git.giacomolli.uk - onechat/blob - daemon.go

   1 package main
   2
   3 import (
   4         "bufio"
   5         "context"
   6         "encoding/json"
   7         "errors"
   8         "fmt"
   9         "io"
  10         "net"
  11         "os"
  12         "path/filepath"
  13         "sync"
  14         "time"
  15 )
  16
  17 const (
  18         // daemonAgentDirName is the directory of the root agent in the data
  19         // directory.
  20         daemonAgentDirName = "agent"
  21
  22         // daemonClientEvents is the number of events that can wait for a slow
  23         // client before the daemon drops it.
  24         daemonClientEvents = 1024
  25 )
  26
  27 // DaemonHub holds the connected clients and sends each event to all of them. It
  28 // runs the WhatsApp commands of the clients on whatsapp. snapshot locks all
  29 // agents and subagents during a backup copy.
  30 type DaemonHub struct {
  31         whatsapp *WhatsApp
  32         snapshot func(context.Context, func() error) error
  33
  34         mutex   sync.Mutex
  35         clients map[*DaemonClient]bool
  36         workers sync.WaitGroup
  37 }
  38
  39 // DaemonClient is one connected client. Its events wait in a channel for the
  40 // writer, so a slow client does not stop the agent.
  41 type DaemonClient struct {
  42         conn   net.Conn
  43         events chan Event
  44         once   sync.Once
  45 }
  46
  47 // Accept serves each client of listener until the listener closes. A new client
  48 // gets a ready event with the name of the agent, and a busy event when a turn
  49 // runs.
  50 func (h *DaemonHub) Accept(ctx context.Context, listener net.Listener, l *TurnLoop) {
  51         for {
  52                 conn, err := listener.Accept()
  53                 if err != nil {
  54                         return
  55                 }
  56
  57                 client := &DaemonClient{
  58                         conn:   conn,
  59                         events: make(chan Event, daemonClientEvents),
  60                 }
  61
  62                 h.mutex.Lock()
  63
  64                 h.clients[client] = true
  65                 h.mutex.Unlock()
  66
  67                 h.workers.Add(2)
  68
  69                 go h.write(client)
  70                 go h.read(ctx, client, l)
  71
  72                 h.Send(client, Event{
  73                         Kind: "ready",
  74                         Text: l.agent.config.Name,
  75                 })
  76
  77                 if l.Busy() {
  78                         h.Send(client, Event{
  79                                 Kind: "busy",
  80                         })
  81                 }
  82         }
  83 }
  84
  85 // write sends the events of client as JSON lines until its channel closes.
  86 func (h *DaemonHub) write(client *DaemonClient) {
  87         defer h.workers.Done()
  88
  89         encoder := json.NewEncoder(client.conn)
  90         for event := range client.events {
  91                 err := encoder.Encode(event)
  92                 if err != nil {
  93                         h.drop(client)
  94                 }
  95         }
  96 }
  97
  98 // read runs the requests of client until it closes its connection.
  99 func (h *DaemonHub) read(ctx context.Context, client *DaemonClient, l *TurnLoop) {
 100         defer h.workers.Done()
 101         defer h.drop(client)
 102
 103         scanner := bufio.NewScanner(client.conn)
 104         scanner.Buffer(make([]byte, 64*1024), socketRequestBytes)
 105
 106         for ctx.Err() == nil && scanner.Scan() {
 107                 var request SocketRequest
 108
 109                 err := json.Unmarshal(scanner.Bytes(), &request)
 110                 if err == nil {
 111                         if request.Type == "backup" {
 112                                 var size int64
 113
 114                                 size, err = backupData(ctx, filepath.Dir(l.agent.dir), request.Text, h.whatsapp, h.snapshot)
 115                                 if err == nil {
 116                                         h.Send(client, Event{
 117                                                 Kind: "backup",
 118                                                 Text: fmt.Sprintf("%s (%d bytes)", request.Text, size),
 119                                         })
 120                                 }
 121                         } else {
 122                                 err = daemonRunRequest(l, h.whatsapp, request, func(text string) {
 123                                         h.Send(client, Event{
 124                                                 Kind: "notice",
 125                                                 Text: text,
 126                                         })
 127                                 })
 128                         }
 129                 }
 130
 131                 if err != nil {
 132                         h.Send(client, Event{
 133                                 Kind: "error",
 134                                 Text: err.Error(),
 135                         })
 136                 }
 137         }
 138 }
 139
 140 // Send gives event to one client, or drops the client when its events do not
 141 // fit.
