X-Git-Url: https://git.giacomolli.uk/onechat/blobdiff_plain/f80b058ae0550a56b4e62f8e65c2a74349b970e4..6e3e9c5c10f3d35bae7e9846334974a09418aef7:/telegram.go?ds=sidebyside

diff --git a/telegram.go b/telegram.go
index 9a24a4e..6d5332d 100644
--- a/telegram.go
+++ b/telegram.go
@@ -14,103 +14,65 @@ import (
 	"strings"
 	"sync"
 	"time"
+	"unicode/utf8"
 )
 
 const (
-	// telegramAPI is the Telegram Bot API service.
-	telegramAPI = "https://api.telegram.org"
+	// telegramAPIURL is the Telegram Bot API service.
+	telegramAPIURL = "https://api.telegram.org"
 
-	// botConfigFileName is the bot config in the data directory. The daemon
-	// runs the bot only when the file exists.
-	botConfigFileName = "telegram.json"
+	// telegramConfigFileName is the bot config in the data directory. The
+	// daemon runs the bot only when the file exists.
+	telegramConfigFileName = "telegram.json"
 
-	// offsetFileName is the update offset of the bot in the agent directory.
-	offsetFileName = "telegram-offset.json"
+	// telegramOffsetFileName is the update offset of the bot in the agent
+	// directory.
+	telegramOffsetFileName = "telegram-offset.json"
 
 	// telegramSource is the source of the user messages from Telegram.
 	telegramSource = "telegram"
 
-	// messageUnits is the most UTF-16 code units of one Telegram message.
-	messageUnits = 4096
+	// telegramMessageUnits is the most UTF-16 code units of one Telegram
+	// message.
+	telegramMessageUnits = 4096
 
-	// typingInterval is how often the bot shows the typing action while a
-	// turn runs. Telegram shows the action for about five seconds.
-	typingInterval = 4 * time.Second
+	// telegramRichBytes is the most bytes of the Markdown of one rich
+	// message. The Bot API allows 32768 characters of rendered text, and
+	// the Markdown of a text has at least as many bytes as the text has
+	// characters.
+	telegramRichBytes = 32768
 
-	// textOnly is the answer to a message that is not text.
-	textOnly = "Only text messages are supported."
-)
-
-// botConfig holds the settings of telegram.json: the bot token and the user ID
-// of the owner. The token is a secret, so the file must have mode 0600.
-type botConfig struct {
-	Token string `json:"token"`
-	Owner int64  `json:"owner"`
-}
-
-// telegramBot connects the private chat of the owner with the agent. It polls
-// the Bot API for new messages and sends each text message of the owner to the
-// agent. It ignores every other update. Its sender sends the replies, the
-// errors, the notices of the bot, and the user messages of the other clients to
-// the chat, and shows the typing action while a turn runs. Only the replies
-// have a format: their Markdown goes as message entities. The rest is plain
-// text. Tool calls do not go to the chat.
-//
-// The offset file holds the ID of the next update. It moves past an update
-// only after the message of the update is durable in the log, so a restart
-// does not lose a message. A crash between the flush and the offset write
-// 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.
-type telegramBot struct {
-	api        *botAPI
-	owner      int64
-	offsetPath string
-	logger     *log.Logger
-	minWait    time.Duration
-	maxWait    time.Duration
-
-	mutex sync.Mutex
-	queue []Event
-	ready chan struct{}
-}
+	// telegramTypingInterval is how often the bot shows the typing action
+	// while a turn runs. Telegram shows the action for about five seconds.
+	telegramTypingInterval = 4 * time.Second
 
-// botPart is the text of one message to the chat, at most messageUnits long,
-// and its entities.
-type botPart struct {
-	text     string
-	entities []botEntity
-}
-
-// offsetFile is the content of the offset file.
-type offsetFile struct {
-	Offset int64 `json:"offset"`
-}
+	// telegramTextOnly is the answer to a message that is not text.
+	telegramTextOnly = "Only text messages are supported."
+)
 
