[codex] Harden context compression history handling (#848)

* Use token threshold for chat compression

* Add compression settings controls

* Use config for chat compression

* Cover protected messages in compression tests

* Remove message-count compression limit

* Harden compression window fallback

* Rebuild stale compression snapshots

* Harden stale compression snapshots

* Update changelog for compression hardening

* Prefer local history session details

packages/client/src/api/hermes/config.ts

@@ -28,6 +28,14 @@ export interface MemoryConfig {
   user_char_limit?: number
 }
 
+export interface CompressionConfig {
+  enabled?: boolean
+  threshold?: number
+  target_ratio?: number
+  protect_last_n?: number
+  protect_first_n?: number
+}
+
 export interface SessionResetConfig {
   mode?: string
   idle_minutes?: number
@@ -47,6 +55,7 @@ export interface AppConfig {
   display?: DisplayConfig
   agent?: AgentConfig
   memory?: MemoryConfig
+  compression?: CompressionConfig
   session_reset?: SessionResetConfig
   privacy?: PrivacyConfig
   approvals?: ApprovalConfig

packages/client/src/components/hermes/settings/CompressionSettings.vue

@@ -0,0 +1,106 @@
+<script setup lang="ts">
+import { NInputNumber, NSwitch, useMessage } from 'naive-ui'
+import { useI18n } from 'vue-i18n'
+import { useSettingsStore } from '@/stores/hermes/settings'
+import SettingRow from './SettingRow.vue'
+
+const settingsStore = useSettingsStore()
+const message = useMessage()
+const { t } = useI18n()
+
+const defaults = {
+  enabled: true,
+  threshold: 0.5,
+  target_ratio: 0.2,
+  protect_last_n: 20,
+  protect_first_n: 3,
+}
+
+const debounceTimers: Record<string, ReturnType<typeof setTimeout>> = {}
+
+function save(values: Record<string, any>) {
+  settingsStore.updateLocal('compression', values)
+  settingsStore.saveSection('compression', values).then(() => {
+    message.success(t('settings.saved'))
+  }).catch(() => {
+    message.error(t('settings.saveFailed'))
+  })
+}
+
+function debouncedSave(key: string, value: any) {
+  settingsStore.updateLocal('compression', { [key]: value })
+  if (debounceTimers[key]) clearTimeout(debounceTimers[key])
+  debounceTimers[key] = setTimeout(async () => {
+    try {
+      await settingsStore.saveSection('compression', { [key]: value })
+      message.success(t('settings.saved'))
+    } catch {
+      message.error(t('settings.saveFailed'))
+    }
+  }, 300)
+}
+</script>
+
+<template>
+  <section class="settings-section">
+    <SettingRow :label="t('settings.compression.enabled')" :hint="t('settings.compression.enabledHint')">
+      <NSwitch
+        :value="settingsStore.compression.enabled ?? defaults.enabled"
+        size="small"
+        @update:value="v => save({ enabled: v })"
+      />
+    </SettingRow>
+    <SettingRow :label="t('settings.compression.threshold')" :hint="t('settings.compression.thresholdHint')">
+      <NInputNumber
+        :value="settingsStore.compression.threshold ?? defaults.threshold"
+        :min="0.1"
+        :max="0.95"
+        :step="0.05"
+        size="small"
+        class="input-sm"
+        @update:value="v => v != null && debouncedSave('threshold', v)"
+      />
+    </SettingRow>
+    <SettingRow :label="t('settings.compression.targetRatio')" :hint="t('settings.compression.targetRatioHint')">
+      <NInputNumber
+        :value="settingsStore.compression.target_ratio ?? defaults.target_ratio"
+        :min="0.05"
+        :max="0.8"
+        :step="0.05"
+        size="small"
+        class="input-sm"
+        @update:value="v => v != null && debouncedSave('target_ratio', v)"
+      />
+    </SettingRow>
+    <SettingRow :label="t('settings.compression.protectLastN')" :hint="t('settings.compression.protectLastNHint')">
+      <NInputNumber
+        :value="settingsStore.compression.protect_last_n ?? defaults.protect_last_n"
+        :min="0"
+        :max="200"
+        :step="1"
+        size="small"
+        class="input-sm"
+        @update:value="v => v != null && debouncedSave('protect_last_n', v)"
+      />
+    </SettingRow>
+    <SettingRow :label="t('settings.compression.protectFirstN')" :hint="t('settings.compression.protectFirstNHint')">
+      <NInputNumber
+        :value="settingsStore.compression.protect_first_n ?? defaults.protect_first_n"
+        :min="0"
+        :max="50"
+        :step="1"
+        size="small"
+        class="input-sm"
+        @update:value="v => v != null && debouncedSave('protect_first_n', v)"
+      />
+    </SettingRow>
+  </section>
+</template>
+
+<style scoped lang="scss">
+@use '@/styles/variables' as *;
+
+.settings-section {
+  margin-top: 16px;
+}
+</style>

packages/client/src/data/changelog.ts

@@ -17,6 +17,7 @@ export const changelog: ChangelogEntry[] = [
       'changelog.new_0_5_30_6',
       'changelog.new_0_5_30_7',
       'changelog.new_0_5_30_8',
+      'changelog.new_0_5_30_9',
     ],
   },
   {

packages/client/src/i18n/locales/de.ts

@@ -549,6 +549,7 @@ jobTriggered: 'Job ausgelost',
       account: 'Konto',
       agent: 'Agent',
       memory: 'Gedachtnis',
+      compression: 'Komprimierung',
       session: 'Sitzung',
       privacy: 'Datenschutz',
       apiServer: 'API-Server',
@@ -599,6 +600,18 @@ jobTriggered: 'Job ausgelost',
       userCharLimit: 'Zeichenlimit fur Benutzerprofil',
       userCharLimitHint: 'Maximale Zeichen fur USER.md',
     },
+    compression: {
+      enabled: 'Komprimierung aktivieren',
+      enabledHint: 'Langen Chatverlauf automatisch komprimieren, bevor der Modellkontext uberschritten wird',
+      threshold: 'Komprimierungsschwelle',
+      thresholdHint: 'Komprimierung starten, wenn geschatzte Token dieses Kontextverhaltnis uberschreiten',
+      targetRatio: 'Zielverhaltnis',
+      targetRatioHint: 'Zielgroße des Verlaufs nach der Komprimierung als Kontextverhaltnis',
+      protectLastN: 'Neueste Nachrichten schutzen',
+      protectLastNHint: 'So viele neueste Nachrichten unkomprimiert lassen',
+      protectFirstN: 'Erste Nachrichten schutzen',
+      protectFirstNHint: 'So viele erste Nachrichten unkomprimiert lassen',
+    },
     session: {
       mode: 'Zurucksetzungsmodus',
       modeHint: 'Ausloser fur Sitzungszurucksetzung',
@@ -957,6 +970,7 @@ jobTriggered: 'Job ausgelost',
     new_0_5_30_6: 'Harden Windows compatibility for path detection, file downloads, and job/update subprocesses so they no longer flash terminal windows',
     new_0_5_30_7: 'Fix config writes and provider presets: validate .env keys, route FUN-Codex through the Responses API, and refresh Z.AI/GLM model lists',
     new_0_5_30_8: 'Polish frontend details including collapsed sidebar layout, short group labels, sidebar divider, and conversation outline styling',
+    new_0_5_30_9: 'Context compression now follows Profile compression settings and hardens stale snapshots by reusing previous summaries with a safe tail instead of recompressing full history',
     new_0_5_5_1: '🎉 Tag der Arbeit! Heute wird nicht gearbeitet, bitte habt Verständnis',
     new_0_5_5_2: 'Verlaufsseite für Hermes-Sitzungshistorie hinzugefügt',
     new_0_5_5_3: 'Verlaufsseite verwaltet Sitzungen unabhängig ohne Störung des aktiven Chats',

packages/client/src/i18n/locales/en.ts

@@ -716,6 +716,7 @@ export default {
       account: 'Account',
       agent: 'Agent',
       memory: 'Memory',
+      compression: 'Compression',
       session: 'Session',
       privacy: 'Privacy',
       apiServer: 'API Server',
@@ -774,6 +775,18 @@ export default {
       userCharLimit: 'User Profile Char Limit',
       userCharLimitHint: 'Max characters for USER.md',
     },
+    compression: {
+      enabled: 'Enable Compression',
+      enabledHint: 'Automatically compress long chat history before it exceeds the model context',
+      threshold: 'Compression Threshold',
+      thresholdHint: 'Start compression when estimated tokens exceed this context ratio',
+      targetRatio: 'Target Ratio',
+      targetRatioHint: 'Target history size after compression as a context ratio',
+      protectLastN: 'Protect Recent Messages',
+      protectLastNHint: 'Keep this many latest messages uncompressed',
+      protectFirstN: 'Protect First Messages',
+      protectFirstNHint: 'Keep this many earliest messages uncompressed',
+    },
     session: {
       mode: 'Reset Mode',
       modeHint: 'Trigger condition for session reset',
@@ -1251,6 +1264,7 @@ export default {
     new_0_5_30_6: 'Harden Windows compatibility for path detection, file downloads, and job/update subprocesses so they no longer flash terminal windows',
     new_0_5_30_7: 'Fix config writes and provider presets: validate .env keys, route FUN-Codex through the Responses API, and refresh Z.AI/GLM model lists',
     new_0_5_30_8: 'Polish frontend details including collapsed sidebar layout, short group labels, sidebar divider, and conversation outline styling',
+    new_0_5_30_9: 'Context compression now follows Profile compression settings and hardens stale snapshots by reusing previous summaries with a safe tail instead of recompressing full history',
 
     new_0_5_6_1: 'Add voice playback feature with Web Speech API: manual button, auto-play toggle, rainbow border animation, and mobile optimization',
     new_0_5_6_2: 'Add robust LLM JSON parser with tolerance for Python format and extract text from streaming events',

