consciousness/src/agent/runner.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.

use anyhow::Result;
use tiktoken_rs::CoreBPE;

use std::io::Write;

use crate::agent::api::ApiClient;
use crate::agent::journal;
use crate::agent::log::ConversationLog;
use crate::agent::tools;
use crate::agent::tools::ProcessTracker;
use crate::agent::types::*;
use crate::agent::ui_channel::{ContextSection, SharedContextState, StatusInfo, StreamTarget, UiMessage, UiSender};

/// 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,
}

pub struct Agent {
    client: ApiClient,
    messages: Vec<Message>,
    tool_defs: Vec<ToolDef>,
    /// Last known prompt token count from the API (tracks context size).
    last_prompt_tokens: u32,
    /// Shared process tracker for bash tool — lets TUI show/kill running commands.
    pub process_tracker: ProcessTracker,
    /// Current reasoning effort level ("none", "low", "high").
    pub reasoning_effort: String,
    /// Persistent conversation log — append-only record of all messages.
    conversation_log: Option<ConversationLog>,
    /// Current context window budget breakdown.
    pub context_budget: ContextBudget,
    /// 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,
    /// Stable session ID for memory-search dedup across turns.
    session_id: String,
}

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

        let context = ContextState {
            system_prompt: system_prompt.clone(),
            personality,
            journal: String::new(),
            working_stack: Vec::new(),
            loaded_nodes: Vec::new(),
        };
        let session_id = format!("poc-agent-{}", chrono::Utc::now().format("%Y%m%d-%H%M%S"));
        let mut agent = Self {
            client,
            messages: Vec::new(),
            tool_defs,
            last_prompt_tokens: 0,
            process_tracker: ProcessTracker::new(),
            reasoning_effort: "none".to_string(),
            conversation_log,
            context_budget: ContextBudget::default(),
            tokenizer,
            context,
            shared_context,
            session_id,
        };

        // Load recent journal entries at startup for orientation
        agent.load_startup_journal();
        agent.load_working_stack();

        agent.push_context(Message::system(system_prompt));
        let rendered = agent.context.render_context_message();
        if !rendered.is_empty() {
            agent.push_context(Message::user(rendered));
        }
        if !agent.context.journal.is_empty() {
            agent.push_context(Message::user(agent.context.journal.clone()));
        }
        agent.measure_budget();
        agent.publish_context_state();
        agent
    }

    /// Run memory search for a given event, returning any output to inject.
    /// Direct library call — no subprocess needed since everything is one crate.
    fn run_hook(&self, event: &str, _prompt: &str) -> Option<String> {
        let transcript_path = self.conversation_log.as_ref()
            .map(|l| l.path().to_string_lossy().to_string())
            .unwrap_or_default();

        let hook_input = serde_json::json!({
            "hook_event_name": event,
            "session_id": self.session_id,
            "transcript_path": transcript_path,
        });

        let text = crate::memory_search::run_hook(&hook_input.to_string());
        if text.trim().is_empty() {
            None
        } else {
            Some(text)
        }
    }

    /// Push a conversation message — stamped and logged.
    fn push_message(&mut self, mut msg: Message) {
        msg.stamp();
        if let Some(ref log) = self.conversation_log {
            if let Err(e) = log.append(&msg) {
                eprintln!("warning: failed to log message: {:#}", e);
            }
        }
        self.messages.push(msg);
    }

    /// Push a context-only message (system prompt, identity context,
    /// journal summaries). Not logged — these are reconstructed on
    /// every startup/compaction.
    fn push_context(&mut self, msg: Message) {
        self.messages.push(msg);
    }

