git.giacomolli.uk - onechat/blob - turn.go

   1 package main
   2
   3 import (
   4         "context"
   5         "errors"
   6         "fmt"
   7         "path/filepath"
   8         "sync"
   9         "time"
  10 )
  11
  12 // newTurnLoop returns the loop of agent with its memory, its subagents, and the
  13 // WhatsApp link of its tools. Its calls go through client, and notify gets each
  14 // event. The clock gives the date of a turn.
  15 func newTurnLoop(agent *Agent, memory *Memory, subagents *Subagents, whatsapp *WhatsApp, client ChainClient, clock func() time.Time, notify func(Event)) (*TurnLoop, error) {
  16         layout, err := loadViewLayout(filepath.Join(agent.dir, viewLayoutFileName))
  17         if err != nil {
  18                 return nil, err
  19         }
  20
  21         l := &TurnLoop{
  22                 agent:  agent,
  23                 memory: memory,
  24                 client: client,
  25                 layout: layout,
  26                 clock:  clock,
  27                 notify: notify,
  28                 wake:   make(chan struct{}, 1),
  29         }
  30
  31         l.runs = &SubagentRuns{
  32                 agent:     agent,
  33                 memory:    memory,
  34                 subagents: subagents,
  35                 layout:    layout,
  36                 client:    client,
  37                 state:     l.state,
  38                 report:    l.Report,
  39                 note:      l.note,
  40                 fail:      l.fail,
  41                 notify:    notify,
  42                 running:   make(map[string]*SubagentRun),
  43         }
  44
  45         l.tools = &Toolbox{
  46                 log:       l,
  47                 chat:      memory,
  48                 subagents: subagents,
  49                 runs:      l.runs,
  50                 whatsapp:  whatsapp,
  51                 workdir:   agent.config.Workdir,
  52         }
  53
  54         return l, nil
  55 }
  56
  57 // Event tells the clients what the agent does as it happens. Kind is the kind
  58 // of a logged message (user, reply, tool, echo, note, work), or one of ready,
  59 // busy, idle, error, and subagent, which tells that the subagent Text starts a
  60 // run. ID is the ID of a logged message. Source names the client that sent a
  61 // user message, such as terminal or telegram, so that each client can show the
  62 // messages of the others, or the subagent of a report. Status tells how the run
  63 // of a report ended: finished, stopped, or failed.
  64 type Event struct {
  65         Kind   string `json:"kind"`
  66         ID     int    `json:"id,omitempty"`
  67         Text   string `json:"text,omitempty"`
  68         Source string `json:"source,omitempty"`
  69         Status string `json:"status,omitempty"`
  70 }
  71
  72 // TurnLoop runs the turns of one agent. A user message goes to the log when it
  73 // comes, and it wakes the loop. A turn waits until every message before it is
  74 // summarized. Then the turn runs a Chain whose first request holds the prompt,
  75 // the tools, and the input of turnInput.
  76 //
  77 // The view goes out as the messages of a ViewLayout, so the first request of a
  78 // turn reads the view of the turn before from the prompt cache. User messages
  79 // that come during a turn join the next request of its chain after the tool
  80 // results, and so do the reports of the subagents. The subagents run until the
  81 // loop ends.
  82 type TurnLoop struct {
  83         agent  *Agent
  84         memory *Memory
  85         client ChainClient
  86         tools  *Toolbox
  87         runs   *SubagentRuns
  88         layout *ViewLayout
  89         clock  func() time.Time
  90         notify func(Event)
  91         wake   chan struct{}
  92
  93         mutex   sync.Mutex
  94         pending []Message
  95         cancel  context.CancelFunc
  96         failure error
  97 }
  98
  99 // Submit logs a user message from the client source and wakes the loop. The
 100 // message joins the next request: the first one of a new turn, or the next one
 101 // of a running turn.
 102 func (l *TurnLoop) Submit(source, text string) (Message, error) {
 103         message, err := l.record("user", source, text)
 104         if err != nil {
 105                 return Message{}, err
 106         }
 107
 108         l.mutex.Lock()
 109         l.pending = append(l.pending, message)
 110         l.mutex.Unlock()
 111
 112         select {
 113         case l.wake <- struct{}{}:
 114         default:
 115         }
 116
 117         return message, nil
 118 }
 119
 120 // Report logs the report text of the subagent name as a work message and tells
 121 // the clients, with the status of the run. Unless the subagent was stopped, the
 122 // report joins the waiting messages and wakes the loop, as a user message of
 123 // Submit does. A failed write stops the loop, because the log refuses all
 124 // writes after it.