packages/client/src/i18n/locales/es.ts

@@ -549,6 +549,7 @@ jobTriggered: 'Job ejecutado',
       account: 'Cuenta',
       agent: 'Agente',
       memory: 'Memoria',
+      compression: 'Compresion',
       session: 'Sesion',
       privacy: 'Privacidad',
       apiServer: 'Servidor API',
@@ -599,6 +600,18 @@ jobTriggered: 'Job ejecutado',
       userCharLimit: 'Limite de caracteres del perfil de usuario',
       userCharLimitHint: 'Maximo de caracteres para USER.md',
     },
+    compression: {
+      enabled: 'Activar compresion',
+      enabledHint: 'Comprimir automaticamente el historial largo antes de superar el contexto del modelo',
+      threshold: 'Umbral de compresion',
+      thresholdHint: 'Iniciar compresion cuando los tokens estimados superen esta proporcion del contexto',
+      targetRatio: 'Proporcion objetivo',
+      targetRatioHint: 'Tamano objetivo del historial tras comprimir como proporcion del contexto',
+      protectLastN: 'Proteger mensajes recientes',
+      protectLastNHint: 'Mantener sin comprimir esta cantidad de mensajes recientes',
+      protectFirstN: 'Proteger primeros mensajes',
+      protectFirstNHint: 'Mantener sin comprimir esta cantidad de mensajes iniciales',
+    },
     session: {
       mode: 'Modo de reinicio',
       modeHint: 'Condicion de activacion del reinicio de sesion',
@@ -953,6 +966,7 @@ jobTriggered: 'Job ejecutado',
     new_0_5_30_6: 'Harden Windows compatibility for path detection, file downloads, and job/update subprocesses so they no longer flash terminal windows',
     new_0_5_30_7: 'Fix config writes and provider presets: validate .env keys, route FUN-Codex through the Responses API, and refresh Z.AI/GLM model lists',
     new_0_5_30_8: 'Polish frontend details including collapsed sidebar layout, short group labels, sidebar divider, and conversation outline styling',
+    new_0_5_30_9: 'Context compression now follows Profile compression settings and hardens stale snapshots by reusing previous summaries with a safe tail instead of recompressing full history',
     new_0_5_5_1: '🎉 ¡Feliz Día del Trabajo! Hoy no se trabaja, agradezcan su comprensión',
     new_0_5_5_2: 'Añadida página de historial para sesiones Hermes',
     new_0_5_5_3: 'La página de historial gestiona sesiones de forma independiente',

packages/client/src/i18n/locales/fr.ts

@@ -549,6 +549,7 @@ jobTriggered: 'Job declenche',
       account: 'Compte',
       agent: 'Agent',
       memory: 'Memoire',
+      compression: 'Compression',
       session: 'Session',
       privacy: 'Confidentialite',
       apiServer: 'Serveur API',
@@ -599,6 +600,18 @@ jobTriggered: 'Job declenche',
       userCharLimit: 'Limite de caracteres du profil utilisateur',
       userCharLimitHint: 'Nombre maximum de caracteres pour USER.md',
     },
+    compression: {
+      enabled: 'Activer la compression',
+      enabledHint: 'Compresser automatiquement un long historique avant de depasser le contexte du modele',
+      threshold: 'Seuil de compression',
+      thresholdHint: 'Demarrer la compression quand les jetons estimes depassent ce ratio de contexte',
+      targetRatio: 'Ratio cible',
+      targetRatioHint: 'Taille cible de l\'historique apres compression comme ratio du contexte',
+      protectLastN: 'Proteger les messages recents',
+      protectLastNHint: 'Garder autant de messages recents non compresses',
+      protectFirstN: 'Proteger les premiers messages',
+      protectFirstNHint: 'Garder autant de premiers messages non compresses',
+    },
     session: {
       mode: 'Mode de reinitialisation',
       modeHint: 'Condition de declenchement de la reinitialisation de session',
@@ -952,6 +965,7 @@ jobTriggered: 'Job declenche',
     new_0_5_30_6: 'Harden Windows compatibility for path detection, file downloads, and job/update subprocesses so they no longer flash terminal windows',
     new_0_5_30_7: 'Fix config writes and provider presets: validate .env keys, route FUN-Codex through the Responses API, and refresh Z.AI/GLM model lists',
     new_0_5_30_8: 'Polish frontend details including collapsed sidebar layout, short group labels, sidebar divider, and conversation outline styling',
+    new_0_5_30_9: 'Context compression now follows Profile compression settings and hardens stale snapshots by reusing previous summaries with a safe tail instead of recompressing full history',
     new_0_5_5_1: '🎉 Joyeuse Fête du Travail! Pas de travail aujourd\'hui, merci de votre compréhension',
     new_0_5_5_2: 'Ajout d\'une page d\'historique pour les sessions Hermes',
     new_0_5_5_3: 'La page d\'historique gère les sessions de manière indépendante',

packages/client/src/i18n/locales/ja.ts

@@ -549,6 +549,7 @@ export default {
       account: 'アカウント',
       agent: 'エージェント',
       memory: 'メモリ',
+      compression: '圧縮',
       session: 'セッション',
       privacy: 'プライバシー',
       apiServer: 'API サーバー',
@@ -599,6 +600,18 @@ export default {
       userCharLimit: 'ユーザープロファイル文字数上限',
       userCharLimitHint: 'USER.md の最大文字数',
     },
+    compression: {
+      enabled: '圧縮を有効化',
+      enabledHint: '長いチャット履歴がモデルコンテキストを超える前に自動圧縮',
+      threshold: '圧縮しきい値',
+      thresholdHint: '推定トークンがこのコンテキスト比率を超えたら圧縮を開始',
+      targetRatio: '目標比率',
+      targetRatioHint: '圧縮後の履歴サイズをコンテキスト比率で指定',
+      protectLastN: '直近メッセージを保護',
+      protectLastNHint: 'この数の最新メッセージは圧縮しない',
+      protectFirstN: '先頭メッセージを保護',
+      protectFirstNHint: 'この数の最初のメッセージは圧縮しない',
+    },
     session: {
       mode: 'リセットモード',
       modeHint: 'セッションリセットのトリガー条件',
@@ -953,6 +966,7 @@ export default {
     new_0_5_30_6: 'Harden Windows compatibility for path detection, file downloads, and job/update subprocesses so they no longer flash terminal windows',
     new_0_5_30_7: 'Fix config writes and provider presets: validate .env keys, route FUN-Codex through the Responses API, and refresh Z.AI/GLM model lists',
     new_0_5_30_8: 'Polish frontend details including collapsed sidebar layout, short group labels, sidebar divider, and conversation outline styling',
+    new_0_5_30_9: 'Context compression now follows Profile compression settings and hardens stale snapshots by reusing previous summaries with a safe tail instead of recompressing full history',
     new_0_5_5_1: '🎉 労働者の日！今日はお休みです、何卒ご理解ください',
     new_0_5_5_2: 'Hermesセッション履歴ページを追加',
     new_0_5_5_3: '履歴ページはアクティブチャットに干渉せずにセッション管理',

packages/client/src/i18n/locales/ko.ts

@@ -549,6 +549,7 @@ export default {
       account: '계정',
       agent: '에이전트',
       memory: '메모리',
+      compression: '압축',
       session: '세션',
       privacy: '개인정보',
       apiServer: 'API 서버',
@@ -599,6 +600,18 @@ export default {
       userCharLimit: '사용자 프로필 문자 제한',
       userCharLimitHint: 'USER.md 최대 문자 수',
     },
+    compression: {
+      enabled: '압축 활성화',
+      enabledHint: '긴 채팅 기록이 모델 컨텍스트를 넘기 전에 자동 압축',
+      threshold: '압축 임계값',
+      thresholdHint: '추정 토큰이 이 컨텍스트 비율을 넘으면 압축 시작',
+      targetRatio: '목표 비율',
+      targetRatioHint: '압축 후 기록 크기를 컨텍스트 비율로 지정',
+      protectLastN: '최근 메시지 보호',
+      protectLastNHint: '이 수만큼 최신 메시지는 압축하지 않음',
+      protectFirstN: '처음 메시지 보호',
+      protectFirstNHint: '이 수만큼 처음 메시지는 압축하지 않음',
+    },
     session: {
       mode: '초기화 모드',
       modeHint: '세션 초기화 트리거 조건',
@@ -953,6 +966,7 @@ export default {
     new_0_5_30_6: 'Harden Windows compatibility for path detection, file downloads, and job/update subprocesses so they no longer flash terminal windows',
     new_0_5_30_7: 'Fix config writes and provider presets: validate .env keys, route FUN-Codex through the Responses API, and refresh Z.AI/GLM model lists',
     new_0_5_30_8: 'Polish frontend details including collapsed sidebar layout, short group labels, sidebar divider, and conversation outline styling',
+    new_0_5_30_9: 'Context compression now follows Profile compression settings and hardens stale snapshots by reusing previous summaries with a safe tail instead of recompressing full history',
     new_0_5_5_1: '🎉 노동절 감사합니다! 오늘은 쉬니까 양해 부탁드립니다',
     new_0_5_5_2: 'Hermes 세션 기록 페이지 추가',
     new_0_5_5_3: '기록 페이지는 독립적으로 세션 관리',

packages/client/src/i18n/locales/pt.ts

@@ -549,6 +549,7 @@ jobTriggered: 'Job acionado',
       account: 'Conta',
       agent: 'Agente',
       memory: 'Memoria',
+      compression: 'Compressao',
       session: 'Sessao',
       privacy: 'Privacidade',
       apiServer: 'Servidor API',
@@ -599,6 +600,18 @@ jobTriggered: 'Job acionado',
       userCharLimit: 'Limite de caracteres do perfil do usuario',
       userCharLimitHint: 'Maximo de caracteres para USER.md',
     },
+    compression: {
+      enabled: 'Ativar compressao',
+      enabledHint: 'Comprimir automaticamente historico longo antes de exceder o contexto do modelo',
+      threshold: 'Limiar de compressao',
+      thresholdHint: 'Iniciar compressao quando tokens estimados excederem esta proporcao do contexto',
+      targetRatio: 'Proporcao alvo',
+      targetRatioHint: 'Tamanho alvo do historico apos compressao como proporcao do contexto',
+      protectLastN: 'Proteger mensagens recentes',
+      protectLastNHint: 'Manter sem compressao esta quantidade de mensagens recentes',
+      protectFirstN: 'Proteger primeiras mensagens',
+      protectFirstNHint: 'Manter sem compressao esta quantidade de mensagens iniciais',
+    },
     session: {
       mode: 'Modo de reinicializacao',
       modeHint: 'Condicao de acionamento para reinicializacao de sessao',
@@ -953,6 +966,7 @@ jobTriggered: 'Job acionado',
     new_0_5_30_6: 'Harden Windows compatibility for path detection, file downloads, and job/update subprocesses so they no longer flash terminal windows',
     new_0_5_30_7: 'Fix config writes and provider presets: validate .env keys, route FUN-Codex through the Responses API, and refresh Z.AI/GLM model lists',
     new_0_5_30_8: 'Polish frontend details including collapsed sidebar layout, short group labels, sidebar divider, and conversation outline styling',
+    new_0_5_30_9: 'Context compression now follows Profile compression settings and hardens stale snapshots by reusing previous summaries with a safe tail instead of recompressing full history',
     new_0_5_5_1: '🎉 Feliz Dia do Trabalhador! Hoje não se trabalha, obrigado pela compreensão',
     new_0_5_5_2: 'Adicionada página de histórico para sessões Hermes',
     new_0_5_5_3: 'Página de histórico gerencia sessões de forma independente',

