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         "log"
  11         "net"
  12         "os"
  13         "path/filepath"
  14         "sync"
  15         "time"
  16 )
  17
  18 const (
  19         // daemonAgentDirName is the directory of the root agent in the data
  20         // directory.
  21         daemonAgentDirName = "agent"
  22
  23         // daemonClientEvents is the number of events that can wait for a slow
  24         // client before the daemon drops it.
  25         daemonClientEvents = 1024
  26 )
  27
  28 // daemon runs "onechat daemon": it owns the root agent and its files and serves
  29 // the clients on the Unix socket in the data directory until ctx ends. When
  30 // telegram.json exists, it also runs the Telegram bot. It also keeps a linked
  31 // WhatsApp device connected. It holds the lock of the token file, so it is the
  32 // only process that refreshes the tokens and opens the WhatsApp session. Its
  33 // log lines go to out, through the loggers of daemonLogger.
  34 func daemon(ctx context.Context, out io.Writer) (err error) {
  35         dir, err := dataDir()
  36         if err != nil {
  37                 return err
  38         }
  39
  40         // Check the socket path before the daemon writes a file, so that a data
  41         // directory with a long path stays empty.
  42         socket := filepath.Join(dir, socketFileName)
  43         err = checkSocketPath(socket)
  44         if err != nil {
  45                 return err
  46         }
  47
  48         config, err := loadConfig(dir)
  49         if err != nil {
  50                 return err
  51         }
  52
  53         botSettings, hasBot, err := loadTelegramConfig(dir)
  54         if err != nil {
  55                 return err
  56         }
  57
  58         tokens, err := lockTokens(dir)
  59         if err != nil {
  60                 return err
  61         }
  62
  63         defer func() {
  64                 err = errors.Join(err, tokens.Close())
  65         }()
  66
  67         whatsappSetDeviceProps()
  68         logger := daemonLogger(out, "onechat: ")
  69         whatsapp, err := openWhatsApp(ctx, dir, time.Now, daemonLogger(out, "whatsapp: "), whatsappNewClient)
  70         if err != nil {
  71                 return err
  72         }
  73
  74         defer func() {
  75                 err = errors.Join(err, whatsapp.Close())
  76         }()
  77
  78         address, auth, err := modelEndpoint(config, dir)
  79         if err != nil {
  80                 return err
  81         }
  82
  83         agent, err := openAgent(filepath.Join(dir, daemonAgentDirName))
  84         if err != nil {
  85                 return err
  86         }
  87
  88         client := newModelClient(address, auth, agent.CacheKey(), filepath.Join(agent.dir, "usage.jsonl"))
  89         memory, err := openMemory(ctx, agent.dir, time.Now, newCompactor(client, agent.Prompt, config.Compactor))
  90         if err != nil {
  91                 return err
  92         }
  93
  94         defer func() {
  95                 err = errors.Join(err, memory.Close())
  96         }()
  97
  98         schedules, err := openSchedules(filepath.Join(agent.dir, scheduleFileName), time.Now)
  99         if err != nil {
 100                 return err
 101         }
 102
 103         subagents := newSubagents(filepath.Join(agent.dir, subagentDirName), time.Now, subagentNameRandom)
 104         defer func() {
 105                 err = errors.Join(err, subagents.Close())
 106         }()
 107
 108         listener, err := daemonListenSocket(socket)
 109         if err != nil {
 110                 return err
 111         }
 112
 113         linked := "WhatsApp is not linked"
 114         if whatsapp.isLinked() {
 115                 linked = "WhatsApp is linked"
 116         }
 117
 118         logger.Printf("%s is ready: socket %s, model %s at %s effort, commands run in %s, %s, active schedules: %d", agent.config.Name, listener.Addr(), agent.config.Model, agent.config.Effort, agent.config.Workdir, linked, schedules.Len())
 119
 120         var telegram *TelegramBot
 121         if hasBot {
 122                 telegram = newTelegramBot(newTelegramAPI(telegramAPIURL, botSettings.Token), botSettings.Owner, filepath.Join(agent.dir, telegramOffsetFileName), daemonLogger(out, "telegram: "))
 123                 logger.Print("the Telegram bot waits for messages of its owner")
 124         }
 125
 126         return daemonServe(ctx, listener, telegram, whatsapp, schedules, func(notify func(Event)) (*TurnLoop, error) {
 127                 return newTurnLoop(agent, memory, subagents, whatsapp, schedules, client, time.Now, notify)
 128         })
 129 }
 130
 131 // DaemonHub holds the connected clients and sends each event to all of them. It
 132 // runs the WhatsApp commands of the clients on whatsapp.
 133 type DaemonHub struct {
 134         whatsapp *WhatsApp
 135
 136         mutex   sync.Mutex
 137         clients map[*DaemonClient]bool
 138         workers sync.WaitGroup
 139 }
 140
 141 // Accept serves each client of listener until the listener closes. A new
 142 // client gets a ready event with the name of the agent, and a busy event when a
 143 // turn runs.
