumulated from stream deltas and
// dispatched after the stream completes.
//
// The DMN (dmn.rs) is the outer loop that decides what prompts
// to send here. This module just handles single turns: prompt
// in, response out, tool calls dispatched.

pub mod api;
pub mod context;
pub mod oneshot;
pub mod tools;

use std::sync::Arc;
use anyhow::Result;
use tiktoken_rs::CoreBPE;

use api::{ApiClient, ToolCall};
use api::types::{ContentPart, Message, MessageContent, Role};
use context::{ConversationEntry, ContextState, ContextBudget};
use tools::{summarize_args, working_stack};

use crate::mind::log::ConversationLog;
use crate::agent::context::{ContextSection, SharedContextState};
use crate::mind::StreamTarget;
use crate::subconscious::learn;

// --- Activity tracking (RAII guards) ---

pub struct ActivityEntry {
    pub id: u64,
    pub label: String,
    pub started: std::time::Instant,
    /// Auto-expires this long after creation (or completion).
    pub expires_at: std::time::Instant,
}

/// RAII guard — marks the activity "(complete)" on drop, starts expiry timer.
pub struct ActivityGuard {
    agent: Arc<tokio::sync::Mutex<Agent>>,
    id: u64,
}

const ACTIVITY_LINGER: std::time::Duration = std::time::Duration::from_secs(5);

impl Drop for ActivityGuard {
    fn drop(&mut self) {
        if let Ok(mut ag) = self.agent.try_lock() {
            if let Some(entry) = ag.activities.iter_mut().find(|a| a.id == self.id) {
                entry.label.push_str(" (complete)");
                entry.expires_at = std::time::Instant::now() + ACTIVITY_LINGER;
            }
        }
    }
}

impl Agent {
    /// Register an activity, returns its ID. Caller creates the guard.
    pub fn push_activity(&mut self, label: impl Into<String>) -> u64 {
        self.expire_activities();
        let id = self.next_activity_id;
        self.next_activity_id += 1;
        self.activities.push(ActivityEntry {
            id, label: label.into(),
            started: std::time::Instant::now(),
            expires_at: std::time::Instant::now() + std::time::Duration::from_secs(3600),
        });
        id
    }

    /// Push a notification — auto-expires after 5 seconds.
    pub fn notify(&mut self, label: impl Into<String>) {
        self.expire_activities();
        let id = self.next_activity_id;
        self.next_activity_id += 1;
        self.activities.push(ActivityEntry {
            id, label: label.into(),
            started: std::time::Instant::now(),
            expires_at: std::time::Instant::now() + ACTIVITY_LINGER,
        });
    }

    /// Remove expired activities.
    pub fn expire_activities(&mut self) {
        let now = std::time::Instant::now();
        self.activities.retain(|a| a.expires_at > now);
    }
}

/// Create an activity guard from outside the lock.
pub fn activity_guard(agent: &Arc<tokio::sync::Mutex<Agent>>, id: u64) -> ActivityGuard {
    ActivityGuard { agent: agent.clone(), id }
}

/// Convenience: lock, push activity, unlock, return guard.
pub async fn start_activity(agent: &Arc<tokio::sync::Mutex<Agent>>, label: impl Into<String>) -> ActivityGuard {
    let id = agent.lock().await.push_activity(label);
    ActivityGuard { agent: agent.clone(), id }
}

/// Result of a single agent turn.
pub struct TurnResult {
    /// The text response (already sent through UI channel).
    #[allow(dead_code)]
    pub text: String,
    /// Whether the model called yield_to_user during this turn.
    pub yield_requested: bool,
    /// Whether any tools (other than yield_to_user) were called.
    pub had_tool_calls: bool,
    /// Number of tool calls that returned errors this turn.
    pub tool_errors: u32,
    /// Model name to switch to after this turn completes.
    pub model_switch: Option<String>,
    /// Agent requested DMN pause (full stop on autonomous behavior).
    pub dmn_pause: bool,
}

/// Accumulated state across tool dispatches within a single turn.
struct DispatchState {
    yield_requested: bool,
    had_tool_calls: bool,
    tool_errors: u32,
    model_switch: Option<String>,
    dmn_pause: bool,
}

impl DispatchState {
    fn new() -> Self {
        Self {
            yield_requested: false, had_tool_calls: false,
            tool_errors: 0, model_switch: None, dmn_pause: false,
        }
    }
}

