Plugins

A model-native meta-agent plugin for DeepSeek Harness: the model synthesizes task-specific workflows at runtime from five agentic workflow patterns (prompt chaining, parallel workers, router, supervisor, evaluation loop) and coordinates tools and subagent

Per-model token cost tracking with configurable cache-hit/miss, output and peak-window pricing, a live session cost bar, and unconfigured-model flags.

Claude-Code-style lifecycle hooks: configured shell commands run before/after model tool calls with a JSON payload on stdin; a non-zero pre-tool exit blocks the call.

First-class Git tools for the model: status / diff / log / commit / branch with validated messages and paths - no bare-shell git calls.

Per-call model, provider, persona, and toolFilter overrides for subagent delegation, with @preset: references and provider/model composite ids.

Composer-attached images are transcribed to text by an OpenAI-compatible vision model before reaching text-only DeepSeek models.

Bounded local memory with CBDC authority gating: SQLite + FTS5 claims, scoped recall with explainable use/verify/ignore decisions and a full audit trail, /memory commands, ≤3-claim/1200-char injection per call, no extra model call.

Animated DSH Web pet with model-adaptive output pacing, tool-state reactions, replaceable character packs, and a bundled skill for creating or extending pet animations.

Lightweight zero-dependency memory plugin for DeepSeek Harness: Markdown-file memory store, L0 catalog injection, and 5 model-facing tools.

DSH 模型群聊：多模型角色群组对话面板。角色绑定不同 provider/model，群内共享对话记录；设置页支持启用/停用。

Prompt-polish template injector for DSH: click ✨ in the composer to inject a "polish my prompt" template (current draft auto-embedded), no send, no network, no model — the main agent rewrites with full context. Replaces narrative-prompt-polish (2026-09-14 migration).

Model-callable compact_agents tool that force-compacts context in every live session (main session, ordinary sub-agents and AgentTeams members alike), queues busy targets and compacts them when the turn ends, reports the shadowed node count and estimated tokens per target, and ships a browser form for the trigger ratio, retained ratio, controlled-phase output budget, auto-continue budget and an end-of-turn pre-compaction ratio that compacts near the trigger line once a turn ends, so a reply no longer opens with a wait for the summarizer.

Automatically highlights key points, definitions, warnings, and questions in completed assistant answers using one separate model call.

Shows a Chinese translation under each English thinking block in the DSH web chat, without modifying the session transcript; per-chunk caching, a configurable translation model, and a settings panel.

Custom modes and custom prompts for DeepSeek Harness (dsh): edit a mode's system prompt on the settings page (it takes effect on the next model step), choose its base mode, switch plugins row by row, and keep several assistants side by side.

Per-reply read-aloud with a multi-provider settings page — MiniMax or any OpenAI-compatible /audio/speech endpoint, with voice, emotion, speed and model selection plus an auto-read toggle.

Per-project memory for DeepSeek Harness. Three Markdown files per workspace — PROJECT.md, DECISIONS.md and SESSIONS.md — are auto-created on the first step of any session in that workspace, re-read from disk and re-injected only when they change (and re-injected if compaction dropped them), and written back by classification through a memory_checkpoint tool, with a bounded end-of-turn nudge when a turn did work but recorded nothing. memory_read and memory_search read the live log and the archive, the latter over a zero-dependency BM25 index. The session log is capped at 200 entries with the overflow moved to SESSIONS-archive.md — never injected, but visible, readable and searchable, and restorable one entry at a time. Settings page: a workspace list with a path filter, read-only tabs for the three memory files plus the archive, reload, export/import of one workspace's memory as a JSON bundle, a two-step clear, a panel that rewrites or removes the AGENTS.md instruction block, and a PROJECT.md staleness reminder. A composer status chip shows that workspace's sync state and creates the files on click. Zero dependencies — no vector store, no embeddings, and no model calls.

Layered memory for DSH whose agent-facing surface is 10 tools: high-privilege writes (memory_add, memory_delete, memory_import), rumination controls (memory_ruminate, _cancel, _status), and a memory graph (memory_search_graph, memory_expand_graph_node), built over an L0-L3 distillation pipeline that recalls and injects relevant memories before each model step.

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.

Reasoning effort slider for DeepSeek Harness — a client (web) bundle that injects a per-session slider into the conversation input, shown when the selected model declares reasoning effort levels. Works with any provider that has reasoning efforts (Bailian

Session instruction anchor for DeepSeek Harness: the first step of a new session injects a freely composed instruction block as a plugin-sourced message, and it is re-anchored after a summarizing compaction, every N turns, or when metered context pressure crosses a threshold. An /anchor command re-anchors on demand without a model call; a settings page manages the paged preset library, injection order, re-anchor source and character limits.

Adds a 3D model library in its own conversation tab, where agents store models built from CadQuery scripts or supplied as tessellated meshes; the library previews them with three-cad-viewer, organises them into folders and exports them to 10 formats including STEP, BREP, STL and DXF.

Dense translucent-glass Universal Palette for DeepSeek Harness Web — federates Commands, Models, Sessions and Conversation Hits with deterministic ranking and verified public-contract interoperability.

dsh plugin: switch named model profiles (default model, think level, per-subagent models) interactively via the host ask-user flow — one implementation for the TUI and web surfaces, a clean fast-fail on headless