 144 func (h *DaemonHub) Accept(ctx context.Context, listener net.Listener, l *TurnLoop) {
 145         for {
 146                 conn, err := listener.Accept()
 147                 if err != nil {
 148                         return
 149                 }
 150
 151                 client := &DaemonClient{
 152                         conn:   conn,
 153                         events: make(chan Event, daemonClientEvents),
 154                 }
 155
 156                 h.mutex.Lock()
 157                 h.clients[client] = true
 158                 h.mutex.Unlock()
 159
 160                 h.workers.Add(2)
 161                 go h.write(client)
 162                 go h.read(ctx, client, l)
 163
 164                 h.Send(client, Event{
 165                         Kind: "ready",
 166                         Text: l.agent.config.Name,
 167                 })
 168
 169                 if l.Busy() {
 170                         h.Send(client, Event{
 171                                 Kind: "busy",
 172                         })
 173                 }
 174         }
 175 }
 176
 177 // write sends the events of client as JSON lines until its channel closes.
 178 func (h *DaemonHub) write(client *DaemonClient) {
 179         defer h.workers.Done()
 180
 181         encoder := json.NewEncoder(client.conn)
 182         for event := range client.events {
 183                 err := encoder.Encode(event)
 184                 if err != nil {
 185                         h.drop(client)
 186                 }
 187         }
 188 }
 189
 190 // read runs the requests of client until it closes its connection.
 191 func (h *DaemonHub) read(ctx context.Context, client *DaemonClient, l *TurnLoop) {
 192         defer h.workers.Done()
 193         defer h.drop(client)
 194
 195         scanner := bufio.NewScanner(client.conn)
 196         scanner.Buffer(make([]byte, 64*1024), socketRequestBytes)
 197         for ctx.Err() == nil && scanner.Scan() {
 198                 var request SocketRequest
 199                 err := json.Unmarshal(scanner.Bytes(), &request)
 200                 if err == nil {
 201                         err = daemonRunRequest(l, h.whatsapp, request, func(text string) {
 202                                 h.Send(client, Event{
 203                                         Kind: "notice",
 204                                         Text: text,
 205                                 })
 206                         })
 207                 }
 208
 209                 if err != nil {
 210                         h.Send(client, Event{
 211                                 Kind: "error",
 212                                 Text: err.Error(),
 213                         })
 214                 }
 215         }
 216 }
 217
 218 // Send gives event to one client, or drops the client when its events do not
 219 // fit.
 220 func (h *DaemonHub) Send(client *DaemonClient, event Event) {
 221         h.mutex.Lock()
 222         defer h.mutex.Unlock()
 223
 224         h.send(client, event)
 225 }
 226
 227 // drop removes client and closes its connection.
 228 func (h *DaemonHub) drop(client *DaemonClient) {
 229         h.mutex.Lock()
 230         defer h.mutex.Unlock()
 231
 232         h.remove(client)
 233 }
 234
 235 // send gives event to client. The caller holds the mutex.
 236 func (h *DaemonHub) send(client *DaemonClient, event Event) {
 237         if !h.clients[client] {
 238                 return
 239         }
 240
 241         select {
 242         case client.events <- event:
 243         default:
 244                 h.remove(client)
 245         }
 246 }
 247
 248 // remove removes client, closes its connection, and ends its writer. The
 249 // caller holds the mutex.
 250 func (h *DaemonHub) remove(client *DaemonClient) {
 251         delete(h.clients, client)
 252         client.once.Do(func() {
 253                 close(client.events)
 254                 _ = client.conn.Close()
 255         })
 256 }
 257
 258 // Broadcast gives event to every client.
 259 func (h *DaemonHub) Broadcast(event Event) {
 260         h.mutex.Lock()
 261         defer h.mutex.Unlock()
 262
 263         for client := range h.clients {
 264                 h.send(client, event)
 265         }
 266 }
 267
 268 // CloseAll drops every client and waits for their readers and writers.
 269 func (h *DaemonHub) CloseAll() {
 270         h.mutex.Lock()
 271         for client := range h.clients {
 272                 h.remove(client)
 273         }
 274         h.mutex.Unlock()
 275
 276         h.workers.Wait()
 277 }
 278
 279 // DaemonClient is one connected client. Its events wait in a channel for the
 280 // writer, so a slow client does not stop the agent.
 281 type DaemonClient struct {
 282         conn   net.Conn
 283         events chan Event
 284         once   sync.Once
 285 }
 286
 287 // daemonListenSocket listens on the Unix socket at path, with mode 0600. The
 288 // caller owns the chat, so a socket file at path is left from a daemon that
 289 // ended, and daemonListenSocket removes it.
