rpi_cli/

modes.rs

//! Output modes. Mirrors the v1-relevant slice of the TS
//! `packages/coding-agent/src/modes/{print-mode,json-event,rpc-mode}.ts` — the
//! three run shapes a harness-backed CLI needs:
//!
//! - [`print`] — single-shot: send the prompt(s), print the final assistant
//!   text (or the error) to stdout, exit. Mirrors TS `runPrintMode` (text mode).
//! - [`json`] — single-shot streaming: emit each harness event as a JSON line
//!   on stdout, then the final outcome. Mirrors TS `runPrintMode`
//!   (`mode === "json"`) + [`json_event::toJsonEvent`].
//! - [`interactive`] — a minimal line-oriented REPL: read prompts from stdin,
//!   run each, print the assistant text, loop until EOF / `/exit`. v1 does NOT
//!   port the TS `InteractiveMode` TUI (`modes/interactive/*` — a full terminal
//!   UI with Ink/React components); this is a deliberately minimal replacement,
//!   documented in the initial port notes (retired).
//!
//! All three drive the same `AgentHarness` via `AgentLane::prompt_text`.

use std::io::{BufRead, IsTerminal, Write};
use std::sync::{Arc, Mutex};

use rpi_agent::events::AgentEvent;
use rpi_ai::types::{AssistantMessage, Content, ImageContent, StopReason, UserMessage};
use rpi_harness::agent_harness::{
    AgentHarness, AgentLane, CompactionOutcome, CompactionResult, HarnessRunOutcome,
};
use rpi_harness::events::{HarnessEvent, RunEndOutcome};
use rpi_harness::session::types::{EntryOrder, LaneRecord, RecordQuery};

use crate::args::Args;
use crate::rpc::protocol as rpc_protocol;
use crate::rpc::writer::{start_rpc_writer, write_agent_event, write_line, RpcWriter};

/// Extract the concatenated text content from an assistant message. Mirrors the
/// TS print-mode loop (`for content of assistantMsg.content if type===text`).
pub fn assistant_text(msg: &AssistantMessage) -> String {
    msg.content
        .iter()
        .filter_map(|c| match c {
            Content::Text(t) => Some(t.text.clone()),
            _ => None,
        })
        .collect()
}

/// The exit code a run's outcome maps to. Mirrors TS print mode: error/aborted
/// ⇒ exit 1; everything else ⇒ 0.
pub fn outcome_exit_code(outcome: &HarnessRunOutcome) -> i32 {
    match outcome {
        HarnessRunOutcome::Failed { .. } | HarnessRunOutcome::Aborted { .. } => 1,
        _ => 0,
    }
}

/// `print` mode: send the initial message (prompt text + inline `@file`
/// expansions), then any follow-up messages, print the final assistant text,
/// return the exit code. Mirrors TS `runPrintMode` (text).
pub async fn print(
    harness: &AgentHarness,
    _args: &Args,
    initial: Option<String>,
    extra_messages: &[String],
    initial_images: Vec<ImageContent>,
) -> i32 {
    let lane: Arc<dyn AgentLane> = harness.lane("main");

    let mut last_exit = 0;
    let mut last_msg: Option<AssistantMessage> = None;

    // The initial prompt (and its `@file` attachments) go in one user message;
    // extra positionals are separate prompts (mirrors the TS loop).
    let mut prompts: Vec<String> = Vec::new();
    if let Some(init) = initial {
        prompts.push(init);
    }
    for m in extra_messages {
        prompts.push(m.clone());
    }

    if prompts.is_empty() {
        // Nothing to do — print mode with no prompt is a no-op success.
        return 0;
    }

    let mut images = initial_images;
    for prompt in prompts {
        match lane.prompt_text(&prompt, std::mem::take(&mut images)).await {
            Ok(result) => {
                last_exit = outcome_exit_code(&result.outcome);
                match &result.outcome {
                    HarnessRunOutcome::Completed { final_message, .. }
                    | HarnessRunOutcome::Aborted { final_message, .. } => {
                        last_msg = Some(final_message.clone());
                    }
                    HarnessRunOutcome::Failed {
                        error,
                        final_message,
                        ..
                    } => {
                        if let Some(m) = final_message {
                            if m.stop_reason == StopReason::Error {
                                if let Some(em) = &m.error_message {
                                    eprintln!("{em}");
                                }
                            }
                        }
                        eprintln!("run failed: {error:?}");
                    }
                    HarnessRunOutcome::Suspended { .. } => {
                        eprintln!("run suspended (deferred) — resume is not supported in v1");
                        last_exit = 1;
                    }
                }
            }
            Err(e) => {
                eprintln!("prompt rejected: {e}");
                return 1;
            }
        }
    }

    // Print the final assistant text to stdout (TS: writeRawStdout text + "\n").
    if let Some(m) = &last_msg {
        match m.stop_reason {
            StopReason::Error => {
                if let Some(em) = &m.error_message {
                    eprintln!("{em}");
                }
                last_exit = 1;
            }
            StopReason::Aborted => {
                eprintln!("request aborted");
                last_exit = 1;
            }
            _ => {
                let text = assistant_text(m);
                let mut out = std::io::stdout();
                let _ = out.write_all(text.as_bytes());
                if !text.ends_with('\n') {
                    let _ = out.write_all(b"\n");
                }
                let _ = out.flush();
            }
        }
    }

    last_exit
}

/// `json` mode: emit each harness event as a JSON line on stdout, run the
/// prompts, then emit a terminal `result` line carrying the outcome + final
/// text. Mirrors TS `runPrintMode` (`mode === "json"`) streaming every event.
pub async fn json(
    harness: &AgentHarness,
    _args: &Args,
    initial: Option<String>,
    extra_messages: &[String],
    initial_images: Vec<ImageContent>,
    mut agent_events: Option<tokio::sync::broadcast::Receiver<AgentEvent>>,
) -> i32 {
    let lane: Arc<dyn AgentLane> = harness.lane("main");
    let collected: Arc<Mutex<Vec<HarnessEvent>>> = Arc::new(Mutex::new(Vec::new()));
    let collected_for_watch = collected.clone();

    // A watch captures every event (RunStart fires inline during prompt_text,
    // before a post-call listener could attach — same reason as the M5g test).
    let mut watch = harness.events().watch(|| ());
    watch.start(Arc::new(move |event: &HarnessEvent| {
        // Emit each event live as JSON, and also buffer for the final summary.
        emit_json_event(event);
        collected_for_watch.lock().unwrap().push(event.clone());
    }));
    // Keep the watch alive for the whole run. Leaking is acceptable for a
    // single-shot CLI process (the bus outlives this scope anyway).
    std::mem::forget(watch);

    // The harness bus carries run lifecycle events; the agent receiver carries
    // the native fine-grained stream (turns, message deltas, and tools).
    let (agent_done_tx, mut agent_done_rx) = tokio::sync::mpsc::unbounded_channel::<()>();
    let agent_event_task = agent_events.take().map(|mut rx| {
        tokio::spawn(async move {
            let done_tx = agent_done_tx;
            loop {
                match rx.recv().await {
                    Ok(event) => {
                        let terminal = event.is_terminal();
                        emit_agent_event(&event);
                        if terminal {
                            let _ = done_tx.send(());
                        }
                    }
                    Err(tokio::sync::broadcast::error::RecvError::Lagged(_)) => continue,
                    Err(tokio::sync::broadcast::error::RecvError::Closed) => break,
                }
            }
        })
    });

    let mut prompts: Vec<String> = Vec::new();
    if let Some(init) = initial {
        prompts.push(init);
    }
    for m in extra_messages {
        prompts.push(m.clone());
    }

    let mut last_exit = 0;
    let mut final_outcome: Option<HarnessRunOutcome> = None;

    let mut images = initial_images;
    for prompt in prompts {
        match lane.prompt_text(&prompt, std::mem::take(&mut images)).await {
            Ok(result) => {
                // The harness resolves after its run outcome, while the
                // broadcast listener may still be scheduling the terminal
                // AgentEnd line. Wait briefly so JSON consumers see the full
                // lifecycle before the final result summary. The timeout is
                // deliberately bounded for custom/older harness emitters.
                let _ = tokio::time::timeout(
                    std::time::Duration::from_millis(250),
                    agent_done_rx.recv(),
                )
                .await;
                last_exit = outcome_exit_code(&result.outcome);
                final_outcome = Some(result.outcome);
            }
            Err(e) => {
                // Emit a structured error line + exit.
                let line = serde_json::json!({
                    "type": "error",
                    "error": e.to_string(),
                });
                println!("{line}");
                if let Some(task) = agent_event_task {
                    task.abort();
                }
                return 1;
            }
        }
    }

    // Terminal result summary.
    let (outcome_str, final_text) = match final_outcome {
        Some(HarnessRunOutcome::Completed { final_message, .. }) => {
            ("completed", Some(assistant_text(&final_message)))
        }
        Some(HarnessRunOutcome::Aborted { final_message, .. }) => {
            ("aborted", Some(assistant_text(&final_message)))
        }
        Some(HarnessRunOutcome::Failed { final_message, .. }) => {
            let t = final_message.as_ref().map(assistant_text);
            ("failed", t)
        }
        Some(HarnessRunOutcome::Suspended { .. }) => ("suspended", None),
        None => ("idle", None),
    };
    let result_line = serde_json::json!({
        "type": "result",
        "outcome": outcome_str,
        "finalText": final_text,
    });
    println!("{result_line}");
    if let Some(task) = agent_event_task {
        task.abort();
    }
    last_exit
}

fn emit_agent_event(event: &AgentEvent) {
    println!("{}", agent_event_json(event));
}

/// `rpc` mode: a headless JSONL command loop (the server-side agent).
///
/// Reads one JSON command per line from stdin (`{id?, type, ...}`), drives the
/// `main` lane, and writes one JSON object per line to stdout:
/// - fine-grained `AgentEvent`s via [`agent_event_json`] (each carries `type`),
/// - a `{"type":"response","id":<id>,"status":"ok"|"error",...}` per command.
///
/// Commands: `prompt`, `resume_pending`, and `compact` (drive the lane),
/// `abort`, `get_state`, `set_model`,
/// `set_thinking_level`, `set_active_tools`, `ping`, `stop`. `stop` (or stdin
/// EOF) ends the loop. This is the process `rpi-server` spawns and that
/// `rpi --connect` talks to. Mirrors TS `runRpcMode` / `modes/rpc`.
///
/// The extension `UiDialog` mailbox is bridged over the same channel: the host
/// emits `{"type":"ui_dialog_request","requestId","toolCallId","ui"}`, and the
/// client answers with `{"type":"ui_dialog_response","requestId","answer"}` (or
/// `"cancelled":true`). A `prompt` runs in its own task so this loop keeps
/// reading — that is what lets `abort` and dialog answers arrive mid-run.
pub async fn rpc(
    harness: &AgentHarness,
    mut agent_events: Option<tokio::sync::broadcast::Receiver<AgentEvent>>,
    model_catalog: Vec<rpi_ai::Model>,
    ui_dialog: rpi_extensions::UiDialogMailbox,
    oauth_snapshot: Option<Arc<rpi_extensions::RegistrySnapshot>>,
) -> i32 {
    rpc_with_io(
        harness,
        agent_events.take(),
        model_catalog,
        ui_dialog,
        oauth_snapshot,
        tokio::io::BufReader::new(tokio::io::stdin()),
        std::io::stdout(),
    )
    .await
}

async fn rpc_with_io<R, W>(
    harness: &AgentHarness,
    mut agent_events: Option<tokio::sync::broadcast::Receiver<AgentEvent>>,
    model_catalog: Vec<rpi_ai::Model>,
    ui_dialog: rpi_extensions::UiDialogMailbox,
    oauth_snapshot: Option<Arc<rpi_extensions::RegistrySnapshot>>,
    input: R,
    output: W,
) -> i32
where
    R: tokio::io::AsyncBufRead + Unpin,
    W: Write + Send + 'static,
{
    let lane: Arc<dyn AgentLane> = harness.lane("main");
    // One blocking writer owns stdout. Producers only enqueue bounded messages,
    // so pipe backpressure never blocks a Tokio worker on IO or a std mutex.
    let (stdout, writer_task, writer_failed) = start_rpc_writer(output);
    let pump_stop = tokio_util::sync::CancellationToken::new();

