X-Git-Url: https://git.giacomolli.uk/onechat/blobdiff_plain/2a276c273c21a1752d541bf87e074c5d1c27e30f..b1be0ee8530df5db8973bfe355f2f6d3149383a0:/tools.go

diff --git a/tools.go b/tools.go
index 0811981..621cd84 100644
--- a/tools.go
+++ b/tools.go
@@ -27,9 +27,9 @@ const (
 	// output, enough for toolClipCharacters/2 characters of 4 bytes.
 	toolClipBytes = 4 * toolClipCharacters / 2
 
-	// toolShellWaitDelay is how long a command can keep its output open
+	// toolShellOutputTimeout is how long a command can keep its output open
 	// after it ends or after its kill.
-	toolShellWaitDelay = 2 * time.Second
+	toolShellOutputTimeout = 2 * time.Second
 )
 
 // ToolLog is the log that the tools write their calls and results to.
@@ -49,11 +49,13 @@ type ToolChat interface {
 // the log as it happens: the call as a tool message, the result as an echo
 // message. The call is durable before its tool runs, so the log never misses an
 // action outside the agent. Commands and files run as the user in the working
-// directory, with no sandbox. Zoom with a name reads the log of a subagent. The
-// toolbox of the agent has the runs of its subagents for spawn and tell, the
-// WhatsApp link for the WhatsApp tools, and the schedules for the schedule
-// tools. A child has no WhatsApp link. The toolbox of a subagent has none of
-// them, so these are unknown tools there.
+// directory, with no sandbox. Zoom with a name reads a child's view or a
+// subagent's log. The toolbox of the agent has the runs of its subagents for
+// spawn and tell, the WhatsApp link for the WhatsApp tools, and the schedules
+// for the schedule tools. Each permanent agent has the parent-child queue. Only
+// the root has child creation and source routes. A child has neither of them
+// nor a WhatsApp link. Only the root can show images to the user. The toolbox
+// of a subagent has none of them, so these are unknown tools there.
 type Toolbox struct {
 	log       ToolLog
 	chat      ToolChat
@@ -62,13 +64,20 @@ type Toolbox struct {
 	whatsapp  *WhatsApp
 	schedules *Schedules
 	workdir   string
+	agentName string
+	messages  *DaemonMessages
+	children  *DaemonChildren
+	routes    *Routes
+	images    *ImageStore
+	imageShow *ImageShow
 }
 
 // Call logs the call of the tool name with its JSON arguments, makes the log
-// durable, runs the tool, and logs its clipped output, which it returns. A
-// failed tool gives its error as the output. An error of Call itself is an
-// error of the log. When ctx ends before the flush, the tool does not run.
-func (b *Toolbox) Call(ctx context.Context, name, arguments string) (string, error) {
+// durable, runs the tool, and logs its clipped text. It returns that text or
+// the text and image parts of a single-message zoom. Pixels never enter the
+// log. A failed tool gives its error as the output. An error of Call itself is
+// an error of the log. When ctx ends before the flush, the tool does not run.
+func (b *Toolbox) Call(ctx context.Context, name, arguments string) (any, error) {
 	_, err := b.log.Append("tool", toolCallText(name, arguments))
 	if err != nil {
 		return "", fmt.Errorf("append tool call: %w", err)
@@ -84,12 +93,26 @@ func (b *Toolbox) Call(ctx context.Context, name, arguments string) (string, err
 		output = toolClip(b.run(ctx, name, arguments))
 	}
 
+	var (
+		result    any = output
+		resultErr error
+	)
+
+	if err == nil {
+		result, resultErr = imageToolOutput(b, name, arguments, output)
+	}
+
+	if resultErr != nil {
+		output = "error: " + resultErr.Error()
+		result = output
+	}
+
 	_, err = b.log.Append("echo", output)
 	if err != nil {
 		return "", fmt.Errorf("append tool result: %w", err)
 	}
 
-	return output, nil
+	return result, nil
 }
 
 // run runs one tool and returns its output, or its error as text.
@@ -156,6 +179,46 @@ func (b *Toolbox) run(ctx context.Context, name, arguments string) string {
 		case err == nil:
 			output, err = b.runs.Spawn(ctx, args.Tasks, args.Effort)
 		}
+	case "agent_create":
+		var args ToolAgentCreateArguments
+
+		err = toolDecodeArguments(arguments, &args)
+		switch {
+		case b.children == nil:
+			err = fmt.Errorf("unknown tool %q", name)
+		case err == nil:
+			output, err = b.children.Create(args.Name, args.Responsibility)
+		}
+	case "route", "unroute", "routes":
+		var args ToolRouteArguments
+
+		switch name {
+		case "routes":
+			err = toolDecodeArguments(arguments, &struct{}{})
+		case "unroute":
+			var source ToolRouteSourceArguments
+
+			err = toolDecodeArguments(arguments, &source)
+			args.Source = source.Source
+		default:
+			err = toolDecodeArguments(arguments, &args)
+		}
+
+		switch {
+		case b.routes == nil:
+			err = fmt.Errorf("unknown tool %q", name)
+		case err != nil:
+		case name == "routes":
+			output = b.routes.List()
+		case name == "route" && (b.messages == nil || b.messages.Child(b.agentName, args.Agent) == nil):
+			err = errors.New("the agent must be a child of the root")
+		default:
+			err = b.routes.Set(args.Source, args.Agent)
+			output = "Route removed."
+			if name == "route" {
+				output = "Route set."
+			}
+		}
 	case "tell":
 		var args ToolTellArguments
 