pub struct Agent {
    client: ApiClient,
    tools: Vec<tools::Tool>,
    /// Last known prompt token count from the API (tracks context size).
    last_prompt_tokens: u32,
    /// Current reasoning effort level ("none", "low", "high").
    pub reasoning_effort: String,
    /// Sampling parameters — adjustable at runtime from the thalamus screen.
    pub temperature: f32,
    pub top_p: f32,
    pub top_k: u32,
    /// Active activities — RAII guards auto-remove on drop.
    pub activities: Vec<ActivityEntry>,
    next_activity_id: u64,
    /// Control tool flags — set by tool handlers, consumed by turn loop.
    pub pending_yield: bool,
    pub pending_model_switch: Option<String>,
    pub pending_dmn_pause: bool,
    /// Persistent conversation log — append-only record of all messages.
    conversation_log: Option<ConversationLog>,
    /// BPE tokenizer for token counting (cl100k_base — close enough
    /// for Claude and Qwen budget allocation, ~85-90% count accuracy).
    tokenizer: CoreBPE,
    /// Mutable context state — personality, working stack, etc.
    pub context: ContextState,
    /// Shared live context summary — TUI reads this directly for debug screen.
    pub shared_context: SharedContextState,
    /// App config — used to reload identity on compaction and model switching.
    pub app_config: crate::config::AppConfig,
    pub prompt_file: String,
    /// Stable session ID for memory-search dedup across turns.
    pub session_id: String,
    /// Incremented on compaction — UI uses this to detect resets.
    pub generation: u64,
    /// Agent orchestration state (surface-observe, journal, reflect).
    /// TODO: move to Session — it's session-level, not agent-level.
    pub agent_cycles: crate::subconscious::subconscious::AgentCycleState,
    /// Shared active tools — Agent writes, TUI reads.
    pub active_tools: crate::agent::tools::SharedActiveTools,
}

fn render_journal(entries: &[context::JournalEntry]) -> String {
    if entries.is_empty() { return String::new(); }
    let mut text = String::from("[Earlier — from your journal]\n\n");
    for entry in entries {
        use std::fmt::Write;
        writeln!(text, "## {}\n{}\n", entry.timestamp.format("%Y-%m-%dT%H:%M"), entry.content).ok();
    }
    text
}

impl Agent {
    pub fn new(
        client: ApiClient,
        system_prompt: String,
        personality: Vec<(String, String)>,
        app_config: crate::config::AppConfig,
        prompt_file: String,
        conversation_log: Option<ConversationLog>,
        shared_context: SharedContextState,
        active_tools: tools::SharedActiveTools,
    ) -> Self {
        let tokenizer = tiktoken_rs::cl100k_base()
            .expect("failed to load cl100k_base tokenizer");

        let context = ContextState {
            system_prompt: system_prompt.clone(),
            personality,
            journal: Vec::new(),
            working_stack: Vec::new(),
            entries: Vec::new(),
        };
        let session_id = format!("consciousness-{}", chrono::Utc::now().format("%Y%m%d-%H%M%S"));
        let agent_cycles = crate::subconscious::subconscious::AgentCycleState::new(&session_id);
        let mut agent = Self {
            client,
            tools: tools::tools(),
            last_prompt_tokens: 0,
            reasoning_effort: "none".to_string(),
            temperature: 0.6,
            top_p: 0.95,
            top_k: 20,
            activities: Vec::new(),
            next_activity_id: 0,
            pending_yield: false,
            pending_model_switch: None,
            pending_dmn_pause: false,
            conversation_log,
            tokenizer,
            context,
            shared_context,
            app_config,
            prompt_file,
            session_id,
            generation: 0,
            agent_cycles,
            active_tools,
        };

        agent.load_startup_journal();
        agent.load_working_stack();
        agent.publish_context_state();
        agent
    }

    /// Assemble the full message list for the API call from typed sources.
    /// System prompt + personality context + journal + conversation messages.
    fn assemble_api_messages(&self) -> Vec<Message> {
        let mut msgs = Vec::new();
        msgs.push(Message::system(&self.context.system_prompt));
        let ctx = self.context.render_context_message();
        if !ctx.is_empty() {
            msgs.push(Message::user(ctx));
        }
        let jnl = render_journal(&self.context.journal);
        if !jnl.is_empty() {
            msgs.push(Message::user(jnl));
        }
        msgs.extend(self.context.entries.iter().map(|e| e.api_message().clone()));
        msgs
    }

    /// Run agent orchestration cycle, returning structured output.
    fn run_agent_cycle(&mut self) -> crate::subconscious::subconscious::AgentCycleOutput {
        let transcript_path = self.conversation_log.as_ref()
            .map(|l| l.path().to_string_lossy().to_string())
            .unwrap_or_default();

        let session = crate::session::HookSession::from_fields(
            self.session_id.clone(),
            transcript_path,
            "UserPromptSubmit".into(),
        );

        self.agent_cycles.trigger(&session);
        std::mem::take(&mut self.agent_cycles.last_output)
    }

