// makes Telegram send the update again, and the message repeats once.
//
// A failed Bot API call waits and tries again, from minWait up to maxWait, or
-// as long as the Bot API asks. It never stops the agent.
+// as long as the Bot API asks. It never stops the agent. A TelegramOutage logs
+// the first failure and the end of each run of failures, and the retries
+// between them log nothing.
type TelegramBot struct {
api *TelegramAPI
owner int64
}
// Notify queues event for the chat. It never blocks, so a slow network does not
-// stop the agent.
+// stop the agent. The WhatsApp batches and the runs of the schedules do not go
+// to the chat: the reply of the agent tells the user what needs the user.
func (b *TelegramBot) Notify(event Event) {
switch event.Kind {
- case "busy", "idle", "reply", "error", "notice":
+ case "busy", "idle", "reply", "error", "notice", "subagent":
+ case "work":
+ if event.Source == whatsappSource || event.Source == scheduleSource {
+ return
+ }
case "user":
if event.Source == telegramSource {
return
// Run polls for the messages of the owner and sends the queued events to the
// chat until ctx ends. It returns an error only when a message or the offset
// cannot be written, because then the bot would lose or repeat messages.
-func (b *TelegramBot) Run(ctx context.Context, l *TurnLoop) error {
+func (b *TelegramBot) Run(ctx context.Context, l *TurnLoop, whatsapp *WhatsApp) error {
ctx, cancel := context.WithCancel(ctx)
defer cancel()
b.send(ctx)
}()
- err := b.poll(ctx, l)
+ err := b.poll(ctx, l, whatsapp)
cancel()
sender.Wait()
// poll gets the updates from the stored offset on and handles each one, until
// ctx ends. Each update is handled to its end also when ctx ends during it, so
// a stop does not repeat its message.
-func (b *TelegramBot) poll(ctx context.Context, l *TurnLoop) error {
+func (b *TelegramBot) poll(ctx context.Context, l *TurnLoop, whatsapp *WhatsApp) error {
offset, err := telegramLoadOffset(b.offsetPath)
if err != nil {
return err
}
+ outage := &TelegramOutage{
+ logger: b.logger,
+ loss: "receive failed",
+ back: "receive back",
+ }
+
wait := b.minWait
for ctx.Err() == nil {
updates, err := b.api.GetUpdates(ctx, offset)
if err != nil {
_, asked := telegramRetryable(err)
- wait = b.pause(ctx, err, wait, asked)
+ wait = b.pause(ctx, outage, err, wait, asked)
continue
}
+ outage.End()
wait = b.minWait
for _, update := range updates {
if ctx.Err() != nil {
return nil
}
- err = b.handle(context.WithoutCancel(ctx), l, update)
+ err = b.handle(context.WithoutCancel(ctx), l, whatsapp, update)
if err != nil {
return err
}
}
// post sends event to the chat. A reply goes as rich messages of its Markdown,
-// split with telegramSplitRich. Every other event goes as plain text, split
-// with telegramSplitMessage.
+// with each < and & outside code escaped by telegramEscapeTags and then split
+// with telegramSplitRich, so that the length limit counts the escaped text.
+// Every other event goes as plain text, split with telegramSplitMessage.
func (b *TelegramBot) post(ctx context.Context, event Event) {
if event.Kind != "reply" {
for _, text := range telegramSplitMessage(telegramText(event), telegramMessageUnits) {
return
}
- for _, markdown := range telegramSplitRich(event.Text, telegramRichBytes) {
+ for _, markdown := range telegramSplitRich(telegramEscapeTags(event.Text), telegramRichBytes) {
b.deliverRich(ctx, markdown)
}
}
-// pause logs err and waits the longer of wait and asked, or until ctx ends. It
-// returns the wait of the next failure: twice wait, up to maxWait.
-func (b *TelegramBot) pause(ctx context.Context, err error, wait, asked time.Duration) time.Duration {
+// pause gives err to outage and waits the longer of wait and asked, or until
+// ctx ends. It returns the wait of the next failure: twice wait, up to maxWait.
