codecraft/server/src/runner.rs

//! The agent loop: stream a model response, run tool calls (with permission
//! checks), feed results back, repeat until the model stops.

use crate::app::{App, PermissionReply};
use crate::permission::{Decision, Evaluator, bash_prefix, split_command};
use crate::prompts::{self, Agent};
use crate::provider::{ChatRequest, Provider, StreamEvent};
use crate::store::UsageRecord;
use crate::tools::{self, Tool, ToolCtx};
use crate::types::{Event, Message, Part, Role, Session, Usage, new_id, now_ms};
use anyhow::{Result, anyhow, bail};
use futures::future::join_all;
use serde_json::Value;
use std::path::PathBuf;
use std::sync::Arc;
use std::time::Instant;
use tokio::sync::mpsc;
use tokio_util::sync::CancellationToken;

const MAX_STEPS: usize = 250;
const MAX_SUBAGENT_DEPTH: usize = 3;

/// What happened to a prompt.
pub enum Submitted {
    Started,
    Queued(crate::store::QueueItem),
}

fn emit_queue(app: &App, session_id: &str) {
    let items = app.store.queue_list(session_id).unwrap_or_default();
    app.emit(Event::QueueUpdated { session_id: session_id.into(), items: serde_json::json!(items) });
}

/// Take all queued messages (one user message's worth of parts).
fn take_queue(app: &App, session_id: &str) -> Vec<Part> {
    let items = app.store.queue_take(session_id).unwrap_or_default();
    if !items.is_empty() {
        emit_queue(app, session_id);
    }
    items.into_iter().flat_map(|i| i.parts).collect()
}

/// Start processing a user prompt. If the session is busy, the prompt (with
/// attachments) is queued and delivered at the agent's next step.
pub fn start_prompt(app: Arc<App>, session_id: String, parts: Vec<Part>) -> Result<()> {
    submit(app, session_id, parts, "user").map(|_| ())
}

pub fn submit(app: Arc<App>, session_id: String, parts: Vec<Part>, source: &str) -> Result<Submitted> {
    // Queueing and the end-of-run check run under the same runtime lock, so a
    // message is never lost or delivered twice.
    let cancel = CancellationToken::new();
    let queued = app.with_runtime(&session_id, |r| -> Result<Option<crate::store::QueueItem>> {
        if r.cancel.is_some() {
            let item = crate::store::QueueItem { id: new_id(), parts: parts.clone(), source: source.into(), created_at: now_ms() };
            app.store.queue_add(&session_id, &item)?;
            Ok(Some(item))
        } else {
            r.cancel = Some(cancel.clone());
            Ok(None)
        }
    })?;
    if let Some(item) = queued {
        emit_queue(&app, &session_id);
        return Ok(Submitted::Queued(item));
    }
    let text_only: String = parts
        .iter()
        .filter_map(|p| if let Part::Text { text } = p { Some(text.as_str()) } else { None })
        .collect::<Vec<_>>()
        .join("\n");
    app.emit(Event::SessionStatus { session_id: session_id.clone(), busy: true });

    tokio::spawn(async move {
        let res = async {
            let msg = Message::new(Role::User, parts);
            app.store.add_message(&session_id, &msg)?;
            app.emit(Event::MessageAdded { session_id: session_id.clone(), message: msg });
            maybe_generate_title(app.clone(), &session_id, &text_only);
            run_session(app.clone(), &session_id, cancel.clone(), 0).await
        }
        .await;
        if let Err(e) = res {
            tracing::warn!("session {session_id}: {e:#}");
            app.emit(Event::Error { session_id: session_id.clone(), message: format!("{e:#}") });
        }
        // Messages queued after the last step start a new turn – unless the user
        // stopped the agent: then they stay queued until sent explicitly.
        let leftover = app.with_runtime(&session_id, |r| {
            r.cancel = None;
            if cancel.is_cancelled() { Vec::new() } else { take_queue(&app, &session_id) }
        });
        app.emit(Event::SessionStatus { session_id: session_id.clone(), busy: false });
        if !leftover.is_empty() {
            let _ = start_prompt(app, session_id, leftover);
        }
    });
    Ok(Submitted::Started)
}

