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

   1 package main
   2
   3 import (
   4         "context"
   5         "errors"
   6         "fmt"
   7         "log/slog"
   8         "path/filepath"
   9         "sync"
  10         "time"
  11 )
  12
  13 // newTurnLoop returns the loop of agent with its memory, its subagents, and the
  14 // WhatsApp link and the schedules of its tools. Its calls go through client,
  15 // and notify gets each event. The clock gives the date of a turn. logger gets
  16 // the failed turns and the ends of the subagents. A child has no WhatsApp
  17 // tools, even when the caller supplies a link.
  18 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) {
  19         layout, err := loadViewLayout(filepath.Join(agent.dir, viewLayoutFileName))
  20         if err != nil {
  21                 return nil, err
  22         }
  23
  24         l := &TurnLoop{
  25                 agent:  agent,
  26                 memory: memory,
  27                 client: client,
  28                 layout: layout,
  29                 clock:  clock,
  30                 logger: logger,
  31                 notify: notify,
  32                 wake:   make(chan struct{}, 1),
  33         }
  34
  35         l.runs = &SubagentRuns{
  36                 agent:     agent,
  37                 memory:    memory,
  38                 subagents: subagents,
  39                 layout:    layout,
  40                 client:    client,
  41                 state:     l.state,
  42                 report:    l.Report,
  43                 note:      l.note,
  44                 fail:      l.fail,
  45                 logger:    logger,
  46                 notify:    notify,
  47                 running:   make(map[string]*SubagentRun),
  48         }
  49
  50         if agent.config.Parent != "" {
  51                 whatsapp = nil
  52         }
  53
  54         l.tools = &Toolbox{
  55                 log:       l,
  56                 chat:      memory,
  57                 subagents: subagents,
  58                 runs:      l.runs,
  59                 whatsapp:  whatsapp,
  60                 schedules: schedules,
  61                 workdir:   agent.config.Workdir,
  62         }
  63
  64         return l, nil
  65 }
  66
  67 // Event tells the clients what the agent does as it happens. Kind is the kind
  68 // of a logged message (user, reply, tool, echo, note, work), or one of ready,
  69 // busy, idle, error, and subagent, which tells that the subagent Text starts a
  70 // run. ID is the ID of a logged message. Source names the client that sent a
  71 // user message, such as terminal or telegram, so that each client can show the
  72 // messages of the others, or the subagent of a report. Status tells how the run
  73 // of a report ended: finished, stopped, or failed.
  74 type Event struct {
  75         Kind   string `json:"kind"`
  76         ID     int    `json:"id,omitempty"`
  77         Text   string `json:"text,omitempty"`
  78         Source string `json:"source,omitempty"`
  79         Status string `json:"status,omitempty"`
  80 }
  81
  82 // TurnLoop runs the turns of one agent. A user message goes to the log when it
  83 // comes, and it wakes the loop. A turn waits until every message before it is
  84 // summarized. Then the turn runs a Chain whose first request holds the prompt,
  85 // the tools, and the input of turnInput.
  86 //
  87 // The view goes out as the messages of a ViewLayout, so the first request of a
  88 // turn reads the view of the turn before from the prompt cache. User messages
  89 // that come during a turn join the next request of its chain after the tool
  90 // results, and so do the reports of the subagents. The subagents run until the
  91 // loop ends. Parent and child messages stay queued until the turn finishes.
  92 // active belongs to the running turn. The mutex protects pending, cancel, and
  93 // failure.
  94 type TurnLoop struct {
  95         agent    *Agent
  96         memory   *Memory
  97         client   ChainClient
  98         tools    *Toolbox
  99         runs     *SubagentRuns
 100         layout   *ViewLayout
 101         clock    func() time.Time
 102         logger   *slog.Logger
 103         notify   func(Event)
 104         wake     chan struct{}
 105         messages *DaemonMessages
 106         active   []Message
 107
 108         mutex   sync.Mutex
 109         pending []Message
 110         cancel  context.CancelFunc
 111         failure error
 112 }
 113
 114 // Submit logs a user message from the client source and wakes the loop. The
 115 // message joins the next request: the first one of a new turn, or the next one
 116 // of a running turn.
