dsh-honcho-sync — Honcho Memory Plugin for DeepSeek Harness
Give your DeepSeek Harness AI persistent memory.
Auto-sync every conversation turn to a self-hosted Honcho service, and equip the AI with 25 tools covering the full official Honcho v3 API surface plus extended endpoints. Configure via environment variables, settings panel, or cordis.patch.yml.
中文
这是什么？
DeepSeek Harness (DSH) ↔ Honcho 记忆桥接插件。 安装后 DSH 自动将每轮对话同步到用户自建的 Honcho 记忆服务（NAS、服务器或云端均可），AI 获得 25 个内置工具，覆盖官方 Honcho API 全量端点 + 扩展端点：检索、推理、画像、会话管理、消息收发、结论、Dream、队列监控。
功能
| 功能 | 说明 |
|---|
| 自动同步 | 每轮对话自动推送到 Honcho（debounce 3s），watermark 去重 |
| 持久化状态 | ~/.dsh/honcho-sync-state.json 保存同步进度，DSH 重启不丢失 cursor |
| Layer-1 上下文注入 | 会话首条消息时自动加载 peer card + representation + session summary |
| 语义去重 | representation 注入前做 IP/关键词 Jaccard / 3-gram 短语重叠检测，过滤与 peer card 重复的事实 |
| 25 个记忆工具 | 覆盖官方 API + 扩展端点：search / dialectic / profile / session / peer / message / conclude / dream / queue |
工具清单（25 个）
| 分类 | 工具 | 说明 |
|---|
| 记忆 | honcho_recall | 7 天窗口语义搜索（快，2-5s） |
| honcho_ask | 辩证推理问答（慢，2-5min） |
| honcho_remember | 保存事实到持久记忆 |
| honcho_context | 读取 session 完整上下文 |
| 搜索 | honcho_search | 跨所有 session 搜索（无日期限制） |
| honcho_session_search | 在指定 session 内搜索 |
| 画像 | honcho_profile | 获取 peer card（身份/属性/关系/指令） |
| honcho_representation | 获取 working representation（跨会话观察） |
| 会话 | honcho_session_list | 列出 workspace 所有 session |
| honcho_session_create | 创建/获取 session |
| honcho_session_clone | 克隆 session |
| honcho_session_peers | 查看 session 关联的 peers |
| honcho_session_summaries | 获取 session 摘要 |
| honcho_session_context | 获取 session 上下文（含 summary） |
| Peer | honcho_peer_list | 列出所有 peers |
| honcho_peer_create | 创建/获取 peer |
| 消息 | honcho_message_send | 向指定 session 发送消息（扩展） |
| honcho_message_get | 按 message_id 读取单条消息（扩展） |
| 结论 | honcho_conclude | 创建显式结论（高置信度事实） |
| honcho_conclude_list | 列出结论 |
| honcho_conclude_query | 语义搜索结论 |
| 管理 | honcho_status | 服务器健康检查 |
| honcho_dream | 触发离线处理（card_refresh / omni） |
| honcho_queue | 查看处理队列状态 |
自有特色功能说明
1. 语义去重链 (Semantic Deduplication)
Layer-1 注入时，对 Honcho representation 做三级过滤：
- 时间块拆分 — 按
[YYYY-MM-DD HH:MM:SS] 前缀拆分为独立块
- Peer Card 重叠检测 — 提取 peer card 中的 IDENTITY/ATTRIBUTE/RELATIONSHIP/INSTRUCTION 事实，与每块内容做语义重叠判断
- 内容哈希去重 — 基于块内容的 60-char 前缀做
Map 去重
_semanticOverlap(a, b) 判断逻辑：
- IP 重叠：双方都包含相同
192.168.x.x IP → 判定重复
- 关键词 Jaccard — 分词后计算 Jaccard 相似度 > 0.3 → 判定重复
- 3-gram 短语重叠 — a 中任意连续 3 词出现在 b 中 → 判定重复
过滤后的 representation 按时间块保留最多 N 条（默认 8 条），并将 UTC 时间转换为 CST (+8h)。
2. 持久化状态 (Persistence)
同步进度写入 ~/.dsh/honcho-sync-state.json，包含每个 Honcho session 的 lastSyncedEventCount。DSH 重启、HMR 热更新后自动恢复 sync cursor，不会重复推送已同步的消息。
3. 动态配置读取 (Dynamic Config)
每次工具调用时通过 resolveConfig() 从 _resolvedSettings 动态读取配置，settings 面板修改后立即生效，无需重启 DSH。
4. 额外 Honcho API 工具
| 工具 | 端点 | 说明 |
|---|
honcho_message_send | POST /sessions/{id}/messages | 向指定 session 发送消息 |