/// Send queued messages now (after a stop or a server restart).
pub fn flush_queue(app: Arc<App>, session_id: &str) -> Result<bool> {
    if app.is_busy(session_id) {
        return Ok(false);
    }
    let parts = take_queue(&app, session_id);
    if parts.is_empty() {
        return Ok(false);
    }
    start_prompt(app, session_id.to_string(), parts)?;
    Ok(true)
}

pub fn abort(app: &App, session_id: &str) {
    app.with_runtime(session_id, |r| {
        if let Some(c) = &r.cancel {
            c.cancel();
        }
    });
}

/// Messages since the last compaction summary. Tool calls left without a
/// result (run interrupted, e.g. by a server restart) get a synthetic result:
/// providers reject histories with unanswered tool calls.
fn active_history(all: Vec<Message>) -> Vec<Message> {
    let start = all
        .iter()
        .rposition(|m| m.parts.iter().any(|p| matches!(p, Part::Summary { .. })))
        .unwrap_or(0);
    let msgs: Vec<Message> = all.into_iter().skip(start).collect();
    let mut out: Vec<Message> = Vec::with_capacity(msgs.len());
    for (i, m) in msgs.iter().enumerate() {
        out.push(m.clone());
        if m.role != Role::Assistant {
            continue;
        }
        let answered: std::collections::HashSet<&str> = msgs
            .get(i + 1)
            .filter(|n| n.role == Role::User)
            .map(|n| {
                n.parts
                    .iter()
                    .filter_map(|p| if let Part::ToolResult { call_id, .. } = p { Some(call_id.as_str()) } else { None })
                    .collect()
            })
            .unwrap_or_default();
        let missing: Vec<Part> = m
            .parts
            .iter()
            .filter_map(|p| match p {
                Part::ToolCall { id, name, .. } if !answered.contains(id.as_str()) => {
                    Some(tool_result(id, name, "Not executed: the run was interrupted (e.g. server restart).".into(), true, None))
                }
                _ => None,
            })
            .collect();
        if missing.is_empty() {
            continue;
        }
        // Results must directly follow the call: merge into the next user message or insert one.
        match msgs.get(i + 1) {
            Some(n) if n.role == Role::User => {
                // The next message is pushed in the next iteration; prepend there.
                let mut next = n.clone();
                let mut parts = missing;
                parts.extend(next.parts);
                next.parts = parts;
                // Replace by pushing now and skipping the original.
                out.push(next);
            }
            _ => out.push(Message::new(Role::User, missing)),
        }
    }
    // Remove duplicates created by the merge above (same id pushed twice: keep the merged one).
    let mut dedup: Vec<Message> = Vec::with_capacity(out.len());
    for m in out {
        if let Some(last) = dedup.last() {
            if last.id == m.id {
                continue;
            }
        }
        dedup.push(m);
    }
    dedup
}

fn normalize_tool_name(s: &str) -> String {
    s.chars().filter(|c| c.is_ascii_alphanumeric()).collect::<String>().to_lowercase()
}

pub fn available_tools(app: &App, agent: &Agent) -> Vec<Arc<dyn Tool>> {
    let cfg = app.config();
    let mut all = tools::builtin_tools();
    if !cfg.images.enabled {
        all.retain(|t| t.name() != "generate_image");
    }
    let (direct, deferred) = app.mcp.tools(&cfg);
    all.extend(direct);
    if !deferred.is_empty() {
        all.push(Arc::new(crate::mcp::McpSearch));
        all.push(Arc::new(crate::mcp::McpCall));
    }
    all.into_iter()
        .filter(|t| !agent.read_only || t.read_only())
        // Craft tools only where they are useful.
        .filter(|t| agent.name == "craft" || !matches!(t.name(), "browser_snapshot" | "image_compare") || agent.tools.is_some())
        .filter(|t| match &agent.tools {
            None => true,
            Some(list) => {
                let n = normalize_tool_name(t.name());
                list.iter().any(|a| {
                    let a = normalize_tool_name(a);
                    a == n || (t.name().starts_with("mcp") && a == "mcp")
                })
            }
        })
        .collect()
}

struct Step {
    message: Message,
    stop_reason: String,
}

/// Boxed with an explicit `Send` bound: subagents recurse into the agent loop,
/// and the declared type breaks the cycle for the compiler.
fn run_session_boxed(
    app: Arc<App>,
    session_id: String,
    cancel: CancellationToken,
    depth: usize,
) -> futures::future::BoxFuture<'static, Result<String>> {
    Box::pin(async move { run_session(app, &session_id, cancel, depth).await })
}