 117 func (l *TurnLoop) Submit(source, text string) (Message, error) {
 118         message, err := l.record("user", source, text)
 119         if err != nil {
 120                 return Message{}, err
 121         }
 122
 123         l.mutex.Lock()
 124
 125         l.pending = append(l.pending, message)
 126         l.mutex.Unlock()
 127
 128         select {
 129         case l.wake <- struct{}{}:
 130         default:
 131         }
 132
 133         return message, nil
 134 }
 135
 136 // Report logs the report text of the subagent name as a work message and tells
 137 // the clients, with the status of the run. Unless the subagent was stopped, the
 138 // report joins the waiting messages and wakes the loop, as a user message of
 139 // Submit does. A failed write stops the loop, because the log refuses all
 140 // writes after it.
 141 func (l *TurnLoop) Report(name, text, status string) {
 142         message, err := l.Work(name, text, status)
 143         switch {
 144         case err != nil:
 145                 l.fail(fmt.Errorf("%w: %w", errChainLogWrite, err))
 146         case status != subagentStopped:
 147                 l.Wake(message)
 148         }
 149 }
 150
 151 // Work logs text as a work message of source, such as the name of a subagent,
 152 // and tells the clients, with status. The message starts no turn until it goes
 153 // to Wake.
 154 func (l *TurnLoop) Work(source, text, status string) (Message, error) {
 155         message, err := l.memory.Append("work", text)
 156         if err != nil {
 157                 return Message{}, err
 158         }
 159
 160         l.notify(Event{
 161                 Kind:   "work",
 162                 ID:     message.ID,
 163                 Text:   text,
 164                 Source: source,
 165                 Status: status,
 166         })
 167
 168         return message, nil
 169 }
 170
 171 // Wake adds the logged messages to the waiting messages and wakes the loop
 172 // once, as a user message of Submit does. They join the first request of a new
 173 // turn together, or the next request of a running turn after the tool results.
 174 func (l *TurnLoop) Wake(messages ...Message) {
 175         if len(messages) == 0 {
 176                 return
 177         }
 178
 179         l.mutex.Lock()
 180
 181         l.pending = append(l.pending, messages...)
 182         l.mutex.Unlock()
 183
 184         select {
 185         case l.wake <- struct{}{}:
 186         default:
 187         }
 188 }
 189
 190 // StopSubagent ends the run of the subagent name and tells if it ran. Its
 191 // report is logged, but it starts no turn.
 192 func (l *TurnLoop) StopSubagent(name string) (bool, error) {
 193         return l.runs.Stop(name)
 194 }
 195
 196 // Interrupt ends the running turn. A running tool is killed, and the turn sends
 197 // no further request.
 198 func (l *TurnLoop) Interrupt() bool {
 199         l.mutex.Lock()
 200         defer l.mutex.Unlock()
 201
 202         if l.cancel == nil {
 203                 return false
 204         }
 205
 206         l.cancel()
 207
 208         return true
 209 }
 210
 211 // Busy tells if a turn runs.
 212 func (l *TurnLoop) Busy() bool {
 213         l.mutex.Lock()
 214         defer l.mutex.Unlock()
 215
 216         return l.cancel != nil
 217 }
 218
 219 // Append logs a message and tells the clients. The toolbox logs its calls and
 220 // results through it.
 221 func (l *TurnLoop) Append(kind, text string) (Message, error) {
 222         return l.record(kind, "", text)
 223 }
 224
 225 // Flush makes the log durable.
 226 func (l *TurnLoop) Flush(ctx context.Context) error {
 227         return l.memory.Flush(ctx)
 228 }
 229
 230 // Run runs a turn each time user messages or reports wait, until ctx ends. It
 231 // returns the error of a failed log write, which stops the agent. At its start
 232 // it resumes the subagents that a stop of the daemon cut, and at its end it
 233 // ends the runs of the subagents and waits for them.