packages/client/src/i18n/locales/zh-TW.ts

@@ -696,6 +696,7 @@ export default {
       account: '帳號',
       agent: '代理',
       memory: '記憶',
+      compression: '上下文壓縮',
       session: '工作階段',
       privacy: '隱私',
       apiServer: 'API 伺服器',
@@ -754,6 +755,18 @@ export default {
       userCharLimit: '使用者畫像字元上限',
       userCharLimitHint: 'USER.md 最大字元數',
     },
+    compression: {
+      enabled: '啟用壓縮',
+      enabledHint: '長對話接近模型上下文上限前自動壓縮歷史',
+      threshold: '壓縮閾值',
+      thresholdHint: '預估 token 超過上下文比例時開始壓縮',
+      targetRatio: '目標比例',
+      targetRatioHint: '壓縮後歷史保留到上下文的目標比例',
+      protectLastN: '保護最近訊息',
+      protectLastNHint: '最近多少則訊息不參與壓縮',
+      protectFirstN: '保護開頭訊息',
+      protectFirstNHint: '最早多少則訊息不參與壓縮',
+    },
     session: {
       mode: '重設模式',
       modeHint: '工作階段重設的觸發條件',
@@ -1243,6 +1256,7 @@ export default {
     new_0_5_30_6: '優化 Windows 相容性：路徑識別、檔案下載、任務/更新子程序不再彈出額外終端視窗',
     new_0_5_30_7: '修復配置寫入和 Provider 預設：加強 .env 校驗，FUN-Codex 改走 Responses API，並更新 Z.AI/GLM 模型列表',
     new_0_5_30_8: '前端體驗細節優化：折疊側邊欄布局、分組短標題、側邊欄分隔線和對話大綱樣式更穩定',
+    new_0_5_30_9: '上下文壓縮現在跟隨 Profile 壓縮配置，並在舊快照失效時複用舊摘要和安全尾部，避免重新壓縮完整歷史',
     new_0_5_6_1: '新增語音播放功能：使用 Web Speech API，支援手動播放按鈕、自動播放開關、彩虹邊框動畫和行動端最佳化',
     new_0_5_6_2: '新增強健的 LLM JSON 解析器，相容 Python 格式並從串流事件中擷取文字',
     new_0_5_6_3: 'Skills 功能增強：使用統計、來源過濾、封存技能、來源追溯和釘選切換',

packages/client/src/i18n/locales/zh.ts

@@ -708,6 +708,7 @@ export default {
       account: '账户',
       agent: '代理',
       memory: '记忆',
+      compression: '上下文压缩',
       session: '会话',
       privacy: '隐私',
       apiServer: 'API 服务器',
@@ -766,6 +767,18 @@ export default {
       userCharLimit: '用户画像字符上限',
       userCharLimitHint: 'USER.md 最大字符数',
     },
+    compression: {
+      enabled: '启用压缩',
+      enabledHint: '长对话接近模型上下文上限前自动压缩历史',
+      threshold: '压缩阈值',
+      thresholdHint: '预计 token 超过上下文比例时开始压缩',
+      targetRatio: '目标比例',
+      targetRatioHint: '压缩后历史保留到上下文的目标比例',
+      protectLastN: '保护最近消息',
+      protectLastNHint: '最近多少条消息不参与压缩',
+      protectFirstN: '保护开头消息',
+      protectFirstNHint: '最早多少条消息不参与压缩',
+    },
     session: {
       mode: '重置模式',
       modeHint: '会话重置的触发条件',
@@ -1253,6 +1266,7 @@ export default {
     new_0_5_30_6: '优化 Windows 兼容性：路径识别、文件下载、任务/更新子进程不再弹出额外终端窗口',
     new_0_5_30_7: '修复配置写入和 Provider 预设：加强 .env 校验，FUN-Codex 改走 Responses API，并更新 Z.AI/GLM 模型列表',
     new_0_5_30_8: '前端体验细节优化：折叠侧边栏布局、分组短标题、侧边栏分隔线和对话大纲样式更稳定',
+    new_0_5_30_9: '上下文压缩现在跟随 Profile 压缩配置，并在旧快照失效时复用旧摘要和安全尾部，避免重新压缩完整历史',
 
     new_0_5_6_1: '新增语音播放功能：使用 Web Speech API，支持手动播放按钮、自动播放开关、彩虹边框动画和移动端优化',
     new_0_5_6_2: '新增健壮的 LLM JSON 解析器，兼容 Python 格式并从流式事件中提取文本',

packages/client/src/stores/hermes/settings.ts

@@ -1,7 +1,7 @@
 import { defineStore } from 'pinia'
 import { ref } from 'vue'
 import * as configApi from '@/api/hermes/config'
-import type { DisplayConfig, AgentConfig, MemoryConfig, SessionResetConfig, PrivacyConfig, ApprovalConfig } from '@/api/hermes/config'
+import type { DisplayConfig, AgentConfig, MemoryConfig, CompressionConfig, SessionResetConfig, PrivacyConfig, ApprovalConfig } from '@/api/hermes/config'
 
 export const useSettingsStore = defineStore('settings', () => {
   const loading = ref(false)
@@ -10,6 +10,7 @@ export const useSettingsStore = defineStore('settings', () => {
   const display = ref<DisplayConfig>({})
   const agent = ref<AgentConfig>({})
   const memory = ref<MemoryConfig>({})
+  const compression = ref<CompressionConfig>({})
   const sessionReset = ref<SessionResetConfig>({})
   const privacy = ref<PrivacyConfig>({})
   const approvals = ref<ApprovalConfig>({})
@@ -32,6 +33,7 @@ export const useSettingsStore = defineStore('settings', () => {
       display.value = data.display || {}
       agent.value = data.agent || {}
       memory.value = data.memory || {}
+      compression.value = data.compression || {}
       sessionReset.value = data.session_reset || {}
       privacy.value = data.privacy || {}
       approvals.value = data.approvals || {}
@@ -58,6 +60,7 @@ export const useSettingsStore = defineStore('settings', () => {
       case 'display': display.value = { ...display.value, ...values }; break
       case 'agent': agent.value = { ...agent.value, ...values }; break
       case 'memory': memory.value = { ...memory.value, ...values }; break
+      case 'compression': compression.value = { ...compression.value, ...values }; break
       case 'session_reset': sessionReset.value = { ...sessionReset.value, ...values }; break
       case 'privacy': privacy.value = { ...privacy.value, ...values }; break
       case 'approvals': approvals.value = { ...approvals.value, ...values }; break
@@ -91,6 +94,7 @@ export const useSettingsStore = defineStore('settings', () => {
       case 'display': display.value = { ...display.value, ...values }; break
       case 'agent': agent.value = { ...agent.value, ...values }; break
       case 'memory': memory.value = { ...memory.value, ...values }; break
+      case 'compression': compression.value = { ...compression.value, ...values }; break
       case 'session_reset': sessionReset.value = { ...sessionReset.value, ...values }; break
       case 'privacy': privacy.value = { ...privacy.value, ...values }; break
       case 'approvals': approvals.value = { ...approvals.value, ...values }; break
@@ -122,7 +126,7 @@ export const useSettingsStore = defineStore('settings', () => {
 
   return {
     loading, saving,
-    display, agent, memory, sessionReset, privacy, approvals,
+    display, agent, memory, compression, sessionReset, privacy, approvals,
     telegram, discord, slack, whatsapp, matrix, wecom, feishu, dingtalk, qqbot, weixin, platforms,
     fetchSettings, saveSection, updateLocal,
   }

packages/client/src/views/hermes/SettingsView.vue

@@ -10,6 +10,7 @@ import { useSettingsStore } from "@/stores/hermes/settings";
 import DisplaySettings from "@/components/hermes/settings/DisplaySettings.vue";
 import AgentSettings from "@/components/hermes/settings/AgentSettings.vue";
 import MemorySettings from "@/components/hermes/settings/MemorySettings.vue";
+import CompressionSettings from "@/components/hermes/settings/CompressionSettings.vue";
 import SessionSettings from "@/components/hermes/settings/SessionSettings.vue";
 import PrivacySettings from "@/components/hermes/settings/PrivacySettings.vue";
 import ModelSettings from "@/components/hermes/settings/ModelSettings.vue";
@@ -49,6 +50,9 @@ onMounted(() => {
           <NTabPane name="memory" :tab="t('settings.tabs.memory')">
             <MemorySettings />
           </NTabPane>
+          <NTabPane name="compression" :tab="t('settings.tabs.compression')">
+            <CompressionSettings />
+          </NTabPane>
           <NTabPane name="session" :tab="t('settings.tabs.session')">
             <SessionSettings />
           </NTabPane>

packages/server/src/controllers/hermes/sessions.ts

@@ -188,7 +188,16 @@ export async function get(ctx: any) {
  * GET /api/hermes/sessions/hermes/:id
  */
 export async function getHermesSession(ctx: any) {
-  // Try database first (consistent with listHermesSessions)
+  // Prefer the Web UI local session store. Hermes state.db can lag behind or
+  // miss messages for Bridge-backed runs, while the local store is the source
+  // used by chat rendering and compression.
+  const localSession = localGetSessionDetail(ctx.params.id)
+  if (localSession && localSession.source !== 'api_server') {
+    ctx.body = { session: localSession }
+    return
+  }
+
+  // Try Hermes state.db next (consistent with listHermesSessions)
   try {
     const session = await getSessionDetailFromDb(ctx.params.id)
     if (session && session.source !== 'api_server') {

packages/server/src/lib/context-compressor/index.ts

@@ -46,6 +46,8 @@ export interface CompressionConfig {
   triggerTokens: number
   /** Summary token target (default: 8000) */
   summaryBudget: number
+  /** Number of earliest messages to keep verbatim (default: 0) */
+  headMessageCount: number
   /** Number of recent messages to keep verbatim (default: 10) */
   tailMessageCount: number
   /** Timeout for LLM summarization call (default: 60_000ms) */
@@ -55,6 +57,7 @@ export interface CompressionConfig {
 export const DEFAULT_COMPRESSION_CONFIG: CompressionConfig = {
   triggerTokens: 100_000,
   summaryBudget: 8_000,
+  headMessageCount: 0,
   tailMessageCount: 10,
   summarizationTimeoutMs: 120_000,
 }
@@ -108,6 +111,54 @@ export function countTokensForModel(text: string, model: string): number {
   }
 }
 