/// Run the loop for a session until the model stops. Returns the final text.
pub async fn run_session(app: Arc<App>, session_id: &str, cancel: CancellationToken, depth: usize) -> Result<String> {
    let mut final_text = String::new();
    for _ in 0..MAX_STEPS {
        if cancel.is_cancelled() {
            break;
        }
        let session = app.session(session_id)?;
        let cwd = PathBuf::from(&session.cwd);
        let roots = app.user_roots(session.owner_id.as_deref());
        let root = app.root_for(session.owner_id.as_deref(), &cwd);
        let agent = prompts::find_agent(&app, &root, &cwd, &session.agent);
        let model_spec = agent.model.clone().unwrap_or_else(|| session.model.clone());
        let (provider, model) = app.registry().resolve(&model_spec)?;

        // Auto-compaction when the last request used most of the context window.
        if let (Some(u), Some(ctx_len)) = (&session.last_usage, model.context) {
            let threshold = app.config().server.auto_compact_threshold;
            if threshold > 0.0 && u.context_tokens() as f64 > ctx_len as f64 * threshold {
                compact(app.clone(), session_id, cancel.clone()).await?;
                continue;
            }
        }

        let tools = available_tools(&app, &agent);
        let history = active_history(app.store.messages(session_id)?);
        let req = ChatRequest {
            system: prompts::system_prompt(&app, &agent, &root, &cwd, &model_spec, &session.permission_mode),
            messages: history,
            tools: tools.iter().map(|t| t.def()).collect(),
            max_tokens: model.max_output.unwrap_or(16_000),
            reasoning: true,
            model: model.clone(),
        };
        let step = stream_step(&app, &session, &agent.name, &model_spec, provider, req, &cancel).await?;
        final_text = step
            .message
            .parts
            .iter()
            .filter_map(|p| if let Part::Text { text } = p { Some(text.as_str()) } else { None })
            .collect::<Vec<_>>()
            .join("");

        let calls: Vec<(String, String, Value)> = step
            .message
            .parts
            .iter()
            .filter_map(|p| match p {
                Part::ToolCall { id, name, input, .. } => Some((id.clone(), name.clone(), input.clone())),
                _ => None,
            })
            .collect();

        if cancel.is_cancelled() {
            // Every tool call needs a result, otherwise the history is invalid.
            if !calls.is_empty() {
                let results = calls
                    .iter()
                    .map(|(id, name, _)| tool_result(id, name, "Aborted by user".into(), true, None))
                    .collect();
                save_user_parts(&app, session_id, results)?;
            }
            break;
        }
        if calls.is_empty() {
            match step.stop_reason.as_str() {
                "max_tokens" | "length" => app.emit(Event::Error {
                    session_id: session_id.into(),
                    message: "The response hit the output token limit (max_output).".into(),
                }),
                "refusal" | "content_filter" => app.emit(Event::Error {
                    session_id: session_id.into(),
                    message: "The model declined to respond (refusal).".into(),
                }),
                _ => {}
            }
            break;
        }

        let base = ToolCtx {
            app: app.clone(),
            session_id: session_id.to_string(),
            call_id: String::new(),
            cwd: cwd.clone(),
            cancel: cancel.clone(),
            agent: agent.name.clone(),
            roots: roots.clone(),
        };
        let by_name = |n: &str| tools.iter().find(|t| t.name() == n).cloned();
        let all_read_only = calls.iter().all(|(_, n, _)| by_name(n).is_some_and(|t| t.read_only()));
        let mut results = Vec::new();
        if all_read_only {
            let futs = calls.iter().map(|(id, name, input)| {
                execute_call(&app, &session, &base, by_name(name), id, name, input.clone(), depth)
            });
            results = join_all(futs).await;
        } else {
            for (id, name, input) in &calls {
                results.push(execute_call(&app, &session, &base, by_name(name), id, name, input.clone(), depth).await);
            }
        }
        // Deliver messages the user queued while we were working (with attachments).
        results.extend(take_queue(&app, session_id));
        save_user_parts(&app, session_id, results)?;
    }
    Ok(final_text)
}

fn tool_result(call_id: &str, name: &str, content: String, is_error: bool, meta: Option<Value>) -> Part {
    Part::ToolResult { call_id: call_id.into(), name: name.into(), content, images: vec![], is_error, meta }
}

fn save_user_parts(app: &App, session_id: &str, parts: Vec<Part>) -> Result<()> {
    let msg = Message::new(Role::User, parts);
    app.store.add_message(session_id, &msg)?;
    app.emit(Event::MessageAdded { session_id: session_id.into(), message: msg });
    Ok(())
}

