X-Git-Url: https://git.giacomolli.uk/onechat/blobdiff_plain/5da3ea038d1e9f9a2634d5ab137aef6989a3a789..1379d7d0d0173416237ddf8ebe820743d2cecfa6:/turn.go?ds=inline

diff --git a/turn.go b/turn.go
index 76f524e..c6a083b 100644
--- a/turn.go
+++ b/turn.go
@@ -5,30 +5,67 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"path/filepath"
 	"strings"
 	"sync"
 	"time"
 )
 
-// modelClient sends model calls.
-type modelClient interface {
-	Create(ctx context.Context, request Request) (Response, error)
+// errTurnLogWrite marks an error of a log write, which ends the loop.
+var errTurnLogWrite = errors.New("log write failed")
+
+// newTurnLoop returns the loop of agent with its memory. Its calls go through
+// client, and notify gets each event. The clock gives the date of a turn.
+func newTurnLoop(agent *Agent, memory *Memory, client TurnClient, clock func() time.Time, notify func(Event)) (*TurnLoop, error) {
+	layout, err := loadViewLayout(filepath.Join(agent.dir, viewLayoutFileName))
+	if err != nil {
+		return nil, err
+	}
+
+	l := &TurnLoop{
+		agent:  agent,
+		memory: memory,
+		client: client,
+		layout: layout,
+		clock:  clock,
+		notify: notify,
+		wake:   make(chan struct{}, 1),
+	}
+
+	l.tools = &Toolbox{
+		log:     l,
+		chat:    memory,
+		workdir: agent.config.Workdir,
+	}
+
+	return l, nil
+}
+
+// TurnClient sends model calls.
+type TurnClient interface {
+	Create(ctx context.Context, request ModelRequest) (ModelResponse, error)
 }
 
 // Event tells the clients what the agent does as it happens. Kind is the kind
-// of a logged message (user, reply, tool, echo), or one of busy, idle, and
-// error. ID is the ID of a logged message.
+// of a logged message (user, reply, tool, echo), or one of ready, busy, idle,
+// and error. ID is the ID of a logged message. Source names the client that
+// sent a user message, such as terminal or telegram, so that each client can
+// show the messages of the others.
 type Event struct {
-	Kind string `json:"kind"`
-	ID   int    `json:"id,omitempty"`
-	Text string `json:"text,omitempty"`
+	Kind   string `json:"kind"`
+	ID     int    `json:"id,omitempty"`
+	Text   string `json:"text,omitempty"`
+	Source string `json:"source,omitempty"`
 }
 
-// loop runs the turns of one agent. A user message goes to the log when it
+// TurnLoop runs the turns of one agent. A user message goes to the log when it
 // comes, and it wakes the loop. A turn waits until every message before it is
-// summarized. Then its first request holds the prompt, the tools, and the
-// input of turnInput. The loop runs the tool calls of each response and sends
-// their results back, until a response has no tool call: the final answer.
+// summarized. Then its first request holds the prompt, the tools, and the input
+// of turnInput. The loop runs the tool calls of each response and sends their
+// results back, until a response has no tool call: the final answer.
+//
+// The view goes out as the messages of a ViewLayout, so the first request of a
+// turn reads the view of the turn before from the prompt cache.
 //
 // Each later request of a turn begins with the whole input of the request
 // before it, then its output items, then the tool results. The output items
@@ -37,11 +74,12 @@ type Event struct {
 // from the prompt cache. User messages that come during a turn join the next
 // request after the tool results. The replies, the tool calls, and the results
 // go to the log as they happen. The thoughts never do.
-type loop struct {
+type TurnLoop struct {
 	agent  *Agent
 	memory *Memory
-	client modelClient
-	tools  *toolbox
+	client TurnClient
+	tools  *Toolbox
+	layout *ViewLayout
 	clock  func() time.Time
 	notify func(Event)
 	wake   chan struct{}
@@ -51,46 +89,11 @@ type loop struct {
 	cancel  context.CancelFunc
 }
 