+function messageTokenEstimate(message: ChatMessage): number {
+  if (typeof message.content === 'string') return countTokens(message.content)
+  if (Array.isArray(message.content)) {
+    return countTokens(message.content.map(block => {
+      if (block.type === 'text') return block.text || ''
+      if (block.type === 'image') return `[Image: ${block.path || ''}]`
+      if (block.type === 'file') return `[File: ${block.path || ''}]`
+      return ''
+    }).join(''))
+  }
+  return 0
+}
+
+function messagesTokenEstimate(messages: ChatMessage[]): number {
+  return messages.reduce((sum, message) => sum + messageTokenEstimate(message), 0)
+}
+
+function truncateTextToTokenBudget(text: string, tokenBudget: number): string {
+  if (tokenBudget <= 0 || countTokens(text) <= tokenBudget) return text
+  let lo = 0
+  let hi = text.length
+  while (lo < hi) {
+    const mid = Math.ceil((lo + hi) / 2)
+    if (countTokens(text.slice(0, mid)) <= tokenBudget) lo = mid
+    else hi = mid - 1
+  }
+  return text.slice(0, lo).trimEnd() + '\n\n[Summary truncated to fit context budget]'
+}
+
+function enforceCompressedBudget(
+  messages: ChatMessage[],
+  triggerTokens: number,
+  summaryIndex: number,
+): ChatMessage[] {
+  if (triggerTokens <= 0 || messagesTokenEstimate(messages) <= triggerTokens) return messages
+
+  const summaryMessage = messages[summaryIndex]
+  if (!summaryMessage || typeof summaryMessage.content !== 'string') return messages
+
+  const summaryOnly = [{ ...summaryMessage }]
+  if (messagesTokenEstimate(summaryOnly) <= triggerTokens) return summaryOnly
+
+  return [{
+    ...summaryMessage,
+    content: truncateTextToTokenBudget(summaryMessage.content, triggerTokens),
+  }]
+}
+
 // ─── Prompts ────────────────────────────────────────────
 
 export const SUMMARY_PREFIX = `[CONTEXT COMPACTION — REFERENCE ONLY] Earlier turns were compacted
@@ -371,6 +422,10 @@ export function pruneOldToolResults(messages: ChatMessage[], keepRecentCount: nu
   return [...pruned, ...tail]
 }
 
+function pruneFallbackToolResults(messages: ChatMessage[], keepRecentCount: number): ChatMessage[] {
+  return pruneOldToolResults(messages, keepRecentCount)
+}
+
 // ─── LLM Summarization ──────────────────────────────────
 
 export async function callSummarizer(
@@ -474,7 +529,7 @@ export class ChatContextCompressor {
     // Check if we have a previous compression snapshot
     const snapshot = sessionId ? getCompressionSnapshot(sessionId) : null
 
-    if (snapshot) {
+    if (snapshot && snapshot.lastMessageIndex >= 0 && snapshot.lastMessageIndex < messages.length) {
       // Has snapshot → incremental compress (merge old summary with new messages)
       logger.info(
         '[context-compressor] session=%s: incremental compress with snapshot at index %d',
@@ -484,6 +539,22 @@ export class ChatContextCompressor {
         messages, snapshot, upstream, apiKey, sessionId!, makeMeta(), summarizer,
       )
     } else {
+      if (snapshot && sessionId) {
+        const fallbackLastMessageIndex = Math.max(-1, messages.length - this.config.tailMessageCount - 1)
+        logger.warn(
+          '[context-compressor] session=%s: stale snapshot index %d for %d messages; using summary plus tail from index %d',
+          sessionId, snapshot.lastMessageIndex, messages.length, fallbackLastMessageIndex,
+        )
+        return this.incrementalCompress(
+          messages,
+          { summary: snapshot.summary, lastMessageIndex: fallbackLastMessageIndex },
+          upstream,
+          apiKey,
+          sessionId,
+          makeMeta(),
+          summarizer,
+        )
+      }
       // No snapshot → full compress (compress all messages)
       logger.info(
         '[context-compressor] session=%s: full compress %d messages',
@@ -504,27 +575,36 @@ export class ChatContextCompressor {
   ): Promise<CompressedResult> {
     const { summary: previousSummary, lastMessageIndex } = snapshot
     const total = messages.length
-    const cleaned = pruneOldToolResults(messages, this.config.tailMessageCount)
+    const headCount = Math.min(this.config.headMessageCount, Math.max(0, lastMessageIndex + 1))
-    const newMessages = cleaned.slice(lastMessageIndex + 1)
+    const head = messages.slice(0, headCount)
+    const newMessages = messages.slice(lastMessageIndex + 1)
     const tailCount = this.config.tailMessageCount
+    const previousSummaryMessage: ChatMessage = { role: 'user', content: SUMMARY_PREFIX + '\n\n' + previousSummary }
+    const assembledWithPrevious = [
+      ...head,
+      previousSummaryMessage,
+      ...newMessages,
+    ]
+    const assembledOverBudget = messagesTokenEstimate(assembledWithPrevious) > this.config.triggerTokens
+    const canKeepTailWindow = newMessages.length > tailCount
 
-    // Keep last N of new messages, compress the rest
+    // If the new segment itself is too small to split but already over budget,
-    const tailStart = Math.max(0, newMessages.length - tailCount)
+    // fold all new messages into the existing summary instead of preserving them verbatim.
+    const tailStart = assembledOverBudget && !canKeepTailWindow
+      ? newMessages.length
+      : Math.max(0, newMessages.length - tailCount)
     const toCompress = newMessages.slice(0, tailStart)
     const tail = newMessages.slice(tailStart)
 
     if (toCompress.length === 0) {
       return {
-        messages: [
+        messages: assembledWithPrevious,
-          { role: 'user', content: SUMMARY_PREFIX + '\n\n' + previousSummary },
-          ...newMessages,
-        ],
         meta: {
           ...meta,
           compressed: true,
           llmCompressed: false,
           summaryTokenEstimate: countTokens(SUMMARY_PREFIX + previousSummary),
-          verbatimCount: newMessages.length,
+          verbatimCount: head.length + newMessages.length,
           compressedStartIndex: lastMessageIndex,
         },
       }
@@ -546,26 +626,32 @@ export class ChatContextCompressor {
       logger.info('[context-compressor] incremental-llm done in %dms, %d chars', Date.now() - t0, summary.length)
     } catch (err: any) {
       logger.warn('[context-compressor] incremental-llm failed: %s — keeping new messages verbatim', err.message)
-      return {
+      const fallback = [
-        messages: [
+        ...head,
-          { role: 'user', content: SUMMARY_PREFIX + '\n\n' + previousSummary },
+        previousSummaryMessage,
         ...newMessages,
-        ],
+      ]
+      const prunedFallback = pruneFallbackToolResults(fallback, this.config.tailMessageCount)
+      const budgetedFallback = enforceCompressedBudget(prunedFallback, this.config.triggerTokens, head.length)
+      return {
+        messages: budgetedFallback,
         meta: {
           ...meta,
           compressed: true,
           llmCompressed: false,
           summaryTokenEstimate: countTokens(SUMMARY_PREFIX + previousSummary),
-          verbatimCount: newMessages.length,
+          verbatimCount: budgetedFallback.length === fallback.length ? head.length + newMessages.length : 0,
           compressedStartIndex: lastMessageIndex,
         },
       }
     }
 
-    const result: ChatMessage[] = [
+    let result: ChatMessage[] = [
+      ...head,
       { role: 'user', content: SUMMARY_PREFIX + '\n\n' + summary },
       ...tail,
     ]
+    result = enforceCompressedBudget(result, this.config.triggerTokens, head.length)
 
     const newLastIndex = lastMessageIndex + tailStart
     if (sessionId) {
@@ -579,7 +665,7 @@ export class ChatContextCompressor {
         compressed: true,
         llmCompressed: true,
         summaryTokenEstimate: countTokens(SUMMARY_PREFIX + summary),
-        verbatimCount: tail.length,
+        verbatimCount: result.length === head.length + 1 + tail.length ? head.length + tail.length : 0,
         compressedStartIndex: newLastIndex,
       },
     }
@@ -594,20 +680,20 @@ export class ChatContextCompressor {
     summarizer?: string | SummarizerOptions,
   ): Promise<CompressedResult> {
     const total = messages.length
-    const cleaned = pruneOldToolResults(messages, this.config.tailMessageCount)
+    const requestedHeadCount = Math.min(this.config.headMessageCount, total)
-    const tailCount = this.config.tailMessageCount
+    const requestedTailCount = this.config.tailMessageCount
-
+    const canKeepProtectedWindows = total > requestedHeadCount + requestedTailCount
-    if (total <= tailCount) {
+    const headCount = canKeepProtectedWindows ? requestedHeadCount : 0
-      return { messages: cleaned, meta }
+    const tailCount = canKeepProtectedWindows ? requestedTailCount : 0
-    }
 
     const tailStart = total - tailCount
-    const toCompress = cleaned.slice(0, tailStart)
+    const head = messages.slice(0, headCount)
-    const tail = cleaned.slice(tailStart)
+    const toCompress = messages.slice(headCount, tailStart)
+    const tail = messages.slice(tailStart)
 
     logger.info(
-      '[context-compressor] [full-llm] compressing messages 0-%d, keeping %d-%d',
+      '[context-compressor] [full-llm] compressing messages %d-%d, keeping first %d and last %d',
-      tailStart - 1, tailStart, total - 1,
+      headCount, tailStart - 1, head.length, tail.length,
     )
 
     const contentToSummarize = serializeForSummary(toCompress)
@@ -624,26 +710,28 @@ export class ChatContextCompressor {
     }
 
     if (!summary) {
-      return { messages: cleaned, meta }
+      return { messages: pruneFallbackToolResults(messages, this.config.tailMessageCount), meta }
     }
 
     const result: ChatMessage[] = []
 
+    result.push(...head)
     result.push({ role: 'user', content: SUMMARY_PREFIX + '\n\n' + summary })
     if (sessionId) {
       saveCompressionSnapshot(sessionId, summary, tailStart - 1, total)
     }
 
     result.push(...tail)
+    const budgetedResult = enforceCompressedBudget(result, this.config.triggerTokens, head.length)
 
     return {
-      messages: result,
+      messages: budgetedResult,
       meta: {
         ...meta,
         compressed: true,
         llmCompressed: !!summary,
         summaryTokenEstimate: summary ? countTokens(SUMMARY_PREFIX + summary) : 0,
-        verbatimCount: tail.length,
+        verbatimCount: budgetedResult.length === result.length ? head.length + tail.length : 0,
         compressedStartIndex: tailStart - 1,
       },
     }

packages/server/src/services/hermes/run-chat/compression.ts

@@ -10,14 +10,88 @@ import {
 import { getCompressionSnapshot } from '../../../db/hermes/compression-snapshot'
 import { ChatContextCompressor, SUMMARY_PREFIX } from '../../../lib/context-compressor'
 import { getModelContextLength } from '../model-context'
+import { readConfigYamlForProfile } from '../../config-helpers'
 import { logger } from '../../logger'
 import { bridgeLogger } from '../../logger'
 import { calcAndUpdateUsage, estimateUsageTokensFromMessages } from './usage'
 import { isAssistantMessageSendable } from './message-format'
-import type { ChatMessage } from '../../../lib/context-compressor'
+import type { ChatMessage, CompressionConfig as CompressorConfig } from '../../../lib/context-compressor'
 import type { SessionState, BridgeCompressionResult } from './types'
 