#[allow(clippy::too_many_arguments)]
async fn execute_call(
    app: &Arc<App>,
    session: &Session,
    base: &ToolCtx,
    tool: Option<Arc<dyn Tool>>,
    call_id: &str,
    name: &str,
    input: Value,
    depth: usize,
) -> Part {
    let Some(tool) = tool else {
        return tool_result(call_id, name, format!("Unknown or unavailable tool '{name}' in this mode."), true, None);
    };
    if let Some(raw) = input.get("__invalid_json") {
        return tool_result(
            call_id,
            name,
            format!("Tool input was not valid JSON (possibly truncated). Retry with a smaller input.\n{}", raw.as_str().unwrap_or("").chars().take(300).collect::<String>()),
            true,
            None,
        );
    }
    if name == "task" && depth >= MAX_SUBAGENT_DEPTH {
        return tool_result(call_id, name, "Maximum subagent depth reached.".into(), true, None);
    }

    let ctx = ToolCtx {
        app: app.clone(),
        session_id: base.session_id.clone(),
        call_id: call_id.into(),
        cwd: base.cwd.clone(),
        cancel: base.cancel.clone(),
        agent: base.agent.clone(),
        roots: base.roots.clone(),
    };
    // mcp_call is judged (and approved) as the MCP tool it calls.
    let call_name = name;
    let (tool, input, name): (Arc<dyn Tool>, Value, String) = if name == "mcp_call" {
        match crate::mcp::resolve_call_target(&ctx, &input) {
            Ok(target) => {
                let args = input.get("arguments").cloned().unwrap_or(serde_json::json!({}));
                let n = target.full_name.clone();
                (target, args, n)
            }
            Err(e) => return tool_result(call_id, call_name, format!("Error: {e:#}"), true, None),
        }
    } else {
        (tool, input, name.to_string())
    };
    let name = name.as_str();

    // Permission check. Approvals are tracked on the top-level session.
    let root = app.root_session_id(&session.id);
    let cfg = app.config();
    let decision = app.with_runtime(&root, |r| {
        Evaluator { cfg: &cfg.permissions, mode: &session.permission_mode, approved: &r.approved }.evaluate(tool.as_ref(), &input)
    });
    match decision {
        Decision::Deny => {
            return tool_result(call_id, call_name, "Denied by the permission rules in the configuration.".into(), true, None);
        }
        Decision::Ask => {
            let diff = tool.preview(&input, &ctx).await;
            let (reply, message) = app
                .request_permission(&session.id, name, tool.title(&input), input.clone(), diff, base.cancel.clone())
                .await;
            match reply {
                PermissionReply::Reject => {
                    let mut text = "The user rejected this tool call.".to_string();
                    if let Some(m) = message.filter(|m| !m.trim().is_empty()) {
                        text.push_str(&format!(" User feedback: {m}"));
                    }
                    return tool_result(call_id, call_name, text, true, None);
                }
                PermissionReply::Always => app.with_runtime(&root, |r| {
                    if name == "bash" {
                        let cmd = input["command"].as_str().unwrap_or("");
                        for part in split_command(cmd).unwrap_or_default() {
                            r.approved.insert(format!("bash:{}", bash_prefix(&part)));
                        }
                    } else {
                        r.approved.insert(name.to_string());
                    }
                }),
                PermissionReply::Once => {}
            }
        }
        Decision::Allow => {}
    }

    let res = if name == "task" {
        // Pass depth through the agent name channel: subagents run with depth+1.
        run_task(app, &ctx, input, depth).await
    } else {
        tool.run(input, &ctx).await
    };
    match res {
        Ok(out) => Part::ToolResult {
            call_id: call_id.into(),
            name: call_name.into(),
            content: out.content,
            images: out.images,
            is_error: out.is_error,
            meta: out.meta,
        },
        Err(e) => tool_result(call_id, call_name, format!("Error: {e:#}"), true, None),
    }
}

async fn run_task(app: &Arc<App>, ctx: &ToolCtx, input: Value, depth: usize) -> Result<tools::ToolOutput> {
    let prompt = tools::arg_str(&input, "prompt")?.to_string();
    let description = input["description"].as_str().unwrap_or("subagent").to_string();
    let agent_name = input["subagent_type"].as_str().unwrap_or("general").to_string();
    let background = input["background"].as_bool().unwrap_or(false);
    let (child, agent) = create_child(app, ctx, &agent_name, &description, prompt)?;
    let meta = serde_json::json!({
        "session_id": child.id, "agent": agent.name, "model": child.model, "background": background
    });