-// functionCall holds the fields of a function call output item.
-type functionCall struct {
-	Type      string `json:"type"`
-	Name      string `json:"name"`
-	Arguments string `json:"arguments"`
-	CallID    string `json:"call_id"`
-}
-
-// functionCallOutput is the input item with the result of a function call.
-type functionCallOutput struct {
-	Type   string `json:"type"`
-	CallID string `json:"call_id"`
-	Output string `json:"output"`
-}
-
-// newLoop returns the loop of agent with its memory. Its calls go through
-// client, and notify gets each event. The clock gives the date of a turn.
-func newLoop(agent *Agent, memory *Memory, client modelClient, clock func() time.Time, notify func(Event)) *loop {
-	l := &loop{
-		agent:  agent,
-		memory: memory,
-		client: client,
-		clock:  clock,
-		notify: notify,
-		wake:   make(chan struct{}, 1),
-	}
-
-	l.tools = &toolbox{
-		log:     l,
-		chat:    memory,
-		workdir: agent.config.Workdir,
-	}
-
-	return l
-}
-
-// Submit logs a user message and wakes the loop. The message joins the next
-// request: the first one of a new turn, or the next one of a running turn.
-func (l *loop) Submit(text string) (Message, error) {
-	message, err := l.Append("user", text)
+// Submit logs a user message from the client source and wakes the loop. The
+// message joins the next request: the first one of a new turn, or the next one
+// of a running turn.
+func (l *TurnLoop) Submit(source, text string) (Message, error) {
+	message, err := l.record("user", source, text)
 	if err != nil {
 		return Message{}, err
 	}
@@ -109,7 +112,7 @@ func (l *loop) Submit(text string) (Message, error) {
 
 // Interrupt ends the running turn. A running tool is killed, and the turn sends
 // no further request.
-func (l *loop) Interrupt() bool {
+func (l *TurnLoop) Interrupt() bool {
 	l.mutex.Lock()
 	defer l.mutex.Unlock()
 
@@ -122,31 +125,28 @@ func (l *loop) Interrupt() bool {
 	return true
 }
 
-// Append logs a message and tells the clients. The toolbox logs its calls and
-// results through it.
-func (l *loop) Append(kind, text string) (Message, error) {
-	message, err := l.memory.Append(kind, text)
-	if err != nil {
-		return Message{}, err
-	}
+// Busy tells if a turn runs.
+func (l *TurnLoop) Busy() bool {
+	l.mutex.Lock()
+	defer l.mutex.Unlock()
 
-	l.notify(Event{
-		Kind: kind,
-		ID:   message.ID,
-		Text: text,
-	})
+	return l.cancel != nil
+}
 
-	return message, nil
+// Append logs a message and tells the clients. The toolbox logs its calls and
+// results through it.
+func (l *TurnLoop) Append(kind, text string) (Message, error) {
+	return l.record(kind, "", text)
 }
 
 // Flush makes the log durable.
-func (l *loop) Flush(ctx context.Context) error {
+func (l *TurnLoop) Flush(ctx context.Context) error {
 	return l.memory.Flush(ctx)
 }
 
 // Run runs a turn each time user messages wait, until ctx ends. It returns the
 // error of a failed log write, which stops the agent.
-func (l *loop) Run(ctx context.Context) error {
+func (l *TurnLoop) Run(ctx context.Context) error {
 	for {
 		for l.waiting() {
 			err := l.turn(ctx)
@@ -163,30 +163,37 @@ func (l *loop) Run(ctx context.Context) error {
 	}
 }
 
-// waiting tells if user messages wait for a request.
-func (l *loop) waiting() bool {
-	l.mutex.Lock()
-	defer l.mutex.Unlock()
+// record logs a message and tells the clients, with the client source of a user
+// message.
+func (l *TurnLoop) record(kind, source, text string) (Message, error) {
+	message, err := l.memory.Append(kind, text)
+	if err != nil {
+		return Message{}, err
+	}
 
-	return len(l.pending) > 0
+	l.notify(Event{
+		Kind:   kind,
+		ID:     message.ID,
+		Text:   text,
+		Source: source,
+	})
+
+	return message, nil
 }
 