    // A completion response is queued only after its final events, so a
    // subsequent prompt cannot overtake the previous run's agent_end.
    let (event_flush_tx, mut event_flush_rx) =
        tokio::sync::mpsc::unbounded_channel::<tokio::sync::oneshot::Sender<()>>();
    // Stream fine-grained AgentEvents as JSONL until the receiver closes.
    let event_task = agent_events.take().map(|mut rx| {
        let out = stdout.clone();
        let stop = pump_stop.clone();
        tokio::spawn(async move {
            loop {
                let received = tokio::select! {
                    _ = stop.cancelled() => {
                        drain_rpc_events(&out, &mut rx).await;
                        break;
                    }
                    Some(done) = event_flush_rx.recv() => {
                        drain_rpc_events(&out, &mut rx).await;
                        let _ = done.send(());
                        continue;
                    }
                    event = rx.recv() => event,
                };
                match received {
                    Ok(event) => write_agent_event(&out, &event).await,
                    Err(tokio::sync::broadcast::error::RecvError::Lagged(_)) => continue,
                    Err(tokio::sync::broadcast::error::RecvError::Closed) => break,
                }
            }
        })
    });

    let event_flush = event_task.as_ref().map(|_| event_flush_tx);

    // Bridge the extension `UiDialog` mailbox over the rpc channel so a
    // permission-gate extension can ask the client a question mid-run: we emit
    // `{"type":"ui_dialog_request",…}` and answer it from a
    // `{"type":"ui_dialog_response",…}` on stdin. Without this a headless run
    // leaves the mailbox detached and every `UiDialog` call fails.
    ui_dialog.attach();
    let dialog_task = {
        let mailbox = ui_dialog.clone();
        let out = stdout.clone();
        let stop = pump_stop.clone();
        tokio::spawn(async move {
            let mut shown = std::collections::HashSet::new();
            while !stop.is_cancelled() {
                while let Some(request) = mailbox.take_pending() {
                    let event = crate::remote::protocol::RemoteEvent::UiDialogRequest {
                        request_id: request.request_id.clone(),
                        tool_call_id: request.tool_call_id,
                        ui: request.ui,
                    };
                    write_line(&out, &serde_json::to_value(event).unwrap()).await;
                    shown.insert(request.request_id);
                }
                let closed: Vec<String> = shown
                    .iter()
                    .filter(|id| !mailbox.is_pending(id))
                    .cloned()
                    .collect();
                for request_id in closed {
                    shown.remove(&request_id);
                    let event = crate::remote::protocol::RemoteEvent::UiDialogClosed { request_id };
                    write_line(&out, &serde_json::to_value(event).unwrap()).await;
                }
                tokio::select! {
                    _ = stop.cancelled() => break,
                    _ = tokio::time::sleep(std::time::Duration::from_millis(30)) => {}
                }
            }
        })
    };

    // A `prompt` runs in a spawned task so this loop keeps reading stdin while
    // the run is active — otherwise `abort` and `ui_dialog_response` could not
    // be delivered mid-run. Only one prompt may be in flight at a time.
    let run_active = Arc::new(std::sync::atomic::AtomicBool::new(false));
    let mut run_task: Option<tokio::task::JoinHandle<()>> = None;
    let mut run_cancel = tokio_util::sync::CancellationToken::new();
    let mut cancel_on_exit = false;

    use tokio::io::AsyncBufReadExt;

    let mut lines = input.lines();
    write_line(&stdout, &rpc_protocol::ready()).await;

    loop {
        let incoming = tokio::select! {
            _ = writer_failed.cancelled() => {
                cancel_on_exit = true;
                break;
            }
            line = lines.next_line() => line,
        };
        let line = match incoming {
            Ok(Some(line)) => line,
            Ok(None) => break, // EOF
            Err(error) => {
                eprintln!("[rpi] rpc stdin error: {error}");
                cancel_on_exit = true;
                break;
            }
        };
        let line = line.trim();
        if line.is_empty() {
            continue;
        }
        let msg: serde_json::Value = match serde_json::from_str(line) {
            Ok(value) => value,
            Err(error) => {
                write_line(
                    &stdout,
                    &rpc_protocol::response(
                        None,
                        "invalid_request",
                        Err(format!("invalid json: {error}")),
                    ),
                )
                .await;
                continue;
            }
        };
        let id = msg.get("id").cloned().unwrap_or(serde_json::Value::Null);
        let kind = msg
            .get("type")
            .and_then(|v| v.as_str())
            .unwrap_or("")
            .to_string();

        if rpc_protocol::SUPPORTED_COMMANDS.contains(&kind.as_str())
            && serde_json::from_value::<rpc_protocol::Request>(msg.clone()).is_err()
        {
            write_line(
                &stdout,
                &rpc_protocol::response(
                    Some(id.clone()),
                    &kind,
                    Err("invalid command payload".to_string()),
                ),
            )
            .await;
            continue;
        }

        match kind.as_str() {
            "ui_dialog_response" => {
                let request_id = msg
                    .get("requestId")
                    .and_then(|v| v.as_str())
                    .unwrap_or("")
                    .to_string();
                let result = if msg
                    .get("cancelled")
                    .and_then(|v| v.as_bool())
                    .unwrap_or(false)
                {
                    ui_dialog.cancel(&request_id)
                } else {
                    let answer = msg
                        .get("answer")
                        .cloned()
                        .unwrap_or(serde_json::Value::Null);
                    ui_dialog.respond(&request_id, answer)
                };
                if msg.get("id").is_some() {
                    let response = match result {
                        Ok(()) => rpc_protocol::response(
                            Some(id),
                            "ui_dialog_response",
                            Ok(Some(
                                serde_json::json!({"accepted":true,"requestId":request_id}),
                            )),
                        ),
                        Err(error) => {
                            rpc_protocol::response(Some(id), "ui_dialog_response", Err(error))
                        }
                    };
                    write_line(&stdout, &response).await;
                } else if let Err(error) = result {
                    eprintln!("[rpi] ui_dialog_response: {error}");
                }
                continue;
            }
            "stop" => {
                cancel_on_exit = true;
                write_line(
                    &stdout,
                    &rpc_protocol::response(
                        Some(id.clone()),
                        "stop",
                        Ok(Some(serde_json::json!({"stopped": true}))),
                    ),
                )
                .await;
                break;
            }
            "prompt" | "resume_pending" | "compact" => {
                let resume = kind == "resume_pending";
                let compact = kind == "compact";
                if run_active.swap(true, std::sync::atomic::Ordering::AcqRel) {
                    write_line(
                        &stdout,
                        &rpc_protocol::response(
                            Some(id.clone()),
                            &kind,
                            Err("a run is already active".to_string()),
                        ),
                    )
                    .await;
                    continue;
                }
                if let Some(task) = run_task.take() {
                    let _ = task.await;
                }
                run_cancel = tokio_util::sync::CancellationToken::new();
                let cancel_task = run_cancel.clone();
                // Resuming spends the tokens the interrupted run never did, so it
                // only proceeds when recovery actually left a pending run.
                if resume && !lane.has_pending_resume().await {
                    run_active.store(false, std::sync::atomic::Ordering::Release);
                    write_line(
                        &stdout,
                        &rpc_protocol::response(
                            Some(id),
                            "resume_pending",
                            Ok(Some(serde_json::json!({"resumed": false}))),
                        ),
                    )
                    .await;
                    continue;
                }
                let content = msg
                    .get("message")
                    .and_then(|v| v.as_str())
                    .map(str::to_owned);
                let lane_task = Arc::clone(&lane);
                let out_task = stdout.clone();
                let active_task = Arc::clone(&run_active);
                let mailbox_task = ui_dialog.clone();
                let flush_events = event_flush.clone();
                let Some(content) = content.or_else(|| (resume || compact).then(String::new))
                else {
                    active_task.store(false, std::sync::atomic::Ordering::Release);
                    write_line(
                        &stdout,
                        &rpc_protocol::response(
                            Some(id),
                            "prompt",
                            Err("prompt requires a string `message`".to_string()),
                        ),
                    )
                    .await;
                    continue;
                };
                write_line(
                    &stdout,
                    &rpc_protocol::response(
                        Some(id),
                        &kind,
                        Ok(Some(serde_json::json!({"accepted": true}))),
                    ),
                )
                .await;
                let instructions = msg
                    .get("customInstructions")
                    .and_then(|v| v.as_str())
                    .map(str::to_owned);
                run_task = Some(tokio::spawn(async move {
                    let result = if compact {
                        rpc_compact_cancellable(&lane_task, instructions, cancel_task).await
                    } else if resume {
                        rpc_resume_cancellable(&lane_task, cancel_task).await
                    } else {
                        rpc_prompt_cancellable(&lane_task, &content, cancel_task).await
                    };
                    if let Some(flush) = flush_events {
                        let (done, drained) = tokio::sync::oneshot::channel();
                        if flush.send(done).is_ok() {
                            let _ = drained.await;
                        }
                    }
                    let (outcome, aborted, error) = match &result {
                        Ok(value) => {
                            let outcome = value
                                .get("outcome")
                                .and_then(|v| v.as_str())
                                .unwrap_or("completed")
                                .to_string();
                            let error = value
                                .get("error")
                                .and_then(|v| v.as_str())
                                .map(str::to_owned);
                            (outcome.clone(), outcome == "aborted", error)
                        }
                        Err(ref error) => ("failed".to_string(), false, Some(error.clone())),
                    };
                    // Recovery may leave a continuable run behind an abort, so the
                    // settlement carries what the client needs to offer "continue".
                    let pending_resume = lane_task.has_pending_resume().await;
                    let summary = result
                        .as_ref()
                        .ok()
                        .and_then(|value| value.get("summary"))
                        .cloned();
                    let settled = serde_json::json!({
                        "type": "agent_settled",
                        "runId": result.as_ref().ok().and_then(|value| value.get("runId")),
                        "outcome": outcome,
                        "aborted": aborted,
                        "error": error,
                        "summary": summary,
                        "pendingResume": pending_resume,
                    });
                    write_line(&out_task, &settled).await;
                    mailbox_task.cancel_all();
                    active_task.store(false, std::sync::atomic::Ordering::Release);
                }));
                continue;
            }
            _ => {}
        }