    if background {
        let app = app.clone();
        let parent_id = ctx.session_id.clone();
        let child_id = child.id.clone();
        let agent_name = agent.name.clone();
        let desc = description.clone();
        // Background tasks outlive the current turn; they are stopped via their own session.
        let cancel = CancellationToken::new();
        tokio::spawn(async move {
            let res = run_child(app.clone(), &child, &agent, cancel, depth + 1).await;
            let body = match res {
                Ok(text) => text,
                Err(e) => format!("The background task failed: {e:#}"),
            };
            let text = format!(
                "<background-task-result agent=\"{agent_name}\" session=\"{child_id}\" description=\"{desc}\">\n{body}\n</background-task-result>"
            );
            // Injected into the running turn, or starts a new turn when the parent is idle.
            if app.store.get_session(&parent_id).ok().flatten().is_some() {
                if let Err(e) = submit(app.clone(), parent_id.clone(), vec![Part::Text { text }], "background") {
                    tracing::warn!("could not deliver background result: {e:#}");
                }
            }
        });
        return Ok(tools::ToolOutput::text(format!(
            "Started background task '{description}' with subagent '{}' (model {}). Its result will arrive \
             automatically as a message when it finishes – continue with other work and do not poll.",
            meta["agent"].as_str().unwrap_or(""),
            meta["model"].as_str().unwrap_or("")
        ))
        .with_meta(meta));
    }

    let text = run_child(app.clone(), &child, &agent, ctx.cancel.child_token(), depth + 1).await?;
    Ok(tools::ToolOutput::text(text).with_meta(meta))
}

/// Entry point used by the task tool when called outside the runner.
pub async fn run_subagent(app: Arc<App>, ctx: &ToolCtx, agent: &str, description: &str, prompt: String) -> Result<(String, String)> {
    let (child, agent) = create_child(&app, ctx, agent, description, prompt)?;
    let text = run_child(app, &child, &agent, ctx.cancel.child_token(), 1).await?;
    Ok((child.id, text))
}

/// Create the subagent session with its first message.
fn create_child(app: &Arc<App>, ctx: &ToolCtx, agent_name: &str, description: &str, prompt: String) -> Result<(Session, Agent)> {
    let parent = app.session(&ctx.session_id)?;
    let root = ctx.root();
    let parent_agent = prompts::find_agent(app, &root, &ctx.cwd, &parent.agent);
    let agents = prompts::agents(app, &root, &ctx.cwd);
    let mut agent = agents
        .iter()
        .find(|a| a.name == agent_name && a.mode != "primary")
        .cloned()
        .ok_or_else(|| {
            let names: Vec<_> = agents.iter().filter(|a| a.mode != "primary").map(|a| a.name.as_str()).collect();
            anyhow!("unknown subagent '{agent_name}'. Available: {}", names.join(", "))
        })?;
    // Read-only parents may only spawn read-only subagents.
    if parent_agent.read_only && !agent.read_only {
        agent = agents.into_iter().find(|a| a.name == "explore").unwrap_or(agent);
    }
    if let Some(m) = &agent.model {
        app.registry().resolve(m).map_err(|e| anyhow!("subagent '{}' has an unusable model: {e:#}", agent.name))?;
    }
    let now = now_ms();
    let child = Session {
        id: new_id(),
        title: description.to_string(),
        cwd: parent.cwd.clone(),
        model: agent.model.clone().unwrap_or_else(|| parent.model.clone()),
        agent: agent.name.clone(),
        permission_mode: parent.permission_mode.clone(),
        parent_id: Some(parent.id.clone()),
        owner_id: parent.owner_id.clone(),
        created_at: now,
        updated_at: now,
        todos: Value::Null,
        last_usage: None,
    };
    let child = app.save_session(child)?;
    let msg = Message::new(Role::User, vec![Part::Text { text: prompt }]);
    app.store.add_message(&child.id, &msg)?;
    app.emit(Event::MessageAdded { session_id: child.id.clone(), message: msg });
    Ok((child, agent))
}