    /// Push a conversation message — stamped and logged.
    fn push_message(&mut self, mut msg: Message) {
        msg.stamp();
        let entry = ConversationEntry::Message(msg);
        self.push_entry(entry);
    }

    fn push_entry(&mut self, entry: ConversationEntry) {
        if let Some(ref log) = self.conversation_log {
            if let Err(e) = log.append(&entry) {
                eprintln!("warning: failed to log entry: {:#}", e);
            }
        }
        self.context.entries.push(entry);
    }

    /// Append streaming text to the last entry (creating a partial
    /// assistant entry if needed). Called by collect_stream per token batch.
    pub fn append_streaming(&mut self, text: &str) {
        if let Some(entry) = self.context.entries.last_mut() {
            let msg = entry.message_mut();
            if msg.role == Role::Assistant {
                msg.append_content(text);
                return;
            }
        }
        // No assistant entry yet — push a new partial one
        self.context.entries.push(ConversationEntry::Message(
            Message::assistant(text),
        ));
    }

    pub fn budget(&self) -> ContextBudget {
        let count_str = |s: &str| self.tokenizer.encode_with_special_tokens(s).len();
        let count_msg = |m: &Message| crate::agent::context::msg_token_count(&self.tokenizer, m);
        let window = crate::agent::context::context_window();
        self.context.budget(&count_str, &count_msg, window)
    }

    /// Send a user message and run the agent loop until the model
    /// produces a text response (no more tool calls). Streams text
    /// and tool activity through the UI channel.
    ///
    /// Takes Arc<Mutex<Agent>> and manages locking internally so the
    /// lock is never held across I/O (API streaming, tool dispatch).
    pub async fn turn(
        agent: Arc<tokio::sync::Mutex<Agent>>,
        user_input: &str,
        target: StreamTarget,
    ) -> Result<TurnResult> {
        // --- Pre-loop setup (lock 1): agent cycle, memories, user input ---
        let active_tools = {
            let mut finished = Vec::new();
            let tools = {
                let mut me = agent.lock().await;

                let cycle = me.run_agent_cycle();
                for key in &cycle.surfaced_keys {
                    if let Some(rendered) = crate::cli::node::render_node(
                        &crate::store::Store::load().unwrap_or_default(), key,
                    ) {
                        let mut msg = Message::user(format!(
                                "<system-reminder>\n--- {} (surfaced) ---\n{}\n</system-reminder>",
                                key, rendered,
                        ));
                        msg.stamp();
                        me.push_entry(ConversationEntry::Memory { key: key.clone(), message: msg });
                    }
                }
                if let Some(ref reflection) = cycle.reflection {
                    me.push_message(Message::user(format!(
                                "<system-reminder>\n--- subconscious reflection ---\n{}\n</system-reminder>",
                                reflection.trim(),
                    )));
                }

                // Collect completed background tool handles — remove from active list
                // but don't await yet (MutexGuard isn't Send).
                let mut tools = me.active_tools.lock().unwrap();
                let mut i = 0;
                while i < tools.len() {
                    if tools[i].handle.is_finished() {
                        finished.push(tools.remove(i));
                    } else {
                        i += 1;
                    }
                }

                me.active_tools.clone()
            };

            // Await finished handles without holding the agent lock
            let mut bg_results = Vec::new();
            for entry in finished {
                if let Ok((call, output)) = entry.handle.await {
                    bg_results.push((call, output));
                }
            }

            // Re-acquire to apply results and push user input
            {
                let mut me = agent.lock().await;
                let mut bg_ds = DispatchState::new();
                for (call, output) in bg_results {
                    me.apply_tool_result(&call, output, &mut bg_ds);
                }
                me.push_message(Message::user(user_input));
            }

            tools
        };

        let mut overflow_retries: u32 = 0;
        let mut empty_retries: u32 = 0;
        let mut ds = DispatchState::new();

        loop {
            // --- Lock 2: assemble messages, start stream ---
            let _thinking = start_activity(&agent, "thinking...").await;
            let (mut rx, _stream_guard) = {
                let me = agent.lock().await;
                let api_messages = me.assemble_api_messages();
                let sampling = api::SamplingParams {
                    temperature: me.temperature,
                    top_p: me.top_p,
                    top_k: me.top_k,
                };
                me.client.start_stream(
                    &api_messages,
                    &me.tools,
                    &me.reasoning_effort,
                    sampling,
                    None,
                )
            };
            // --- Lock released ---