        let result = match kind.as_str() {
            "abort" if run_active.load(std::sync::atomic::Ordering::Acquire) => {
                // This also covers abort arriving before the task enters the lane.
                run_cancel.cancel();
                Ok(serde_json::json!({"aborted": true}))
            }
            "abort" => lane
                .abort()
                .await
                .map(|_| serde_json::json!({"aborted": true}))
                .map_err(|e| e.to_string()),
            "get_state" => Ok(rpc_get_state(&lane, harness).await),
            "get_messages" => rpc_get_messages(&lane, harness, &msg).await,
            "get_history" => rpc_get_history(&lane, harness, &msg).await,
            "get_stats" => rpc_get_stats(&lane, harness).await,
            "get_queue" => rpc_get_queue(&lane, harness, &msg).await,
            "queue_message" => rpc_queue_message(&lane, &msg).await,
            "cancel_queued" => rpc_cancel_queued(&lane, &msg).await,
            "clear_queue" => rpc_clear_queue(&lane).await,
            "navigate" => rpc_navigate(&lane, &msg).await,
            "get_available_models" => Ok(rpc_get_available_models(&model_catalog)),
            "get_auth_status" => Ok(rpc_get_auth_status(&model_catalog, oauth_snapshot.as_ref())),
            "oauth" => rpc_oauth(&oauth_snapshot, &msg),
            "set_api_key" => rpc_set_api_key(&msg).await,
            "set_model" => rpc_set_model(&lane, &model_catalog, &msg).await,
            "get_available_thinking_levels" => Ok(rpc_get_available_thinking_levels(&lane).await),
            "set_thinking_level" => rpc_set_thinking_level(&lane, &msg).await,
            "get_available_tools" => Ok(rpc_get_available_tools(&lane, harness).await),
            "set_active_tools" => rpc_set_active_tools(&lane, &msg).await,
            "ping" => Ok(serde_json::json!({"pong": true})),
            other => Err(format!("unknown command type: {other}")),
        };

        let response = rpc_protocol::response(Some(id.clone()), &kind, result.map(Some));
        write_line(&stdout, &response).await;
    }

    // EOF closes the input half of the protocol, not an accepted prompt.
    // There can be no more dialog answers, so release approvals, then await the
    // run before SessionShutdown can unload plugins or the runtime exits.
    ui_dialog.detach();
    if cancel_on_exit {
        run_cancel.cancel();
    }
    let mut exit_code = 0;
    if let Some(task) = run_task {
        if let Err(error) = task.await {
            eprintln!("[rpi] rpc prompt task failed: {error}");
            exit_code = 1;
        }
    }
    pump_stop.cancel();
    let _ = dialog_task.await;
    if let Some(task) = event_task {
        let _ = task.await;
    }
    // All producers have joined; close the queue and flush every queued line.
    drop(stdout);
    match writer_task.await {
        Ok(Ok(())) => {}
        error => {
            eprintln!("[rpi] rpc output failed: {error:?}");
            exit_code = 1;
        }
    }
    exit_code
}

async fn drain_rpc_events(
    out: &RpcWriter,
    events: &mut tokio::sync::broadcast::Receiver<AgentEvent>,
) {
    loop {
        match events.try_recv() {
            Ok(event) => write_agent_event(out, &event).await,
            Err(tokio::sync::broadcast::error::TryRecvError::Lagged(_)) => continue,
            Err(_) => break,
        }
    }
}

/// Cancellation is delivered even if abort/stop arrives before prompt_text has
/// registered a run. Once active, await its normal aborted outcome and cleanup.
async fn rpc_prompt_cancellable(
    lane: &Arc<dyn AgentLane>,
    content: &str,
    cancel: tokio_util::sync::CancellationToken,
) -> Result<serde_json::Value, String> {
    let prompt = rpc_prompt_text(lane, content);
    tokio::pin!(prompt);
    tokio::select! {
        biased;
        _ = cancel.cancelled() => {
            match lane.abort().await {
                Ok(_) => prompt.await,
                Err(_) => Err("Operation aborted".to_string()),
            }
        }
        result = &mut prompt => result,
    }
}

/// Compact the branch, unless the caller cancels first.
///
/// Compaction is its own operation kind in the harness: it spends a provider
/// call on the summary, so it takes the same single-run slot and reports the
/// same settlement shape a prompt does.
async fn rpc_compact_cancellable(
    lane: &Arc<dyn AgentLane>,
    instructions: Option<String>,
    cancel: tokio_util::sync::CancellationToken,
) -> Result<serde_json::Value, String> {
    let compact = async {
        match lane.compact(instructions.as_deref()).await {
            Ok(result) => Ok(compaction_json(&result)),
            Err(error) => Err(error.to_string()),
        }
    };
    tokio::pin!(compact);
    tokio::select! {
        biased;
        _ = cancel.cancelled() => {
            let _ = lane.abort().await;
            compact.await
        }
        result = &mut compact => result,
    }
}

/// Reduce a compaction result to what a client needs: the verdict, and a line
/// a transcript can print.
fn compaction_json(result: &CompactionResult) -> serde_json::Value {
    let mut value = match &result.outcome {
        CompactionOutcome::Completed { leaf_id, entry } => {
            let (tokens_before, retained) = match entry {
                rpi_harness::session::types::Entry::Compaction(compaction) => {
                    (compaction.tokens_before, compaction.retained_tail.len())
                }
                _ => (0, 0),
            };
            serde_json::json!({
                "outcome": "compacted",
                "leafId": leaf_id,
                "tokensBefore": tokens_before,
                "retainedMessages": retained,
                "summary": format!(
                    "Conversation compacted · {} tokens before, {} messages kept.",
                    tokens_before, retained
                ),
            })
        }
        CompactionOutcome::Declined { leaf_id } => serde_json::json!({
            "outcome": "declined",
            "leafId": leaf_id,
            "summary": "Nothing to compact yet.",
        }),
        CompactionOutcome::Aborted { leaf_id } => serde_json::json!({
            "outcome": "aborted",
            "leafId": leaf_id,
            "summary": "Compaction aborted.",
        }),
        CompactionOutcome::Failed { leaf_id, error } => serde_json::json!({
            "outcome": "failed",
            "leafId": leaf_id,
            "error": error.message,
            "summary": format!("Compaction failed: {}", error.message),
        }),
    };
    value["runId"] = serde_json::json!(result.run_id);
    value
}

/// Drive the run recovery left behind, unless the caller cancels first.
///
/// `Ok(None)` from the lane means there was nothing to resume: the run is
/// reported as `skipped` so the client can settle without inventing an outcome.
async fn rpc_resume_cancellable(
    lane: &Arc<dyn AgentLane>,
    cancel: tokio_util::sync::CancellationToken,
) -> Result<serde_json::Value, String> {
    let resume = async {
        match lane.resume_pending().await {
            Ok(Some(result)) => {
                let (outcome, final_text, error) = rpc_outcome_str(&result.outcome);
                Ok(serde_json::json!({
                    "runId": result.run_id,
                    "outcome": outcome,
                    "finalText": final_text,
                    "error": error,
                    "resumed": true,
                }))
            }
            Ok(None) => Ok(serde_json::json!({"outcome": "skipped", "resumed": false})),
            Err(error) => Err(error.to_string()),
        }
    };
    tokio::pin!(resume);
    tokio::select! {
        biased;
        _ = cancel.cancelled() => {
            let _ = lane.abort().await;
            resume.await
        }
        result = &mut resume => result,
    }
}

fn rpc_outcome_str(outcome: &HarnessRunOutcome) -> (&'static str, Option<String>, Option<String>) {
    match outcome {
        HarnessRunOutcome::Completed { final_message, .. } => {
            ("completed", Some(assistant_text(final_message)), None)
        }
        HarnessRunOutcome::Aborted { final_message, .. } => {
            ("aborted", Some(assistant_text(final_message)), None)
        }
        // A failed run can end with nothing to say, which is exactly when the
        // reason matters: rate limits and provider errors live here.
        HarnessRunOutcome::Failed {
            final_message,
            error,
            ..
        } => (
            "failed",
            final_message.as_ref().map(assistant_text),
            Some(error.message.clone()),
        ),
        HarnessRunOutcome::Suspended { .. } => ("suspended", None, None),
    }
}

async fn rpc_prompt_text(
    lane: &Arc<dyn AgentLane>,
    content: &str,
) -> Result<serde_json::Value, String> {
    let result = lane
        .prompt_text(content, Vec::new())
        .await
        .map_err(|e| e.to_string())?;
    let (outcome, final_text, error) = rpc_outcome_str(&result.outcome);
    Ok(
        serde_json::json!({"runId":result.run_id, "outcome": outcome, "finalText": final_text, "error": error}),
    )
}

async fn rpc_get_state(lane: &Arc<dyn AgentLane>, harness: &AgentHarness) -> serde_json::Value {
    let model = lane.get_model().await.ok();
    let thinking = lane
        .get_thinking_level()
        .await
        .ok()
        .map(|t| serde_json::to_value(t).unwrap_or(serde_json::Value::Null));
    let active_tools = lane.get_active_tools().await.unwrap_or_default();
    let leaf_id = lane.get_leaf_id().await.ok().flatten();
    let context = rpc_context_usage(lane, harness).await;
    let active_operation = lane
        .snapshot()
        .await
        .ok()
        .and_then(|snapshot| snapshot.active)
        .map(|active| {
            serde_json::json!({
                "runId": active.run_id,
                "lane": active.lane,
                "kind": format!("{:?}", active.kind).to_lowercase(),
            })
        });
    let queued = lane.queued_messages().await.unwrap_or_default();
    let stats = rpc_get_stats(lane, harness).await.ok();
    serde_json::json!({
        "modelId": model.as_ref().map(|model| model.id.clone()),
        "provider": model.as_ref().map(|model| model.provider.clone()),
        "thinkingLevel": thinking,
        "activeTools": active_tools,
        "leafId": leaf_id,
        "context": context,
        "sessionId": harness.session().storage().metadata().id,
        "activeOperation": active_operation,
        "queue": {
            "steering": queued.steering,
            "followUp": queued.follow_up,
            "steeringCount": queued.steering.len(),
            "followUpCount": queued.follow_up.len(),
        },
        "stats": stats,
    })
}

fn rpc_now_ms() -> i64 {
    std::time::SystemTime::now()
        .duration_since(std::time::UNIX_EPOCH)
        .map(|duration| duration.as_millis() as i64)
        .unwrap_or_default()
}

fn rpc_message_text(value: &serde_json::Value) -> String {
    let Some(content) = value.get("content") else {
        return String::new();
    };
    if let Some(text) = content.as_str() {
        return text.to_string();
    }
    content
        .as_array()
        .into_iter()
        .flatten()
        .filter_map(|block| block.get("text").and_then(|text| text.as_str()))
        .collect::<Vec<_>>()
        .join("\\n")
}

