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