import (
"context"
"errors"
+ "fmt"
+ "log/slog"
"path/filepath"
"sync"
"time"
)
-// 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 ChainClient, clock func() time.Time, notify func(Event)) (*TurnLoop, 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,
+// 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
client: client,
layout: layout,
clock: clock,
+ logger: logger,
notify: notify,
wake: make(chan struct{}, 1),
}
+ l.runs = &SubagentRuns{
+ agent: agent,
+ memory: memory,
+ subagents: subagents,
+ layout: layout,
+ client: client,
+ state: l.state,
+ report: l.Report,
+ note: l.note,
+ fail: l.fail,
+ logger: logger,
+ notify: notify,
+ running: make(map[string]*SubagentRun),
+ }
+
+ if agent.config.Parent != "" {
+ whatsapp = nil
+ }
+
l.tools = &Toolbox{
- log: l,
- chat: memory,
- workdir: agent.config.Workdir,
+ log: l,
+ chat: memory,
+ subagents: subagents,
+ runs: l.runs,
+ whatsapp: whatsapp,
+ schedules: schedules,
+ workdir: agent.config.Workdir,
+ agentName: agent.config.Name,
}
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), 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.
+// 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
// 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.
+// 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
- layout *ViewLayout
- clock func() time.Time
- notify func(Event)
- wake chan 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
}
// Submit logs a user message from the client source and wakes the loop. The
}
l.mutex.Lock()
+
l.pending = append(l.pending, message)
l.mutex.Unlock()
return message, nil
}
+// Report logs the report text of the subagent name as a work message and tells
+// the clients, with the status of the run. Unless the subagent was stopped, the
+// report joins the waiting messages and wakes the loop, as a user message of
+// Submit does. A failed write stops the loop, because the log refuses all
+// writes after it.
+func (l *TurnLoop) Report(name, text, status string) {
+ message, err := l.Work(name, text, status)
+ switch {
+ case err != nil:
+ l.fail(fmt.Errorf("%w: %w", errChainLogWrite, err))
+ case status != subagentStopped:
+ l.Wake(message)
+ }
+}
+
+// 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.
+func (l *TurnLoop) Work(source, text, status string) (Message, error) {
+ message, err := l.memory.Append("work", text)
+ if err != nil {
+ return Message{}, err
+ }
+
+ l.notify(Event{
+ Kind: "work",
+ ID: message.ID,
+ Text: text,
+ Source: source,
+ Status: status,
+ })
+
+ return message, nil
+}
+
+// Wake adds the logged messages to the waiting messages and wakes the loop
+// once, as a user message of Submit does. They join the first request of a new
+// turn together, or the next request of a running turn after the tool results.
+func (l *TurnLoop) Wake(messages ...Message) {
+ if len(messages) == 0 {
+ return
+ }
+
+ l.mutex.Lock()
+
+ l.pending = append(l.pending, messages...)
+ l.mutex.Unlock()
+
+ select {
+ case l.wake <- struct{}{}:
+ default:
+ }
+}
+
+// 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) {
+ return l.runs.Stop(name)
+}
+
// Interrupt ends the running turn. A running tool is killed, and the turn sends
// no further request.
func (l *TurnLoop) Interrupt() bool {
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.
+// Run runs a turn each time user messages or reports wait, until ctx ends. It
+// returns the error of a failed log write, which stops the agent. At its start
+// it resumes the subagents that a stop of the daemon cut, and at its end it
+// ends the runs of the subagents and waits for them.
func (l *TurnLoop) Run(ctx context.Context) error {
+ defer l.runs.Close()
+
+ err := l.runs.Resume(ctx)
+ if err != nil {
+ return fmt.Errorf("resume subagents: %w", err)
+ }
+
for {
- for l.waiting() {
- err := l.turn(ctx)
+ for ctx.Err() == nil && l.waiting() {
+ err := l.failed()
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)
}
}
case <-ctx.Done():
return ctx.Err()
}
+
+ err := l.failed()
+ if err != nil {
+ 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) {
return message, nil
}
-// waiting tells if user messages wait for a request.
+// waiting tells if user messages or reports wait for a request.
func (l *TurnLoop) waiting() bool {
l.mutex.Lock()
defer l.mutex.Unlock()
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()
+ defer l.mutex.Unlock()
+
+ return l.failure
+}
+
// 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 {
+ runCtx := ctx
ctx, cancel := context.WithCancel(ctx)
defer cancel()
l.mutex.Lock()
+
l.cancel = cancel
l.mutex.Unlock()
defer func() {
l.mutex.Lock()
+
l.cancel = nil
l.mutex.Unlock()
}()
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 ctx.Err() == nil {
+ l.logger.Error("turn failed", "err", err)
+ }
+
l.notify(Event{
Kind: "error",
Text: turnError(ctx, err),
err = nil
}
+ if l.messages != nil {
+ l.logger.Info("turn done")
+ }
+
l.notify(Event{
Kind: "idle",
})
}
// 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 {
- return nil
+ return "", 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 {
- return err
+ return "", fmt.Errorf("read turn prompt: %w", err)
}
skip := make(map[int]bool, len(messages))
view, err := l.layout.Split(l.memory.ViewLines(skip))
if err != nil {
- return err
+ return "", fmt.Errorf("split turn view: %w", err)
}
chain := &Chain{
Model: l.agent.config.Model,
Effort: l.agent.config.Effort,
Instructions: instructions,
- Tools: promptTools(),
+ Tools: l.agent.Tools(),
},
take: l.take,
}
- _, err = chain.Run(ctx, turnInput(view, l.state(), messages))
+ reply, err := chain.Run(ctx, turnInput(view, l.state(), messages))
+ if err != nil {
+ return "", fmt.Errorf("run turn chain: %w", err)
+ }
- return err
+ return reply, nil
}
// take returns the user messages that wait for a request, and removes them.
messages := l.pending
l.pending = nil
+ if l.messages != nil {
+ l.active = append(l.active, messages...)
+ }
return messages
}
// state returns the per-turn state: the end of the view, and the date and time
-// of the turn.
+// of the turn, then the names and responsibilities of its children.
func (l *TurnLoop) state() string {
- return "</chat>\n\nNow: " + l.clock().Format("Monday 2006-01-02 15:04 -07:00")
+ 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 state
}
// turnError returns the text of the error that ended a turn.
projects / onechat / blobdiff