Sync agent queue replacements before success

[onechat] / turn.go

diff --git a/turn.go b/turn.go
index 837ac7c519ecbdd6760bdbc370c7522a622f5422..c08e94a0eed7d94a22e787d7bbaf97a69ce89289 100644 (file)
--- a/turn.go
+++ b/turn.go
@@ -4,15 +4,66 @@ import (
        "context"
        "errors"
        "fmt"
        "context"
        "errors"
        "fmt"

+       "log/slog"

        "path/filepath"
        "sync"
        "time"
 )
 
        "path/filepath"
        "sync"
        "time"
 )
 

+// Event tells the clients what the agent does as it happens. Kind is the kind
+// of a logged message (user, reply, tool, echo, note, work), or one of ready,
+// busy, idle, error, and subagent, which tells that the subagent Text starts a
+// run. 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, or the subagent of a report. Status tells how the run
+// of a report ended: finished, stopped, or failed.
+type Event struct {
+       Kind   string `json:"kind"`
+       ID     int    `json:"id,omitempty"`
+       Text   string `json:"text,omitempty"`
+       Source string `json:"source,omitempty"`
+       Status string `json:"status,omitempty"`
+}
+
+// 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 the turn runs a Chain whose first request holds the prompt,
+// the tools, and the input of turnInput.
+//
+// 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. User messages
+// that come during a turn join the next request of its chain after the tool
+// results, and so do the reports of the subagents. The subagents run until the
+// loop ends. Parent and child messages stay queued until the turn finishes.
+// active belongs to the running turn. The mutex protects pending, cancel, and
+// failure.
+type TurnLoop struct {
+       agent     *Agent
+       memory    *Memory
+       client    ChainClient
+       tools     *Toolbox
+       runs      *SubagentRuns
+       layout    *ViewLayout
+       clock     func() time.Time
+       logger    *slog.Logger
+       notify    func(Event)
+       wake      chan struct{}
+       messages  *DaemonMessages
+       openChild func(context.Context, *Agent) (*DaemonAgent, error)
+       active    []Message
+
+       mutex   sync.Mutex
+       pending []Message
+       cancel  context.CancelFunc
+       failure error
+}
+

 // newTurnLoop returns the loop of agent with its memory, its subagents, and the
 // WhatsApp link and the schedules of its tools. Its calls go through client,
 // newTurnLoop returns the loop of agent with its memory, its subagents, and the
 // WhatsApp link and the schedules of its tools. Its calls go through client,

-// and notify gets each event. The clock gives the date of a turn.
-func newTurnLoop(agent *Agent, memory *Memory, subagents *Subagents, whatsapp *WhatsApp, schedules *Schedules, client ChainClient, clock func() time.Time, notify func(Event)) (*TurnLoop, error) {
+// and notify gets each event. The clock gives the date of a turn. logger gets
+// the failed turns and the ends of the subagents. A child has no WhatsApp
+// tools, even when the caller supplies a link.
+func newTurnLoop(agent *Agent, memory *Memory, subagents *Subagents, whatsapp *WhatsApp, schedules *Schedules, client ChainClient, clock func() time.Time, logger *slog.Logger, notify func(Event)) (*TurnLoop, error) {

        layout, err := loadViewLayout(filepath.Join(agent.dir, viewLayoutFileName))
        if err != nil {
                return nil, err
        layout, err := loadViewLayout(filepath.Join(agent.dir, viewLayoutFileName))
        if err != nil {
                return nil, err


@@ -24,6 +75,7 @@ func newTurnLoop(agent *Agent, memory *Memory, subagents *Subagents, whatsapp *W
                client: client,
                layout: layout,
                clock:  clock,
                client: client,
                layout: layout,
                clock:  clock,

+               logger: logger,

                notify: notify,
                wake:   make(chan struct{}, 1),
        }
                notify: notify,
                wake:   make(chan struct{}, 1),
        }


@@ -38,10 +90,15 @@ func newTurnLoop(agent *Agent, memory *Memory, subagents *Subagents, whatsapp *W
                report:    l.Report,
                note:      l.note,
                fail:      l.fail,
                report:    l.Report,
                note:      l.note,
                fail:      l.fail,

+               logger:    logger,

                notify:    notify,
                running:   make(map[string]*SubagentRun),
        }
 
                notify:    notify,
                running:   make(map[string]*SubagentRun),
        }
 

+       if agent.config.Parent != "" {
+               whatsapp = nil
+       }
+

        l.tools = &Toolbox{
                log:       l,
                chat:      memory,
        l.tools = &Toolbox{
                log:       l,
                chat:      memory,