    /// Measure context window usage by category. Uses the BPE tokenizer
    /// for direct token counting (no chars/4 approximation).
    fn measure_budget(&mut self) {
        let mut id_tokens: usize = 0;
        let mem_tokens: usize = 0;
        let mut jnl_tokens: usize = 0;
        let mut conv_tokens: usize = 0;
        let mut in_conversation = false;

        for msg in &self.messages {
            let tokens = crate::agent::context::msg_token_count(&self.tokenizer, msg);

            if in_conversation {
                conv_tokens += tokens;
                continue;
            }

            match msg.role {
                Role::System => id_tokens += tokens,
                Role::User => {
                    let text = msg.content_text();
                    if text.starts_with("[Earlier in this conversation") {
                        jnl_tokens += tokens;
                    } else if text.starts_with("Your context was just rebuilt") {
                        jnl_tokens += tokens;
                    } else if jnl_tokens == 0 && conv_tokens == 0 {
                        // Static identity context (before any journal/conversation)
                        id_tokens += tokens;
                    } else {
                        in_conversation = true;
                        conv_tokens += tokens;
                    }
                }
                _ => {
                    in_conversation = true;
                    conv_tokens += tokens;
                }
            }
        }

        self.context_budget = ContextBudget {
            identity_tokens: id_tokens,
            memory_tokens: mem_tokens,
            journal_tokens: jnl_tokens,
            conversation_tokens: conv_tokens,
            window_tokens: crate::agent::context::model_context_window(&self.client.model),
        };
    }

    /// 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.
    pub async fn turn(
        &mut self,
        user_input: &str,
        ui_tx: &UiSender,
        target: StreamTarget,
    ) -> Result<TurnResult> {
        // Run poc-hook (memory search, notifications, context check)
        if let Some(hook_output) = self.run_hook("UserPromptSubmit", user_input) {
            let enriched = format!("{}\n\n<system-reminder>\n{}\n</system-reminder>",
                user_input, hook_output);
            self.push_message(Message::user(enriched));
        } else {
            self.push_message(Message::user(user_input));
        }

        let mut overflow_retries: u32 = 0;
        let mut empty_retries: u32 = 0;
        let mut ds = DispatchState {
            yield_requested: false,
            had_tool_calls: false,
            tool_errors: 0,
            model_switch: None,
            dmn_pause: false,
        };

        loop {
            let _ = ui_tx.send(UiMessage::Activity("thinking...".into()));
            let api_result = self
                .client
                .chat_completion_stream(
                    &self.messages,
                    Some(&self.tool_defs),
                    ui_tx,
                    target,
                    &self.reasoning_effort,
                )
                .await;

            // Context overflow → compact and retry (max 2 attempts)
            // Stream error → retry with backoff (max 2 attempts)
            let (msg, usage) = match api_result {
                Err(e) if crate::agent::context::is_context_overflow(&e) && overflow_retries < 2 => {
                    overflow_retries += 1;
                    let _ = ui_tx.send(UiMessage::Info(format!(
                        "[context overflow — compacting and retrying ({}/2)]",
                        overflow_retries,
                    )));
                    self.emergency_compact();
                    continue;
                }
                Err(e) if crate::agent::context::is_stream_error(&e) && empty_retries < 2 => {
                    empty_retries += 1;
                    let _ = ui_tx.send(UiMessage::Info(format!(
                        "[stream error: {} — retrying ({}/2)]",
                        e, empty_retries,
                    )));
                    tokio::time::sleep(std::time::Duration::from_secs(2)).await;
                    continue;
                }
                other => other?,
            };

            // Strip ephemeral tool calls (journal) that the API has
            // now processed. They're persisted to disk; no need to keep
            // them in the conversation history burning tokens.
            self.strip_ephemeral_tool_calls();

            if let Some(usage) = &usage {
                self.last_prompt_tokens = usage.prompt_tokens;
                self.measure_budget();
                self.publish_context_state();
                let _ = ui_tx.send(UiMessage::StatusUpdate(StatusInfo {
                    dmn_state: String::new(), // filled by main loop
                    dmn_turns: 0,
                    dmn_max_turns: 0,
                    prompt_tokens: usage.prompt_tokens,
                    completion_tokens: usage.completion_tokens,
                    model: self.client.model.clone(),
                    turn_tools: 0, // tracked by TUI from ToolCall messages
                    context_budget: self.context_budget.status_string(),
                }));
            }

