プラグイン

Complement to official Creator mode: stage, validate, live-mount via ctx.plugin, score, and solidify or roll back in-memory tool-plugin trials.

Minimal-anchored agent preset: request #1 keeps the real bash + str_replace_editor pair with auto-injected context stripped, then promotion unlocks the full Standard catalog and posts a one-shot background-jobs notice against sleep-polling habits.

Collapsible right-docked web terminal (xterm.js) bridged to a host PTY over WebSocket: bash/zsh switching, resizable panel, and fold/unfold that keeps the session alive.

RLM mode for dsh: recursive sub-agents via a unified rlm() function — spawn, await, and chain child agents as first-class Python calls. Adds a persistent IPython kernel with tools.* bindings and context-as-variable.

Dependency-topology guard: pre-install conflict prediction and post-install detection of duplicate @deepseek-ai/dsh-* copies, version drift, and vendored core services that cause Symbol-key crashes like "Cannot read properties of undefined (reading 'prepare')".

Verifies that Cordis-managed runtime resources return to baseline after isolated plugin teardown, with deterministic cleanup receipts.

Two-way Feishu (Lark) chat bridge for DeepSeek Harness over the official WebSocket long connection: messages route to the bound workspace's latest session and replies return automatically; bot credentials and workspace routing are configurable from the settings page.

Email reminders from your DSH agent when you're away: fires at goal completion, blockage, before asking the user a decision, and at long-task milestones via a bundled zero-dependency Node SMTP sender.

Short Web Audio cues when a session needs your attention (approval, question, or plan review) and when a task finishes; synthesized inside the page in two selectable styles with a volume setting — no popups, no OS notifications, no audio files, no network — and one bundle that covers DSH 0.1.7+ (desktop & web) and 0.1.0-rc.6.

Desktop notifications and tones when a task finishes, waits for approval, or errors — native on macOS, Windows and Linux with zero runtime dependencies.

Web UI のワンクリック起動：Windows デスクトップの Start-DSH-Web.cmd をダブルクリックすると dsh web が起動し（準備完了までポーリングして）ブラウザを自動で開きます。web_launcher モデルツール（install / open / status）付き。

メタデータのみのDSHターン、モデル、ツール、承認、リトライ、エラーイベントをTuning Enginesへエクスポート。ガバナンス付きトレース、ポリシー評価、コスト分析、Work Sessionレビューに利用可能。ディスク永続化されたリトライキュー付き。

Injects allowlisted GitHub Actions and GitLab CI run metadata into agent context without reading logs, credentials, or provider APIs.

GitHub workbench MCP server (dsh bundle): 23 MCP tools for repos, issues, PRs, code review, and search, with zero runtime dependencies and PAT/OAuth auth.

Git and GitHub panel for the DSH web profile: stage/unstage/discard/diff, branches, commit, push/pull/sync, log, file browser, workflow board, and GitHub OAuth with PR creation.

Main-brain orchestration preset: delegation-first persona, root-session execution gate, and portable AI-LAW rule skill.

Declarative task-DAG orchestration for DSH: validates dependency graphs, runs topological task layers in parallel through workflow-backed subagents, and propagates failures deterministically.

Workflow tools for the local aflare binary: generate, validate and run local-first deterministic YAML workflow DAGs (WAL crash recovery, Saga compensation) through the local aflare binary, with 300+ templates built in.

Document workshop skill pack: four pure-prompt Agent Skills that turn a codebase into READMEs, PR descriptions, changelogs and code reviews, installed into dsh via a dsh.bundle plugin row.

DSH project scaffold preset: strict 5-phase init runbook, engineering standards and six runnable starter templates (node-ts / react-vite / python / go / spring-boot / monorepo).

Give text-only models vision: a describe_image tool that delegates images to a vision model from your dsh model list over the harness's own LLM runtime.

Hands repetitive text and vision labor (OCR, image analysis, comparison) to a locally running Unsloth Desktop (Unsloth Studio) server through unsloth_run and unsloth_vision tools; pure HTTP client, never spawns or owns processes.

Hands repetitive text and vision labor (OCR, image analysis, comparison) to a local KoboldCpp (llama.cpp) server through koboldcpp_run and koboldcpp_vision tools, with on-demand server lifecycle management.

Human-in-the-loop tools (`human_task` / `human_task_ready_check`): pause the agent until the user performs a real-world action or returns an observation (GUI operation, game testing, visual verification, hardware checks), gated by session consent and an AFK presence check, returning a structured JSON result via a web task dialog.