feat: group chat session lifecycle, typing recovery, mention highlighting (#186)

* feat: restore group chat system with Socket.IO and SQLite persistence

- GroupChatServer: Socket.IO server with room management, message history, typing indicators
- SQLite storage for rooms, messages, and agent configuration
- AgentClients: manages AI agent connections via socket.io-client, forwards @mentions to Hermes gateway
- REST API: room CRUD, agent management, invite codes
- Agent auto-restoration on server restart
- Tests for all REST endpoints

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* docs: add context-engine design document for group chat compression

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: handle special-character session search

* fix: keep unicode dotted session search on quoted FTS path

* feat: add context engine and group chat frontend UI

- Context engine: three-zone compression (head/tail/summary) with LLM
  summarization, incremental updates, TTL cache, and graceful degradation
- Frontend: group chat page with Socket.IO client, room sidebar, message
  list, agent/member display, create/join-by-code modals
- Integration: wire context engine into agent-clients before /v1/runs
- Refactor ChatStorage to use global DB (getDb/ensureTable) with gc_ prefix
- Add i18n keys for group chat to all 8 locales
- Add sidebar nav entry and router for group chat page

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: remove leftover main branch code from merge conflict resolution

The `isNumericQuery`, `hasUnsafeChars`, and `runLikeContentSearch` functions
no longer exist — they were replaced by HEAD's `shouldUseLiteralContentSearch`
and `runLiteralContentSearch`. This dead code block caused a TypeScript
compile error after the merge.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: install missing socket.io dep and type ack params

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: enable WebSocket proxy and fix socket.io transport for group chat

- Add ws: true to Vite proxy config so WebSocket upgrade requests
  are forwarded to the backend
- Allow both polling and websocket transports on server and client
  (polling as fallback when WebSocket upgrade fails through proxy)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix: separate socket.io path from REST routes for group chat

socket.io was mounted at /api/hermes/group-chat which intercepted all
REST requests to /api/hermes/group-chat/rooms etc, returning
"Transport unknown". Changed socket.io path to /api/hermes/group-chat/ws
to avoid conflicts.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* feat: improve group chat UI, agent management, and socket.io reliability

- Redesign GroupChatPanel with Naive UI, stacked agent avatars, and popover management
- Match GroupChatInput style with single chat input, add IME composition handling
- Add agent add/remove per room with profile selection and duplicate prevention
- Use @multiavatar for SVG avatar generation with caching
- Decouple joinRoom from socket.io, use REST API for data loading
- Switch socket.io to default path with /group-chat namespace to avoid proxy conflicts
- Restore agent connections after server is listening
- Add getRoomDetail REST endpoint and duplicate agent prevention (409)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* feat: server-side @mention routing with context compression status and queue

- Move @mention detection from agent socket listeners to server-side processMentions()
- Add per-room processing lock to block mention dispatch during compression
- Queue mentions during processing, drain only the latest when ready
- Emit context_status events (compressing/replying/ready) to room via Socket.IO
- Frontend displays compression status indicator above input
- Token-based compression trigger (100k threshold) with CJK-aware estimation
- Fix compressor type errors (countTokens parameter type)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* feat: improve group chat profile handling and session sync

Refine group chat room/session behavior with per-room compression controls, sidebar updates, and better stale session cleanup so multi-profile group chat state stays consistent.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* feat: group chat improvements — session lifecycle, typing recovery, mention highlighting

- Fix cross-profile session deletion with deferred delete queue
- Move saveSessionProfile to after gateway response confirmation
- Replace all console.log with logger in group-chat modules
- Add server-side typing/context_status state tracking for room rejoin
- Fix @ mention popup position to follow cursor
- Add @ mention highlighting (blue) in chat message content
- Fix mention regex to match all occurrences after HTML tags
- Enable esbuild minify and treeShaking
- Move @multiavatar/multiavatar to devDependencies
- Add i18n keys for group chat features
- Update tests for new functionality

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* chore: bump version to 0.4.5 and move @multiavatar to devDependencies

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Zhicheng Han <zhicheng.han@mathematik.uni-goettingen.de>

packages/server/src/services/hermes/context-engine/prompt.ts

@@ -0,0 +1,82 @@
+// ─── Agent Identity Instructions ────────────────────────────
+
+import type { MemberInfo } from './types'
+
+interface AgentInstructionsParams {
+    agentName: string
+    roomName: string
+    agentDescription: string
+    memberNames: string[]
+    members: MemberInfo[]
+}
+
+export function buildAgentInstructions(params: AgentInstructionsParams): string {
+    let memberSection: string
+    if (params.members.length > 0) {
+        memberSection = params.members
+            .map(m => m.description ? `- ${m.name}: ${m.description}` : `- ${m.name}`)
+            .join('\n')
+    } else if (params.memberNames.length > 0) {
+        memberSection = params.memberNames.map(n => `- ${n}`).join('\n')
+    } else {
+        memberSection = '- 未知'
+    }
+
+    return `你是"${params.agentName}"，群聊房间"${params.roomName}"中的 AI 助手。
+
+你的角色：${params.agentDescription}
+
+当前房间成员：
+${memberSection}
+
+规则：
+- 有人用 @${params.agentName} 提及你时才需要回复，重点回应提及你的人。
+- 回答简洁、对群聊有帮助。
+- 不要假装是人类，需要时明确表明自己是 AI。
+- 对话历史中包含多个人的消息，每条消息前标有发送者名字。
+- 对话开头可能包含之前的对话摘要，用于提供更早的上下文。
+- 回复最新一条提及你的消息。
+- 如果需要其他 agent 协作或明确回复某个人，使用 @名字 来提及对方。
+- 自行判断对话是否已经结束——如果问题已解决、达成共识、或对方只是陈述不需要回复，则不要再 @任何人，直接结束回复，避免产生无意义的循环对话。`
+}
+
+// ─── Summarization Prompts ─────────────────────────────────
+
+export function buildSummarizationSystemPrompt(): string {
+    return `你是一个群聊对话的摘要助手。请创建一份结构化摘要，帮助 AI 助手快速理解完整的对话上下文并智能回复。
+
+使用以下格式：
+
+当前话题：
+- 现在在聊什么，目标是什么
+
+已知结论：
+- 已达成哪些共识，哪些问题已经回答过
+
+待回复消息：
+- 还剩谁的问题没回，下一步要做什么
+
+关键人物：
+- 人名、角色、引用关系
+
+重要上下文：
+- 不要丢时间线和立场变化
+- 少写废话，多保留"可行动信息"
+- 重点保留：谁说了什么、结论是什么、下一步是什么
+- 关键的 URL、代码片段、错误信息、约束条件
+
+规则：
+- 基于事实，不要编造信息。
+- 保持简洁（500 字以内）。
+- 聚焦于帮助 AI 回复下一条消息的可行动信息。
+- 使用与对话相同的语言。
+- 不要回复对话内容，只输出摘要。`
+}
+
+export function buildFullSummaryPrompt(): string {
+    return '请对上方对话创建一份简洁的摘要。只输出摘要内容。'
+}
+
+export function buildIncrementalUpdatePrompt(): string {
+    return '对话自上次摘要后有了新的内容。请更新摘要，整合新消息。保持相同格式，更新所有部分。只输出更新后的摘要。'
+}