X-Git-Url: http://git.giacomolli.uk/onechat/blobdiff_plain/4b063ca25798c78cc03d1d07f67b02c94c95248d..97729fc59a2fbd73e3f8a424cae4c825f5ba885b:/telegram.go

diff --git a/telegram.go b/telegram.go
index dabf2e8..9dbe86e 100644
--- a/telegram.go
+++ b/telegram.go
@@ -8,11 +8,13 @@ import (
 	"fmt"
 	"io"
 	"log"
+	"net/http"
 	"os"
 	"path/filepath"
 	"strings"
 	"sync"
 	"time"
+	"unicode/utf8"
 )
 
 const (
@@ -32,6 +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
@@ -50,9 +58,10 @@ type botConfig struct {
 // 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, and the user messages of the other clients to the chat as plain text,
-// and shows the typing action while a turn runs. Tool calls do not go to the
-// chat.
+// 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
@@ -74,6 +83,13 @@ type telegramBot struct {
 	ready chan struct{}
 }
 
+// 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"`
@@ -156,7 +172,7 @@ func newTelegramBot(api *botAPI, owner int64, offsetPath string, logger *log.Log
 // stop the agent.
 func (b *telegramBot) Notify(event Event) {
 	switch event.Kind {
-	case "busy", "idle", "reply", "error":
+	case "busy", "idle", "reply", "error", "notice":
 	case "user":
 		if event.Source == telegramSource {
 			return
@@ -228,7 +244,7 @@ func (b *telegramBot) send(ctx context.Context) {
 		case "idle":
 			busy = false
 		default:
-			for _, part := range splitMessage(botText(event), messageUnits) {
+			for _, part := range botParts(event) {
 				b.deliver(ctx, part)
 			}
 
@@ -260,16 +276,26 @@ func (b *telegramBot) typing(ctx context.Context) {
 	_ = b.api.SendTyping(ctx, b.owner)
 }
 
-// deliver sends text to the chat. It tries again after each failure that a
-// later try can fix, until ctx ends. It drops the text after any other failure.
-func (b *telegramBot) deliver(ctx context.Context, text string) {
+// 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, text)
+		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)
@@ -379,7 +405,7 @@ func (b *telegramBot) handle(ctx context.Context, l *loop, update botUpdate) err
 // notice queues a short answer of the bot for the chat.
 func (b *telegramBot) notice(text string) {
 	b.Notify(Event{
-		Kind: "reply",
+		Kind: "notice",
 		Text: text,
 	})
 }
@@ -422,6 +448,23 @@ 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 {
@@ -435,6 +478,38 @@ 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
@@ -480,3 +555,83 @@ 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, "", ""
+}