 142 func (h *DaemonHub) Send(client *DaemonClient, event Event) {
 143         h.mutex.Lock()
 144         defer h.mutex.Unlock()
 145
 146         h.send(client, event)
 147 }
 148
 149 // drop removes client and closes its connection.
 150 func (h *DaemonHub) drop(client *DaemonClient) {
 151         h.mutex.Lock()
 152         defer h.mutex.Unlock()
 153
 154         h.remove(client)
 155 }
 156
 157 // send gives event to client. The caller holds the mutex.
 158 func (h *DaemonHub) send(client *DaemonClient, event Event) {
 159         if !h.clients[client] {
 160                 return
 161         }
 162
 163         select {
 164         case client.events <- event:
 165         default:
 166                 h.remove(client)
 167         }
 168 }
 169
 170 // remove removes client, closes its connection, and ends its writer. The caller
 171 // holds the mutex.
 172 func (h *DaemonHub) remove(client *DaemonClient) {
 173         delete(h.clients, client)
 174         client.once.Do(func() {
 175                 close(client.events)
 176
 177                 // The client is removed already, so a close failure cannot
 178                 // restore it.
 179                 client.conn.Close() // nolint: errcheck
 180         })
 181 }
 182
 183 // Broadcast gives event to every client.
 184 func (h *DaemonHub) Broadcast(event Event) {
 185         h.mutex.Lock()
 186         defer h.mutex.Unlock()
 187
 188         for client := range h.clients {
 189                 h.send(client, event)
 190         }
 191 }
 192
 193 // CloseAll drops every client and waits for their readers and writers.
 194 func (h *DaemonHub) CloseAll() {
 195         h.mutex.Lock()
 196
 197         for client := range h.clients {
 198                 h.remove(client)
 199         }
 200
 201         h.mutex.Unlock()
 202
 203         h.workers.Wait()
 204 }
 205
 206 // daemonRunRequest runs one request of a client. The answers of a WhatsApp
 207 // command go to answer.
 208 func daemonRunRequest(l *TurnLoop, whatsapp *WhatsApp, request SocketRequest, answer func(text string)) error {
 209         switch request.Type {
 210         case "send":
 211                 if request.Text == "" {
 212                         return errors.New("empty message")
 213                 }
 214
 215                 _, err := l.Submit(terminalSource, request.Text)
 216
 217                 return err
 218         case "image":
 219                 if !filepath.IsAbs(request.Path) {
 220                         return errors.New("use /image PATH [caption]")
 221                 }
 222
 223                 data, err := imageReadFile(request.Path)
 224                 if err != nil {
 225                         return fmt.Errorf("read terminal image: %w", err)
 226                 }
 227
 228                 name, err := l.agent.Images().Save(data)
 229                 if err != nil {
 230                         return fmt.Errorf("store terminal image: %w", err)
 231                 }
 232
 233                 _, err = l.Submit(terminalSource, imageText([]string{name}, request.Text))
 234                 if err != nil {
 235                         return fmt.Errorf("submit terminal image: %w", err)
 236                 }
 237
 238                 return nil
 239         case "interrupt":
 240                 if !l.Interrupt() {
 241                         return errors.New("no turn runs")
 242                 }
 243
 244                 return nil
 245         case "whatsapp":
 246                 whatsapp.Command(request.Text, answer)
 247
 248                 return nil
 249         case "stop":
 250                 found, err := l.StopSubagent(request.Text)
 251                 if err == nil && !found {
 252                         err = fmt.Errorf("no subagent %s runs", request.Text)
 253                 }
 254
 255                 return err
 256         default:
 257                 return fmt.Errorf("unknown request %q", request.Type)
 258         }
 259 }
 260
 261 // daemon runs "onechat daemon": it owns the root, its children, and their
 262 // files, and serves the clients on the Unix socket in the data directory until
 263 // ctx ends. When telegram.json exists, it also runs the Telegram bot. It also
 264 // keeps a linked WhatsApp device connected. It holds the lock of the token
 265 // file, so it is the only process that refreshes the tokens and opens the
 266 // WhatsApp session. Its log lines go to out, through the loggers of
 267 // daemonLogger.
 268 func daemon(ctx context.Context, out io.Writer) (err error) {
 269         dir, err := dataDir()
 270         if err != nil {
 271                 return err
 272         }
 273
 274         // Check the socket path before the daemon writes a file, so that a data
 275         // directory with a long path stays empty.
