feat(agent): add extensible tag interception system and inline reactions

Refactor the attachment tag extraction into a generic TagResolver/StreamTagExtractor
system that supports multiple custom tags. Implement <reactions> tag allowing the
agent to embed emoji reactions directly in text responses, dispatched as side-effects
through the channel reactor interface.

- Add TagResolver interface and StreamTagExtractor streaming state machine
- Refactor AttachmentsStreamExtractor as backward-compatible wrapper
- Add reactionsResolver and ReactionDeltaAction stream event
- Wire reaction dispatch in Go channel inbound processor
- Fix .gitignore to scope compiled binary patterns to repo root

.gitignore

@@ -91,8 +91,8 @@ Thumbs.db
 tmp/
 
 # compiled files
-memoh
-agent
+/memoh
+/agent
 
 docs/docs/.vitepress/cache
 .pnpm-store

apps/web/src/composables/api/useChat.types.ts

@@ -46,7 +46,7 @@ export interface StreamEvent {
     | 'text_start' | 'text_delta' | 'text_end'
     | 'reasoning_start' | 'reasoning_delta' | 'reasoning_end'
     | 'tool_call_start' | 'tool_call_end'
-    | 'attachment_delta'
+    | 'attachment_delta' | 'reaction_delta'
     | 'agent_start' | 'agent_end' | 'agent_abort'
     | 'processing_started' | 'processing_completed' | 'processing_failed'
     | 'error'

cmd/agent/main.go

@@ -448,6 +448,7 @@ func provideChannelManager(log *slog.Logger, registry *channel.Registry, channel
 	if mw := channelRouter.IdentityMiddleware(); mw != nil {
 		mgr.Use(mw)
 	}
+	channelRouter.SetReactor(mgr)
 	return mgr
 }
 

internal/channel/inbound/channel.go

@@ -39,6 +39,10 @@ type RouteResolver interface {
 	ResolveConversation(ctx context.Context, input route.ResolveInput) (route.ResolveConversationResult, error)
 }
 
+type channelReactor interface {
+	React(ctx context.Context, botID string, channelType channel.ChannelType, req channel.ReactRequest) error
+}
+
 type mediaIngestor interface {
 	Ingest(ctx context.Context, input media.IngestInput) (media.Asset, error)
 	// GetByStorageKey resolves an asset by reading its sidecar JSON.
@@ -55,6 +59,7 @@ type ChannelInboundProcessor struct {
 	routeResolver RouteResolver
 	message       messagepkg.Writer
 	mediaService  mediaIngestor
+	reactor       channelReactor
 	inboxService  *inbox.Service
 	registry      *channel.Registry
 	logger        *slog.Logger
@@ -114,6 +119,14 @@ func (p *ChannelInboundProcessor) SetMediaService(mediaService mediaIngestor) {
 	p.mediaService = mediaService
 }
 
+// SetReactor configures the channel reactor for handling inline emoji reactions.
+func (p *ChannelInboundProcessor) SetReactor(reactor channelReactor) {
+	if p == nil {
+		return
+	}
+	p.reactor = reactor
+}
+
 // SetStreamObserver configures an observer that receives copies of all stream
 // events produced for non-local channels (e.g. Telegram, Feishu). This enables
 // cross-channel visibility in the WebUI without coupling adapters to the hub.
@@ -443,6 +456,10 @@ func (p *ChannelInboundProcessor) HandleInbound(ctx context.Context, cfg channel
 					}
 					assetMu.Unlock()
 				}
+				if event.Type == channel.StreamEventReaction && len(event.Reactions) > 0 {
+					p.dispatchReactions(ctx, identity.BotID, msg.Channel, target, sourceMessageID, event.Reactions)
+					continue
+				}
 				if pushErr := stream.Push(ctx, events[i]); pushErr != nil {
 					streamErr = pushErr
 					break
@@ -861,6 +878,7 @@ type gatewayStreamEnvelope struct {
 	Input       json.RawMessage `json:"input"`
 	Result      json.RawMessage `json:"result"`
 	Attachments json.RawMessage `json:"attachments"`
+	Reactions   json.RawMessage `json:"reactions"`
 }
 
