Plugins

DSH hardware sandbox plugin: assemble a 3D scene of ESP32 / BME280 / PIR / LED components with protocol- and voltage-checked port wiring, run real Python project code in the loop against simulated I2C/GPIO hardware with fault injection, let networked device simulators report over HTTP to a real backend (validated end-to-end against a live seat-reservation system), observe a real device read-only and map its state onto the scene, and add support for new boards through data-driven pin definitions. Ships 16 hw_* agent tools plus a bundled SKILL.md.

A memory subsystem for DeepSeek Harness: full ontological memory (event/fact/rule three-ontology) plus a goal system (goal.js) — belief values written via Bridge trigger words, goal activation scored at read time (0.25 domain + 0.35 bridge + 0.40 sim) to recover goal/value long-range associations, journal ledger, episodic archive, low-confidence decay, memory_append/read/consolidate plus linghun_goal_update/get/state closed loop. Independent of the original dsh-linghun line. 16/16 on LoCoMo-Plus.

Long-term memory scoped to the DSH workspace. A session opens with a one-line-per-memory index - bodies are fetched on demand by the `memory` tool, so the same character budget holds about three times as many entries - and when a turn ends a purely local search over the last user message pulls the matching memories into the next briefing at no model cost. Optional auto-capture distils a stretch of conversation into structured memories; in draft mode they wait in an approval queue instead of landing directly. A portrait layer keeps a workspace and a global picture, rewritten from the journal and the records, with the same draft-or-direct choice. Settings → Memory holds the archive, list and timeline views, a health check (duplicate clusters, stale entries, broken links), the supersede version chain with rollback, and a read-only scan that imports existing WorkBuddy / Qoder memory files.

Cuts session input tokens three ways, each with its own switch in a panel under the composer: lossless structure-aware compression of tool results on `tools/post-execute`, then a per-tool head/tail elision that always names a recovery path through the shared spill store; tool-catalogue visibility per agent on `agent/created`; and a per-workspace memory document that recall matches against each new message and injects through `agent.inject`, so the transcript is not re-read. Measured 41.76% fewer prompt tokens over 149 recorded requests, with a bundled script that re-measures any profile.

Keeps the images you pasted or attached in the composer, per session: they are mirrored locally in the browser and handed back through the official draft channel when you return to the conversation or the composer remounts, and dropped as soon as the message is sent or you remove them.

Composer prompt enhancement for DeepSeek Harness: a button beside the model selector rewrites the draft through one auxiliary LLM call and writes the result back in place, with composer undo left intact; a /prompt-enhance slash command returns the rewritten prompt as a command result line without touching the draft; the enhancement call can be pinned to its own fast provider and model in the profile patch; the rewrite policy is faithful - original intent, language and information are kept, code, commands and paths are preserved verbatim, and nothing is invented or padded with placeholder text.

Binds a Feishu/Lark bot to DeepSeek Harness over a WebSocket long connection: incoming messages run through an agent session and stream back to the chat, chat commands such as /attach manage the live session, and feishu_* tools let the model push messages.

Run one deterministic skill-routing pass on every user instruction, so the skill catalog is consulted before the model answers instead of after it remembers. On the host it scores the live catalog from the instruction's keywords and the catalog metadata, with no LLM or embedding call, and appends a single producer-tagged message to the step: a ranked shortlist by default, the winning SKILL.md body once the score clears a threshold, or an instruction to call a named router skill exactly once. Unrelated instructions stay silent.

Trip-planning agent preset for mainland China that answers in Chinese: queries 12306 train availability, Fliggy and Tuniu flight and hotel prices, and real Amap transit routes, checks closures and weather, and delivers a single offline HTML roadbook with draggable maps, three hotel price tiers and an itemized budget. Ships 9 skills, an environment doctor that walks users through missing API keys, and a patch for the 12306 MCP server; it plans and queries but never books or pays.

File-backed post office between DeepSeek Harness sessions: every letter is sealed with an HMAC and a body digest, recipients are resolved through a roster interface, and delivery is offline-first, so an undeliverable letter waits in the mailbox instead of being lost.

AI companion with a learned body clock, real offline sleep, its own memory, and self-reported affect delivered through the same judge math as the pi version.

Prompt optimizer for the composer: a button in the tool row rewrites your draft through a model channel you already configured, streaming into a panel that floats above the composer without taking up layout. Your draft changes only when you click Accept — and Accept is refused while the answer is still streaming, so a half-finished reply can never be written back. It asks first when the draft carries @ references or attachments, which a whole-draft replace would flatten into plain text. No global keyboard shortcuts, and no API key ever reaches the browser: the host half makes the call.

