feat: v0.5.16 - migrate to Responses API (#586)

* refactor: migrate from /v1/runs to /v1/responses streaming API Replace EventSource-based polling with direct SSE streaming via the /v1/responses endpoint across all server-side callers (chat-run-socket, context-compressor, gateway-client, agent-clients). Messages are now written to DB in real-time during streaming, eliminating post-run sync. Frontend chat store adds tool_call_id tracking for deduplication. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> * chore: bump version to 0.5.16 and add changelog - Persist real API usage to usage table on response.completed - Remove unused codex_reasoning_items field from message schema - Fix unused variable warnings in chat-run-socket - Bump version to 0.5.16 - Add changelog entries for 0.5.16 (8 locales) Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
2026-05-10 02:49:58 +08:00
parent a36c0a3095
commit 50122c5ff8
19 changed files with 843 additions and 866 deletions
@@ -1,6 +1,6 @@
 {
  "name": "hermes-web-ui",
-  "version": "0.5.15",
+  "version": "0.5.16",
  "description": "Self-hosted AI chat dashboard for Hermes Agent — multi-model (Claude, GPT, Gemini, DeepSeek) web UI with Telegram, Discord, Slack, WhatsApp integration",
  "repository": {
    "type": "git",
@@ -5,6 +5,16 @@ export interface ChangelogEntry {
 }

