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
  48 // client gets a ready event with the name of the agent, and a busy event when a
  49 // turn 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
 171 // caller 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                 _ = client.conn.Close()
 178         })
 179 }
 180
 181 // Broadcast gives event to every client.
 182 func (h *DaemonHub) Broadcast(event Event) {
 183         h.mutex.Lock()
 184         defer h.mutex.Unlock()
 185
 186         for client := range h.clients {
 187                 h.send(client, event)
 188         }
 189 }
 190
 191 // CloseAll drops every client and waits for their readers and writers.
 192 func (h *DaemonHub) CloseAll() {
 193         h.mutex.Lock()
 194
 195         for client := range h.clients {
 196                 h.remove(client)
 197         }
 198
 199         h.mutex.Unlock()
 200
 201         h.workers.Wait()
 202 }
 203
 204 // daemon runs "onechat daemon": it owns the root, its children, and their
 205 // files, and serves the clients on the Unix socket in the data directory until
 206 // ctx ends. When telegram.json exists, it also runs the Telegram bot. It also
 207 // keeps a linked WhatsApp device connected. It holds the lock of the token
 208 // file, so it is the only process that refreshes the tokens and opens the
 209 // WhatsApp session. Its log lines go to out, through the loggers of
 210 // daemonLogger.
 211 func daemon(ctx context.Context, out io.Writer) (err error) {
 212         dir, err := dataDir()
 213         if err != nil {
 214                 return err
 215         }
 216
 217         // Check the socket path before the daemon writes a file, so that a data
 218         // directory with a long path stays empty.
 219         socket := filepath.Join(dir, socketFileName)
 220         err = checkSocketPath(socket)
 221         if err != nil {
 222                 return err
 223         }
 224
 225         config, err := loadConfig(dir)
 226         if err != nil {
 227                 return err
 228         }
 229
 230         botSettings, hasBot, err := loadTelegramConfig(dir)
 231         if err != nil {
 232                 return err
 233         }
 234
 235         tokens, err := lockTokens(dir)
 236         if err != nil {
 237                 return err
 238         }
 239
 240         defer func() {
 241                 err = errors.Join(err, tokens.Close())
 242         }()
 243
 244         whatsappSetDeviceProps()
 245
 246         logger := daemonLogger(out, "onechat: ")
 247         whatsapp, err := openWhatsApp(ctx, dir, time.Now, daemonLogger(out, "whatsapp: "), whatsappNewClient)
 248         if err != nil {
 249                 return err
 250         }
 251
 252         defer func() {
 253                 err = errors.Join(err, whatsapp.Close())
 254         }()
 255
 256         address, auth, err := modelEndpoint(config, dir)
 257         if err != nil {
 258                 return err
 259         }
 260
 261         agents, err := openDaemonAgents(ctx, dir, address, auth, config.Compactor, time.Now, logger)
 262         if err != nil {
 263                 return err
 264         }
 265
 266         defer func() {
 267                 err = errors.Join(err, agents.Close())
 268         }()
 269
 270         whatsapp.inbox.voice = &VoiceTranscriber{config: config.Voice}
 271
 272         root := agents[0]
 273         agent := root.agent
 274         schedules := root.schedules
 275
 276         listener, err := daemonListenSocket(socket)
 277         if err != nil {
 278                 return err
 279         }
 280
 281         linked := "not linked"
 282         if whatsapp.isLinked() {
 283                 linked = "linked"
 284         }
 285
 286         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())
 287
 288         var telegram *TelegramBot
 289
 290         if hasBot {
 291                 telegram = newTelegramBot(newTelegramAPI(telegramAPIURL, botSettings.Token), botSettings.Owner, filepath.Join(agent.dir, telegramOffsetFileName), daemonLogger(out, "telegram: "))
 292
 293                 telegram.voice = &VoiceTranscriber{config: config.Voice}
 294
 295                 logger.Info("Telegram bot started")
 296         }
 297
 298         return daemonServe(ctx, listener, telegram, whatsapp, schedules, func(notify func(Event)) (*TurnLoop, error) {
 299                 return root.Loop(whatsapp, notify)
 300         }, agents[1:]...)
