Plugins

A full-viewport donkey-mill background for the DSH Web UI: an animated top-down donkey or a Chinese-character form circles the millstone, accelerating with the model's live token throughput and plodding slowly when idle; a bottom-right panel switches form, adjusts opacity and hides the scene, persisting in localStorage.

DSH plugin: lifetime token counter in the conversation header, summed across every local session.

给 dsh web 部署加一个右上角电源按钮：刷新 / 重启 / 定时 / 关闭。重启后自动把新 token 的链接打开——dsh web 每次启动都换 token，不打开就得自己去日志里翻。macOS 走 launchd、Linux 走 systemd、其它平台自己 respawn。

Manage a LAN or Tailscale reachable dsh web instance from inside DSH. dsh web listens on loopback by default; this adds a Settings > Web access panel that detects the machine's Tailscale and LAN addresses, starts dsh web bound to the one you pick, and shows the token URL to open on a phone or another machine. Also stops and restarts it, tails the child log, persists the bind address and port, and registers two agent tools (dsh_web_control, dsh_web_config). Zero runtime dependencies. Anyone who can reach the bound address and holds the token URL gets your dsh's full privileges (files and commands); LAN mode is plain HTTP, Tailscale is encrypted by the tunnel, and the token is single-use and changes on every start.

Multi-machine job board for DSH: 14 board tools, 4 shell tools, a GUI panel, and a dispatcher that hands queued tasks to remote hosts through dsh-remote-agent. Pick the target machine and model from the panel or the board_options tool; rename machines with display aliases while task ownership stays keyed to the host identity (machine). Failed tasks can be retried with their original goal/workspace/model, and the panel checks npm for a newer plugin version. Single-machine use starts its own loopback board; multi-machine use needs the bundled Python board server (tools/board_server.py, started with GP_BOARD_TOKEN set, binding 0.0.0.0 by default) and dsh-remote-agent hosts. The sh_* tools only reach directories on the allowedRoots list.

Moves reasoning that a provider leaks into the visible text channel back into the reasoning channel, and emits a text block when a whole turn of answer text lands only in the reasoning channel. Catches three shapes: `<thinking>`-tagged text, text that continues a reasoning tail which ended mid-token, and an answer-only reasoning turn that stops. Healthy streams pass through unchanged, block order is preserved, and each of the four rules can be switched off in one JSON config.

Carries a DSH environment between machines through one private GitHub or Gitee repository: third-party plugins (reinstalled from a version, or carried as a tarball or source tree), the four profile configuration files, and folder snapshots. A restore resolves each folder through an absolute path, a registered root, or a confirmed new path, compares file by file, leaves locally edited files alone, backs up anything it replaces, and refuses writes that escape a root, cross a junction, or land on a directory the plugin does not own. Configuration files and folder contents are uploaded as they are (no secret scrubbing) and the plugin does not check that the repository is private; the access token stays in the credentials service, and on first start a GIT_SYNC_* token left by dsh-plugin-git-sync is copied over. Needs Node 22.12 or later.

GitHub Trending masonry feed in the DSH sidebar: daily, weekly and monthly boards with category chips, concise project intros, and one-click starring that syncs to your GitHub account or saves to a local collection. It reads github.com and raw.githubusercontent.com, sends README excerpts to your current model and text to Google Translate when a translation is needed, and stars with the GitHub token you connect (or GITHUB_TOKEN / GH_TOKEN); its HTTP routes have no DSH access check.

Continuous voice conversation for DSH: auto sentence splitting, a 2-second confirm window before sending, and spoken replies you can interrupt by speaking. Speech is recognised by DSH's own speech service by default (a local Whisper bridge or an OpenAI-compatible endpoint can be set instead), the recognised text is corrected by a DSH model call, and replies are read aloud through Microsoft Edge's online text-to-speech service, so reply text is sent to Microsoft. A daily check of PyPI and Hugging Face for faster-whisper updates is on by default; settings and any API keys are stored in plain text in ~/.dsh/voice-tutor-config.json. Its /voice-tutor routes require an authenticated DSH session (admission is checked per request), API keys are never returned, and applying an update needs a one-time confirmation token.

Configuration bundle that connects Microsoft's official @playwright/mcp to your running Chrome through the Playwright extension, so the agent can drive your logged-in tabs. Install from a local copy: put your extension token in cordis.patch.yml first (with the token, connections skip the extension's approval prompt, so keep it private). Needs Node.js and the Playwright Chrome extension; the default command is the Windows form (cmd.exe /c npx), so change it to npx on macOS or Linux. It starts @playwright/mcp@latest through npx.

Experience and skill engine with a sidebar console and exp_record / exp_recall / exp_feedback / exp_consolidate / exp_type tools, stored in a local JSON file. Optional Git backup (off until you configure a remote) pushes the library to a remote you choose; GitHub sign-in uses a device code or a token kept in local settings.

Dispatch a task to a real DSH session by workspace path or session id; results land in a persistent task ledger that can be queried without triggering the model. It serves HTTP on 127.0.0.1:7391 and requires a bearer token it generates on first run (kept in ~/.dsh/relay/token), rejects non-loopback Host and foreign Origin, and stores task results as JSON files under ~/.dsh/relay/tasks; an optional notifyUrl receives each finished result.