-const compressor = new ChatContextCompressor()
+interface RunChatCompressionConfig {
+  enabled: boolean
+  triggerTokens: number
+  compressor: Partial<CompressorConfig>
+}
+
+function isSnapshotUsable(
+  snapshot: { lastMessageIndex: number } | null,
+  history: ChatMessage[],
+): boolean {
+  return !!snapshot && snapshot.lastMessageIndex >= 0 && snapshot.lastMessageIndex < history.length
+}
+
+function buildSnapshotHistory(
+  snapshot: { summary: string; lastMessageIndex: number } | null,
+  history: ChatMessage[],
+  compressionConfig?: Partial<CompressorConfig>,
+): ChatMessage[] | null {
+  if (!snapshot) return null
+  const headCount = compressionConfig?.headMessageCount || 0
+  const tailCount = compressionConfig?.tailMessageCount || 0
+  const protectedHead = headCount > 0 ? history.slice(0, headCount) : []
+  const summaryMessage = { role: 'user', content: SUMMARY_PREFIX + '\n\n' + snapshot.summary } as ChatMessage
+
+  if (isSnapshotUsable(snapshot, history)) {
+    return [
+      ...protectedHead,
+      summaryMessage,
+      ...history.slice(snapshot.lastMessageIndex + 1),
+    ]
+  }
+
+  const tailStart = Math.max(protectedHead.length, history.length - tailCount)
+  return [
+    ...protectedHead,
+    summaryMessage,
+    ...history.slice(tailStart),
+  ]
+}
+
+function clampRatio(value: unknown, fallback: number, min: number, max: number): number {
+  const n = typeof value === 'number' && Number.isFinite(value) ? value : fallback
+  return Math.min(max, Math.max(min, n))
+}
+
+function clampInt(value: unknown, fallback: number, min: number, max: number): number {
+  const n = typeof value === 'number' && Number.isFinite(value) ? Math.floor(value) : fallback
+  return Math.min(max, Math.max(min, n))
+}
+
+async function getRunChatCompressionConfig(profile: string, contextLength: number): Promise<RunChatCompressionConfig> {
+  let raw: Record<string, any> = {}
+  try {
+    raw = (await readConfigYamlForProfile(profile))?.compression || {}
+  } catch (err) {
+    logger.warn(err, '[context-compress] failed to read compression config for profile %s, using defaults', profile)
+  }
+
+  const threshold = clampRatio(raw.threshold, 0.5, 0.05, 0.95)
+  const targetRatio = clampRatio(raw.target_ratio, 0.2, 0.01, 0.8)
+  const protectLastN = clampInt(raw.protect_last_n, 20, 0, 500)
+  const protectFirstN = clampInt(raw.protect_first_n, 3, 0, 100)
+
+  return {
+    enabled: raw.enabled !== false,
+    triggerTokens: Math.floor(contextLength * threshold),
+    compressor: {
+      triggerTokens: Math.floor(contextLength * threshold),
+      summaryBudget: Math.max(1_000, Math.floor(contextLength * targetRatio)),
+      headMessageCount: protectFirstN,
+      tailMessageCount: protectLastN,
+    },
+  }
+}
 
 /**
  * Load conversation history from DB with full message structure (user/assistant/tool).
@@ -77,12 +151,7 @@ export function estimateSnapshotAwareHistoryUsage(
   history: ChatMessage[],
 ): { messageCount: number; tokenCount: number } {
   const snapshot = getCompressionSnapshot(sessionId)
-  const messages = snapshot
+  const messages = buildSnapshotHistory(snapshot, history) || history
-    ? [
-        { role: 'user', content: SUMMARY_PREFIX + snapshot.summary },
-        ...history.slice(snapshot.lastMessageIndex + 1),
-      ]
-    : history
   const usage = estimateUsageTokensFromMessages(messages)
   return {
     messageCount: messages.length,
@@ -108,29 +177,45 @@ export async function buildCompressedHistory(
       model: modelContext.model,
       provider: modelContext.provider,
     })
-    const triggerTokens = Math.floor(contextLength / 2)
+    const compressionConfig = await getRunChatCompressionConfig(profile, contextLength)
+    const triggerTokens = compressionConfig.triggerTokens
+    if (!compressionConfig.enabled) {
+      logger.info('[context-compress] session=%s: compression disabled by config', sessionId)
+      return history
+    }
     const cState = getOrCreateSession(sessionMap, sessionId)
     const assembledTokens = await calcAndUpdateUsage(sessionId, cState, emit)
-    const totalTokens = assembledTokens.inputTokens + assembledTokens.outputTokens
+    let totalTokens = assembledTokens.inputTokens + assembledTokens.outputTokens
     const snapshot = getCompressionSnapshot(sessionId)
+    const staleSnapshot = snapshot && !isSnapshotUsable(snapshot, history)
+    if (staleSnapshot) {
+      logger.warn('[context-compress] session=%s: stale snapshot index %d for %d history messages; using summary plus safe tail',
+        sessionId, snapshot.lastMessageIndex, history.length)
+      const staleHistory = buildSnapshotHistory(snapshot, history, compressionConfig.compressor) || history
+      const staleUsage = estimateUsageTokensFromMessages(staleHistory)
+      totalTokens = staleUsage.inputTokens + staleUsage.outputTokens
+    }
 
-    if (snapshot) {
+    if (snapshot && !staleSnapshot) {
       const newMessages = history.slice(snapshot.lastMessageIndex + 1)
       logger.info('[context-compress] session=%s: snapshot at %d, %d new messages, assembled ~%d tokens (threshold %d)',
         sessionId, snapshot.lastMessageIndex, newMessages.length, totalTokens, triggerTokens)
-      if (totalTokens <= triggerTokens && newMessages.length <= 150) {
+      if (totalTokens <= triggerTokens) {
-        history = [
+        history = buildSnapshotHistory(snapshot, history, compressionConfig.compressor) || history
-          { role: 'user', content: SUMMARY_PREFIX + '\n\n' + snapshot.summary },
-          ...newMessages,
-        ] as ChatMessage[]
       } else {
-        history = await compressHistory(history, newMessages, sessionId, upstream, apiKey, cState, totalTokens, emit, sessionMap, modelContext)
+        history = await compressHistory(history, newMessages, sessionId, upstream, apiKey, cState, totalTokens, emit, sessionMap, modelContext, compressionConfig.compressor)
+      }
+    } else if (snapshot && staleSnapshot) {
+      if (totalTokens <= triggerTokens) {
+        history = buildSnapshotHistory(snapshot, history, compressionConfig.compressor) || history
+      } else {
+        history = await compressHistory(history, null, sessionId, upstream, apiKey, cState, totalTokens, emit, sessionMap, modelContext, compressionConfig.compressor)
       }
     } else if (history.length > 4) {
-      if (totalTokens <= triggerTokens && history.length <= 150) {
+      if (totalTokens <= triggerTokens) {
         logger.info('[context-compress] session=%s: %d messages, ~%d tokens — under threshold, skip', sessionId, history.length, totalTokens)
       } else {
-        history = await compressHistory(history, null, sessionId, upstream, apiKey, cState, totalTokens, emit, sessionMap, modelContext)
+        history = await compressHistory(history, null, sessionId, upstream, apiKey, cState, totalTokens, emit, sessionMap, modelContext, compressionConfig.compressor)
       }
     }
 
@@ -152,6 +237,7 @@ export async function compressHistory(
   emit: (event: string, payload: any) => void,
   sessionMap: Map<string, SessionState>,
   modelContext: { model?: string | null; provider?: string | null } = {},
+  compressionConfig?: Partial<CompressorConfig>,
 ): Promise<ChatMessage[]> {
   const msgCount = newMessagesOnly ? newMessagesOnly.length : history.length
   pushState(sessionMap, sessionId, 'compression.started', {
@@ -163,6 +249,7 @@ export async function compressHistory(
 
   try {
     const session = getSession(sessionId)
+    const compressor = new ChatContextCompressor({ config: compressionConfig })
     const result = await compressor.compress(history, upstream, apiKey, sessionId, {
       profile: session?.profile,
       model: modelContext.model || session?.model,
@@ -244,6 +331,8 @@ export async function forceCompressBridgeHistory(
   const upstream = ''
   const apiKey = undefined
   const session = getSession(sessionId)
+  const contextLength = getModelContextLength({ profile, model: session?.model, provider: session?.provider })
+  const compressionConfig = await getRunChatCompressionConfig(session?.profile || profile, contextLength)
   const beforeUsage = estimateSnapshotAwareHistoryUsage(sessionId, history)
   const totalTokens = beforeUsage.tokenCount
   bridgeLogger.info({
@@ -256,6 +345,7 @@ export async function forceCompressBridgeHistory(
     snapshotAware: true,
   }, '[chat-run-socket] bridge forced compression started')
 
+  const compressor = new ChatContextCompressor({ config: compressionConfig.compressor })
   const result = await compressor.compress(history, upstream, apiKey, sessionId, {
     profile: session?.profile || profile,
     model: session?.model,

packages/server/src/services/hermes/run-chat/handle-api-run.ts

@@ -38,7 +38,7 @@ export async function loadSessionStateFromDb(sid: string, _sessionMap: Map<strin
     let inputTokens: number
     let outputTokens: number
     const snapshot = getCompressionSnapshot(sid)
-    if (snapshot) {
+    if (snapshot && snapshot.lastMessageIndex >= 0 && snapshot.lastMessageIndex < messages.length) {
       const newMessages = messages.slice(snapshot.lastMessageIndex + 1)
       const newUsage = estimateUsageTokensFromMessages(newMessages)
       inputTokens = countTokens(SUMMARY_PREFIX + snapshot.summary) +

packages/server/src/services/hermes/run-chat/usage.ts

@@ -53,7 +53,7 @@ export async function calcAndUpdateUsage(
     const snapshot = getCompressionSnapshot(sid)
     let inputTokens: number
     let outputTokens: number
-    if (snapshot && msgs.length) {
+    if (snapshot && msgs.length && snapshot.lastMessageIndex >= 0 && snapshot.lastMessageIndex < msgs.length) {
       const newMessages = msgs.slice(snapshot.lastMessageIndex + 1)
       const newUsage = estimateUsageTokensFromMessages(newMessages)
       inputTokens = countTokens(SUMMARY_PREFIX + snapshot.summary) +

tests/server/context-compressor.test.ts

@@ -3,6 +3,8 @@ import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
 const getCompressionSnapshotMock = vi.fn()
 const saveCompressionSnapshotMock = vi.fn()
 const deleteCompressionSnapshotMock = vi.fn()
+const bridgeRequestMock = vi.fn()
+const bridgeDestroyMock = vi.fn()
 
 vi.mock('../../packages/server/src/services/logger', () => ({
   logger: {
@@ -19,6 +21,13 @@ vi.mock('../../packages/server/src/db/hermes/compression-snapshot', () => ({
   deleteCompressionSnapshot: deleteCompressionSnapshotMock,
 }))
 
+vi.mock('../../packages/server/src/services/hermes/agent-bridge', () => ({
+  AgentBridgeClient: class {
+    request = bridgeRequestMock
+    destroy = bridgeDestroyMock
+  },
+}))
+
 describe('ChatContextCompressor', () => {
   let originalFetch: typeof global.fetch
 