 301 }
 302
 303 // daemonListenSocket listens on the Unix socket at path, with mode 0600. The
 304 // caller owns the chat, so a socket file at path is left from a daemon that
 305 // ended, and daemonListenSocket removes it.
 306 func daemonListenSocket(path string) (net.Listener, error) {
 307         err := checkSocketPath(path)
 308         if err != nil {
 309                 return nil, err
 310         }
 311
 312         err = os.Remove(path)
 313         if err != nil && !errors.Is(err, os.ErrNotExist) {
 314                 return nil, fmt.Errorf("remove old socket: %w", err)
 315         }
 316
 317         listener, err := net.Listen("unix", path)
 318         if err != nil {
 319                 return nil, fmt.Errorf("listen on socket: %w", err)
 320         }
 321
 322         err = os.Chmod(path, 0o600)
 323         if err != nil {
 324                 return nil, errors.Join(fmt.Errorf("set socket mode: %w", err), listener.Close())
 325         }
 326
 327         return listener, nil
 328 }
 329
 330 // daemonServe runs the loop that start makes and serves the clients of listener
 331 // until ctx ends or the loop fails. It closes listener. It runs whatsapp, which
 332 // logs its batches and notes in the loop and tells all clients its notices, and
 333 // a failure of whatsapp stops the agent. It runs schedules, which log their
 334 // runs in the loop, and a failure of schedules stops the agent too. When bot is
 335 // not nil, it runs the bot too, and a failure of the bot stops the agent. Each
 336 // child runs its own loop and schedules with no client events. A failed child
 337 // stops the daemon, and backups lock all loops before they copy data.
 338 func daemonServe(ctx context.Context, listener net.Listener, bot *TelegramBot, whatsapp *WhatsApp, schedules *Schedules, start func(notify func(Event)) (*TurnLoop, error), children ...*DaemonAgent) error {
 339         ctx, cancel := context.WithCancel(ctx)
 340         defer cancel()
 341
 342         hub := &DaemonHub{
 343                 whatsapp: whatsapp,
 344                 clients:  make(map[*DaemonClient]bool),
 345         }
 346
 347         notify := hub.Broadcast
 348         if bot != nil {
 349                 notify = func(event Event) {
 350                         hub.Broadcast(event)
 351                         bot.Notify(event)
 352                 }
 353         }
 354
 355         l, err := start(notify)
 356         if err != nil {
 357                 return errors.Join(err, listener.Close())
 358         }
 359
 360         if bot != nil {
 361                 l.tools.imageShow.bot = bot
 362         }
 363
 364         loops := []*TurnLoop{l}
 365         for _, child := range children {
 366                 loop, err := child.Loop(nil, func(Event) {
 367                 })
 368                 if err != nil {
 369                         return errors.Join(fmt.Errorf("start child %q: %w", child.agent.config.Name, err), listener.Close())
 370                 }
 371
 372                 loops = append(loops, loop)
 373         }
 374
 375         messages, err := openDaemonMessages(loops)
 376         if err != nil {
 377                 return errors.Join(fmt.Errorf("open agent messages: %w", err), listener.Close())
 378         }
 379
 380         hub.snapshot = func(ctx context.Context, snapshot func() error) error {
 381                 return messages.SnapshotAgents(ctx, snapshot)
 382         }
 383
 384         runtime := &DaemonChildren{
 385                 ctx:      ctx,
 386                 cancel:   cancel,
 387                 root:     l,
 388                 messages: messages,
 389         }
 390
 391         l.tools.children = runtime
 392
 393         var workers sync.WaitGroup
 394
 395         childErrors := make([]error, len(children))
 396         for i, child := range children {
 397                 workers.Add(1)
 398
 399                 go func() {
 400                         defer workers.Done()
 401
 402                         childErrors[i] = daemonRunChild(ctx, loops[i+1], child.schedules)
 403                         if childErrors[i] != nil {
 404                                 child.logger.Error("child stopped", "err", childErrors[i])
 405                         }
 406
 407                         cancel()