            // --- Stream loop (no lock) ---
            let sr = api::collect_stream(
                &mut rx, &agent, &active_tools,
            ).await;
            let api::StreamResult {
                content, tool_calls, usage, finish_reason,
                error: stream_error, display_buf, in_tool_call,
            } = sr;
            // --- Stream complete ---

            // --- Lock 3: process results ---
            let (msg, pending) = {
                let mut me = agent.lock().await;

                // Handle stream errors with retry logic
                if let Some(e) = stream_error {
                    let err = anyhow::anyhow!("{}", e);
                    if crate::agent::context::is_context_overflow(&err) && overflow_retries < 2 {
                        overflow_retries += 1;
                        me.notify(format!("context overflow — retrying ({}/2)", overflow_retries));
                        me.compact();
                        continue;
                    }
                    if crate::agent::context::is_stream_error(&err) && empty_retries < 2 {
                        empty_retries += 1;
                        me.notify(format!("stream error — retrying ({}/2)", empty_retries));
                        drop(me);
                        tokio::time::sleep(std::time::Duration::from_secs(2)).await;
                        continue;
                    }
                    return Err(err);
                }

                if finish_reason.as_deref() == Some("error") {
                    let detail = if content.is_empty() { "no details".into() } else { content };
                    return Err(anyhow::anyhow!("model stream error: {}", detail));
                }

                // Flush remaining display buffer to streaming entry
                if !in_tool_call && !display_buf.is_empty() {
                    me.append_streaming(&display_buf);
                }

                // Pop the streaming entry — the proper entry gets pushed below
                // via build_response_message which handles tool calls, leaked
                // tool calls, etc. sync_from_agent handles the swap.
                if let Some(entry) = me.context.entries.last() {
                    if entry.message().role == Role::Assistant && entry.message().timestamp.is_none() {
                        me.context.entries.pop();
                    }
                }

                let msg = api::build_response_message(content, tool_calls);

                if let Some(usage) = &usage {
                    me.last_prompt_tokens = usage.prompt_tokens;
                    me.publish_context_state();
                }

                // Empty response — nudge and retry
                let has_content = msg.content.is_some();
                let has_tools = msg.tool_calls.as_ref().map_or(false, |tc| !tc.is_empty());
                if !has_content && !has_tools {
                    if empty_retries < 2 {
                        empty_retries += 1;
                        dbglog!(
                            "empty response, injecting nudge and retrying ({}/2)",
                            empty_retries,
                        );
                        me.push_message(Message::user(
                            "[system] Your previous response was empty. \
                             Please respond with text or use a tool."
                        ));
                        continue;
                    }
                } else {
                    empty_retries = 0;
                }

                // Collect non-background tool calls fired during streaming
                let mut tools_guard = active_tools.lock().unwrap();
                let mut non_bg = Vec::new();
                let mut i = 0;
                while i < tools_guard.len() {
                    if !tools_guard[i].background {
                        non_bg.push(tools_guard.remove(i));
                    } else {
                        i += 1;
                    }
                }
                (msg, non_bg)
            };

            if !pending.is_empty() {
                agent.lock().await.push_message(msg.clone());

                // Drop lock before awaiting tool handles
                let mut results = Vec::new();
                for entry in pending {
                    if let Ok(r) = entry.handle.await {
                        results.push(r);
                    }
                }
                // Reacquire to apply results
                let mut me = agent.lock().await;
                for (call, output) in results {
                    me.apply_tool_result(&call, output, &mut ds);
                }
                me.publish_context_state();
                continue;
            }

            // Tool calls (structured API path)
            if let Some(ref tool_calls) = msg.tool_calls {
                if !tool_calls.is_empty() {
                    agent.lock().await.push_message(msg.clone());
                    let calls: Vec<ToolCall> = tool_calls.clone();
                    // Drop lock before tool dispatch
                    for call in &calls {
                        Agent::dispatch_tool_call_unlocked(
                            &agent, &active_tools, call, &mut ds,
                        ).await;
                    }
                    continue;
                }
            }

            // Genuinely text-only response
            let text = msg.content_text().to_string();
            let mut me = agent.lock().await;
            me.push_message(msg);

            // Drain pending control flags
            if me.pending_yield { ds.yield_requested = true; me.pending_yield = false; }
            if me.pending_model_switch.is_some() { ds.model_switch = me.pending_model_switch.take(); }
            if me.pending_dmn_pause { ds.dmn_pause = true; me.pending_dmn_pause = false; }

            return Ok(TurnResult {
                text,
                yield_requested: ds.yield_requested,
                had_tool_calls: ds.had_tool_calls,
                tool_errors: ds.tool_errors,
                model_switch: ds.model_switch,
                dmn_pause: ds.dmn_pause,
            });
        }
    }