@@ -50,72 +107,26 @@ func newTurnLoop(agent *Agent, memory *Memory, subagents *Subagents, whatsapp *W
                whatsapp:  whatsapp,
                schedules: schedules,
                workdir:   agent.config.Workdir,
                whatsapp:  whatsapp,
                schedules: schedules,
                workdir:   agent.config.Workdir,

+               agentName: agent.config.Name,
+               images:    agent.Images(),

        }
 
        }
 

-       return l, nil
-}
-
-// Event tells the clients what the agent does as it happens. Kind is the kind
-// of a logged message (user, reply, tool, echo, note, work), or one of ready,
-// busy, idle, error, and subagent, which tells that the subagent Text starts a
-// run. 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, or the subagent of a report. Status tells how the run
-// of a report ended: finished, stopped, or failed.
-type Event struct {
-       Kind   string `json:"kind"`
-       ID     int    `json:"id,omitempty"`
-       Text   string `json:"text,omitempty"`
-       Source string `json:"source,omitempty"`
-       Status string `json:"status,omitempty"`
-}
-
-// 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 the turn runs a Chain whose first request holds the prompt,
-// the tools, and the input of turnInput.
-//
-// 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. User messages
-// that come during a turn join the next request of its chain after the tool
-// results, and so do the reports of the subagents. The subagents run until the
-// loop ends.
-type TurnLoop struct {
-       agent  *Agent
-       memory *Memory
-       client ChainClient
-       tools  *Toolbox
-       runs   *SubagentRuns
-       layout *ViewLayout
-       clock  func() time.Time
-       notify func(Event)
-       wake   chan struct{}
-
-       mutex   sync.Mutex
-       pending []Message
-       cancel  context.CancelFunc
-       failure error
-}
-
-// 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
+       if agent.config.Parent == "" {
+               l.tools.imageShow = &ImageShow{store: agent.Images()}

        }
 
        }
 

-       l.mutex.Lock()
-       l.pending = append(l.pending, message)
-       l.mutex.Unlock()
+       return l, nil
+}

 
 

-       select {
-       case l.wake <- struct{}{}:
-       default:
+// state returns the per-turn state: the end of the view, and the date and time
+// of the turn, then the names and responsibilities of its children.
+func (l *TurnLoop) state() string {
+       state := "</chat>\n\nNow: " + l.clock().Format("Monday 2006-01-02 15:04 -07:00")
+       if l.messages != nil {
+               state += l.messages.ChildState(l.agent.config.Name)

        }
 
        }
 

-       return message, nil
+       return state

 }
 
 // Report logs the report text of the subagent name as a work message and tells
 }
 
 // Report logs the report text of the subagent name as a work message and tells


@@ -133,6 +144,26 @@ func (l *TurnLoop) Report(name, text, status string) {
        }
 }
 
        }
 }
 

+// note logs a note and tells the clients. A note starts no turn.
+func (l *TurnLoop) note(text string) error {
+       _, err := l.record("note", "", text)
+
+       return err
+}
+
+// fail stops the loop with err.
+func (l *TurnLoop) fail(err error) {
+       l.mutex.Lock()
+
+       l.failure = errors.Join(l.failure, err)
+       l.mutex.Unlock()
+
+       select {
+       case l.wake <- struct{}{}:
+       default:
+       }
+}
+

 // Work logs text as a work message of source, such as the name of a subagent,
 // and tells the clients, with status. The message starts no turn until it goes
 // to Wake.
 // Work logs text as a work message of source, such as the name of a subagent,
 // and tells the clients, with status. The message starts no turn until it goes
 // to Wake.


@@ -162,6 +193,7 @@ func (l *TurnLoop) Wake(messages ...Message) {
        }
 
        l.mutex.Lock()
        }
 
        l.mutex.Lock()

+

        l.pending = append(l.pending, messages...)
        l.mutex.Unlock()
 
        l.pending = append(l.pending, messages...)
        l.mutex.Unlock()
 


@@ -171,6 +203,46 @@ func (l *TurnLoop) Wake(messages ...Message) {
        }
 }
 
        }
 }
 

+// 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
+       }
+
+       l.notify(Event{
+               Kind:   kind,
+               ID:     message.ID,
+               Text:   text,
+               Source: source,
+       })
+
+       return message, nil
+}
+
+// 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
+       }
+
+       l.mutex.Lock()
+
+       l.pending = append(l.pending, message)
+       l.mutex.Unlock()
+
+       select {
+       case l.wake <- struct{}{}:
+       default:
+       }
+
+       return message, nil
+}
+

