メモリ・ナレッジ · DeepSeek Harness プラグイン — 35 ページ

会话清理：在运行中的 web 运行时里彻底删除 DSH 会话——实时 store 条目、工作区记录、磁盘产物与投影缓存行一并清掉，并在会话行菜单里加一个删除入口。

DSH 技能管理器：诊断、新建、编辑、开关、搬运磁盘上的技能（SKILL.md），补上官方 provider 静默跳过坏文件的那半边，并给模型一个 skill_manager 工具

Superpowers（移植自 obra/superpowers）：把整套软件开发方法论技能与会话引导语作为 dsh 插件提供，技能在运行时注册，首轮对话自动注入引导。

Local SAG personal knowledge base plugin for DeepSeek Harness

Global knowledge base for the dsh web GUI: upload PDF/Word/Excel/PPT/TXT/MD/CSV/HTML documents, pure-JS text extraction + BM25 keyword index persisted to disk, a right-edge management panel, automatic retrieval injected into every session (agent/pre-step) plus a knowledge_base_search model tool, and source citations. Zero native/vector dependencies; embedding backend can be layered on later.

Installable Ontology Kernel for DeepSeek Harness with durable local storage and typed browser remotes.

Host-only DSH foundation helpers: default workspace, resource inspection, and CLIProxyAPI catalog bridging.

Cross-tool session migration plugin for the DeepSeek Harness (dsh): high-fidelity ingest, continue, knowledge extraction, bidirectional export, privacy redaction.

Context Assembler DSH V0.99 — Context Assembler plugin for DeepSeek Harness (dsh): context compaction, cache-friendly topic-block management, water-pressure topic splitting, tool trace/rewrite, handoff planning and reality recall injection.

Spend the context window like a budget on long tasks: load files by graduated reads (head first, sections on demand, targeted search before full reads), store discovered facts to a notes file the moment they are established so nothing is re-derived, and re-state the running task plan at checkpoints — the difference between finishing a long task and drowning at 80%.

What will your agent's context cost before it runs? Audit instruction files, skills, MCP tool schemas and memory per source — with duplicates, always-on versus on-demand costs, and a budget you can enforce in CI.

DeerFlow 2.0 (SuperAgent Harness) as a DeepSeek Harness (dsh) plugin — a Lead Agent that decomposes tasks into sub-tasks and dispatches them to pluggable skills (research, data-analysis, slide, web, image, video) with web search/crawl, sandboxed Python, reliability middleware, and cross-session memory, packaged per the deepseek-harness plugin spec.

Doco knowledge-base native plugin for DeepSeek Harness (dsh) — search, read, outline, spreadsheet cells and save-draft your Doco KB with block-level addressing. Provides the `doco` service consumed by doco-memory-dsh.

存档点管理器：最后的保活机制 + 试错回滚工具。定时/事件/主人指令创建健康存档点（记忆库+灵魂+校验和），验证器确保存档可启动，出事后一键恢复最近健康点。

DSH web plugin: multi-agent commander. Open Claude / opencode / codex agents with role & skill definitions, watch every agent's live terminal in a right-side radar panel, and let DeepSeek dispatch tasks through agent_* tools. Shared memory: 记忆库.md / 任务看板.md / 交接区 in the project directory.

多 Agent 管理与配置：新建 Agent 预设、编辑人格 / 知识库注入 / 知识库工具开关 / MCP 按 Agent 收窄，并在 Agent 预设卡片上提供编辑入口

DSH 长期记忆插件 + 配套 CLI：全局偏好（跨对话）与对话记忆（仅本会话）双层注入、实时采集、只读数据浏览与会话内面板。零运行时依赖、能力探测与熔断，面向跨版本升级设计。

Claw persistent memory: vault, daily notes, review, flush. 完整留存：金库、日记、回顾、压缩前冲洗。

每日反思插件：固定时间（默认凌晨 12 点）向爱丽丝发反思提醒，结合当天记忆按 6 维进化棱镜自审。信号送达，反思归爱丽丝。

Agent terminal for the DSH web GUI: sidebar entry + center panel that runs real external agent CLIs (claude code, hermes, ...) interactively inside a local PTY, rendered with xterm.js. 100% fidelity — the page IS the agent's own TUI. Sessions survive tab switches / panel close (host-side scrollback replay on re-attach), loopback-fenced REST + WebSocket routes.

DSH ↔ agentmemory session-memory bridge (backend-only). Mirrors session lifecycle into agentmemory with standard hookTypes (compression-friendly, dedup-safe), exposes memory_recall / memory_remember tools, and injects memory context via agent/pre-step. Configuration comes from the cordis.yml row.

Agent-plane tool plugin: connects dsh to the agentmemory REST API (search / remember / forget / health)

Closes ai-memory's consolidation loop for DSH: when a DSH session ends, calls ai-memory finalize-session so its captured observations become searchable wiki pages the other harnesses ai-memory serves can retrieve. Skips subagents, queues concurrent session ends, and sweeps sessions abandoned by a forced restart.

Lightweight arrow-key history recall for DSH.