            // Empty response — model returned finish=stop with no content
            // or tool calls. Inject a nudge so the retry has different input.
            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;
                    let _ = ui_tx.send(UiMessage::Debug(format!(
                        "empty response, injecting nudge and retrying ({}/2)",
                        empty_retries,
                    )));
                    self.push_message(Message::user(
                        "[system] Your previous response was empty. \
                         Please respond with text or use a tool."
                    ));
                    continue;
                }
                // After max retries, fall through — return the empty response
            } else {
                empty_retries = 0;
            }

            // Structured tool calls from the API
            if let Some(ref tool_calls) = msg.tool_calls {
                if !tool_calls.is_empty() {
                    self.push_message(msg.clone());
                    for call in tool_calls {
                        self.dispatch_tool_call(call, None, ui_tx, &mut ds)
                            .await;
                    }
                    continue;
                }
            }

            // No structured tool calls — check for leaked tool calls
            // (Qwen sometimes outputs <tool_call> XML as text).
            let text = msg.content_text().to_string();
            let leaked = crate::agent::parsing::parse_leaked_tool_calls(&text);

            if !leaked.is_empty() {
                let _ = ui_tx.send(UiMessage::Debug(format!(
                    "recovered {} leaked tool call(s) from text",
                    leaked.len()
                )));
                // Strip tool call XML and thinking tokens from the message
                // so they don't clutter the conversation history.
                let cleaned = crate::agent::parsing::strip_leaked_artifacts(&text);
                let mut clean_msg = msg.clone();
                clean_msg.content = if cleaned.trim().is_empty() {
                    None
                } else {
                    Some(MessageContent::Text(cleaned))
                };
                self.push_message(clean_msg);
                for call in &leaked {
                    self.dispatch_tool_call(call, Some("recovered"), ui_tx, &mut ds)
                        .await;
                }
                continue;
            }

            // Genuinely text-only response
            let _ = ui_tx.send(UiMessage::Activity(String::new()));
            self.push_message(msg);

            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 single tool call: send UI annotations, run the tool,
    /// push results into the conversation, handle images.
    async fn dispatch_tool_call(
        &mut self,
        call: &ToolCall,
        tag: Option<&str>,
        ui_tx: &UiSender,
        ds: &mut DispatchState,
    ) {
        let args: serde_json::Value =
            serde_json::from_str(&call.function.arguments).unwrap_or_default();

        let args_summary = summarize_args(&call.function.name, &args);
        let label = match tag {
            Some(t) => format!("calling: {} ({})", call.function.name, t),
            None => format!("calling: {}", call.function.name),
        };
        let _ = ui_tx.send(UiMessage::Activity(label));
        let _ = ui_tx.send(UiMessage::ToolCall {
            name: call.function.name.clone(),
            args_summary: args_summary.clone(),
        });
        let _ = ui_tx.send(UiMessage::ToolStarted {
            id: call.id.clone(),
            name: call.function.name.clone(),
            detail: args_summary,
        });

        // Handle working_stack tool — needs &mut self for context state
        if call.function.name == "working_stack" {
            let result = tools::working_stack::handle(&args, &mut self.context.working_stack);
            let output = tools::ToolOutput {
                text: result.clone(),
                is_yield: false,
                images: Vec::new(),
                model_switch: None,
                dmn_pause: false,
            };
            let _ = ui_tx.send(UiMessage::ToolResult {
                name: call.function.name.clone(),
                result: output.text.clone(),
            });
            let _ = ui_tx.send(UiMessage::ToolFinished { id: call.id.clone() });
            self.push_message(Message::tool_result(&call.id, &output.text));
            ds.had_tool_calls = true;
            