 type gatewayStreamDoneData struct {
@@ -954,6 +972,14 @@ func mapStreamChunkToChannelEvents(chunk conversation.StreamChunk) ([]channel.St
 		return []channel.StreamEvent{
 			{Type: channel.StreamEventAttachment, Attachments: attachments},
 		}, finalMessages, nil
+	case "reaction_delta":
+		reactions := parseReactionDelta(envelope.Reactions)
+		if len(reactions) == 0 {
+			return nil, finalMessages, nil
+		}
+		return []channel.StreamEvent{
+			{Type: channel.StreamEventReaction, Reactions: reactions},
+		}, finalMessages, nil
 	case "agent_start":
 		return []channel.StreamEvent{
 			{
@@ -1935,6 +1961,73 @@ func parseAttachmentDelta(raw json.RawMessage) []channel.Attachment {
 	return attachments
 }
 
+// parseReactionDelta converts raw JSON reaction data to channel ReactRequests.
+func parseReactionDelta(raw json.RawMessage) []channel.ReactRequest {
+	if len(raw) == 0 {
+		return nil
+	}
+	var items []struct {
+		Emoji string `json:"emoji"`
+	}
+	if err := json.Unmarshal(raw, &items); err != nil {
+		return nil
+	}
+	reactions := make([]channel.ReactRequest, 0, len(items))
+	for _, item := range items {
+		emoji := strings.TrimSpace(item.Emoji)
+		if emoji == "" {
+			continue
+		}
+		reactions = append(reactions, channel.ReactRequest{
+			Emoji: emoji,
+		})
+	}
+	return reactions
+}
+
+// dispatchReactions sends emoji reactions to the channel for the source message.
+func (p *ChannelInboundProcessor) dispatchReactions(
+	ctx context.Context,
+	botID string,
+	channelType channel.ChannelType,
+	target string,
+	sourceMessageID string,
+	reactions []channel.ReactRequest,
+) {
+	if p.reactor == nil {
+		return
+	}
+	target = strings.TrimSpace(target)
+	sourceMessageID = strings.TrimSpace(sourceMessageID)
+	if target == "" || sourceMessageID == "" {
+		if p.logger != nil {
+			p.logger.Warn("cannot dispatch reactions: missing target or source message ID",
+				slog.String("bot_id", botID),
+				slog.String("channel", channelType.String()),
+			)
+		}
+		return
+	}
+	for _, reaction := range reactions {
+		req := channel.ReactRequest{
+			Target:    target,
+			MessageID: sourceMessageID,
+			Emoji:     reaction.Emoji,
+		}
+		if err := p.reactor.React(ctx, strings.TrimSpace(botID), channelType, req); err != nil {
+			if p.logger != nil {
+				p.logger.Warn("inline reaction failed",
+					slog.String("bot_id", botID),
+					slog.String("channel", channelType.String()),
+					slog.String("emoji", reaction.Emoji),
+					slog.String("message_id", sourceMessageID),
+					slog.Any("error", err),
+				)
+			}
+		}
+	}
+}
+
 // buildRouteMetadata extracts user/conversation information for route metadata persistence.
 func buildRouteMetadata(msg channel.InboundMessage, identity InboundIdentity) map[string]any {
 	m := make(map[string]any)

internal/channel/types.go

@@ -101,6 +101,7 @@ const (
 	StreamEventAttachment          StreamEventType = "attachment"
 	StreamEventAgentStart          StreamEventType = "agent_start"
 	StreamEventAgentEnd            StreamEventType = "agent_end"
+	StreamEventReaction            StreamEventType = "reaction"
 	StreamEventProcessingStarted   StreamEventType = "processing_started"
 	StreamEventProcessingCompleted StreamEventType = "processing_completed"
 	StreamEventProcessingFailed    StreamEventType = "processing_failed"
@@ -146,6 +147,7 @@ type StreamEvent struct {
 	ToolCall    *StreamToolCall        `json:"tool_call,omitempty"`
 	Phase       StreamPhase            `json:"phase,omitempty"`
 	Attachments []Attachment           `json:"attachments,omitempty"`
+	Reactions   []ReactRequest         `json:"reactions,omitempty"`
 	Metadata    map[string]any         `json:"metadata,omitempty"`
 }
 