 export const changelog: ChangelogEntry[] = [
+  {
+    version: '0.5.16',
+    date: '2026-05-10',
+    changes: [
+      'changelog.new_0_5_16_1',
+      'changelog.new_0_5_16_2',
+      'changelog.new_0_5_16_3',
+      'changelog.new_0_5_16_4',
+    ],
+  },
  {
    version: '0.5.15',
    date: '2026-05-09',
@@ -665,6 +665,10 @@ jobTriggered: 'Job ausgelost',
    new_0_5_15_9: 'Hermes Markdown-Medien-Rendering und Sync-Wiederholung korrigiert',
    new_0_5_15_10: 'Upstream-Umgebungsvariablenabhängigkeit entfernt',
    new_0_5_15_11: 'Wenn die Kanban-Funktion nicht verfügbar ist, updaten Sie bitte hermes-agent',
+    new_0_5_16_1: 'Chat-Streaming von /v1/runs auf /v1/responses API migriert für geringere Latenz',
+    new_0_5_16_2: 'Echte API-Nutzung (Tokens, Cache, Reasoning) in Nutzungsstatistik-Tabelle speichern',
+    new_0_5_16_3: 'QQ-Gruppen-QR-Code zur Website-Navigationsleiste hinzugefügt',
+    new_0_5_16_4: 'Unbenutztes codex_reasoning_items-Feld aus dem Nachrichtenschema entfernt',
    new_0_5_13_1: 'Nachrichtenwarteschlange für sequenzielle Run-Verarbeitung hinzugefügt, um gleichzeitige Konflikte zu vermeiden',
    new_0_5_13_2: 'Zwei-Ebenen-Skills-Verzeichnisstruktur mit Sonstige-Kategorie für flache Skills unterstützt',
    new_0_5_13_3: 'Temporäre Sitzungen (eph_*) beim Start-Sync filtern, um interne Sitzungen nicht zu importieren',
@@ -912,6 +912,10 @@ export default {
    new_0_5_15_9: 'Fix Hermes markdown media rendering and sync retry',
    new_0_5_15_10: 'Refactor to remove upstream env dependency',
    new_0_5_15_11: 'If the Kanban feature is not available, please upgrade hermes-agent',
+    new_0_5_16_1: 'Migrate chat streaming from /v1/runs to /v1/responses API for lower latency',
+    new_0_5_16_2: 'Persist real API usage (tokens, cache, reasoning) to usage table',
+    new_0_5_16_3: 'Add QQ group QR code to website navigation bar',
+    new_0_5_16_4: 'Remove unused codex_reasoning_items field from message schema',
    new_0_5_13_1: 'Add message queue for sequential run processing to prevent concurrent request conflicts',
    new_0_5_13_2: 'Support two-level skills directory structure with misc category for flat skills',
    new_0_5_13_3: 'Filter out ephemeral sessions during startup sync to avoid importing internal sessions',
@@ -661,6 +661,10 @@ jobTriggered: 'Job ejecutado',
    new_0_5_15_9: 'Corregido renderizado de medios Markdown y reintento de sincronización',
    new_0_5_15_10: 'Eliminada dependencia de variables de entorno upstream',
    new_0_5_15_11: 'Si la función Kanban no está disponible, actualice hermes-agent',
+    new_0_5_16_1: 'Migrar streaming de chat de /v1/runs a /v1/responses API para menor latencia',
+    new_0_5_16_2: 'Persistir uso real de API (tokens, caché, razonamiento) en tabla de estadísticas',
+    new_0_5_16_3: 'Añadir código QR del grupo QQ a la barra de navegación del sitio web',
+    new_0_5_16_4: 'Eliminar campo codex_reasoning_items no utilizado del esquema de mensajes',
    new_0_5_13_1: 'Cola de mensajes para procesamiento secuencial de ejecuciones, evitando conflictos concurrentes',
    new_0_5_13_2: 'Soporte para estructura de directorios de skills de dos niveles con categoría misc',
    new_0_5_13_3: 'Filtrar sesiones efímeras (eph_*) durante la sincronización de inicio',
@@ -660,6 +660,10 @@ jobTriggered: 'Job declenche',
    new_0_5_15_9: 'Correction rendu média Markdown et retry synchronisation',
    new_0_5_15_10: 'Suppression dépendance variable d\'environnement amont',
    new_0_5_15_11: 'Si la fonction Kanban n\'est pas disponible, veuillez mettre à niveau hermes-agent',
+    new_0_5_16_1: 'Migration du streaming de chat de /v1/runs vers l\'API /v1/responses pour une latence réduite',
+    new_0_5_16_2: 'Persistance de l\'utilisation réelle de l\'API (tokens, cache, raisonnement) dans la table des statistiques',
+    new_0_5_16_3: 'Ajout du code QR du groupe QQ dans la barre de navigation du site',
+    new_0_5_16_4: 'Suppression du champ codex_reasoning_items inutilisé du schéma de messages',
    new_0_5_13_1: 'File d\'attente de messages pour le traitement séquentiel des exécutions, évitant les conflits concurrents',
    new_0_5_13_2: 'Prise en charge de la structure de répertoire de skills à deux niveaux avec catégorie divers',
    new_0_5_13_3: 'Filtrer les sessions éphémères (eph_*) lors de la synchronisation au démarrage',
@@ -661,6 +661,10 @@ export default {
    new_0_5_15_9: 'Hermes Markdownメディアレンダリングと同期リトライを修正',
    new_0_5_15_10: 'アップストリーム環境変数依存をリファクタリングで削除',
    new_0_5_15_11: 'カンバン機能が使用できない場合は、hermes-agent をアップグレードしてください',
+    new_0_5_16_1: 'チャットストリーミングを /v1/runs から /v1/responses API に移行し、レイテンシを削減',
+    new_0_5_16_2: '実際の API 使用量（トークン、キャッシュ、推論）を統計テーブルに保存',
+    new_0_5_16_3: 'ウェブサイトのナビゲーションバーにQQグループのQRコードを追加',
+    new_0_5_16_4: 'メッセージスキーマから未使用の codex_reasoning_items フィールドを削除',
    new_0_5_13_1: 'メッセージキューによる順次実行処理で同時リクエストの競合を防止',
    new_0_5_13_2: '2階層スキルディレクトリ構造をサポート、フラットスキルは「その他」カテゴリに分類',
    new_0_5_13_3: '起動同期時に一時セッション（eph_*）をフィルタリング',
@@ -661,6 +661,10 @@ export default {
    new_0_5_15_9: 'Hermes Markdown 미디어 렌더링 및 동기화 재시도 수정',
    new_0_5_15_10: '업스트림 환경 변수 종속성 제거',
    new_0_5_15_11: '칸반 기능을 사용할 수 없는 경우 hermes-agent를 업그레이드하세요',
+    new_0_5_16_1: '채팅 스트리밍을 /v1/runs에서 /v1/responses API로 마이그레이션하여 지연 시간 단축',
+    new_0_5_16_2: '실제 API 사용량(토큰, 캐시, 추론)을 사용량 통계 테이블에 저장',
+    new_0_5_16_3: '웹사이트 내비게이션 바에 QQ 그룹 QR 코드 추가',
+    new_0_5_16_4: '메시지 스키마에서 사용하지 않는 codex_reasoning_items 필드 제거',
    new_0_5_13_1: '메시지 큐를 통한 순차 실행 처리로 동시 요청 충돌 방지',
    new_0_5_13_2: '2단계 스킬 디렉토리 구조 지원, 플랫 스킬은 기타 카테고리로 분류',
    new_0_5_13_3: '시작 동기화 시 임시 세션(eph_*) 필터링',
@@ -661,6 +661,10 @@ jobTriggered: 'Job acionado',
    new_0_5_15_9: 'Corrigido renderização de mídia Markdown e retry de sincronização',
    new_0_5_15_10: 'Removida dependência de variável de ambiente upstream',
    new_0_5_15_11: 'Se o recurso Kanban não estiver disponível, atualize o hermes-agent',
+    new_0_5_16_1: 'Migrar streaming de chat de /v1/runs para /v1/responses API para menor latência',
+    new_0_5_16_2: 'Persistir uso real da API (tokens, cache, raciocínio) na tabela de estatísticas',
+    new_0_5_16_3: 'Adicionar código QR do grupo QQ à barra de navegação do site',
+    new_0_5_16_4: 'Remover campo codex_reasoning_items não utilizado do esquema de mensagens',
    new_0_5_13_1: 'Fila de mensagens para processamento sequencial de execuções, evitando conflitos concorrentes',
    new_0_5_13_2: 'Suporte à estrutura de diretório de skills de dois níveis com categoria diversos',
    new_0_5_13_3: 'Filtrar sessões efêmeras (eph_*) durante a sincronização na inicialização',
@@ -914,6 +914,10 @@ export default {
    new_0_5_15_9: '修复 Hermes Markdown 媒体渲染与同步重试',
    new_0_5_15_10: '重构移除上游环境变量依赖',
    new_0_5_15_11: '如果看板功能无法使用，请升级 hermes-agent',
+    new_0_5_16_1: '聊天流式接口从 /v1/runs 迁移至 /v1/responses，降低延迟',
+    new_0_5_16_2: '持久化真实 API 用量（token、缓存、推理）到用量统计表',
+    new_0_5_16_3: '官网导航栏新增 QQ 群二维码',
+    new_0_5_16_4: '移除消息 schema 中未使用的 codex_reasoning_items 字段',
    new_0_5_13_1: '新增消息队列，顺序处理运行请求，避免并发冲突',
    new_0_5_13_2: '支持二级 Skills 目录结构，扁平化 Skill 归入"杂项"分类',
    new_0_5_13_3: '启动同步时过滤临时会话（eph_*），避免导入内部会话',
@@ -27,6 +27,7 @@ export interface Message {
  content: string
  timestamp: number
  toolName?: string
+  toolCallId?: string
  toolPreview?: string
  toolArgs?: string
  toolResult?: string
@@ -156,6 +157,7 @@ function mapHermesMessages(msgs: HermesMessage[]): Message[] {
          content: '',
          timestamp: Math.round(msg.timestamp * 1000),
          toolName: tc.function?.name || 'tool',
+          toolCallId: tc.id,
          toolArgs: tc.function?.arguments || undefined,
          toolStatus: 'done',
        })
@@ -191,6 +193,7 @@ function mapHermesMessages(msgs: HermesMessage[]): Message[] {
        content: '',
        timestamp: Math.round(msg.timestamp * 1000),
        toolName,
+        toolCallId: tcId || undefined,
        toolArgs,
        toolPreview: typeof preview === 'string' ? preview.slice(0, 100) || undefined : undefined,
        toolResult: msg.content || undefined,
@@ -910,6 +913,7 @@ export const useChatStore = defineStore('chat', () => {
            case 'tool.started': {
              runHadToolActivity = true
              const msgs = getSessionMsgs(sid)
+              const toolCallId = (evt as any).tool_call_id as string | undefined
              const last = activeAssistantMessageId
                ? msgs.find(m => m.id === activeAssistantMessageId)
                : msgs[msgs.length - 1]
@@ -917,13 +921,27 @@ export const useChatStore = defineStore('chat', () => {
                updateMessage(sid, last.id, { isStreaming: false })
              }
              activeAssistantMessageId = null
+              const existingTool = toolCallId
+                ? msgs.find(m => m.role === 'tool' && m.toolCallId === toolCallId)
+                : null
+              if (existingTool) {
+                updateMessage(sid, existingTool.id, {
+                  toolName: evt.tool || evt.name,
+                  toolArgs: typeof (evt as any).arguments === 'string' ? (evt as any).arguments : existingTool.toolArgs,
+                  toolPreview: evt.preview || existingTool.toolPreview,
+                  toolStatus: existingTool.toolStatus || 'running',
+                })
+                break
+              }
              addMessage(sid, {
                id: uid(),
                role: 'tool',
                content: '',
                timestamp: Date.now(),
                toolName: evt.tool || evt.name,
+                toolCallId,
                toolPreview: evt.preview,
+                toolArgs: typeof (evt as any).arguments === 'string' ? (evt as any).arguments : undefined,
                toolStatus: 'running',
              })

