Plugins

A DeepSeek Harness plugin that freezes tool declarations to improve prompt cache stability.

A social deduction game played against your own model configurations. A table has N seats (six by default, one of them yours) and some of the rest are played by agents bound to different provider, model, temperature, system prompt and disguise settings; you interrogate them and cross-check the answers to work out which seats are not human. Seat identity and each agent model configuration stay hidden in the host state until a seat is judged, then the disguise parameters are revealed. Every round exports a per-seat conversation JSONL plus the judgment record, and the same seed with the same move script replays byte-identical, so runs with different models or disguise strengths can be diffed.

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.

Deterministic prompt-injection interception: tool results from file reads, repo searches and web fetches are scanned before they enter model context; high-severity findings quarantine-wrap or block, and credential-material files quarantine on the path alone.

Tool-surface snapshot and drift gate: baselines the full visible tool registry (name/description/parameters, MCP tools included) into a committable JSON file and warns loudly when an upgrade adds, removes or reshapes a tool.

Watches a turn's thinking and output streams and cancels the turn when the model starts repeating itself. Four detectors: a verbatim cycle, a counting loop whose lines differ only by a number, repeated lines separated by other text, and a user-supplied phrase list. A catch raises a manual-dismiss alarm card in the global overlay with an optional tone, and a Monitor tab holds the switches, thresholds and hit history. Detection is local string comparison with no model calls, so it costs no tokens.

Turns a streaming assistant reply into short blips, one per character at a configurable minimum spacing, for the reply text only, and keeps sounding when the window is minimized, covered, or in a background tab. Two built-in voices, or a custom one shaped by waveform, or loaded from an audio file.

Adds an Omni-modal settings section to DeepSeek Harness that points the harness at any OpenAI-compatible image and video endpoint: switchable config profiles with separate endpoint, API key, model, path and size for each side, free-form extra request parameters, a connectivity test that does not spend generation quota, and three model-visible tools — generate_image, generate_video and check_video — with async video tasks that survive restarts.

Code retrieval engine exposed as an MCP server: four tools (code_search, code_outline, code_index, index_status) over a SQLite FTS5 index with BM25 ranking, CJK bigram indexing, declaration-aware tree-sitter chunking, incremental reindexing, and a three-layer secret filter. The bigram index matches Chinese-language queries against English code, and the default path needs no vector store or embedding model. Requires a local Python 3.10-3.12 engine installed by scripts/install.sh, and CODERAG_ROOT set in the profile patch.

Normalizes every tool schema before it reaches the model, rewriting it into the strict OpenAPI 3.0 subset that Gemini function declarations accept — a type on every node, items on every array, no empty enum entry, $ref inlined — so a single offending MCP or built-in tool no longer fails the whole request with HTTP 400.

A context usage ring with configurable character-based automatic compaction, manual compaction, protected user requirements and recoverable original history for DSH 0.2.0-rc.2.

Yukimi Sajo theme with illustrated backgrounds, interactive characters, and an optional Windows desktop companion.

The Quick Control panel core for DSH Web: a standalone floating shortcut panel whose switches other plugins register through the quickControl service, with the skin system and zh/en i18n. The dock-base Workbench integration ships separately as the dock-flash adapter.

In-place Markdown editing for the right-sidebar document preview: a file keeps the shipped renderer's look, clicking a paragraph, heading, list, fence, or table cell puts the caret exactly where you clicked, and only the edited range is written back through ctx.fs.writeText, optionally guarded by the version it read.

Plays an intro full-frame when you open DSH, when you open a new conversation, or every time you open the one you pinned. The built-in footage is themed on a Wandering Earth-style space station and the quantum AI MOSS (550W): a red single eye ignites while the Earth’s curve slides past the viewport. Four clips ship embedded in the code — no separate video files — and the library lets you tick any number to rotate at random (never the same clip twice in a row), preview the rotation, or swap in your own videos. Sound plays without a click; the pin button sits beside Settings.

Select text in the DSH web client and a translate button appears at the end of the selection; clicking it translates that text in place. Translation goes through the host llm service, so it reuses the model the current session already has selected and needs no separate API key; an OpenAI-compatible endpoint can be configured instead. The result opens in a popover that stays inside the viewport, with one-click copy, a row of language buttons for instant retranslation, and smart Chinese-to-English / other-to-Chinese as the default. Configuration is stored host-side and its API key is never sent to the browser.

Replaces the chat model picker with a collapsible provider-grouped list (640x720 max) that shows each model name, provider name, provider ID and exact model ID, plus a manually managed uncommon list and a separate favorites list. Classification is user-driven only: no usage tracking and no auto-sorting. Search covers model name, model ID, provider name, provider ID and account notes.

Adds two visual pages to DSH Settings. MCP servers are kept as a JSON source of truth that rewrites the profile patch managed block, with YAML validation and timestamped backups, plus import of a pasted mcpServers JSON that first lists what will be added or overwritten. Thinking unlock probes each OpenAI-compatible gateway for the reasoning efforts it really declares and writes the per-level config back while leaving other hand-written YAML byte ranges untouched; the available model list syncs from a separate button that previews additions and removals, then writes levels and models in one atomic pass, and can also run on the page schedule.

Customs declaration document consistency check: field-level findings with rule IDs.

Discord bridge loaded into the same Host as the desktop app or dsh web, with no separate process or profile — chat with sessions in any workspace from DMs or threads, attach to sessions open on the desktop, and answer tool approvals and model questions from whichever side responds first.

Session-scoped delivery receipts for local artifacts: compare current file hashes with a recorded baseline, keep agent claims separate from explicit user acceptance, flag changed or unavailable files, and export a Markdown receipt through a tool, slash command, and conversation panel.

Registers a Tavily search provider for the ctx.web seam and switches web_search to it, with a comma-separated apiKey list that rotates on 401/403/429/5xx and transport errors.

Low-token web search that returns only {title, url, snippet} per hit — vendor extras such as answer, score and raw_content never enter the context. Five interchangeable backends (Tavily, Brave, SearXNG, Google CSE, Exa) with automatic failover to the next usable one in fallback / chain order when the primary fails, deduplicating repeated providers. Returned URLs and snippets are sanitized of tracking parameters.

Cross-session and cross-project long-term memory: a non-blocking silent sedimentation dispatch on turn/end writes into a local SQLite store (FTS5 trigram full-text index with insert/update/delete sync triggers) that later turns recall from; entries are scoped per project alongside a global tree, and the plugin ships its own embedded web dashboard (it starts the HTTP server itself, no separate process) for browsing, graphing and controlled deletion.