Plugins

Fills in the vision capability that DSH's "fetch models" step drops, so image-capable models stop being declared text-only, and marks them in the model name.

Explicit, project-scoped memory handoff: when you open a new conversation, one model call that you confirm curates the project's earlier conversations into a bounded brief and injects it into the conversation's first step. No silent background calls; a Memory panel shows the brief and lets you edit notes.

Full-screen ice-and-fire boot intro for DSH: a frost-to-ember boot sequence with typed bilingual logs and a hairline progress line, then Ygritte's five-step identity authorization. Esc skips, Enter or click fast-forwards, and a settings page switches it off/quick/full or replays it; the host half paints the opening frame while the document parses, so the built-in loading card never shows.

OpenFOAM case doctor for DeepSeek-Harness. Six agent tools: audit a case for 25 classes of silent failure (steadyState paired with timeStep writes, limiters set to 1 that let k/epsilon go negative, unbounded advection of transport scalars, nNonOrthogonalCorrectors 0 on a skewed mesh, loose solver tolerances, blockMesh vertex/block mismatch, stray time directories, purgeWrite 0); read any dictionary as JSON; survey a sweep of cases; run checkMesh/blockMesh/snappyHexMesh/foamRun inside WSL with log parsing; parse logs into the metrics that matter (aspect ratio, non-orthogonality, skewness, residual floor, fatal errors). Deterministic text analysis — no LLM — validated against a real 82k-cell street-canyon snappy case.

Beta: a project supervisor for DeepSeek Harness — standing rules plus a Web panel (three switches, dispatch modes) that splits a goal into parallel teammate dialogs, checks engineering baselines, interviews new-project requirements and protects beginners. The standing rules come from the repo kit installer, an optional separate step.

Drives the DSH right-sidebar built-in browser from the agent side with 38 tools (open, navigate, read page text and interactive elements, accessibility tree, click, fill, select, check, key press, scroll, resize the browser frame by dragging its corner, with the page inside reflowing responsively so a narrowed frame shows the mobile layout, screenshot, full-page screenshot, PDF, markdown export, scrape to JSON, download, cookies, per-step operation records with replay and action whitelisting), serves the built-in web_search from a self-hosted SearXNG over an SSH tunnel with Tavily and a real-browser fallback, and registers a fetch provider that retries blocked GitHub URLs through measured mirrors and sniffs charset from the response header so GBK and GB2312 pages decode correctly.

Connects a ComfyUI server to the agent through five draw_* tools (status, workflows, models, generate, reverse) and a sidebar panel. It ships no workflows and assumes nothing about your models: every API-format JSON you drop into the workflow directory is inspected to find where prompt, negative, width, height, steps, cfg, seed and LoRA live, and a sidecar file can override what was detected. LoRA injection splices into any graph. All paths are configuration, editable from the panel and saved to a settings file, so it works against a remote or already-running ComfyUI. Headless auto-start and idle shutdown are opt-in and off by default.

Drives a local MaaEnd (MXU) client over its loopback HTTP API: 10 maend_* agent tools, a sidebar dashboard panel, and startup-mute / success-gated close-game steps in the task chain. Requires a local MaaEnd install with its MXU service running.

Adds a connection-test panel under Settings -> Models: reads the deployment's model catalog and probes each configured model with one minimal chat round-trip (a short "ping" with a small output cap) through the Host LLM service, verifying connectivity with a real request. Each row reports success/failure with the harness failure code (NO_ADAPTER, MISSING_CREDENTIAL, AUTH, RATE_LIMIT, TIMEOUT, ...) plus first-token/total latency; single-model and concurrent test-all (3) are supported. Credentials are resolved inside the Host process and never reach the browser; probes are independent per model behind a same-origin-fenced route with a body-size cap. Pure JS, no build step, MIT.

A transparent always-on-top desktop fish that mirrors the current DSH session - turn, step, tool calls, todos, background jobs and team tasks - with optional login autostart, an option to stay after DSH exits, and a right-click hide. Windows only.

Adds a sparkle button beside the model picker that rewrites the current draft into a spec sized to the draft — two or three steps for a small ask, a longer checklist only when the request has several parts — and puts the original text back when pressed again. Greetings and small talk pass through untouched without calling the model.

Turns the composer send button into a tomato button while a task is running and the draft is empty. One or two taps hand the model a hidden note on its next step asking it to reflect on the approach and to finish by asking the user how to continue; three taps within a second stop the run first and send the same kind of note. Thrown-tomato animations and sound can be switched off in settings, which also counts every tomato thrown.

