Plugins

会话净化：扫描本地会话记录，剥掉被模型抄进回复里的输入信封（<user_message>/<turn_meta>），备份后重写，可一键还原。

Context-budgeted local memory for DeepSeek Harness: fixed ~600-token cache-friendly injection (session/daily snapshots), three-stage sleep consolidation, spaced repetition, temporal validity, subagent-driven L0 mental-model rewrite, typed memories (fact/s

Read-only behavior audit for DSH sessions: flags unsafe actions in 8 categories (A–H) with rules for 14 subclasses (the 41-subclass taxonomy file is not shipped, so a built-in 14-subclass fallback is used), anchors every finding to a real event seq, and gives deep links plus the full tool-call / data-access chain of any assistant reply (parameters, results, files read and written, domains, sandbox and approval context). Every finding is drill-downable.

Extra folders registered for a workspace become real workspace roots: the session runtime context lists them, a workspace_roots tool reports each root top-level entries, git branch and AGENTS.md/README/manifests, and file tools read, write, patch, move and delete inside them without approval prompts. Folders are edited per workspace from a sidebar hover action or from Settings. On Windows a confined shell stays inside the workspace, because ACL confinement cannot take extra roots.

Per-turn git checkpoints of a workspace: revert one turn’s changes (or go back to the end of an earlier turn), previewed file by file, and forget that part of the conversation from the model context.

Configure automation Hooks for user and project scopes to run commands on session, tool, permission, and context-compaction events. Manage Hook authorization manually in Hooks setting.

Standalone DeepSeek Harness port of Google’s Elephant-Goldfish Model with seven slash commands (/egm-*) for design, documentation, and fresh-context reviews using native DSH subagents. Requires no other plugins, but it registers the same /egm-* commands as other EGM ports, so install only one EGM plugin per profile.

Task-pack builder for DeepSeek Harness: compress large files into one brief (default budget 8000 tokens, estimated) that a subagent can read alone.

Local Markdown notes panel: import any local folder, browse its Markdown tree in the sidebar, edit files with auto-save, and search across every imported folder. Shows read-only Git context for each file — status, diff against HEAD with per-line change bars, and line blame — and exposes six notes_* tools so the agent works on the same folders. Files stay where they are on disk; the plugin never commits, stages or checks out.

Drives the Edge already open on your machine through a local WebSocket bridge and an unpacked MV3 extension: the agent reads the page as Markdown, dumps its interactive elements with stable refs, clicks and types with pre-action assertions and a three-valued effect state, and can capture pages from the extension popup as self-contained HTML snapshots, filed into a kb.py literature knowledge base when one is configured (kbRoot or DSH_KB_ROOT). Seven browser tools stay unregistered until edge_bridge is started, so an idle session pays no context cost.

Five Brave Search tools — web/news/image/video results, token-budgeted page context, cited AI answers, local places, and autocomplete/spellcheck — plus a brave search provider for the built-in web_search tool. Needs a Brave Search API key (BRAVE_SEARCH_API_KEY; some endpoints need a higher plan). Based on Brave's official Search API skills; it uses the same plugin id as BuffMcBigHuge/dsh-brave-search, so install only one of the two.

DSH Lean Context indexes available tools locally and preselects a clear match or finds tools on demand.

A local mirrored A/B benchmark for context compression, not another compressor. It runs one fixed read-only scenario twice with the same model and temperature — compression off, then on — and reports input tokens, turns, tool-call success, fact coverage, answer similarity and a per-line diff, with JSON, a self-contained SVG and raw warnings. 18 node --test tests, zero dependencies. The first measured run: 63,029 → 6,927 locally estimated input tokens (−89.0%), 3 → 3 turns, facts 3/3 → 3/3, answer similarity 1.00 on one sample.

Three-tier memory (personal, project, session) built on the Hy-Memory engine, with relevant memories injected before each step. Ships seven memory tools (search, add, list, get, update, delete, forget), a settings panel with per-tier lists, editing, change history and duplicate cleanup, real token accounting per write, session summaries written by the engine at session end, and a user profile refreshed from recent memories. Memories can be exported to and imported from MemHub. Needs Python 3.11+ once to install the engine, or uv, which fetches 3.12 itself; uninstalling keeps the stored memories.

DeepSeek Harness plugin: a Task Instructions button beside the input box pins reusable task requirements to a session, and the instruction survives context compression.

Deterministic, zero-LLM compaction engine for DeepSeek Harness: the checkpoint is compiled locally from the session log with no summarizing round-trip, and dropped tool traffic becomes seq reference lines the model can restore with recall, which helps on small context windows. Ships as two packages: install dsh-fast-compact and dsh-fast-compact-preset, add both to the profile's bundles, pick the Fast Compact (standard) preset and start a new session. Needs dsh >=0.1.5-rc.2 <0.3.0.

Permanently deletes one session from its sidebar context menu after a confirmation dialog: the log directory under the DSH home, the projection-cache thumbnail, its id in every workspace.json, and its archived and pinned entries. Sessions resumed since launch are detached from both the sessions and the agents stores first, so the deleted row does not return to the sidebar as an orphan. Refuses while the session is running a turn, and never touches the workspace folder itself.

Context profiles for new sessions: pick which tools and which system prompt each new conversation receives. Deferred tools stay as name+summary placeholders, keeping the tool list and the prompt prefix cache stable.

A Conductor tab and commands for a chat's context: /new replaces the context in place with a model-written handoff, /rewind restores the chat and the agent's file edits to an earlier prompt, plus /drop, pins, an automatic compaction level, recovery after failed turns, silent messages between chats and agent and task orchestration. Supersedes dsh-orchestrator, dsh-force-compact and dsh-continue, so do not install it alongside them. Requires DSH >= 0.2.1-rc.5, Node >= 22.5, the web profile and dsh-better-sidebar.

Volumes and handoffs for DeepSeek Harness: group sessions per project and hand state to a fresh session through a generated handoff document instead of replaying context.

A learning chat tree for branching chat: fork selected text with inherited context, continue each native session, and return to your reading position.

Switches the tone of DSH replies from Settings → 人格模式 or the /persona command (/persona catgirl, /persona off), effective from the next turn. Only the active persona is injected, as runtime context, so the system prompt keeps a single stable short strategy; with the mode off the injected persona text is 0 characters. Facts, code and engineering discipline are unchanged.

Adds a Settings section to adjust the advertised context window of each connected WorkBuddy and Trae model, writing the override into the connector's context budgets.

Adds a Task Progress tab to the conversation view, next to the Context tab, with a progress bar for the session's todo list and completed items ordered by completion time.