            // Re-render the context message so the model sees the updated stack
            if !result.starts_with("Error:") {
                self.refresh_context_message();
            }
            return;
        }

        // Handle memory tools — needs &mut self for node tracking
        if call.function.name.starts_with("memory_") {
            let result = tools::memory::dispatch(&call.function.name, &args, None);
            let text = match &result {
                Ok(s) => s.clone(),
                Err(e) => format!("Error: {:#}", e),
            };

            // Track loaded/updated nodes
            if result.is_ok() {
                match call.function.name.as_str() {
                    "memory_render" | "memory_links" => {
                        if let Some(key) = args.get("key").and_then(|v| v.as_str()) {
                            if let Some(node) = crate::agent::memory::MemoryNode::load(key) {
                                // Replace if already tracked, otherwise add
                                if let Some(existing) = self.context.loaded_nodes.iter_mut()
                                    .find(|n| n.key == node.key) {
                                    *existing = node;
                                } else {
                                    self.context.loaded_nodes.push(node);
                                }
                            }
                        }
                    }
                    "memory_write" => {
                        if let Some(key) = args.get("key").and_then(|v| v.as_str()) {
                            if let Some(node) = crate::agent::memory::MemoryNode::load(key) {
                                // Refresh if already tracked
                                if let Some(existing) = self.context.loaded_nodes.iter_mut()
                                    .find(|n| n.key == node.key) {
                                    *existing = node;
                                }
                                // Don't auto-add writes — only renders register nodes
                            }
                        }
                    }
                    _ => {}
                }
            }

            let output = tools::ToolOutput {
                text,
                is_yield: false,
                images: Vec::new(),
                model_switch: None,
                dmn_pause: false,
            };
            let _ = ui_tx.send(UiMessage::ToolResult {
                name: call.function.name.clone(),
                result: output.text.clone(),
            });
            let _ = ui_tx.send(UiMessage::ToolFinished { id: call.id.clone() });
            self.push_message(Message::tool_result(&call.id, &output.text));
            ds.had_tool_calls = true;
            if output.text.starts_with("Error:") {
                ds.tool_errors += 1;
            }
            return;
        }

        let output =
            tools::dispatch(&call.function.name, &args, &self.process_tracker).await;

        if output.is_yield {
            ds.yield_requested = true;
        } else {
            ds.had_tool_calls = true;
        }
        if output.model_switch.is_some() {
            ds.model_switch = output.model_switch;
        }
        if output.dmn_pause {
            ds.dmn_pause = true;
        }
        if output.text.starts_with("Error:") {
            ds.tool_errors += 1;
        }

        let _ = ui_tx.send(UiMessage::ToolResult {
            name: call.function.name.clone(),
            result: output.text.clone(),
        });
        let _ = ui_tx.send(UiMessage::ToolFinished { id: call.id.clone() });

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

        if !output.images.is_empty() {
            // Only one live image in context at a time — age out any
            // previous ones to avoid accumulating ~90KB+ per image.
            self.age_out_images();
            self.push_message(Message::user_with_images(
                "Here is the image you requested:",
                &output.images,
            ));
        }
    }

    /// Build context state summary for the debug screen.
    pub fn context_state_summary(&self) -> 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 — split into per-entry children
        {
            let mut journal_children = Vec::new();
            let mut current_header = String::new();
            let mut current_body = String::new();
            for line in self.context.journal.lines() {
                if line.starts_with("## ") {
                    if !current_header.is_empty() {
                        let body = std::mem::take(&mut current_body);
                        let preview: String = body.lines().next().unwrap_or("").chars().take(60).collect();
                        journal_children.push(ContextSection {
                            name: format!("{}: {}", current_header, preview),
                            tokens: count(&body),
                            content: body,
                            children: Vec::new(),
                        });
                    }
                    current_header = line.trim_start_matches("## ").to_string();
                    current_body.clear();
                } else {
                    if !current_body.is_empty() || !line.is_empty() {
                        current_body.push_str(line);
                        current_body.push('\n');
                    }
                }
            }
            if !current_header.is_empty() {
                let preview: String = current_body.lines().next().unwrap_or("").chars().take(60).collect();
                journal_children.push(ContextSection {
                    name: format!("{}: {}", current_header, preview),
                    tokens: count(&current_body),
                    content: current_body,
                    children: Vec::new(),
                });
            }
            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)
            .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,
        });

