Plugins

Imports sessions from pi, opencode, codex, claude-code and WorkBuddy into DSH, unifies AGENTS.md, skills, project and per-agent memory across harnesses with symlinks instead of copies, and migrates Codex and WorkBuddy automations into DSH scheduled tasks.

Folds degenerate reasoning loops out of the model-visible history so they are not replayed upstream on the next turn. Detects two shapes — one sentence hammered repeatedly, and short-sentence rotation — and replaces the matched span in place; a companion escape valve aborts a turn that will not stop on its own.

Auto-compaction that writes a handoff note first: as a session approaches a token threshold it asks the model to save a project checkpoint into the workspace, then optionally starts a fresh session seeded with a continue instruction and parks the old one. The note must be written within the current compaction cycle and its conclusions section must actually change, so a stale or merely touched file cannot slip through.

Visual conversation map for DeepSeek Harness — sessions as cards on a whiteboard: drag to arrange, double-click to chat, draw an edge to fork with injected context. ADHD-friendly spatial memory. · DSH 可视化对话地图

Long-term memory for DSH: conversations auto-distilled into atomic facts, scene summaries and a persona profile, with relevant memories injected before every model step. Zero-config out of the box — hybrid BM25 + vector retrieval, optional offline local embeddings, chat/work family separation.

Two-suit VTuber theme for Joi: Flowers and Library are separate token systems for light and dark, with a standing portrait and a whale girl napping on the headline of the new-session page, two chibi companions above the composer that follow the turn state, a duo sidebar mark, and a context meter shaped as a ripening mandarin; a third card returns the stock DeepSeek look with the plugin still installed.

Presents DSH Web as a playable, full-screen first-person Minecraft-style voxel interface—not a color-only skin: users walk through a persistent Three.js block world and access native workspaces, sessions, conversations, tools, model selection, permissions, context information, and the composer through in-world facilities and a workbench.

Semantic retrieval, resource management, auto-recall (user + agent dual spaces) and session memory for DeepSeek Harness: memsearch/memfind/memread/membrowse/memgrep/memglob over viking:// URIs, memadd/memremove resource management, memqueue observer status, memcommit session capture, and memlearn skill playbook minting.

Terminal-style input history for the web composer: edge-first arrow-key recall with exact draft/caret restore, browser-local persisted history, Ctrl+R reverse search, and sliding-context awareness; 0.5.0 adds a smart input layer — cross-session snippets (/save, /load), prompt templates with variables, reuse insights, and compaction-summary highlighting.

Conversation rendering: every fenced block is rendered by the plugin itself (markdown for text fences, language-banner code blocks with highlight.js for others), mermaid fences become zoomable diagram cards, plus context-body rendering, whole-content copy and a MarkdownView export (npm: `dsh-md-render`).

Context lens: visualizes token usage of every request (context composition, KV cache hit rate) and enforces per-turn / per-session budgets with warning and blocking (npm: `dsh-my-context`).

Replaces the chat composer with an editor-style input panel on CodeMirror 6: line numbers, edit + markdown-preview split, a floating window mode and a full-height sidebar tab — one input surface at a time. Drafts ride the official input machine; the official voice-input button is rendered when the host provides one, and the stats row and context meter read the same data as the official dock.

Anatomy panel for the Agent-Body organs: serves a live vitals dashboard at /anatomy inside the harness itself - organs, heartbeat, healing ledger, self-training counters and the real pulse stream - read straight from the runtime files rather than retyped. Host-only, no client bundle and no build step.

Zero-residence context engine: large tool payloads are evicted out of the live context into a queryable store and only deterministic pointers are left behind, so per-turn prompt tokens stop growing with the transcript. Any masked content is reconstructed verbatim from the session log on demand, by tool-call id or session seq.

The organ kernel of Agent-Body: plugins organized as one body. Each command is turned into a nerve impulse that deterministically innervates the organ which should handle it and names the capability to use; a heartbeat with a variable interval (faster on alerts, slower when idle) pumps operator directives, proprioception and homeostasis warnings system-wide; reflex arcs fire with zero model calls; failures go through deterministic attribution (missing tool / arguments / permission / timeout / network), a prescription, and a re-check loop, with argument errors never auto-retried. A missing organ is compensated by the one with the highest capability overlap, and tool schemas are gated by the current intent (84.7% measured reduction, reproducible in-repo). This entry is the kernel plugin package inside the agent-body repository.

The memory organ of Agent-Body: after 2 minutes of silence it enters light sleep and after 5 minutes deep sleep, where deterministic rules distil real experience into memory cards (pitfalls, playbooks, hotspots, unresolved issues, facts). Incoming commands recall related cards by keyword and inject them into context; cards unused for long decay on a half-life and are archived rather than deleted. It also de-noises body-wide homeostasis warnings: repeats inside a window are counted instead of re-printed, and a warning that keeps recurring without ever healing converges into a known steady-state offset and goes quiet. Consolidation and recall make no model calls.

Per-call cache billing inside the context-meter popover: current step, round and session totals with hit/miss/output tokens and amounts, automatic Beijing working-hour peak/off-peak pricing, per-model rates, and third-party relays billed too.

Kiro model provider with Builder ID, Identity Center, Google, GitHub and credential-import sign-in, automatic token and profile refresh, account usage, live model discovery, selectable reasoning efforts, streaming tool calls, provider stop-reason and context-overflow handling, and token accounting that uses Kiro's exact counters when it sends them and otherwise scales the context-usage percentage it does send by the model's advertised window.

Automatic context compression for DeepSeek Harness: compresses tool outputs before they reach the model and keeps every lossy compression reversible via CCR retrieval tools.

Finder-style session labels: a hover button next to archive and pin, plus the native context menu, assign coloured labels; titles can take the first label's colour or a label symbol; view options gain a filter by label applied before grouping. Labels are stored by the Host and shared across workspaces. Needs DSH 0.2.2-rc.1 or newer and Node.js 24.9+ (node:sqlite); the interface is available in Chinese and English only.

Same-origin in mechanism with the two loudest lines in context compression. The code-skeleton gate follows the skeletonization approach of Headroom (Apache-2.0), whose published headline is 20% fewer tokens for coding agents and 60–95% fewer tokens for JSON, same answers. The estimator channel follows TokenPilot (arXiv:2606.17016), which reports up to 60% lower cost for long-session agents. Both figures are theirs, quoted as-is; this plugin ships no benchmark of its own and claims no reduction of its own. What it adds for DeepSeek Harness: choose a compression profile, set the Auto Compact trigger level and toggle code-skeleton compression from one settings section, with exact DeepSeek V4 tokenizer measurement, same-revision count verification, and fail-open behaviour that keeps the original tool results on unsupported models. Within a turn it supports skeletonized information compression with a low original-text retrieval rate; at the tail of a turn it supports a turn-tail intent summary implemented purely by content substitution; context content stays safe, and the whole process runs inside the official JSONL compression.

更好的右键：DeepSeek Harness 应用封装端的完整原生风格上下文菜单

Aggregated multi-upstream web search for DeepSeek Harness: replaces native web_search with model-native grounding and agent search APIs, source attribution, and rich citation cards.

Business knowledge-network context for DeepSeek Harness: bind a session to an OpenBKN knowledge network, query its objects and metrics through managed MCP tools, and open trace-backed provenance (execution facts and a business graph) for each answer. Ships with the OpenBKN Runtime; the provenance views need an OpenBKN enterprise license.