async fn rpc_get_history(
    lane: &Arc<dyn AgentLane>,
    harness: &AgentHarness,
    msg: &serde_json::Value,
) -> Result<serde_json::Value, String> {
    let limit = msg
        .get("limit")
        .and_then(serde_json::Value::as_u64)
        .unwrap_or(100)
        .clamp(1, 500) as usize;
    let after_seq = msg.get("afterSeq").and_then(serde_json::Value::as_u64);
    let before_seq = msg.get("beforeSeq").and_then(serde_json::Value::as_u64);
    let mut records = harness
        .session()
        .find_records(&RecordQuery {
            lane: Some(lane.name().to_string()),
            after_seq,
            order: Some(EntryOrder::OldestFirst),
            ..Default::default()
        })
        .await
        .map_err(|error| error.to_string())?;
    if let Some(before) = before_seq {
        records.retain(|record| record.seq() < before);
    }
    let total = records.len();
    let has_more = records.len() > limit;
    if before_seq.is_some() {
        if has_more {
            records = records.split_off(records.len() - limit);
        }
    } else if after_seq.is_some() {
        records.truncate(limit);
    } else if has_more {
        records = records.split_off(records.len() - limit);
    }
    let first_seq = records.first().map(LaneRecord::seq);
    let last_seq = records.last().map(LaneRecord::seq);
    let open_operations = harness
        .session()
        .find_open_operations(lane.name(), Some(50))
        .await
        .map_err(|error| error.to_string())?;
    Ok(serde_json::json!({
        "records": records.iter().map(|record| serde_json::to_value(record).unwrap_or(serde_json::Value::Null)).collect::<Vec<_>>(),
        "total": total,
        "limit": limit,
        "afterSeq": after_seq,
        "beforeSeq": before_seq,
        "firstSeq": first_seq,
        "lastSeq": last_seq,
        "hasMore": has_more,
        "leafId": lane.get_leaf_id().await.ok().flatten(),
        "openOperations": open_operations.iter().map(|record| serde_json::to_value(record).unwrap_or(serde_json::Value::Null)).collect::<Vec<_>>(),
    }))
}

async fn rpc_get_stats(
    lane: &Arc<dyn AgentLane>,
    harness: &AgentHarness,
) -> Result<serde_json::Value, String> {
    let stats = lane.get_stats().await.map_err(|error| error.to_string())?;
    let records = harness
        .session()
        .find_records(&RecordQuery {
            lane: Some(lane.name().to_string()),
            order: Some(EntryOrder::OldestFirst),
            ..Default::default()
        })
        .await
        .map_err(|error| error.to_string())?;
    let mut operation_counts = std::collections::BTreeMap::<String, u64>::new();
    for record in &records {
        let kind = match record {
            LaneRecord::OperationStarted(started) => {
                let key = started.intent.kind();
                *operation_counts.entry(key.to_string()).or_default() += 1;
                if matches!(
                    &started.intent,
                    rpi_harness::session::types::OperationIntent::Run {
                        resume_data: Some(_),
                        ..
                    }
                ) {
                    *operation_counts.entry("resume".to_string()).or_default() += 1;
                }
                continue;
            }
            LaneRecord::OperationFinished(_) => "operation_finished",
            LaneRecord::RetryPending(_) => "retry_pending",
            LaneRecord::Usage(_) => "usage",
            _ => continue,
        };
        *operation_counts.entry(kind.to_string()).or_default() += 1;
    }
    let count = |name: &str| operation_counts.get(name).copied().unwrap_or_default();
    let model = lane.get_model().await.ok();
    let input = stats.cached_tokens + stats.uncached_tokens;
    Ok(serde_json::json!({
        "messageCount": stats.message_count,
        "operationCount": count("operation_finished"),
        "retryCount": count("retry_pending"),
        "compactionCount": count("compaction"),
        "resumeCount": count("resume"),
        "operationCounts": operation_counts,
        "provider": model.as_ref().map(|value| value.provider.clone()),
        "modelId": model.as_ref().map(|value| value.id.clone()),
        "cachedTokens": stats.cached_tokens,
        "uncachedTokens": stats.uncached_tokens,
        "totalTokens": stats.total_tokens,
        "costTotal": stats.cost_total,
        "cacheHitRatio": if input > 0 { Some(stats.cached_tokens as f64 / input as f64) } else { None::<f64> },
    }))
}

async fn rpc_get_queue(
    lane: &Arc<dyn AgentLane>,
    harness: &AgentHarness,
    msg: &serde_json::Value,
) -> Result<serde_json::Value, String> {
    let include_completed = msg
        .get("includeCompleted")
        .and_then(serde_json::Value::as_bool)
        .unwrap_or(false);
    let records = harness
        .session()
        .find_records(&RecordQuery {
            lane: Some(lane.name().to_string()),
            order: Some(EntryOrder::OldestFirst),
            ..Default::default()
        })
        .await
        .map_err(|error| error.to_string())?;
    let mut cancelled = std::collections::HashMap::<String, u64>::new();
    for record in &records {
        if let LaneRecord::QueueCancelled(record) = record {
            cancelled.insert(record.entry_id.clone(), record.base.seq);
        }
    }
    let mut items = Vec::new();
    for record in records {
        let LaneRecord::QueueEnqueued(record) = record else {
            continue;
        };
        let Some(entry_id) = record.target.get("id").and_then(|value| value.as_str()) else {
            continue;
        };
        let (status, completed_seq) = if let Some(seq) = cancelled.get(entry_id) {
            ("cancelled", Some(*seq))
        } else if let Some(entry) = harness.get_entry(entry_id).await.ok().flatten() {
            ("consumed", Some(entry.base().seq))
        } else {
            ("queued", None)
        };
        if status != "queued" && !include_completed {
            continue;
        }
        let message = record
            .target
            .get("message")
            .cloned()
            .unwrap_or(serde_json::Value::Null);
        items.push(serde_json::json!({
            "entryId": entry_id,
            "queue": record.queue,
            "status": status,
            "text": rpc_message_text(&message),
            "message": message,
            "runId": record.run_id,
            "seq": record.base.seq,
            "completedSeq": completed_seq,
        }));
    }
    Ok(
        serde_json::json!({"items": items, "count": items.len(), "includeCompleted": include_completed}),
    )
}

async fn rpc_queue_message(
    lane: &Arc<dyn AgentLane>,
    msg: &serde_json::Value,
) -> Result<serde_json::Value, String> {
    let text = msg
        .get("message")
        .and_then(serde_json::Value::as_str)
        .filter(|text| !text.trim().is_empty())
        .ok_or_else(|| "queue_message requires a non-empty string `message`".to_string())?;
    let message = rpi_agent::message::AgentMessage::User(UserMessage::new(text, rpc_now_ms()));
    let queue = msg
        .get("deliverAs")
        .and_then(serde_json::Value::as_str)
        .unwrap_or("followUp");
    let result = match queue {
        "steer" => lane.steer(message).await,
        "followUp" | "follow_up" => lane.follow_up(message).await,
        "nextTurn" | "next_run" => lane.next_run(message).await,
        other => return Err(format!("unknown queue kind: {other}")),
    }
    .map_err(|error| error.to_string())?;
    Ok(serde_json::json!({
        "accepted": true,
        "entryId": result.entry_id,
        "queue": queue,
    }))
}

async fn rpc_cancel_queued(
    lane: &Arc<dyn AgentLane>,
    msg: &serde_json::Value,
) -> Result<serde_json::Value, String> {
    let entry_id = msg
        .get("entryId")
        .and_then(serde_json::Value::as_str)
        .filter(|value| !value.is_empty())
        .ok_or_else(|| "cancel_queued requires `entryId`".to_string())?;
    let result = lane
        .cancel_queued(entry_id)
        .await
        .map_err(|error| error.to_string())?;
    Ok(serde_json::json!({
        "entryId": entry_id,
        "outcome": format!("{:?}", result.outcome).to_lowercase(),
    }))
}

async fn rpc_clear_queue(lane: &Arc<dyn AgentLane>) -> Result<serde_json::Value, String> {
    let result = lane
        .clear_queue()
        .await
        .map_err(|error| error.to_string())?;
    Ok(serde_json::json!({
        "cleared": result.steering.len() + result.follow_up.len(),
        "steering": result.steering,
        "followUp": result.follow_up,
    }))
}

async fn rpc_navigate(
    lane: &Arc<dyn AgentLane>,
    msg: &serde_json::Value,
) -> Result<serde_json::Value, String> {
    let target_id = msg
        .get("targetId")
        .and_then(serde_json::Value::as_str)
        .filter(|value| !value.is_empty());
    if lane
        .snapshot()
        .await
        .map_err(|error| error.to_string())?
        .active
        .is_some()
    {
        return Err("cannot navigate while an operation is active".to_string());
    }
    let result = lane
        .navigate_tree(target_id, false, None, None)
        .await
        .map_err(|error| error.to_string())?;
    let (outcome, leaf_id, error) = match result.outcome {
        rpi_harness::agent_harness::NavigationOutcome::Completed { new_leaf_id, .. } => {
            ("completed", new_leaf_id, None)
        }
        rpi_harness::agent_harness::NavigationOutcome::Declined { leaf_id } => {
            ("declined", leaf_id, None)
        }
        rpi_harness::agent_harness::NavigationOutcome::Aborted { leaf_id } => {
            ("aborted", leaf_id, None)
        }
        rpi_harness::agent_harness::NavigationOutcome::Failed { leaf_id, error } => {
            ("failed", leaf_id, Some(error.message))
        }
    };
    Ok(serde_json::json!({
        "runId": result.run_id,
        "outcome": outcome,
        "leafId": leaf_id,
        "error": error,
    }))
}

/// How full the model's context is, measured the way compaction measures it.
///
/// The branch is projected into model input first (that is what applies a
/// compaction boundary), so the number matches what the next provider call will
/// actually send rather than how many messages the transcript happens to hold.
async fn rpc_context_usage(lane: &Arc<dyn AgentLane>, harness: &AgentHarness) -> serde_json::Value {
    use rpi_harness::compaction::{build_session_context, estimate_context_tokens};
    use rpi_harness::session::types::{BranchBounds, EntryOrder, EntryQuery};

    let entries = harness
        .find_entries_on_branch(
            &EntryQuery {
                order: Some(EntryOrder::OldestFirst),
                ..Default::default()
            },
            &BranchBounds::default(),
        )
        .await
        .unwrap_or_default();
    let estimate = estimate_context_tokens(&build_session_context(&entries).messages);
    let window = lane
        .get_model()
        .await
        .ok()
        .map(|model| model.context_window)
        .unwrap_or(0);
    let settings = harness.compaction_settings();
    let window = window as i64;
    let threshold = window - settings.reserve_tokens;
    let percent = if window > 0 {
        (estimate.tokens as f64) * 100.0 / (window as f64)
    } else {
        0.0
    };
    serde_json::json!({
        "tokens": estimate.tokens,
        "window": window,
        "threshold": threshold,
        "percent": (percent * 10.0).round() / 10.0,
        "reserveTokens": settings.reserve_tokens,
        "keepRecentTokens": settings.keep_recent_tokens,
        "compactionEnabled": settings.enabled,
        // False once a real provider usage report backs the number; until then
        // it is a character-based estimate.
        "estimated": estimate.last_usage_index.is_none(),
        "trailingTokens": estimate.trailing_tokens,
    })
}