packages/agent/src/agent.ts

@@ -25,11 +25,13 @@ import { system, schedule, heartbeat, subagentSystem } from './prompts'
 import { AuthFetcher } from './types'
 import { createModel } from './model'
 import {
-  extractAttachmentsFromText,
   stripAttachmentsFromMessages,
   dedupeAttachments,
-  AttachmentsStreamExtractor,
+  attachmentsResolver,
 } from './utils/attachments'
+import type { ContainerFileAttachment } from './types/attachment'
+import { reactionsResolver, type ReactionItem } from './utils/reactions'
+import { StreamTagExtractor, extractTagsFromText, type TagEvent } from './utils/tag-extractor'
 import { createImagePartFromAttachment } from './utils/image-parts'
 import type { GatewayInputAttachment } from './types/attachment'
 import { getMCPTools } from './tools/mcp'
@@ -330,10 +332,16 @@ export const createAgent = (
       basePrepareStep: () => ({ system: systemPrompt }),
     })
     const stepUsages = buildStepUsages(steps)
-    const { cleanedText, attachments: textAttachments } =
-      extractAttachmentsFromText(text)
+    const tagResolvers = [attachmentsResolver, reactionsResolver]
+    const { cleanedText, events } = extractTagsFromText(text, tagResolvers)
+    const textAttachments = events
+      .filter((e) => e.tag === 'attachments')
+      .flatMap((e) => e.data as ContainerFileAttachment[])
+    const reactions = events
+      .filter((e) => e.tag === 'reactions')
+      .flatMap((e) => e.data as ReactionItem[])
     const { messages: strippedMessages, attachments: messageAttachments } =
-      stripAttachmentsFromMessages(response.messages)
+      stripAttachmentsFromMessages(response.messages, [reactionsResolver])
     const allAttachments = dedupeAttachments([
       ...textAttachments,
       ...messageAttachments,
@@ -348,6 +356,7 @@ export const createAgent = (
       usage,
       text: cleanedText,
       attachments: allAttachments,
+      reactions,
       skills: getEnabledSkills(),
     }
   }
@@ -472,12 +481,34 @@ export const createAgent = (
     return 'Model stream failed'
   }
 