@@ -933,9 +951,10 @@ export const useChatStore = defineStore('chat', () => {
            case 'tool.completed': {
              runHadToolActivity = true
              const msgs = getSessionMsgs(sid)
-              const toolMsgs = msgs.filter(
-                m => m.role === 'tool' && m.toolStatus === 'running',
-              )
+              const toolCallId = (evt as any).tool_call_id as string | undefined
+              const toolMsgs = toolCallId
+                ? msgs.filter(m => m.role === 'tool' && m.toolCallId === toolCallId)
+                : msgs.filter(m => m.role === 'tool' && m.toolStatus === 'running')
              if (toolMsgs.length > 0) {
                const last = toolMsgs[toolMsgs.length - 1]
                // Check if tool errored
@@ -944,6 +963,7 @@ export const useChatStore = defineStore('chat', () => {
                updateMessage(sid, last.id, {
                  toolStatus: hasError ? 'error' : 'done',
                  toolDuration: duration,
+                  toolResult: typeof (evt as any).output === 'string' ? (evt as any).output : undefined,
                })
              }

@@ -1326,6 +1346,7 @@ export const useChatStore = defineStore('chat', () => {
        case 'tool.started': {
          runHadToolActivity = true
          const msgs = getSessionMsgs(sid)
+          const toolCallId = (evt as any).tool_call_id as string | undefined
          const last = activeAssistantMessageId
            ? msgs.find(m => m.id === activeAssistantMessageId)
            : msgs[msgs.length - 1]
@@ -1333,13 +1354,27 @@ export const useChatStore = defineStore('chat', () => {
            updateMessage(sid, last.id, { isStreaming: false })
          }
          activeAssistantMessageId = null
+          const existingTool = toolCallId
+            ? msgs.find(m => m.role === 'tool' && m.toolCallId === toolCallId)
+            : null
+          if (existingTool) {
+            updateMessage(sid, existingTool.id, {
+              toolName: evt.tool || evt.name,
+              toolArgs: typeof (evt as any).arguments === 'string' ? (evt as any).arguments : existingTool.toolArgs,
+              toolPreview: evt.preview || existingTool.toolPreview,
+              toolStatus: existingTool.toolStatus || 'running',
+            })
+            break
+          }
          addMessage(sid, {
            id: uid(),
            role: 'tool',
            content: '',
            timestamp: Date.now(),
            toolName: evt.tool || evt.name,
+            toolCallId,
            toolPreview: evt.preview,
+            toolArgs: typeof (evt as any).arguments === 'string' ? (evt as any).arguments : undefined,
            toolStatus: 'running',
          })

@@ -1349,12 +1384,16 @@ export const useChatStore = defineStore('chat', () => {
        case 'tool.completed': {
          runHadToolActivity = true
          const msgs = getSessionMsgs(sid)
-          const toolMsgs = msgs.filter(m => m.role === 'tool' && m.toolStatus === 'running')
+          const toolCallId = (evt as any).tool_call_id as string | undefined
+          const toolMsgs = toolCallId
+            ? msgs.filter(m => m.role === 'tool' && m.toolCallId === toolCallId)
+            : msgs.filter(m => m.role === 'tool' && m.toolStatus === 'running')
          if (toolMsgs.length > 0) {
            const hasError = (evt as any).error === true
            updateMessage(sid, toolMsgs[toolMsgs.length - 1].id, {
              toolStatus: hasError ? 'error' : 'done',
              toolDuration: (evt as any).duration,
+              toolResult: typeof (evt as any).output === 'string' ? (evt as any).output : undefined,
            })
          }

@@ -70,7 +70,6 @@ export const MESSAGES_SCHEMA: Record<string, string> = {
  reasoning: 'TEXT',
  reasoning_details: 'TEXT',
  reasoning_content: 'TEXT',
-  codex_reasoning_items: 'TEXT',
 }

 export const MESSAGES_INDEX = 'CREATE INDEX IF NOT EXISTS idx_messages_session_id ON messages(session_id)'
@@ -45,7 +45,6 @@ export interface HermesMessageRow {
  finish_reason: string | null
  reasoning: string | null
  reasoning_details?: string | null
-  codex_reasoning_items?: string | null
  reasoning_content?: string | null
 }