    /// Dispatch a tool call without holding the agent lock across I/O.
    /// Used by `turn()` which manages its own locking.
    async fn dispatch_tool_call_unlocked(
        agent: &Arc<tokio::sync::Mutex<Agent>>,
        active_tools: &crate::agent::tools::SharedActiveTools,
        call: &ToolCall,
        ds: &mut DispatchState,
    ) {
        let args: serde_json::Value = match serde_json::from_str(&call.function.arguments) {
            Ok(v) => v,
            Err(e) => {
                let err = format!("Error: malformed tool call arguments: {e}");
                let _act = start_activity(agent, format!("rejected: {} (bad args)", call.function.name)).await;
                let mut me = agent.lock().await;
                me.apply_tool_result(call, err, ds);
                return;
            }
        };

        let args_summary = summarize_args(&call.function.name, &args);
        let _calling = start_activity(agent, format!("calling: {}", call.function.name)).await;

        // Spawn tool, track it
        let call_clone = call.clone();
        let agent_handle = agent.clone();
        let handle = tokio::spawn(async move {
            let output = tools::dispatch_with_agent(&call_clone.function.name, &args, Some(agent_handle)).await;
            (call_clone, output)
        });
        active_tools.lock().unwrap().push(
            tools::ActiveToolCall {
                id: call.id.clone(),
                name: call.function.name.clone(),
                detail: args_summary,
                started: std::time::Instant::now(),
                background: false,
                handle,
            }
        );

        // Pop it back and await — no agent lock held
        let entry = {
            let mut tools = active_tools.lock().unwrap();
            tools.pop().unwrap()
        };
        if let Ok((call, output)) = entry.handle.await {
            // Brief lock to apply result
            let mut me = agent.lock().await;
            me.apply_tool_result(&call, output, ds);
        }
    }

    /// Apply a completed tool result to conversation state.
    fn apply_tool_result(
        &mut self,
        call: &ToolCall,
        output: String,
        ds: &mut DispatchState,
    ) {
        let args: serde_json::Value =
            serde_json::from_str(&call.function.arguments).unwrap_or_default();

        ds.had_tool_calls = true;
        if output.starts_with("Error:") {
            ds.tool_errors += 1;
        }

        self.active_tools.lock().unwrap().retain(|t| t.id != call.id);

        // Tag memory_render results for context deduplication
        if call.function.name == "memory_render" && !output.starts_with("Error:") {
            if let Some(key) = args.get("key").and_then(|v| v.as_str()) {
                let mut msg = Message::tool_result(&call.id, &output);
                msg.stamp();
                self.push_entry(ConversationEntry::Memory { key: key.to_string(), message: msg });
                self.publish_context_state();
                return;
            }
        }

        self.push_message(Message::tool_result(&call.id, &output));
    }

    /// Build context state summary for the debug screen.
    pub fn context_state_summary(&self, memory_scores: Option<&learn::MemoryScore>) -> Vec<ContextSection> {
        let count = |s: &str| self.tokenizer.encode_with_special_tokens(s).len();

        let mut sections = Vec::new();

        // System prompt
        sections.push(ContextSection {
            name: "System prompt".into(),
            tokens: count(&self.context.system_prompt),
            content: self.context.system_prompt.clone(),
            children: Vec::new(),
        });

        // Personality — parent with file children
        let personality_children: Vec<ContextSection> = self.context.personality.iter()
            .map(|(name, content)| ContextSection {
                name: name.clone(),
                tokens: count(content),
                content: content.clone(),
                children: Vec::new(),
            })
            .collect();
        let personality_tokens: usize = personality_children.iter().map(|c| c.tokens).sum();
        sections.push(ContextSection {
            name: format!("Personality ({} files)", personality_children.len()),
            tokens: personality_tokens,
            content: String::new(),
            children: personality_children,
        });

        // Journal
        {
            let journal_children: Vec<ContextSection> = self.context.journal.iter()
                .map(|entry| {
                    let preview: String = entry.content.lines()
                        .find(|l| !l.trim().is_empty())
                        .unwrap_or("").chars().take(60).collect();
                    ContextSection {
                        name: format!("{}: {}", entry.timestamp.format("%Y-%m-%dT%H:%M"), preview),
                        tokens: count(&entry.content),
                        content: entry.content.clone(),
                        children: Vec::new(),
                    }
                })
                .collect();
            let journal_tokens: usize = journal_children.iter().map(|c| c.tokens).sum();
            sections.push(ContextSection {
                name: format!("Journal ({} entries)", journal_children.len()),
                tokens: journal_tokens,
                content: String::new(),
                children: journal_children,
            });
        }