+  function* emitTagEvents(events: TagEvent[]): Generator<AgentStreamAction> {
+    for (const event of events) {
+      switch (event.tag) {
+        case 'attachments': {
+          const attachments = dedupeAttachments(event.data as ContainerFileAttachment[]) as ContainerFileAttachment[]
+          if (attachments.length) {
+            yield { type: 'attachment_delta', attachments }
+          }
+          break
+        }
+        case 'reactions': {
+          const reactions = event.data as ReactionItem[]
+          if (reactions.length) {
+            yield { type: 'reaction_delta', reactions }
+          }
+          break
+        }
+      }
+    }
+  }
+
   async function* stream(input: AgentInput): AsyncGenerator<AgentStreamAction> {
     const userPrompt = generateUserPrompt(input)
     const messages = [...input.messages, userPrompt]
     input.skills.forEach((skill) => enableSkill(skill))
     const systemPrompt = await generateSystemPrompt()
-    const attachmentsExtractor = new AttachmentsStreamExtractor()
+    const tagResolvers = [attachmentsResolver, reactionsResolver]
+    const tagExtractor = new StreamTagExtractor(tagResolvers)
     const textLoopGuard = loopDetectionEnabled
       ? createTextLoopGuard({
         consecutiveHitsToAbort: LOOP_DETECTED_STREAK_THRESHOLD,
@@ -598,9 +629,7 @@ export const createAgent = (
             }
             break
           case 'text-delta': {
-            const { visibleText, attachments } = attachmentsExtractor.push(
-              chunk.text,
-            )
+            const { visibleText, events } = tagExtractor.push(chunk.text)
             if (visibleText) {
               if (textLoopProbeBuffer) {
                 textLoopProbeBuffer.push(visibleText)
@@ -610,16 +639,11 @@ export const createAgent = (
                 delta: visibleText,
               }
             }
-            if (attachments.length) {
-              yield {
-                type: 'attachment_delta',
-                attachments,
-              }
-            }
+            yield* emitTagEvents(events)
             break
           }
           case 'text-end': {
-            const remainder = attachmentsExtractor.flushRemainder()
+            const remainder = tagExtractor.flushRemainder()
             if (remainder.visibleText) {
               if (textLoopProbeBuffer) {
                 textLoopProbeBuffer.push(remainder.visibleText)
@@ -632,20 +656,15 @@ export const createAgent = (
             if (textLoopProbeBuffer) {
               textLoopProbeBuffer.flush()
             }
-            if (remainder.attachments.length) {
-              yield {
-                type: 'attachment_delta',
-                attachments: remainder.attachments,
-              }
-            }
+            yield* emitTagEvents(remainder.events)
             yield {
               type: 'text_end',
               metadata: chunk,
             }
             break
           }
-          case 'tool-call':
-            const remainder = attachmentsExtractor.flushRemainder()
+          case 'tool-call': {
+            const remainder = tagExtractor.flushRemainder()
             if (remainder.visibleText) {
               if (textLoopProbeBuffer) {
                 textLoopProbeBuffer.push(remainder.visibleText)
@@ -658,12 +677,7 @@ export const createAgent = (
             if (textLoopProbeBuffer) {
               textLoopProbeBuffer.flush()
             }
-            if (remainder.attachments.length) {
-              yield {
-                type: 'attachment_delta',
-                attachments: remainder.attachments,
-              }
-            }
+            yield* emitTagEvents(remainder.events)
             yield {
               type: 'tool_call_start',
               toolName: chunk.toolName,
@@ -672,7 +686,8 @@ export const createAgent = (
               metadata: chunk,
             }
             break
-          case 'tool-result':
+          }
+          case 'tool-result': {
             const shouldAbortForToolLoop = toolLoopAbortCallIds.delete(chunk.toolCallId)
             yield {
               type: 'tool_call_end',
@@ -691,6 +706,7 @@ export const createAgent = (
               throw new Error(TOOL_LOOP_DETECTED_ABORT_MESSAGE)
             }
             break
+          }
           case 'file':
             yield {
               type: 'attachment_delta',
@@ -710,6 +726,7 @@ export const createAgent = (
   
       const { messages: strippedMessages } = stripAttachmentsFromMessages(
         result.messages,
+        [reactionsResolver],
       )
       yield {
         type: 'agent_end',

packages/agent/src/prompts/system.ts

@@ -206,6 +206,22 @@ Rules:
 - No extra text inside ${quote('<attachments>...</attachments>')}
 - The block can appear anywhere in your response; it will be parsed and stripped from visible text
 
+## Reactions
+
+To react with an emoji to the message you are replying to, use this format in your direct response:
+
+${block([
+  '<reactions>',
+  '- 👍',
+  '</reactions>',
+].join('\n'))}
+
+Rules:
+- One emoji per line, prefixed by ${quote('- ')}
+- The block can appear anywhere in your response; it will be parsed and stripped from visible text
+- This reacts to the **source message** of the current conversation (the message you are responding to)
+- For reacting to messages in other channels or removing reactions, use the ${quote('react')} tool instead
+
 ## Schedule Tasks
 
 You can create and manage schedule tasks via cron.

packages/agent/src/types/action.ts

@@ -1,6 +1,7 @@
 import { LanguageModelUsage, ModelMessage } from 'ai'
 import { AgentInput } from './agent'
 import { AgentAttachment } from './attachment'
+import { ReactionItem } from '../utils/reactions'
 
 export interface BaseAction {
   type: string
@@ -39,6 +40,11 @@ export interface AttachmentDeltaAction extends BaseAction {
   attachments: AgentAttachment[]
 }
 
+export interface ReactionDeltaAction extends BaseAction {
+  type: 'reaction_delta'
+  reactions: ReactionItem[]
+}
+
 export interface TextEndAction extends BaseAction {
   type: 'text_end'
 }
@@ -84,6 +90,7 @@ export type AgentStreamAction =
   | TextStartAction
   | TextDeltaAction
   | AttachmentDeltaAction
+  | ReactionDeltaAction
   | TextEndAction
   | ToolCallStartAction
   | ToolCallEndAction

packages/agent/src/utils/attachments.ts

@@ -3,9 +3,12 @@ import type {
   AgentAttachment,
   ContainerFileAttachment,
 } from '../types/attachment'
+import type { TagResolver } from './tag-extractor'
+import { StreamTagExtractor, extractTagsFromText } from './tag-extractor'
 