/// Run a subagent session to completion, honouring the subagent limits.
async fn run_child(app: Arc<App>, child: &Session, agent: &Agent, cancel: CancellationToken, depth: usize) -> Result<String> {
    // Registered like a normal run so the user can stop it from its own session.
    app.with_runtime(&child.id, |r| r.cancel = Some(cancel.clone()));
    app.emit(Event::SessionStatus { session_id: child.id.clone(), busy: true });
    // Only first-level subagents count against the limits: nested subagents would
    // otherwise deadlock waiting for a slot their own parent holds.
    let _slot = if depth <= 1 {
        let (slot, waited) = app.subagent_slot(&agent.name, agent.max_parallel, &child.id, &cancel).await;
        if waited {
            app.emit(Event::Waiting { session_id: child.id.clone(), reason: None });
        }
        slot
    } else {
        None
    };
    let res = if cancel.is_cancelled() {
        Err(anyhow!("aborted"))
    } else {
        run_session_boxed(app.clone(), child.id.clone(), cancel, depth).await
    };
    app.with_runtime(&child.id, |r| r.cancel = None);
    app.emit(Event::SessionStatus { session_id: child.id.clone(), busy: false });
    let text = res?;
    Ok(if text.trim().is_empty() { "(subagent returned no text)".into() } else { text })
}

fn is_retryable(e: &anyhow::Error) -> bool {
    let s = format!("{e:#}");
    s.contains("HTTP 429") || s.contains("HTTP 5") || s.contains("overloaded") || s.contains("connection") || s.contains("timed out")
}

/// One model request with streaming to the UI, usage recording and retries.
async fn stream_step(
    app: &Arc<App>,
    session: &Session,
    agent: &str,
    model_spec: &str,
    provider: Arc<dyn Provider>,
    req: ChatRequest,
    cancel: &CancellationToken,
) -> Result<Step> {
    let mut attempt = 0;
    loop {
        attempt += 1;
        let msg_id = new_id();
        let (tx, mut rx) = mpsc::channel::<StreamEvent>(256);
        let p = provider.clone();
        let r = req.clone();
        let c = cancel.clone();
        // Wait for a free slot on the model/host; not counted as time to first token.
        let registry = app.registry();
        if registry.is_saturated(model_spec) {
            app.emit(Event::Waiting { session_id: session.id.clone(), reason: Some(format!("model:{model_spec}")) });
        }
        let slot = tokio::select! {
            s = registry.acquire(model_spec) => s,
            _ = cancel.cancelled() => bail!("aborted"),
        };
        app.emit(Event::Waiting { session_id: session.id.clone(), reason: None });
        let handle = tokio::spawn(async move {
            let _slot = slot;
            p.stream(r, tx, c).await
        });

        let started = Instant::now();
        let mut ttft: Option<u128> = None;
        let mut parts: Vec<Part> = Vec::new();
        let mut usage = Usage::default();
        let mut stop_reason = String::new();
        while let Some(ev) = rx.recv().await {
            if ttft.is_none() && !matches!(ev, StreamEvent::Usage(_) | StreamEvent::Done { .. }) {
                ttft = Some(started.elapsed().as_millis());
            }
            match ev {
                StreamEvent::TextDelta(t) => {
                    match parts.last_mut() {
                        Some(Part::Text { text }) => text.push_str(&t),
                        _ => parts.push(Part::Text { text: t.clone() }),
                    }
                    app.emit(Event::TextDelta { session_id: session.id.clone(), message_id: msg_id.clone(), text: t });
                }
                StreamEvent::ReasoningDelta(t) => {
                    match parts.last_mut() {
                        Some(Part::Reasoning { text, signature: None }) => text.push_str(&t),
                        _ => parts.push(Part::Reasoning { text: t.clone(), signature: None }),
                    }
                    app.emit(Event::ReasoningDelta { session_id: session.id.clone(), message_id: msg_id.clone(), text: t });
                }
                StreamEvent::ReasoningSignature(sig) => match parts.last_mut() {
                    Some(Part::Reasoning { signature, .. }) if signature.is_none() => *signature = Some(sig),
                    // Thinking without visible text (display omitted / redacted).
                    _ => parts.push(Part::Reasoning { text: String::new(), signature: Some(sig) }),
                },
                StreamEvent::ToolCall { id, name, input, extra } => {
                    app.emit(Event::ToolStarted {
                        session_id: session.id.clone(),
                        message_id: msg_id.clone(),
                        call_id: id.clone(),
                        name: name.clone(),
                        input: input.clone(),
                    });
                    parts.push(Part::ToolCall { id, name, input, extra });
                }
                StreamEvent::Usage(u) => usage = u,
                StreamEvent::Done { stop_reason: s } => stop_reason = s,
            }
        }
        let result = handle.await.map_err(|e| anyhow!("provider task failed: {e}"))?;
        let duration = started.elapsed().as_millis() as u64;
        let (provider_name, model_id) = model_spec.split_once('/').unwrap_or(("", model_spec));
        let record = |error: bool, usage: &Usage| UsageRecord {
            ts: now_ms(),
            session_id: session.id.clone(),
            root_session_id: app.root_session_id(&session.id),
            provider: provider_name.into(),
            model: model_id.into(),
            agent: agent.into(),
            input_tokens: usage.input_tokens,
            output_tokens: usage.output_tokens,
            cache_read_tokens: usage.cache_read_tokens,
            cache_write_tokens: usage.cache_write_tokens,
            ttft_ms: ttft.unwrap_or(duration as u128) as u64,
            duration_ms: duration,
            cost: req.model.cost(usage),
            error,
        };