 408                 }()
 409         }
 410
 411         workers.Add(2)
 412
 413         go func() {
 414                 defer workers.Done()
 415
 416                 hub.Accept(ctx, listener, l)
 417         }()
 418
 419         var whatsappErr error
 420
 421         go func() {
 422                 defer workers.Done()
 423
 424                 whatsappErr = whatsapp.Run(ctx, l, notify)
 425
 426                 cancel()
 427         }()
 428
 429         var schedulesErr error
 430
 431         workers.Add(1)
 432
 433         go func() {
 434                 defer workers.Done()
 435
 436                 schedulesErr = schedules.Run(ctx, l)
 437
 438                 cancel()
 439         }()
 440
 441         var botErr error
 442
 443         if bot != nil {
 444                 workers.Add(1)
 445
 446                 go func() {
 447                         defer workers.Done()
 448
 449                         botErr = bot.Run(ctx, l, whatsapp)
 450
 451                         cancel()
 452                 }()
 453         }
 454
 455         err = l.Run(ctx)
 456         if errors.Is(err, context.Canceled) {
 457                 err = nil
 458         }
 459
 460         cancel()
 461
 462         err = errors.Join(err, listener.Close())
 463
 464         hub.CloseAll()
 465         workers.Wait()
 466
 467         return errors.Join(err, whatsappErr, schedulesErr, botErr, errors.Join(childErrors...), runtime.Close())
 468 }
 469
 470 // daemonRunChild runs a child's turn loop and its schedules until either stops.
 471 // Its events stay in its own log.
 472 func daemonRunChild(ctx context.Context, loop *TurnLoop, schedules *Schedules) error {
 473         ctx, cancel := context.WithCancel(ctx)
 474         defer cancel()
 475
 476         done := make(chan error, 1)
 477         go func() {
 478                 done <- schedules.Run(ctx, loop)
 479
 480                 cancel()
 481         }()
 482
 483         err := loop.Run(ctx)
 484         if errors.Is(err, context.Canceled) {
 485                 err = nil
 486         }
 487
 488         cancel()
 489
 490         return errors.Join(err, <-done)
 491 }
 492
 493 // daemonRunRequest runs one request of a client. The answers of a WhatsApp
 494 // command go to answer.
 495 func daemonRunRequest(l *TurnLoop, whatsapp *WhatsApp, request SocketRequest, answer func(text string)) error {
 496         switch request.Type {
 497         case "send":
 498                 if request.Text == "" {
 499                         return errors.New("empty message")
 500                 }
 501
 502                 _, err := l.Submit(terminalSource, request.Text)
 503
 504                 return err
 505         case "image":
 506                 if !filepath.IsAbs(request.Path) {
 507                         return errors.New("use /image PATH [caption]")
 508                 }
 509
 510                 data, err := imageReadFile(request.Path)
 511                 if err != nil {
 512                         return fmt.Errorf("read terminal image: %w", err)
 513                 }
 514
 515                 name, err := l.agent.Images().Save(data)
 516                 if err != nil {
 517                         return fmt.Errorf("store terminal image: %w", err)
 518                 }
 519
 520                 _, err = l.Submit(terminalSource, imageText([]string{name}, request.Text))
 521                 if err != nil {
 522                         return fmt.Errorf("submit terminal image: %w", err)
 523                 }
 524
 525                 return nil
 526         case "interrupt":
 527                 if !l.Interrupt() {
 528                         return errors.New("no turn runs")
 529                 }
 530
 531                 return nil
 532         case "whatsapp":
 533                 whatsapp.Command(request.Text, answer)
 534
 535                 return nil
 536         case "stop":
 537                 found, err := l.StopSubagent(request.Text)
 538                 if err == nil && !found {
 539                         err = fmt.Errorf("no subagent %s runs", request.Text)
 540                 }
 541
 542                 return err
 543         default:
 544                 return fmt.Errorf("unknown request %q", request.Type)
 545         }
 546 }