 234 func (l *TurnLoop) Run(ctx context.Context) error {
 235         defer l.runs.Close()
 236
 237         err := l.runs.Resume(ctx)
 238         if err != nil {
 239                 return fmt.Errorf("resume subagents: %w", err)
 240         }
 241
 242         for {
 243                 for ctx.Err() == nil && l.waiting() {
 244                         err := l.failed()
 245                         if err != nil {
 246                                 return fmt.Errorf("check turn loop failure: %w", err)
 247                         }
 248
 249                         err = l.turn(ctx)
 250                         if err != nil {
 251                                 return fmt.Errorf("run turn: %w", err)
 252                         }
 253                 }
 254
 255                 select {
 256                 case <-l.wake:
 257                 case <-ctx.Done():
 258                         return ctx.Err()
 259                 }
 260
 261                 err := l.failed()
 262                 if err != nil {
 263                         return fmt.Errorf("check turn loop failure: %w", err)
 264                 }
 265         }
 266 }
 267
 268 // note logs a note and tells the clients. A note starts no turn.
 269 func (l *TurnLoop) note(text string) error {
 270         _, err := l.record("note", "", text)
 271
 272         return err
 273 }
 274
 275 // record logs a message and tells the clients, with the client source of a user
 276 // message.
 277 func (l *TurnLoop) record(kind, source, text string) (Message, error) {
 278         message, err := l.memory.Append(kind, text)
 279         if err != nil {
 280                 return Message{}, err
 281         }
 282
 283         l.notify(Event{
 284                 Kind:   kind,
 285                 ID:     message.ID,
 286                 Text:   text,
 287                 Source: source,
 288         })
 289
 290         return message, nil
 291 }
 292
 293 // waiting tells if user messages or reports wait for a request.
 294 func (l *TurnLoop) waiting() bool {
 295         l.mutex.Lock()
 296         defer l.mutex.Unlock()
 297
 298         return len(l.pending) > 0
 299 }
 300
 301 // fail stops the loop with err.
 302 func (l *TurnLoop) fail(err error) {
 303         l.mutex.Lock()
 304
 305         l.failure = errors.Join(l.failure, err)
 306         l.mutex.Unlock()
 307
 308         select {
 309         case l.wake <- struct{}{}:
 310         default:
 311         }
 312 }
 313
 314 // failed returns the error of a failed write of a report.
 315 func (l *TurnLoop) failed() error {
 316         l.mutex.Lock()
 317         defer l.mutex.Unlock()
 318
 319         return l.failure
 320 }
 321
 322 // turn runs one turn. A failed or interrupted model call ends the turn and
 323 // tells the clients. A failure also goes to the logger, so that it is not lost
 324 // when no client is connected, but an interrupt does not. The messages of the
 325 // turn stay in the log and the view, and the next turn starts only with a new
 326 // message, so a lasting failure such as a needed login does not repeat. Only a
 327 // failed log or queue write is an error of turn. A daemon stop leaves taken
 328 // parent and child messages queued for restart. A user interrupt completes them
 329 // without a report.