@@ -121,7 +120,6 @@ function mapMessageRow(row: Record<string, unknown>): HermesMessageRow {
    finish_reason: row.finish_reason != null ? String(row.finish_reason) : null,
    reasoning: row.reasoning != null ? String(row.reasoning) : null,
    reasoning_details: row.reasoning_details != null ? String(row.reasoning_details) : null,
-    codex_reasoning_items: row.codex_reasoning_items != null ? String(row.codex_reasoning_items) : null,
    reasoning_content: row.reasoning_content != null ? String(row.reasoning_content) : null,
  }
 }
@@ -343,21 +341,20 @@ export function addMessage(msg: {
  reasoning?: string | null
  reasoning_details?: string | null
  reasoning_content?: string | null
-  codex_reasoning_items?: string | null
 }): number | undefined {
  if (!isSqliteAvailable()) return undefined
  const db = getDb()!
  const toolCallsJson = msg.tool_calls ? JSON.stringify(msg.tool_calls) : null
  const result = db.prepare(
-    `INSERT INTO ${MESSAGES_TABLE} (session_id, role, content, tool_call_id, tool_calls, tool_name, timestamp, token_count, finish_reason, reasoning, reasoning_details, reasoning_content, codex_reasoning_items)
-     VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`,
+    `INSERT INTO ${MESSAGES_TABLE} (session_id, role, content, tool_call_id, tool_calls, tool_name, timestamp, token_count, finish_reason, reasoning, reasoning_details, reasoning_content)
+     VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`,
  ).run(
    msg.session_id, msg.role, msg.content,
    msg.tool_call_id ?? null, toolCallsJson, msg.tool_name ?? null,
    msg.timestamp ?? Math.floor(Date.now() / 1000),
    msg.token_count ?? null, msg.finish_reason ?? null,
    msg.reasoning ?? null, msg.reasoning_details ?? null,
-    msg.reasoning_content ?? null, msg.codex_reasoning_items ?? null,
+    msg.reasoning_content ?? null,
  )
  return result.lastInsertRowid as number
 }
@@ -375,13 +372,12 @@ export function addMessages(msgs: Array<{
  reasoning?: string | null
  reasoning_details?: string | null
  reasoning_content?: string | null
-  codex_reasoning_items?: string | null
 }>): void {
  if (!isSqliteAvailable() || msgs.length === 0) return
  const db = getDb()!
  const insert = db.prepare(
-    `INSERT INTO ${MESSAGES_TABLE} (session_id, role, content, tool_call_id, tool_calls, tool_name, timestamp, token_count, finish_reason, reasoning, reasoning_details, reasoning_content, codex_reasoning_items)
-     VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`,
+    `INSERT INTO ${MESSAGES_TABLE} (session_id, role, content, tool_call_id, tool_calls, tool_name, timestamp, token_count, finish_reason, reasoning, reasoning_details, reasoning_content)
+     VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`,
  )
  db.exec('BEGIN')
  try {
@@ -393,7 +389,7 @@ export function addMessages(msgs: Array<{
        msg.timestamp ?? Math.floor(Date.now() / 1000),
        msg.token_count ?? null, msg.finish_reason ?? null,
        msg.reasoning ?? null, msg.reasoning_details ?? null,
-        msg.reasoning_content ?? null, msg.codex_reasoning_items ?? null,
+        msg.reasoning_content ?? null,
      )
    }
    db.exec('COMMIT')
@@ -53,7 +53,6 @@ export interface HermesMessageRow {
  finish_reason: string | null
  reasoning: string | null
  reasoning_details?: string | null
-  codex_reasoning_items?: string | null
  reasoning_content?: string | null
 }

@@ -350,7 +349,6 @@ function mapMessageRow(row: Record<string, unknown>): HermesMessageRow {
    finish_reason: normalizeNullableString(row.finish_reason),
    reasoning,
    reasoning_details: normalizeNullableString(row.reasoning_details),
-    codex_reasoning_items: normalizeNullableString(row.codex_reasoning_items),
    reasoning_content: normalizeNullableString(row.reasoning_content),
  }
 }
@@ -13,7 +13,6 @@
 * 6. Save snapshot: last_message_index = index where compression ends
 */

-import { EventSource } from 'eventsource'
 import { encodingForModel, getEncoding } from 'js-tiktoken'
 import { logger } from '../../services/logger'
 import {
@@ -21,7 +20,6 @@ import {
  saveCompressionSnapshot,
  deleteCompressionSnapshot,
 } from '../../db/hermes/compression-snapshot'
-import { getDb } from '../../db/index'

 // ─── Types ───────────────────────────────────────────────

@@ -376,8 +374,6 @@ export async function callSummarizer(
  previousSummary?: string,
  profile?: string,
 ): Promise<string> {
-  const sessionId = `compress_${Date.now().toString(36)}_${Math.random().toString(36).slice(2, 8)}`
-
  const convHistory: Array<{ role: string; content: string }> = [...history]

  if (previousSummary) {
@@ -390,88 +386,57 @@ export async function callSummarizer(
  const headers: Record<string, string> = { 'Content-Type': 'application/json' }
  if (apiKey) headers['Authorization'] = `Bearer ${apiKey}`

-  const res = await fetch(`${upstream}/v1/runs`, {
+  const res = await fetch(`${upstream.replace(/\/$/, '')}/v1/responses`, {
    method: 'POST',
    headers,
    body: JSON.stringify({
      input: prompt,
      conversation_history: convHistory,
-      session_id: sessionId,
+      stream: true,
+      store: false,
    }),
    signal: AbortSignal.timeout(timeoutMs),
  })

  if (!res.ok) {
-    throw new Error(`Summarization run failed: ${res.status}`)
+    throw new Error(`Summarization response failed: ${res.status}`)
  }

-  const { run_id } = await res.json() as { run_id: string }
+  if (!res.body) {
+    throw new Error('Summarization response stream missing')
+  }

-  return new Promise<string>((resolve, reject) => {
-    const timer = setTimeout(() => {
-      source.close()
-      reject(new Error('Summarization timed out'))
-    }, timeoutMs)
+  let output = ''
+  for await (const frame of readSseFrames(res.body)) {
+    let parsed: any
+    try {
+      parsed = JSON.parse(frame.data)
+    } catch {
+      continue
+    }
+    const eventType = parsed.type || frame.event || parsed.event

-    const eventsUrl = new URL(`${upstream}/v1/runs/${run_id}/events`)
-
-    // Use Authorization header instead of query parameter for better compatibility
-    const eventSourceInit: any = apiKey ? {
-      fetch: (url: string, init: any = {}) => fetch(url, {
-        ...init,
-        headers: {
-          ...(init.headers || {}),
-          Authorization: `Bearer ${apiKey}`,
-        },
-      }),
-    } : {}
-
-    // @ts-ignore - eventsource library types are too strict
-    const source = new EventSource(eventsUrl.toString(), eventSourceInit)
-
-    source.onmessage = (event: MessageEvent) => {
-      try {
-        const parsed = JSON.parse(event.data)
-        if (parsed.event === 'run.completed') {
-          clearTimeout(timer)
-          source.close()
-          deleteCompressSession(sessionId, profile).catch(() => { })
-          const output = parsed.output
-          if (!output || typeof output !== 'string' || output.trim() === '') {
-            reject(new Error('Empty summarization response'))
-            return
-          }
-          resolve(output.trim())
-        } else if (parsed.event === 'run.failed') {
-          clearTimeout(timer)
-          source.close()
-          deleteCompressSession(sessionId, profile).catch(() => { })
-          reject(new Error(parsed.error || 'Summarization run failed'))
-        }
-      } catch { /* ignore parse errors */ }
+    if (eventType === 'response.output_text.delta' && parsed.delta) {
+      output += parsed.delta
+      continue
    }