/// The window `get_messages` returns, expressed as a slice range plus whether an
/// older page still exists. `before` is the index of the `beforeId` entry.
fn page_window(
    total: usize,
    limit: usize,
    before: Option<usize>,
) -> (std::ops::Range<usize>, bool) {
    match before {
        Some(index) => {
            let start = index.saturating_sub(limit);
            (start..index, start > 0)
        }
        None => {
            let skip = total.saturating_sub(limit);
            (skip..total, skip > 0)
        }
    }
}

/// Read-only snapshot of the current branch, oldest first. Messages are
/// projected with the same `AgentMessage` encoding the event stream uses, so a
/// client can render a transcript without understanding session entries.
async fn rpc_get_messages(
    lane: &Arc<dyn AgentLane>,
    harness: &AgentHarness,
    msg: &serde_json::Value,
) -> Result<serde_json::Value, String> {
    use rpi_harness::session::types::{BranchBounds, Entry, EntryOrder, EntryQuery};

    const DEFAULT_LIMIT: usize = 400;
    let limit = msg
        .get("limit")
        .and_then(|value| value.as_u64())
        .map(|value| value as usize)
        .unwrap_or(DEFAULT_LIMIT)
        .max(1);
    // `beforeId` pages backwards: the returned window ends immediately before
    // the entry with that id, so a client can walk a long branch one page at a
    // time instead of always re-reading the tail.
    let before_id = msg
        .get("beforeId")
        .and_then(|value| value.as_str())
        .filter(|value| !value.is_empty())
        .map(str::to_owned);
    let entries = harness
        .find_entries_on_branch(
            &EntryQuery {
                order: Some(EntryOrder::OldestFirst),
                ..Default::default()
            },
            &BranchBounds::default(),
        )
        .await
        .map_err(|error| error.to_string())?;
    let mut items: Vec<(String, serde_json::Value)> = Vec::new();
    for entry in entries {
        match entry {
            Entry::Message(entry) => {
                let mut value = serde_json::to_value(&entry.message).map_err(|e| e.to_string())?;
                // The session entry id, not the message's own id: it is what
                // `beforeId` keys on, and what makes rows stable across reloads.
                value["entryId"] = serde_json::json!(entry.base.id);
                items.push((entry.base.id, value));
            }
            // Compaction rewrites what the model sees, not the branch, so the
            // transcript keeps its messages. Marking the boundary is what lets a
            // client explain why the context suddenly shrank.
            Entry::Compaction(compaction) => items.push((
                compaction.base.id.clone(),
                serde_json::json!({
                    "role": "compaction",
                    "entryId": compaction.base.id,
                    "tokensBefore": compaction.tokens_before,
                    "retainedMessages": compaction.retained_tail.len(),
                    "timestamp": compaction.base.timestamp,
                }),
            )),
            _ => {}
        }
    }
    let total = items.len();
    let before_index = match &before_id {
        Some(before) => match items.iter().position(|(id, _)| id == before) {
            Some(index) => Some(index),
            None => return Err(format!("beforeId not found: {before}")),
        },
        None => None,
    };
    let (range, has_more) = page_window(total, limit, before_index);
    let window = items[range].to_vec();
    let first_id = window.first().map(|(id, _)| id.clone());
    let last_id = window.last().map(|(id, _)| id.clone());
    let messages = window
        .iter()
        .map(|(_, value)| value.clone())
        .collect::<Vec<_>>();
    // Read-only branch position: the leaf the snapshot ends on, and whether
    // recovery left an interrupted run that a client may want to surface.
    let leaf_id = lane.get_leaf_id().await.ok().flatten();
    let pending_resume = lane.has_pending_resume().await;
    Ok(serde_json::json!({
        "messages": messages,
        "total": total,
        "truncated": has_more,
        "limit": limit,
        "firstId": first_id,
        "lastId": last_id,
        "hasMore": has_more,
        "beforeId": before_id,
        "leafId": leaf_id,
        "pendingResume": pending_resume,
    }))
}

fn rpc_auth_provider_ids(model_catalog: &[rpi_ai::Model]) -> std::collections::BTreeSet<String> {
    let mut providers =
        std::collections::BTreeSet::from(["anthropic".to_string(), "openai".to_string()]);
    providers.extend(model_catalog.iter().map(|model| model.provider.clone()));
    if let Ok(models) = crate::config::load_models_config() {
        providers.extend(models.providers.keys().cloned());
    }
    providers
}

fn rpc_get_auth_status(
    model_catalog: &[rpi_ai::Model],
    oauth_snapshot: Option<&Arc<rpi_extensions::RegistrySnapshot>>,
) -> serde_json::Value {
    let store = crate::config::read_auth().unwrap_or_default();
    let models = crate::config::load_models_config().unwrap_or_default();
    let manifests = oauth_snapshot.map_or_else(Vec::new, |snapshot| {
        snapshot
            .providers()
            .iter()
            .filter_map(|provider| snapshot.provider_manifest(&provider.provider_id).cloned())
            .collect()
    });
    let auth_states = crate::credentials::statuses(&models, &store, &manifests);
    let providers = rpc_auth_provider_ids(model_catalog)
        .into_iter()
        .map(|provider| {
            let credential = store.get(&provider);
            let state = auth_states.get(&provider).cloned().unwrap_or(
                rpi_runtime::credentials::AuthState {
                    status: rpi_runtime::credentials::AuthStatus::Missing,
                    source: crate::credentials::CredentialSource::Missing,
                },
            );
            let oauth = matches!(credential, Some(crate::config::Credential::Oauth { .. }));
            let oauth_available =
                oauth_snapshot.is_some_and(|snapshot| snapshot.oauth_provider(&provider).is_some());
            serde_json::json!({
                "provider": provider,
                "ready": state.is_ready(),
                "source": state.source.as_json_str(),
                "status": state.status,
                "oauth": oauth,
                "oauthAvailable": oauth_available,
            })
        })
        .collect::<Vec<_>>();
    serde_json::json!({"providers": providers})
}

fn rpc_oauth(
    snapshot: &Option<Arc<rpi_extensions::RegistrySnapshot>>,
    msg: &serde_json::Value,
) -> Result<serde_json::Value, String> {
    let snapshot = snapshot
        .as_ref()
        .ok_or_else(|| "OAuth is unavailable without an extension provider".to_string())?;
    let provider = msg
        .get("provider")
        .and_then(serde_json::Value::as_str)
        .map(str::trim)
        .filter(|value| !value.is_empty())
        .ok_or_else(|| "oauth requires a provider".to_string())?;
    let action = msg
        .get("action")
        .and_then(serde_json::Value::as_str)
        .map(str::trim)
        .filter(|value| matches!(*value, "begin" | "exchange" | "refresh" | "revoke"))
        .ok_or_else(|| "oauth requires a valid action".to_string())?;
    let input = msg
        .get("input")
        .cloned()
        .unwrap_or_else(|| serde_json::json!({}));
    let raw = rpi_extensions::request_oauth(snapshot, provider, action, input)
        .map_err(|_| "OAuth provider request failed".to_string())?;
    let response = rpi_extensions::OAuthActionResponse::parse(raw)
        .map_err(|_| "OAuth provider returned an invalid response".to_string())?;
    match response.kind {
        rpi_extensions::OAuthResponseKind::Credential => {
            let credential = response
                .credential()
                .map_err(|_| "OAuth provider returned an invalid credential".to_string())?
                .ok_or_else(|| "OAuth provider returned no credential".to_string())?;
            crate::config::upsert_oauth_token(
                provider,
                credential.access,
                credential.refresh,
                credential.expires_at,
            )
            .map_err(|_| "OAuth credentials could not be saved".to_string())?;
            Ok(serde_json::json!({
                "provider": provider,
                "status": "authenticated"
            }))
        }
        rpi_extensions::OAuthResponseKind::Interaction => Ok(serde_json::json!({
            "provider": provider,
            "status": "interaction",
            "display": oauth_display(&response.payload)
        })),
        rpi_extensions::OAuthResponseKind::Success => Ok(serde_json::json!({
            "provider": provider,
            "status": "success"
        })),
        rpi_extensions::OAuthResponseKind::Error => {
            Err("OAuth provider rejected the request".to_string())
        }
    }
}

/// Keep provider-defined OAuth UI data deliberately narrow. In particular,
/// token-like fields never cross the RPC boundary, even if a plugin includes
/// them in an interaction payload by mistake.
fn oauth_display(payload: &serde_json::Value) -> serde_json::Value {
    let display = payload.get("display").unwrap_or(payload);
    let Some(object) = display.as_object() else {
        return serde_json::json!({});
    };
    let mut safe = serde_json::Map::new();
    for key in [
        "title",
        "url",
        "instructions",
        "user_code",
        "verification_uri",
    ] {
        if let Some(value) = object.get(key) {
            safe.insert(key.to_string(), value.clone());
        }
    }
    serde_json::Value::Object(safe)
}

async fn rpc_set_api_key(msg: &serde_json::Value) -> Result<serde_json::Value, String> {
    let provider = msg
        .get("provider")
        .and_then(serde_json::Value::as_str)
        .map(str::trim)
        .filter(|value| !value.is_empty())
        .ok_or_else(|| "set_api_key requires a provider".to_string())?;
    let api_key = msg
        .get("apiKey")
        .and_then(serde_json::Value::as_str)
        .map(str::trim)
        .filter(|value| !value.is_empty())
        .ok_or_else(|| "set_api_key requires a non-empty apiKey".to_string())?;
    crate::config::upsert_credential(
        provider,
        crate::config::Credential::ApiKey {
            key: Some(api_key.to_string()),
            env: None,
        },
    )
    .map_err(|error| error.to_string())?;
    Ok(serde_json::json!({
        "provider": provider,
        "ready": true,
        "source": "auth.json",
        "requiresRestart": true,
    }))
}

fn rpc_get_available_models(catalog: &[rpi_ai::Model]) -> serde_json::Value {
    let models = catalog
        .iter()
        .map(crate::rpc::protocol::model_descriptor)
        .collect::<Vec<_>>();
    serde_json::json!({"models": models})
}

async fn rpc_get_available_thinking_levels(lane: &Arc<dyn AgentLane>) -> serde_json::Value {
    let levels = lane
        .get_model()
        .await
        .map(|model| model.supported_thinking_levels())
        .unwrap_or_default();
    serde_json::json!({"levels": levels})
}

async fn rpc_get_available_tools(
    lane: &Arc<dyn AgentLane>,
    harness: &AgentHarness,
) -> serde_json::Value {
    let active = lane.get_active_tools().await.unwrap_or_default();
    let tools = harness
        .get_tools()
        .await
        .map(|tools| {
            tools
                .iter()
                .map(|tool| tool.tool.schema().name.clone())
                .collect::<Vec<_>>()
        })
        .unwrap_or_else(|_| active.clone());
    serde_json::json!({"tools": tools, "activeTools": active})
}