honcho_message_get | GET /sessions/{id}/messages/{msg_id} | 按 ID 读取单条消息 |
honcho_session_context | GET /sessions/{id}/context | 获取 session 上下文（含 summary） |
honcho_peer_list | POST /peers/list | 列出所有 peers |
这些端点在 Honcho v3 API 中可用但未被上游 v0.6.0 覆盖。
安装
前提： 已部署 Honcho 记忆服务
方式 1：从 GitHub 直接安装（推荐）
dsh plugin --profile web add link:https://github.com/nanpaidashi/dsh-honcho-sync dsh web
方式 2：本地手动安装
# 1. 克隆仓库
git clone https://github.com/nanpaidashi/dsh-honcho-sync.git
cd dsh-honcho-sync
npm run build # cp src/index.mjs dist/index.js
# 2. 复制文件到 DSH profile
cp dist/index.js ~/.dsh/profiles/web/honcho-sync.mjs
# 3. 在 ~/.dsh/profiles/web/cordis.patch.yml 中添加：
# - insert:
# - id: honcho-sync
# name: './honcho-sync.mjs'
# 4. 重启 DSH
systemctl --user restart dsh-web
配置
必须配置 HONCHO_URL 和 HONCHO_WORKSPACE，二选一：
方式 1：环境变量（推荐）
export HONCHO_URL="http://your-honcho-server:8000" # 必填 — Honcho API 地址
export HONCHO_WORKSPACE="hermes" # 必填 — Honcho workspace 名称
export HONCHO_USER_PEER="user" # 可选 — 默认 "user"
export HONCHO_AGENT_PEER="agent" # 可选 — 默认 "agent"
然后在 ~/.dsh/profiles/web/cordis.patch.yml 中引用：
- insert:
- id: honcho-sync
name: './honcho-sync.mjs'
config:
honchoUrl: __FROM_ENV__
workspace: ''
userPeer: user
agentPeer: agent
debounceMs: 3000
autoRecall: true
recallBudget: 2000
autoSync: true
messageMaxChars: 25000
injectionMaxChars: 4000
其中 honchoUrl: __FROM_ENV__ 表示从 HONCHO_URL 环境变量读取。
方式 2：直接写在 cordis.patch.yml
- insert:
- id: honcho-sync
name: './honcho-sync.mjs'
config:
honchoUrl: "http://192.168.0.4:8000"
workspace: "hermes"
userPeer: "user"
agentPeer: "agent"
debounceMs: 3000
autoRecall: true
recallBudget: 2000
autoSync: true
messageMaxChars: 25000
injectionMaxChars: 4000
方式 3：通过 DSH settings 服务（运行时配置）
插件通过 DSH settings 服务注册了 honcho-memory 命名空间。如果 DSH 版本支持 settings UI，可以在设置界面直接修改：
- honchoUrl — Honcho API 地址
- workspace — Workspace 名称
- userPeer / agentPeer — Peer 标识符
- debounceMs — 同步防抖延迟（100ms+，默认 3000）
- autoRecall — 是否自动 recall（默认 true）
- recallBudget — Recall token 预算（默认 2000）
- autoSync — 是否自动同步（默认 true）
- messageMaxChars — 同步消息最大字符数（默认 25000）
- injectionMaxChars — 首条注入最大字符数（默认 4000）
- reprMaxObs — representation 保留最大观察数（默认 8）
- reprTimeoutMs — representation 请求超时（默认 8000ms）
- cardTimeoutMs — peer card 请求超时（默认 5000ms）
架构
DSH Session
│
├─ session/event ──→ 自动同步（debounce + watermark）──→ Honcho API (POST messages)
│ ↓
│ ~/.dsh/honcho-sync-state.json 持久化
│
├─ 25 个 honcho_* 工具 ──→ AI 可调用的全量 API 封装
│
└─ 首条 user message ──→ Layer-1 注入（语义去重后的 card + representation + summary）
同步策略
- Per-day session ID：
dsh-<cwd>-YYYY-MM-DD，每天一个 Honcho session，保持检索空间小
- Watermark 去重：记录已同步的 event count，只推送增量
- Debounce 3s：避免高频写入
- 持久化状态：
~/.dsh/honcho-sync-state.json 保存 sync cursor，DSH 重启不丢失
参数调优建议
| 参数 | 默认值 | 说明 |
|---|
debounceMs | 3000 | 对话结束后等待多久才推送 Honcho。多轮快速对话可适当增大 |
messageMaxChars | 25000 | 单次同步的最大消息字符数。对话很长时调大，节省 Honcho token 时调小 |
