Plugins

Annotate replies and documents for the next message, edit past user messages and rerun from that point, and use keyboard shortcuts to navigate user messages or recall previous inputs in the composer.

An isolated translation pipeline where each module is its own model call carrying only its own prompt and its own input — no session history, no tool schemas, no other module's content. Build chains from draggable cards, edit any module's prompt in place, override the model per module, or start from one of two precision presets. The same chain is exposed to the agent as a translate_flow tool.

Single/multi-line composer toggle for the DSH Web GUI: an editable shortcut (registered through the harness shortcut service) — Cmd/Ctrl+E on desktop builds and Cmd/Ctrl+Shift+E in the browser, with Linux browsers unbound until you pick one — switches the composer between its native single-line shape and a four-line-plus draft area; a small tab on the composer shoulder always shows the current shortcut and target mode; while multi-line, Enter inserts a newline and Shift+Enter sends, and the mode is remembered per device.

Shows an aggregated session-status dot in front of every folder in DSH Web's sidebar workspace tree — waiting for input, just finished, or running — and bubbles the sessions of nested workspaces up to their parent folder.

A polish button in the composer trailing controls that rewrites the draft in place with the model. Context is matched by topic, not by recency: anchors from the draft (paths, identifiers, quoted text, CJK bigrams) pick the earlier messages about the same task, so a session that interleaves several tasks does not leak another task goal or files into the rewrite — and the prompt explicitly lets the model ignore the extract when it does not fit. The result goes back through the same undo-able editor path a human edit uses. Drafts holding reference chips (@file, folder, session) or a claimed /command, and drafts mid-submission, disable the button instead of silently dropping references.

Composer history that restores the whole input state: composer contents, image and file attachments, @ references and slash commands, plus the model, thinking effort, permission and agent preset used.

Persistent prompt history for the DSH composer — Up/Down recalls sent prompts across restarts, Up with text typed filters by prefix, and cleared drafts are not recorded.

Keyboard shortcut for the DSH voice input composer button: start recording, press again to stop and transcribe.

Windows desktop control for DeepSeek Harness agents: window and UI Automation automation, local Windows OCR, direct browser reads over the Chrome DevTools Protocol (DOM text, JavaScript, selector clicks), humanised mouse and keyboard input, browser tab management, and a live control panel. It probes the machine on first run, writes its own config, and degrades gracefully when a capability is missing. Windows 10/11 only.

Auditable token-cost ledger for DeepSeek Harness: wraps the llm/stream waterfall and records every call with provider, model, input/output/cache-hit/cache-write tokens, usage source and integer nano-CNY cost from an explicit price table. Aggregates by hour, day and session; exports JSON or CSV with a no-write dry-run preview; keeps a manually entered official platform bill in a parallel column with actual-minus-estimate diff. Unknown provider/model routes stay explicitly unpriced. Zero runtime dependencies; no balance query, notifications or animations.

Per-call token meter for the DSH composer: a colour-coded pill under the input and a dashboard above it showing every model call (total versus billable input+output), sub-agent spend, tool overhead and cache-versus-fresh input, read only from local session logs.

/cmd <command> in the composer runs that shell command in the session working directory, sends the command and its output to the model as a user message, and renders the input as a terminal card with exit status, expandable output and copy buttons; a bare /cmd opens an interactive terminal whose transcript is sent to the model with Ctrl+Enter. Runs with its own terminal permissions (danger-full-access) by default; set useSessionPolicy: true to use the session sandbox.

A floating in-conversation find bar docked over the DSH composer (Ctrl/⌘+F or the magnifier beside the input): hits list with a match counter as you type, Enter jumps to the next hit, hovering a hit pops a scrollable full-text card with a one-click copy of the whole message, the up arrow in an empty composer recalls previously sent messages one at a time with their image and file attachments restored as drafts, hovering any image in the conversation shows a copy button that writes it to the clipboard as PNG, and every finished answer gains a regenerate button while the newest also gets continue. Client-side only, searches the messages already loaded in the conversation, and never jumps.

Mirror an iOS Simulator or Android emulator into DSH as floating overlay windows, several at once, and tap, drag or type on the mirrored screen with the input forwarded back to the device.

Thinking-sorting plugin. A switch beside the composer puts the session into a three-stage mode — sort, critique, research — and each stage has a checklist of what it should produce before you press the advance button; finished ideas are filed by genre into Settings → Thoughts Center. A floating notebook holds half-formed sparks and opens a new session with the chosen one already written into the input box.

Local wake-word voice input for DSH. Say the wake word, wait for the chime, then speak; text is sent to a DSH session. Audio is processed locally by default and not stored. Requires the DSH speech input service (speechToText) to be available.

Tree-style questioning for DSH: click "Ask in tree" on a selection to add a node to the right-sidebar tree and ask inline under it; clicking another node makes it the current parent, so later selections nest there, Alt+click opens a root node, Ctrl+click appends another passage to the same node, double-click jumps back to the quoted text, a red/green dot marks each node as understood or not, and the panel switches between a list view and a horizontal branch graph. "Add to composer" instead puts the quote into the input box as a reference chip, without creating a node. Nodes live in localStorage and nothing is injected into the prompt.

Adds an Optimize button to the composer that rewrites a one-line idea into a structured product brief: a positioning sentence, 3–6 key features, visual references and design tokens; it sits beside your original text in a compare dialog where you can accept it into the input, generate another version, or cancel and keep the draft untouched. No settings page and no API keys — it calls the model channel you already configured.

Adds a full-screen dialog for writing long messages: Markdown on the left rendered live by marked on the right, with a draggable splitter that remembers its ratio. Confirming writes the text back to the composer draft, keeping command tokens and @-reference chips as a fixed prefix; Esc asks whether to discard or write back before closing.

Effort slider in the model menu, a live tokens-per-second and turn-total readout beside the running status, a per-answer input/output/cached report, and a completion notification when a turn finishes, for the DeepSeek Harness Web GUI.

Two buttons beside the chat input: one rewrites the draft with the current directory in view, the other inserts a role template from 85 curated entries generated from the bundled `agents/` role library, which the settings card can extend with your own `.md` role files.

Inspects ONNX models (inputs/outputs, shapes, opset, operators, size) and benchmarks them with ONNX Runtime on the local machine or on a remote target over SSH, reporting average/P50/P95/P99 latency, FPS and the execution provider actually used, with PASS/FAIL gates for FPS, P95 latency and model size. Requires Python 3.10+ with numpy, onnx, onnxruntime, and psutil installed (pip install numpy onnx onnxruntime psutil); remote targets need python3, numpy, and onnxruntime.

Plays a voice alert and a browser notification when a turn finishes, asks for input, or fails, and can push the same summary to a Feishu bot webhook as a card.

Extends the composer model selector menu with a third Context window row, preset capacities (64K–2M), custom input, provider-wide sync, and adaptive compaction guard.