X-Git-Url: http://git.giacomolli.uk/onechat/blobdiff_plain/77f29e78734dc16cfcea5f41b8e0185f8b33534a..2a276c273c21a1752d541bf87e074c5d1c27e30f:/daemon.go?ds=inline

diff --git a/daemon.go b/daemon.go
index e534aea..3a54b1a 100644
--- a/daemon.go
+++ b/daemon.go
@@ -24,12 +24,13 @@ const (
 	daemonClientEvents = 1024
 )
 
-// daemon runs "onechat daemon": it owns the root agent and its files and serves
-// the clients on the Unix socket in the data directory until ctx ends. When
-// telegram.json exists, it also runs the Telegram bot. It also keeps a linked
-// WhatsApp device connected. It holds the lock of the token file, so it is the
-// only process that refreshes the tokens and opens the WhatsApp session. Its
-// log lines go to out, through the loggers of daemonLogger.
+// daemon runs "onechat daemon": it owns the root, its children, and their
+// files, and serves the clients on the Unix socket in the data directory until
+// ctx ends. When telegram.json exists, it also runs the Telegram bot. It also
+// keeps a linked WhatsApp device connected. It holds the lock of the token
+// file, so it is the only process that refreshes the tokens and opens the
+// WhatsApp session. Its log lines go to out, through the loggers of
+// daemonLogger.
 func daemon(ctx context.Context, out io.Writer) (err error) {
 	dir, err := dataDir()
 	if err != nil {
@@ -80,30 +81,18 @@ func daemon(ctx context.Context, out io.Writer) (err error) {
 		return err
 	}
 
-	agent, err := openAgent(filepath.Join(dir, daemonAgentDirName))
-	if err != nil {
-		return err
-	}
-
-	client := newModelClient(address, auth, agent.CacheKey(), filepath.Join(agent.dir, "usage.jsonl"))
-	memory, err := openMemory(ctx, agent.dir, time.Now, newCompactor(client, agent.Prompt, config.Compactor), logger)
+	agents, err := openDaemonAgents(ctx, dir, address, auth, config.Compactor, time.Now, logger)
 	if err != nil {
 		return err
 	}
 
 	defer func() {
-		err = errors.Join(err, memory.Close())
+		err = errors.Join(err, agents.Close())
 	}()
 
-	schedules, err := openSchedules(filepath.Join(agent.dir, scheduleFileName), time.Now, logger)
-	if err != nil {
-		return err
-	}
-
-	subagents := newSubagents(filepath.Join(agent.dir, subagentDirName), time.Now, subagentNameRandom)
-	defer func() {
-		err = errors.Join(err, subagents.Close())
-	}()
+	root := agents[0]
+	agent := root.agent
+	schedules := root.schedules
 
 	listener, err := daemonListenSocket(socket)
 	if err != nil {
@@ -115,7 +104,7 @@ func daemon(ctx context.Context, out io.Writer) (err error) {
 		linked = "linked"
 	}
 
-	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())
+	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())
 
 	var telegram *TelegramBot
 
@@ -126,14 +115,16 @@ func daemon(ctx context.Context, out io.Writer) (err error) {
 	}
 
 	return daemonServe(ctx, listener, telegram, whatsapp, schedules, func(notify func(Event)) (*TurnLoop, error) {
-		return newTurnLoop(agent, memory, subagents, whatsapp, schedules, client, time.Now, logger, notify)
-	})
+		return root.Loop(whatsapp, notify)
+	}, agents[1:]...)
 }
 
 // DaemonHub holds the connected clients and sends each event to all of them. It
-// runs the WhatsApp commands of the clients on whatsapp.
+// runs the WhatsApp commands of the clients on whatsapp. snapshot locks all
+// agents and subagents during a backup copy.
 type DaemonHub struct {
 	whatsapp *WhatsApp
+	snapshot func(context.Context, func() error) error
 
 	mutex   sync.Mutex
 	clients map[*DaemonClient]bool
@@ -207,7 +198,7 @@ func (h *DaemonHub) read(ctx context.Context, client *DaemonClient, l *TurnLoop)
 			if request.Type == "backup" {
 				var size int64
 