recallBudget | 2000 | Layer-1 注入时 recall 的 token 预算。peer card 通常 500-1500 tokens |
injectionMaxChars | 4000 | Layer-1 注入的最大字符数。包含 peer card + representation + session summary |
reprMaxObs | 8 | representation 注入时保留的最大观察数（语义去重后） |
reprTimeoutMs | 8000 | representation POST 请求超时（毫秒） |
cardTimeoutMs | 5000 | peer card GET 请求超时（毫秒） |
依赖
- Honcho — 需自行部署
- DeepSeek Harness ≥ 0.1.0-rc.6
许可证
MIT
English
What is this?
A bridge plugin connecting DeepSeek Harness to Honcho memory. After installation, DSH auto-syncs every conversation turn to a self-hosted Honcho service (NAS, server, or cloud), and the AI gains 25 built-in tools covering the full official Honcho v3 API surface plus extended endpoints: search, dialectic reasoning, peer profiles, session management, message send/retrieve, conclusions, dream scheduling, and queue monitoring.
Tools (25)
| Category | Tool | Description |
|---|
| Memory | honcho_recall | 7-day window semantic search (fast, 2-5s) |
| honcho_ask | Dialectic reasoning Q&A (slow, 2-5min) |
| honcho_remember | Save facts to persistent memory |
| honcho_context | Read full session context |
| Search | honcho_search | Cross-all-sessions search (no date filter) |
| honcho_session_search | Search within a specific session |
| Profile | honcho_profile | Get peer card (identity/attributes/relationships) |
| honcho_representation | Get working representation (cross-session observations) |
| Session | honcho_session_list | List all sessions in workspace |
| honcho_session_create | Create/get a session |
| honcho_session_clone | Clone a session |
| honcho_session_peers | List session peers |
| honcho_session_summaries | Get session summaries |
| honcho_session_context | Get session context with summary |
| Peer | honcho_peer_list | List all peers |
| honcho_peer_create | Create/get a peer |
| Message | honcho_message_send | Send message to a session (extended) |
| honcho_message_get | Get single message by ID (extended) |
| Conclude | honcho_conclude | Create explicit conclusions |
| honcho_conclude_list | List conclusions |
Unique Features
1. Semantic Deduplication
Before injecting Honcho representation into DSH context, the plugin applies a 3-stage filter:
- Time-block splitting — splits representation by
[YYYY-MM-DD HH:MM:SS] prefixes
- Peer card overlap detection — extracts IDENTITY/ATTRIBUTE/RELATIONSHIP/INSTRUCTION facts from peer cards and checks for semantic overlap with each block
- Content-based deduplication — deduplicates blocks sharing the same 60-char content prefix
_semanticOverlap(a, b) checks:
- IP overlap — both contain the same
192.168.x.x IP → duplicate
- Keyword Jaccard similarity — tokenized words Jaccard > 0.3 → duplicate
- 3-gram phrase overlap — any consecutive 3-word sequence from a found in b → duplicate
Result: representation is trimmed to max N observations (default 8), with UTC times converted to CST (+8h).