async fn rpc_set_model(
    lane: &Arc<dyn AgentLane>,
    catalog: &[rpi_ai::Model],
    msg: &serde_json::Value,
) -> Result<serde_json::Value, String> {
    let id = msg
        .get("modelId")
        .or_else(|| msg.get("model"))
        .and_then(|v| v.as_str())
        .ok_or("set_model requires `modelId`")?;
    let provider = msg.get("provider").and_then(|v| v.as_str());
    let model = catalog
        .iter()
        .find(|m| {
            m.id.eq_ignore_ascii_case(id)
                && provider.is_none_or(|provider| m.provider.eq_ignore_ascii_case(provider))
        })
        .cloned()
        .ok_or_else(|| format!("unknown model: {id}"))?;
    lane.set_model(model).await.map_err(|e| e.to_string())?;
    Ok(serde_json::json!({"modelId": id, "provider": provider}))
}

async fn rpc_set_thinking_level(
    lane: &Arc<dyn AgentLane>,
    msg: &serde_json::Value,
) -> Result<serde_json::Value, String> {
    let value = msg
        .get("level")
        .or_else(|| msg.get("thinkingLevel"))
        .ok_or("set_thinking_level requires `level`")?;
    let level: rpi_ai::types::ThinkingLevel =
        serde_json::from_value(value.clone()).map_err(|_| {
            "invalid thinking level (off|minimal|low|medium|high|xhigh|max)".to_string()
        })?;
    lane.set_thinking_level(level)
        .await
        .map_err(|e| e.to_string())?;
    Ok(serde_json::json!({"thinkingLevel": level}))
}

async fn rpc_set_active_tools(
    lane: &Arc<dyn AgentLane>,
    msg: &serde_json::Value,
) -> Result<serde_json::Value, String> {
    let tools = msg
        .get("tools")
        .and_then(|v| v.as_array())
        .map(|arr| {
            arr.iter()
                .filter_map(|v| v.as_str().map(str::to_owned))
                .collect::<Vec<_>>()
        })
        .ok_or("set_active_tools requires a `tools` array")?;
    lane.set_active_tools(tools.clone())
        .await
        .map_err(|e| e.to_string())?;
    Ok(serde_json::json!({"activeTools": tools}))
}

/// Stable JSON projection for the fine-grained agent lifecycle stream.
///
/// Delegates to the shared [`crate::remote::protocol::RemoteEvent`] so the
/// server and the `--connect` client agree on the wire shape by construction.
fn agent_event_json(event: &AgentEvent) -> serde_json::Value {
    crate::rpc::events::project(event)
}

/// Emit a single harness event as a JSON line on stdout. Mirrors the TS
/// `toJsonEvent` projection (here a lossy but stable shape: `type` + the event
/// payload's key fields).
fn emit_json_event(event: &HarnessEvent) {
    let line = match event {
        HarnessEvent::CompactionProgress(e) => {
            serde_json::json!({ "type": "compaction_progress", "lane": e.lane, "manual": e.manual, "active": e.active })
        }
        HarnessEvent::RunStart(e) => serde_json::json!({
            "type": "run_start",
            "lane": e.lane,
            "runId": e.run_id,
        }),
        HarnessEvent::RunEnd(e) => serde_json::json!({
            "type": "run_end",
            "lane": e.lane,
            "runId": e.run_id,
            "outcome": run_end_outcome_str(e.outcome),
            "leafId": e.leaf_id,
        }),
    };
    println!("{line}");
}

fn run_end_outcome_str(o: RunEndOutcome) -> &'static str {
    match o {
        RunEndOutcome::Completed => "completed",
        RunEndOutcome::Aborted => "aborted",
        RunEndOutcome::Failed => "failed",
    }
}

/// `interactive` mode: uses TUI if terminal supports it, falls back to minimal REPL.
///
/// `event_rx` carries the live `AgentEvent` stream (drained by the TUI to
/// render streaming responses). The REPL fallback ignores it.
///
/// `model_catalog` is the resolved provider's full model list, passed through
/// so the TUI's `/model` selector can display available models (read-only —
/// v1 does not switch models mid-session; see the initial port notes (retired)).
pub async fn interactive(
    harness: &AgentHarness,
    event_rx: Option<tokio::sync::broadcast::Receiver<rpi_agent::AgentEvent>>,
    args: &Args,
    model_catalog: Vec<rpi_ai::Model>,
    initial: Option<String>,
    extra_messages: &[String],
    initial_images: Vec<ImageContent>,
    theme: Option<&str>,
    no_themes: bool,
    reload_context: &crate::session::ReloadContext,
) -> i32 {
    // ---- Session lifecycle (P1): BeforeTuiStart ----
    // `BeforeTuiStart`: extensions may prepare (connect a server, load a
    // resource) before the interactive UI initializes. Runs with a per-event
    // timeout; a handler returning `EVENT_HANDLER_ABORT` **vetoes** startup —
    // we print the reason and exit without entering the TUI. No-op when no
    // extension subscribes.
    if let Some(reason) = crate::session::dispatch_session_event_async(
        reload_context,
        rpi_plugin_sdk::EventTag::BeforeTuiStart,
    )
    .await
    {
        eprintln!("[rpi] startup vetoed by extension: {reason}");
        return crate::app::EXIT_VETOED;
    }

    // Check if TUI is supported
    let force_tui = std::env::var("RPI_FORCE_TUI")
        .map(|v| v == "1")
        .unwrap_or(false);
    let code = if force_tui || crate::interactive_tui::is_tui_supported() {
        // Use TUI-based interactive mode
        crate::interactive_tui::interactive_tui(
            harness,
            event_rx,
            args,
            model_catalog,
            initial,
            extra_messages,
            initial_images,
            theme,
            no_themes,
            reload_context,
        )
        .await
    } else {
        // Fall back to simple REPL
        interactive_repl(harness, args, initial, extra_messages, initial_images).await
    };

    // `SessionShutdown`: the interactive session closed (TUI/REPL returned) —
    // extensions release resources, disconnect, persist state. Fired after the
    // UI returns so it always runs (even on early-return paths inside the TUI).
    // A veto here is **advisory** (the session is already closing) — warn and
    // continue.
    if let Some(reason) = crate::session::dispatch_session_event_async(
        reload_context,
        rpi_plugin_sdk::EventTag::SessionShutdown,
    )
    .await
    {
        tracing::warn!(
            "[rpi] extension vetoed SessionShutdown (ignored, session closing): {reason}"
        );
    }

    code
}

/// Simple REPL-based interactive mode (fallback for non-TTY environments).
pub async fn interactive_repl(
    harness: &AgentHarness,
    #[allow(unused_variables)] args: &Args,
    initial: Option<String>,
    extra_messages: &[String],
    initial_images: Vec<ImageContent>,
) -> i32 {
    // Debug: confirm we entered REPL mode
    let lane: Arc<dyn AgentLane> = harness.lane("main");
    let stdin = std::io::stdin();
    let is_tty = stdin.is_terminal();

    if is_tty {
        println!(
            "rpi interactive (v1 minimal REPL). Type /exit to quit, /abort to cancel a run.\n"
        );
    }

    // Run the initial prompt + extra messages first (same as print mode).
    let mut prompts: Vec<String> = Vec::new();
    if let Some(init) = initial {
        prompts.push(init);
    }
    for m in extra_messages {
        prompts.push(m.clone());
    }
    let mut images = initial_images;
    for prompt in prompts {
        if let Err(code) = run_one(&lane, &prompt, std::mem::take(&mut images)).await {
            return code;
        }
    }

    // Then read lines from stdin until EOF / `/exit`.
    let mut line = String::new();
    loop {
        if is_tty {
            print!("> ");
            let _ = std::io::stdout().flush();
        }
        line.clear();
        match stdin.lock().read_line(&mut line) {
            Ok(0) => break, // EOF
            Ok(_) => {}
            Err(_) => break,
        }
        let trimmed = line.trim();
        if trimmed.is_empty() {
            continue;
        }
        if trimmed == "/exit" || trimmed == "/quit" {
            break;
        }
        if trimmed == "/abort" {
            let _ = lane.abort().await;
            eprintln!("(aborted)");
            continue;
        }
        if let Err(code) = run_one(&lane, trimmed, Vec::new()).await {
            return code;
        }
    }
    0
}

/// Run a single prompt in interactive mode, printing the assistant reply (or
/// the error). Returns `Ok(())` on success/soft-failure, `Err(exit_code)` on a
/// hard rejection.
async fn run_one(
    lane: &Arc<dyn AgentLane>,
    prompt: &str,
    images: Vec<ImageContent>,
) -> Result<(), i32> {
    match lane.prompt_text(prompt, images).await {
        Ok(result) => {
            match &result.outcome {
                HarnessRunOutcome::Completed { final_message, .. }
                | HarnessRunOutcome::Aborted { final_message, .. } => {
                    let text = assistant_text(final_message);
                    if !text.is_empty() {
                        println!("{text}");
                    }
                }
                HarnessRunOutcome::Failed {
                    error,
                    final_message,
                    ..
                } => {
                    if let Some(m) = final_message {
                        if let Some(em) = &m.error_message {
                            eprintln!("error: {em}");
                        }
                    }
                    eprintln!("run failed: {error:?}");
                }
                HarnessRunOutcome::Suspended { .. } => {
                    eprintln!("run suspended (deferred) — resume not supported in v1");
                }
            }
            Ok(())
        }
        Err(e) => {
            eprintln!("prompt rejected: {e}");
            Err(1)
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use rpi_agent::events::AgentEvent;
    use rpi_agent::types::AgentToolResult;
    use rpi_ai::types::{
        AssistantMessage, Content, StopReason, TextContent, TextContentType, Usage,
    };
    use rpi_harness::session::types::OperationError;

    #[test]
    fn rpc_auth_status_shares_header_readiness_and_expired_token_state() {
        let _guard = crate::config::test_support::env_lock().lock().unwrap();
        struct RestoreConfig(Option<std::ffi::OsString>);
        impl Drop for RestoreConfig {
            fn drop(&mut self) {
                match self.0.take() {
                    Some(value) => std::env::set_var(crate::config::CONFIG_DIR_ENV, value),
                    None => std::env::remove_var(crate::config::CONFIG_DIR_ENV),
                }
            }
        }
        let _restore = RestoreConfig(std::env::var_os(crate::config::CONFIG_DIR_ENV));
        let directory = tempfile::tempdir().unwrap();
        std::env::set_var(crate::config::CONFIG_DIR_ENV, directory.path());
        std::fs::write(directory.path().join("models.json"), serde_json::json!({"providers": {
            "header-gateway": {"api":"openai-completions", "headers":{"Authorization":"Bearer header-secret"}, "models":[{"id":"model"}]},
            "expired-account": {"api":"openai-completions", "models":[]}
        }}).to_string()).unwrap();
        std::fs::write(directory.path().join("auth.json"), serde_json::json!({
            "expired-account": {"type":"oauth", "access":"expired-secret", "refresh":"refresh-secret", "expires":1}
        }).to_string()).unwrap();
        let status = rpc_get_auth_status(&[], None);
        let providers = status["providers"].as_array().unwrap();
        let header = providers
            .iter()
            .find(|row| row["provider"] == "header-gateway")
            .unwrap();
        assert_eq!(header["ready"], true);
        assert_eq!(header["source"], "headers");
        assert_eq!(header["status"], "ready");
        let expired = providers
            .iter()
            .find(|row| row["provider"] == "expired-account")
            .unwrap();
        assert_eq!(expired["ready"], false);
        assert_eq!(expired["status"], "expired");
        assert!(!status.to_string().contains("secret"));
    }