-// take returns the user messages that wait for a request, and removes them.
-func (l *loop) take() []Message {
+// waiting tells if user messages wait for a request.
+func (l *TurnLoop) waiting() bool {
 	l.mutex.Lock()
 	defer l.mutex.Unlock()
 
-	messages := l.pending
-	l.pending = nil
-
-	return messages
+	return len(l.pending) > 0
 }
 
 // turn runs one turn. A failed or interrupted model call ends the turn and
 // tells the clients. The messages of the turn stay in the log and the view, and
 // the next turn starts only with a new message, so a lasting failure such as a
 // needed login does not repeat. Only a failed log write is an error of turn.
-func (l *loop) turn(ctx context.Context) error {
+func (l *TurnLoop) turn(ctx context.Context) error {
 	ctx, cancel := context.WithCancel(ctx)
 	defer cancel()
 
@@ -205,7 +212,7 @@ func (l *loop) turn(ctx context.Context) error {
 	})
 
 	err := l.chain(ctx)
-	if err != nil && !errors.Is(err, errLogWrite) {
+	if err != nil && !errors.Is(err, errTurnLogWrite) {
 		l.notify(Event{
 			Kind: "error",
 			Text: turnError(ctx, err),
@@ -221,11 +228,8 @@ func (l *loop) turn(ctx context.Context) error {
 	return err
 }
 
-// errLogWrite marks an error of a log write, which ends the loop.
-var errLogWrite = errors.New("log write failed")
-
 // chain runs the request chain of a turn until the final answer.
-func (l *loop) chain(ctx context.Context) error {
+func (l *TurnLoop) chain(ctx context.Context) error {
 	messages := l.take()
 	if len(messages) == 0 {
 		return nil
@@ -246,14 +250,20 @@ func (l *loop) chain(ctx context.Context) error {
 		skip[message.ID] = true
 	}
 
-	input := turnInput(l.memory.ViewWithout(skip), l.state(), messages)
-	for {
-		response, err := l.client.Create(ctx, Request{
+	view, err := l.layout.Split(l.memory.ViewLines(skip))
+	if err != nil {
+		return err
+	}
+
+	input := turnInput(view, l.state(), messages)
+	for turn := "first"; ; turn = "next" {
+		response, err := l.client.Create(ctx, ModelRequest{
 			Model:        l.agent.config.Model,
 			Effort:       l.agent.config.Effort,
 			Instructions: instructions,
-			Tools:        agentTools(),
+			Tools:        promptTools(),
 			Input:        input,
+			Turn:         turn,
 		})
 		if err != nil {
 			return err
@@ -270,21 +280,38 @@ func (l *loop) chain(ctx context.Context) error {
 	}
 }
 
+// take returns the user messages that wait for a request, and removes them.
+func (l *TurnLoop) take() []Message {
+	l.mutex.Lock()
+	defer l.mutex.Unlock()
+
+	messages := l.pending
+	l.pending = nil
+
+	return messages
+}
+
+// state returns the per-turn state: the end of the view, and the date and time
+// of the turn.
+func (l *TurnLoop) state() string {
+	return "</chat>\n\nNow: " + l.clock().Format("Monday 2006-01-02 15:04 -07:00")
+}
+
 // handle logs the replies of response and runs its tool calls. It returns the
 // input of the next request: input, the output items, the tool results, and
 // the user messages that came in the meantime. After the final answer it
 // returns no input.
-func (l *loop) handle(ctx context.Context, input []any, response Response) ([]any, error) {
+func (l *TurnLoop) handle(ctx context.Context, input []any, response ModelResponse) ([]any, error) {
 	var outputs []any
 	for _, raw := range response.Output {
-		item, err := inputItem(raw)
+		item, err := turnInputItem(raw)
 		if err != nil {
 			return nil, err
 		}
 
 		input = append(input, item)
 
-		var call functionCall
+		var call TurnFunctionCall
 		err = json.Unmarshal(raw, &call)
 		if err != nil {
 			return nil, fmt.Errorf("decode output item: %w", err)
@@ -292,7 +319,7 @@ func (l *loop) handle(ctx context.Context, input []any, response Response) ([]an
 
 		switch call.Type {
 		case "message":
-			text, err := outputText([]json.RawMessage{raw})
+			text, err := modelOutputText([]json.RawMessage{raw})
 			if err != nil {
 				return nil, err
 			}
@@ -304,15 +331,15 @@ func (l *loop) handle(ctx context.Context, input []any, response Response) ([]an
 
