Memory and knowledge · DeepSeek Harness plugins — Page 13

Claude Code-compatible output style selection for the DeepSeek Harness

Load Claude Code plugin.json manifests as in-memory dsh plugin mounts

Manage Claude Code plugin state files (marketplaces, installs, per-scope settings) in byte-shape parity with the real CLI

Structured progress state + post-compact context rebuild: derives a typed per-session progress brief from session events (goal fold, todo snapshot, verified bash receipts, last user message) and re-injects it after a successful compaction/end; default off

Serena-first prompt steering on the CC preset: a registered policy section plus replace-not-mutate appends to the upstream tool:read / tool:grep sections while a serena MCP server is ready

CC host-plane first-prompt session-title provider on the haiku cheap lane

Cache-stable tool ordering on the system-prompt/assemble waterfall: each scope's emitted tool sequence keeps its positions and new tools append lexicographically at the tail

CC-style manage_skill tool: create/update/delete/list learned skills under $DSH_HOME/learned-skills via the shared LearnedSkillStore, gated by the cc-learn.enabled setting

NotebookEdit model tool that edits Jupyter notebook (.ipynb) cells over the ctx.fs seam, aligned with Claude Code's NotebookEditTool replace/insert/delete, cell addressing, and read-before-write semantics

Deferred tool registration and a ToolSearch model tool that loads heavy tools on demand

Sleep model tool over a cooperative AbortSignal wait: align with Claude Code's SleepTool duration, interrupt-cancel, and concurrency semantics

StructuredOutput tool that validates model input against a caller-supplied JSON schema and returns it verbatim, aligned with Claude Code's SyntheticOutputTool

Tool Use Summary (TUS) pipeline: fire-and-forget cheap-lane digests of large tool results, stored in an append-only per-session ledger and consumed at compaction time

CC-style web_fetch replacement with an optional prompt: summarizes the fetched page on the cheap lane (haiku alias) when configured; passing a prompt without haiku fails before fetching

CC-parity dynamic workflow tool: inline meta extraction, saved-source resolution, async launch with a session-scoped run registry and exactly-once completion delivery

Tool registry and execution pipeline for the DeepSeek Harness

Local-first personal memory and knowledge management for DeepSeek Harness

agentmemory for DeepSeek Harness (dsh): full memory_* tools, capture hooks, and context injection over the local REST server

Cross-session plaintext memory plugin for the DeepSeek Harness — deterministic BM25 recall, writes take effect immediately, signal-driven injection, no embeddings

DeepSeek Harness bundle that connects Agentero to dsh

Enterprise workbench capabilities for the DeepSeek Harness — AI employees, knowledge bases, skills, MCP servers and plugins, all managed from local state

把 Codex（OpenAI Codex CLI / Desktop）的 MCP、技能、全局配置、记忆以适配 DSH 的形式迁移进 DeepSeek Harness（DSH 插件）

Adaptive Context Plane core bundle: Evidence Ledger + Governance + Context Composer + User Model (explicit)

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.