Files
Memoh/internal/agent/tools/tts.go
T
Acbox Liu 5cfbaa40e2 refactor(agent): replace XML tag extraction with tool-based send/react/speak (#330)
* refactor(agent): replace XML tag extraction with tool-based send/react/speak

Remove the <attachments>, <reactions>, and <speech> XML tag extraction
system from the agent streaming pipeline. Instead, the send/react/speak
tools now handle both same-conversation and cross-conversation delivery:

- send: omit target to deliver attachments in the current conversation;
  specify target for cross-channel messaging
- react: omit target to react in the current conversation
- speak: omit target to speak in the current conversation

Backend changes:
- Add StreamEmitter callback to tools.SessionContext so tools can push
  attachment/reaction/speech events directly into the agent stream
- Wire emitter in agent.go for both streaming and non-streaming paths
- Remove StreamTagExtractor, DefaultTagResolvers, emitTagEvents, and
  delete internal/agent/tags.go entirely
- Remove StripAgentTags calls from assistant_output.go
- Add IsSameConversation detection in messaging executor; same-conv
  sends pass raw paths through the emitter for downstream ingestion
- Auto-resolve relative paths (e.g. "IDENTITY.md" -> "/data/IDENTITY.md")
- Add Metadata propagation through the full attachment chain
  (tools.Attachment -> agent.FileAttachment -> parseAttachmentDelta)
- Update system_chat.md and _contacts.md prompts

Frontend changes (apps/web):
- Hide send/react/speak tool_call blocks when result indicates
  delivered to current conversation
- Defer attachment_delta blocks to end of message (flush on stream
  completion) for consistent positioning with DB-loaded history

* fix(agent): speak tool emits synthesized audio directly as voice attachment

Instead of emitting speech_delta (which requires downstream re-synthesis),
the speak tool now emits the already-synthesized audio as an attachment_delta
with voice type. This avoids double TTS synthesis and eliminates dependency
on ttsService being configured on the inbound processor.

Also fixes speak on WebUI where ReplyTarget is empty (same fix as send).
2026-04-04 20:55:03 +08:00