        // Working stack — instructions + items as children
        let instructions = std::fs::read_to_string(working_stack::instructions_path())
            .unwrap_or_default();
        let mut stack_children = vec![ContextSection {
            name: "Instructions".into(),
            tokens: count(&instructions),
            content: instructions,
            children: Vec::new(),
        }];
        for (i, item) in self.context.working_stack.iter().enumerate() {
            let marker = if i == self.context.working_stack.len() - 1 { "→" } else { " " };
            stack_children.push(ContextSection {
                name: format!("{} [{}] {}", marker, i, item),
                tokens: count(item),
                content: String::new(),
                children: Vec::new(),
            });
        }
        let stack_tokens: usize = stack_children.iter().map(|c| c.tokens).sum();
        sections.push(ContextSection {
            name: format!("Working stack ({} items)", self.context.working_stack.len()),
            tokens: stack_tokens,
            content: String::new(),
            children: stack_children,
        });

        // Memory nodes — extracted from Memory entries in the conversation
        let memory_entries: Vec<&ConversationEntry> = self.context.entries.iter()
            .filter(|e| e.is_memory())
            .collect();
        if !memory_entries.is_empty() {
            let node_children: Vec<ContextSection> = memory_entries.iter()
                .map(|entry| {
                    let key = match entry {
                        ConversationEntry::Memory { key, .. } => key.as_str(),
                        _ => unreachable!(),
                    };
                    let text = entry.message().content_text();
                    // Show node weight from graph (updated by incremental scorer)
                    let graph_weight = crate::hippocampus::store::Store::load().ok()
                        .and_then(|s| s.nodes.get(key).map(|n| n.weight));
                    // Show full matrix score if available
                    let matrix_score = memory_scores
                        .and_then(|s| s.memory_weights.iter()
                            .find(|(k, _)| k == key)
                            .map(|(_, v)| *v));
                    let label = match (graph_weight, matrix_score) {
                        (Some(w), Some(s)) => format!("{} (w:{:.2} score:{:.1})", key, w, s),
                        (Some(w), None) => format!("{} (w:{:.2})", key, w),
                        (None, Some(s)) => format!("{} (score:{:.1})", key, s),
                        (None, None) => key.to_string(),
                    };
                    ContextSection {
                        name: label,
                        tokens: count(text),
                        content: String::new(),
                        children: Vec::new(),
                    }
                })
                .collect();
            let node_tokens: usize = node_children.iter().map(|c| c.tokens).sum();
            sections.push(ContextSection {
                name: format!("Memory nodes ({} loaded)", memory_entries.len()),
                tokens: node_tokens,
                content: String::new(),
                children: node_children,
            });
        }

        // Conversation — each message as a child
        let conv_messages = &self.context.entries;
        let conv_children: Vec<ContextSection> = conv_messages.iter().enumerate()
            .map(|(i, entry)| {
                let m = entry.message();
                let text = m.content.as_ref()
                    .map(|c| c.as_text().to_string())
                    .unwrap_or_default();
                let tool_info = m.tool_calls.as_ref().map(|tc| {
                    tc.iter()
                        .map(|c| c.function.name.clone())
                        .collect::<Vec<_>>()
                        .join(", ")
                });
                let label = if entry.is_memory() {
                    if let ConversationEntry::Memory { key, .. } = entry {
                        format!("[memory: {}]", key)
                    } else { unreachable!() }
                } else {
                    match &tool_info {
                        Some(tools) => format!("[tool_call: {}]", tools),
                        None => {
                            let preview: String = text.chars().take(60).collect();
                            let preview = preview.replace('\n', " ");
                            if text.len() > 60 { format!("{}...", preview) } else { preview }
                        }
                    }
                };
                let tokens = count(&text);
                let cfg = crate::config::get();
                let role_name = if entry.is_memory() { "mem".to_string() } else {
                    match m.role {
                        Role::Assistant => cfg.assistant_name.clone(),
                        Role::User => cfg.user_name.clone(),
                        Role::Tool => "tool".to_string(),
                        Role::System => "system".to_string(),
                    }
                };
                // Show which memories were important for this response
                let children = if m.role == Role::Assistant {
                    memory_scores
                        .map(|s| s.important_memories_for_entry(i))
                        .unwrap_or_default()
                        .into_iter()
                        .map(|(key, score)| ContextSection {
                            name: format!("← {} ({:.1})", key, score),
                            tokens: 0,
                            content: String::new(),
                            children: Vec::new(),
                        })
                        .collect()
                } else {
                    Vec::new()
                };
                ContextSection {
                    name: format!("[{}] {}: {}", i, role_name, label),
                    tokens,
                    content: text,
                    children,
                }
            })
            .collect();
        let conv_tokens: usize = conv_children.iter().map(|c| c.tokens).sum();
        sections.push(ContextSection {
            name: format!("Conversation ({} messages)", conv_children.len()),
            tokens: conv_tokens,
            content: String::new(),
            children: conv_children,
        });

        sections
    }

    /// Load recent journal entries at startup for orientation.
    /// Uses the same budget logic as compaction but with empty conversation.
    /// Only parses the tail of the journal file (last 64KB) for speed.
    fn load_startup_journal(&mut self) {
        let store = match crate::store::Store::load() {
            Ok(s) => s,
            Err(_) => return,
        };