 // StopSubagent ends the run of the subagent name and tells if it ran. Its
 // report is logged, but it starts no turn.
 func (l *TurnLoop) StopSubagent(name string) (bool, error) {
 // StopSubagent ends the run of the subagent name and tells if it ran. Its
 // report is logged, but it starts no turn.
 func (l *TurnLoop) StopSubagent(name string) (bool, error) {


@@ -224,10 +296,15 @@ func (l *TurnLoop) Run(ctx context.Context) error {
        }
 
        for {
        }
 
        for {

-               for l.waiting() {
-                       err := l.turn(ctx)
+               for ctx.Err() == nil && l.waiting() {
+                       err := l.failed()

                        if err != nil {
                        if err != nil {

-                               return err
+                               return fmt.Errorf("check turn loop failure: %w", err)
+                       }
+
+                       err = l.turn(ctx)
+                       if err != nil {
+                               return fmt.Errorf("run turn: %w", err)

                        }
                }
 
                        }
                }
 


@@ -239,36 +316,11 @@ func (l *TurnLoop) Run(ctx context.Context) error {
 
                err := l.failed()
                if err != nil {
 
                err := l.failed()
                if err != nil {

-                       return err
+                       return fmt.Errorf("check turn loop failure: %w", err)

                }
        }
 }
 
                }
        }
 }
 

-// note logs a note and tells the clients. A note starts no turn.
-func (l *TurnLoop) note(text string) error {
-       _, err := l.record("note", "", text)
-
-       return err
-}
-
-// 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
-       }
-
-       l.notify(Event{
-               Kind:   kind,
-               ID:     message.ID,
-               Text:   text,
-               Source: source,
-       })
-
-       return message, nil
-}
-

 // waiting tells if user messages or reports wait for a request.
 func (l *TurnLoop) waiting() bool {
        l.mutex.Lock()
 // waiting tells if user messages or reports wait for a request.
 func (l *TurnLoop) waiting() bool {
        l.mutex.Lock()


@@ -277,18 +329,6 @@ func (l *TurnLoop) waiting() bool {
        return len(l.pending) > 0
 }
 
        return len(l.pending) > 0
 }
 

-// fail stops the loop with err.
-func (l *TurnLoop) fail(err error) {
-       l.mutex.Lock()
-       l.failure = errors.Join(l.failure, err)
-       l.mutex.Unlock()
-
-       select {
-       case l.wake <- struct{}{}:
-       default:
-       }
-}
-

 // failed returns the error of a failed write of a report.
 func (l *TurnLoop) failed() error {
        l.mutex.Lock()
 // failed returns the error of a failed write of a report.
 func (l *TurnLoop) failed() error {
        l.mutex.Lock()


@@ -298,19 +338,26 @@ func (l *TurnLoop) failed() error {
 }
 
 // turn runs one turn. A failed or interrupted model call ends the turn and
 }
 
 // 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.
+// tells the clients. A failure also goes to the logger, so that it is not lost
+// when no client is connected, but an interrupt does not. 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 or queue write is an error of turn. A daemon stop leaves taken
+// parent and child messages queued for restart. A user interrupt completes them
+// without a report.

 func (l *TurnLoop) turn(ctx context.Context) error {
 func (l *TurnLoop) turn(ctx context.Context) error {

+       runCtx := ctx

        ctx, cancel := context.WithCancel(ctx)
        defer cancel()
 
        l.mutex.Lock()
        ctx, cancel := context.WithCancel(ctx)
        defer cancel()
 
        l.mutex.Lock()

+

        l.cancel = cancel
        l.mutex.Unlock()
 
        defer func() {
                l.mutex.Lock()
        l.cancel = cancel
        l.mutex.Unlock()
 
        defer func() {
                l.mutex.Lock()

+

                l.cancel = nil
                l.mutex.Unlock()
        }()
                l.cancel = nil
                l.mutex.Unlock()
        }()


@@ -319,8 +366,24 @@ func (l *TurnLoop) turn(ctx context.Context) error {
                Kind: "busy",
        })
 
                Kind: "busy",
        })
 