@@ -163,9 +226,22 @@ func (b *Toolbox) run(ctx context.Context, name, arguments string) string {
 		switch {
 		case b.runs == nil:
 			err = fmt.Errorf("unknown tool %q", name)
+		case err == nil && b.messages != nil && b.messages.Child(b.agentName, args.Name) != nil:
+			err = b.messages.Tell(b.agentName, args.Name, args.Message)
+			output = "Sent. " + args.Name + "'s reply will reach you as a message."
 		case err == nil:
 			output, err = b.runs.Tell(ctx, args.Name, args.Message)
 		}
+	case "image_show":
+		var args ToolImageShowArguments
+
+		err = toolDecodeArguments(arguments, &args)
+		switch {
+		case b.imageShow == nil:
+			err = fmt.Errorf("unknown tool %q", name)
+		case err == nil:
+			output, err = b.imageShow.Show(ctx, args.Image, args.Caption)
+		}
 	case "whatsapp_find":
 		var args ToolWhatsAppFindArguments
 
@@ -184,7 +260,7 @@ func (b *Toolbox) run(ctx context.Context, name, arguments string) string {
 		case b.whatsapp == nil:
 			err = fmt.Errorf("unknown tool %q", name)
 		case err == nil:
-			output, err = b.whatsapp.Send(ctx, args.ID, args.Chat, args.Text)
+			output, err = b.whatsapp.Send(ctx, args.ID, args.Chat, args.Text, args.Image)
 		}
 	case "schedule_set":
 		var args ToolScheduleSetArguments
@@ -225,8 +301,15 @@ func (b *Toolbox) run(ctx context.Context, name, arguments string) string {
 	return output
 }
 
-// zoomAgent returns the whole log of the subagent name.
+// zoomAgent returns a child's view or the whole log of a subagent.
 func (b *Toolbox) zoomAgent(ctx context.Context, name string) (string, error) {
+	if b.messages != nil {
+		child := b.messages.Child(b.agentName, name)
+		if child != nil {
+			return strings.Join(child.memory.ViewLines(nil), ""), nil
+		}
+	}
+
 	subagent, err := b.subagents.Get(ctx, name)
 	if err != nil {
 		return "", err
@@ -307,7 +390,7 @@ func (b *Toolbox) shell(ctx context.Context, command string) string {
 	cmd.Cancel = func() error {
 		return syscall.Kill(-cmd.Process.Pid, syscall.SIGKILL)
 	}
-	cmd.WaitDelay = toolShellWaitDelay
+	cmd.WaitDelay = toolShellOutputTimeout
 
 	output := &ToolHeadTail{}
 	cmd.Stdout = output
@@ -410,8 +493,9 @@ func (b *Toolbox) path(path string) string {
 	return filepath.Join(b.workdir, path)
 }
 
-// ToolZoomArguments are the arguments of zoom. A Name that is not nil names
-// the subagent whose whole log zoom gives, and then ID and N are not used.
+// ToolZoomArguments are the arguments of zoom. A Name that is not nil names the
+// child whose view or the subagent whose whole log zoom gives. ID and N are
+// then not used.
 type ToolZoomArguments struct {
 	ID   int     `json:"id"`
 	N    int     `json:"n"`
@@ -453,12 +537,35 @@ type ToolSpawnArguments struct {
 	Effort *string  `json:"effort"`
 }
 
+// ToolAgentCreateArguments are the arguments of agent_create.
+type ToolAgentCreateArguments struct {
+	Name           string `json:"name"`
+	Responsibility string `json:"responsibility"`
+}
+
+// ToolRouteArguments are the arguments of route.
+type ToolRouteArguments struct {
+	Source string `json:"source"`
+	Agent  string `json:"agent"`
+}
+
+// ToolRouteSourceArguments are the arguments of unroute.
+type ToolRouteSourceArguments struct {
+	Source string `json:"source"`
+}
+
 // ToolTellArguments are the arguments of tell.
 type ToolTellArguments struct {
 	Name    string `json:"name"`
 	Message string `json:"message"`
 }
 
+// ToolImageShowArguments are the arguments of image_show.
+type ToolImageShowArguments struct {
+	Image   string `json:"image"`
+	Caption string `json:"caption"`
+}
+
 // ToolWhatsAppFindArguments are the arguments of whatsapp_find.
 type ToolWhatsAppFindArguments struct {
 	Name string `json:"name"`
@@ -466,9 +573,10 @@ type ToolWhatsAppFindArguments struct {
 
 // ToolWhatsAppSendArguments are the arguments of whatsapp_send.
 type ToolWhatsAppSendArguments struct {
-	ID   string `json:"id"`
-	Chat string `json:"chat"`
-	Text string `json:"text"`
+	ID    string `json:"id"`
+	Chat  string `json:"chat"`
+	Text  string `json:"text"`
+	Image string `json:"image"`
 }
 
 // ToolScheduleSetArguments are the arguments of schedule_set. Exactly one of At