-const ATTACHMENTS_START = '<attachments>'
-const ATTACHMENTS_END = '</attachments>'
+// ---------------------------------------------------------------------------
+// Helpers
+// ---------------------------------------------------------------------------
 
 /**
  * Get a unique key for deduplication of attachments.
@@ -39,26 +42,41 @@ export const parseAttachmentPaths = (content: string): string[] => {
     .filter(Boolean)
 }
 
+// ---------------------------------------------------------------------------
+// TagResolver for <attachments>
+// ---------------------------------------------------------------------------
+
+export const attachmentsResolver: TagResolver<ContainerFileAttachment> = {
+  tag: 'attachments',
+  parse(content: string): ContainerFileAttachment[] {
+    const paths = Array.from(new Set(parseAttachmentPaths(content)))
+    return paths.map((path): ContainerFileAttachment => ({ type: 'file', path }))
+  },
+}
+
+// ---------------------------------------------------------------------------
+// Batch extraction (backward-compatible wrapper)
+// ---------------------------------------------------------------------------
+
 /**
  * Extract all `<attachments>...</attachments>` blocks from a text string.
  * Returns the cleaned text (blocks removed) and the parsed file attachments.
  */
 export const extractAttachmentsFromText = (text: string): { cleanedText: string; attachments: ContainerFileAttachment[] } => {
-  const paths: string[] = []
-  const cleanedText = text.replace(
-    /<attachments>([\s\S]*?)<\/attachments>/g,
-    (_match, inner: string) => { 
-      paths.push(...parseAttachmentPaths(inner))
-      return ''
-    }
-  )
-  const uniquePaths = Array.from(new Set(paths))
+  const { cleanedText, events } = extractTagsFromText(text, [attachmentsResolver])
+  const attachments = events
+    .filter((e) => e.tag === 'attachments')
+    .flatMap((e) => e.data as ContainerFileAttachment[])
   return {
-    cleanedText: cleanedText.replace(/\n{3,}/g, '\n\n').trim(),
-    attachments: uniquePaths.map((path): ContainerFileAttachment => ({ type: 'file', path })),
+    cleanedText,
+    attachments: dedupeAttachments(attachments) as ContainerFileAttachment[],
   }
 }
 
+// ---------------------------------------------------------------------------
+// Message-level stripping
+// ---------------------------------------------------------------------------
+
 /**
  * Type guard: checks whether a content part is a TextPart.
  */
@@ -72,13 +90,25 @@ const isTextPart = (part: unknown): part is TextPart => {
 }
 
 /**
- * Strip `<attachments>` blocks from all assistant messages in a message list.
+ * Strip all registered tag blocks from assistant messages in a message list.
+ * Accepts additional resolvers to strip beyond `<attachments>` (e.g. `<reactions>`).
  * Returns the cleaned messages and a deduplicated list of attachments found.
  */
 export const stripAttachmentsFromMessages = (
-  messages: ModelMessage[]
+  messages: ModelMessage[],
+  extraResolvers: TagResolver[] = [],
 ): { messages: ModelMessage[]; attachments: ContainerFileAttachment[] } => {
   const allAttachments: ContainerFileAttachment[] = []
+  const resolvers: TagResolver[] = [attachmentsResolver, ...extraResolvers]
+
+  const cleanText = (text: string): string => {
+    const { cleanedText, events } = extractTagsFromText(text, resolvers)
+    const attachments = events
+      .filter((e) => e.tag === 'attachments')
+      .flatMap((e) => e.data as ContainerFileAttachment[])
+    allAttachments.push(...attachments)
+    return cleanedText
+  }
 
   const stripped = messages.map((msg): ModelMessage => {
     if (msg.role !== 'assistant') return msg
@@ -87,17 +117,13 @@ export const stripAttachmentsFromMessages = (
     const { content } = assistantMsg
 