        // Find oldest message timestamp in conversation log
        let oldest_msg_ts = self.conversation_log.as_ref()
            .and_then(|log| log.oldest_timestamp());

        // Get journal entries from the memory graph
        let mut journal_nodes: Vec<_> = store.nodes.values()
            .filter(|n| n.node_type == crate::store::NodeType::EpisodicSession)
            .collect();
        let mut dbg = std::fs::OpenOptions::new().create(true).append(true)
                .open("/tmp/poc-journal-debug.log").ok();
        macro_rules! dbg_log {
            ($($arg:tt)*) => {
                if let Some(ref mut f) = dbg { use std::io::Write; let _ = writeln!(f, $($arg)*); }
            }
        }
        dbg_log!("[journal] {} nodes, oldest_msg={:?}", journal_nodes.len(), oldest_msg_ts);

        journal_nodes.sort_by_key(|n| n.created_at);
        if let Some(first) = journal_nodes.first() {
            dbg_log!("[journal] first created_at={}", first.created_at);
        }
        if let Some(last) = journal_nodes.last() {
            dbg_log!("[journal] last created_at={}", last.created_at);
        }

        // Find the cutoff index — entries older than conversation, plus one overlap
        let cutoff_idx = if let Some(cutoff) = oldest_msg_ts {
            let cutoff_ts = cutoff.timestamp();
            dbg_log!("[journal] cutoff timestamp={}", cutoff_ts);
            let mut idx = journal_nodes.len();
            for (i, node) in journal_nodes.iter().enumerate() {
                if node.created_at >= cutoff_ts {
                    idx = i + 1;
                    break;
                }
            }
            idx
        } else {
            journal_nodes.len()
        };
        dbg_log!("[journal] cutoff_idx={}", cutoff_idx);

        // Walk backwards from cutoff, accumulating entries within 15% of context
        let count = |s: &str| self.tokenizer.encode_with_special_tokens(s).len();
        let context_window = crate::agent::context::context_window();
        let journal_budget = context_window * 15 / 100;
        dbg_log!("[journal] budget={} tokens ({}*15%)", journal_budget, context_window);

        let mut entries = Vec::new();
        let mut total_tokens = 0;

        for node in journal_nodes[..cutoff_idx].iter().rev() {
            let tokens = count(&node.content);
            if total_tokens + tokens > journal_budget && !entries.is_empty() {
                break;
            }
            entries.push(context::JournalEntry {
                timestamp: chrono::DateTime::from_timestamp(node.created_at, 0)
                    .unwrap_or_default(),
                content: node.content.clone(),
            });
            total_tokens += tokens;
        }
        entries.reverse();
        dbg_log!("[journal] loaded {} entries, {} tokens", entries.len(), total_tokens);

        if entries.is_empty() {
            dbg_log!("[journal] no entries!");
            return;
        }

        self.context.journal = entries;
        dbg_log!("[journal] context.journal now has {} entries", self.context.journal.len());
    }

    /// Called after any change to context state (working stack, etc).
    /// Load working stack from disk.
    fn load_working_stack(&mut self) {
        if let Ok(data) = std::fs::read_to_string(working_stack::file_path()) {
            if let Ok(stack) = serde_json::from_str::<Vec<String>>(&data) {
                self.context.working_stack = stack;
            }
        }
    }

    /// Push the current context summary to the shared state for the TUI to read.
    pub fn publish_context_state(&self) {
        self.publish_context_state_with_scores(None);
    }

    pub fn publish_context_state_with_scores(&self, memory_scores: Option<&learn::MemoryScore>) {
        let summary = self.context_state_summary(memory_scores);
        if let Ok(mut dbg) = std::fs::OpenOptions::new().create(true).append(true)
                .open("/tmp/poc-journal-debug.log") {
            use std::io::Write;
            for s in &summary {
                let _ = writeln!(dbg, "[publish] {} ({} tokens, {} children)", s.name, s.tokens, s.children.len());
            }
        }
        if let Ok(mut state) = self.shared_context.write() {
            *state = summary;
        }
    }

