Плагины

Folds long composer drafts into a compact summary card once a character threshold is reached, without changing message submission.

Tool-health sentinel for DSH: cross-session tool success/latency/error history, chronic-failure warning injected into the system prompt, report tool + /tool-health command + GET /dsh-tool-health/summary.

Folds streaming reasoning and tool calls into collapsed step bars and collapses each finished turn into a single summary bar; step bars show native tool icons, filenames, edited line counts, and failure counts, finished turns restore duration, first-token time, token usage, generation speed, and cache hit metrics from recorded Session events, and preferences persist on the Host. Changes browser presentation only. Requires DSH 0.1.5-alpha.1.

Decision log for DeepSeek Harness: record key decisions (chosen option + why) into a versionable DECISIONS.md via decision_log / decision_list / decision_audit tools, and auto-inject the summary into future sessions so decisions stay traceable and are never re-litigated.

Регистрирует /wrapped, который читает собственную статистику сессии и отрисовывает доступную для публикации тёмную SVG-карточку-сводку — время инструментов против времени размышления, скорость декодирования и вердикт по соотношению инструментов и размышления.

Детектор качества компактификации / галлюцинаций: когда фрагмент разговора компактифицируется, проверяет по двум осям, что ключевая информация сохранилась — ТОЧНОСТЬ (утверждения в сводке должны дословно встречаться в исходных листовых событиях, иначе помечаются как сфабрикованные) и C

Нагрузочное тестирование с Grafana k6: сначала смок-тест, затем запуск скриптов и чтение процентилей из JSON-сводки.

Отслеживание баланса учётной записи DeepSeek и стоимости моделей с официальными тарифами пик/внепик, трёхуровневый биллинг токенов (промах кэша, попадание кэша, вывод), разбивка по моделям и сессиям, сводная панель в поле ввода с деталями при наведении, экспорт в CSV и панель настроек.

Глобальный опытный плагин DSH (скелет): автоматически обнаруживает локальные зависимости/удобные пути/сеть/диск и оседает их как межсессионный профиль окружения, подписывается на session/event для свёртки статистики по сессиям, внедряет малозатратное резюме в системный промпт каждого хода, инструмент env_profile запрашивает полный объём по запросу — уменьшает повторные обнаружения, ускоряет сессии, снижает расход токенов.

Локальное структурированное зрение для текстовых агентов: изображения уходят в локальный OpenAI-совместимый VLM и возвращаются как JSON-доказательства (сводка, дословный OCR, области вёрстки, сущности и связи, цвета, явная неуверенность), с защитой от галлюцинаций и необязательным мостом вставки и загрузки; без облачных затрат, изображения не покидают машину.

Предоставляет инструмент bili_summary: метаданные видео Bilibili, таймлайн субтитров и захват кадров на основе sharp со встроенной обработкой ошибок и плавной деградацией.

Режимы отображения для deepseek harness, включая Verbose, Normal, Summary Mode

Монитор использования DeepSeek для DeepSeek Harness: запрос баланса аккаунта и сводка использования токенов в виде инструментов, вызываемых моделью.

Compaction integrity: directive lines (never/must/必须 style markers) extracted from the compacted range are verified against the summary — verbatim or by content-word coverage — and every constraint the session silently forgot is reported and written to a JSONL ledger; pinned constraints from config are checked on every compaction.

Plays a voice alert and a browser notification when a turn finishes, asks for input, or fails, and can push the same summary to a Feishu bot webhook as a card.

Lets the model actually see a video: local ffmpeg samples timestamped stills and faster-whisper transcribes the audio track, so the agent reads real pixels frame by frame instead of trusting a text summary. Fully local, nothing uploaded; reports honestly when the audio has no speech. Windows only; needs Python 3.10+ and ffmpeg installed, and the first video_eyes_setup run downloads about 700 MB (faster-whisper and its model).

Share omp (oh-my-pi) project memory with DeepSeek Harness: DSH injects the same memory_summary.md and learned.md window omp reads, and its learn tool writes back to the very same learned.md. Conservative by default - never deletes, refuses to evict when omp's 100-lesson cap would be hit, byte-faithful writes with a CAS re-check, and it will not create omp's memory directory on its own. Verified black-box: one prompt sent to omp headless and to a one-shot DSH task yields the same answer, while a control run without memory abstains on both sides.

Token usage dashboard and per-session tree statistics: provides time-range filtering, 6 summary metric cards, GitHub-style activity heatmap, model stacked-bar charts with cache hit rate curves, and Top-5 model/provider rankings; enhances the native session status bar to display aggregated token usage across main agents and all subagents with detailed breakdowns.

Progress caption above the composer while a turn runs: the step number, what it is doing (from tool calls, or a local 17-rule pass over the thinking text — never a model call), and running / succeeded / failed with a plain-language failure reason, including host-reported turn failures and retry backoff. Adds a per-turn summary line and a settings sheet for size, colour, animation, the Claude-style marker and a thinking sub-line. No network requests, no model calls, no writes to DSH data.

Folds the durable todo and goal events the dsh pet ignores into a percentage bubble on the pet, and falls back to a one-line activity summary for a session that never wrote a todo list. Requires the dsh web pet plugin (@linxin666/dsh-pet >= 0.4.0).

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.

Daily and weekly task manager: startup popup reminder, a marquee strip above the composer, one-click circle completion, editable task cards, and a Sunday AI weekly summary, plus a sidebar DeepSeek API usage widget with balance card, today/cumulative tokens and cost, per-model breakdown and a 7-day chart. Persists as JSON under $DSH_HOME.

Generates release material from what a local repository can actually support: a hero image, a feature grid and bilingual copy rendered as SVG and HTML from `README.md`, `package.json` and optional screenshots. Facts are quoted from those files — name, version, license, summary, features, install snippet — and a field that is missing is left unrendered rather than filled with placeholder text. Every feature carries a source citation such as `README.md#L42`, and a claim the repository cannot substantiate is linted out before anything is written, with the reasoning left in `lint-report.json`. `manifest.json` records a sha256 for each input and output, and `--now` or `SOURCE_DATE_EPOCH` makes a run byte-identical. Also ships a GUI panel that regenerates from inside a session.

Cross-session long-term memory: a hybrid-retrieval memory store (BM25 plus CJK bigrams), chunked long-document search, chat-log extraction and distillation, and context-summary archiving and browsing, with important memories and the latest summaries auto-injected into the prompt.