        if let Err(e) = result {
            let _ = app.store.add_usage(&record(true, &usage));
            if parts.is_empty() && attempt < 4 && is_retryable(&e) && !cancel.is_cancelled() {
                let wait = std::time::Duration::from_secs(2u64.pow(attempt));
                app.emit(Event::Error {
                    session_id: session.id.clone(),
                    message: format!("{e:#} – retrying in {}s", wait.as_secs()),
                });
                tokio::time::sleep(wait).await;
                continue;
            }
            return Err(e);
        }
        if cancel.is_cancelled() {
            // Keep what was streamed, but drop unfinished tool calls and unsigned thinking.
            parts.retain(|p| matches!(p, Part::Text { .. }));
        }
        if usage.input_tokens + usage.output_tokens > 0 || !cancel.is_cancelled() {
            let _ = app.store.add_usage(&record(false, &usage));
        }
        let mut message = Message::new(Role::Assistant, parts);
        message.id = msg_id;
        message.model = Some(model_spec.to_string());
        message.usage = Some(usage.clone());
        if !message.parts.is_empty() {
            app.store.add_message(&session.id, &message)?;
            app.emit(Event::MessageAdded { session_id: session.id.clone(), message: message.clone() });
        }
        let mut s = app.session(&session.id)?;
        s.last_usage = Some(usage);
        app.save_session(s)?;
        return Ok(Step { message, stop_reason });
    }
}

/// Collect a complete, tool-free answer (titles, summaries).
async fn complete_text(app: &App, spec: &str, provider: Arc<dyn Provider>, req: ChatRequest, cancel: CancellationToken) -> Result<(String, Usage)> {
    let slot = app.registry().acquire(spec).await;
    let (tx, mut rx) = mpsc::channel(256);
    let handle = tokio::spawn(async move {
        let _slot = slot;
        provider.stream(req, tx, cancel).await
    });
    let mut text = String::new();
    let mut usage = Usage::default();
    while let Some(ev) = rx.recv().await {
        match ev {
            StreamEvent::TextDelta(t) => text.push_str(&t),
            StreamEvent::Usage(u) => usage = u,
            _ => {}
        }
    }
    handle.await??;
    Ok((text, usage))
}

fn small_model(app: &App, session: &Session) -> String {
    let cfg = app.config();
    if !cfg.server.small_model.is_empty() && app.registry().resolve(&cfg.server.small_model).is_ok() {
        cfg.server.small_model.clone()
    } else {
        session.model.clone()
    }
}

fn record_side_usage(app: &App, session: &Session, spec: &str, model: &crate::config::ModelConfig, usage: &Usage, agent: &str) {
    let (p, m) = spec.split_once('/').unwrap_or(("", spec));
    let _ = app.store.add_usage(&UsageRecord {
        ts: now_ms(),
        session_id: session.id.clone(),
        root_session_id: app.root_session_id(&session.id),
        provider: p.into(),
        model: m.into(),
        agent: agent.into(),
        input_tokens: usage.input_tokens,
        output_tokens: usage.output_tokens,
        cache_read_tokens: usage.cache_read_tokens,
        cache_write_tokens: usage.cache_write_tokens,
        ttft_ms: 0,
        duration_ms: 0,
        cost: model.cost(usage),
        error: false,
    });
}