-				size, err = backupData(ctx, filepath.Dir(l.agent.dir), request.Text, h.whatsapp, l.runs.Snapshot)
+				size, err = backupData(ctx, filepath.Dir(l.agent.dir), request.Text, h.whatsapp, h.snapshot)
 				if err == nil {
 					h.Send(client, Event{
 						Kind: "backup",
@@ -337,8 +328,10 @@ func daemonListenSocket(path string) (net.Listener, error) {
 // logs its batches and notes in the loop and tells all clients its notices, and
 // a failure of whatsapp stops the agent. It runs schedules, which log their
 // runs in the loop, and a failure of schedules stops the agent too. When bot is
-// not nil, it runs the bot too, and a failure of the bot stops the agent.
-func daemonServe(ctx context.Context, listener net.Listener, bot *TelegramBot, whatsapp *WhatsApp, schedules *Schedules, start func(notify func(Event)) (*TurnLoop, error)) error {
+// not nil, it runs the bot too, and a failure of the bot stops the agent. Each
+// child runs its own loop and schedules with no client events. A failed child
+// stops the daemon, and backups lock all loops before they copy data.
+func daemonServe(ctx context.Context, listener net.Listener, bot *TelegramBot, whatsapp *WhatsApp, schedules *Schedules, start func(notify func(Event)) (*TurnLoop, error), children ...*DaemonAgent) error {
 	ctx, cancel := context.WithCancel(ctx)
 	defer cancel()
 
@@ -360,8 +353,39 @@ func daemonServe(ctx context.Context, listener net.Listener, bot *TelegramBot, w
 		return errors.Join(err, listener.Close())
 	}
 
+	loops := []*TurnLoop{l}
+	for _, child := range children {
+		loop, err := child.Loop(nil, func(Event) {
+		})
+		if err != nil {
+			return errors.Join(fmt.Errorf("start child %q: %w", child.agent.config.Name, err), listener.Close())
+		}
+
+		loops = append(loops, loop)
+	}
+
+	hub.snapshot = func(ctx context.Context, snapshot func() error) error {
+		return daemonSnapshot(ctx, loops, snapshot)
+	}
+
 	var workers sync.WaitGroup
 
+	childErrors := make([]error, len(children))
+	for i, child := range children {
+		workers.Add(1)
+
+		go func() {
+			defer workers.Done()
+
+			childErrors[i] = daemonRunChild(ctx, loops[i+1], child.schedules)
+			if childErrors[i] != nil {
+				child.logger.Error("child stopped", "err", childErrors[i])
+			}
+
+			cancel()
+		}()
+	}
+
 	workers.Add(2)
 
 	go func() {
@@ -418,7 +442,30 @@ func daemonServe(ctx context.Context, listener net.Listener, bot *TelegramBot, w
 	hub.CloseAll()
 	workers.Wait()
 
-	return errors.Join(err, whatsappErr, schedulesErr, botErr)
+	return errors.Join(err, whatsappErr, schedulesErr, botErr, errors.Join(childErrors...))
+}
+
+// daemonRunChild runs a child's turn loop and its schedules until either stops.
+// Its events stay in its own log.
+func daemonRunChild(ctx context.Context, loop *TurnLoop, schedules *Schedules) error {
+	ctx, cancel := context.WithCancel(ctx)
+	defer cancel()
+
+	done := make(chan error, 1)
+	go func() {
+		done <- schedules.Run(ctx, loop)
+
+		cancel()
+	}()
+
+	err := loop.Run(ctx)
+	if errors.Is(err, context.Canceled) {
+		err = nil
+	}
+
+	cancel()
+
+	return errors.Join(err, <-done)
 }
 
 // daemonRunRequest runs one request of a client. The answers of a WhatsApp