-// loadBotConfig reads telegram.json in dir. It returns false when the file
+// loadTelegramConfig reads telegram.json in dir. It returns false when the file
 // does not exist. An error never holds the token.
-func loadBotConfig(dir string) (botConfig, bool, error) {
-	path := filepath.Join(dir, botConfigFileName)
+func loadTelegramConfig(dir string) (TelegramConfig, bool, error) {
+	path := filepath.Join(dir, telegramConfigFileName)
 	info, err := os.Stat(path)
 	if errors.Is(err, os.ErrNotExist) {
-		return botConfig{}, false, nil
+		return TelegramConfig{}, false, nil
 	}
 
 	if err != nil {
-		return botConfig{}, false, fmt.Errorf("read bot config: %w", err)
+		return TelegramConfig{}, false, fmt.Errorf("read bot config: %w", err)
 	}
 
 	if info.Mode().Perm()&0o077 != 0 {
-		return botConfig{}, false, fmt.Errorf("%s holds the bot token, so its mode must be 0600", botConfigFileName)
+		return TelegramConfig{}, false, fmt.Errorf("%s holds the bot token, so its mode must be 0600", telegramConfigFileName)
 	}
 
 	data, err := os.ReadFile(path)
 	if err != nil {
-		return botConfig{}, false, fmt.Errorf("read bot config: %w", err)
+		return TelegramConfig{}, false, fmt.Errorf("read bot config: %w", err)
 	}
 
-	var config botConfig
+	var config TelegramConfig
 	decoder := json.NewDecoder(bytes.NewReader(data))
 	decoder.DisallowUnknownFields()
 	err = decoder.Decode(&config)
@@ -119,21 +81,43 @@ func loadBotConfig(dir string) (botConfig, bool, error) {
 	}
 
 	if err != nil {
-		return botConfig{}, false, fmt.Errorf("%s is not valid JSON with the fields token and owner", botConfigFileName)
+		return TelegramConfig{}, false, fmt.Errorf("%s is not valid JSON with the fields token and owner", telegramConfigFileName)
 	}
 
 	err = config.validate()
 	if err != nil {
-		return botConfig{}, false, fmt.Errorf("%s: %w", botConfigFileName, err)
+		return TelegramConfig{}, false, fmt.Errorf("%s: %w", telegramConfigFileName, err)
 	}
 
 	return config, true, nil
 }
 
+// newTelegramBot returns the bot of owner on api. The offset file is at
+// offsetPath, and logger gets the failures. No log line holds the text of a
+// message.
+func newTelegramBot(api *TelegramAPI, owner int64, offsetPath string, logger *log.Logger) *TelegramBot {
+	return &TelegramBot{
+		api:        api,
+		owner:      owner,
+		offsetPath: offsetPath,
+		logger:     logger,
+		minWait:    time.Second,
+		maxWait:    time.Minute,
+		ready:      make(chan struct{}, 1),
+	}
+}
+
+// TelegramConfig holds the settings of telegram.json: the bot token and the
+// user ID of the owner. The token is a secret, so the file must have mode 0600.
+type TelegramConfig struct {
+	Token string `json:"token"`
+	Owner int64  `json:"owner"`
+}
+
 // validate makes sure that the token has the form of a bot token and that the
 // owner is a user ID. The token goes into a URL path, so it can only hold the
 // characters of a bot token.
-func (c botConfig) validate() error {
+func (c TelegramConfig) validate() error {
 	id, secret, found := strings.Cut(c.Token, ":")
 	if !found || id == "" || secret == "" || strings.Trim(id, "0123456789") != "" || strings.Trim(secret, "0123456789abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ_-") != "" {
 		return errors.New("token does not have the form of a bot token from BotFather")
@@ -146,24 +130,38 @@ func (c botConfig) validate() error {
 	return nil
 }
 
-// newTelegramBot returns the bot of owner on api. The offset file is at
-// offsetPath, and logger gets the failures. No log line holds the text of a
-// message.
-func newTelegramBot(api *botAPI, owner int64, offsetPath string, logger *log.Logger) *telegramBot {
-	return &telegramBot{
-		api:        api,
-		owner:      owner,
-		offsetPath: offsetPath,
-		logger:     logger,
-		minWait:    time.Second,
-		maxWait:    time.Minute,
-		ready:      make(chan struct{}, 1),
-	}
+// TelegramBot connects the private chat of the owner with the agent. It polls
+// the Bot API for new messages and sends each text message of the owner to the
+// agent. It ignores every other update. Its sender sends the replies, the
+// errors, the notices of the bot, and the user messages of the other clients to
+// the chat, and shows the typing action while a turn runs. Only the replies
+// have a format: their Markdown goes as rich messages, which Telegram renders
+// itself, and as message entities only when Telegram refuses a rich message.
+// The rest is plain text. Tool calls do not go to the chat.
+//
+// The offset file holds the ID of the next update. It moves past an update
+// only after the message of the update is durable in the log, so a restart
+// does not lose a message. A crash between the flush and the offset write
+// 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.
+type TelegramBot struct {
+	api        *TelegramAPI
+	owner      int64
+	offsetPath string
+	logger     *log.Logger
+	minWait    time.Duration
+	maxWait    time.Duration
+
+	mutex sync.Mutex
+	queue []Event
+	ready chan struct{}
 }
 