    #[test]
    fn oauth_display_drops_token_like_fields() {
        let display = oauth_display(&serde_json::json!({
            "display": {
                "title": "Sign in",
                "url": "https://example.test/authorize",
                "instructions": ["Paste the code"],
                "access_token": "secret-access",
                "refresh_token": "secret-refresh",
            }
        }));
        assert_eq!(display["title"], "Sign in");
        assert!(display.get("access_token").is_none());
        assert!(display.get("refresh_token").is_none());
    }

    #[derive(Clone, Default)]
    struct RpcOutput(Arc<Mutex<Vec<u8>>>);

    impl Write for RpcOutput {
        fn write(&mut self, bytes: &[u8]) -> std::io::Result<usize> {
            self.0.lock().unwrap().extend_from_slice(bytes);
            Ok(bytes.len())
        }
        fn flush(&mut self) -> std::io::Result<()> {
            Ok(())
        }
    }

    impl RpcOutput {
        fn messages(&self) -> Vec<serde_json::Value> {
            String::from_utf8(self.0.lock().unwrap().clone())
                .unwrap()
                .lines()
                .map(|line| serde_json::from_str(line).unwrap())
                .collect()
        }

        async fn wait_for(
            &self,
            predicate: impl Fn(&serde_json::Value) -> bool,
        ) -> serde_json::Value {
            tokio::time::timeout(std::time::Duration::from_secs(3), async {
                loop {
                    if let Some(message) = self.messages().into_iter().find(&predicate) {
                        return message;
                    }
                    tokio::time::sleep(std::time::Duration::from_millis(1)).await;
                }
            })
            .await
            .expect("RPC message not received")
        }
    }

    async fn rpc_test_harness(
        text: &str,
        tokens_per_second: f64,
    ) -> (AgentHarness, Arc<rpi_ai::providers::faux::FauxProvider>) {
        let provider = rpi_ai::providers::faux::FauxProvider::new(
            rpi_ai::providers::faux::FauxScript::new()
                .with_text(text)
                .with_tokens_per_second(tokens_per_second),
        );
        let options = rpi_harness::types::AgentHarnessOptions {
            model: provider.default_model().clone(),
            models: vec![provider.clone()],
            ..Default::default()
        };
        (AgentHarness::create(options).await.unwrap(), provider)
    }

    fn start_test_rpc(
        harness: AgentHarness,
        ui: rpi_extensions::UiDialogMailbox,
    ) -> (
        tokio::io::DuplexStream,
        RpcOutput,
        tokio::task::JoinHandle<i32>,
    ) {
        let (client, input) = tokio::io::duplex(4096);
        let output = RpcOutput::default();
        let captured = output.clone();
        let task = tokio::spawn(async move {
            rpc_with_io(
                &harness,
                None,
                vec![],
                ui,
                None,
                tokio::io::BufReader::new(input),
                captured,
            )
            .await
        });
        (client, output, task)
    }

    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
    async fn rpc_eof_waits_for_accepted_prompt() {
        use tokio::io::AsyncWriteExt;
        let (harness, _) = rpc_test_harness("completed", 100.0).await;
        let ui = rpi_extensions::UiDialogMailbox::new();
        let (mut client, output, task) = start_test_rpc(harness, ui.clone());
        client
            .write_all(b"{\"type\":\"prompt\",\"id\":1,\"message\":\"test\"}\n")
            .await
            .unwrap();
        client.shutdown().await.unwrap();
        assert_eq!(
            tokio::time::timeout(std::time::Duration::from_secs(3), task)
                .await
                .unwrap()
                .unwrap(),
            0
        );
        let response = output
            .messages()
            .into_iter()
            .find(|m| m["id"] == 1)
            .unwrap();
        assert_eq!(response["command"], "prompt");
        assert_eq!(response["data"]["accepted"], true);
        assert!(output.messages().iter().any(|message| {
            message["type"] == "agent_settled" && message["outcome"] == "completed"
        }));
        assert!(!ui.is_attached());
    }

    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
    async fn rpc_compaction_settlement_identifies_its_durable_operation() {
        let (harness, _) = rpc_test_harness("unused", 0.0).await;
        let output = RpcOutput::default();
        let code = rpc_with_io(
            &harness,
            None,
            vec![],
            rpi_extensions::UiDialogMailbox::new(),
            None,
            tokio::io::BufReader::new(&b"{\"type\":\"compact\",\"id\":1}\n"[..]),
            output.clone(),
        )
        .await;
        assert_eq!(code, 0);
        let messages = output.messages();
        let settled = messages
            .iter()
            .find(|message| message["type"] == "agent_settled")
            .unwrap();
        assert_eq!(settled["outcome"], "declined");
        let run_id = settled["runId"]
            .as_str()
            .expect("compaction settlement needs its durable run id");
        let records = harness
            .session()
            .find_records(&Default::default())
            .await
            .unwrap();
        assert!(records.iter().any(|record| matches!(record,
            rpi_harness::session::types::LaneRecord::OperationFinished(finished)
                if finished.run_id == run_id)));
    }

    #[tokio::test]
    async fn rpc_eof_drains_queued_final_events() {
        let harness = AgentHarness::create(Default::default()).await.unwrap();
        let (tx, rx) = tokio::sync::broadcast::channel(256);
        for _ in 0..100 {
            tx.send(AgentEvent::AgentStart).unwrap();
        }
        tx.send(AgentEvent::AgentEnd { messages: vec![] }).unwrap();
        let output = RpcOutput::default();
        let code = rpc_with_io(
            &harness,
            Some(rx),
            vec![],
            rpi_extensions::UiDialogMailbox::new(),
            None,
            tokio::io::BufReader::new(&b""[..]),
            output.clone(),
        )
        .await;
        assert_eq!(code, 0);
        let messages = output.messages();
        assert_eq!(
            messages.len(),
            102,
            "ready plus every queued event must be flushed"
        );
        assert_eq!(messages.last().unwrap()["type"], "agent_end");
    }

    struct BackpressuredOutput {
        entered: Arc<std::sync::atomic::AtomicBool>,
        release: Arc<std::sync::atomic::AtomicBool>,
        output: RpcOutput,
    }

    impl Write for BackpressuredOutput {
        fn write(&mut self, bytes: &[u8]) -> std::io::Result<usize> {
            self.entered
                .store(true, std::sync::atomic::Ordering::SeqCst);
            while !self.release.load(std::sync::atomic::Ordering::SeqCst) {
                std::thread::sleep(std::time::Duration::from_millis(1));
            }
            self.output.write(bytes)
        }
        fn flush(&mut self) -> std::io::Result<()> {
            Ok(())
        }
    }

    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
    async fn rpc_output_backpressure_is_bounded_and_keeps_workers_responsive() {
        use std::sync::atomic::{AtomicBool, Ordering};
        let release = Arc::new(AtomicBool::new(false));
        struct ReleaseOnDrop(Arc<AtomicBool>);
        impl Drop for ReleaseOnDrop {
            fn drop(&mut self) {
                self.0.store(true, Ordering::SeqCst);
            }
        }
        let _release = ReleaseOnDrop(release.clone());
        let entered = Arc::new(AtomicBool::new(false));
        let output = RpcOutput::default();
        let (writer, task, _) = start_rpc_writer(BackpressuredOutput {
            entered: entered.clone(),
            release: release.clone(),
            output: output.clone(),
        });
        write_line(&writer, &serde_json::json!({"id": 0})).await;
        tokio::time::timeout(std::time::Duration::from_secs(2), async {
            while !entered.load(Ordering::SeqCst) {
                tokio::task::yield_now().await;
            }
        })
        .await
        .unwrap();
        let producer_writer = writer.clone();
        let producer = tokio::spawn(async move {
            for id in 1..=100 {
                write_line(&producer_writer, &serde_json::json!({"id": id})).await;
            }
        });
        tokio::time::timeout(
            std::time::Duration::from_millis(150),
            tokio::time::sleep(std::time::Duration::from_millis(20)),
        )
        .await
        .unwrap();
        assert!(
            !producer.is_finished(),
            "the bounded queue must apply backpressure"
        );
        release.store(true, Ordering::SeqCst);
        producer.await.unwrap();
        drop(writer);
        task.await.unwrap().unwrap();
        let messages = output.messages();
        assert_eq!(messages.len(), 101);
        for (index, message) in messages.iter().enumerate() {
            assert_eq!(message["id"], index);
        }
    }

    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
    async fn rpc_completion_response_follows_final_event_under_backpressure() {
        use std::sync::atomic::{AtomicBool, Ordering};
        let (harness, _) = rpc_test_harness("completed", 0.0).await;
        let (emitter, events) = rpi_agent::events::BroadcastEmitter::new(256);
        harness
            .set_agent_emitter(Some(Arc::new(emitter)))
            .await
            .unwrap();
        let output = RpcOutput::default();
        let release = Arc::new(AtomicBool::new(false));
        let release_writer = release.clone();
        let release_thread = std::thread::spawn(move || {
            std::thread::sleep(std::time::Duration::from_millis(50));
            release_writer.store(true, Ordering::SeqCst);
        });
        let code = rpc_with_io(
            &harness,
            Some(events),
            vec![],
            rpi_extensions::UiDialogMailbox::new(),
            None,
            tokio::io::BufReader::new(
                &b"{\"type\":\"prompt\",\"id\":1,\"message\":\"test\"}\n"[..],
            ),
            BackpressuredOutput {
                entered: Arc::new(AtomicBool::new(false)),
                release,
                output: output.clone(),
            },
        )
        .await;
        release_thread.join().unwrap();
        assert_eq!(code, 0);
        let messages = output.messages();
        let end = messages
            .iter()
            .position(|message| message["type"] == "agent_end")
            .unwrap();
        let settled = messages
            .iter()
            .position(|message| message["type"] == "agent_settled")
            .unwrap();
        assert!(end < settled);
        let run_id = messages[settled]["runId"].as_str().unwrap();
        assert!(!run_id.is_empty(), "settlement identifies its durable run");
        let records = harness
            .session()
            .find_records(&Default::default())
            .await
            .unwrap();
        assert!(records.iter().any(|record| matches!(record,
            rpi_harness::session::types::LaneRecord::OperationFinished(finished)
                if finished.run_id == run_id)));
        assert!(messages.iter().any(|message| {
            message["type"] == "response" && message["data"]["accepted"] == true
        }));
        assert!(messages.iter().any(|message| {
            message["type"] == "agent_settled" && message["outcome"] == "completed"
        }));
    }