Replaces LLM summarization with deterministic text compilation, so compaction makes no model calls and keeps original tokens; elided content stays retrievable through seq pointers. Based on TsFreddie/dsh-compaction-instant.

Folds older oversized tool results into spill files and replaces them with short, deterministic stubs at step boundaries, with no LLM call: the stub keeps the first 200 and last 80 characters plus the sequence number, tool name, line/char/size counts and the original file path, and each fold is a real surface replacement accompanied by compaction/prune shadow pricing so token counts stay honest. Triggers on context pressure at 25% of the model window (falling back to 32768 tokens when the window cannot be resolved) or on five accumulated candidates; the last three turns, error results and rich content are never folded. The original is recovered from the session log rather than the spill file, so /folds, /unfold <seq> and /unfold all keep working after spill cleanup; session_fold_recall lets the model fetch a folded result by sequence number or by keyword, and session_fold_diagnose reports the effective settings. /spill-prune removes orphaned spill directories, preview only unless --apply. A Settings page exposes 12 of the 13 options as draft-then-save fields; they are volatile, so saved changes apply without restarting the plugin.

Puts a domain bar under the session header — a domain groups skills and tools and toggles as one block, so one click loads or unloads every member. Hovering a chip lists its members (skills, then tools) in a height-capped panel that scrolls. The right end of the bar is a compaction-threshold counter (− 5 +, 100k tokens per step) that rewrites the live compaction engine policies, so the compaction trigger line changes on the next turn without restarting dsh; a degraded fallback turns the number amber and is reported in the snapshot. Settings maintains domains and members with search, skill/tool tabs, all/selected/unselected filters and a copyable AI grouping prompt. Tool removal uses tools.restrict plus a guard; hiding a skill re-registers it with modelInvocable false, so it leaves the model catalog while /name still works. Sessions of another preset are left untouched.

Read-only session dashboard for DeepSeek Harness over the local ~/.dsh data: a sidebar overview of every session (tokens, estimated cost, turns, tool calls, model mix, daily activity, searchable and sortable session table) and an in-conversation Dashboard view that shows the current conversation turn by turn with per-turn tokens, cost, timing and tools. Zero dependencies, no network.

A Desktop 0.2.2-rc.1 adaptation of DSH WebTokens that registers a signed-in Chrome DeepSeek web session as the deepseek-web model in DeepSeek Harness, with tool-call transport, reply validation, a Chrome bridge extension, plus install and rollback scripts. It cannot be installed with the one-line plugin add shown on this page: it needs its own installer (node scripts/install-desktop.mjs, which backs up the whole profile first). Prompts, project content and tool results are sent to the DeepSeek website through your Chrome session; the author has tested it on macOS only.

Runs a four-task, 150-point model intelligence benchmark inside DeepSeek Harness (Qin 60, Pompeii 45, Pelican 22.5, Candy 22.5): each selected model × reasoning-effort pair executes every task in its own isolated subagent session, Qin and Pelican are graded by the selected judge model against rubrics, Pompeii and Candy programmatically, and batches keep scores, artifacts and cost. Adds a 0–1000 real-world performance score per model × effort computed from local session logs across seven dimensions (output speed, first-token latency, reasoning speed, tool cadence, tool latency, tool success rate, cache hit), with baseline deltas and controlled speed runs, plus a sidebar page with ranking, performance, history and settings views. Each batch spends real tokens of the candidate and judge models, and candidate subagents get file and bash tools inside an isolated run directory.

Adds a recap card under the composer for the current session: turns, steps, tool calls grouped by name, the uncached / cache-read / cache-write / output token split, the cache-hit rate, and a cost range computed from the official Flash off-peak and peak list prices. Exports the same card as Markdown or a standalone HTML page.

Gives the agent its own context usage through a context_usage tool plus one reading line per turn, with an editable template.

Provider-aware account balance chip for the DeepSeek Harness composer dock: DeepSeek, OpenRouter, Moonshot/Kimi, Zhipu/GLM, MiniMax and Alibaba Cloud Token Plan, with peak/off-peak and night-discount countdowns.

An Eisenhower-matrix to-do panel for the DSH sidebar: importance × urgency quadrants that fill the panel, inline quick-add with a hand-drawn due-date calendar, drag to reorder or move between quadrants, a draggable center handle for quadrant proportions, one-click switch to the completed view, and random encouragement toasts that tell praised on-time finishes apart from overdue ones. Minimal styling uses only DSH theme tokens; bilingual UI.

A draggable task progress panel with local estimates, independent teammate progress, optional AI observations, and probe cost diagnostics. Release candidate for the Web profile only; needs Node 24.9 or later and Harness >=0.2.2-rc.1, tested on 0.2.2-rc.2. AI observations are off by default and, once enabled, spend tokens of the session's provider. It uses the same package name and mount id (dsh-task-progress) as chen8923/dsh-task-progress, so install only one of them.

Live monitor of a local Strata model server inside the DeepSeek Harness GUI: state, token rate, VRAM, GPU and CPU temperatures, and the last request, next to the message composer.