        // Loaded memory nodes — tracked by memory tools
        if !self.context.loaded_nodes.is_empty() {
            let node_children: Vec<ContextSection> = self.context.loaded_nodes.iter()
                .map(|node| {
                    let rendered = node.render();
                    ContextSection {
                        name: format!("{} (v{}, w={:.2}, {} links)",
                            node.key, node.version, node.weight, node.links.len()),
                        tokens: count(&rendered),
                        content: String::new(), // don't duplicate in debug view
                        children: Vec::new(),
                    }
                })
                .collect();
            let node_tokens: usize = node_children.iter().map(|c| c.tokens).sum();
            sections.push(ContextSection {
                name: format!("Memory nodes ({} loaded)", self.context.loaded_nodes.len()),
                tokens: node_tokens,
                content: String::new(),
                children: node_children,
            });
        }

        // Conversation — each message as a child
        let conv_start = self.messages.iter()
            .position(|m| m.role == Role::Assistant || m.role == Role::Tool)
            .unwrap_or(self.messages.len());
        let conv_messages = &self.messages[conv_start..];
        let conv_children: Vec<ContextSection> = conv_messages.iter().enumerate()
            .map(|(i, msg)| {
                let text = msg.content.as_ref()
                    .map(|c| c.as_text().to_string())
                    .unwrap_or_default();
                let tool_info = msg.tool_calls.as_ref().map(|tc| {
                    tc.iter()
                        .map(|c| c.function.name.clone())
                        .collect::<Vec<_>>()
                        .join(", ")
                });
                let label = match (&msg.role, &tool_info) {
                    (_, Some(tools)) => format!("[tool_call: {}]", tools),
                    _ => {
                        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 role_name = match msg.role {
                    Role::Assistant => "PoC",
                    Role::User => "Kent",
                    Role::Tool => "tool",
                    Role::System => "system",
                };
                ContextSection {
                    name: format!("[{}] {}: {}", conv_start + i, role_name, label),
                    tokens,
                    content: text,
                    children: Vec::new(),
                }
            })
            .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 journal_path = journal::default_journal_path();
        let entries = journal::parse_journal_tail(&journal_path, 64 * 1024);
        if entries.is_empty() {
            return;
        }

        let count = |s: &str| self.tokenizer.encode_with_special_tokens(s).len();
        let context_message = self.context.render_context_message();

        let plan = crate::agent::context::plan_context(
            &self.context.system_prompt,
            &context_message,
            &[],  // no conversation yet
            &entries,
            &self.client.model,
            &count,
        );

        self.context.journal = crate::agent::context::render_journal_text(&entries, &plan);
    }

    /// Re-render the context message in self.messages from live ContextState.
    /// Called after any change to context state (working stack, etc).
    fn refresh_context_message(&mut self) {
        let rendered = self.context.render_context_message();
        // The context message is the first user message (index 1, after system prompt)
        if self.messages.len() >= 2 && self.messages[1].role == Role::User {
            self.messages[1] = Message::user(rendered);
        }
        self.publish_context_state();
        self.save_working_stack();
    }

    /// Persist working stack to disk.
    fn save_working_stack(&self) {
        if let Ok(json) = serde_json::to_string(&self.context.working_stack) {
            let _ = std::fs::write(WORKING_STACK_FILE, json);
        }
    }

    /// Load working stack from disk.
    fn load_working_stack(&mut self) {
        if let Ok(data) = std::fs::read_to_string(WORKING_STACK_FILE) {
            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.
    fn publish_context_state(&self) {
        if let Ok(mut state) = self.shared_context.write() {
            *state = self.context_state_summary();
        }
    }

