Plugins

Expands every session in the DSH Desktop sidebar. Its control page (/dsh-sidebar-expand-all) applies four equal-length patches to the app.asar workspace client bundle, with backup and one-click restore; after that the "Expand remaining N sessions" row reveals all hidden sessions in one click, collapsing returns to exactly 5, and a workspace reopens fully expanded instead of resetting to five. Needs an app restart, and a DSH update reverts the patch.

Suggested next-step prompts under every finished agent turn in the DeepSeek Harness chat: the model proposes a few followup prompts at the end of each turn, they render as clickable cards under that turn, and clicking one sends it verbatim as your next message. Each card stays attached to the turn that produced it across reloads and restarts, the generation language can be chosen from 15 languages or follow the conversation, and a settings page in the dsh settings covers the language, how many suggestions per turn, the header and the in-app update check. Requires DeepSeek Harness 0.2.x (web profile) and Node.js 20+.

Adds a compact mid-turn steering box above the composer for text, images, and files, with read and pending status, cancel, and unfinished-draft rescue.

Connects DeepSeek Harness to LRC's local structured memory: distills conclusions every 8 turns, drops near-duplicates by character-level Jaccard similarity (≥ 0.85), and writes them to a local LRC store. Five tools — lrc_recall, lrc_forget, lrc_remember_now, lrc_handoff and lrc_status — cover on-demand recall, retracting wrong entries, immediate writes, session handoff and a status line. On first run it downloads a per-platform LRC binary (win32-x64, darwin-arm64 or linux-x64, SHA256-verified) and starts an isolated instance, so the LRC desktop app is not required; data stays in $DSH_HOME/cache/lrc-plugin/data unless LRC_BASE points at an existing instance.

Workspace-wide conversation tree for DeepSeek Harness: every conversation and branch on one canvas with its source turn, own token usage and notes; compare two branches, bring a summary back, export a Markdown outline, archive (undoable through DSH's native archive), and delete a whole subtree (macOS and Windows only; deletion runs immediately and is permanent).

A local small-model habit hint plugin for DeepSeek Harness that reads recent turns, extracts the user habits, key points and hints, and injects them into the conversation system prompt. It needs a Python virtualenv with torch/transformers/peft plus a user-supplied Qwen3-0.6B base model and LoRA adapter, none of which are bundled; its default scope only runs in sessions that companion-memory has enabled, and with the shipped placeholder paths it silently injects nothing.

Renders a Live2D Cubism 4 model as the DSH web UI background: a transparent click-through canvas that covers the window without blocking any control. Eyes, head and body follow the pointer; expressions track the agent state (thinking, running a tool, turn finished, error, waiting) and the conversation (message arrives, reply ends, question asked); after an idle delay the character shows speech bubbles and plays motions and expressions drawn at random from pools you tick. Settings list the models under ~/.dsh/live2d-stage/models, expose position, scale, opacity, gaze strength and every interval and probability, and open a resizable live preview of the same canvas. Expressions and motions a model3.json omits are detected from the model folder; one sample model is included.

Judges the right reasoning level locally from the current draft and recent turns and switches it in one click; warns when the context grows long and tells apart "new topic" from "same unfinished task" — less overthinking on simple tasks, enough reasoning on hard ones.

Wakes a session when a long background task finishes after the turn has ended, and lets the model schedule its own wake-ups (once, interval, or cron with jitter) under quiet hours and a daily budget.

CRT idle screensaver for DSH pages: when no session is running and the page has seen no mouse or keyboard input for ten minutes (configurable), a retro CRT plate covers the window with scanlines and a looping inversion sweep. An in-window click or key press dismisses it, the wheel only locks scrolling, and the whole frame drifts slightly after the cursor; rendering is fully GPU-composited frame by frame. An adaptation of justawebsite/dsh-splash-crt (MIT), turned from a boot splash into an idle screensaver.

A whale or pixel-crab pet that perches on the DSH input box and reacts to the current session (working, awaiting approval, done, error), eats the tokens each turn spends and keeps a diary.

Session-scoped roleplay for DeepSeek Harness: seven fixed character slots in the system prompt, on-demand lore recall, a per-session author's note, a per-turn depth note, and injection preview. Zero runtime dependencies. Rewritten from ZelinW1/dsh-cosplay 0.5.1; do not install it alongside the upstream plugin — both register the same cosplay-core bundle row.