Enlarges the DSH web settings dialog (default 1180x920, 20px steps, remembered per browser) and adds an Advanced Settings page that resizes it and hosts Host operations: restart dsh with a 3-second delay and a busy confirmation, restart once all sessions have been idle for 10s (cancellable, repeat clicks idempotent), last start time with uptime, and a dsh version/update check that compares npm dist-tags across selectable channels without upgrading. The same bundle replaces the sidebar brand area with a machine name, adds a read-only provider, model and thinking-level cell beside the composer stats pills, and a switch that registers or unregisters the three dsh_restart_* agent tools at runtime. Zero dependencies, no build step; all routes sit behind the connection trust fence.

Per-conversation fast mode for multi-step tasks that batches independent tool calls and drops redundant steps. Pooled over 9 blocks and 132 valid paired runs it is about 18% less total wall-clock time (0.825x, per-cell median 0.813x, 95% bootstrap interval 0.747-0.914) and 15-25% faster on the shapes it fits (codeimpl median 0.709x, faster in 13 of 15 pairs; shellpipeline median 0.844x), with hidden tests still passing. Short repairs and data reshaping do not pay for the reminder (debugfix mean 1.287x with output 1.29x longer, perfrefactor mean 1.140x), so try it once per task shape. The earlier "26-28%" claim is retracted.

Web mascot that acts out each live tool call (read, write, edit, search, run, plan, sub-agent, wait, error) with matching step cards in chat; ships two swappable character packs.

Cross-session memory for DeepSeek Harness: a three-tier store (behavior notes, deep details, session archive) plus a 600-character core card injected every step, work/companion/learning state lines, multi-person and multi-group records with visibility partitions, session-scoped suppression or deletion of memory fragments, summaries and fact candidates that can route to a separate model, and backup plus self-check tools. Data is plain text under the home directory, and entries can carry images and audio stored by content hash.

Per-step context ledger for dsh with opt-in gates: hides heavy tool groups from the tool table, externalizes long tool results to disk, and answers duplicate reads with a pointer instead of re-sending. In the author's A/B test the tool-table gate cut 16,110 first-step prompt tokens (-49.7%), and about -23% of per-step context on long sessions.

A deliberately small phone remote for a running desktop DSH: allowlisted session read/write, message send and permission approval, host-rendered pages with no build step, QR pairing, and tunnel-URL recovery.

Bidirectional GitHub repository sync for DSH: git sync with conflict rollback and pre-push checks, Actions/CI monitoring with per-step run detail, repository add/remove/rename, a web panel with local and remote file browsers, scheduled automation, and six agent tools.

Generic monitor plugin for DSH Desktop: an HTTP polling rule engine. Poll any endpoint on schedule (multi-step, variable extraction, JS condition expressions); when conditions hit, wake a visible agent session to act — zero model calls otherwise. Configure everything in a sidebar GUI panel: create/edit/pause/test/fire-now/run history, same UX as the built-in scheduler plugin.

Arms a watch on a directory and delivers a batched, deduplicated diff when the tree goes quiet, filtering out the writes the session made itself; delivery is per watch, either waking a new turn or queueing for the next step.

Teaching-first CTF solving skill: narrates each step's pattern, tool choice and underlying mechanism, then writes a reproducible writeup under solves/<challenge>/ gated by a nine-section check.

Write, check, and publish a Chinese Quarto Book: nine-step chapter writing, structure/link/figure checks, a consistent theme, and render acceptance. Loads only in type: book repositories; requires Quarto CLI and Python 3.

Windows x64 only, 0.5.4 preview: two scoped tools (dragon_browser and dragon_desktop) that drive a visible browser through the DOM or CDP and Windows apps through UI Automation, with per-origin and per-executable read/control scopes, batch steps, a physical-Esc emergency stop and a redacted action log. 0.5.4 only widens the DSH compatibility range to 0.2.0-rc.2 through 0.2.1.x, because 0.5.3 pinned it to an exact DSH version and was silently skipped on DSH Desktop 0.2.1. Two long-lived native processes remove the per-action process spawn, so a single key step takes about 0.7 s on the same machine; when another window steals the foreground the step is aborted instead of typing into the wrong place. The browser side ships its own Edge/Chrome launcher and CDP client, so it no longer needs another browser plugin or the Electron runtime. It is not an OS sandbox; logins, admin windows and secure desktops stay with the user.