 // Notify queues event for the chat. It never blocks, so a slow network does not
 // stop the agent.
-func (b *telegramBot) Notify(event Event) {
+func (b *TelegramBot) Notify(event Event) {
 	switch event.Kind {
 	case "busy", "idle", "reply", "error", "notice":
 	case "user":
@@ -187,7 +185,7 @@ func (b *telegramBot) Notify(event Event) {
 // 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 *loop) error {
+func (b *TelegramBot) Run(ctx context.Context, l *TurnLoop) error {
 	ctx, cancel := context.WithCancel(ctx)
 	defer cancel()
 
@@ -207,10 +205,10 @@ func (b *telegramBot) Run(ctx context.Context, l *loop) error {
 }
 
 // send sends each queued event to the chat in order until ctx ends. While a
-// turn runs, it shows the typing action every typingInterval and after each
-// message, because a message ends the action.
-func (b *telegramBot) send(ctx context.Context) {
-	ticker := time.NewTicker(typingInterval)
+// turn runs, it shows the typing action every telegramTypingInterval and after
+// each message, because a message ends the action.
+func (b *TelegramBot) send(ctx context.Context) {
+	ticker := time.NewTicker(telegramTypingInterval)
 	defer ticker.Stop()
 
 	busy := false
@@ -233,13 +231,11 @@ func (b *telegramBot) send(ctx context.Context) {
 		case "busy":
 			busy = true
 			b.typing(ctx)
-			ticker.Reset(typingInterval)
+			ticker.Reset(telegramTypingInterval)
 		case "idle":
 			busy = false
 		default:
-			for _, part := range botParts(event) {
-				b.deliver(ctx, part)
-			}
+			b.post(ctx, event)
 
 			if busy {
 				b.typing(ctx)
@@ -248,79 +244,11 @@ func (b *telegramBot) send(ctx context.Context) {
 	}
 }
 
-// next removes the first queued event and returns it.
-func (b *telegramBot) next() (Event, bool) {
-	b.mutex.Lock()
-	defer b.mutex.Unlock()
-
-	if len(b.queue) == 0 {
-		return Event{}, false
-	}
-
-	event := b.queue[0]
-	b.queue = b.queue[1:]
-
-	return event, true
-}
-
-// typing shows the typing action once. A failure is not logged, because the
-// next message or poll logs the state of the network.
-func (b *telegramBot) typing(ctx context.Context) {
-	_ = b.api.SendTyping(ctx, b.owner)
-}
-
-// deliver sends part to the chat. When Telegram refuses a part with entities
-// (HTTP 400), it sends the text of the part again once with no entities, so a
-// 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 botPart) {
-	wait := b.minWait
-	for {
-		err := b.api.SendMessage(ctx, b.owner, part.text, part.entities)
-		if err == nil || ctx.Err() != nil {
-			return
-		}
-
-		var answer *botError
-		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)
-			part.entities = nil
-
-			continue
-		}
-
-		retry, asked := retryable(err)
-		if !retry {
-			b.logger.Printf("message dropped: %v", err)
-			return
-		}
-
-		wait = b.pause(ctx, err, wait, asked)
-	}
-}
-
-// 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 {
-	delay := max(wait, asked)
-	b.logger.Printf("%v (next try in %s)", err, delay)
-
-	timer := time.NewTimer(delay)
-	defer timer.Stop()
-
-	select {
-	case <-timer.C:
-	case <-ctx.Done():
-	}
-
-	return min(2*wait, b.maxWait)
-}
-
 // 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 *loop) error {
-	offset, err := loadOffset(b.offsetPath)
+func (b *TelegramBot) poll(ctx context.Context, l *TurnLoop) error {
+	offset, err := telegramLoadOffset(b.offsetPath)
 	if err != nil {
 		return err
 	}
@@ -333,7 +261,7 @@ func (b *telegramBot) poll(ctx context.Context, l *loop) error {
 		}
 
 		if err != nil {
-			_, asked := retryable(err)
+			_, asked := telegramRetryable(err)
 			wait = b.pause(ctx, err, wait, asked)
 
 			continue
@@ -351,7 +279,7 @@ func (b *telegramBot) poll(ctx context.Context, l *loop) error {
 			}
 
 			offset = update.UpdateID + 1
-			err = saveOffset(b.offsetPath, offset)
+			err = telegramSaveOffset(b.offsetPath, offset)
 			if err != nil {
 				return err
 			}
@@ -361,12 +289,70 @@ func (b *telegramBot) poll(ctx context.Context, l *loop) error {
 	return nil
 }
 