-    source.onerror = () => {
-      clearTimeout(timer)
-      source.close()
-      deleteCompressSession(sessionId, profile).catch(() => { })
-      reject(new Error('Summarization SSE connection error'))
+    if (eventType === 'response.completed') {
+      const response = parsed.response || parsed
+      const finalText = extractResponseText(response)
+      if (!output && finalText) output = finalText
+      if (!output || output.trim() === '') {
+        throw new Error('Empty summarization response')
+      }
+      return output.trim()
    }
-  })
-}

-/** Enqueue compression session for later deletion instead of deleting immediately */
-async function deleteCompressSession(sessionId: string, profile?: string): Promise<void> {
-  try {
-    const db = getDb()
-    if (!db) return
-    const now = Date.now()
-    db.prepare(
-      `INSERT INTO gc_pending_session_deletes (session_id, profile_name, status, attempt_count, last_error, created_at, updated_at, next_attempt_at)
-       VALUES (?, ?, 'pending', 0, NULL, ?, ?, 0)
-       ON CONFLICT(session_id) DO NOTHING`,
-    ).run(sessionId, profile || 'default', now, now)
-  } catch { /* best-effort */ }
+    if (eventType === 'response.failed') {
+      throw new Error(parsed.error?.message || parsed.error || 'Summarization response failed')
+    }
+  }
+
+  throw new Error('Summarization response stream ended without a terminal event')
 }

 // ─── Main Compressor ────────────────────────────────────
@@ -665,3 +630,63 @@ export class ChatContextCompressor {
    deleteCompressionSnapshot(sessionId)
  }
 }
+
+async function* readSseFrames(stream: ReadableStream<Uint8Array>): AsyncGenerator<{ event?: string; data: string }> {
+  const decoder = new TextDecoder()
+  const reader = stream.getReader()
+  let buffer = ''
+
+  try {
+    while (true) {
+      const { done, value } = await reader.read()
+      if (done) break
+      buffer += decoder.decode(value, { stream: true })
+
+      let boundary = buffer.indexOf('\n\n')
+      while (boundary >= 0) {
+        const raw = buffer.slice(0, boundary)
+        buffer = buffer.slice(boundary + 2)
+        const frame = parseSseFrame(raw)
+        if (frame?.data) yield frame
+        boundary = buffer.indexOf('\n\n')
+      }
+    }
+
+    buffer += decoder.decode()
+    const frame = parseSseFrame(buffer)
+    if (frame?.data) yield frame
+  } finally {
+    reader.releaseLock()
+  }
+}
+
+function parseSseFrame(raw: string): { event?: string; data: string } | null {
+  let event: string | undefined
+  const data: string[] = []
+  for (const line of raw.split(/\r?\n/)) {
+    if (!line || line.startsWith(':')) continue
+    if (line.startsWith('event:')) {
+      event = line.slice(6).trim()
+    } else if (line.startsWith('data:')) {
+      data.push(line.slice(5).trimStart())
+    }
+  }
+  if (data.length === 0) return null
+  return { event, data: data.join('\n') }
+}
+
+function extractResponseText(response: any): string {
+  const output = Array.isArray(response?.output) ? response.output : []
+  const parts: string[] = []
+  for (const item of output) {
+    if (item.type !== 'message') continue
+    const content = Array.isArray(item.content) ? item.content : []
+    for (const part of content) {
+      if (part.type === 'output_text' || part.type === 'text') {
+        parts.push(part.text || '')
+      }
+    }
+  }
+  if (parts.length > 0) return parts.join('')
+  return typeof response?.output_text === 'string' ? response.output_text : ''
+}
@@ -1,4 +1,3 @@
-import { EventSource } from 'eventsource'
 import type { StoredMessage, GatewayCaller } from './types'
 import {
    buildSummarizationSystemPrompt,
@@ -6,12 +5,11 @@ import {
    buildIncrementalUpdatePrompt,
 } from './prompt'
 import { updateUsage } from '../../../db/hermes/usage-store'
-import { getSessionDetailFromDbWithProfile } from '../../../db/hermes/sessions-db'
 import { logger } from '../../logger'

 /**
- * Calls Hermes /v1/runs to produce LLM-generated summaries.
- * Uses non-streaming EventSource to wait for run.completed.
+ * Calls Hermes /v1/responses to produce LLM-generated summaries.
+ * The context engine owns history assembly; Responses storage/chaining is not used.
 */
 export class GatewaySummarizer implements GatewayCaller {
    private timeoutMs: number
@@ -29,13 +27,11 @@ export class GatewaySummarizer implements GatewayCaller {
        profile: string,
        previousSummary?: string,
    ): Promise<{ summary: string; sessionId: string }> {
-        // Build conversation_history from messages
        const history: Array<{ role: string; content: string }> = messages.map(m => ({
            role: 'user',
            content: `[${m.senderName}]: ${m.content}`,
        }))

-        // Inject previous summary for incremental update
        if (previousSummary) {
            history.unshift(
                { role: 'user', content: `[Previous summary]\n${previousSummary}` },
@@ -47,10 +43,7 @@ export class GatewaySummarizer implements GatewayCaller {
            ? buildIncrementalUpdatePrompt()
            : buildFullSummaryPrompt()

-        const sessionId = Date.now().toString(36) + Math.random().toString(36).slice(2, 8)
-
-        // POST /v1/runs
-        const res = await fetch(`${upstream}/v1/runs`, {
+        const res = await fetch(`${upstream.replace(/\/$/, '')}/v1/responses`, {
            method: 'POST',
            headers: {
                'Content-Type': 'application/json',
@@ -60,98 +53,122 @@ export class GatewaySummarizer implements GatewayCaller {
                input: userPrompt,
                instructions: systemPrompt || buildSummarizationSystemPrompt(),
                conversation_history: history,
-                session_id: sessionId,
+                stream: true,
+                store: false,
            }),
            signal: AbortSignal.timeout(this.timeoutMs),
        })

        if (!res.ok) {
-            throw new Error(`Summarization run failed: ${res.status}`)
+            throw new Error(`Summarization response failed: ${res.status}`)
+        }
+        if (!res.body) {
+            throw new Error('Summarization response stream missing')
        }

-        const { run_id } = await res.json() as { run_id: string }
+        let output = ''
+        for await (const frame of readSseFrames(res.body)) {
+            let parsed: any
+            try {
+                parsed = JSON.parse(frame.data)
+            } catch {
+                continue
+            }
+            const eventType = parsed.type || frame.event || parsed.event