 125 func (l *TurnLoop) Report(name, text, status string) {
 126         message, err := l.Work(name, text, status)
 127         switch {
 128         case err != nil:
 129                 l.fail(fmt.Errorf("%w: %w", errChainLogWrite, err))
 130         case status != subagentStopped:
 131                 l.Wake(message)
 132         }
 133 }
 134
 135 // Work logs text as a work message of source, such as the name of a subagent,
 136 // and tells the clients, with status. The message starts no turn until it goes
 137 // to Wake.
 138 func (l *TurnLoop) Work(source, text, status string) (Message, error) {
 139         message, err := l.memory.Append("work", text)
 140         if err != nil {
 141                 return Message{}, err
 142         }
 143
 144         l.notify(Event{
 145                 Kind:   "work",
 146                 ID:     message.ID,
 147                 Text:   text,
 148                 Source: source,
 149                 Status: status,
 150         })
 151
 152         return message, nil
 153 }
 154
 155 // Wake adds the logged messages to the waiting messages and wakes the loop
 156 // once, as a user message of Submit does. They join the first request of a new
 157 // turn together, or the next request of a running turn after the tool results.
 158 func (l *TurnLoop) Wake(messages ...Message) {
 159         if len(messages) == 0 {
 160                 return
 161         }
 162
 163         l.mutex.Lock()
 164         l.pending = append(l.pending, messages...)
 165         l.mutex.Unlock()
 166
 167         select {
 168         case l.wake <- struct{}{}:
 169         default:
 170         }
 171 }
 172
 173 // StopSubagent ends the run of the subagent name and tells if it ran. Its
 174 // report is logged, but it starts no turn.
 175 func (l *TurnLoop) StopSubagent(name string) (bool, error) {
 176         return l.runs.Stop(name)
 177 }
 178
 179 // Interrupt ends the running turn. A running tool is killed, and the turn sends
 180 // no further request.
 181 func (l *TurnLoop) Interrupt() bool {
 182         l.mutex.Lock()
 183         defer l.mutex.Unlock()
 184
 185         if l.cancel == nil {
 186                 return false
 187         }
 188
 189         l.cancel()
 190
 191         return true
 192 }
 193
 194 // Busy tells if a turn runs.
 195 func (l *TurnLoop) Busy() bool {
 196         l.mutex.Lock()
 197         defer l.mutex.Unlock()
 198
 199         return l.cancel != nil
 200 }
 201
 202 // Append logs a message and tells the clients. The toolbox logs its calls and
 203 // results through it.
 204 func (l *TurnLoop) Append(kind, text string) (Message, error) {
 205         return l.record(kind, "", text)
 206 }
 207
 208 // Flush makes the log durable.
 209 func (l *TurnLoop) Flush(ctx context.Context) error {
 210         return l.memory.Flush(ctx)
 211 }
 212
 213 // Run runs a turn each time user messages or reports wait, until ctx ends. It
 214 // returns the error of a failed log write, which stops the agent. At its start
 215 // it resumes the subagents that a stop of the daemon cut, and at its end it
 216 // ends the runs of the subagents and waits for them.
 217 func (l *TurnLoop) Run(ctx context.Context) error {
 218         defer l.runs.Close()
 219
 220         err := l.runs.Resume(ctx)
 221         if err != nil {
 222                 return fmt.Errorf("resume subagents: %w", err)
 223         }
 224
 225         for {
 226                 for l.waiting() {
 227                         err := l.turn(ctx)
 228                         if err != nil {
 229                                 return err
 230                         }
 231                 }
 232
 233                 select {
 234                 case <-l.wake:
 235                 case <-ctx.Done():
 236                         return ctx.Err()
 237                 }
 238
 239                 err := l.failed()
 240                 if err != nil {
 241                         return err
 242                 }
 243         }
 244 }
 245
 246 // note logs a note and tells the clients. A note starts no turn.
 247 func (l *TurnLoop) note(text string) error {
 248         _, err := l.record("note", "", text)
 249
 250         return err
 251 }
 252
 253 // record logs a message and tells the clients, with the client source of a user
 254 // message.
 255 func (l *TurnLoop) record(kind, source, text string) (Message, error) {
 256         message, err := l.memory.Append(kind, text)
 257         if err != nil {
 258                 return Message{}, err
 259         }
 260
 261         l.notify(Event{
 262                 Kind:   kind,
 263                 ID:     message.ID,
 264                 Text:   text,
 265                 Source: source,
 266         })
 267
 268         return message, nil
 269 }
 270
 271 // waiting tells if user messages or reports wait for a request.