+// next removes the first queued event and returns it.
+func (b *TelegramBot) next() (Event, bool) {
+	b.mutex.Lock()
+	defer b.mutex.Unlock()
+
+	if len(b.queue) == 0 {
+		return Event{}, false
+	}
+
+	event := b.queue[0]
+	b.queue = b.queue[1:]
+
+	return event, true
+}
+
+// typing shows the typing action once. A failure is not logged, because the
+// next message or poll logs the state of the network.
+func (b *TelegramBot) typing(ctx context.Context) {
+	_ = b.api.SendTyping(ctx, b.owner)
+}
+
+// post sends event to the chat. A reply goes as rich messages of its Markdown,
+// 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) {
+			b.deliver(ctx, TelegramPart{
+				text: text,
+			})
+		}
+
+		return
+	}
+
+	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 {
+	delay := max(wait, asked)
+	b.logger.Printf("%v (next try in %s)", err, delay)
+
+	timer := time.NewTimer(delay)
+	defer timer.Stop()
+
+	select {
+	case <-timer.C:
+	case <-ctx.Done():
+	}
+
+	return min(2*wait, b.maxWait)
+}
+
 // 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
 // that only text is supported. Every other update is ignored, and its content
 // is not logged.
-func (b *telegramBot) handle(ctx context.Context, l *loop, update botUpdate) error {
+func (b *TelegramBot) handle(ctx context.Context, l *TurnLoop, 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")
@@ -375,7 +361,7 @@ func (b *telegramBot) handle(ctx context.Context, l *loop, update botUpdate) err
 
 	switch {
 	case message.Text == "":
-		b.notice(textOnly)
+		b.notice(telegramTextOnly)
 	case strings.TrimSpace(message.Text) == stopCommand:
 		if !l.Interrupt() {
 			b.notice("No turn runs.")
@@ -395,17 +381,93 @@ func (b *telegramBot) handle(ctx context.Context, l *loop, update botUpdate) err
 	return nil
 }
 
+// deliver sends part to the chat. When Telegram refuses a part with entities
+// (HTTP 400), it sends the text of the part again once with no entities, so a
+// 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) {
+	wait := b.minWait
+	for {
+		err := b.api.SendMessage(ctx, b.owner, part.text, part.entities)
+		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)
+			part.entities = nil
+
+			continue
+		}
+
+		retry, asked := telegramRetryable(err)
+		if !retry {
+			b.logger.Printf("message dropped: %v", err)
+			return
+		}
+
+		wait = b.pause(ctx, err, wait, asked)
+	}
+}
+
+// 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) {
+	wait := b.minWait
+	for {
+		err := b.api.SendRichMessage(ctx, b.owner, markdown)
+		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(telegramUnescapeTags(markdown)) {
+				b.deliver(ctx, part)
+			}
+
+			return
+		}
+
+		retry, asked := telegramRetryable(err)
+		if !retry {
+			b.logger.Printf("message dropped: %v", err)
+			return
+		}
+
+		wait = b.pause(ctx, err, wait, asked)
+	}
+}
+
 // notice queues a short answer of the bot for the chat.
-func (b *telegramBot) notice(text string) {
+func (b *TelegramBot) notice(text string) {
 	b.Notify(Event{
 		Kind: "notice",
 		Text: text,
 	})
 }
 
-// loadOffset reads the update offset at path. A missing file gives 0, which
-// asks for every update that Telegram keeps.
-func loadOffset(path string) (int64, error) {
+// TelegramPart is the text of one message to the chat, at most
+// telegramMessageUnits long, and its entities.
+type TelegramPart struct {
+	text     string
+	entities []TelegramEntity
+}
+
+// TelegramOffsetFile is the content of the offset file.
+type TelegramOffsetFile struct {
+	Offset int64 `json:"offset"`
+}
+
+// telegramLoadOffset reads the update offset at path. A missing file gives 0,
+// which asks for every update that Telegram keeps.
+func telegramLoadOffset(path string) (int64, error) {
 	data, err := os.ReadFile(path)
 	if errors.Is(err, os.ErrNotExist) {
 		return 0, nil
@@ -415,7 +477,7 @@ func loadOffset(path string) (int64, error) {
 		return 0, fmt.Errorf("read update offset: %w", err)
 	}
 