    #[tokio::test]
    async fn rpc_output_failure_closes_without_waiting_for_more_input() {
        struct BrokenOutput;
        impl Write for BrokenOutput {
            fn write(&mut self, _: &[u8]) -> std::io::Result<usize> {
                Err(std::io::Error::new(
                    std::io::ErrorKind::BrokenPipe,
                    "closed client output",
                ))
            }
            fn flush(&mut self) -> std::io::Result<()> {
                Ok(())
            }
        }
        let harness = AgentHarness::create(Default::default()).await.unwrap();
        let (_client, input) = tokio::io::duplex(4096);
        let ui = rpi_extensions::UiDialogMailbox::new();
        let code = tokio::time::timeout(
            std::time::Duration::from_secs(2),
            rpc_with_io(
                &harness,
                None,
                vec![],
                ui.clone(),
                None,
                tokio::io::BufReader::new(input),
                BrokenOutput,
            ),
        )
        .await
        .unwrap();
        assert_eq!(code, 1);
        assert!(!ui.is_attached());
    }

    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
    async fn rpc_stop_aborts_and_joins_active_prompt() {
        use tokio::io::AsyncWriteExt;
        let (harness, provider) = rpc_test_harness("still streaming the response", 20.0).await;
        let lane = harness.lane("main");
        let (mut client, output, task) =
            start_test_rpc(harness, rpi_extensions::UiDialogMailbox::new());
        client
            .write_all(b"{\"type\":\"prompt\",\"id\":1,\"message\":\"test\"}\n")
            .await
            .unwrap();
        tokio::time::timeout(std::time::Duration::from_secs(3), async {
            while provider
                .state()
                .call_count
                .load(std::sync::atomic::Ordering::SeqCst)
                == 0
            {
                tokio::task::yield_now().await;
            }
        })
        .await
        .unwrap();
        client
            .write_all(b"{\"type\":\"stop\",\"id\":2}\n")
            .await
            .unwrap();
        assert_eq!(
            tokio::time::timeout(std::time::Duration::from_secs(3), task)
                .await
                .unwrap()
                .unwrap(),
            0
        );
        let messages = output.messages();
        assert!(messages.iter().any(|message| {
            message["type"] == "agent_settled"
                && (message["outcome"] == "aborted" || message["outcome"] == "failed")
        }));
        assert_eq!(
            messages.iter().find(|m| m["id"] == 2).unwrap()["data"]["stopped"],
            true
        );
        assert!(
            lane.abort().await.is_err(),
            "RPC returned with an active lane"
        );
    }

    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
    async fn rpc_immediate_stop_cancels_prompt_before_or_after_lane_start() {
        use tokio::io::AsyncWriteExt;
        let (harness, _) = rpc_test_harness("not completed", 20.0).await;
        let lane = harness.lane("main");
        let (mut client, output, task) =
            start_test_rpc(harness, rpi_extensions::UiDialogMailbox::new());
        client.write_all(b"{\"type\":\"prompt\",\"id\":1,\"message\":\"test\"}\n{\"type\":\"stop\",\"id\":2}\n").await.unwrap();
        assert_eq!(
            tokio::time::timeout(std::time::Duration::from_secs(3), task)
                .await
                .unwrap()
                .unwrap(),
            0
        );
        let messages = output.messages();
        assert!(messages.iter().any(|message| {
            message["type"] == "agent_settled"
                && (message["outcome"] == "aborted" || message["outcome"] == "failed")
        }));
        assert!(lane.abort().await.is_err());
    }

    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
    async fn rpc_dialog_answers_and_ping_are_read_while_prompt_runs() {
        use tokio::io::AsyncWriteExt;
        let (harness, _) =
            rpc_test_harness("the streamed response finishes after the dialog", 30.0).await;
        let ui = rpi_extensions::UiDialogMailbox::new();
        let (mut client, output, task) = start_test_rpc(harness, ui.clone());
        output.wait_for(|m| m["type"] == "ready").await;
        client
            .write_all(b"{\"type\":\"prompt\",\"id\":1,\"message\":\"test\"}\n")
            .await
            .unwrap();
        ui.handle(serde_json::json!({"requestId":"approval", "toolCallId":"call", "ui":{}}))
            .unwrap();
        output.wait_for(|m| m["type"] == "ui_dialog_request").await;
        client.write_all(b"{\"type\":\"ui_dialog_response\",\"requestId\":\"approval\",\"answer\":true}\n{\"type\":\"ping\",\"id\":2}\n").await.unwrap();
        output.wait_for(|m| m["id"] == 2).await;
        assert_eq!(ui.poll("approval").unwrap()["answer"], true);
        output
            .wait_for(|m| m["type"] == "ui_dialog_closed" && m["requestId"] == "approval")
            .await;
        client.shutdown().await.unwrap();
        assert_eq!(
            tokio::time::timeout(std::time::Duration::from_secs(3), task)
                .await
                .unwrap()
                .unwrap(),
            0
        );
        assert!(output.messages().iter().any(|message| {
            message["type"] == "agent_settled" && message["outcome"] == "completed"
        }));
    }

    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
    async fn rpc_notifies_client_when_an_approval_is_cancelled() {
        use tokio::io::AsyncWriteExt;
        let harness = AgentHarness::create(Default::default()).await.unwrap();
        let ui = rpi_extensions::UiDialogMailbox::new();
        let (mut client, output, task) = start_test_rpc(harness, ui.clone());
        output.wait_for(|m| m["type"] == "ready").await;
        ui.handle(serde_json::json!({"requestId":"cancelled", "toolCallId":"c", "ui":{}}))
            .unwrap();
        output.wait_for(|m| m["type"] == "ui_dialog_request").await;
        ui.cancel_tool_call("c");
        output
            .wait_for(|m| m["type"] == "ui_dialog_closed" && m["requestId"] == "cancelled")
            .await;
        // The notification must not consume the cancellation meant for the plugin.
        assert_eq!(ui.poll("cancelled").unwrap()["status"], "cancelled");
        client.shutdown().await.unwrap();
        assert_eq!(task.await.unwrap(), 0);
    }

    #[tokio::test]
    async fn rpc_accepts_native_dialog_responses_without_consuming_plugin_answers() {
        use crate::remote::protocol::RemoteCommand;
        use tokio::io::AsyncWriteExt;
        let harness = AgentHarness::create(Default::default()).await.unwrap();
        let ui = rpi_extensions::UiDialogMailbox::new();
        let (mut client, output, task) = start_test_rpc(harness, ui.clone());
        output.wait_for(|m| m["type"] == "ready").await;
        for (request_id, answer, response_id) in [
            (
                "answered",
                Some(serde_json::json!({"values":["No"],"kind":"selector"})),
                5,
            ),
            ("cancelled", None, 6),
        ] {
            ui.handle(serde_json::json!({"requestId":request_id,"toolCallId":request_id,"ui":{}}))
                .unwrap();
            output
                .wait_for(|m| m["type"] == "ui_dialog_request" && m["requestId"] == request_id)
                .await;
            let command = RemoteCommand::UiDialogResponse {
                request_id: request_id.into(),
                cancelled: answer.is_none(),
                answer,
            }
            .with_id(response_id);
            client
                .write_all(format!("{command}\n").as_bytes())
                .await
                .unwrap();
            let ack = output.wait_for(|m| m["id"] == response_id).await;
            assert_eq!(ack["data"]["accepted"], true);
            let result = ui.poll(request_id).unwrap();
            if request_id == "answered" {
                assert_eq!(result["answer"]["values"][0], "No");
            } else {
                assert_eq!(result["status"], "cancelled");
            }
        }
        client.shutdown().await.unwrap();
        assert_eq!(task.await.unwrap(), 0);
    }

    fn assistant(text: &str, stop: StopReason) -> AssistantMessage {
        AssistantMessage {
            role: rpi_ai::types::AssistantRole,
            content: vec![Content::Text(TextContent {
                kind: TextContentType,
                text: text.into(),
                text_signature: None,
            })],
            api: rpi_ai::Api::AnthropicMessages,
            provider: "anthropic".into(),
            model: "claude-sonnet-5".into(),
            response_model: None,
            response_id: None,
            usage: Usage::zero(),
            stop_reason: stop,
            deferred: None,
            error_message: None,
            raw_stop_reason: None,
            end_turn: None,
            timestamp: 0,
        }
    }

    #[test]
    fn assistant_text_concatenates_text_blocks() {
        let m = assistant("hello", StopReason::Stop);
        assert_eq!(assistant_text(&m), "hello");
    }

    #[test]
    fn outcome_exit_code_maps_failed_aborted_to_1() {
        let failed = HarnessRunOutcome::Failed {
            leaf_id: "l".into(),
            error: OperationError {
                code: "boom".into(),
                message: "boom".into(),
            },
            final_entry_id: None,
            final_message: None,
        };
        assert_eq!(outcome_exit_code(&failed), 1);
        let completed = HarnessRunOutcome::Completed {
            leaf_id: "l".into(),
            final_entry_id: "e".into(),
            final_message: assistant("ok", StopReason::Stop),
        };
        assert_eq!(outcome_exit_code(&completed), 0);
    }

    #[test]
    fn run_end_outcome_str_roundtrip() {
        assert_eq!(run_end_outcome_str(RunEndOutcome::Completed), "completed");
        assert_eq!(run_end_outcome_str(RunEndOutcome::Aborted), "aborted");
        assert_eq!(run_end_outcome_str(RunEndOutcome::Failed), "failed");
    }

    #[test]
    fn agent_event_projection_keeps_terminal_and_tool_payloads() {
        let end = agent_event_json(&AgentEvent::AgentEnd { messages: vec![] });
        assert_eq!(end["type"], "agent_end");
        assert_eq!(end["messages"], serde_json::json!([]));

        let tool = agent_event_json(&AgentEvent::ToolExecutionEnd {
            tool_call_id: "call-1".into(),
            tool_name: "read".into(),
            result: AgentToolResult::text("hello"),
            is_error: false,
        });
        assert_eq!(tool["type"], "tool_execution_end");
        assert_eq!(tool["result"]["content"][0]["text"], "hello");
        assert_eq!(tool["result"]["terminate"], false);

        let retry = agent_event_json(&AgentEvent::RetryScheduled {
            attempt: 3,
            max_retries: 10,
            delay_ms: 8_000,
            error: "503 service unavailable".into(),
        });
        assert_eq!(retry["type"], "retry_scheduled");
        assert_eq!(retry["attempt"], 3);
        assert_eq!(retry["maxRetries"], 10);
        assert_eq!(retry["delayMs"], 8_000);
        assert_eq!(retry["error"], "503 service unavailable");
    }
}

#[cfg(test)]
mod page_window_tests {
    use super::page_window;

    #[test]
    fn tail_is_kept_without_before_id() {
        // 12 items, limit 5 -> keep the last 5, more behind.
        assert_eq!(page_window(12, 5, None), (7..12, true));
        // Fewer items than the limit -> everything, nothing behind.
        assert_eq!(page_window(3, 5, None), (0..3, false));
    }

    #[test]
    fn before_id_opens_the_window_immediately_before_it() {
        // Page back from index 10 with limit 4 -> [6,10), more behind.
        assert_eq!(page_window(12, 4, Some(10)), (6..10, true));
        // Same, but reaching the very start -> no more behind.
        assert_eq!(page_window(12, 10, Some(10)), (0..10, false));
    }

    #[test]
    fn limit_larger_than_the_branch_never_underflows() {
        assert_eq!(page_window(5, 100, Some(2)), (0..2, false));
        assert_eq!(page_window(5, 100, None), (0..5, false));
    }
}