-        try {
-            const output = await this.pollForResult(upstream, apiKey, run_id, sessionId, roomId, profile)
-            return { summary: output, sessionId }
-        } finally {
-            // Note: session cleanup is handled by the caller (compressor.ts)
+            if (eventType === 'response.output_text.delta' && parsed.delta) {
+                output += parsed.delta
+                continue
+            }
+
+            if (eventType === 'response.completed') {
+                const response = parsed.response || parsed
+                const finalText = extractResponseText(response)
+                if (!output && finalText) output = finalText
+
+                const usage = response.usage || {}
+                updateUsage(roomId, {
+                    inputTokens: usage.input_tokens ?? usage.inputTokens ?? 0,
+                    outputTokens: usage.output_tokens ?? usage.outputTokens ?? 0,
+                    cacheReadTokens: usage.cache_read_tokens ?? usage.cacheReadTokens ?? 0,
+                    cacheWriteTokens: usage.cache_write_tokens ?? usage.cacheWriteTokens ?? 0,
+                    reasoningTokens: usage.reasoning_tokens ?? usage.reasoningTokens ?? 0,
+                    model: response.model || '',
+                    profile,
+                })
+                logger.debug(`[GatewaySummarizer] Recorded response usage for compression room ${roomId} (profile=${profile}): input=${usage.input_tokens ?? 0}, output=${usage.output_tokens ?? 0}`)
+
+                if (!output || output.trim() === '') {
+                    throw new Error('Empty summarization response')
+                }
+                return { summary: output.trim(), sessionId: '' }
+            }
+
+            if (eventType === 'response.failed') {
+                throw new Error(parsed.error?.message || parsed.error || 'Summarization response failed')
+            }
        }
+
+        throw new Error('Summarization response stream ended without a terminal event')
    }
-
-    private pollForResult(upstream: string, apiKey: string | null, runId: string, sessionId: string, roomId: string, profile: string): Promise<string> {
-        return new Promise<string>((resolve, reject) => {
-            const timer = setTimeout(() => {
-                source.close()
-                reject(new Error('Summarization timed out'))
-            }, this.timeoutMs)
-
-            const eventsUrl = new URL(`${upstream}/v1/runs/${runId}/events`)
-
-            // Use Authorization header instead of query parameter for better compatibility
-            const eventSourceInit: any = apiKey ? {
-                fetch: (url: string, init: any = {}) => fetch(url, {
-                  ...init,
-                  headers: {
-                    ...(init.headers || {}),
-                    Authorization: `Bearer ${apiKey}`,
-                  },
-                }),
-              } : {}
-
-            // @ts-ignore - eventsource library types are too strict
-            const source = new EventSource(eventsUrl.toString(), eventSourceInit)
-
-            source.onmessage = async (event: MessageEvent) => {
-                try {
-                    const parsed = JSON.parse(event.data)
-                    if (parsed.event === 'run.completed') {
-                        clearTimeout(timer)
-
-                        // Record usage data from Hermes state.db BEFORE closing source
-                        // This ensures we fetch usage before sessionCleaner can delete it
-                        try {
-                            const detail = await getSessionDetailFromDbWithProfile(sessionId, profile)
-                            if (detail) {
-                                updateUsage(roomId, {
-                                    inputTokens: detail.input_tokens,
-                                    outputTokens: detail.output_tokens,
-                                    cacheReadTokens: detail.cache_read_tokens,
-                                    cacheWriteTokens: detail.cache_write_tokens,
-                                    reasoningTokens: detail.reasoning_tokens,
-                                    model: detail.model,
-                                    profile,
-                                })
-                                logger.debug(`[GatewaySummarizer] Recorded usage for compression room ${roomId} (session ${sessionId}, profile=${profile}): input=${detail.input_tokens}, output=${detail.output_tokens}`)
-                            } else {
-                                logger.warn(`[GatewaySummarizer] Failed to get session detail for ${sessionId} (profile=${profile})`)
-                            }
-                        } catch (err: any) {
-                            logger.warn(err, '[GatewaySummarizer] Failed to record usage from DB')
-                        }
-
-                        source.close()
-
-                        const output = parsed.output
-                        if (!output || typeof output !== 'string' || output.trim() === '') {
-                            reject(new Error('Empty summarization response'))
-                            return
-                        }
-                        resolve(output.trim())
-                    } else if (parsed.event === 'run.failed') {
-                        clearTimeout(timer)
-                        source.close()
-                        reject(new Error(parsed.error || 'Summarization run failed'))
-                    }
-                } catch { /* ignore parse errors for non-JSON events */ }
-            }
-
-            source.onerror = () => {
-                clearTimeout(timer)
-                source.close()
-                reject(new Error('Summarization SSE connection error'))
-            }
-        })
-    }
-
+}
+
+async function* readSseFrames(stream: ReadableStream<Uint8Array>): AsyncGenerator<{ event?: string; data: string }> {
+    const decoder = new TextDecoder()
+    const reader = stream.getReader()
+    let buffer = ''
+
+    try {
+        while (true) {
+            const { done, value } = await reader.read()
+            if (done) break
+            buffer += decoder.decode(value, { stream: true })
+
+            let boundary = buffer.indexOf('\n\n')
+            while (boundary >= 0) {
+                const raw = buffer.slice(0, boundary)
+                buffer = buffer.slice(boundary + 2)
+                const frame = parseSseFrame(raw)
+                if (frame?.data) yield frame
+                boundary = buffer.indexOf('\n\n')
+            }
+        }
+
+        buffer += decoder.decode()
+        const frame = parseSseFrame(buffer)
+        if (frame?.data) yield frame
+    } finally {
+        reader.releaseLock()
+    }
+}
+
+function parseSseFrame(raw: string): { event?: string; data: string } | null {
+    let event: string | undefined
+    const data: string[] = []
+    for (const line of raw.split(/\r?\n/)) {
+        if (!line || line.startsWith(':')) continue
+        if (line.startsWith('event:')) {
+            event = line.slice(6).trim()
+        } else if (line.startsWith('data:')) {
+            data.push(line.slice(5).trimStart())
+        }
+    }
+    if (data.length === 0) return null
+    return { event, data: data.join('\n') }
+}
+
+function extractResponseText(response: any): string {
+    const output = Array.isArray(response?.output) ? response.output : []
+    const parts: string[] = []
+    for (const item of output) {
+        if (item.type !== 'message') continue
+        const content = Array.isArray(item.content) ? item.content : []
+        for (const part of content) {
+            if (part.type === 'output_text' || part.type === 'text') {
+                parts.push(part.text || '')
+            }
+        }
+    }
+    if (parts.length > 0) return parts.join('')
+    return typeof response?.output_text === 'string' ? response.output_text : ''
 }
@@ -1,12 +1,8 @@
 import { io, Socket } from 'socket.io-client'
-import { EventSource } from 'eventsource'
 import { getToken } from '../../../services/auth'
 import type { GatewayManager } from '../gateway-manager'
-import { deleteSession as hermesDeleteSession } from '../hermes-cli'
-import { getActiveProfileName } from '../hermes-profile'
 import { logger } from '../../../services/logger'
 import { updateUsage } from '../../../db/hermes/usage-store'
-import { getSessionDetailFromDbWithProfile } from '../../../db/hermes/sessions-db'