 272 func (l *TurnLoop) waiting() bool {
 273         l.mutex.Lock()
 274         defer l.mutex.Unlock()
 275
 276         return len(l.pending) > 0
 277 }
 278
 279 // fail stops the loop with err.
 280 func (l *TurnLoop) fail(err error) {
 281         l.mutex.Lock()
 282         l.failure = errors.Join(l.failure, err)
 283         l.mutex.Unlock()
 284
 285         select {
 286         case l.wake <- struct{}{}:
 287         default:
 288         }
 289 }
 290
 291 // failed returns the error of a failed write of a report.
 292 func (l *TurnLoop) failed() error {
 293         l.mutex.Lock()
 294         defer l.mutex.Unlock()
 295
 296         return l.failure
 297 }
 298
 299 // turn runs one turn. A failed or interrupted model call ends the turn and
 300 // tells the clients. The messages of the turn stay in the log and the view, and
 301 // the next turn starts only with a new message, so a lasting failure such as a
 302 // needed login does not repeat. Only a failed log write is an error of turn.
 303 func (l *TurnLoop) turn(ctx context.Context) error {
 304         ctx, cancel := context.WithCancel(ctx)
 305         defer cancel()
 306
 307         l.mutex.Lock()
 308         l.cancel = cancel
 309         l.mutex.Unlock()
 310
 311         defer func() {
 312                 l.mutex.Lock()
 313                 l.cancel = nil
 314                 l.mutex.Unlock()
 315         }()
 316
 317         l.notify(Event{
 318                 Kind: "busy",
 319         })
 320
 321         err := l.chain(ctx)
 322         if err != nil && !errors.Is(err, errChainLogWrite) {
 323                 l.notify(Event{
 324                         Kind: "error",
 325                         Text: turnError(ctx, err),
 326                 })
 327
 328                 err = nil
 329         }
 330
 331         l.notify(Event{
 332                 Kind: "idle",
 333         })
 334
 335         return err
 336 }
 337
 338 // chain runs the request chain of a turn until the final answer.
 339 func (l *TurnLoop) chain(ctx context.Context) error {
 340         messages := l.take()
 341         if len(messages) == 0 {
 342                 return nil
 343         }
 344
 345         err := l.memory.Wait(ctx)
 346         if err != nil {
 347                 return err
 348         }
 349
 350         instructions, err := l.agent.Prompt()
 351         if err != nil {
 352                 return err
 353         }
 354
 355         skip := make(map[int]bool, len(messages))
 356         for _, message := range messages {
 357                 skip[message.ID] = true
 358         }
 359
 360         view, err := l.layout.Split(l.memory.ViewLines(skip))
 361         if err != nil {
 362                 return err
 363         }
 364
 365         chain := &Chain{
 366                 client: l.client,
 367                 tools:  l.tools,
 368                 request: ModelRequest{
 369                         Model:        l.agent.config.Model,
 370                         Effort:       l.agent.config.Effort,
 371                         Instructions: instructions,
 372                         Tools:        promptTools(),
 373                 },
 374                 take: l.take,
 375         }
 376
 377         _, err = chain.Run(ctx, turnInput(view, l.state(), messages))
 378
 379         return err
 380 }
 381
 382 // take returns the user messages that wait for a request, and removes them.
 383 func (l *TurnLoop) take() []Message {
 384         l.mutex.Lock()
 385         defer l.mutex.Unlock()
 386
 387         messages := l.pending
 388         l.pending = nil
 389
 390         return messages
 391 }
 392
 393 // state returns the per-turn state: the end of the view, and the date and time
 394 // of the turn.
 395 func (l *TurnLoop) state() string {
 396         return "</chat>\n\nNow: " + l.clock().Format("Monday 2006-01-02 15:04 -07:00")
 397 }
 398
 399 // turnError returns the text of the error that ended a turn.
 400 func turnError(ctx context.Context, err error) string {
 401         if ctx.Err() != nil {
 402                 return "interrupted"
 403         }
 404
 405         return err.Error()
 406 }
 407
 408 // turnInput returns the input of the first request of a turn, stable parts
 409 // first: a user message for each view message, one with the per-turn state,
 410 // and one for each new message, whole. The view leaves out the leaf lines of
 411 // the new messages.
 412 func turnInput(view []string, state string, messages []Message) []any {
 413         input := make([]any, 0, len(view)+1+len(messages))
 414         for _, text := range view {
 415                 input = append(input, modelUserMessage(text))
 416         }
 417
 418         input = append(input, modelUserMessage(state))
 419         for _, message := range messages {
 420                 input = append(input, modelUserMessage(message.Text))
 421         }
 422
 423         return input
 424 }