プラグイン

メモリプロトコル強制プラグイン: opencode-memory MCP サーバーをブリッジし、ツール呼び出し前に memory_weave を強制し、ターンを自動取り込み、エージェント手順にメモリコンテキストを注入します。

Codex スタイルの自動隠蔽式右端 turn rail：edge-peek 表示／非表示、per-turn tooltip、rail 内検索、長時間セッション向けスクロール可能な 30-turn ウィンドウ。

ask_user_question 向けの協調的なタイムアウトガード。見失われた、または未回答の質問は、ターンを永遠に停止させる代わりに、構造化された ASK_TIMEOUT として解決されます。

DeepSeek Harnessセッションを日次、週次、引継ぎ、記事レポートに変換。検証可能なHTMLエクスポートと保存ファイルを再チェックするレシート付き。

DeepSeek Harness 向けの、リプレイを考慮したターンごとのモデル使用量とセッションコストの表示。

Native Windows notifications when an unawaited background job settles (completed, failed, or stopped), when a conversation turn ends, or when the agent waits for a question answer or tool approval.

Cost ledger tab for DSH: per-turn token accounting converted to CNY, stacked daily charts styled after the DeepSeek console, and a built-in Alibaba Bailian price book covering 142 models (Qwen, DeepSeek, Kimi, GLM, MiniMax). Models no catalogue covers stay explicitly unpriced instead of being counted as zero, and a manual price fills them in with one click.

Composer-side helpers for the DSH Web UI: next-message prediction shown as grey ghost text that Tab fills in, a one-line per-turn recap above the input (placeable below instead) that never enters the model context, a right-sidebar panel for follow-ups from past sessions, recap history and side questions with their own model picker, and config-driven reply length, style and actionable framing.

Predicts your next message after each completed turn and shows it as grey ghost text in the empty composer: Tab or → fills it in without sending, Esc hides it, and a rebindable shortcut (Alt+\ by default) predicts on demand or steps through several candidates. Off until enabled in its settings page, which has three one-click modes (balanced, saver, quality), an optional left-sidebar entry, daily request and token limits, masking of keys, tokens and e-mail addresses before sending, and a 7-day usage summary; short messages repeated in the session and a list of pinned phrases come after the model suggestion when stepping through candidates, without a model call, and the interface is available in nine languages.

Predicts your next instruction from the current conversation and shows it as the composer placeholder; press Tab to turn it into an editable draft.

Plays a chime or a short melody when an agent turn finishes, with five synthesized tones plus a custom audio upload, optional spoken announcement, optional system notification, and per-subagent control; settings are stored on the host.

Reviews each finished turn, distills reusable skills as drafts, runs code-level checks over them, and holds each draft until the user approves or rejects it with a directive; every write is recorded in a SHA-256 ledger that supports rolling back a single entry. On by default, it reviews finished turns with the current session model (or an OpenAI-compatible endpoint set through DSH_EVOLUTION_BASE_URL), sending a bounded excerpt of the conversation, and writes drafts to ~/.dsh/skills/.staging until you promote them. Its bundle patch lowers DSH's skill-catalog description limit from 500 to 80 characters for every skill, and it starts an unauthenticated HTTP server on 0.0.0.0:18890 (DSH_EVOLUTION_PORT) that can list and promote drafts, change settings and trigger reviews.

Derives the index from your Markdown memory files on every turn instead of keeping a hand-written index table, so renames, splits and deletions never leave a stale entry. Retrieval ranks by library tags, section tags, title tokens and body text, then injects the matching line ranges rather than whole files. Writing is model-proposed and user-confirmed through a sticky note; entries are appended after the target section's format is detected. Ships a structure panel showing each library and section with line ranges and token estimates, a starter template for an empty memory directory, and optional appearance controls.

Local-first memory governance: distills conversations into categorized memories, surfaces conflicts for human arbitration, and previews what gets injected next turn.

Long-term memory for DSH where an LLM decides what to keep: turns pool while the session is idle, then the main chat model distills them into one-line entries in a single batched call; vector dedup runs before insert, so near-duplicates and saturated topics are dropped; recall uses hybrid vector plus FTS5 search over tiered hot, warm and cold storage; ships remember, recall, forget and prune tools. Zero dependencies, Node only.