@@ -27,6 +36,10 @@ describe('ChatContextCompressor', () => {
     getCompressionSnapshotMock.mockReset()
     saveCompressionSnapshotMock.mockReset()
     deleteCompressionSnapshotMock.mockReset()
+    bridgeRequestMock.mockReset()
+    bridgeDestroyMock.mockReset()
+    bridgeRequestMock.mockRejectedValue(new Error('summarizer failed'))
+    bridgeDestroyMock.mockResolvedValue(undefined)
   })
 
   afterEach(() => {
@@ -42,7 +55,6 @@ describe('ChatContextCompressor', () => {
     }))
 
     getCompressionSnapshotMock.mockReturnValue(null)
-    global.fetch = vi.fn(async () => ({ ok: false, status: 500 })) as any
 
     const result = await compressor.compress(messages, 'http://upstream', undefined, 's1')
 
@@ -66,7 +78,6 @@ describe('ChatContextCompressor', () => {
       lastMessageIndex: 1,
       messageCountAtTime: 2,
     })
-    global.fetch = vi.fn(async () => ({ ok: false, status: 500 })) as any
 
     const result = await compressor.compress(messages, 'http://upstream', undefined, 's1')
 
@@ -109,4 +120,331 @@ describe('ChatContextCompressor', () => {
     expect(result.meta.compressedStartIndex).toBe(3)
     expect(saveCompressionSnapshotMock).not.toHaveBeenCalled()
   })
+
+  it('keeps configured first and last messages during full compression', async () => {
+    const { ChatContextCompressor, SUMMARY_PREFIX } = await import('../../packages/server/src/lib/context-compressor')
+    const compressor = new ChatContextCompressor({
+      config: { headMessageCount: 2, tailMessageCount: 3, summaryBudget: 1000 },
+    })
+    const messages = Array.from({ length: 10 }, (_, i) => ({
+      role: i % 2 === 0 ? 'user' : 'assistant',
+      content: `message ${i}`,
+    }))
+
+    getCompressionSnapshotMock.mockReturnValue(null)
+    bridgeRequestMock.mockResolvedValue({
+      status: 'completed',
+      result: { final_response: 'compressed summary' },
+    })
+
+    const result = await compressor.compress(messages, 'http://upstream', undefined, 's1')
+
+    expect(result.messages.map(m => m.content)).toEqual([
+      'message 0',
+      'message 1',
+      `${SUMMARY_PREFIX}\n\ncompressed summary`,
+      'message 7',
+      'message 8',
+      'message 9',
+    ])
+    expect(result.meta.compressed).toBe(true)
+    expect(result.meta.llmCompressed).toBe(true)
+    expect(result.meta.verbatimCount).toBe(5)
+    expect(saveCompressionSnapshotMock).toHaveBeenCalledWith('s1', 'compressed summary', 6, 10)
+  })
+
+  it('does not pre-prune tool results before sending them to the summarizer', async () => {
+    const { ChatContextCompressor } = await import('../../packages/server/src/lib/context-compressor')
+    const compressor = new ChatContextCompressor({
+      config: { headMessageCount: 0, tailMessageCount: 1, summaryBudget: 1000 },
+    })
+    const longToolOutput = `${'x'.repeat(180)}KEEP_MARKER${'y'.repeat(180)}`
+    const messages = [
+      {
+        role: 'assistant',
+        content: 'calling terminal',
+        tool_calls: [{ id: 'call_1', type: 'function', function: { name: 'terminal', arguments: '{}' } }],
+      },
+      { role: 'tool', name: 'terminal', tool_call_id: 'call_1', content: longToolOutput },
+      { role: 'user', content: 'tail' },
+    ]
+
+    getCompressionSnapshotMock.mockReturnValue(null)
+    bridgeRequestMock.mockResolvedValue({
+      status: 'completed',
+      result: { final_response: 'compressed summary' },
+    })
+
+    await compressor.compress(messages, 'http://upstream', undefined, 's1')
+
+    const request = bridgeRequestMock.mock.calls[0][0]
+    const serializedHistory = JSON.stringify(request.conversation_history)
+    expect(serializedHistory).toContain('KEEP_MARKER')
+    expect(serializedHistory).not.toContain('[terminal] ')
+  })
+
+  it('keeps protected head tool results verbatim after successful full compression', async () => {
+    const { ChatContextCompressor, SUMMARY_PREFIX } = await import('../../packages/server/src/lib/context-compressor')
+    const compressor = new ChatContextCompressor({
+      config: { headMessageCount: 2, tailMessageCount: 1, summaryBudget: 1000 },
+    })
+    const longToolOutput = `${'head-tool-output '.repeat(30)}KEEP_HEAD_TOOL`
+    const messages = [
+      {
+        role: 'assistant',
+        content: 'calling terminal',
+        tool_calls: [{ id: 'call_1', type: 'function', function: { name: 'terminal', arguments: '{}' } }],
+      },
+      { role: 'tool', name: 'terminal', tool_call_id: 'call_1', content: longToolOutput },
+      { role: 'user', content: 'middle' },
+      { role: 'assistant', content: 'tail' },
+    ]
+
+    getCompressionSnapshotMock.mockReturnValue(null)
+    bridgeRequestMock.mockResolvedValue({
+      status: 'completed',
+      result: { final_response: 'compressed summary' },
+    })
+
+    const result = await compressor.compress(messages, 'http://upstream', undefined, 's1')
+
+    expect(result.messages.map(m => m.content)).toEqual([
+      'calling terminal',
+      longToolOutput,
+      `${SUMMARY_PREFIX}\n\ncompressed summary`,
+      'tail',
+    ])
+  })
+
+  it('uses the previous summary plus a safe tail when an existing snapshot index is stale', async () => {
+    const { ChatContextCompressor, SUMMARY_PREFIX } = await import('../../packages/server/src/lib/context-compressor')
+    const compressor = new ChatContextCompressor({
+      config: { headMessageCount: 2, tailMessageCount: 3, summaryBudget: 1000 },
+    })
+    const messages = Array.from({ length: 8 }, (_, i) => ({
+      role: i % 2 === 0 ? 'user' : 'assistant',
+      content: `message ${i}`,
+    }))
+
+    getCompressionSnapshotMock.mockReturnValue({
+      summary: 'stale previous summary',
+      lastMessageIndex: 20,
+      messageCountAtTime: 21,
+    })
+    bridgeRequestMock.mockResolvedValue({
+      status: 'completed',
+      result: { final_response: 'rebuilt summary' },
+    })
+
+    const result = await compressor.compress(messages, 'http://upstream', undefined, 's1')
+
+    expect(deleteCompressionSnapshotMock).not.toHaveBeenCalled()
+    expect(bridgeRequestMock).not.toHaveBeenCalled()
+    expect(result.messages.map(m => m.content)).toEqual([
+      'message 0',
+      'message 1',
+      `${SUMMARY_PREFIX}\n\nstale previous summary`,
+      'message 5',
+      'message 6',
+      'message 7',
+    ])
+    expect(saveCompressionSnapshotMock).not.toHaveBeenCalled()
+  })
+
+  it('folds a stale snapshot safe tail into a new summary when it still exceeds budget', async () => {
+    const { ChatContextCompressor, SUMMARY_PREFIX } = await import('../../packages/server/src/lib/context-compressor')
+    const compressor = new ChatContextCompressor({
+      config: { triggerTokens: 800, headMessageCount: 2, tailMessageCount: 3, summaryBudget: 1000 },
+    })
+    const largeTail = 'tail-token '.repeat(200)
+    const messages = [
+      { role: 'user', content: 'message 0' },
+      { role: 'assistant', content: 'message 1' },
+      { role: 'user', content: 'message 2' },
+      { role: 'assistant', content: 'message 3' },
+      { role: 'user', content: 'message 4' },
+      { role: 'assistant', content: largeTail },
+      { role: 'user', content: largeTail },
+      { role: 'assistant', content: largeTail },
+    ]
+
+    getCompressionSnapshotMock.mockReturnValue({
+      summary: 'stale previous summary',
+      lastMessageIndex: 20,
+      messageCountAtTime: 21,
+    })
+    bridgeRequestMock.mockResolvedValue({
+      status: 'completed',
+      result: { final_response: 'updated stale summary' },
+    })
+
+    const result = await compressor.compress(messages, 'http://upstream', undefined, 's1')
+
+    expect(deleteCompressionSnapshotMock).not.toHaveBeenCalled()
+    expect(bridgeRequestMock).toHaveBeenCalledTimes(1)
+    expect(result.messages.map(m => m.content)).toEqual([
+      'message 0',
+      'message 1',
+      `${SUMMARY_PREFIX}\n\nupdated stale summary`,
+    ])
+    expect(saveCompressionSnapshotMock).toHaveBeenCalledWith('s1', 'updated stale summary', 7, 8)
+  })
+
+  it('compresses the full history when protected windows cover all messages', async () => {
+    const { ChatContextCompressor, SUMMARY_PREFIX } = await import('../../packages/server/src/lib/context-compressor')
+    const compressor = new ChatContextCompressor({
+      config: { headMessageCount: 3, tailMessageCount: 20, summaryBudget: 1000 },
+    })
+    const messages = Array.from({ length: 20 }, (_, i) => ({
+      role: i % 2 === 0 ? 'user' : 'assistant',
+      content: `message ${i}`,
+    }))
+
+    getCompressionSnapshotMock.mockReturnValue(null)
+    bridgeRequestMock.mockResolvedValue({
+      status: 'completed',
+      result: { final_response: 'compressed all messages' },
+    })
+
+    const result = await compressor.compress(messages, 'http://upstream', undefined, 's1')
+
+    expect(bridgeRequestMock).toHaveBeenCalledTimes(1)
+    expect(result.messages.map(m => m.content)).toEqual([
+      `${SUMMARY_PREFIX}\n\ncompressed all messages`,
+    ])
+    expect(result.meta.compressed).toBe(true)
+    expect(result.meta.llmCompressed).toBe(true)
+    expect(result.meta.verbatimCount).toBe(0)
+    expect(result.meta.compressedStartIndex).toBe(19)
+    expect(saveCompressionSnapshotMock).toHaveBeenCalledWith('s1', 'compressed all messages', 19, 20)
+  })
+
+  it('drops protected messages when compressed output still exceeds the trigger budget', async () => {
+    const { ChatContextCompressor, SUMMARY_PREFIX } = await import('../../packages/server/src/lib/context-compressor')
+    const compressor = new ChatContextCompressor({
+      config: { triggerTokens: 200, headMessageCount: 2, tailMessageCount: 2, summaryBudget: 100 },
+    })
+    const largeText = 'tail-token '.repeat(500)
+    const messages = [
+      { role: 'user', content: 'head 0' },
+      { role: 'assistant', content: 'head 1' },
+      { role: 'user', content: 'middle 2' },
+      { role: 'assistant', content: 'middle 3' },
+      { role: 'user', content: largeText },
+      { role: 'assistant', content: largeText },
+    ]
+
+    getCompressionSnapshotMock.mockReturnValue(null)
+    bridgeRequestMock.mockResolvedValue({
+      status: 'completed',
+      result: { final_response: 'short summary' },
+    })
+
+    const result = await compressor.compress(messages, 'http://upstream', undefined, 's1')
+
+    expect(result.messages.map(m => m.content)).toEqual([
+      `${SUMMARY_PREFIX}\n\nshort summary`,
+    ])
+    expect(result.meta.compressed).toBe(true)
+    expect(result.meta.llmCompressed).toBe(true)
+    expect(result.meta.verbatimCount).toBe(0)
+  })
+
+  it('truncates the summary when the summary alone exceeds the trigger budget', async () => {
+    const { ChatContextCompressor, SUMMARY_PREFIX, countTokens } = await import('../../packages/server/src/lib/context-compressor')
+    const compressor = new ChatContextCompressor({
+      config: { triggerTokens: 120, headMessageCount: 2, tailMessageCount: 2, summaryBudget: 100 },
+    })
+    const messages = Array.from({ length: 6 }, (_, i) => ({
+      role: i % 2 === 0 ? 'user' : 'assistant',
+      content: `message ${i}`,
+    }))
+    const longSummary = 'summary-token '.repeat(500)
+
+    getCompressionSnapshotMock.mockReturnValue(null)
+    bridgeRequestMock.mockResolvedValue({
+      status: 'completed',
+      result: { final_response: longSummary },
+    })
+
+    const result = await compressor.compress(messages, 'http://upstream', undefined, 's1')
+
+    expect(result.messages).toHaveLength(1)
+    expect(String(result.messages[0].content)).toContain('[Summary truncated to fit context budget]')
+    expect(String(result.messages[0].content).startsWith(SUMMARY_PREFIX)).toBe(true)
+    expect(countTokens(String(result.messages[0].content))).toBeLessThanOrEqual(140)
+    expect(result.meta.verbatimCount).toBe(0)
+  })
+
+  it('keeps configured first messages when incremental compression reuses an existing snapshot', async () => {
+    const { ChatContextCompressor, SUMMARY_PREFIX } = await import('../../packages/server/src/lib/context-compressor')
+    const compressor = new ChatContextCompressor({
+      config: { headMessageCount: 2, tailMessageCount: 10 },
+    })
+    const messages = Array.from({ length: 6 }, (_, i) => ({
+      role: i % 2 === 0 ? 'user' : 'assistant',
+      content: `message ${i}`,
+    }))
+
+    getCompressionSnapshotMock.mockReturnValue({
+      summary: 'previous summary',
+      lastMessageIndex: 3,
+      messageCountAtTime: 4,
+    })
+
+    const result = await compressor.compress(messages, 'http://upstream', undefined, 's1')
+
+    expect(bridgeRequestMock).not.toHaveBeenCalled()
+    expect(result.messages.map(m => m.content)).toEqual([
+      'message 0',
+      'message 1',
+      `${SUMMARY_PREFIX}\n\nprevious summary`,
+      'message 4',
+      'message 5',
+    ])
+    expect(result.meta.verbatimCount).toBe(4)
+    expect(saveCompressionSnapshotMock).not.toHaveBeenCalled()
+  })
+
+  it('folds all new messages into the summary when incremental tail protection would exceed budget', async () => {
+    const { ChatContextCompressor, SUMMARY_PREFIX } = await import('../../packages/server/src/lib/context-compressor')
+    const compressor = new ChatContextCompressor({
+      config: { triggerTokens: 1000, headMessageCount: 3, tailMessageCount: 20, summaryBudget: 100 },
+    })
+    const largeText = 'new-token '.repeat(80)
+    const messages = [
+      { role: 'user', content: 'head 0' },
+      { role: 'assistant', content: 'head 1' },
+      { role: 'user', content: 'head 2' },
+      ...Array.from({ length: 20 }, (_, i) => ({
+        role: i % 2 === 0 ? 'user' : 'assistant',
+        content: `${largeText}${i}`,
+      })),
+    ]
+
+    getCompressionSnapshotMock.mockReturnValue({
+      summary: 'previous summary',
+      lastMessageIndex: 2,
+      messageCountAtTime: 3,
+    })
+    bridgeRequestMock.mockResolvedValue({
+      status: 'completed',
+      result: { final_response: 'updated summary' },
+    })
+
+    const result = await compressor.compress(messages, 'http://upstream', undefined, 's1')
+
+    expect(bridgeRequestMock).toHaveBeenCalledTimes(1)
+    expect(result.messages.map(m => m.content)).toEqual([
+      'head 0',
+      'head 1',
+      'head 2',
+      `${SUMMARY_PREFIX}\n\nupdated summary`,
+    ])
+    expect(result.meta.compressed).toBe(true)
+    expect(result.meta.llmCompressed).toBe(true)
+    expect(result.meta.verbatimCount).toBe(3)
+    expect(result.meta.compressedStartIndex).toBe(22)
+    expect(saveCompressionSnapshotMock).toHaveBeenCalledWith('s1', 'updated summary', 22, 23)
+  })
 })