     if (typeof content === 'string') {
-      const { cleanedText, attachments } = extractAttachmentsFromText(content)
-      allAttachments.push(...attachments)
-      return { ...assistantMsg, content: cleanedText }
+      return { ...assistantMsg, content: cleanText(content) }
     }
 
     if (Array.isArray(content)) {
       const newParts = content.map(part => {
         if (!isTextPart(part)) return part
-        const { cleanedText, attachments } = extractAttachmentsFromText(part.text)
-        allAttachments.push(...attachments)
-        return { ...part, text: cleanedText }
+        return { ...part, text: cleanText(part.text) }
       })
       return { ...assistantMsg, content: newParts }
     }
@@ -112,7 +138,7 @@ export const stripAttachmentsFromMessages = (
 }
 
 // ---------------------------------------------------------------------------
-// Streaming extractor
+// Streaming extractor (backward-compatible wrapper)
 // ---------------------------------------------------------------------------
 
 export interface AttachmentsStreamResult {
@@ -121,81 +147,35 @@ export interface AttachmentsStreamResult {
 }
 
 /**
- * Incremental state-machine that intercepts `<attachments>...</attachments>`
- * blocks from a stream of text deltas. Text outside those blocks is passed
- * through as `visibleText`; completed blocks are emitted as `attachments`.
+ * Backward-compatible streaming extractor that delegates to {@link StreamTagExtractor}.
+ * Intercepts `<attachments>...</attachments>` blocks from a stream of text deltas.
  */
 export class AttachmentsStreamExtractor {
-  private state: 'text' | 'attachments' = 'text'
-  private buffer = ''
-  private attachmentsBuffer = ''
+  private inner: StreamTagExtractor
 
-  /**
-   * Feed a new text delta into the extractor.
-   */
-  push(delta: string): AttachmentsStreamResult {
-    this.buffer += delta
-    let visible = ''
-    const attachments: ContainerFileAttachment[] = []
-
-    while (this.buffer.length > 0) {
-      if (this.state === 'text') {
-        const idx = this.buffer.indexOf(ATTACHMENTS_START)
-        if (idx === -1) {
-          // Emit everything except a small tail that could be the start of the opening tag.
-          const keep = Math.min(this.buffer.length, ATTACHMENTS_START.length - 1)
-          const emit = this.buffer.slice(0, this.buffer.length - keep)
-          visible += emit
-          this.buffer = this.buffer.slice(this.buffer.length - keep)
-          break
-        }
-
-        visible += this.buffer.slice(0, idx)
-        this.buffer = this.buffer.slice(idx + ATTACHMENTS_START.length)
-        this.attachmentsBuffer = ''
-        this.state = 'attachments'
-        continue
-      }
-
-      // state === 'attachments'
-      const endIdx = this.buffer.indexOf(ATTACHMENTS_END)
-      if (endIdx === -1) {
-        const keep = Math.min(this.buffer.length, ATTACHMENTS_END.length - 1)
-        const take = this.buffer.slice(0, this.buffer.length - keep)
-        this.attachmentsBuffer += take
-        this.buffer = this.buffer.slice(this.buffer.length - keep)
-        break
-      }
-
-      this.attachmentsBuffer += this.buffer.slice(0, endIdx)
-      const paths = parseAttachmentPaths(this.attachmentsBuffer)
-      for (const path of paths) {
-        attachments.push({ type: 'file', path })
-      }
-      this.buffer = this.buffer.slice(endIdx + ATTACHMENTS_END.length)
-      this.attachmentsBuffer = ''
-      this.state = 'text'
-    }
-
-    return { visibleText: visible, attachments: dedupeAttachments(attachments) as ContainerFileAttachment[] }
+  constructor() {
+    this.inner = new StreamTagExtractor([attachmentsResolver])
   }
 