fn maybe_generate_title(app: Arc<App>, session_id: &str, text: &str) {
    let Ok(session) = app.session(session_id) else { return };
    if session.parent_id.is_some() || !session.title.is_empty() {
        return;
    }
    let text: String = text.chars().take(2000).collect();
    tokio::spawn(async move {
        let spec = small_model(&app, &session);
        let Ok((provider, mut model)) = app.registry().resolve(&spec) else { return };
        model.effort = Some("low".into());
        let req = ChatRequest {
            system: "Write a short title (max 6 words) for a coding session that starts with the user's message below. Use the user's language. Reply with the title only, no quotes or punctuation at the end.".into(),
            messages: vec![Message::new(Role::User, vec![Part::Text { text }])],
            tools: vec![],
            max_tokens: 2000,
            reasoning: false,
            model: model.clone(),
        };
        match complete_text(&app, &spec, provider, req, CancellationToken::new()).await {
            Ok((title, usage)) => {
                record_side_usage(&app, &session, &spec, &model, &usage, "title");
                let title = title.lines().find(|l| !l.trim().is_empty()).unwrap_or("").trim().trim_matches('"').to_string();
                if let (false, Ok(mut s)) = (title.is_empty(), app.session(&session.id)) {
                    if s.title.is_empty() {
                        s.title = title.chars().take(80).collect();
                        let _ = app.save_session(s);
                    }
                }
            }
            Err(e) => tracing::debug!("title generation failed: {e:#}"),
        }
    });
}

const COMPACT_PROMPT: &str = "Summarize this coding session so far so that the work can continue seamlessly from the summary alone. Include: the user's goals and explicit requests (verbatim where important), decisions made, files created/changed (with paths) and their current state, commands/tests run and results, open problems, and the exact next steps. Be thorough but concise; use headings and bullet points.";

/// Replace the active history with a summary (the old messages stay visible in the UI).
pub async fn compact(app: Arc<App>, session_id: &str, cancel: CancellationToken) -> Result<()> {
    let session = app.session(session_id)?;
    let history = active_history(app.store.messages(session_id)?);
    if history.is_empty() {
        return Ok(());
    }
    let spec = session.model.clone();
    let (provider, model) = app.registry().resolve(&spec)?;
    let mut messages = history;
    // Strip images to keep the summary request small.
    for m in &mut messages {
        m.parts.retain(|p| !matches!(p, Part::Image { .. }));
        for p in &mut m.parts {
            if let Part::ToolResult { images, .. } = p {
                images.clear();
            }
        }
    }
    messages.push(Message::new(Role::User, vec![Part::Text { text: COMPACT_PROMPT.into() }]));
    let req = ChatRequest {
        system: "You are a precise summarizer of software engineering sessions.".into(),
        messages,
        tools: vec![],
        max_tokens: 16_000,
        reasoning: false,
        model: model.clone(),
    };
    let (summary, usage) = complete_text(&app, &spec, provider, req, cancel).await?;
    record_side_usage(&app, &session, &spec, &model, &usage, "compaction");
    if summary.trim().is_empty() {
        bail!("compaction produced an empty summary");
    }
    save_user_parts(&app, session_id, vec![Part::Summary { text: summary }])?;
    let mut s = app.session(session_id)?;
    s.last_usage = None;
    // Files must be re-read after compaction before editing.
    app.with_runtime(session_id, |r| r.read_files.clear());
    app.save_session(s)?;
    Ok(())
}

#[cfg(test)]
mod tests {
    use super::*;

    fn call(id: &str) -> Part {
        Part::ToolCall { id: id.into(), name: "bash".into(), input: serde_json::json!({}), extra: None }
    }

    #[test]
    fn dangling_tool_calls_get_results() {
        let a = Message::new(Role::Assistant, vec![call("1"), call("2")]);
        let r = Message::new(Role::User, vec![tool_result("1", "bash", "ok".into(), false, None)]);
        let a2 = Message::new(Role::Assistant, vec![call("3")]);
        let h = active_history(vec![Message::new(Role::User, vec![Part::Text { text: "hi".into() }]), a, r, a2]);
        // user, assistant(1,2), user(result 2 + result 1), assistant(3), user(result 3)
        assert_eq!(h.len(), 5);
        let ids = |m: &Message| m.parts.iter().filter_map(|p| if let Part::ToolResult { call_id, .. } = p { Some(call_id.clone()) } else { None }).collect::<Vec<_>>();
        assert_eq!(ids(&h[2]), vec!["2", "1"]);
        assert_eq!(ids(&h[4]), vec!["3"]);
    }
}