tests/server/run-chat-compression.test.ts

@@ -0,0 +1,349 @@
+import { beforeEach, describe, expect, it, vi } from 'vitest'
+
+const getSessionDetailMock = vi.fn()
+const getSessionMock = vi.fn()
+const getCompressionSnapshotMock = vi.fn()
+const getModelContextLengthMock = vi.fn()
+const calcAndUpdateUsageMock = vi.fn()
+const estimateUsageTokensFromMessagesMock = vi.fn()
+const compressorCompressMock = vi.fn()
+const readConfigYamlForProfileMock = vi.fn()
+const compressorConstructorMock = vi.fn()
+
+vi.mock('../../packages/server/src/db/hermes/session-store', () => ({
+  getSessionDetail: getSessionDetailMock,
+  getSession: getSessionMock,
+}))
+
+vi.mock('../../packages/server/src/db/hermes/compression-snapshot', () => ({
+  getCompressionSnapshot: getCompressionSnapshotMock,
+}))
+
+vi.mock('../../packages/server/src/lib/context-compressor', () => ({
+  SUMMARY_PREFIX: '[Previous context summary]',
+  ChatContextCompressor: class {
+    constructor(opts?: any) {
+      compressorConstructorMock(opts)
+    }
+    compress = compressorCompressMock
+  },
+}))
+
+vi.mock('../../packages/server/src/services/hermes/model-context', () => ({
+  getModelContextLength: getModelContextLengthMock,
+}))
+
+vi.mock('../../packages/server/src/services/config-helpers', () => ({
+  readConfigYamlForProfile: readConfigYamlForProfileMock,
+}))
+
+vi.mock('../../packages/server/src/services/logger', () => ({
+  logger: {
+    info: vi.fn(),
+    warn: vi.fn(),
+    error: vi.fn(),
+    debug: vi.fn(),
+  },
+  bridgeLogger: {
+    info: vi.fn(),
+    warn: vi.fn(),
+    error: vi.fn(),
+    debug: vi.fn(),
+  },
+}))
+
+vi.mock('../../packages/server/src/services/hermes/run-chat/usage', () => ({
+  calcAndUpdateUsage: calcAndUpdateUsageMock,
+  estimateUsageTokensFromMessages: estimateUsageTokensFromMessagesMock,
+}))
+
+vi.mock('../../packages/server/src/services/hermes/run-chat/message-format', () => ({
+  isAssistantMessageSendable: vi.fn(() => true),
+}))
+
+describe('run chat compression trigger', () => {
+  beforeEach(() => {
+    getSessionDetailMock.mockReset()
+    getSessionMock.mockReset()
+    getCompressionSnapshotMock.mockReset()
+    getModelContextLengthMock.mockReset()
+    calcAndUpdateUsageMock.mockReset()
+    estimateUsageTokensFromMessagesMock.mockReset()
+    compressorCompressMock.mockReset()
+    compressorConstructorMock.mockReset()
+    readConfigYamlForProfileMock.mockReset()
+
+    getSessionMock.mockReturnValue({ id: 'session-1', profile: 'default' })
+    getModelContextLengthMock.mockReturnValue(200_000)
+    calcAndUpdateUsageMock.mockResolvedValue({ inputTokens: 1_000, outputTokens: 0 })
+    estimateUsageTokensFromMessagesMock.mockReturnValue({ inputTokens: 0, outputTokens: 0 })
+    getCompressionSnapshotMock.mockReturnValue(null)
+    readConfigYamlForProfileMock.mockResolvedValue({})
+  })
+
+  it('does not compress long low-token history just because it has more than 150 messages', async () => {
+    const messages = Array.from({ length: 152 }, (_, index) => ({
+      id: index + 1,
+      session_id: 'session-1',
+      role: index === 151 ? 'user' : index % 2 === 0 ? 'user' : 'assistant',
+      content: `m${index}`,
+      timestamp: index + 1,
+      tool_call_id: null,
+      tool_calls: null,
+      tool_name: null,
+      finish_reason: null,
+      reasoning_content: null,
+    }))
+    getSessionDetailMock.mockReturnValue({ messages })
+
+    const { buildCompressedHistory } = await import('../../packages/server/src/services/hermes/run-chat/compression')
+    const history = await buildCompressedHistory(
+      'session-1',
+      'default',
+      'http://upstream',
+      undefined,
+      vi.fn(),
+      new Map(),
+    )
+
+    expect(history).toHaveLength(151)
+    expect(history[0]).toEqual({ role: 'user', content: 'm0' })
+    expect(history.at(-1)).toEqual({ role: 'user', content: 'm150' })
+    expect(compressorCompressMock).not.toHaveBeenCalled()
+  })
+
+  it('uses configured threshold before triggering compression', async () => {
+    const messages = Array.from({ length: 10 }, (_, index) => ({
+      id: index + 1,
+      session_id: 'session-1',
+      role: index === 9 ? 'user' : index % 2 === 0 ? 'user' : 'assistant',
+      content: `message ${index}`,
+      timestamp: index + 1,
+      tool_call_id: null,
+      tool_calls: null,
+      tool_name: null,
+      finish_reason: null,
+      reasoning_content: null,
+    }))
+    getSessionDetailMock.mockReturnValue({ messages })
+    readConfigYamlForProfileMock.mockResolvedValue({
+      compression: { threshold: 0.25, target_ratio: 0.1, protect_last_n: 7, protect_first_n: 2 },
+    })
+    calcAndUpdateUsageMock.mockResolvedValue({ inputTokens: 60_000, outputTokens: 0 })
+    compressorCompressMock.mockResolvedValue({
+      messages: [{ role: 'user', content: 'compressed' }],
+      meta: {
+        compressed: true,
+        llmCompressed: true,
+        totalMessages: 9,
+        summaryTokenEstimate: 1,
+        verbatimCount: 0,
+        compressedStartIndex: 0,
+      },
+    })
+
+    const { buildCompressedHistory } = await import('../../packages/server/src/services/hermes/run-chat/compression')
+    const history = await buildCompressedHistory(
+      'session-1',
+      'default',
+      'http://upstream',
+      undefined,
+      vi.fn(),
+      new Map(),
+    )
+
+    expect(history).toEqual([{ role: 'user', content: 'compressed' }])
+    expect(compressorCompressMock).toHaveBeenCalledWith(
+      expect.any(Array),
+      'http://upstream',
+      undefined,
+      'session-1',
+      expect.objectContaining({ profile: 'default' }),
+    )
+  })
+
+  it('merges partial compression config with defaults', async () => {
+    const messages = Array.from({ length: 10 }, (_, index) => ({
+      id: index + 1,
+      session_id: 'session-1',
+      role: index === 9 ? 'user' : index % 2 === 0 ? 'user' : 'assistant',
+      content: `message ${index}`,
+      timestamp: index + 1,
+      tool_call_id: null,
+      tool_calls: null,
+      tool_name: null,
+      finish_reason: null,
+      reasoning_content: null,
+    }))
+    getSessionDetailMock.mockReturnValue({ messages })
+    readConfigYamlForProfileMock.mockResolvedValue({
+      compression: { protect_last_n: 5 },
+    })
+    calcAndUpdateUsageMock.mockResolvedValue({ inputTokens: 120_000, outputTokens: 0 })
+    compressorCompressMock.mockResolvedValue({
+      messages: [{ role: 'user', content: 'compressed' }],
+      meta: {
+        compressed: true,
+        llmCompressed: true,
+        totalMessages: 9,
+        summaryTokenEstimate: 1,
+        verbatimCount: 0,
+        compressedStartIndex: 0,
+      },
+    })
+
+    const { buildCompressedHistory } = await import('../../packages/server/src/services/hermes/run-chat/compression')
+    await buildCompressedHistory(
+      'session-1',
+      'default',
+      'http://upstream',
+      undefined,
+      vi.fn(),
+      new Map(),
+    )
+
+    expect(compressorConstructorMock).toHaveBeenCalledWith({
+      config: {
+        triggerTokens: 100_000,
+        summaryBudget: 40_000,
+        headMessageCount: 3,
+        tailMessageCount: 5,
+      },
+    })
+    expect(compressorCompressMock).toHaveBeenCalledTimes(1)
+  })
+
+  it('uses stale snapshot summary plus safe tail instead of full history when under threshold', async () => {
+    const messages = Array.from({ length: 10 }, (_, index) => ({
+      id: index + 1,
+      session_id: 'session-1',
+      role: index === 9 ? 'user' : index % 2 === 0 ? 'user' : 'assistant',