 290 func daemonListenSocket(path string) (net.Listener, error) {
 291         err := checkSocketPath(path)
 292         if err != nil {
 293                 return nil, err
 294         }
 295
 296         err = os.Remove(path)
 297         if err != nil && !errors.Is(err, os.ErrNotExist) {
 298                 return nil, fmt.Errorf("remove old socket: %w", err)
 299         }
 300
 301         listener, err := net.Listen("unix", path)
 302         if err != nil {
 303                 return nil, fmt.Errorf("listen on socket: %w", err)
 304         }
 305
 306         err = os.Chmod(path, 0o600)
 307         if err != nil {
 308                 return nil, errors.Join(fmt.Errorf("set socket mode: %w", err), listener.Close())
 309         }
 310
 311         return listener, nil
 312 }
 313
 314 // daemonServe runs the loop that start makes and serves the clients of listener
 315 // until ctx ends or the loop fails. It closes listener. It runs whatsapp, which
 316 // logs its batches and notes in the loop and tells all clients its notices, and
 317 // a failure of whatsapp stops the agent. It runs schedules, which log their
 318 // runs in the loop, and a failure of schedules stops the agent too. When bot is
 319 // not nil, it runs the bot too, and a failure of the bot stops the agent.
 320 func daemonServe(ctx context.Context, listener net.Listener, bot *TelegramBot, whatsapp *WhatsApp, schedules *Schedules, start func(notify func(Event)) (*TurnLoop, error)) error {
 321         ctx, cancel := context.WithCancel(ctx)
 322         defer cancel()
 323
 324         hub := &DaemonHub{
 325                 whatsapp: whatsapp,
 326                 clients:  make(map[*DaemonClient]bool),
 327         }
 328
 329         notify := hub.Broadcast
 330         if bot != nil {
 331                 notify = func(event Event) {
 332                         hub.Broadcast(event)
 333                         bot.Notify(event)
 334                 }
 335         }
 336
 337         l, err := start(notify)
 338         if err != nil {
 339                 return errors.Join(err, listener.Close())
 340         }
 341
 342         var workers sync.WaitGroup
 343         workers.Add(2)
 344         go func() {
 345                 defer workers.Done()
 346
 347                 hub.Accept(ctx, listener, l)
 348         }()
 349
 350         var whatsappErr error
 351         go func() {
 352                 defer workers.Done()
 353
 354                 whatsappErr = whatsapp.Run(ctx, l, notify)
 355                 cancel()
 356         }()
 357
 358         var schedulesErr error
 359         workers.Add(1)
 360         go func() {
 361                 defer workers.Done()
 362
 363                 schedulesErr = schedules.Run(ctx, l)
 364                 cancel()
 365         }()
 366
 367         var botErr error
 368         if bot != nil {
 369                 workers.Add(1)
 370                 go func() {
 371                         defer workers.Done()
 372
 373                         botErr = bot.Run(ctx, l, whatsapp)
 374                         cancel()
 375                 }()
 376         }
 377
 378         err = l.Run(ctx)
 379         if errors.Is(err, context.Canceled) {
 380                 err = nil
 381         }
 382
 383         cancel()
 384         err = errors.Join(err, listener.Close())
 385         hub.CloseAll()
 386         workers.Wait()
 387
 388         return errors.Join(err, whatsappErr, schedulesErr, botErr)
 389 }
 390
 391 // daemonRunRequest runs one request of a client. The answers of a WhatsApp
 392 // command go to answer.
 393 func daemonRunRequest(l *TurnLoop, whatsapp *WhatsApp, request SocketRequest, answer func(text string)) error {
 394         switch request.Type {
 395         case "send":
 396                 if request.Text == "" {
 397                         return errors.New("empty message")
 398                 }
 399
 400                 _, err := l.Submit(terminalSource, request.Text)
 401
 402                 return err
 403         case "interrupt":
 404                 if !l.Interrupt() {
 405                         return errors.New("no turn runs")
 406                 }
 407
 408                 return nil
 409         case "whatsapp":
 410                 whatsapp.Command(request.Text, answer)
 411
 412                 return nil
 413         case "stop":
 414                 found, err := l.StopSubagent(request.Text)
 415                 if err == nil && !found {
 416                         err = fmt.Errorf("no subagent %s runs", request.Text)
 417                 }
 418
 419                 return err
 420         default:
 421                 return fmt.Errorf("unknown request %q", request.Type)
 422         }
 423 }
 424
 425 // daemonLogger returns the logger of the part of the daemon with prefix, such
 426 // as "telegram: ", on out. The journal adds the time to each line and sets
 427 // JOURNAL_STREAM for a service, so then the lines have no time stamp. Otherwise
 428 // each line has the standard time stamp.
 429 func daemonLogger(out io.Writer, prefix string) *log.Logger {
 430         if os.Getenv("JOURNAL_STREAM") != "" {
 431                 return log.New(out, prefix, 0)
 432         }
 433
 434         return log.New(out, prefix, log.LstdFlags)
 435 }
 436
 437 // durationText returns d for a log line: rounded to the second, or to the
 438 // millisecond when it is shorter than a second.
 439 func durationText(d time.Duration) string {
 440         if d < time.Second {
 441                 return d.Round(time.Millisecond).String()
 442         }
 443
 444         return d.Round(time.Second).String()
 445 }