-       err := l.chain(ctx)
+       if l.messages != nil {
+               l.logger.Info("turn started")
+       }
+
+       reply, err := l.chain(ctx)
+       if l.messages != nil && !errors.Is(err, errChainLogWrite) && (err == nil || runCtx.Err() == nil) {
+               finishErr := l.messages.Finish(l, l.active, reply)
+               if finishErr != nil {
+                       err = errors.Join(err, fmt.Errorf("%w: finish agent messages: %w", errChainLogWrite, finishErr))
+               }
+       }
+
+       l.active = nil

        if err != nil && !errors.Is(err, errChainLogWrite) {
        if err != nil && !errors.Is(err, errChainLogWrite) {

+               if ctx.Err() == nil {
+                       l.logger.Error("turn failed", "err", err)
+               }
+

                l.notify(Event{
                        Kind: "error",
                        Text: turnError(ctx, err),
                l.notify(Event{
                        Kind: "error",
                        Text: turnError(ctx, err),


@@ -329,6 +392,10 @@ func (l *TurnLoop) turn(ctx context.Context) error {
                err = nil
        }
 
                err = nil
        }
 

+       if l.messages != nil {
+               l.logger.Info("turn done")
+       }
+

        l.notify(Event{
                Kind: "idle",
        })
        l.notify(Event{
                Kind: "idle",
        })


@@ -337,20 +404,20 @@ func (l *TurnLoop) turn(ctx context.Context) error {
 }
 
 // chain runs the request chain of a turn until the final answer.
 }
 
 // chain runs the request chain of a turn until the final answer.

-func (l *TurnLoop) chain(ctx context.Context) error {
+func (l *TurnLoop) chain(ctx context.Context) (string, error) {

        messages := l.take()
        if len(messages) == 0 {
        messages := l.take()
        if len(messages) == 0 {

-               return nil
+               return "", nil

        }
 
        err := l.memory.Wait(ctx)
        if err != nil {
        }
 
        err := l.memory.Wait(ctx)
        if err != nil {

-               return err
+               return "", fmt.Errorf("wait for turn memory: %w", err)

        }
 
        instructions, err := l.agent.Prompt()
        if err != nil {
        }
 
        instructions, err := l.agent.Prompt()
        if err != nil {

-               return err
+               return "", fmt.Errorf("read turn prompt: %w", err)

        }
 
        skip := make(map[int]bool, len(messages))
        }
 
        skip := make(map[int]bool, len(messages))


@@ -360,7 +427,7 @@ func (l *TurnLoop) chain(ctx context.Context) error {
 
        view, err := l.layout.Split(l.memory.ViewLines(skip))
        if err != nil {
 
        view, err := l.layout.Split(l.memory.ViewLines(skip))
        if err != nil {

-               return err
+               return "", fmt.Errorf("split turn view: %w", err)

        }
 
        chain := &Chain{
        }
 
        chain := &Chain{


@@ -370,14 +437,26 @@ func (l *TurnLoop) chain(ctx context.Context) error {
                        Model:        l.agent.config.Model,
                        Effort:       l.agent.config.Effort,
                        Instructions: instructions,
                        Model:        l.agent.config.Model,
                        Effort:       l.agent.config.Effort,
                        Instructions: instructions,

-                       Tools:        promptTools(),
+                       Tools:        l.agent.Tools(),

                },
                },

-               take: l.take,
+               take:   l.take,
+               images: l.agent.Images(),

        }
 
        }
 

-       _, err = chain.Run(ctx, turnInput(view, l.state(), messages))
+       input := turnInput(view, l.state(), messages)
+       for n, message := range messages {
+               input[len(input)-len(messages)+n], err = imageUserMessage(l.agent.Images(), message.Text)
+               if err != nil {
+                       return "", fmt.Errorf("read turn images: %w", err)
+               }
+       }

 
 

-       return err
+       reply, err := chain.Run(ctx, input)
+       if err != nil {
+               return "", fmt.Errorf("run turn chain: %w", err)
+       }
+
+       return reply, nil

 }
 
 // take returns the user messages that wait for a request, and removes them.
 }
 
 // take returns the user messages that wait for a request, and removes them.


@@ -387,16 +466,13 @@ func (l *TurnLoop) take() []Message {
 
        messages := l.pending
        l.pending = nil
 
        messages := l.pending
        l.pending = nil

+       if l.messages != nil {
+               l.active = append(l.active, messages...)
+       }

 
        return messages
 }
 
 
        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")
-}
-

 // turnError returns the text of the error that ended a turn.
 func turnError(ctx context.Context, err error) string {
        if ctx.Err() != nil {
 // turnError returns the text of the error that ended a turn.
 func turnError(ctx context.Context, err error) string {
        if ctx.Err() != nil {