 276         socket := filepath.Join(dir, socketFileName)
 277         err = checkSocketPath(socket)
 278         if err != nil {
 279                 return err
 280         }
 281
 282         config, err := loadConfig(dir)
 283         if err != nil {
 284                 return err
 285         }
 286
 287         botSettings, hasBot, err := loadTelegramConfig(dir)
 288         if err != nil {
 289                 return err
 290         }
 291
 292         tokens, err := lockTokens(dir)
 293         if err != nil {
 294                 return err
 295         }
 296
 297         defer func() {
 298                 err = errors.Join(err, tokens.Close())
 299         }()
 300
 301         whatsappSetDeviceProps()
 302
 303         logger := daemonLogger(out, "onechat: ")
 304         whatsapp, err := openWhatsApp(ctx, dir, time.Now, daemonLogger(out, "whatsapp: "), whatsappNewClient)
 305         if err != nil {
 306                 return err
 307         }
 308
 309         defer func() {
 310                 err = errors.Join(err, whatsapp.Close())
 311         }()
 312
 313         address, auth, err := modelEndpoint(config, dir)
 314         if err != nil {
 315                 return err
 316         }
 317
 318         agents, err := openDaemonAgents(ctx, dir, address, auth, config.Compactor, time.Now, logger)
 319         if err != nil {
 320                 return err
 321         }
 322
 323         defer func() {
 324                 err = errors.Join(err, agents.Close())
 325         }()
 326
 327         whatsapp.inbox.voice = &VoiceTranscriber{config: config.Voice}
 328
 329         root := agents[0]
 330         agent := root.agent
 331         schedules := root.schedules
 332
 333         listener, err := daemonListenSocket(socket)
 334         if err != nil {
 335                 return err
 336         }
 337
 338         linked := "not linked"
 339         if whatsapp.isLinked() {
 340                 linked = "linked"
 341         }
 342
 343         logger.Info("ready", "version", version, "socket", listener.Addr(), "model", agent.config.Model, "effort", agent.config.Effort, "workdir", agent.config.Workdir, "whatsapp", linked, "schedules", schedules.Len(), "children", agents.ChildNames())
 344
 345         var telegram *TelegramBot
 346
 347         if hasBot {
 348                 telegram = newTelegramBot(newTelegramAPI(telegramAPIURL, botSettings.Token), botSettings.Owner, filepath.Join(agent.dir, telegramOffsetFileName), daemonLogger(out, "telegram: "))
 349
 350                 telegram.voice = &VoiceTranscriber{config: config.Voice}
 351
 352                 logger.Info("Telegram bot started")
 353         }
 354
 355         return daemonServe(ctx, listener, telegram, whatsapp, schedules, func(notify func(Event)) (*TurnLoop, error) {
 356                 return root.Loop(whatsapp, notify)
 357         }, agents[1:]...)
 358 }
 359
 360 // daemonListenSocket listens on the Unix socket at path, with mode 0600. The
 361 // caller owns the chat, so a socket file at path is left from a daemon that
 362 // ended, and daemonListenSocket removes it.
 363 func daemonListenSocket(path string) (net.Listener, error) {
 364         err := checkSocketPath(path)
 365         if err != nil {
 366                 return nil, err
 367         }
 368
 369         err = os.Remove(path)
 370         if err != nil && !errors.Is(err, os.ErrNotExist) {
 371                 return nil, fmt.Errorf("remove old socket: %w", err)
 372         }
 373
 374         listener, err := net.Listen("unix", path)
 375         if err != nil {
 376                 return nil, fmt.Errorf("listen on socket: %w", err)
 377         }
 378
 379         err = os.Chmod(path, 0o600)
 380         if err != nil {
 381                 return nil, errors.Join(fmt.Errorf("set socket mode: %w", err), listener.Close())
 382         }
 383
 384         return listener, nil
 385 }
 386
 387 // daemonServe runs the loop that start makes and serves the clients of listener
 388 // until ctx ends or the loop fails. It closes listener. It runs whatsapp, which
 389 // logs its batches and notes in the loop and tells all clients its notices, and
 390 // a failure of whatsapp stops the agent. It runs schedules, which log their
 391 // runs in the loop, and a failure of schedules stops the agent too. When bot is
 392 // not nil, it runs the bot too, and a failure of the bot stops the agent. Each
 393 // child runs its own loop and schedules with no client events. A failed child
 394 // stops the daemon, and backups lock all loops before they copy data.