+      content: `message ${index}`,
+      timestamp: index + 1,
+      tool_call_id: null,
+      tool_calls: null,
+      tool_name: null,
+      finish_reason: null,
+      reasoning_content: null,
+    }))
+    getSessionDetailMock.mockReturnValue({ messages })
+    getCompressionSnapshotMock.mockReturnValue({
+      summary: 'old summary',
+      lastMessageIndex: 99,
+      messageCountAtTime: 100,
+    })
+    readConfigYamlForProfileMock.mockResolvedValue({
+      compression: { protect_first_n: 2, protect_last_n: 3 },
+    })
+    estimateUsageTokensFromMessagesMock.mockReturnValue({ inputTokens: 1_000, outputTokens: 0 })
+
+    const { buildCompressedHistory } = await import('../../packages/server/src/services/hermes/run-chat/compression')
+    const history = await buildCompressedHistory(
+      'session-1',
+      'default',
+      'http://upstream',
+      undefined,
+      vi.fn(),
+      new Map(),
+    )
+
+    expect(history.map(m => m.content)).toEqual([
+      'message 0',
+      'message 1',
+      '[Previous context summary]\n\nold summary',
+      'message 6',
+      'message 7',
+      'message 8',
+    ])
+    expect(compressorCompressMock).not.toHaveBeenCalled()
+  })
+
+  it('compresses stale snapshot safe tail instead of full history when stale assembly exceeds threshold', async () => {
+    const messages = Array.from({ length: 10 }, (_, index) => ({
+      id: index + 1,
+      session_id: 'session-1',
+      role: index === 9 ? 'user' : index % 2 === 0 ? 'user' : 'assistant',
+      content: `message ${index}`,
+      timestamp: index + 1,
+      tool_call_id: null,
+      tool_calls: null,
+      tool_name: null,
+      finish_reason: null,
+      reasoning_content: null,
+    }))
+    getSessionDetailMock.mockReturnValue({ messages })
+    getCompressionSnapshotMock.mockReturnValue({
+      summary: 'old summary',
+      lastMessageIndex: 99,
+      messageCountAtTime: 100,
+    })
+    readConfigYamlForProfileMock.mockResolvedValue({
+      compression: { protect_first_n: 2, protect_last_n: 3 },
+    })
+    estimateUsageTokensFromMessagesMock.mockReturnValue({ inputTokens: 120_000, outputTokens: 0 })
+    compressorCompressMock.mockResolvedValue({
+      messages: [{ role: 'user', content: 'updated stale compressed' }],
+      meta: {
+        compressed: true,
+        llmCompressed: true,
+        totalMessages: 9,
+        summaryTokenEstimate: 1,
+        verbatimCount: 0,
+        compressedStartIndex: 8,
+      },
+    })
+
+    const { buildCompressedHistory } = await import('../../packages/server/src/services/hermes/run-chat/compression')
+    const history = await buildCompressedHistory(
+      'session-1',
+      'default',
+      'http://upstream',
+      undefined,
+      vi.fn(),
+      new Map(),
+    )
+
+    expect(history).toEqual([{ role: 'user', content: 'updated stale compressed' }])
+    expect(compressorCompressMock).toHaveBeenCalledWith(
+      expect.arrayContaining([{ role: 'user', content: 'message 0' }]),
+      'http://upstream',
+      undefined,
+      'session-1',
+      expect.objectContaining({ profile: 'default' }),
+    )
+  })
+
+  it('does not compress when compression is disabled', async () => {
+    const messages = Array.from({ length: 10 }, (_, index) => ({
+      id: index + 1,
+      session_id: 'session-1',
+      role: index === 9 ? 'user' : index % 2 === 0 ? 'user' : 'assistant',
+      content: `message ${index}`,
+      timestamp: index + 1,
+      tool_call_id: null,
+      tool_calls: null,
+      tool_name: null,
+      finish_reason: null,
+      reasoning_content: null,
+    }))
+    getSessionDetailMock.mockReturnValue({ messages })
+    readConfigYamlForProfileMock.mockResolvedValue({
+      compression: { enabled: false, threshold: 0.01 },
+    })
+    calcAndUpdateUsageMock.mockResolvedValue({ inputTokens: 180_000, outputTokens: 0 })
+
+    const { buildCompressedHistory } = await import('../../packages/server/src/services/hermes/run-chat/compression')
+    const history = await buildCompressedHistory(
+      'session-1',
+      'default',
+      'http://upstream',
+      undefined,
+      vi.fn(),
+      new Map(),
+    )
+
+    expect(history).toHaveLength(9)
+    expect(compressorCompressMock).not.toHaveBeenCalled()
+    expect(calcAndUpdateUsageMock).not.toHaveBeenCalled()
+  })
+})

tests/server/sessions-controller.test.ts

@@ -206,6 +206,81 @@ describe('session conversations controller', () => {
     expect(ctx.body).toEqual({ error: 'Conversation not found' })
   })
 
+  it('prefers local session detail for Hermes history detail when available', async () => {
+    localGetSessionDetailMock.mockReturnValue({
+      id: 'cli-1',
+      source: 'cli',
+      title: 'Local complete',
+      messages: [
+        { id: 1, session_id: 'cli-1', role: 'user', content: 'local full message', timestamp: 1 },
+      ],
+    })
+    getSessionDetailFromDbMock.mockResolvedValue({
+      id: 'cli-1',
+      source: 'cli',
+      title: 'Hermes incomplete',
+      messages: [],
+    })
+
+    const mod = await import('../../packages/server/src/controllers/hermes/sessions')
+    const ctx: any = { params: { id: 'cli-1' }, body: null }
+    await mod.getHermesSession(ctx)
+
+    expect(localGetSessionDetailMock).toHaveBeenCalledWith('cli-1')
+    expect(getSessionDetailFromDbMock).not.toHaveBeenCalled()
+    expect(getSessionMock).not.toHaveBeenCalled()
+    expect(ctx.body.session).toMatchObject({
+      id: 'cli-1',
+      title: 'Local complete',
+      messages: [{ content: 'local full message' }],
+    })
+  })
+
+  it('falls back to Hermes state.db when local history detail is missing', async () => {
+    localGetSessionDetailMock.mockReturnValue(null)
+    getSessionDetailFromDbMock.mockResolvedValue({
+      id: 'hermes-1',
+      source: 'cli',
+      title: 'Hermes detail',
+      messages: [
+        { id: 1, session_id: 'hermes-1', role: 'user', content: 'from hermes', timestamp: 1 },
+      ],
+    })
+
+    const mod = await import('../../packages/server/src/controllers/hermes/sessions')
+    const ctx: any = { params: { id: 'hermes-1' }, body: null }
+    await mod.getHermesSession(ctx)
+
+    expect(localGetSessionDetailMock).toHaveBeenCalledWith('hermes-1')
+    expect(getSessionDetailFromDbMock).toHaveBeenCalledWith('hermes-1')
+    expect(getSessionMock).not.toHaveBeenCalled()
+    expect(ctx.body.session).toMatchObject({
+      id: 'hermes-1',
+      title: 'Hermes detail',
+      messages: [{ content: 'from hermes' }],
+    })
+  })
+
+  it('does not return api_server sessions from the Hermes history detail endpoint', async () => {
+    localGetSessionDetailMock.mockReturnValue({
+      id: 'api-1',
+      source: 'api_server',
+      title: 'API Server',
+      messages: [{ id: 1, session_id: 'api-1', role: 'user', content: 'local api', timestamp: 1 }],
+    })
+    getSessionDetailFromDbMock.mockResolvedValue(null)
+    getSessionMock.mockResolvedValue(null)
+
+    const mod = await import('../../packages/server/src/controllers/hermes/sessions')
+    const ctx: any = { params: { id: 'api-1' }, body: null }
+    await mod.getHermesSession(ctx)
+
+    expect(localGetSessionDetailMock).toHaveBeenCalledWith('api-1')
+    expect(getSessionDetailFromDbMock).toHaveBeenCalledWith('api-1')
+    expect(ctx.status).toBe(404)
+    expect(ctx.body).toEqual({ error: 'Session not found' })
+  })
+
   it('returns native state.db usage analytics for the requested period', async () => {
     const today = new Date().toISOString().slice(0, 10)
     getLocalUsageStatsMock.mockReturnValue({