Plugins

Available on npm as dsh-doudizhu-lan. Three-player LAN Dou Dizhu (斗地主) played from a floating window: one machine opens a room on its own port, the other two join with a six-character room code plus an address. Empty seats can be filled by computer players in four switchable strengths (easy, medium, hard, hell), so one or two people can still play a whole hand; the deal starts as soon as the three seats are filled. 54 cards in four suits, bidding of 1/2/3 or pass, the landlord takes the three bottom cards and holds 20, play runs clockwise, and the first-bid seat rotates clockwise between hands. The rule engine is pure and covers 11 legal shapes, with big joker > small joker > 2 > A > K down to 3, suits never compared, straights, pair-straights and planes excluding the 2s and the jokers, and a rocket that beats everything. The table itself is built to be read at a glance: three seat cards side by side in play order (you, then your next and previous seats) with role, hand count and a bar that runs out, one line that says whose hand is on the table and what would beat it, the bottom cards as text, an in-place play history that folds into a single summary line and unfolds into every trick including the passes, plus hint and auto-play buttons that pick the cheapest legal answer out of the local rule enumeration. Nicknames can be changed mid-hand (twelve characters, duplicates get a suffix). Computers run inside the authority and are handed only their own hand, the public record and the seat counts, so no hand can leak; the authority is the only place that decides which card ids leave it. Computers are driven by the room itself, so a one-person table keeps playing with nobody else on the LAN port. Humans have no turn timer, the collapsed floating pill can be dragged around the screen, and 179 unit tests pin the rules, the wire format, the theme tokens, the drag gesture, the history grouping and the computer seats.

Registers a janitor_scan tool that returns a deterministic, reviewable cleanup plan for session debris (plan docs, scratch scripts) with per-file verdicts and evidence; scan-only, never deletes.

Linux desktop notifications and a chime for DeepSeek Harness, sent by the host process so they still arrive with the browser page closed: turn completion, pending questions, tool approvals and turn errors, delivered through notify-send and paplay with a canberra-gtk-play fallback. A browser half adds an in-page toast, a Web Notification and a synthesized Web Audio chime as fallback; every switch lives in a Settings page or a JSON file.

Turn-handoff reminders in the DSH Web UI with a modal or top banner, six browser-synthesized chimes, configurable auto-dismiss, and a sidebar speaker toggle with an unread session list.

Shows a focus-free notice card in the corner when a turn ends or DSH is waiting on an approval or a question.

An agent preset (Hermit/小寄) that turns DSH into an always-on personal hub: a scoped butler persona, a dispatch tool that delegates multi-step tasks to background sub-agents with per-task model routing, async completion reports woven back into chat, plus a credit budget gate, a P/M confirmation chain and an M1 event log.

Converts Office and PDF files in the DSH workspace to Markdown locally — no network, no API quota — before the model reads them. Writes each .md next to its source so only a file path enters the context, and returns byte size, line count, token estimate and a heading outline, for a single file or a whole directory.

Registers one `ocr` tool for text-only models: it runs the offline script bundled in the repo over a single image or PDF and returns the recognized text — Windows native OCR first, then RapidOCR, then Tesseract. Recognition runs locally with no network request and no API key. Needs Python 3.11 on PATH plus `pip install winocr pillow` (winocr is the preferred engine; RapidOCR and Tesseract are fallbacks). Per-line results for images, per-page results for PDFs.

Complementary web-search provider: every web_search merges the DeepSeek Search API (primary) with serper.dev Google results (secondary) at a fixed 2:1 ratio. Both sides are direct HTTP retrievals with no auxiliary model turn: URL-deduped sources, long DeepSeek summaries clipped into snippets, per-backend timeouts and graceful degradation to whichever backend answers.

Rewinds a session to before a chosen user turn, from a button beside Copy under your own message — the model-visible surface is replaced in place and the interface hides the rewound turns, without patching DSH source or deleting the original session log.

/compact and auto-compaction summarize with the model selected in the session, not the model that finished the last turn.

Session budget gate for DeepSeek Harness: prices every turn against the official price table, warns at 80%, stops the task at 100%, and writes a handoff snapshot so the next session can continue the work.

Shows DeepSeek API balance and per-turn conversation cost directly in the session UI.

Read-only clinic for DSH session logs: 22 named diagnostics (token waste, retry and failure loops, context bloat, cross-session amnesia, idle MCP servers and skills), each reported with a replayable evidence locator such as session.v4.jsonl.zstd#L101-L110, and each finding turned into an executable prescription — a profile-patch YAML snippet or a command with current to suggested values, risk, verification and source — plus a feedback command that retires a prescription you reject, after which the report states it was removed by your feedback. Zero dependencies, default zero egress, session files are never written.

A prompt optimizer in the composer input tool row: pick a category and a template, then ✨ rewrites the draft in place with one-click undo. Pick from the Basic, Context and Image categories; you can cancel a run, and it shows elapsed time and token usage. Ships 15 templates covering task-instruction rewriting, step planning, safety-research framing, system-prompt role cards, and text-to-image / image-to-image prompts; context templates read the recent turns of the session for coreference resolution. Calls go through the host model service, and a settings section exposes the model and the budgets. Output guards check that user-written constraints survive, warning rather than auto-retrying. Adapted for dsh 0.1.7 and 0.2.

Open and copy local files and web links from a chat turn.

Adds a live status badge to the session header and delivers guidance added mid-task without interrupting the turn.

Input prediction: the likely continuation of your draft appears as grey ghost text after the caret and Tab accepts it, and after each assistant turn 1-3 clickable follow-up suggestions appear above the input box and send on click. Ghost completion, suggestions and the suggestion count can be toggled in Settings, Built-in plugins.

Conversation navigation with full-history turns, collapsible process steps, final-result and structured question-and-answer keys, and pointer-following piano interaction.

Subagent model control: a Jev classifier labels each turn, and the subagent request itself is rewritten to the provider, model and reasoning effort that the label maps to, at the agent/request hook; a per-session inherit/auto/manual selector, one prompt per role injected ahead of the task, and offline replay of past verdicts.

Sends a macOS Notification Center alert when a session finishes or the agent is blocked on an answer, an approval, or an error. The banner carries the last reply and plays your system alert sound, and clicking it returns you to the DSH page.

Warm pixel theme for the DSH Web UI: cream-paper and arcade-phosphor palettes, hard-edged pixel chrome with bundled pixel fonts, CRT motion on streaming and state changes, three intensity tiers, and a pixel loading bar while a turn runs.

Turns an image folder into a frosted-glass global wallpaper for the DSH web UI, keeping the official palette.

Tracks this turn cost under the input box, today tokens and spend in the sidebar with peak and idle pricing, and pins the latest question to the top of the conversation.