 			_, err = l.Append("reply", text)
 			if err != nil {
-				return nil, fmt.Errorf("%w: %w", errLogWrite, err)
+				return nil, fmt.Errorf("%w: %w", errTurnLogWrite, err)
 			}
 		case "function_call":
 			output, err := l.tools.Call(ctx, call.Name, call.Arguments)
 			if err != nil {
-				return nil, fmt.Errorf("%w: %w", errLogWrite, err)
+				return nil, fmt.Errorf("%w: %w", errTurnLogWrite, err)
 			}
 
-			outputs = append(outputs, functionCallOutput{
+			outputs = append(outputs, TurnFunctionCallOutput{
 				Type:   "function_call_output",
 				CallID: call.CallID,
 				Output: output,
@@ -326,32 +353,58 @@ func (l *loop) handle(ctx context.Context, input []any, response Response) ([]an
 
 	input = append(input, outputs...)
 	for _, message := range l.take() {
-		input = append(input, userMessage(message.Text))
+		input = append(input, modelUserMessage(message.Text))
 	}
 
 	return input, nil
 }
 
-// state returns the per-turn state: the date and time of the turn.
-func (l *loop) state() string {
-	return "Now: " + l.clock().Format("Monday 2006-01-02 15:04 -07:00")
+// TurnFunctionCall holds the fields of a function call output item.
+type TurnFunctionCall struct {
+	Type      string `json:"type"`
+	Name      string `json:"name"`
+	Arguments string `json:"arguments"`
+	CallID    string `json:"call_id"`
+}
+
+// TurnFunctionCallOutput is the input item with the result of a function call.
+type TurnFunctionCallOutput struct {
+	Type   string `json:"type"`
+	CallID string `json:"call_id"`
+	Output string `json:"output"`
+}
+
+// turnError returns the text of the error that ended a turn.
+func turnError(ctx context.Context, err error) string {
+	if ctx.Err() != nil {
+		return "interrupted"
+	}
+
+	return err.Error()
 }
 
-// turnInput returns the input of the first request of a turn: one user message
-// with the view, the per-turn state, and each new message whole. The view
-// leaves out the leaf lines of the new messages, which come whole after it.
-func turnInput(view, state string, messages []Message) []any {
-	parts := []string{view + "\n", state + "\n"}
+// turnInput returns the input of the first request of a turn, stable parts
+// first: a user message for each view message, one with the per-turn state,
+// and one for each new message, whole. The view leaves out the leaf lines of
+// the new messages.
+func turnInput(view []string, state string, messages []Message) []any {
+	input := make([]any, 0, len(view)+1+len(messages))
+	for _, text := range view {
+		input = append(input, modelUserMessage(text))
+	}
+
+	input = append(input, modelUserMessage(state))
 	for _, message := range messages {
-		parts = append(parts, message.Text)
+		input = append(input, modelUserMessage(message.Text))
 	}
 
-	return []any{userMessage(parts...)}
+	return input
 }
 
-// inputItem returns an output item as an input item for the next request. The
-// calls do not store their items, so the item goes back whole, without its id.
-func inputItem(raw json.RawMessage) (map[string]any, error) {
+// turnInputItem returns an output item as an input item for the next request.
+// The calls do not store their items, so the item goes back whole, without its
+// id.
+func turnInputItem(raw json.RawMessage) (map[string]any, error) {
 	var item map[string]any
 	err := json.Unmarshal(raw, &item)
 	if err != nil {
@@ -362,12 +415,3 @@ func inputItem(raw json.RawMessage) (map[string]any, error) {
 
 	return item, nil
 }
-
-// turnError returns the text of the error that ended a turn.
-func turnError(ctx context.Context, err error) string {
-	if ctx.Err() != nil {
-		return "interrupted"
-	}
-
-	return err.Error()
-}