Plugins

Qwen Token Plan CN Responses provider for DeepSeek Harness via the OpenAI Responses API, with Qwen server-side tools, DSH local function tools, text/reasoning/tool-call multi-turn history, persisted image attachments, and a per-release official model catalog snapshot.

Adds deliverable-file buttons to the end of each agent turn (opens files with the OS default app) and a floating dot rail for jumping between conversation turns.

DSH Web GUI 会话问题导航条：聊天区左缘一问一杠，随滚动高亮当前问题、悬停显示问答预览气泡、点击把该问平滑滚到视口顶（未渲染的更早历史自动翻页加载），无需改动 dsh 本体源码

An agent preset that turns a one-line idea into requirements, tech plan and maintenance baseline — every step is restated and confirmed by you before proceeding. No scenario presets: 73 atomic skills (6 master skills + a verify dispatcher + 66 sub-skills) load by context.

Ask a sidebar assistant about selected assistant text: an ephemeral right-side panel answers from a snapshot of the main conversation's LLM history with streaming Markdown, collapsible thinking and multi-turn follow-ups, without touching the main conversation.

Native desktop alerts when dsh needs the human — approvals, questions, plan reviews, turn completion, goals, errors and workflow runs — plus a Settings tab for every option, hot-applied.

Collapses each settled turn's process rows into one expandable disclosure in the DeepSeek Harness Web conversation while keeping the final answer visible.

Turns the dsh web UI into a native Windows desktop window, with a system tray, autostart option and a protocol file that the bundled client reads.

Turns attached files into versioned evidence: `search_documents` builds a private local index (SQLite FTS5 after a startup capability probe, dependency-free JS fallback otherwise) and returns compact evidence blocks carrying an exact coordinate — PDF page, PPTX slide, text/DOCX line range, or quoted XLSX `Sheet!Range` — which `read_document` expands only while the content version still matches. Contiguous CJK runs are indexed as overlapping bigrams and queried as phrases, so word order is preserved; uploaded raster images take the native vision attachment path instead.

Session code collector for DeepSeek Harness: a composer-side panel gathering every code block the model produced in the current session, with by-language / by-file views, per-file version history, live search, per-entry copy, copy-all, download-all as Markdown, and click-to-jump back to the exact turn (Alt+Shift+C).

Zero-config per-turn git checkpoints: after every completed agent turn in a git workspace, stages all changes into one tagged \[turnsnap] commit (serialized queue, sandbox-policy aware), so any turn is a rollback point.

Lets the AI insert streaming-generated UI into the conversation or a side panel, written in TSX — already one of the model's native languages, so there is no JSON schema to teach it. Rendered without flicker, state kept across frames with no hard remount, and a click in the card driving the next chat turn. Inline is where streaming matters most; the side Canvas is the more agentic shape. Cards get agentic capabilities of their own — run bash, curl, call the model, read and write files.

Memory triage for DeepSeek Harness: at the end of each turn it decides which entries to keep and which to archive, routes project-specific ones into a project bucket while global ones stay in the main store, merges near-duplicate entries, and falls back to an LLM when the rules are unsure — automatic actions are recoverable archives only, never deletions.

A source library for DeepSeek Harness: 72 feeds on an hourly timer, a reader with transcripts and translation, and a publisher that turns what it collects into a podcast, a digest, or a report

Turn codebase smells into an executable, priority-ordered refactoring plan: file-length / deep-nesting / function-length / god-object heuristics plus a staged runbook (read-only, no auto-rewrite).

Vision browser tool: drives real Chrome over CDP with browser-use and reads the page with deepseek-v4-flash-vision-exp, so canvas text, text baked into images and values in rendered charts are readable; returns JSON validated against a caller-supplied schema and reports per-run token cost.

Edit an earlier question and run again from that point in the DeepSeek Harness web GUI: forks a child session whose history ends before the edited turn and prefills the composer, leaving the original conversation untouched

Time-traveling stream rules for DeepSeek Harness — project rules stay dormant until a regex matches the live token stream, then the request is aborted, the rule injected as a system reminder, and the request retried from the same point. Zero per-turn cont

Periodic prompt injection for the DeepSeek Harness Web GUI: every N model invocations (conversation turns and tool-call steps both count) the plugin injects a user-configured prompt as a model-visible user message; the first injection happens at conversat

Freeze a working sequence into a hosted production line. Turn a sequence of tool calls that already works into a named pipeline that runs on our servers on a schedule, with a receipt per run — plus recurring watch tasks that notify you only when a result

Fork any session into a different agent preset from the conversation header: a preset picker creates a new child session mounted on the chosen preset, inheriting the source session's completed turns.

LiteLLM /v1/search-backed web search provider for the DeepSeek Harness web capability seam (ctx.web). One search is one HTTP request — no model turn.

DeepSeek Harness plugin: a vision worker over Cloudflare Workers AI (@cf/moonshotai/kimi-k2.6) that routes image requests from text-only callers, returns a versioned righthand.vision.v1 envelope, and supports follow-up questions.

LLM-as-a-Verifier for dsh with the Best-of-N conversation mode built in: rank N candidates with a fine-grained verifier (expected grade over the logprob distribution), on demand via the verify tool or automatically on every turn of a Best-of-N session. In