-	var file offsetFile
+	var file TelegramOffsetFile
 	err = json.Unmarshal(data, &file)
 	if err != nil {
 		return 0, fmt.Errorf("%s: %w", path, err)
@@ -424,9 +486,9 @@ func loadOffset(path string) (int64, error) {
 	return file.Offset, nil
 }
 
-// saveOffset replaces the update offset at path.
-func saveOffset(path string, offset int64) error {
-	data, err := json.Marshal(offsetFile{
+// telegramSaveOffset replaces the update offset at path.
+func telegramSaveOffset(path string, offset int64) error {
+	data, err := json.Marshal(TelegramOffsetFile{
 		Offset: offset,
 	})
 	if err != nil {
@@ -441,26 +503,35 @@ func saveOffset(path string, offset int64) error {
 	return nil
 }
 
-// botParts returns the messages of event for the chat. The Markdown of a reply
-// becomes text with entities. Every other event is plain text. A reply that has
-// no text left after the conversion, such as an empty code block, goes as its
-// Markdown.
-func botParts(event Event) []botPart {
-	text, entities := botText(event), []botEntity(nil)
-	if event.Kind == "reply" {
-		text, entities = markdownEntities(event.Text)
-	}
+// telegramSplitMessage cuts text into parts of at most limit UTF-16 code units,
+// the unit of the Telegram message limit. It cuts after the last line break or
+// space of a part when one is in its second half, and else at the limit. It
+// never cuts a character.
+func telegramSplitMessage(text string, limit int) []string {
+	var parts []string
+	for {
+		end := telegramPrefixEnd(text, limit)
+		if end == len(text) {
+			return append(parts, text)
+		}
 
-	if strings.TrimSpace(text) == "" {
-		text, entities = event.Text, nil
-	}
+		cut := strings.LastIndexByte(text[:end], '\n') + 1
+		if cut <= end/2 {
+			cut = strings.LastIndexByte(text[:end], ' ') + 1
+		}
 
-	return splitEntities(text, entities, messageUnits)
+		if cut <= end/2 {
+			cut = end
+		}
+
+		parts = append(parts, text[:cut])
+		text = text[cut:]
+	}
 }
 
-// botText returns the text of event for the chat. A user message of another
-// client has its source as a mark.
-func botText(event Event) string {
+// telegramText returns the text of event for the chat. A user message of
+// another client has its source as a mark.
+func telegramText(event Event) string {
 	switch event.Kind {
 	case "user":
 		return "[" + event.Source + "] " + event.Text
@@ -471,67 +542,47 @@ func botText(event Event) string {
 	}
 }
 
-// splitEntities cuts text into parts with splitMessage, and gives each part the
-// entities in it, with offsets from the start of the part. An entity that
-// crosses a cut becomes one entity in each part.
-func splitEntities(text string, entities []botEntity, limit int) []botPart {
-	var parts []botPart
-	start := 0
-	for _, part := range splitMessage(text, limit) {
-		end := start + utf16Units(part)
-		var kept []botEntity
-		for _, entity := range entities {
-			from := max(entity.Offset, start)
-			to := min(entity.Offset+entity.Length, end)
-			if to <= from {
-				continue
-			}
+// telegramSplitRich cuts the Markdown of a reply into parts of at most limit
+// bytes. It cuts after a blank line outside a code block, so a table, a code
+// block, or a paragraph stays whole. When no such cut fits in a part, it cuts
+// 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, 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, fence, opening := telegramRichCut(markdown, limit)
+		if fence == "" {
+			parts = append(parts, markdown[:cut])
+			markdown = markdown[cut:]
 
-			entity.Offset = from - start
-			entity.Length = to - from
-			kept = append(kept, entity)
+			continue
 		}
 
-		parts = append(parts, botPart{
-			text:     part,
-			entities: kept,
-		})
-
-		start = end
+		parts = append(parts, markdown[:cut]+fence+"\n")
+		markdown = opening + "\n" + markdown[cut:]
 	}
 
