X-Git-Url: https://git.giacomolli.uk/onechat/blobdiff_plain/f80b058ae0550a56b4e62f8e65c2a74349b970e4..84fc815feb485b35eec53a8c7863c6770fc65e02:/telegram.go?ds=sidebyside

diff --git a/telegram.go b/telegram.go
index 9a24a4e..dd62ef0 100644
--- a/telegram.go
+++ b/telegram.go
@@ -14,6 +14,7 @@ import (
 	"strings"
 	"sync"
 	"time"
+	"unicode/utf8"
 )
 
 const (
@@ -33,6 +34,12 @@ const (
 	// messageUnits is the most UTF-16 code units of one Telegram message.
 	messageUnits = 4096
 
+	// richMessageBytes 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.
+	richMessageBytes = 32768
+
 	// 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
@@ -53,8 +60,9 @@ type botConfig struct {
 // 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.
+// 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
@@ -237,9 +245,7 @@ func (b *telegramBot) send(ctx context.Context) {
 		case "idle":
 			busy = false
 		default:
-			for _, part := range botParts(event) {
-				b.deliver(ctx, part)
-			}
+			b.post(ctx, event)
 
 			if busy {
 				b.typing(ctx)
@@ -269,6 +275,57 @@ 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,
+// split with splitRich. Every other event goes as plain text, split with
+// splitMessage.
+func (b *telegramBot) post(ctx context.Context, event Event) {
+	if event.Kind != "reply" {
+		for _, text := range splitMessage(botText(event), messageUnits) {
+			b.deliver(ctx, botPart{
+				text: text,
+			})
+		}
+
+		return
+	}
+
+	for _, markdown := range splitRich(event.Text, richMessageBytes) {
+		b.deliverRich(ctx, markdown)
+	}
+}
+
+// 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.
+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 rejected(err) {
+			b.logger.Printf("rich message refused, the reply goes with entities: %v", err)
+			for _, part := range replyParts(markdown) {
+				b.deliver(ctx, part)
+			}
+
+			return
+		}
+
+		retry, asked := retryable(err)
+		if !retry {
+			b.logger.Printf("message dropped: %v", err)
+			return
+		}
+
+		wait = b.pause(ctx, err, wait, asked)
+	}
+}
+
 // 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
@@ -441,23 +498,6 @@ 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)
-	}
-
-	if strings.TrimSpace(text) == "" {
-		text, entities = event.Text, nil
-	}
-
-	return splitEntities(text, entities, messageUnits)
-}
-
 // 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 {
@@ -471,38 +511,6 @@ 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
-			}
-
-			entity.Offset = from - start
-			entity.Length = to - from
-			kept = append(kept, entity)
-		}
-
-		parts = append(parts, botPart{
-			text:     part,
-			entities: kept,
-		})
-
-		start = end
-	}
-
-	return parts
-}
-
 // 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
@@ -548,3 +556,127 @@ func prefixEnd(text string, limit int) int {
 
 	return len(text)
 }
+
+// splitRich 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. It never cuts a character.
+func splitRich(markdown string, limit int) []string {
+	var parts []string
+	for len(markdown) > limit {
+		cut, fence, opening := richCut(markdown, limit)
+		if fence == "" {
+			parts = append(parts, markdown[:cut])
+			markdown = markdown[cut:]
+
+			continue
+		}
+
+		parts = append(parts, markdown[:cut]+fence+"\n")
+		markdown = opening + "\n" + markdown[cut:]
+	}
+
+	return append(parts, markdown)
+}
+
+// richCut returns where splitRich 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 richCut(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 := codeFence(text)
+		switch {
+		case fence != "":
+			if closesFence(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--
+	}
+
+	return cut, "", ""
+}
+
+// replyParts 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 replyParts(markdown string) []botPart {
+	text, entities := markdownEntities(markdown)
+	if strings.TrimSpace(text) == "" {
+		text, entities = markdown, nil
+	}
+
+	return splitEntities(text, entities, messageUnits)
+}
+
+// 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
+			}
+
+			entity.Offset = from - start
+			entity.Length = to - from
+			kept = append(kept, entity)
+		}
+
+		parts = append(parts, botPart{
+			text:     part,
+			entities: kept,
+		})
+
+		start = end
+	}
+
+	return parts
+}