Plays a completion chime — the same sound as OpenAI Codex — on turn end or when the agent needs you (approval, question, plan review). Subagent turns are skipped; mute and volume are adjustable.

Text-to-speech for the dsh-go native phone UI: long-press an assistant reply to read it aloud through Xiaomi MiMo TTS or ByteDance Doubao Seed TTS. The reply is segmented and pre-synthesized in the background after each turn, so one tap hands the app a single continuous MP3 instead of one short chunk per tap; playback has a stable lease so stop works, a status strip and progress bar give live feedback, and audio can be exported to phone storage. Keys stay on the device (0600) and are never rendered, committed, or sent to the model.

Delegates independent tasks from DeepSeek Harness to the local OpenAI Codex CLI and returns the result to the current conversation. Provides four tools (task execution, status check, danger-mode toggle, enable/disable switch), enforces a real-path working-directory whitelist with symlink-escape protection, defaults to Codex workspace-write sandbox mode, passes the task body as a single argv argument to avoid shell injection, kills the whole Codex process group on timeout or cancel, and caps output length. Requires Node.js >= 20 and a Codex CLI build supporting codex exec.

Turn rough ideas into precise prompts for 19 image/video generation, writing, code and text models, with multi-model enhancement, prompt scoring, image-to-prompt and a periodic self-updating knowledge base.

Add a "Delete session…" item to the sidebar session row menu that moves the session log and projection cache to the OS recycle bin after confirmation, then removes it from the workspace list. Windows only: on macOS and Linux it returns an error and changes nothing; a session loaded in the current run leaves the sidebar only after a DSH restart.

Turns a past session into a reusable Skill offline: it extracts the verified steps, the files that were edited, failure/fix pairs and the final checks, each tagged with its session event number.

Quota takeover for DeepSeek Harness on top of Our Free Model: compiles your platform order and pinned model tiers into one ordered fallback chain, skipping platforms that lack the model, sit in cooldown or are out of quota, writes the result back as Our Free Model combos, and after a quota or rate-limit failure re-plans and injects a follow-up continue message into the interrupted session so the turn resumes on the next platform; the chain ends with your configured aggregate fallback model or the official DeepSeek quota.

Floating card above the composer with the conversation's spend, the account balance and the current peak/off-peak billing period, plus a per-answer cost figure under every finished turn.

Native UI cards instead of Markdown: the assistant calls render_ui and the answer renders as a card in the conversation — 29 block types (metrics, tables, charts, timelines, kanban, flow charts, clickable actions), a content auto-layout mode, four placements (inline, process, dock board above the composer, turn-tail next-step card), a schema-backed settings form, and an optional automatic next-step card. Cards are rebuilt from the recorded tool arguments, so they survive replay and fork; 253 offline tests, no build step.

Prompt optimizer inside the DSH composer. One click rewrites the current draft with the model you already selected: it classifies the scenario (paper / agent task / daily question / English writing), returns a compact three-check diagnosis panel (is acceptance testable? missing information or capability? raising depth or widening scope?), and lets you pin a scenario manually. The draft is replaced in place, with no extra chat turn.

Second-model gate on the assistant turn, installed with one command: the bundle installs the settings panel and mounts the engine out of the same package, so both halves come up together. A separate judge call scores the draft against the numbered criteria (0-10 each); any criterion under 6 fails the turn and sends it back for a rewrite, up to 3 rounds by default. Ships in dry-run mode, which only records verdicts to an audit log until you switch it off. The web panel edits the criteria, the thresholds and the judge prompt.

Adds a "Handoff session" item to the session context menu: pick an older session, choose how the new one should read it, and a fresh session is created with the handoff instruction pre-filled in its composer while the old session stays untouched. Three read modes — the AI decides what to read, a six-part summary written into a .md file, or just the last 20 turns. The second tab turns those .md files into a cleanup list grouped by workspace, deletable once the receiving session has run 3 turns or the handoff is 7 days old. All file, record and turn-count operations run in the host process; the browser half only renders and requests. Requires Windows and the DSH desktop app (it uses Electron shell.openPath and explorer.exe/cmd.exe).