-	return parts
+	return append(parts, markdown)
 }
 
-// splitMessage cuts text into parts of at most limit UTF-16 code units, the
-// unit of the Telegram message limit. It cuts after the last line break or
-// space of a part when one is in its second half, and else at the limit. It
-// never cuts a character.
-func splitMessage(text string, limit int) []string {
-	var parts []string
-	for {
-		end := prefixEnd(text, limit)
-		if end == len(text) {
-			return append(parts, text)
-		}
-
-		cut := strings.LastIndexByte(text[:end], '\n') + 1
-		if cut <= end/2 {
-			cut = strings.LastIndexByte(text[:end], ' ') + 1
-		}
-
-		if cut <= end/2 {
-			cut = end
-		}
-
-		parts = append(parts, text[:cut])
-		text = text[cut:]
+// telegramReplyParts returns the messages of the Markdown of a reply with
+// message entities. A reply that has no text left after the conversion, such as
+// an empty code block, goes as its Markdown.
+func telegramReplyParts(markdown string) []TelegramPart {
+	text, entities := telegramMarkdownEntities(markdown)
+	if strings.TrimSpace(text) == "" {
+		text, entities = markdown, nil
 	}
+
+	return telegramSplitEntities(text, entities, telegramMessageUnits)
 }
 
-// prefixEnd returns the byte length of the longest prefix of text with at most
-// limit UTF-16 code units.
-func prefixEnd(text string, limit int) int {
+// telegramPrefixEnd returns the byte length of the longest prefix of text with
+// at most limit UTF-16 code units.
+func telegramPrefixEnd(text string, limit int) int {
 	units := 0
 	for index, char := range text {
 		size := 1
@@ -548,3 +599,96 @@ func prefixEnd(text string, limit int) int {
 
 	return len(text)
 }
+
+// telegramRichCut returns where telegramSplitRich cuts markdown, which is
+// longer than limit. For a cut in a code block, it also returns the fence that
+// closes the block and the line that opens it. Each part before a cut holds
+// text other than space, and a cut in a code block leaves room for the fence,
+// so each cut makes the rest shorter.
+func telegramRichCut(markdown string, limit int) (int, string, string) {
+	blank, line, code := 0, 0, 0
+	var fence, opening, inFence, inOpening string
+	for start := 0; start < limit; {
+		end := strings.IndexByte(markdown[start:], '\n')
+		if end < 0 || start+end+1 > limit {
+			break
+		}
+
+		end += start + 1
+		text := strings.TrimRight(markdown[start:end], "\r\n")
+		_, opens := telegramCodeFence(text)
+		switch {
+		case fence != "":
+			if telegramClosesFence(text, fence) {
+				fence = ""
+			} else if end > len(opening)+1 && end+len(fence)+1 <= limit {
+				code, inFence, inOpening = end, fence, opening
+			}
+		case opens != "":
+			fence, opening = opens, text
+		case strings.TrimSpace(markdown[:end]) == "":
+		case strings.TrimSpace(text) == "":
+			blank = end
+		}
+
+		if fence == "" && strings.TrimSpace(markdown[:end]) != "" {
+			line = end
+		}
+
+		start = end
+	}
+
+	switch {
+	case blank > 0:
+		return blank, "", ""
+	case line > 0:
+		return line, "", ""
+	case code > 0:
+		return code, inFence, inOpening
+	}
+
+	cut := limit
+	for !utf8.RuneStart(markdown[cut]) {
+		cut--
+	}
+
+	start := max(cut-len("&amp;")+1, 0)
+	entity := strings.LastIndexByte(markdown[start:cut], '&')
+	if entity >= 0 && start+entity > 0 {
+		cut = start + entity
+	}
+
+	return cut, "", ""
+}
+
+// telegramSplitEntities cuts text into parts with telegramSplitMessage, and
+// gives each part the entities in it, with offsets from the start of the part.
+// An entity that crosses a cut becomes one entity in each part.
+func telegramSplitEntities(text string, entities []TelegramEntity, limit int) []TelegramPart {
+	var parts []TelegramPart
+	start := 0
+	for _, part := range telegramSplitMessage(text, limit) {
+		end := start + telegramUTF16Units(part)
+		var kept []TelegramEntity
+		for _, entity := range entities {
+			from := max(entity.Offset, start)
+			to := min(entity.Offset+entity.Length, end)
+			if to <= from {
+				continue
+			}
+
+			entity.Offset = from - start
+			entity.Length = to - from
+			kept = append(kept, entity)
+		}
+
+		parts = append(parts, TelegramPart{
+			text:     part,
+			entities: kept,
+		})
+
+		start = end
+	}
+
+	return parts
+}