    /// Replace base64 image data in older messages with text placeholders.
    /// Only the most recent image stays live — each new image ages out
    /// all previous ones. The tool result message (right before each image
    /// message) already records what was loaded, so no info is lost.
    fn age_out_images(&mut self) {
        for msg in &mut self.messages {
            if let Some(MessageContent::Parts(parts)) = &msg.content {
                let has_images = parts.iter().any(|p| matches!(p, ContentPart::ImageUrl { .. }));
                if !has_images {
                    continue;
                }
                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 — see tool result above for details]",
                            );
                        }
                    }
                }
                msg.content = Some(MessageContent::Text(replacement));
            }
        }
    }

    /// Strip ephemeral tool calls from the conversation history.
    ///
    /// Ephemeral tools (like journal) persist their output to disk,
    /// so the tool call + result don't need to stay in the context
    /// window. We keep them for exactly one API round-trip (the model
    /// needs to see the result was acknowledged), then strip them.
    ///
    /// If an assistant message contains ONLY ephemeral tool calls,
    /// the entire message and its tool results are removed. If mixed
    /// with non-ephemeral calls, we leave it (rare case, small cost).
    fn strip_ephemeral_tool_calls(&mut self) {
        // Collect IDs of tool calls to strip
        let mut strip_ids: Vec<String> = Vec::new();
        let mut strip_msg_indices: Vec<usize> = Vec::new();

        for (i, msg) in self.messages.iter().enumerate() {
            if msg.role != Role::Assistant {
                continue;
            }
            let calls = match &msg.tool_calls {
                Some(c) if !c.is_empty() => c,
                _ => continue,
            };

            let all_ephemeral = calls.iter().all(|c| {
                c.function.name == tools::journal::TOOL_NAME
            });

            if all_ephemeral {
                strip_msg_indices.push(i);
                for call in calls {
                    strip_ids.push(call.id.clone());
                }
            }
        }

        if strip_ids.is_empty() {
            return;
        }

        // Remove in reverse order to preserve indices
        self.messages.retain(|msg| {
            // Strip the assistant messages we identified
            if msg.role == Role::Assistant {
                if let Some(calls) = &msg.tool_calls {
                    if calls.iter().all(|c| strip_ids.contains(&c.id)) {
                        return false;
                    }
                }
            }
            // Strip matching tool results
            if msg.role == Role::Tool {
                if let Some(ref id) = msg.tool_call_id {
                    if strip_ids.contains(id) {
                        return false;
                    }
                }
            }
            true
        });
    }

    /// Last prompt token count reported by the API.
    pub fn last_prompt_tokens(&self) -> u32 {
        self.last_prompt_tokens
    }

    /// Build context window from conversation messages + journal.
    /// Used by both compact() (in-memory messages) and restore_from_log()
    /// (conversation log). The context window is always:
    ///   identity + journal summaries + raw recent messages
    pub fn compact(&mut self, new_system_prompt: String, new_personality: Vec<(String, String)>) {
        self.context.system_prompt = new_system_prompt;
        self.context.personality = new_personality;
        self.do_compact();
    }

    /// Internal compaction — rebuilds context window from current messages.
    fn do_compact(&mut self) {
        // Find where actual conversation starts (after system + context)
        let conv_start = self
            .messages
            .iter()
            .position(|m| m.role == Role::Assistant || m.role == Role::Tool)
            .unwrap_or(self.messages.len());

        let conversation: Vec<Message> = self.messages[conv_start..].to_vec();
        let (messages, journal) = crate::agent::context::build_context_window(
            &self.context,
            &conversation,
            &self.client.model,
            &self.tokenizer,
        );
        self.context.journal = journal;
        self.messages = messages;
        self.last_prompt_tokens = 0;
        self.measure_budget();
        self.publish_context_state();
    }

