Plugins

Session rotation for long conversations: when a session reaches its step budget, at the end of the turn the plugin opens a new session in the same workspace, posts a handoff summary it assembles itself from local session events, and archives the old one immediately or after a retention window. Per-workspace step limits, a minimum-steps floor, a cooldown, a rotation cap and a dry-run switch are configurable in the settings page, and the summary is built with no model call.

Nested session folders for the DSH sidebar: create folders per workspace, nest them and drag to reorder, move sessions by drag or context menu, click a folder activity count to open the matching sessions, and tag sessions with a header tag filter. Undo, rename, archive, search and keyboard navigation work on the tree.

A broom button beside the context-usage readout under the web composer. Its frame draws the current occupancy, shifting green to yellow to red as the session fills, with stronger wording and a faster pulse as pressure rises; clicking it compacts the session, preferring a native compact or compress command when one is registered and otherwise folding the oldest ranges that contain no user message through the plugin host route.

Adds a right-click context menu to delivered file cards, the changed-files card, inline file references and http(s) links in DSH chat. Actions are open with the default app, open with the system chooser, reveal in the file manager, copy the file itself, copy its path, save a copy to a chosen folder, download, and open a link in the browser. Delivery cards are addressed by Session coordinates the Host re-derives from the session log; inline references and the changed-files card are addressed by an absolute path the Host re-verifies.

Adds one system-prompt section that limits delegation to subagents and teammates to two cases: adversarial review of freshly produced output, and one-shot context reading that returns a summary.

Port of MiMo Code's infinite-context memory: as the window fills it asks the model in the background (the current provider route unless writerProvider/writerModel is set) to write a structured checkpoint and a project MEMORY.md, then, ahead of DSH's native compaction, replaces the visible window with a rebuilt dump of tasks, checkpoint, recent user messages, memory and notes, falling back to a MiMo-style compaction. Memory files live under $DSH_HOME/mimo-context (per session, per project, plus a read-only global one); it adds mimo_* memory, history, checkpoint and rebuild tools and /mimo-* commands, with dream/distill off by default. Needs DSH 0.2.2-rc.1 or a later 0.2.x.

Externalizes oversized assistant and tool messages in long sessions into project files while keeping a recent window, and archives reasoning from empty turns as a visible message so long tasks do not stall on context size.

Adds a recoverable delete to the session context menu: the Host archives the session, detaches it from its project and moves the session directory plus the Markdown files that session wrote into $DSH_HOME/session-trash; a session with running work is refused.

Permanently delete DSH session records from the sidebar context menu, removing the session log directory, the projection cache entry and the workspace registration in one action. A confirmation dialog guards every delete. Its /dsh-session-purger routes, including permanent delete, are mounted with webServer.register and carry no DSH Host/Origin admission check, so a web page open in your browser can delete session records through them.

Adds a right-click menu with five original-image actions — open in a tab, copy the PNG, add to the composer without sending, reveal the file, download — to Codex-generated images and to images you send yourself, as a companion to dsh-codex-subscription.

Makes a DeepSeek Harness agent reason in Simplified Chinese: each turn it appends one runtime-context instruction at the end of the conversation, and it can keep a stable declaration in the system prompt as well. Settings pick Simplified Chinese or off, and whether that declaration is included.

Seal proven command chains into reusable executable tool scripts for DSH agents. DSH 代码脚本固存，留住模型成功自实现的工作流。

Godot game development workbench for DeepSeek Harness. Windows only, preview release: plans a game (GDD, tasks, acceptance checks), changes project files through controlled read/write tools, diagnoses the project and runs it for evidence. Runtime and import checks launch your configured Godot as a normal process with your permissions (the AppContainer sandbox applies only in the restrictedProduction option); the official Godot 4 standard build can be downloaded on request after confirmation (SHA-256 checked), creating a project registers it in Godot's project list, and the selected project context is sent to your configured model.