 330 func (l *TurnLoop) turn(ctx context.Context) error {
 331         runCtx := ctx
 332         ctx, cancel := context.WithCancel(ctx)
 333         defer cancel()
 334
 335         l.mutex.Lock()
 336
 337         l.cancel = cancel
 338         l.mutex.Unlock()
 339
 340         defer func() {
 341                 l.mutex.Lock()
 342
 343                 l.cancel = nil
 344                 l.mutex.Unlock()
 345         }()
 346
 347         l.notify(Event{
 348                 Kind: "busy",
 349         })
 350
 351         if l.messages != nil {
 352                 l.logger.Info("turn started")
 353         }
 354
 355         reply, err := l.chain(ctx)
 356         if l.messages != nil && !errors.Is(err, errChainLogWrite) && (err == nil || runCtx.Err() == nil) {
 357                 finishErr := l.messages.Finish(l, l.active, reply)
 358                 if finishErr != nil {
 359                         err = errors.Join(err, fmt.Errorf("%w: finish agent messages: %w", errChainLogWrite, finishErr))
 360                 }
 361         }
 362
 363         l.active = nil
 364         if err != nil && !errors.Is(err, errChainLogWrite) {
 365                 if ctx.Err() == nil {
 366                         l.logger.Error("turn failed", "err", err)
 367                 }
 368
 369                 l.notify(Event{
 370                         Kind: "error",
 371                         Text: turnError(ctx, err),
 372                 })
 373
 374                 err = nil
 375         }
 376
 377         if l.messages != nil {
 378                 l.logger.Info("turn done")
 379         }
 380
 381         l.notify(Event{
 382                 Kind: "idle",
 383         })
 384
 385         return err
 386 }
 387
 388 // chain runs the request chain of a turn until the final answer.
 389 func (l *TurnLoop) chain(ctx context.Context) (string, error) {
 390         messages := l.take()
 391         if len(messages) == 0 {
 392                 return "", nil
 393         }
 394
 395         err := l.memory.Wait(ctx)
 396         if err != nil {
 397                 return "", fmt.Errorf("wait for turn memory: %w", err)
 398         }
 399
 400         instructions, err := l.agent.Prompt()
 401         if err != nil {
 402                 return "", fmt.Errorf("read turn prompt: %w", err)
 403         }
 404
 405         skip := make(map[int]bool, len(messages))
 406         for _, message := range messages {
 407                 skip[message.ID] = true
 408         }
 409
 410         view, err := l.layout.Split(l.memory.ViewLines(skip))
 411         if err != nil {
 412                 return "", fmt.Errorf("split turn view: %w", err)
 413         }
 414
 415         chain := &Chain{
 416                 client: l.client,
 417                 tools:  l.tools,
 418                 request: ModelRequest{
 419                         Model:        l.agent.config.Model,
 420                         Effort:       l.agent.config.Effort,
 421                         Instructions: instructions,
 422                         Tools:        l.agent.Tools(),
 423                 },
 424                 take: l.take,
 425         }
 426
 427         reply, err := chain.Run(ctx, turnInput(view, l.state(), messages))
 428         if err != nil {
 429                 return "", fmt.Errorf("run turn chain: %w", err)
 430         }
 431
 432         return reply, nil
 433 }
 434
 435 // take returns the user messages that wait for a request, and removes them.
 436 func (l *TurnLoop) take() []Message {
 437         l.mutex.Lock()
 438         defer l.mutex.Unlock()
 439
 440         messages := l.pending
 441         l.pending = nil
 442         if l.messages != nil {
 443                 l.active = append(l.active, messages...)
 444         }
 445
 446         return messages
 447 }
 448
 449 // state returns the per-turn state: the end of the view, and the date and time
 450 // of the turn.
 451 func (l *TurnLoop) state() string {
 452         return "</chat>\n\nNow: " + l.clock().Format("Monday 2006-01-02 15:04 -07:00")
 453 }
 454
 455 // turnError returns the text of the error that ended a turn.
 456 func turnError(ctx context.Context, err error) string {
 457         if ctx.Err() != nil {
 458                 return "interrupted"
 459         }
 460
 461         return err.Error()
 462 }
 463
 464 // turnInput returns the input of the first request of a turn, stable parts
 465 // first: a user message for each view message, one with the per-turn state,
 466 // and one for each new message, whole. The view leaves out the leaf lines of
 467 // the new messages.
 468 func turnInput(view []string, state string, messages []Message) []any {
 469         input := make([]any, 0, len(view)+1+len(messages))
 470         for _, text := range view {
 471                 input = append(input, modelUserMessage(text))
 472         }
 473
 474         input = append(input, modelUserMessage(state))
 475         for _, message := range messages {
 476                 input = append(input, modelUserMessage(message.Text))
 477         }
 478
 479         return input
 480 }