    /// Emergency compaction using stored config — called on context overflow.
    fn emergency_compact(&mut self) {
        self.do_compact();
    }

    /// 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,
        system_prompt: String,
        personality: Vec<(String, String)>,
    ) -> bool {
        self.context.system_prompt = system_prompt;
        self.context.personality = personality;

        let all_messages = match &self.conversation_log {
            Some(log) => match log.read_tail(512 * 1024) {
                Ok(msgs) if !msgs.is_empty() => {
                    dbglog!("[restore] read {} messages from log tail", msgs.len());
                    msgs
                }
                Ok(_) => {
                    dbglog!("[restore] log exists but is empty");
                    return false;
                }
                Err(e) => {
                    dbglog!("[restore] failed to read log: {}", e);
                    return false;
                }
            },
            None => {
                dbglog!("[restore] no conversation log configured");
                return false;
            }
        };

        // Filter out system/context messages — we only want the
        // actual conversation (user prompts, assistant responses,
        // tool calls/results)
        let conversation: Vec<Message> = all_messages
            .into_iter()
            .filter(|m| m.role != Role::System)
            .collect();
        dbglog!("[restore] {} messages after filtering system", conversation.len());

        let (messages, journal) = crate::agent::context::build_context_window(
            &self.context,
            &conversation,
            &self.client.model,
            &self.tokenizer,
        );
        dbglog!("[restore] journal text: {} chars, {} lines",
            journal.len(), journal.lines().count());
        self.context.journal = journal;
        self.messages = messages;
        dbglog!("[restore] built context window: {} messages", self.messages.len());
        self.last_prompt_tokens = 0;
        self.measure_budget();
        self.publish_context_state();
        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 history for persistence.
    pub fn messages(&self) -> &[Message] {
        &self.messages
    }

    /// Mutable access to conversation history (for /retry).
    pub fn messages_mut(&mut self) -> &mut Vec<Message> {
        &mut self.messages
    }

    /// Restore from a saved conversation.
    pub fn restore(&mut self, messages: Vec<Message>) {
        self.messages = messages;
    }
}

// Context window building, token counting, and error classification
// live in context.rs


/// Create a short summary of tool args for the tools pane header.
fn summarize_args(tool_name: &str, args: &serde_json::Value) -> String {
    match tool_name {
        "read_file" | "write_file" | "edit_file" => args["file_path"]
            .as_str()
            .unwrap_or("")
            .to_string(),
        "bash" => {
            let cmd = args["command"].as_str().unwrap_or("");
            if cmd.len() > 60 {
                let end = cmd.char_indices()
                    .map(|(i, _)| i)
                    .take_while(|&i| i <= 60)
                    .last()
                    .unwrap_or(0);
                format!("{}...", &cmd[..end])
            } else {
                cmd.to_string()
            }
        }
        "grep" => {
            let pattern = args["pattern"].as_str().unwrap_or("");
            let path = args["path"].as_str().unwrap_or(".");
            format!("{} in {}", pattern, path)
        }
        "glob" => args["pattern"]
            .as_str()
            .unwrap_or("")
            .to_string(),
        "view_image" => {
            if let Some(pane) = args["pane_id"].as_str() {
                format!("pane {}", pane)
            } else {
                args["file_path"].as_str().unwrap_or("").to_string()
            }
        }
        "journal" => {
            let entry = args["entry"].as_str().unwrap_or("");
            if entry.len() > 60 {
                format!("{}...", &entry[..60])
            } else {
                entry.to_string()
            }
        }
        "yield_to_user" => args["message"]
            .as_str()
            .unwrap_or("")
            .to_string(),
        "switch_model" => args["model"]
            .as_str()
            .unwrap_or("")
            .to_string(),
        "pause" => String::new(),
        _ => String::new(),
    }
}

// Parsing functions (parse_leaked_tool_calls, strip_leaked_artifacts)
// and their tests live in parsing.rs