 395 func daemonServe(ctx context.Context, listener net.Listener, bot *TelegramBot, whatsapp *WhatsApp, schedules *Schedules, start func(notify func(Event)) (*TurnLoop, error), children ...*DaemonAgent) error {
 396         ctx, cancel := context.WithCancel(ctx)
 397         defer cancel()
 398
 399         hub := &DaemonHub{
 400                 whatsapp: whatsapp,
 401                 clients:  make(map[*DaemonClient]bool),
 402         }
 403
 404         notify := hub.Broadcast
 405         if bot != nil {
 406                 notify = func(event Event) {
 407                         hub.Broadcast(event)
 408                         bot.Notify(event)
 409                 }
 410         }
 411
 412         l, err := start(notify)
 413         if err != nil {
 414                 return errors.Join(err, listener.Close())
 415         }
 416
 417         if bot != nil {
 418                 l.tools.imageShow.bot = bot
 419         }
 420
 421         loops := []*TurnLoop{l}
 422         for _, child := range children {
 423                 loop, err := child.Loop(nil, func(Event) {
 424                 })
 425                 if err != nil {
 426                         return errors.Join(fmt.Errorf("start child %q: %w", child.agent.config.Name, err), listener.Close())
 427                 }
 428
 429                 loops = append(loops, loop)
 430         }
 431
 432         messages, err := openDaemonMessages(loops)
 433         if err != nil {
 434                 return errors.Join(fmt.Errorf("open agent messages: %w", err), listener.Close())
 435         }
 436
 437         hub.snapshot = func(ctx context.Context, snapshot func() error) error {
 438                 return messages.SnapshotAgents(ctx, snapshot)
 439         }
 440
 441         runtime := &DaemonChildren{
 442                 ctx:      ctx,
 443                 cancel:   cancel,
 444                 root:     l,
 445                 messages: messages,
 446         }
 447
 448         l.tools.children = runtime
 449
 450         var workers sync.WaitGroup
 451
 452         childErrors := make([]error, len(children))
 453         for i, child := range children {
 454                 workers.Add(1)
 455
 456                 go func() {
 457                         defer workers.Done()
 458
 459                         childErrors[i] = daemonRunChild(ctx, loops[i+1], child.schedules)
 460                         if childErrors[i] != nil {
 461                                 child.logger.Error("child stopped", "err", childErrors[i])
 462                         }
 463
 464                         cancel()
 465                 }()
 466         }
 467
 468         workers.Add(2)
 469
 470         go func() {
 471                 defer workers.Done()
 472
 473                 hub.Accept(ctx, listener, l)
 474         }()
 475
 476         var whatsappErr error
 477
 478         go func() {
 479                 defer workers.Done()
 480
 481                 whatsappErr = whatsapp.Run(ctx, l, notify)
 482
 483                 cancel()
 484         }()
 485
 486         var schedulesErr error
 487
 488         workers.Add(1)
 489
 490         go func() {
 491                 defer workers.Done()
 492
 493                 schedulesErr = schedules.Run(ctx, l)
 494
 495                 cancel()
 496         }()
 497
 498         var botErr error
 499
 500         if bot != nil {
 501                 workers.Add(1)
 502
 503                 go func() {
 504                         defer workers.Done()
 505
 506                         botErr = bot.Run(ctx, l, whatsapp)
 507
 508                         cancel()
 509                 }()
 510         }
 511
 512         err = l.Run(ctx)
 513         if errors.Is(err, context.Canceled) {
 514                 err = nil
 515         }
 516
 517         cancel()
 518
 519         err = errors.Join(err, listener.Close())
 520
 521         hub.CloseAll()
 522         workers.Wait()
 523
 524         return errors.Join(err, whatsappErr, schedulesErr, botErr, errors.Join(childErrors...), runtime.Close())
 525 }
 526
 527 // daemonRunChild runs a child's turn loop and its schedules until either stops.
 528 // Its events stay in its own log.
 529 func daemonRunChild(ctx context.Context, loop *TurnLoop, schedules *Schedules) error {
 530         ctx, cancel := context.WithCancel(ctx)
 531         defer cancel()
 532
 533         done := make(chan error, 1)
 534         go func() {
 535                 done <- schedules.Run(ctx, loop)
 536
 537                 cancel()
 538         }()
 539
 540         err := loop.Run(ctx)
 541         if errors.Is(err, context.Canceled) {
 542                 err = nil
 543         }
 544
 545         cancel()
 546
 547         return errors.Join(err, <-done)
 548 }