Sound cues played by the host operating system, not the browser: three bundled WAV files fire through the machine's own audio (SoundPlayer on Windows, afplay on macOS, paplay/aplay on Linux) when an agent's turn completes, when the context is compacted and when the agent is waiting for your answer; subagent turns can be included, cues fired closer than a configurable interval are skipped, and each bundled sound can be replaced with your own file via plugin config.

On-demand MCP tool loading. Only the mcp__* tools listed in the coreTools whitelist (max 5) stay in a session's tool set; every other MCP tool definition is actually removed from the request through the session-scoped tools.restrict() API and comes back only after the model calls lazy_load_tool. A tools/pre-execute gate turns calls to unloaded tools into a load-it-first error, and activation resets when the session ends.

Manual context compaction for DSH Agent Teams, adding a scoped compact_context tool and the /compact-agent command.

Replaces LLM summarization with deterministic text compilation, so compaction makes no model calls and keeps original tokens; elided content stays retrievable through seq pointers. Based on TsFreddie/dsh-compaction-instant.

Adds a Q&A agent preset that answers each new question from a clean context: earlier exchanges leave the model input via a surface replacement, with no extra model call and no summary residue. A /qa follow command and a composer toggle switch to follow-up mode, and /clear archives the session and opens a fresh one under the same preset.

Manual session handoff: when you decide a session has run out of usable context, one click turns its event log into a compact handoff document (goal, done, next, project facts), creates a new session seeded with it and wakes it to continue; the new session can inherit the session-bus name as <name>-handoff. No threshold, no timer, no automatic trigger.

Folds older oversized tool results into spill files and replaces them with short, deterministic stubs at step boundaries, with no LLM call: the stub keeps the first 200 and last 80 characters plus the sequence number, tool name, line/char/size counts and the original file path, and each fold is a real surface replacement accompanied by compaction/prune shadow pricing so token counts stay honest. Triggers on context pressure at 25% of the model window (falling back to 32768 tokens when the window cannot be resolved) or on five accumulated candidates; the last three turns, error results and rich content are never folded. The original is recovered from the session log rather than the spill file, so /folds, /unfold <seq> and /unfold all keep working after spill cleanup; session_fold_recall lets the model fetch a folded result by sequence number or by keyword, and session_fold_diagnose reports the effective settings. /spill-prune removes orphaned spill directories, preview only unless --apply. A Settings page exposes 12 of the 13 options as draft-then-save fields; they are volatile, so saved changes apply without restarting the plugin.

Puts a domain bar under the session header — a domain groups skills and tools and toggles as one block, so one click loads or unloads every member. Hovering a chip lists its members (skills, then tools) in a height-capped panel that scrolls. The right end of the bar is a compaction-threshold counter (− 5 +, 100k tokens per step) that rewrites the live compaction engine policies, so the compaction trigger line changes on the next turn without restarting dsh; a degraded fallback turns the number amber and is reported in the snapshot. Settings maintains domains and members with search, skill/tool tabs, all/selected/unselected filters and a copyable AI grouping prompt. Tool removal uses tools.restrict plus a guard; hiding a skill re-registers it with modelInvocable false, so it leaves the model catalog while /name still works. Sessions of another preset are left untouched.

Exposes the chat models of a locally signed-in QwenWork desktop app as a Harness provider, with a multi-account pool, session affinity and failover, pre-expiry credential renewal, a health watchdog, per-model enable/context-window/input-type settings, and account status with credit balance. Unofficial local integration that reuses your own signed-in client credential; your own account only. Windows only: needs the QwenWork desktop app signed in, and keeps its account pool in a DPAPI-protected file under %LOCALAPPDATA%. The author warns that reusing the client credential may conflict with the vendor's terms.

Gives the agent its own context usage through a context_usage tool plus one reading line per turn, with an editable template.

Delivers one required user-role runtime context that makes the agent reason in Simplified Chinese and keep a whale-loli in-thinking persona (calls itself 本鱼, lazy-arrogant tone) in both reasoning and the final reply, while proper nouns, identifiers, paths, commands, URLs and error text stay verbatim; persona, extra proper nouns and context order are configurable. Enabled by default for every agent in the profile, subagents included.