Project-level memory built around a ledger whose structure the AI designs per project; the plugin only handles detection, injection, write-back, archiving and git. Memories are split by meaning: the active ledger holds current state, the turn journal is append-only, and decisions, the goal tree and lessons are one-card-per-file (with a multi-dimension index over the card folders); the fundamental principles file can be changed only by the user. There is a single compaction mode: before compacting, the AI sorts the work into three buckets - what must stay in context and is re-injected verbatim, what is written to the project's _stash/ for later lookup, and what is dropped. When a conversation is opened in a container workspace holding several projects, it can be locked to one sub-project (the lock lives in the plugin's own home and survives restarts; it leaves no marker in your project). Also an activity log with five levels, where L4 stores the user's own words verbatim (never injected, queryable on demand). Before each compaction a verbatim capsule (user quotes, raw errors, changed files) is written to disk, and both the capsule and its source material survive a restart; a handoff document lets a new conversation resume the previous one. 21 tools, 12 subcommands, 39 settings (one of them switches whether the settings panel or each project's manifest wins). Runs locally, no runtime dependencies.

Browser provider for DeepSeek Harness: hand it a web task — open a page, click, type, fill a form, read what is on it — and it drives a real browser step by step, seeing an indexed table of what can be done on the current page instead of raw HTML or screenshots. Each step is one bundled-Jev decision rather than a full model turn (1.1 s median per step versus 2.8–3.6 s for chat-model turns in the repo benchmark), and the main conversation receives only the conclusion, so its context does not grow with the page. Targets come from the page it just read and a changed page is re-read rather than clicked blind, and it runs in the browser that Session owns, leaving your everyday Chrome untouched unless you switch to attach mode.

Vision understanding + image generation/editing for DeepSeek Harness (read images with Qwen/OpenAI/Anthropic/Gemini, generate and edit with Qwen-Image). A rework of liustack/modlens (MIT). Remove @liustack/modlens first — both register the same vision providers and image interception hooks, so co-installation breaks image reading.

Adds subscription OAuth and provider Plan API-key logins through Pi adapters, using a DSH-owned credential store.

Persists unsent composer attachments per session, so images and files pasted into the input box survive page switches, conversation switches and app restarts: the bytes are mirrored in IndexedDB and handed back through the official createDrafts/addAttachments channel, and are dropped as soon as the message is sent or you remove them.

Grounded web search for DeepSeek Harness through OpenRouter-compatible gateways: the built-in web_search tool runs on the same endpoint and key as your chat models, returns citeable sources, and is configurable from the Plugins settings page.

DSH 插件市场：浏览官方与社区插件，装之前先给出兼容性结论，一键交给官方安装管线。Plugin market for DeepSeek Harness: browse official and community DSH plugins with a compatibility verdict, and install them through the official plugin manager.

AI-agent-driven automated testing platform for DeepSeek Harness (dsh): drives the harness agent through long-lived multi-round sessions to write and run Markdown test cases, file bug reports, fix the code and re-verify the fix, with a per-project serial queue, a development board and built-in load testing.

Context fullness bands, the per-session compaction threshold derived from the model window, per-step cost computed from cache read and cache write rates, cache-hit, DeepSeek balance with the peak/off-peak tariff, and a one-click handoff. It never calls a model and spends no tokens.

Bidi-safe RTL rendering for mixed Arabic and English text, a full Arabic interface pack (3,307 strings) registered through the DSH locale service, and three thmanyah cuts — Sans for the interface, Serif Text for reading, Serif Display for headings — placed behind whatever font the app itself is set to.

Read-only Skill inspector for DeepSeek Harness. It shows which Skills a conversation actually loaded — from observed loading evidence only — and which Skills the current environment can discover, and gives each Skill a detail page whose left column switches between the framework, the evidence, validation, evaluation, step evidence, run logic, the before/after of a modification when there is one, the definition facts and the raw SKILL.md. Declaration and run stay apart: the body is split into structure, declared workflow and progressive disclosure, the run logic is read from local receipts, and every declared step carries its own evidence — no observed evidence is reported as exactly that, not as executed or not executed, and the run's instruction hash sits beside the current definition as match, mismatch or unavailable. The evidence view gives each subject one of three states (declared, observed, unavailable), walks definition → load → use → outcome, names what each stage cannot reach, declines to attribute an outcome to the Skill, and can copy the snapshot out as byte-stable JSON. Validation checks the Agent Skills open standard and the platform profiles separately, each finding on a rule id and labelled standard compliance or platform compatibility. Evaluation keeps repeatable cases with the conditions each run was recorded under and the four stages of runtime evidence, comparing two runs of one case item by item without scoring them. The trigger boundary keeps should-trigger and should-not-trigger prompts, has you run them in a real new session, then records what was observed — whether a load record appeared and whether it was model-invoked or user-explicit; a typed /name is not a natural trigger, and an absent load record is reported as not observed, not as not triggered. The same projections are exposed to the agent as a read-only skill_review tool, which returns a structured review packet and says what it omitted when the packet will not fit whole. The inspector, the evidence chain and these reviews are computed on this machine from the definition and the session receipts; the Chinese reading version sends the body through the host's model when you ask for it, keeping code fences, paths and frontmatter verbatim and storing the result on this machine. A Skill can also be cloned into a local copy. It does not write back to the source Skill and does not run its scripts.

Image generation with a sidebar gallery page and three agent tools (paint_image, paint_images, edit_painting); preset providers include ModelScope, Google Gemini, OpenAI, SiliconFlow, Volcengine Seedream, Alibaba DashScope, xAI and Zhipu, plus any OpenAI-compatible endpoint, each with its own proxy setting.