+func (b *TelegramBot) pause(ctx context.Context, outage *TelegramOutage, err error, wait, asked time.Duration) time.Duration {
delay := max(wait, asked)
- b.logger.Printf("%v (next try in %s)", err, delay)
+ outage.Fail(err, delay)
timer := time.NewTimer(delay)
defer timer.Stop()
// handle runs one update. A text message of the owner in the private chat goes
// to the agent, and handle returns when it is durable in the log. The line
-// /stop interrupts the turn. Any other message of the owner gets the answer
+// /stop interrupts the turn, the line /stop NAME stops the subagent NAME, and a
+// line that starts with /whatsapp runs a WhatsApp command on whatsapp, whose
+// answers go to the chat only. Any other message of the owner gets the answer
// that only text is supported. Every other update is ignored, and its content
// is not logged.
-func (b *TelegramBot) handle(ctx context.Context, l *TurnLoop, update TelegramUpdate) error {
+func (b *TelegramBot) handle(ctx context.Context, l *TurnLoop, whatsapp *WhatsApp, update TelegramUpdate) error {
message := update.Message
if message == nil || message.Chat.Type != "private" || message.Chat.ID != b.owner || message.From == nil || message.From.ID != b.owner {
- b.logger.Print("update ignored: not a message of the owner in the private chat")
+ b.logger.Print("message ignored: not from the owner")
return nil
}
+ name, stop := stopSubagentName(message.Text)
+ args, isWhatsApp := whatsappCommandArgs(message.Text)
switch {
case message.Text == "":
b.notice(telegramTextOnly)
if !l.Interrupt() {
b.notice("No turn runs.")
}
+ case stop:
+ found, err := l.StopSubagent(name)
+ if err != nil {
+ return fmt.Errorf("stop subagent %s: %w", name, err)
+ }
+
+ if !found {
+ b.notice("No subagent " + name + " runs.")
+ }
+ case isWhatsApp:
+ whatsapp.Command(args, b.notice)
default:
_, err := l.Submit(telegramSource, message.Text)
if err != nil {
// reply is not lost to a bad entity. It tries again after each failure that a
// later try can fix, until ctx ends. It drops the part after any other failure.
func (b *TelegramBot) deliver(ctx context.Context, part TelegramPart) {
+ outage := b.sendOutage()
wait := b.minWait
for {
err := b.api.SendMessage(ctx, b.owner, part.text, part.entities)
+ retry, asked := telegramRetryable(err)
+ if !retry {
+ outage.End()
+ }
+
if err == nil || ctx.Err() != nil {
return
}
var answer *TelegramError
if len(part.entities) > 0 && errors.As(err, &answer) && answer.Status == http.StatusBadRequest {
- b.logger.Printf("entities refused, the message goes again as plain text: %v", err)
+ b.logger.Printf("reply sent as plain text: %v", err)
part.entities = nil
continue
}
- retry, asked := telegramRetryable(err)
if !retry {
- b.logger.Printf("message dropped: %v", err)
+ b.logger.Printf("reply lost: %v", err)
return
}
- wait = b.pause(ctx, err, wait, asked)
+ wait = b.pause(ctx, outage, err, wait, asked)
}
}
-// deliverRich sends markdown to the chat as a rich message. When the Bot API
-// refuses it with HTTP 400 or 404, the message was not sent, so deliverRich
-// sends the same Markdown once with message entities, and the reply is not
-// lost and does not repeat. It tries again after each failure that a later try
-// can fix, until ctx ends. It drops the message after any other failure.