Long-term memory for DSH: rule gates capture durable facts from conversation turns into a single SQLite database with tiered hot, warm and cold storage; recall injects hybrid vector plus FTS5 hits into the prompt each turn; ships remember, recall, forget and cold-memory prune tools. Zero dependencies, Node only.

Shrinks the first history page of long DSH sessions to at most 100 messages, targeting the latest two turns, and fills the viewport in the background with the built-in load-earlier control when the trimmed transcript is shorter than the window.

Connects the Solmirror open platform to DeepSeek Harness as MCP tools covering image and video generation, a model catalog, cost estimates, reference uploads, result queries and media URLs. Generation jobs submit as native dsh background tasks that notify on completion and run in parallel within one message, with in-chat video playback and an image wall. A settings page takes the platform API key, server URL, per-group model visibility and a test-connection check; the key is stored in the local dsh credential store. Reference uploads (image, video, audio) return reusable ids kept one day unused or three days after last use — 50 MB per file, video up to 30 s, audio up to 60 s. Usage bills platform points at in-site prices, with quotes before a run and automatic refunds on failure. Files passed to upload_reference (jpg/png/webp/gif, mp4/mov/webm, mp3/wav, up to 50 MB) leave your machine: images and videos go straight to Volcengine ImageX/VOD storage with temporary credentials issued by Solmirror, audio through the Solmirror server; its HTTP routes (/plugins/solmirror/...) have no DSH access check.

Connect DeepSeek Harness to the TeleAgent (China Telecom TeleAI desktop app) models through the app's own local kernel after signing in there, keeping DSH's streaming, multi-turn, images, tool calls, reasoning and points balance. macOS and Windows only, with TeleAgent 2.7.1+ installed and signed in. It reads the app's saved login token and points cookies (macOS Keychain / Windows DPAPI), starts the app's own kernel binary as a local child process, and needs the gateway apiKey the desktop app ships with, supplied through TELEAGENT_API_KEY or a local.secrets.json in the plugin folder. Daily points claiming is on by default; adding or switching accounts backs up and rewrites the app's login files. It uses the same package name and mount id as the listed alvinhuang0509-design/dsh-connect-teleagent, so install only one. The author's disclaimer asks you to follow TeleAgent's terms of use.

A market card pinned to the DSH sidebar: live price and 24-hour change for crypto pairs, A-share / Hong Kong / US stocks and world indices, FX majors and London spot gold and silver. An optional long or short cost basis plus a position size turn the change column into a P&L percentage and a cash amount; symbols, order and refresh interval are chosen in DSH settings. No login and no API key. Quotes come from public endpoints of Gate.io, Eastmoney and Sina, and the symbols you choose are sent to them on every refresh.

Keep working when a turn ends without finishing, and only stop once the reply ends with a completion keyword.

Bridges campfire group chats into DeepSeek Harness, delivering messages addressed to your member into a DSH session and posting the reply back to the group. Needs a self-hosted campfire server plus a member handle and channel token; anyone who @-mentions that member in the configured chat can start a turn.

Validates every tool call against that tool's own parameters schema before it runs, and denies with the exact path and type. The host renders a tool's parameters into the prompt but never checks a call against them, so a wrong-typed value reaches the tool body: a number lands in path.join and the caller gets a raw TypeError instead of an answer. This plugin hooks tools/pre-execute, resolves the definition, and returns a deny whose reason reads "demo_tool.dir: expected string, got number (123)". Covers primitive types and unions, required, enum/const, nested properties, array items, additionalProperties:false, and oneOf/anyOf/allOf. Fails open on purpose: an internal error, an unknown tool, or a schema construct it cannot read allows the call. mode deny or warn, plus a skip list. Two read-only tools, arg_guard_status and arg_guard_check.

Disable plugins per workspace: a disabled plugin stops contributing tools to the model tool list and prompt sections to the system prompt, cutting the per-turn schema cost. A plugin is switched as a whole, with packs evaluated by workspace binding, then default.