storkit/server/src/matrix/bot.rs

use crate::llm::providers::claude_code::{ClaudeCodeProvider, ClaudeCodeResult};
use crate::slog;
use matrix_sdk::{
    Client,
    config::SyncSettings,
    event_handler::Ctx,
    room::Room,
    ruma::{
        OwnedRoomId, OwnedUserId,
        events::room::message::{
            MessageType, OriginalSyncRoomMessageEvent, RoomMessageEventContent,
        },
    },
};
use std::path::{Path, PathBuf};
use std::sync::Arc;
use tokio::sync::watch;

use super::config::BotConfig;

/// Shared context injected into Matrix event handlers.
#[derive(Clone)]
pub struct BotContext {
    pub bot_user_id: OwnedUserId,
    pub target_room_id: OwnedRoomId,
    pub project_root: PathBuf,
}

/// Connect to the Matrix homeserver, join the configured room, and start
/// listening for messages. Runs the full Matrix sync loop — call from a
/// `tokio::spawn` task so it doesn't block the main thread.
pub async fn run_bot(config: BotConfig, project_root: PathBuf) -> Result<(), String> {
    let store_path = project_root.join(".story_kit").join("matrix_store");
    let client = Client::builder()
        .homeserver_url(&config.homeserver)
        .sqlite_store(&store_path, None)
        .build()
        .await
        .map_err(|e| format!("Failed to build Matrix client: {e}"))?;

    // Login
    client
        .matrix_auth()
        .login_username(&config.username, &config.password)
        .initial_device_display_name("Story Kit Bot")
        .await
        .map_err(|e| format!("Matrix login failed: {e}"))?;

    let bot_user_id = client
        .user_id()
        .ok_or_else(|| "No user ID after login".to_string())?
        .to_owned();

    slog!("[matrix-bot] Logged in as {bot_user_id}");

    // Parse and join the configured room
    let target_room_id: OwnedRoomId = config
        .room_id
        .parse()
        .map_err(|_| format!("Invalid room ID '{}'", config.room_id))?;

    // Try to join the room with a timeout. Conduit sometimes hangs or
    // returns errors on join if the bot is already a member.
    match tokio::time::timeout(
        std::time::Duration::from_secs(10),
        client.join_room_by_id(&target_room_id),
    )
    .await
    {
        Ok(Ok(_)) => slog!("[matrix-bot] Joined room {target_room_id}"),
        Ok(Err(e)) => slog!("[matrix-bot] Join room error (may already be a member): {e}"),
        Err(_) => slog!("[matrix-bot] Join room timed out (may already be a member)"),
    }

    let ctx = BotContext {
        bot_user_id,
        target_room_id,
        project_root,
    };

    // Register event handler and inject shared context
    client.add_event_handler_context(ctx);
    client.add_event_handler(on_room_message);

    slog!("[matrix-bot] Starting Matrix sync loop");

    // This blocks until the connection is terminated or an error occurs.
    client
        .sync(SyncSettings::default())
        .await
        .map_err(|e| format!("Matrix sync error: {e}"))?;

    Ok(())
}

/// Matrix event handler for room messages. Each invocation spawns an
/// independent task so the sync loop is not blocked by LLM calls.
async fn on_room_message(
    ev: OriginalSyncRoomMessageEvent,
    room: Room,
    Ctx(ctx): Ctx<BotContext>,
) {
    slog!(
        "[matrix-bot] Event received: room={} sender={} target={}",
        room.room_id(),
        ev.sender,
        ctx.target_room_id
    );

    // Only handle messages in the configured room
    if room.room_id() != &*ctx.target_room_id {
        slog!("[matrix-bot] Ignoring message from wrong room");
        return;
    }

    // Ignore the bot's own messages to prevent echo loops
    if ev.sender == ctx.bot_user_id {
        return;
    }

    // Only handle plain text messages
    let MessageType::Text(text_content) = ev.content.msgtype else {
        return;
    };

    let user_message = text_content.body.clone();
    slog!("[matrix-bot] Message from {}: {user_message}", ev.sender);

    // Spawn a separate task so the Matrix sync loop is not blocked while we
    // wait for the LLM response (which can take several seconds).
    tokio::spawn(async move {
        handle_message(room, ctx, user_message).await;
    });
}

async fn handle_message(room: Room, ctx: BotContext, user_message: String) {
    match call_claude_code(&ctx.project_root, &user_message).await {
        Ok(response) => {
            let _ = room
                .send(RoomMessageEventContent::text_plain(response))
                .await;
        }
        Err(e) => {
            slog!("[matrix-bot] LLM error: {e}");
            let _ = room
                .send(RoomMessageEventContent::text_plain(format!(
                    "Error processing your request: {e}"
                )))
                .await;
        }
    }
}

/// Call Claude Code with the user's message.
///
/// Uses the same `ClaudeCodeProvider` as the web UI chat. Claude Code manages
/// its own tools (including MCP tools) natively — no separate tool schemas or
/// HTTP self-calls needed.
async fn call_claude_code(
    project_root: &Path,
    user_message: &str,
) -> Result<String, String> {
    let provider = ClaudeCodeProvider::new();

    // Create a cancel channel that never fires — the bot doesn't support
    // mid-request cancellation (Matrix messages are fire-and-forget).
    let (cancel_tx, mut cancel_rx) = watch::channel(false);
    // Keep the sender alive for the duration of the call.
    let _cancel_tx = cancel_tx;

    // Collect text tokens into the final response. We don't stream to Matrix
    // (each message is posted as a single reply), so we just accumulate.
    let response_text = Arc::new(std::sync::Mutex::new(String::new()));
    let response_clone = Arc::clone(&response_text);

    let ClaudeCodeResult { messages, .. } = provider
        .chat_stream(
            user_message,
            &project_root.to_string_lossy(),
            None, // No session resumption for now (see story 182)
            &mut cancel_rx,
            move |token| {
                response_clone.lock().unwrap().push_str(token);
            },
            |_thinking| {}, // Discard thinking tokens
            |_activity| {}, // Discard activity signals
        )
        .await?;

    // Prefer the accumulated streamed text. If nothing was streamed (e.g.
    // Claude Code returned only tool calls with no final text), fall back to
    // extracting the last assistant message from the structured result.
    let streamed = response_text.lock().unwrap().clone();
    if !streamed.is_empty() {
        return Ok(streamed);
    }

    // Fallback: find the last assistant message
    let last_text = messages
        .iter()
        .rev()
        .find(|m| m.role == crate::llm::types::Role::Assistant && !m.content.is_empty())
        .map(|m| m.content.clone())
        .unwrap_or_default();

    if last_text.is_empty() {
        Err("Claude Code returned no response text".to_string())
    } else {
        Ok(last_text)
    }
}

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

    #[test]
    fn bot_context_is_clone() {
        // BotContext must be Clone for the Matrix event handler injection.
        fn assert_clone<T: Clone>() {}
        assert_clone::<BotContext>();
    }

    #[tokio::test]
    async fn call_claude_code_returns_error_when_claude_not_installed() {
        // When `claude` binary is not in PATH (or returns an error), the
        // provider should return an Err rather than panic.
        let fake_root = PathBuf::from("/tmp/nonexistent_project_root");
        let result = call_claude_code(&fake_root, "hello").await;
        // We expect either an error (claude not found) or a valid response
        // if claude happens to be installed. Both are acceptable — the key
        // property is that it doesn't panic.
        let _ = result;
    }
}