+// deliverRich sends markdown, escaped with telegramEscapeTags, to the chat as a
+// rich message. When the Bot API refuses it with HTTP 400 or 404, the message
+// was not sent, so deliverRich sends the text of the part once with message
+// entities, and the reply is not lost and does not repeat. Message entities
+// read no HTML, so the fallback gets the Markdown of telegramUnescapeTags, as
+// the model wrote it. It tries again after each failure that a later try can
+// fix, until ctx ends. It drops the message after any other failure.
func (b *TelegramBot) deliverRich(ctx context.Context, markdown string) {
+ outage := b.sendOutage()
wait := b.minWait
for {
err := b.api.SendRichMessage(ctx, b.owner, markdown)
+ retry, asked := telegramRetryable(err)
+ if !retry {
+ outage.End()
+ }
+
if err == nil || ctx.Err() != nil {
return
}
if telegramRejected(err) {
- b.logger.Printf("rich message refused, the reply goes with entities: %v", err)
- for _, part := range telegramReplyParts(markdown) {
+ b.logger.Printf("reply sent with simple formatting: %v", err)
+ for _, part := range telegramReplyParts(telegramUnescapeTags(markdown)) {
b.deliver(ctx, part)
}
return
}
- retry, asked := telegramRetryable(err)
if !retry {
- b.logger.Printf("message dropped: %v", err)
+ b.logger.Printf("reply lost: %v", err)
return
}
- wait = b.pause(ctx, err, wait, asked)
+ wait = b.pause(ctx, outage, err, wait, asked)
+ }
+}
+
+// sendOutage returns a new outage of the messages to the chat.
+func (b *TelegramBot) sendOutage() *TelegramOutage {
+ return &TelegramOutage{
+ logger: b.logger,
+ loss: "send failed",
+ back: "send back",
}
}
})
}
+// TelegramOutage is one run of failures of a kind of Bot API call. The first
+// failure logs loss, its cause, and the wait of the next try. The retries log
+// nothing. The next answer of Telegram logs back and the length of the outage.
+// loss and back say what the user loses and gets again.
+type TelegramOutage struct {
+ logger *log.Logger
+ loss string
+ back string
+ start time.Time
+}
+
+// Fail logs the start of an outage with err and delay, unless an outage runs.
+func (o *TelegramOutage) Fail(err error, delay time.Duration) {
+ if !o.start.IsZero() {
+ return
+ }
+
+ o.start = time.Now()
+ o.logger.Printf("%s: %v (retry in %s)", o.loss, err, durationText(delay))
+}
+
+// End logs the end of the outage that runs, with its length.
+func (o *TelegramOutage) End() {
+ if o.start.IsZero() {
+ return
+ }
+
+ o.logger.Printf("%s after %s", o.back, durationText(time.Since(o.start)))
+ o.start = time.Time{}
+}
+
// TelegramPart is the text of one message to the chat, at most
// telegramMessageUnits long, and its entities.
type TelegramPart struct {
}
// telegramText returns the text of event for the chat. A user message of
-// another client has its source as a mark.
+// another client has its source as a mark. A subagent shows only its start and
+// how its run ended, because the reply of the agent carries the content of the
+// report.
func telegramText(event Event) string {
switch event.Kind {
case "user":
return "[" + event.Source + "] " + event.Text
+ case "subagent":
+ return "[" + event.Text + "] started"
+ case "work":
+ return "[" + event.Source + "] " + event.Status
case "error":
return "Error: " + event.Text
default:
// after the last line break outside a code block, then after the last line
// break in a code block, which it closes at the end of the part and opens again
// in the next part with the same fence line. A line longer than limit is cut at
-// the limit. It never cuts a character.
+// the limit, or before an & in the four bytes before the limit, so that it does
+// not cut an entity of telegramEscapeTags. It never cuts a character.
func telegramSplitRich(markdown string, limit int) []string {
var parts []string
for len(markdown) > limit {
cut--
}
+ start := max(cut-len("&")+1, 0)
+ entity := strings.LastIndexByte(markdown[start:cut], '&')
+ if entity >= 0 && start+entity > 0 {
+ cut = start + entity
+ }
+
return cut, "", ""
}
projects / onechat / blobdiff