 // ─── Types ────────────────────────────────────────────────────

@@ -186,29 +182,6 @@ class AgentClient {
        }
    }

-    private async deleteSession(sessionId: string): Promise<void> {
-        try {
-            const sessionProfile = this.storage?.getSessionProfile?.(sessionId)
-            const currentProfile = getActiveProfileName()
-
-            if (sessionProfile && sessionProfile.profile_name !== currentProfile) {
-                // Cross-profile: enqueue deferred delete, don't switch profile
-                this.storage?.enqueuePendingSessionDelete?.(sessionId, sessionProfile.profile_name)
-                logger.info(`[AgentClients] ${this.name}: cross-profile deferred delete session ${sessionId} (session=${sessionProfile.profile_name}, active=${currentProfile})`)
-                return
-            }
-
-            // Same profile or no mapping: delete directly
-            const ok = await hermesDeleteSession(sessionId)
-            if (ok) {
-                this.storage?.deleteSessionProfile?.(sessionId)
-            }
-            logger.debug(`[AgentClients] ${this.name}: delete session ${sessionId} (profile=${this.profile}) → ${ok ? 'ok' : 'failed'}`)
-        } catch (err: any) {
-            logger.warn(`[AgentClients] ${this.name}: failed to delete session ${sessionId}: ${err.message}`)
-        }
-    }
-
    // ─── Hermes Gateway Integration ────────────────────────────

    /**
@@ -235,8 +208,6 @@ class AgentClient {
            return
        }

-        const sessionId = Date.now().toString(36) + Math.random().toString(36).slice(2, 8)
-
        try {
            // Notify room that agent is typing
            this.startTyping(roomId)
@@ -290,8 +261,7 @@ class AgentClient {

            // Strip @mention from input — agent already knows it was mentioned
            const input = msg.content.replace(new RegExp(`@${this.name.replace(/[.*+?^${}()|[\]\\]/g, '\\$&')}\\s*`, 'gi'), '').trim() || msg.content
-            // Start a run on Hermes gateway
-            const runRes = await fetch(`${upstream}/v1/runs`, {
+            const responseRes = await fetch(`${upstream.replace(/\/$/, '')}/v1/responses`, {
                method: 'POST',
                headers: {
                    'Content-Type': 'application/json',
@@ -299,126 +269,81 @@ class AgentClient {
                },
                body: JSON.stringify({
                    input,
-                    session_id: sessionId,
                    ...(conversationHistory.length > 0 ? { conversation_history: conversationHistory } : {}),
                    ...(instructions ? { instructions } : {}),
+                    stream: true,
+                    store: false,
                }),
                signal: AbortSignal.timeout(120000),
            })

-            if (!runRes.ok) {
-                const text = await runRes.text().catch(() => '')
-                logger.error(`[AgentClients] ${this.name}: gateway run failed (${runRes.status}): ${text}`)
+            if (!responseRes.ok) {
+                const text = await responseRes.text().catch(() => '')
+                logger.error(`[AgentClients] ${this.name}: gateway response failed (${responseRes.status}): ${text}`)
                this.stopTyping(roomId)
+                onStatus?.('ready')
                return
            }

-            const runData = await runRes.json() as any
-            const run_id = runData.run_id
-            logger.debug(`[AgentClients] ${this.name}: run started, response=%j`, runData)
-            if (!run_id) {
-                logger.error(`[AgentClients] ${this.name}: no run_id in response`)
+            if (!responseRes.body) {
+                logger.error(`[AgentClients] ${this.name}: gateway response stream missing`)
                this.stopTyping(roomId)
+                onStatus?.('ready')
                return
            }