-  /**
-   * Flush any remaining buffered content. Call this when the stream ends.
-   * If an `<attachments>` block was not properly closed, the raw text is
-   * returned as `visibleText` to avoid data loss.
-   */
-  flushRemainder(): AttachmentsStreamResult {
-    if (this.state === 'text') {
-      const out = this.buffer
-      this.buffer = ''
-      return { visibleText: out, attachments: [] }
+  push(delta: string): AttachmentsStreamResult {
+    const { visibleText, events } = this.inner.push(delta)
+    const attachments = events
+      .filter((e) => e.tag === 'attachments')
+      .flatMap((e) => e.data as ContainerFileAttachment[])
+    return {
+      visibleText,
+      attachments: dedupeAttachments(attachments) as ContainerFileAttachment[],
+    }
+  }
+
+  flushRemainder(): AttachmentsStreamResult {
+    const { visibleText, events } = this.inner.flushRemainder()
+    const attachments = events
+      .filter((e) => e.tag === 'attachments')
+      .flatMap((e) => e.data as ContainerFileAttachment[])
+    return {
+      visibleText,
+      attachments: dedupeAttachments(attachments) as ContainerFileAttachment[],
     }
-    // Unclosed attachments block — treat it as literal text to avoid data loss.
-    const out = `${ATTACHMENTS_START}${this.attachmentsBuffer}${this.buffer}`
-    this.state = 'text'
-    this.buffer = ''
-    this.attachmentsBuffer = ''
-    return { visibleText: out, attachments: [] }
   }
 }

packages/agent/src/utils/reactions.ts

@@ -0,0 +1,28 @@
+import type { TagResolver } from './tag-extractor'
+
+export interface ReactionItem {
+  emoji: string
+}
+
+/**
+ * Parse emoji entries from the inner content of a `<reactions>` block.
+ * Each line should be formatted as `- 👍`.
+ */
+export const parseReactionEmojis = (content: string): string[] => {
+  return content
+    .split('\n')
+    .map(line => line.trim())
+    .map(line => {
+      if (!line.startsWith('-')) return ''
+      return line.slice(1).trim()
+    })
+    .filter(Boolean)
+}
+
+export const reactionsResolver: TagResolver<ReactionItem> = {
+  tag: 'reactions',
+  parse(content: string): ReactionItem[] {
+    const emojis = Array.from(new Set(parseReactionEmojis(content)))
+    return emojis.map((emoji): ReactionItem => ({ emoji }))
+  },
+}