2. Persistent State
Sync progress saved to ~/.dsh/honcho-sync-state.json with lastSyncedEventCount per Honcho session. Survives DSH restarts and HMR — no duplicate pushes.
3. Dynamic Config Resolution
Each tool call reads config via resolveConfig() from _resolvedSettings. Settings panel changes take effect immediately without restarting DSH.
4. Extended Honcho API Tools
| Tool | Endpoint | Description |
|---|
honcho_message_send | POST /sessions/{id}/messages | Send a message to a session |
honcho_message_get | GET /sessions/{id}/messages/{msg_id} | Get single message by ID |
honcho_session_context | GET /sessions/{id}/context | Get session context with summary |
honcho_peer_list | POST /peers/list | List all peers |
Installation
Method 1: Direct from GitHub (Recommended)
dsh plugin --profile web add link:https://github.com/nanpaidashi/dsh-honcho-sync dsh web
Method 2: Manual local installation
git clone https://github.com/nanpaidashi/dsh-honcho-sync.git
cd dsh-honcho-sync
npm run build
cp dist/index.js ~/.dsh/profiles/web/honcho-sync.mjs
# Add to ~/.dsh/profiles/web/cordis.patch.yml:
# - insert:
# - id: honcho-sync
# name: './honcho-sync.mjs'
systemctl --user restart dsh-web
Configuration
HONCHO_URL and HONCHO_WORKSPACE are required — configure via one of:
Method 1: Environment Variables (Recommended)
export HONCHO_URL="http://your-honcho-server:8000" # Required
export HONCHO_WORKSPACE="hermes" # Required
export HONCHO_USER_PEER="user" # Optional, default "user"
export HONCHO_AGENT_PEER="agent" # Optional, default "agent"
Then in ~/.dsh/profiles/web/cordis.patch.yml:
- insert:
- id: honcho-sync
name: './honcho-sync.mjs'
config:
honchoUrl: __FROM_ENV__
workspace: ''
debounceMs: 3000
autoRecall: true
recallBudget: 2000
autoSync: true
messageMaxChars: 25000
injectionMaxChars: 4000
honchoUrl: __FROM_ENV__ reads from the HONCHO_URL environment variable.
Method 2: Direct in cordis.patch.yml
- insert:
- id: honcho-sync
name: './honcho-sync.mjs'
config:
honchoUrl: "http://192.168.0.4:8000"
workspace: "hermes"
Method 3: DSH Settings Service (Runtime)
The plugin registers a honcho-memory namespace via the DSH settings service. If your DSH version supports a settings UI, you can modify all parameters at runtime without restarting.
Architecture
DSH Session
│
├─ session/event ──→ Auto-sync (debounce + watermark) ──→ Honcho API
│ ↓
│ ~/.dsh/honcho-sync-state.json persistence
│
├─ 25 honcho_* tools ──→ Full API surface callable by AI
│
└─ First user message ──→ Layer-1 injection (deduplicated card + representation + summary)
Sync Strategy
- Per-day session ID:
dsh-<cwd>-YYYY-MM-DD, one Honcho session per day
- Watermark deduplication: tracks synced event count, only pushes deltas
- Debounce 3s: avoids high-frequency writes
- Persistent state:
~/.dsh/honcho-sync-state.json saves sync cursor across restarts
Parameter Tuning
| Parameter | Default | Description |
|---|
debounceMs | 3000 | Delay before pushing to Honcho after conversation ends |
messageMaxChars | 25000 | Max characters per sync push |
recallBudget | 2000 | Token budget for Layer-1 context recall |
injectionMaxChars | 4000 | Max chars for Layer-1 injection (card + representation + summary) |
reprMaxObs | 8 | Max observations kept after semantic deduplication |
reprTimeoutMs | 8000 | Representation POST timeout (ms) |
cardTimeoutMs | 5000 | Peer card GET timeout (ms) |
Dependencies
- Honcho — Self-hosted memory service
- DeepSeek Harness ≥ 0.1.0-rc.6
License
MIT