-            // Save session-to-profile mapping after gateway confirms the run
-            const actualSessionId = runData.session_id || sessionId
-            if (!this.storage) {
-                logger.warn(`[AgentClients] ${this.name}: storage is null, cannot save session profile for ${actualSessionId}`)
-            } else {
-                this.storage.saveSessionProfile(actualSessionId, roomId, this.agentId, this.profile)
-                logger.debug(`[AgentClients] ${this.name}: saved session profile ${actualSessionId} → profile=${this.profile}`)
-            }
-
-            // Stream events from Hermes
-            const eventsUrl = new URL(`${upstream}/v1/runs/${run_id}/events`)
-            logger.debug(`[AgentClients] ${this.name}: streaming events from ${eventsUrl}`)
-
-            // Use Authorization header instead of query parameter for better compatibility
-            const eventSourceInit: any = apiKey ? {
-                fetch: (url: string, init: any = {}) => fetch(url, {
-                    ...init,
-                    headers: {
-                        ...(init.headers || {}),
-                        Authorization: `Bearer ${apiKey}`,
-                    },
-                }),
-            } : {}
-
-            // @ts-ignore - eventsource library types are too strict
-            const source = new EventSource(eventsUrl.toString(), eventSourceInit)
-
            let fullContent = ''
-
-            source.onmessage = async (e: any) => {
+            for await (const frame of readSseFrames(responseRes.body)) {
+                let parsed: any
                try {
-                    const parsed = JSON.parse(e.data)
-                    logger.debug(`[AgentClients] ${this.name}: event=${parsed.event}`)
-
-                    if (parsed.event === 'run.completed') {
-                        // Record usage data from Hermes state.db BEFORE closing source
-                        // This ensures we fetch usage before deleteSession can delete it
-                        try {
-                            const detail = await getSessionDetailFromDbWithProfile(actualSessionId, this.profile)
-                            if (detail) {
-                                updateUsage(roomId, {
-                                    inputTokens: detail.input_tokens,
-                                    outputTokens: detail.output_tokens,
-                                    cacheReadTokens: detail.cache_read_tokens,
-                                    cacheWriteTokens: detail.cache_write_tokens,
-                                    reasoningTokens: detail.reasoning_tokens,
-                                    model: detail.model,
-                                    profile: this.profile,
-                                })
-                                logger.debug(`[AgentClients] Recorded usage for room ${roomId} (session ${actualSessionId}, profile=${this.profile}): input=${detail.input_tokens}, output=${detail.output_tokens}`)
-                            } else {
-                                logger.warn(`[AgentClients] Failed to get session detail for ${actualSessionId} (profile=${this.profile})`)
-                            }
-                        } catch (err: any) {
-                            logger.warn(err, '[AgentClients] Failed to record usage from DB')
-                        }
-
-                        source.close()
-                        logger.debug(`[AgentClients] ${this.name}: run completed, content length=${fullContent.length}`)
-                        if (fullContent) {
-                            this.stopTyping(roomId)
-                            this.sendMessage(roomId, fullContent)
-                        }
-                        this.deleteSession(actualSessionId).catch(() => { })
-                        onStatus?.('ready')
-                        return
-                    }
-
-                    if (parsed.event === 'run.failed') {
-                        source.close()
-                        logger.error(`[AgentClients] ${this.name}: run failed`)
-                        this.stopTyping(roomId)
-                        this.deleteSession(actualSessionId).catch(() => { })
-                        onStatus?.('ready')
-                        return
-                    }
-
-                    // Accumulate message deltas
-                    if (parsed.event === 'message.delta' && parsed.delta) {
-                        fullContent += parsed.delta
-                    }
+                    parsed = JSON.parse(frame.data)
                } catch {
-                    // ignore parse errors
+                    continue
+                }
+                const eventType = parsed.type || frame.event || parsed.event
+                logger.debug(`[AgentClients] ${this.name}: event=${eventType}`)
+
+                if (eventType === 'response.output_text.delta' && parsed.delta) {
+                    fullContent += parsed.delta
+                    continue
+                }
+
+                if (eventType === 'response.completed') {
+                    const response = parsed.response || parsed
+                    const finalText = extractResponseText(response)
+                    if (!fullContent && finalText) fullContent = finalText
+                    const usage = response.usage || {}
+                    updateUsage(roomId, {
+                        inputTokens: usage.input_tokens ?? usage.inputTokens ?? 0,
+                        outputTokens: usage.output_tokens ?? usage.outputTokens ?? 0,
+                        cacheReadTokens: usage.cache_read_tokens ?? usage.cacheReadTokens ?? 0,
+                        cacheWriteTokens: usage.cache_write_tokens ?? usage.cacheWriteTokens ?? 0,
+                        reasoningTokens: usage.reasoning_tokens ?? usage.reasoningTokens ?? 0,
+                        model: response.model || '',
+                        profile: this.profile,
+                    })
+                    logger.debug(`[AgentClients] ${this.name}: response completed, content length=${fullContent.length}`)
+                    if (fullContent) {
+                        this.stopTyping(roomId)
+                        this.sendMessage(roomId, fullContent)
+                    }
+                    onStatus?.('ready')
+                    return
+                }
+
+                if (eventType === 'response.failed') {
+                    logger.error(`[AgentClients] ${this.name}: response failed`)
+                    this.stopTyping(roomId)
+                    onStatus?.('ready')
+                    return
                }
            }
-
-            source.onerror = (err: any) => {
-                logger.error(err, `[AgentClients] ${this.name}: EventSource error`)
-                source.close()
-                this.stopTyping(roomId)
-                this.deleteSession(actualSessionId).catch(() => { })
-                onStatus?.('ready')
-            }
+            logger.warn(`[AgentClients] ${this.name}: response stream ended without terminal event`)
+            this.stopTyping(roomId)
+            onStatus?.('ready')
        } catch (err: any) {
            logger.error(`[AgentClients] ${this.name}: error handling message: ${err.message}`)
            this.stopTyping(roomId)
-            this.deleteSession(sessionId).catch(() => { })
            onStatus?.('ready')
        }
    }
@@ -460,6 +385,66 @@ class AgentClient {
    }
 }

+async function* readSseFrames(stream: ReadableStream<Uint8Array>): AsyncGenerator<{ event?: string; data: string }> {
+    const decoder = new TextDecoder()
+    const reader = stream.getReader()
+    let buffer = ''
+
+    try {
+        while (true) {
+            const { done, value } = await reader.read()
+            if (done) break
+            buffer += decoder.decode(value, { stream: true })
+
+            let boundary = buffer.indexOf('\n\n')
+            while (boundary >= 0) {
+                const raw = buffer.slice(0, boundary)
+                buffer = buffer.slice(boundary + 2)
+                const frame = parseSseFrame(raw)
+                if (frame?.data) yield frame
+                boundary = buffer.indexOf('\n\n')
+            }
+        }
+
+        buffer += decoder.decode()
+        const frame = parseSseFrame(buffer)
+        if (frame?.data) yield frame
+    } finally {
+        reader.releaseLock()
+    }
+}
+
+function parseSseFrame(raw: string): { event?: string; data: string } | null {
+    let event: string | undefined
+    const data: string[] = []
+    for (const line of raw.split(/\r?\n/)) {
+        if (!line || line.startsWith(':')) continue
+        if (line.startsWith('event:')) {
+            event = line.slice(6).trim()
+        } else if (line.startsWith('data:')) {
+            data.push(line.slice(5).trimStart())
+        }
+    }
+    if (data.length === 0) return null
+    return { event, data: data.join('\n') }
+}
+
+function extractResponseText(response: any): string {
+    const output = Array.isArray(response?.output) ? response.output : []
+    const parts: string[] = []
+    for (const item of output) {
+        if (item.type !== 'message') continue
+        const content = Array.isArray(item.content) ? item.content : []
+        for (const part of content) {
+            if (part.type === 'output_text' || part.type === 'text') {
+                parts.push(part.text || '')
+            }
+        }
+    }
+    if (parts.length > 0) return parts.join('')
+    return typeof response?.output_text === 'string' ? response.output_text : ''
+}
+
 // ─── AgentClients (roomId -> agents) ──────────────────────────

 export class AgentClients {
@@ -107,7 +107,6 @@ async function syncProfileSessions(profile: string): Promise<{
            reasoning: msg.reasoning,
            reasoning_details: msg.reasoning_details,
            reasoning_content: msg.reasoning_content,
-            codex_reasoning_items: msg.codex_reasoning_items,
          })
        }