packages/agent/src/utils/tag-extractor.ts

@@ -0,0 +1,183 @@
+/**
+ * Generic extensible tag-interception system.
+ *
+ * Register TagResolver instances (e.g. attachments, reactions) and both the
+ * batch extractor and the streaming state-machine will intercept the
+ * corresponding `<tag>...</tag>` blocks, stripping them from visible text and
+ * forwarding the parsed payload through {@link TagEvent} objects.
+ */
+
+// ---------------------------------------------------------------------------
+// Public interfaces
+// ---------------------------------------------------------------------------
+
+export interface TagResolver<T = unknown> {
+  tag: string
+  parse(content: string): T[]
+}
+
+export interface TagEvent {
+  tag: string
+  data: unknown[]
+}
+
+export interface TagStreamResult {
+  visibleText: string
+  events: TagEvent[]
+}
+
+// ---------------------------------------------------------------------------
+// Batch extractor
+// ---------------------------------------------------------------------------
+
+/**
+ * Extract all registered tag blocks from a complete text string.
+ * Returns the cleaned text (blocks removed) and a list of tag events.
+ */
+export function extractTagsFromText(
+  text: string,
+  resolvers: TagResolver[],
+): { cleanedText: string; events: TagEvent[] } {
+  const events: TagEvent[] = []
+  let cleaned = text
+  for (const resolver of resolvers) {
+    const open = `<${resolver.tag}>`
+    const close = `</${resolver.tag}>`
+    const pattern = new RegExp(
+      `${escapeRegExp(open)}([\\s\\S]*?)${escapeRegExp(close)}`,
+      'g',
+    )
+    cleaned = cleaned.replace(pattern, (_match, inner: string) => {
+      const parsed = resolver.parse(inner)
+      if (parsed.length > 0) {
+        events.push({ tag: resolver.tag, data: parsed })
+      }
+      return ''
+    })
+  }
+  return {
+    cleanedText: cleaned.replace(/\n{3,}/g, '\n\n').trim(),
+    events,
+  }
+}
+
+// ---------------------------------------------------------------------------
+// Streaming extractor
+// ---------------------------------------------------------------------------
+
+interface ResolverMeta {
+  resolver: TagResolver
+  openTag: string
+  closeTag: string
+}
+
+/**
+ * Incremental state-machine that intercepts multiple `<tag>...</tag>` blocks
+ * from a stream of text deltas.
+ *
+ * Text outside registered blocks is passed through as `visibleText`; completed
+ * blocks are emitted as {@link TagEvent} entries.
+ */
+export class StreamTagExtractor {
+  private metas: ResolverMeta[]
+  private maxOpenLen: number
+  private state: 'text' | 'inside' = 'text'
+  private activeMeta: ResolverMeta | null = null
+  private buffer = ''
+  private tagBuffer = ''
+
+  constructor(resolvers: TagResolver[]) {
+    this.metas = resolvers.map((resolver) => ({
+      resolver,
+      openTag: `<${resolver.tag}>`,
+      closeTag: `</${resolver.tag}>`,
+    }))
+    this.maxOpenLen = Math.max(...this.metas.map((m) => m.openTag.length), 0)
+  }
+
+  push(delta: string): TagStreamResult {
+    this.buffer += delta
+    let visible = ''
+    const events: TagEvent[] = []
+
+    while (this.buffer.length > 0) {
+      if (this.state === 'text') {
+        let earliestIdx = -1
+        let matchedMeta: ResolverMeta | null = null
+
+        for (const meta of this.metas) {
+          const idx = this.buffer.indexOf(meta.openTag)
+          if (idx !== -1 && (earliestIdx === -1 || idx < earliestIdx)) {
+            earliestIdx = idx
+            matchedMeta = meta
+          }
+        }
+
+        if (earliestIdx === -1) {
+          const keep = Math.min(this.buffer.length, this.maxOpenLen - 1)
+          const emit = this.buffer.slice(0, this.buffer.length - keep)
+          visible += emit
+          this.buffer = this.buffer.slice(this.buffer.length - keep)
+          break
+        }
+
+        visible += this.buffer.slice(0, earliestIdx)
+        this.buffer = this.buffer.slice(earliestIdx + matchedMeta!.openTag.length)
+        this.tagBuffer = ''
+        this.activeMeta = matchedMeta
+        this.state = 'inside'
+        continue
+      }
+
+      // state === 'inside'
+      const closeTag = this.activeMeta!.closeTag
+      const endIdx = this.buffer.indexOf(closeTag)
+      if (endIdx === -1) {
+        const keep = Math.min(this.buffer.length, closeTag.length - 1)
+        const take = this.buffer.slice(0, this.buffer.length - keep)
+        this.tagBuffer += take
+        this.buffer = this.buffer.slice(this.buffer.length - keep)
+        break
+      }
+
+      this.tagBuffer += this.buffer.slice(0, endIdx)
+      const parsed = this.activeMeta!.resolver.parse(this.tagBuffer)
+      if (parsed.length > 0) {
+        events.push({ tag: this.activeMeta!.resolver.tag, data: parsed })
+      }
+      this.buffer = this.buffer.slice(endIdx + closeTag.length)
+      this.tagBuffer = ''
+      this.activeMeta = null
+      this.state = 'text'
+    }
+
+    return { visibleText: visible, events }
+  }
+
+  /**
+   * Flush remaining buffered content. Call when the stream ends.
+   * Unclosed tag blocks are returned as literal `visibleText` to avoid data loss.
+   */
+  flushRemainder(): TagStreamResult {
+    if (this.state === 'text') {
+      const out = this.buffer
+      this.buffer = ''
+      return { visibleText: out, events: [] }
+    }
+    const meta = this.activeMeta!
+    const out = `${meta.openTag}${this.tagBuffer}${this.buffer}`
+    this.state = 'text'
+    this.buffer = ''
+    this.tagBuffer = ''
+    this.activeMeta = null
+    return { visibleText: out, events: [] }
+  }
+}
+
+// ---------------------------------------------------------------------------
+// Helpers
+// ---------------------------------------------------------------------------
+
+function escapeRegExp(s: string): string {
+  return s.replace(/[.*+?^${}()|[\]\\]/g, '\\$&')
+}