169 lines
5.5 KiB
Go

package tools
import (
"context"
"encoding/base64"
"errors"
"fmt"
"log/slog"
"strings"
sdk "github.com/memohai/twilight-ai/sdk"
"github.com/memohai/memoh/internal/channel"
"github.com/memohai/memoh/internal/settings"
ttspkg "github.com/memohai/memoh/internal/tts"
)
const ttsMaxTextLen = 500
// TTSSender sends outbound messages through the channel manager.
type TTSSender interface {
Send(ctx context.Context, botID string, channelType channel.ChannelType, req channel.SendRequest) error
}
// TTSChannelResolver parses platform name to channel type.
type TTSChannelResolver interface {
ParseChannelType(raw string) (channel.ChannelType, error)
}
type TTSProvider struct {
logger *slog.Logger
settings *settings.Service
tts *ttspkg.Service
sender TTSSender
resolver TTSChannelResolver
}
func NewTTSProvider(log *slog.Logger, settingsSvc *settings.Service, ttsSvc *ttspkg.Service, sender TTSSender, resolver TTSChannelResolver) *TTSProvider {
if log == nil {
log = slog.Default()
}
return &TTSProvider{
logger: log.With(slog.String("tool", "tts")),
settings: settingsSvc,
tts: ttsSvc,
sender: sender,
resolver: resolver,
}
}
func (p *TTSProvider) Tools(ctx context.Context, session SessionContext) ([]sdk.Tool, error) {
if session.IsSubagent || p.settings == nil || p.tts == nil || p.sender == nil || p.resolver == nil {
return nil, nil
}
botID := strings.TrimSpace(session.BotID)
if botID == "" {
return nil, nil
}
botSettings, err := p.settings.GetBot(ctx, botID)
if err != nil {
return nil, nil
}
if strings.TrimSpace(botSettings.TtsModelID) == "" {
return nil, nil
}
sess := session
return []sdk.Tool{
{
Name: "speak",
Description: "Send a voice message. When target is omitted, speaks in the current conversation. When target is specified, sends to that channel/person. Synthesizes text to speech and delivers as audio.",
Parameters: map[string]any{
"type": "object",
"properties": map[string]any{
"text": map[string]any{"type": "string", "description": "The text to convert to speech (max 500 characters)"},
"platform": map[string]any{"type": "string", "description": "Channel platform name. Defaults to current session platform."},
"target": map[string]any{"type": "string", "description": "Channel target (chat/group/thread ID). Optional — omit to speak in the current conversation. Use get_contacts to find targets for other conversations."},
"reply_to": map[string]any{"type": "string", "description": "Message ID to reply to. The voice message will reference this message on the platform."},
},
"required": []string{"text"},
},
Execute: func(execCtx *sdk.ToolExecContext, input any) (any, error) {
return p.execSpeak(execCtx.Context, sess, inputAsMap(input))
},
},
}, nil
}
func (p *TTSProvider) execSpeak(ctx context.Context, session SessionContext, args map[string]any) (any, error) {
botID := strings.TrimSpace(session.BotID)
if botID == "" {
return nil, errors.New("bot_id is required")
}
text := strings.TrimSpace(StringArg(args, "text"))
if text == "" {
return nil, errors.New("text is required")
}
if len([]rune(text)) > ttsMaxTextLen {
return nil, errors.New("text too long, max 500 characters")
}
channelType, err := p.resolvePlatform(args, session)
if err != nil {
return nil, err
}
target := FirstStringArg(args, "target")
if target == "" {
target = strings.TrimSpace(session.ReplyTarget)
}
isSameConv := target == "" || session.IsSameConversation(channelType.String(), target)
botSettings, err := p.settings.GetBot(ctx, botID)
if err != nil {
return nil, errors.New("failed to load bot settings")
}
if botSettings.TtsModelID == "" {
return nil, errors.New("bot has no TTS model configured")
}
audioData, contentType, synthErr := p.tts.Synthesize(ctx, botSettings.TtsModelID, text, nil)
if synthErr != nil {
return nil, fmt.Errorf("speech synthesis failed: %s", synthErr.Error())
}
dataURL := fmt.Sprintf("data:%s;base64,%s", contentType, base64.StdEncoding.EncodeToString(audioData))
// Same-conversation: emit the synthesized audio as a voice attachment.
if isSameConv && session.Emitter != nil {
session.Emitter(ToolStreamEvent{
Type: StreamEventAttachment,
Attachments: []Attachment{{
Type: "voice",
URL: dataURL,
Mime: contentType,
Size: int64(len(audioData)),
}},
})
return map[string]any{
"ok": true,
"delivered": "current_conversation",
}, nil
}
if target == "" {
return nil, errors.New("target is required for cross-conversation speak")
}
msg := channel.Message{
Attachments: []channel.Attachment{{Type: channel.AttachmentVoice, URL: dataURL, Mime: contentType, Size: int64(len(audioData))}},
}
if replyTo := FirstStringArg(args, "reply_to"); replyTo != "" {
msg.Reply = &channel.ReplyRef{MessageID: replyTo}
}
if err := p.sender.Send(ctx, botID, channelType, channel.SendRequest{Target: target, Message: msg}); err != nil {
return nil, err
}
return map[string]any{
"ok": true, "bot_id": botID, "platform": channelType.String(), "target": target,
"instruction": "Voice message delivered successfully. You have completed your response. Please STOP now and do not call any more tools.",
}, nil
}
func (p *TTSProvider) resolvePlatform(args map[string]any, session SessionContext) (channel.ChannelType, error) {
platform := FirstStringArg(args, "platform")
if platform == "" {
platform = strings.TrimSpace(session.CurrentPlatform)
}
if platform == "" {
return "", errors.New("platform is required")
}
return p.resolver.ParseChannelType(platform)
}