    /// Replace base64 image data in older messages with text placeholders.
    /// Keeps the 2 most recent images live (enough for motion/comparison).
    /// The tool result message before each image records what was loaded.
    pub fn age_out_images(&mut self) {
        // Find image entries newest-first, skip 1 (caller is about to add another)
        let to_age: Vec<usize> = self.context.entries.iter().enumerate()
            .rev()
            .filter(|(_, e)| {
                if let Some(MessageContent::Parts(parts)) = &e.message().content {
                    parts.iter().any(|p| matches!(p, ContentPart::ImageUrl { .. }))
                } else { false }
            })
            .map(|(i, _)| i)
            .skip(1) // keep 1 existing + 1 about to be added = 2 live
            .collect();

        for i in to_age {
            let msg = self.context.entries[i].message_mut();
            if let Some(MessageContent::Parts(parts)) = &msg.content {
                let mut replacement = String::new();
                for part in parts {
                    match part {
                        ContentPart::Text { text } => {
                            if !replacement.is_empty() { replacement.push('\n'); }
                            replacement.push_str(text);
                        }
                        ContentPart::ImageUrl { .. } => {
                            if !replacement.is_empty() { replacement.push('\n'); }
                            replacement.push_str("[image aged out]");
                        }
                    }
                }
                msg.content = Some(MessageContent::Text(replacement));
            }
        }
        self.generation += 1;
    }

    /// Strip ephemeral tool calls from the conversation history.
    ///
    /// Last prompt token count reported by the API.
    pub fn last_prompt_tokens(&self) -> u32 {
        self.last_prompt_tokens
    }

    /// Rebuild the context window: reload identity, dedup, trim, reload journal.
    pub fn compact(&mut self) {
        // Reload identity from config
        match crate::config::reload_for_model(&self.app_config, &self.prompt_file) {
            Ok((system_prompt, personality)) => {
                self.context.system_prompt = system_prompt;
                self.context.personality = personality;
            }
            Err(e) => {
                eprintln!("warning: failed to reload identity: {:#}", e);
            }
        }

        let before = self.context.entries.len();
        let before_mem = self.context.entries.iter().filter(|e| e.is_memory()).count();
        let before_conv = before - before_mem;

        // Dedup memory, trim to budget, reload journal
        let entries = self.context.entries.clone();
        self.context.entries = crate::agent::context::trim_entries(
            &self.context,
            &entries,
            &self.tokenizer,
        );

        let after = self.context.entries.len();
        let after_mem = self.context.entries.iter().filter(|e| e.is_memory()).count();
        let after_conv = after - after_mem;

        dbglog!("[compact] entries: {} → {} (mem: {} → {}, conv: {} → {})",
            before, after, before_mem, after_mem, before_conv, after_conv);

        let budget = self.budget();
        dbglog!("[compact] budget: {}", budget.status_string());

        self.load_startup_journal();
        self.generation += 1;
        self.last_prompt_tokens = 0;
        self.publish_context_state();
    }

    /// Restore from the conversation log. Builds the context window
    /// the same way compact() does — journal summaries for old messages,
    /// raw recent messages. This is the unified startup path.
    /// Returns true if the log had content to restore.
    pub fn restore_from_log(&mut self) -> bool {
        let entries = match &self.conversation_log {
            Some(log) => match log.read_tail(64 * 1024 * 1024) {
                Ok(entries) if !entries.is_empty() => entries,
                _ => return false,
            },
            None => return false,
        };

        // Load extra — compact() will dedup, trim, reload identity + journal
        let all: Vec<_> = entries.into_iter()
            .filter(|e| e.message().role != Role::System)
            .collect();
        let mem_count = all.iter().filter(|e| e.is_memory()).count();
        let conv_count = all.len() - mem_count;
        dbglog!("[restore] loaded {} entries from log (mem: {}, conv: {})",
            all.len(), mem_count, conv_count);
        self.context.entries = all;
        self.compact();
        // Estimate prompt tokens from budget so status bar isn't 0 on startup
        let b = self.budget();
        self.last_prompt_tokens = b.used() as u32;
        true
    }

    /// Replace the API client (for model switching).
    pub fn swap_client(&mut self, new_client: ApiClient) {
        self.client = new_client;
    }

    /// Get the model identifier.
    pub fn model(&self) -> &str {
        &self.client.model
    }

    /// Get the conversation entries for persistence.
    pub fn entries(&self) -> &[ConversationEntry] {
        &self.context.entries
    }

    /// Mutable access to conversation entries (for /retry).
    pub fn client_clone(&self) -> ApiClient {
        self.client.clone()
    }

    pub fn entries_mut(&mut self) -> &mut Vec<ConversationEntry> {
        &mut self.context.entries
    }

}